Cancers 15 03981 v2

cancers
Systematic Review
Deep Learning for Lung Cancer Diagnosis, Prognosis and
Prediction Using Histological and Cytological Images:
A Systematic Review
Athena Davri 1, * , Effrosyni Birbas 2 , Theofilos Kanavos 2 , Georgios Ntritsos 3,4 , Nikolaos Giannakeas 4, * ,
Alexandros T. Tzallas 4 and Anna Batistatou 1
1 Department of Pathology, Faculty of Medicine, School of Health Sciences, University of Ioannina,

45500 Ioannina, Greece; abatista@uoi.gr
2 Faculty of Medicine, School of Health Sciences, University of Ioannina, 45110 Ioannina, Greece;
faybirbas@gmail.com (E.B.); kanavosus@gmail.com (T.K.)
3 Department of Hygiene and Epidemiology, Faculty of Medicine, School of Health Sciences,
University of Ioannina, 45110 Ioannina, Greece; gntritsos@uoi.gr
4 Department of Informatics and Telecommunications, University of Ioannina, 47100 Arta, Greece;
tzallas@uoi.gr
* Correspondence: athinaadav@outlook.com (A.D.); giannakeas@uoi.gr (N.G.)
Simple Summary: Lung cancer is one of the most common and deadly malignancies worldwide.
Microscopic examination of histological and cytological lung specimens can be a challenging and
time-consuming process. Most of the time, accurate diagnosis and classification require histochemical
and specific immunohistochemical staining. Currently, Artificial Intelligence-based methods show
remarkable advances and potential in Pathology and can guide lung cancer diagnosis, subtyping,
prognosis prediction, mutational status characterization, and PD-L1 expression estimation, perform-
ing with high accuracy rates. This systematic review aims to provide an overview of the current
advances in Deep Learning-based methods on lung cancer by using histological and cytological
images.
Citation: Davri, A.; Birbas, E.;
Kanavos, T.; Ntritsos, G.; Giannakeas,
N.; Tzallas, A.T.; Batistatou, A. Deep
Abstract: Lung cancer is one of the deadliest cancers worldwide, with a high incidence rate, especially
Learning for Lung Cancer Diagnosis, in tobacco smokers. Lung cancer accurate diagnosis is based on distinct histological patterns com-
Prognosis and Prediction Using bined with molecular data for personalized treatment. Precise lung cancer classification from a single
Histological and Cytological Images: H&E slide can be challenging for a pathologist, requiring most of the time additional histochemical
A Systematic Review. Cancers 2023, and special immunohistochemical stains for the final pathology report. According to WHO, small
15, 3981. https://doi.org/10.3390/ biopsy and cytology specimens are the available materials for about 70% of lung cancer patients
cancers15153981 with advanced-stage unresectable disease. Thus, the limited available diagnostic material necessi-
Academic Editors: Rune Matthiesen tates its optimal management and processing for the completion of diagnosis and predictive testing
and Akiteru Goto according to the published guidelines. During the new era of Digital Pathology, Deep Learning
offers the potential for lung cancer interpretation to assist pathologists’ routine practice. Herein,
Received: 29 June 2023
we systematically review the current Artificial Intelligence-based approaches using histological and
Revised: 27 July 2023
cytological images of lung cancer. Most of the published literature centered on the distinction between
Accepted: 3 August 2023
Published: 5 August 2023
lung adenocarcinoma, lung squamous cell carcinoma, and small cell lung carcinoma, reflecting the
realistic pathologist’s routine. Furthermore, several studies developed algorithms for lung adenocar-
cinoma predominant architectural pattern determination, prognosis prediction, mutational status
characterization, and PD-L1 expression status estimation.
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland. Keywords: lung cancer; histopathology; histology; cytology; PD-L1; Digital Pathology; artificial
This article is an open access article intelligence; deep learning; convolutional neural networks; CNN
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Cancers 2023, 15, 3981. https://doi.org/10.3390/cancers15153981 https://www.mdpi.com/journal/cancers

Cancers 2023, 15, 3981 2 of 33
1. Introduction
Lung cancer is one of the most prevalent cancers worldwide, characterized by a high
mortality rate, reaching up to 18% of total cancer-related deaths, with cigarette smoking
being the leading cause [1]. Lung cancer is a heterogeneous disease, mainly classified
as non-small cell lung carcinoma (NSCLC) and small cell lung carcinoma (SCLC) [2].
NSCLC constitutes the majority of lung cancer cases (85%) and is further classified into
adenocarcinoma (ADC), squamous cell carcinoma (SCC), and large cell carcinoma (LCC),
while the remaining 15% accounts for SCLC, which is characterized by neuroendocrine
differentiation.
In the era of personalized medicine, lung cancer diagnosis and accurate classification
strongly rely on cytological and histological subtyping by microscopic evaluation with
standard histochemical stains and ancillary immunohistochemical staining [3]. Molecular
testing is also necessary for personalized therapeutic targeting and monitoring for patients’
stratification to targeted therapy and immunotherapy [4,5]. According to published guide-
lines by the College of American Pathologists, the International Association for the Study
of Lung Cancer, and the Association for Molecular Pathology, patients with advanced lung
cancer with an ADC component should be tested for epidermal growth factor receptor
(EGFR) mutations, anaplastic lymphoma kinase (ALK) and ROSproto oncogene 1 (ROS-1)
rearrangements, v-Raf murine sarcoma viral oncogene homolog B (BRAF) Val600Glu
(BRAFV600E), Ret Proto-Oncogene (RET) rearrangements, mesenchymal-epithelial tran-
sition (MET) exon 14 skipping mutations, Kirsten rat sarcoma (KRAS) mutations, and
neurotrophic tyrosine kinase receptor fusions (NTRK1-3) [2,6]. Advanced-stage non-
neuroendocrine carcinomas should be tested for programmed cell death ligand 1 (PD-L1)
expression status as patients with a PD-L1 Tumor Proportion Score (TPS) ≥ 50% are eligible
for first-line treatment with the anti-PD-L1 therapy, pembrolizumab. Immunohistochemical
assays are available for PD-L1 and ALK expression status detection [7–10]. Currently, reflex-
ordered testing for lung cancer is gaining ground, underlining the necessity of collaboration
between pathologists and oncologists. Although reflex testing is not feasible to perform in
many laboratories, it can provide additional valuable information, detect rare molecular
alterations, and minimize testing turnaround time [3,11].
In the last decade, Deep learning (DL) approaches, including mostly Convolutional
Neural Networks (CNNs), are increasingly valuable in Pathology. Limitations concerning
the shortage of pathologists worldwide, subjectivity in diagnosis, and intra- and inter-
observer variability could be overcome with the aid of DL models. Recent advances in
lung cancer pathology leverage image analysis potential for cancer diagnosis from hema-
toxylin and eosin (H&E) whole slide images (WSIs) [12,13]. Considering that small biopsy
and cytology specimens are the available material for 70% of lung cancer patients with
advanced unresectable disease, DL methods could guide the diagnosis with high accuracy,
minimizing the need for additional special stains required for differential diagnosis and
preserving the already limited material for molecular testing [2,14,15].
In this review, we systematically outline the current implications of DL algorithms for
lung cancer diagnosis, prognosis, and prediction using both histological and cytological
images. We further summarized the extracted data into distinct categories based on the
classification problem, presenting for each study the dataset details, the employed technical
method and methodology, as well as the performance metrics. The different categories
have been structured to be informative for both pathologists and cytologists, can provide
a detailed analysis and a comprehensive guide of the existing DL applications for lung
cancer, and offer valuable information to researchers for further study.
2. Materials and Methods

The systematic review followed the recommendations of the Preferred Reporting
Items for Systematic Reviews and Meta-Analyses (PRISMA) [16]. The protocol has not
been registered.
Cancers 2023, 15, 3981 3 of 33
2.1. Search Strategy

We systematically searched PubMed from inception to 31 March 2023 for primary
studies developing a DL model for the histopathological or cytopathological interpretation
of malignant lesions in lung specimens. For this purpose, we used the following algorithm:
(convolutional neural network* OR CNN OR deep learning) AND lung AND (cancer
OR neoplas* OR carcinoma OR adenocarcinoma OR malign* OR tumor*) AND (histolog*
OR histopatholog* OR eosin OR slide* OR section* OR immunohistochem* OR biop* OR
microscop* OR cytolog* OR cytopatholog* OR immunocytochem*).
2.2. Study Eligibility Criteria

Eligible articles were considered on the basis of the following criteria. We included
studies, the aim of which was to develop at least one DL model for the histopathological or
cytopathological assessment of malignant lesions in lung specimens. Eligible applications
of the DL models included diagnosis, subtyping, prognosis evaluation, characterization
of mutational status, and prediction of PD-L1 expression. Articles that presented in vitro
models or used non-histological/cytological data as well as reviews/meta-analyses, edi-
torials, letters, and invited opinions, were excluded. In addition, articles not available in
English and those referring to organisms other than humans were deemed ineligible.
2.3. Study Selection

All citations collected by the previously mentioned methodology were independently
screened by four authors, who were properly trained before the process started, using
the Rayyan web application [17]. Three of these researchers were scientifically capable of
evaluating the medical aspect of the query, and one of them was a CNN expert, able to
assess the technical part. During the screening period, the researchers met regularly to
discuss disagreements and continue training. Conflicts were resolved by consensus. The
full texts of potentially eligible articles were later retrieved for further evaluation.
2.4. Data Extraction

To facilitate the data extraction process, we specially designed a spreadsheet form,
which all researchers could access to import data from all the eligible articles. From each
paper, we manually extracted information regarding the first author, year of publication,
aim of medical research, technical method, classification details, dataset, and performance
metrics.
3. Results
Our systematic search returned 357 articles, 127 of which were selected for full-text
assessment. Ultimately, 96 articles met our criteria of eligibility and were included in our
study. A detailed description of the study selection process can be found in the PRISMA
flow diagram presented in Figure 1.
At first, the included studies were divided, based on the used dataset, into histology
and cytology sections. Further categorization of the histology section into diagnosis,
lung cancer classification, NSCLC subtyping, ADC predominant architectural patterns
classification, prediction of prognosis and survival, and prediction of molecular alterations
subsections was made based on the classification problem. Studies performing DL for the
PD-L1 expression status estimation were summarized in a particular section.
Cancers 2023, 15, 3981Cancers 2023, 15, 3981 4 of 33 4
Figure of
Figure 1. Flow diagram 1. Flow diagram
the study of theprocess
selection study selection process
illustrating illustrating
the systematic the systematic
search search and sc
and screening
ingthe
strategy along with strategy
number along with the number
of included of included
and excluded and excluded studies.
studies.
3.1. Histology 3.1. Histology

3.1.1. Diagnosis3.1.1. Diagnosis
Jain et al. usedJainDLetarchitectures for detectingforlung
al. used DL architectures cancerlung
detecting from histopathological
cancer from histopathologica
images pre-processed for size, normalization,
ages pre-processed and noiseand
for size, normalization, removal
noise [18].
removal Three
[18].datasets
Three datasets w
were included achieving high-performance
included achieving rates with
high-performance an accuracy
rates of over 97%.
with an accuracy Jiao
of over et al.
97%. Jiao et al.
proposed a rapid and efficient method for tumor classification called Deep
posed a rapid and efficient method for tumor classification called Deep Embedding-bEmbedding-
based Logistic Logistic
Regression (DELR)(DELR)
Regression [19]. DELR was was
[19]. DELR applied in three
applied different
in three different datasets
datasets (colore
(colorectal, lung, and breast cancer) and achieved
lung, and breast cancer) and achieved an area under the curve (AUC)
area under the curve (AUC) of over of 0.95 fo
over 0.95 for allthree
three datasets.
datasets. In lung
In lung cancer,
cancer, the dataset
the dataset consisted
consisted of 338ofregions
338 regions of (ROIs)
of interest
interest (ROIs),cluding
including
ADC ADCand and
SCCSCC images.
images. Moreover,
Moreover, KanavatiKanavati et al. trained
et al. trained a CNN ato disting
CNN to distinguish
lung lung carcinoma
carcinoma from non-neoplastic
from non-neoplastic tissue tissue
based onbased
the on the EfficientNet-
EfficientNet-B3 architecture
B3 architectureAfter
[20]. training,
After training,
the CNN the CNN
was was
tested ontested on four independent
four independent datasets and datasets
attained an AU
and attained anmore AUC of more
than than 0.97, demonstrating
0.97, demonstrating its feasibility its feasibility of generalization.
of generalization. Multiple Instance Lear
Multiple Instance Learning
(MIL) (MIL) was
was employed for employed for the sametask
the same classification classification
without thetask needwithout
for manual ann
the need for manual annotations
tions by pathologistsby [21].
pathologists [21]. Aclassification
A multi-organ multi-organusing
classification using
weakly supervised lear
weakly supervisedwaslearning
performed wasbyperformed
Tsuneki etby al. Tsuneki et al. on transbronchial
on transbronchial lung biopsy WSIs lung[22].
biopsy
The AUC va
WSIs [22]. The AUC
of thevalues of the three
three different different
balanced balanced
training training
datasets datasets
collected fromcollected
medicalfrom
institutions w
0.879–0.933
medical institutions (Table 1). (Table 1).
were 0.879–0.933
Cancers 2023, 15, 3981 5 of 33
Table 1. Characteristics of studies developing models for lung cancer diagnosis on histological
images.
1st Author, Year Technical Method Classification Dataset Performance Metrics

15,000 images from LZ2500 Acc: 97.10% in LZ2500
Kernel PCA combined dataset, 215 tiles from the dataset, 98.00% in
Jain, 2022 Binary:
with Fast Deep Belief NLST dataset and NLST dataset and
[18] cancerous/normal cells
Neural Network 1634 images from NCI 97.50% in NCI
Genomic dataset Genomic dataset
Custom Architecture
Civit-Masot, 2022 Binary: 15,000 images from LC25000 Overall Acc: 99.69%
with 3 Convolution and
[23] benign/malignant dataset using 50 epochs
2 dense layers
Weakly supervised 8896 slides from Mita, Wajiro, Acc: 85.30%
Tsuneki, 2022 Binary:
learning using Shinkuki, Shinkomonji and Se: 88.50%
[22] ADC/non-ADC
EfficientNet-B1 Shinmizumaki Hospitals Sp: 82.50%
Standard Attention
Acc: 90.00%
Moranguinho, MIL approach using
Binary: 3220 samples from TCGA AUC: 0.94
2021 attention module and
tumor/normal dataset Gated Attention
[21] Grad-Cam algorithm
Acc: 91.20%
AUC: 0.95
DELR: deep feature
extraction and active
Jiao, 2021 Binary:
learning for sample 338 ROIs from TCGA dataset AUC: >0.95
[19] tumor/non-tumor
selection in Logistic
Regression
4204 WSIs from Kyushu
Medical Center, 500 WSIs
Weakly supervised Weakly supervised
Binary: from International University
Kanavati, 2020 learning employing AUC: 0.97–0.98
carcinoma/non- of Health and Welfare, Mital
[20] EfficientNet-B3 Fully supervised
neoplastic Hospital, 680 WSIs from
architecture AUC: 0.88–0.96
TCGA dataset and 500 WSIs
from TCIA dataset
Abbreviations: Acc, Accuracy; AUC, Area Under the Curve; DELR, Deep Embedding-based Logistic Regression;
MIL, Multiple Instance Learning; NCI, National Cancer Institute; NLST, National Lung Screening Trial; PCA,
Principal Component Analysis; ROI, Region Of Interest; Se, Sensitivity; Sp, Specificity; TCGA, The Cancer Genome
Atlas; TCIA, The Cancer Imaging Archive; WSI, Whole Slide Image.
3.1.2. Lung Cancer Classification

A common classification problem among all papers included refers to lung cancer
tissue classification into the main categories of ADC, SCC, and SCLC according to WHO
guidelines. Kanavati et al. developed a CNN for lung cancer subtyping (ADC, SCC, SCLC,
and non-neoplastic tissue) trained on transbronchial biopsy (TBLB) images with mainly
poorly differentiated carcinomas [24]. Their model was tested on four validation cohorts
(one with TBLB specimens and three with surgical resections), performing with AUC over
0.9 on all datasets. The same classification problem was employed with weakly supervised
CNN, including a smaller dataset from hospital archives and The Cancer Genome Atlas
(TCGA) database [25]. The model had an overall accuracy of 97.3% and achieved an AUC
of 0.856 in the TCGA cohort. In addition, three common CNNs (Inceptionv3, VGG-16,
InceptionResNetV2) were used for lung cancer classification on TMAs. The InceptionV3
model achieved the highest performance; however, many cases of ADC and SCC were
misclassified [26]. In a retrospective study by Yang et al., a six-type classifier model was
designed for lung cancer (ADC, SCC, SCLC) as well as other lung diseases (pulmonary
tuberculosis, organizing pneumonia) subtyping on H&E-stained slides [27]. The proposed
classification task achieved great performance and consistency with experienced patholo-
gists. In a different study, Yang et al. introduced a CNN for subtyping lung cancer in five
Cancers 2023, 15, 3981 6 of 33
classes, namely ADC, SCC, SCLC, large cell neuroendocrine carcinoma (LCNEC), and non-
tumor [28]. The customized model performed similarly or better than the pre-trained ones,
although existing limitations of the study, such as the use of patches instead of WSIs and
the limited dataset, resulted in moderate classification accuracies. Likewise, Kosaraju et al.
applied a novel DL framework for classifying ADC, SCC, SCLC, and LCNEC, achieving an
AUC of 0.96 [29]. The studies of Yang and Kosaraju et al. were the only ones that included
LCNEC in the classifiers representing the realistic diagnostic practice for a pathologist. Ilié
et al. applied a DL algorithm for distinguishing SCLC, LCNEC, and atypical carcinoid
(AC) [30]. A number of 150 H&E WSIs were included, and the model was in great agree-
ment when compared to expert and general pathologists, achieving an AUC of 0.93. Lastly,
in their recent study, Chen et al. proposed an immunohistochemical phenotype prediction
system for upgrading the classification of lung cancer into ADC, SCC, and SCLC [31].
The WSI-based Immunohistochemical Feature Prediction System (WIFPS) discriminated
lung cancer types on H&E slides based on the positive or negative expression scoring of
characteristic biomarkers for each class (ADC: TTF-1, CK7, and Napsin-A; SCC: CK5/6,
p40, and p63; SCLC: CD56, Synaptophysin, Chromogranin A, and TTF-1). The agreement
between the WIFPS model and pathologists achieved high to almost perfect consistency
(Cohen’s kappa value of 0.7903–0.8891) in validation sets and the AUC in surgical and
biopsy images was over 0.8 in all validation cohorts. In addition, ALK prediction status
achieved an AUC of 0.917; however, programmed cell death protein 1 (PD-1), PD-L1, KRAS,
and EGFR status did not reach high performance (Table 2).
Table 2. Characteristics of studies developing models for lung cancer classification and non-small
cell lung cancer subclassification on histological images.

Lung cancer classification
ResNet-152
5-class: ADC/SCC/SCLC/ 205 WSIs from Gyeongsang Novel CNN model
Yang, 2022 VGG-19
LCNEC/ National University Acc: 75.03%
[27] Xception
non-tumor Hospital Macro-average AUC: 0.90
NASNetLarge
Binary: 1101 WSIs from First Tissue classification
EfficientNet-B5 normal/tumor tissue Affiliated Hospital of Sun Micro-average AUC: 0.98
WSI-based IHC feature Binary: Yat-sen University, Macro-average AUC: 0.99
Chen, 2022
prediction system: a novel negative/positive Shenzhen People’s Hospital Biomarkers expression
[31]
DL model based on expression of biomarkers and Cancer Center of AUC: 0.53–0.95
EfficientNet-B5 3-class: Guangzhou Medical 3-class
ADC/SCC/SCLC University Acc: 90.00%
DEEP-HIPO: 113 WSIs from Gyeongsang
Kosaraju, 2022 two magnifications (20× 4-class: ADC/SCC/SCLC/ National University
AUC: 0.96
[28] and 5×), based on LCNEC Hospital and 657 ADC WSIs
CAT-NET with 19 layers from TCGA dataset
150 NET and 25 poorly Acc: 98.00% (95% CI:
differentiated ADC WSIs 93.70–1.00%)
Ilié, 2022 4-class: SCLC/LCNEC/AC/
HALO-AI from Laboratory of Clinical AUC: 0.93
[30] poorly differentiated ADC
and Experimental Pathology F1-score: 0.99 (95% CI:
of Nice University Hospital 0.94–1.00)
EfficientNet-B5-based deep
learning model
1059 WSIs from First
6-class: AUC: 0.97 in Sun Yat-sen
Affiliated Hospital of Sun
EfficientNet-B5-based and ADC/SCC/SCLC/ University dataset 1, 0.92 in
Yang, 2021 Yat-sen University, 212 WSIs
ResNet-50-based DL pulmonary Sun Yat-sen University dataset
[26] from Shenzhen People’s
model tuberculosis/organizing 2, 0.96 in Shenzhen People’s
Hospital, and 422 WSIs from
pneumonia/normal lung Hospital dataset and 0.98 in
TCGA dataset
TCGA dataset
ICC: >0.87
Cancers 2023, 15, 3981 7 of 33
Table 2. Cont.

Independent TBLB dataset of
1723 WSIs from Kyushu 83 indeterminate WSIs
Medical Center, 500 WSIs AUC: 0.99
Kanavati, 2021 Combination of 4-class: ADC/SCC/SCLC/
from Mita Hospital and 905 1 independent TBLB and
[24] EfficientNet-B1 and RNN non-neoplastic
NSCLC WSIs from TCGA 3 independent surgical resection
dataset datasets of 2407 WSIs
AUC: 0.94–0.99
Acc: 97.30% in Sun Yat-sen
939 WSIs from Sun Yat-sen
University Cancer Center
Wang, 2020 4-class: ADC/SCC/SCLC/ University Cancer Center
Modification of VGG-16 dataset and 82.00% in TCGA
[25] normal and 500 WSIs from TCGA
dataset
dataset
AUC: 0.86 in TCGA dataset
InceptionV3 with weights trained
on the training dataset
VGG-16 270 cases from Institute of
Kriegsmann, 2020 4-class: ADC/SCC/SCLC/ Acc: 86.00% in validation
InceptionV3 Pathology, University Clinic
[29] skeletal muscle dataset using 20 epochs and
InceptionResNetV2 Heidelberg
85.00% in validation dataset
using 50 epochs
NSCLC subclassification
In testing phase using 80% of the
5-class: dataset for training and 20% for
MPADL-LC3 algorithm
Mengash, 2023 lung ADC/lung SCC/lung 25,000 images from LC25000 testing
based on MobileNet and
[32] benign tissue/colon dataset Acc: 99.42% for lung ADC,
DBN
ADC/colon benign tissue 99.28% for lung SCC and
99.30% for lung benign tissue
ANN with fusion features of
5-class:
ANN VGG-19 and handcrafted
Al-Jabbar, 2023 lung ADC/lung SCC/lung 25,000 images from LC25000
GooLeNet Acc: 99.60% for lung ADC,
[33] benign tissue/colon dataset
VGG-19 99.80% for lung SCC and
ADC/colon benign tissue
A novel Overall metrics
Wang, 2023 993 WSIs from TCGA
multiplex-detection-based Binary: ADC/SCC Acc: 90.52%
[34] dataset
MIL model AUC: 0.96
HistoROI: a
Patil, 2023 1034 WSIs from TCGA
ResNet18-based 6-class Binary: ADC/SCC AUC: 0.93
[35] dataset
classifier
Average overall metrics
5-class: lung ADC/lung Acc: 99.94%
El-Ghany, 2023 SCC/lung benign 25,000 images from LC25000 Sp: 99.96%
ResNet 101
[36] tissue/colon ADC/colon dataset Pr: 99.84%
benign tissue Re: 99.85%
F1-score: 99.84%
2071 WSIs from 435 patients Five-fold cross-validation
from the CPTAC dataset, Acc: 91.20% ± 2.50%
Zheng, 2022 Graph-based modules 3-class: 2082 WSIs from 996 patients AUC: 0.98
[37] with ResNet ADC/SCC/normal from TCGA dataset and External test data
665 WSIs from 345 patients Acc: 82.30% ± 1.00%
from NLST dataset AUC: 0.93
3-class:
3-class
infiltra-
766 lung WSIs from First Acc 98.33%
SE-ResNet-50 with novel tion/microinfiltration/normal
Liu, 2022 Hospital of Baiqiu’en and 5-class
activation function 5-class:
[38] 25,000 images from LC25000 Acc: 99.96%
CroRELU lung ADC/lung
dataset Se: 99.86%
SCC/normal lung/colon
Pr: 99.87%
ADC/normal colon
ShuffleNet, SqueezeNet, 5-class:
Acc: 99.30% for lung ADC,
Attallah, 2022 and MobileNet: lung ADC/lung SCC/lung 25,000 images from LC25000
99.00% for lung SCC and
[39] 3 pre-trained lightweight benign tissue/colon dataset
CNN models ADC/colon benign tissue
Cancers 2023, 15, 3981 8 of 33
Table 2. Cont.

Colour CNN classifier
Overall Acc: 97.11% using
Custom Architecture with
Civit-Masot, 2022 3-class: 15,000 images from LC25000 50 epochs
3 Convolution and 2 dense
[23] ADC/SCC/benign dataset Greyscale CNN classifier
layers
Overall Acc: 94.01% using
50 epochs
DSC: 93.50% for segmenting
A custom architecture
SCC and 89.00% for
consisting of
312 images from 36 patients segmenting ADC
Wang, 2022 5 Convolution and 3 Fully 3 class:
from Qilu Hospital of Acc: 97.80% in classifying SCC
[40] Connected layers along ADC/SCC/normal
Shandong University versus normal tissue and
with a segmentation
100.00% in classifying ADC
branch for up-sampling
versus normal tissue
AUROC: 0.96 at maximum
dataset size for
941 WSIs from TCGA
non-uncertainty quantification
Dolezal, 2022 CNN models based on dataset, 1.306 from CPTAC
Binary: ADC/SCC models
[37] Xception architecture dataset and 190 slides from
AUROC: 0.98 at maximum
Mayo Clinic dataset
dataset size for uncertainty
quantification models
Based on WSIs
Acc: 99.00% in the training
dataset, 87.00% in validation
dataset, 85.00% in the test
dataset, 85.00% in the external
132 slides from Dijon
validation cohort and 75.00%
University Hospital,
in TCGA dataset
Binary: 65 slides from Caen
Le Page, 2021 A novel CNN model Based on virtual TMAs
squamous/non-squamous University Hospital,
[41] based on InceptionV3 Acc: 99.00% in training dataset,
NSCLC 60 slides from TCGA
83.00% in validation dataset,
database and 1 cytological
88.00% in test dataset, 92.00%
pericardium specimen
in external validation cohort
and 83.00% in TCGA dataset
AUC: 0.94 in external
validation cohort and 0.77 in
TCGA dataset
LungDIG: Combination of Acc: 87.10%
Wang, 2021 Binary: 988 samples with both CNV
InceptionV3 with AUC: 92.70%
[42] ADC/SCC and histological data
multilayer perceptron F1-Score: 87.60%
ROI localization
F1-score: 0.88
AUC: 0.96
MR-EM-CNN:
Zhao, 2021 Binary: ROI/non-ROI 2125 slides from TCGA NSCLC classification
Hierarchical multiscale
[43] Binary: ADC/SCC dataset Se: 94.74%
features on EM-CNN
Sp: 85.83%
F1-score: 0.90
AUC: 0.96
Validation Sample
Dehkharghanian, 735 WSIs from TCGA
KimiaNet-22: a DL model Pr: 92.00%
2021 Binary: ADC/SCC dataset and 23 WSIs from
based on DenseNet Re: 91.00%
[44] Grand River Hospital
F1-score: 0.91
5-class: lung ADC/lung
Acc: 99.73% for lung ADC,
Toğaçar, 2021 DarkNet-19 combined SCC/lung benign 25,000 images from LC25000
99.74% for lung SCC, 99.98%
[45] with YOLO and SVM tissue/colon ADC/colon dataset
for lung benign tissue
benign tissue
Histology Classifier
Carrillo-Perez, 1420 WSIs and
Acc: 86.03%
2021 Merging ResNet-18 3-class: ADC/SCC/healthy 980 RNA-sequencing data
F1-Score: 83.39%
[46] from TCGA dataset
AUC: 0.95
Cancers 2023, 15, 3981 9 of 33
Table 2. Cont.

Public NSCLC WSI dataset
(TCGA and TCIA)
AUC: 0.96 ± 0.02 using 100%,
131 resection and 110 biopsy
0.95 ± 0.02 using 75% and
NSCLC WSIs from Brigham
Clustering-Constrained- 0.94 ± 0.02 using 50% of cases
and Women’s Hospital,
Lu, 2021 Attention MIL: a novel Binary: in training dataset
993 NSCLC WSIs from
[47] DL-based weakly ADC/SCC Independent NSCLC WSI dataset
TCGA dataset, and 974
supervised model (Brigham and Women’s Hospital)
NSCLC WSIs from TCIA
AUC: 0.94 ± 0.02 using 100%,
dataset
0.92 ± 0.01 using 75% and
0.88 ± 0.02 using 50% of cases
in training dataset
9662 WSIs from
2843 patients from Taipei
Medical University Hospital,
Chen, 2021 MIL combined with 3-class: Taipei Municipal Wanfang AUC: 0.96 for ADC and 0.94
[48] ResNet-50 ADC/SCC/non-cancer Hospital and Taipei Medical for SCC
University Shuang-Ho
Hospital and 532 WSIs from
TCGA dataset
5-class: Testing dataset
Custom CNN architecture
lung ADC/lung Acc: 96.33%
Masud, 2021 consisting of 25,000 images from LC25000
SCC/benign lung Pr: 96.39%
[49] 3 Convolution and 1 Fully dataset
tissue/colon ADC/benign Re: 96.37%
Connected layers
colonic tissue F1-score: 96.38%
InceptionV3, ResNet-50,
VGG-19, MobileNetV2,
Wang, 2021 3-class: NSCLC WSIs from TCGA AUC: 0.98 for ADC, 0.98 for
ShuffleNetV2 and
[50] ADC/SCC/normal dataset SCC and 0.99 for normal
MNASNET on HEAL
Platform
Kobayashi, 2020 A proposed modification 950 patients from Pan-Lung
Binary: ADC/SCC Acc: ~80.00%
[51] to Diet Networks Cancer dataset
Tumor/normal
Xu, 2020 Hierarchical multiscale Binary: tumor/normal 2125 images from TCGA AUC: 1.00
[52] features on EM-CNN Binary: ADC/SCC dataset ADC/SCC
AUC: 0.97
ADC/benign
AUC: 0.95–0.97 in TCGA test
dataset and 0.92–0.94 in ICGC
Binary: ADC/benign
test dataset
Binary: SCC/benign
SCC/benign
Binary: ADC/SCC
AUC: 0.94–0.99 in TCGA test
3-class: terminal respiratory
AlexNet dataset and >0.97 in ICGC test
unit/proximal- 884 WSIs from TCGA
Yu, 2020 GoogLeNet dataset
inflammatory/proximal- dataset and 125 images from
[53] VGGNet-16 ADC/SCC
proliferative ADC ICGC dataset
ResNet-50 AUC: 0.88–0.93 in TCGA test
transcriptome subtype
dataset and 0.73–0.86 in ICGC
4-class: classi-
test dataset
cal/basal/secretor/primitive
ADC transcriptome subtype
SCC transcriptome subtype
AUC: 0.77–0.89
SCC transcriptome subtype
AUC: ~0.70
Acc: 90.50% using 20% labeled
Graph temporal patients, 91.00% using 35%
2904 NSCLC image patches
Shi, 2019 ensembling: a novel labeled patients, 91.10% using
Binary: ADC/SCC from WSIs of 42 patients
[54] semi-supervised CNN 50% labeled patients and
from TCGA
model based on AlexNet 94.00% using all labeled
patients
Cancers 2023, 15, 3981 10 of 33
Table 2. Cont.

CNN-basic
InceptionV3-Last
layer-4000 steps InceptionV1-Fine tune
InceptionV3-Last 1273 images from TMAD Acc: 92% for TMAD images,
Khosravi, 2018
layer-12,000 steps Binary: ADC/SCC and 3149 from TCGA 100% for TCGA intra-images
[55]
InceptionV1-Fine tune dataset and 83% for TCGA
Inception-ResNetV2-Last inter-images
layer
InceptionV3-Fine tune
1634 WSIs from Genetic
Binary
Data Commons database
Coudray, 2018 Binary: tumor/normal AUC: 0.99
InceptionV3 and 340 slides from New
[56] 3-class: normal/ADC/SCC 3-class
York University Langone
AUC: 0.97
Medical Center
14 different combinations
of EM-based MIL
Hou, 2016 3-class: ADC/SCC/ADC 718 WSIs from 641 patients
approach with CNN and Acc: 79.80%
[57] with mixed subtypes from TCGA dataset
multiclass logistic
regression or SVM
Abbreviations: AC, Atypical Carcinoid; Acc, Accuracy; ADC, Adenocarcinoma; ANN, Artificial Neural Network;
AUC, Area Under the Curve; AUROC, Area Under the Receiver Operating Characteristic; CI, Confidence
Interval; CNN, Convolutional Neural Network; CNV, Copy Number Variation; CPTAC, Clinical Proteomic
Tumor Analysis Consortium; DSC, Dice Similarity Coefficient; EM, Expectation-Maximization; ICC, Interclass
Correlation Coefficient; ICGC, International Cancer Genome Consortium; LCNEC, Large Cell Neuroendocrine
Carcinoma; MIL, Multiple Instance Learning; MR, Multi-Resolution; NET, Neuroendocrine Tumor; NLST, National
Lung Screening Trial; NSCLC, Non-Small Cell Lung Cancer; Pr, Precision; Re, Recall; RNA, Ribonucleic Acid;
RNN, Recurrent Neural Network; ROI, Region Of Interest; SCC, Squamous Cell Carcinoma; SCLC, Small Cell
Lung Cancer; Se, Sensitivity; Sp, Specificity; SVM, Support Vector Machine; TBLB, Transbronchial Lung Biopsy;
TCGA, The Cancer Genome Atlas; TCIA, The Cancer Imaging Archive; TMA, Tissue Microarray; TMAD, Tissue
Microarray Database; WSI, Whole Slide Image.
3.1.3. NSCLC Subtypes Classification

The diagnosis between ADC and SCC from a single H&E slide from a small biopsy
or cytological material can be challenging. Thus, for precise diagnosis, additional staining
for immunohistochemical biomarkers, such as TTF-1, CK5/6, CK7, pan keratin, p40, and
p63, and histochemical stains, such as periodic acid-Schiff (PAS), must be performed. Sev-
eral studies have addressed binary classification problems concerning NSCLC subtyping
from H&E slides for an accurate and fast diagnosis. The majority of these mainly include
ADC and SCC WSIs, mostly from the TCGA dataset, whereas the classification task is
performed by a CNN or a combination of the state-of-the-art CNN architectures with
varying approaches and techniques [37,47,51–56,58,59]. Moreover, NSCLC subtyping was
combined with genomic data, namely copy number variations (CNVs), from TCGA [42].
The authors demonstrated that their proposed LungDIG model could be of great impor-
tance not only for ADC and SCC diagnosis but also for stratifying patients for targeted
therapies, as the performance metrics of the model were higher when WSI and CNV data
were combined compared to when WSI or CNV features were used alone. Zhao et al.
developed a weakly supervised DL model to localize ROIs on WSIs (AUC of 0.9602) and
then accurately subtype NSCLC into ADC and SCC with high sensitivity and specificity
rates (0.9474 and 0.8583, respectively) [43]. In another study extracting prominent deep
features (DFs) for each histopathological image, classification accuracy was better, and the
authors identified 15 DFs with the ability to classify lung cancer with an accuracy of over
85% [44]. The generalizability of the model was feasible in distinguishing ADC from SCC
on 21 non-pulmonary carcinomas; however, classification accuracy reached 56% in the ex-
ternal validation cohort. Hou et al. performed a classification task of NSCLC subtyping into
ADC, SCC, and ADC with mixed subtypes [57]. Their proposed framework was trained
and tested on a TCGA dataset with a classification accuracy of 0.798. Masud et al. designed
a classification framework for diagnosing lung and colon cancer from histopathological
Cancers 2023, 15, 3981 11 of 33
images from the LC25000 dataset [49]. The model achieved a peak classification accuracy of
96.33%; however, the lung ADC class had a higher misclassification rate. The same problem
using the LC25000 dataset was employed by other authors, with an overall accuracy of
99% [32,33,36,39]. DarkNet-19 model reached accuracies of 97.57%, 99.87%, and 97.73%
in classifying ADC, benign, and SCC images, respectively, while the overall accuracy of
the model was 99.69% [45]. Likewise, Civit-Masot et al. employed Explainable Artificial
Intelligent (AI) Technologies [23]. Liu et al. used AI along with activation function for
cancer infiltration screening on histopathological images [38]. Their method was further
utilized for lung cancer classification (ADC and SCC) using the LC25000 database, present-
ing good generalization ability. In a more recent study, Liu et al. proposed a novel method
for automated detection of lung ADC infiltration using 780 images with sensitivity and
specificity of 93.10% and 96.43%, respectively [60]. Utilizing a combination of molecular
and histological data (gene expression data and WSIs, respectively) as input for NSCLC
classification, Carrillo Perez et al. demonstrated that the fusion model could provide robust
information for decision-making to targeted therapies [46]. Wang et al. proposed a platform
for the automated classification of NSCLC into ADC, SCC, and normal regions as well as
for prediction of mutational status of 10 frequently mutated genes in ADC [50]. The model
predicted with an AUC of 0.824 the EGFR mutational status on ADC H&E WSIs. Similarly,
a model for NSCLC subtyping (ADC, SCC, normal regions) achieved an AUC of 0.97 [56].
The authors trained the model to predict the mutational status in lung ADC slides. Of the
ten frequently mutated genes in ADC, STK11 and KRAS had the highest AUC (0.845 and
0.814, respectively). An annotation-free DL method for the subtyping of NSCLC slides
achieved high performance for ADC and SCC (AUC of 0.9594 and 0.9414, respectively)
and could be employed in clinical practice as it overcomes the time-consuming process
of annotations and limitations concerning the capacity/memory of WSIs [48]. Wang et al.
developed a DL model to perform cancer lesion region segmentation and histological
subtype classification on ADC and SCC slides [40]. The model showed high classifica-
tion performance metrics (accuracy was 100% and 95.1%, sensitivity was 95.0 and 100.0%,
and specificity was 95.2 and 100.0% for SCC and ADC classification tasks, respectively).
Classification of transcriptomic lung ADC (bronchioid, squamoid, and magnoid) and/or
SCC (primitive, classic, secretory, and basal) subtypes was performed by Yu and Antonio
et al. [53,61]. In the first study, classification was performed on both ADC and SCC, result-
ing in a significant correlation between the transcriptomic subtype and the histopathology
classification scores and achieving AUC of 0.771–0.892 and approximately 0.7 for ADC and
SCC, respectively, with the employment of four CNNs. In the study of Antonio et al., ADC
transcriptome subtype classification resulted in a classification accuracy of 98.9%. Lastly, Le
Page et al. tried to distinguish squamous from non-squamous lung carcinoma from initial
cytology and small biopsy specimens [41]. Their model performed with good classification
accuracy, while the accuracy was slightly increased in the external validation cohorts when
tissue microarrays (TMAs) were selected (accuracy rates of 0.78 in biopsies versus 0.82 in
TMAs). Finally, two recent studies performed a binary classification between ADC and
SCC using over 900 WSIs from TCGA and achieving an AUC of over 0.90 [34,35] (Table 2).
3.1.4. Lung ADC Predominant Architectural Patterns Classification

ADC cases exhibit various histological patterns. According to the WHO, there are
five distinct histological subtypes (lepidic, acinar, papillary, micropapillary, and solid) that
must be included in a pathology report when the material is a resection specimen [2]. The
detection of ADC predominant architectural patterns has been the scope of several research
papers (Table 3). The study by Sadhwani et al. performed a classification problem including
six histological subtypes (acinar, lepidic, solid, papillary, micropapillary, cribriform) and
then combined the predicted output with clinical data (smoking status, age, etc.) for tumor
mutational burden (TMB) status prediction [62]. The AUC for ADC predominant architec-
tural patterns classification was 0.93 and 0.92 for TCGA and the external validation cohort,
respectively, while for the TMB status prediction, it was 0.71. Furthermore, a six-class
Cancers 2023, 15, 3981 12 of 33
problem (lepidic, acinar, papillary, micropapillary, solid, benign) for lung ADC histological
subtypes classification in lung ADC WSIs was in moderate agreement with pathologists’
estimations [63]. In a similar study, ADC histological patterns were classified into five
categories (solid, micropapillary, acinar, cribriform, non-tumor) using three different CNN
architectures [64]. The best classification accuracy was 89.24%, while, in the study of Di-
Palma et al., the histological classification of the known five patterns of lung ADC resulted
in a classification accuracy of 94.51% [65]. Xiao et al. created a novel framework combining
CNNs and graph convolutional networks for quantitative estimation of histopathological
growth patterns in lung ADC slides [66]. Another lung ADC subtyping problem was
performed by Sheikh et al. achieving a high accuracy rate of 0.946 and outperforming the
state-of-the-art models [67]. In a different study, Gao et al. collected slides from ADC with
micropapillary patterns and performed a binary classification problem for detecting the
presence of a micropapillary pattern in ADC slides [68]. Maleki et al. investigated how
several possible methodological errors, such as oversampling and data augmentation, can
lead to poor generalizability performance and performed a binary classification task for the
distinction of solid and acinar predominant histologic subtypes in ADC H&E slides [69].
Table 3. Characteristics of studies developing models for the identification of lung adenocarcinoma
predominant architectural pattern.

Supervised model
Pr: 76.20%
Gao, 2022 Binary: micropapillary/non- ADC WSIs from Shandong Re: 88.40%
Inspired by YOLOv5
[68] micropapillary Provincial Hospital Semi-supervised model
Pr: 77.50%
Re: 89.60%
LAD-GCN
5-class: 243 images from 243 patients Acc: 90.35%
Xiao, 2022 GCNs combined with
lepidic/acinar/papillary/ from Shandong Provincial Pr: 86.53–98.34%
[66] VGG-16
micropapillary/solid Hospital Re: 85.80–98.78%
F1-score: 0.86–0.99
Unsupervised deep Acc: 94.60%
5-class: 31 WSIs from
Sheikh, 2022 learning model which Se: 94.10%
lepidic/acinar/papillary/ Dartmouth-Hitchcock
[67] employs stacked Pr: 94.20%
micropapillary/solid Medical Center
autoencoders F1-score: 0.94
Four novel CNN 110 WSIs from
Maleki, 2022 Binary:
models based on Dartmouth-Hitchcock Acc: 65.90–99.90%
[69] solid/acinar
ResNet-50 Medical Center
ADC WSIs from TCGA
InceptionV3 and Deep 9-class: aci-
dataset and 50 ADC WSIs
features extraction nar/lepidic/solid/papillary/
Sadhwani, 2021 for external validation from AUC: 0.93 in TCGA dataset and
combined with logistic micropapillary cribri-
[62] an independent pathology 0.92 in external validation dataset
regression in two form/necrosis/leukocyte
laboratory in the United
stages aggregates/other
States
269 slides from TCGA Acc: 94.51% (95% CI: 92.77–96.20%)
5-class:
DiPalma, 2021 MIL approach using dataset and Pr: 80.41% (95% CI: 70.55–89.56%)
lepidic/acinar/papillary/
[65] ResNet Dartmouth-Hitchcock Re: 81.67% (95% CI: 71.20–90.43%)
micropapillary/solid
Medical Center F1-score 0.80 (95% CI: 0.71–0.88)
6-class: 422 ADC WSIs from
Wei, 2019
ResNet-18 lepidic/acinar/papillary/ Dartmouth-Hitchcock AUC: 0.97–1.00
[63]
micropapillary/solid/benign Medical Center
50 cases from Cedars-Sinai
GoogLeNet,
5-class: Medical Center in Los
ResNet-50 and
Gertych, 2019 solid/micropapillary/ Angeles, 33 cases from
modified AlexNet Overall Acc: 89.24%
[64] acinar/cribriform/non- Military Institute of
developed in Caffe
tumor Medicine in Warsaw and
engine
27 cases from TCGA dataset
Abbreviations: Acc, Accuracy; ADC, Adenocarcinoma; AUC, Area Under the Curve; CI, Confidence Interval;
CNN, Convolutional Neural Network; GCN, Graph Convolutional Network; MIL, Multiple Instance Learning; Pr,
Precision; Re, Recall; Se, Sensitivity; TCGA, The Cancer Genome Atlas; WSI, Whole Slide Image.
Cancers 2023, 15, 3981 13 of 33
3.1.5. Prediction of Prognosis and Survival

The quantification and evaluation of the tumor microenvironment (TME) features from
histopathological images, derived by the spatial distribution of different cell types (lympho-
cytes, stromal cells), the density of stromal cells, etc., provide valuable information not only
for immune therapy response but also for the probability of survival [70]. TME plays an
important role in immunotherapy response as well as in cancer progression and metastasis
in lung cancer. Several studies have aimed to develop algorithms for TME characterization
of lung cancer pathology images to predict response to targeted therapies and extract prog-
nostic value. Barmpoutis et al. proposed a methodology to identify and quantify tertiary
lymphoid structures (TLS) in lung cancer H&E images [71]. Segmentation of lymphocytes
showed that their density within a TLS region was 3-fold higher than lymphocytes outside
TLS regions. Their study had high sensitivity and specificity rates and could be used as
a prognostic feature to predict response to immunotherapy. DeepRePath was proposed
for prognosis prediction in patients with early-stage ADC [72]. On the external validation
cohort, DeepRePath had an AUC of 0.76, while histopathological features, such as necrosis
or atypical nuclei, were associated with a higher probability of recurrence. The same model,
DeepRePath, was employed by Wu et al. for predicting the recurrence risk of lung cancer,
achieving an AUC of 0.79 on a small testing cohort [73]. In the study of Wang et al., cell type
classification into tumor cells, stromal cells, and lymphocytes achieved great classification
accuracy [74]. TME analysis for spatial distribution estimation associated TME with overall
survival (OS) and could provide valuable information about the patient’s prognosis. In a
similar framework, Wang et al. proposed a CNN for a 6-class classification problem to iden-
tify different cell types nuclei for estimating TME and its prognostic value [75]. The derived
features from the TME analysis were indicators of OS. For instance, higher karyorrhexis
density was associated with worse survival outcomes, while higher stromal nuclei density
was associated with better survival outcomes. Moreover, segmentation of cell nuclei on
H&E WSIs was applied to identify and quantify tumor-infiltrating lymphocytes (TILs) for
prognostic value on NSCLC patients [76]. The authors highlighted the potential of their
proposed model for quantifying TILs, instead of immunohistochemical staining (CD8), for
assisting pathologists. Likewise, the quantitative and spatial localization characteristics
of TILs and tumor cells were evaluated for OS and relapse-free survival (RFS) in NSCLC
cohorts [77]. From 10 immune checkpoint proteins, galectin-9 and OX40L had the higher
relative contribution to OS (33.55%) and RFS (29.02%), respectively, while the percentage
of positive tumor cells and the distance between positive TILs and positive tumor cells
contributed the most to predict OS. A two-step approach of a DL method was proposed by
Pham et al. for detecting lung cancer lymph node metastasis [78]. The proposed approach
was developed to eliminate false positive results by performing a first classification task
for distinguishing reactive lymphoid follicles from lung cancer in lymph nodes. In the
study of Raczkowski
˛ et al., tumor prevalence and TME composition were used as input for
predicting survival and gene mutations in lung ADC cases [79]. The prediction of OS on the
lung dataset was evaluated according to clinical and demographic data [80]. The proposed
weakly supervised and annotation-free CNN achieved a C-index of 0.7033, and features
such as TILs, necrosis, and inflamed stromal regions were identified as prognostic factors
associated with poor outcomes. Estimation of lung ADC tumor cellularity for genetic
tests by pathologists could be improved by DL support. Sakamoto et al. showed that
tumor cellularity can be estimated with minimum deviation from the ground truth when
pathologists and AI scores are combined [81]. Pathologists’ estimations deviated from
the ground truth by approximately 15%, implying over- or under-estimations; however,
false positive results were obtained by AI when cell blocks were evaluated. Prediction
of lung ADC recurrence in several predominant subtypes, including acinar and papillary
carcinoma, after complete resection achieved an accuracy of 90.9% in H&E WSIs from
55 patients [82]. The density of cancer epithelium and cancer stroma lymphocytes was
calculated in H&E slides from lung ADC cases to predict patients’ survival [83]. Low score
rates were associated with significantly superior OS and disease-free survival in patients
Cancers 2023, 15, 3981 14 of 33
with ADC. The authors also included RNA transcripts to determine the TILs infiltration
between the high-risk and low-risk groups revealing that patients in the low-risk group
had a higher proportion of CD8+ T cells, activated CD4+ memory T cells, and plasma cells
versus those in the high-risk group. Slides of lung ADC immunohistochemically stained
for CD3, CD8, and CD20 were used for the detection and quantification of immune cell
biomarkers [84]. High sensitivity and specificity rates were recorded in discriminating T
cells, considering the immunostaining intensity variables and the presence of anthracotic
pigment in the tissue slides. In a recent study, a DL method was employed for predicting
aneuploidy from lung ADC WSIs performing nuclei segmentation and using a single-cell
analysis [85] (Table 4).
Table 4. Characteristics of studies developing models for the prediction of lung cancer prognosis
using histological data.
1st Author, Year Aim of Study Technical Method Classification Dataset Performance Metrics
Overall metrics in the test
dataset
MIM (MLP IN MLP): 3-class: infiltra- 780 images from Acc: 95.31%
Liu, 2023 ADC
a novel deep tion/microinfiltration/ the First Hospital of Jilin Se: 93.10%
[60] prognosis prediction
learning-based model normal University Sp: 96.43%
F1-score: 93.10%
Pr: 93.09%
In lung ADC test dataset
Transformer-guided
Acc: 77.60%
Yu, 2023 ADC MIL with both Binary: negative/positive Slides from 339 patients from
F1-score: 79.50%
[85] prognosis prediction handcrafted and deep aneuploidy TCGA dataset
Cohen’s kappa: 0.55
features
AUC: 0.82
Qaiser, 2022 Lung cancer ResNet-18 along with 1122 WSIs from 410 patients
Binary: high/low OS C-index: 0.70
[80] prognosis prediction attention mechanism from NLST dataset
HoVer-Net PanNuke Aug
model
HR: 0.30 (95% CI:
WSIs from CoNSeP, PanNuke,
Shvetsov, 2022 NSCLC prognosis 0.15–0.60)
HoVer-Net Binary: high-TIL/low-TIL MoNuSAC and UiT-TILs
[76] prediction HoVer-Net MoNuSAC Aug
datasets
model
HR: 0.27 (95% CI:
0.14–0.53)
Integrated score in the
Binary: tumor/non-tumor training dataset
EfficientUnet: a area 1859 NSCLC TMAs from AUC: 0.90 for OS and 0.85
Guo, 2021 NSCLC prognosis combination of Binary: positive/negative Medical University of Gdansk for RFS
[77] prediction EfficientNet and Unet tumor cell staining and 214 NSCLC WSIs from Res-score in the external
ResNet Binary: positive/negative Shanghai Pulmonary Hospital validation dataset
TILs staining AUC: 0.80–0.87 for OS and
0.83–0.94 for RFS
In terms of OS
HR: 2.68 in discovery
cohort, 3.05 in validation
Patients from Guangdong cohort 1, 2.39 in validation
ADC Provincial People’s Hospital, cohort 2 and 1.99 in
Pan, 2022 ResNet-50
prognosis Binary: high-risk/low-risk Shanxi Cancer Hospital, validation cohort 3
[83] HoVer-Net
prediction Yunnan Cancer Hospital and In terms of DFS
TCGA HR: 2.07 in discovery
cohort, 1.54 in validation
cohort 1, and 3.80 in
validation cohort 2
469 ADC slides from TCGA
Levy-Jurgenson, ADC 5 deep learning
Binary: low/high dataset and mRNA/miRNA
2020 prognosis models based on Log rank p-value: 0.07
heterogeneity index expression data from GDC
[86] prediction InceptionV3
database
208 images from 135 patients
ADC
Wang, 2020 from NLST dataset and 431 HR: 2.23 (95% CI:
prognosis Mask-RCNN Binary: high-risk/low-risk
[75] histological images from 372 1.37–3.65)
prediction
patients from TCGA dataset
Cancers 2023, 15, 3981 15 of 33
Table 4. Cont.
1st Author, Year Aim of Study Technical Method Classification Dataset Performance Metrics
from TCGA dataset,
from NLST dataset, 102 images Log rank p-value: <0.01 in
ADC ConvPath: A custom
Wang, 2019 from 102 patients from Chinese TCGA dataset and 0.03 in
prognosis architecture with Binary: high-risk/low-risk
[74] Academy of Medical Sciences Chinese Academy of
prediction 2 convolution layers
dataset and 130 images from Medical Sciences dataset
112 patients from Special
Program of Research
Excellence dataset
DeepLRHE: a novel Se: 84.00%
Lung cancer deep learning model Sp: 67.00%
Wu, 2020
recurrence and consisting of a CNN Binary: high-risk/low-risk 211 images from TCGA dataset Pr: 78.00%
[73]
metastasis prediction and a ResNet F1-score: 81.00%
component AUC: 0.79
Custom Architecture
consisting of 3
Se: 91.70%
Hattori, 2022 ADC recurrence Convolution and 1 Binary: presence/absence WSIs from 55 stage IB ADC
Sp: 90.20%
[82] prediction Fully Connected layer of recurrence patients
Acc: 90.90%
in different color
spaces
3923 slides from 5 St. Mary’s
hospitals affiliated with the
ADC DeepRePath: a novel Binary: high/low
Shim, 2021 Catholic University of Korea in
recurrence CNN model based on probability of recurrence HR: 5.56
[72] Seoul, Incheon, Uijeongbu,
prediction ResNet-50 within 3 years
Bucheon, and Yeouido and
1067 WSIs from TCGA dataset
DeepLRHE: a novel
Binary: positive/negative AUC: 0.87 for TP53, 0.84
Lung cancer deep learning model
Yang, 2021 expression of TP53, EGFR, for EGFR, 0.78 for
immunotherapy consisting of a CNN 180 WSIs from TCGA dataset
[87] DNMT3A, PBRM1 and DNMT3A, 0.75 for PBRM1
efficacy prediction and a ResNet
STK11 and 0.71 for STK11
component
Sp: 92.87% with Se: 95.00%
Lung cancer TLS Combination of Slides from 18 patients from
Barmpoutis, 2021 Binary: TLS/non-TLS Sp: 88.79% with Se: 98.00%
identification and DeepLadV3 with Norfolk and Norwich
[71] region Sp: 84.32% with Se: 99.00%
quantification Inception-ResNetV2 University Hospital
AUROC: 0.96
190 melanoma slides from
Combination of TCGA-SKCM dataset and
Hu, 2021 Anti-PD-L1 response Binary: AUC: 0.65 (95% CI:
Xception, PCA, and 55 NSCLC slides from
[88] prediction response/non-response 49.40–78.40%)
SVM Guangdong Province Cancer
Hospital
Abbreviations: Acc, Accuracy; ADC, Adenocarcinoma; AUC, Area Under the Curve; AUROC, Area Under the
Receiver Operating Characteristic; CI, Confidence Interval; CNN, Convolutional Neural Network; DFS, Disease-
Free Survival; GDC, Genomic Data Commons; HR, Hazard Ratio; miRNA, micro-Ribonucleic Acid; mRNA,
messenger Ribonucleic Acid; NLST, National Lung Screening Trial; NSCLC, Non-Small Cell Lung Cancer; OS,
Overall Survival; Pr, Precision; RCNN, Regional-Convolutional Neural Network; RFS, Relapse-Free Survival; Se,
Sensitivity; Sp, Specificity; TCGA, The Cancer Genome Atlas; TIL, Tumor-Infiltrating Lymphocyte; TLS, Tertiary
Lymphoid Structures; TMA, Tissue Microarray; WSI, Whole Slide Image.
3.1.6. Prediction of Significant Molecular Alterations

Molecular detection of prognostic and predictive biomarkers in specific histological
subtypes can predict favorable responses to targeted therapy and treatment. The detection
of significant molecular alterations on immunohistochemistry (IHC) slides using DL algo-
rithms was the scope of several studies. Concerning ALK rearrangements prediction, in
the study by Terada et al., the commercially available HALO-AI platform and DenseNet
were employed in IHC slides achieving a maximum AUC of 0.73 (in the resolution of
1.0 µm/pix) [89]. Another study aimed to predict mutations (EGFR, BRAF, TP53, STK11,
and KRAS) based on Next Generation Sequencing (NGS) data and H&E WSIs from ADC
samples with several deep neural network-based models [90]. Predicting EGFR and TP53
mutations achieved better performance compared to the remaining genes involved in the
study. In the study of Wang et al., the proposed model for predicting the mutational status
of 10 frequently mutated genes in ADC slides had the best performance for EGFR muta-
tional status with an AUC of 0.824 [50]. Similarly, Coudray et al. trained the Inceptionv3
network to predict the mutational status of 10 genes in lung ADC, with STK11 and KRAS
having the highest AUC of 0.845 and 0.814, respectively [56]. In addition, high performance
Cancers 2023, 15, 3981 16 of 33
was recorded for TP53 and EFGR biomarkers prediction, with AUC of 0.87 and 0.84, respec-
tively, in the study of Yang et al. [87]. However, the model was not validated on an external
cohort, and only 180 WSIs from the TCGA database were used. MET, FGFR1, and FGFR2
mutations were predicted with accuracies of 86.3%, 83.2%, and 82.1%, respectively [91]. The
recent study by Mayer et al. was the first to employ DL for predicting ROS1 rearrangement
directly from H&E WSIs [92]. ROS1 rearrangement prediction reached sensitivity and
specificity of 100% and 98.48%. Moreover, the characterization of intra-tumor heterogeneity
in ADC by gene expression levels was associated with patients’ survival [86]. In the lung
cancer dataset, the highest AUC was detected for miR-17-5p microRNA, followed by KRAS
and CD274 (PD-L1). Another study determined TMB value (low or high) according to a
selected threshold in lung ADC WSIs. TMB value was predicted for each area of the image,
reflecting the heterogeneity of TMB [93]. No significant correlation between the TMB status
and the tumor stage of the patient was noted, while the performance of the DL model was
relatively low, with an AUC of 0.641. Likewise, the prediction of TMB in 50 SCC H&E
images achieved an AUC of 0.65 (Table 5) [94].
Table 5. Characteristics of studies developing models for the prediction of lung cancer mutational
status using histological data.

AUC: 0.87
Pao, 2023 An attention-based MIL Binary:
2099 specimens NPV: 95.40%
[95] model based on ResNet50 mutated/wild-type EGFR
PPV: 41.00%
Per-patient metrics for the optimal
model (VGG16)
VGG16
Dammak, 2023 50 slides from TCGA AUC: 0.65
Xception Binary: high/low TMB
[94] dataset Acc: 65.00%
NASNet-Large
Se: 77.00%
Sp: 43.00%
Se: 100% for both ALK and ROS1
Sp: 100% for ALK and 98.57%
Slides from
GANs along with Binary: positive/negative for ROS1
Mayer, 2022 234 advanced-stage
unsupervised and ALK and ROS1 NPV: 100% for both ALK and
[92] NSCLC patients from
semi-supervised learning rearrangement ROS1
Sheba Medical Center
PPV: 100% for ALK and 50.50%
for ROS1
With 50% probability threshold
AUC: 0.73 (95% CI: 0.65–0.82)
Acc: 73.00%
300 patients from
Terada, 2022 DenseNet via the Binary: positive/negative Se: 73.00%
Shizuoka Cancer Center,
[89] HALO-AI platform ALK rearrangement Sp: 73.00%
Shizuoka, Japan
PPV: 73.00%
NPV: 73.00%
F1-score: 37.00%
Internal test dataset from
Dartmouth-Hitchcock Medical
Center
747 WSIs from AUC: 0.80 (95% CI: 0.69–0.90) for
Binary:
232 patients from EGFR and 0.71 (95% CI:
Tomita, 2022 mutated/wild-type BRAF,
ResNet-18, EfficientNet-B0 Dartmouth-Hitchcock 0.61–0.81) for TP53
[90] EGFR, KRAS, STK11, and
Medical Center and 111 External test dataset from CPTAC-3
TP53
cases from CPTAC-3 study study
AUC: 0.69 (95% CI: 0.62–0.75) for
EGFR and 0.68 (95% CI:
0.60–0.75) for TP53
Binary:
Samples from 55 tumors
mutated/wild-type ALK,
from the Medical
Raczkowski,
˛ 2022 ARA-CNN inspired by BRAF, DDR2, EGFR,
University of Lublin, AUC: up to 0.74 for PDGFRB
[79] ResNet and DarkNet KEAP1, KRAS, MET,
Poland, and 467 images
PIK3CA, RET, ROS1,
from TCGA dataset
STK11, TP53 and PDGFRB
Cancers 2023, 15, 3981 17 of 33
Table 5. Cont.

427 WSIs from
Niu, 2022
ResNet-18 Binary: high/low TMB 427 patients from TCGA AUC: 0.64
[93]
dataset
Binary:
100,000 images from
Li, 2022 mutated/wild-type
Fine-tuned pre-trained NCT-CRC-100k dataset
[96] STK11, TP53, LRP1B, NF1, AUC
Xception model and 900 ADC WSIs from
FAT1, FAT4, KEAP1,
TCGA dataset
EGFR and KRAS
AUC: 0.63 for STK11, 0.77 for
InceptionV3, ResNet-50, Binary:
KEAP1, 0.70 for NF1, 0.72 for
VGG-19, MobileNetV2, mutated/wild-type
Wang, 2021 NSCLC WSIs from TCGA TP53, 0.82 for EGFR, 0.55 for
ShuffleNetV2 and STK11, KEAP1, NF1, TP53,
[50] dataset FAT1, 0.69 for FAT4, 0.76 for
MNASNET on HEAL EGFR, FAT1, FAT4, LRP1B,
LRP1B, 0.54 for SETBP1, 0.66 for
Platform SETBP1 and KRAS
KRAS
Acc: 72.00% for AKT1, 83.00%
for FGFR1, 82.00% for FGFR2,
Binary:
DeepIMLH: a novel CNN 79.00% for HRAS and 86.00% for
Huang, 2021 mutated/wild-type AKT1, 180 WSIs from TCGA
model based on ResNet MET
[91] FGFR1, FGFR2, HRAS and dataset
concept AUC: 0.83 for FGFR1, 0.82 for
MET
FGFR2, 0.79 for HRAS and 0.86
for MET
ADC WSIs from TCGA
InceptionV3 and Deep dataset and 50 ADC WSIs
Sadhwani, 2021 features extraction for external validation
Binary: low/high TMB AUC: 0.71 (95% CI: 0.63–0.79)
[62] combined with logistic from an independent
regression in two stages pathology laboratory in
the United States
Binary: 1634 WSIs from Genetic AUC: 0.64 for NF1, 0.64 for FAT4,
mutated/wild-type NF1, Data Commons database 0.66 for LRP1B, 0.68 for KEAP1,
Coudray, 2018
InceptionV3 FAT4, LRP1B, KEAP1, and 340 slides from New 0.73 for KRAS, 0.75 for FAT1,
[56]
KRAS, FAT1, TP53, SETB1, York University Langone 0.76 for TP53, 0.78 for SETB1,
EGFR and STK11 Medical Center 0.83 for EGFR and 0.86 for STK11
CNN, Convolutional Neural Network; CPTAC, Clinical Proteomic Tumor Analysis Consortium; GAN, Generative
Adversarial Network; MIL, Multiple Instance Learning; NPV, Negative Predictive Value; Non-Small Cell Lung
Cancer; PPV, Positive Predictive Value; Se, Sensitivity; Sp, Specificity; TCGA, The Cancer Genome Atlas; TMB,
Tumor Mutation Burden; WSI, Whole Slide Image.
3.2. Cytology
Cytological specimens from the lung are frequently the only available diagnostic ma-
terial. However, by its nature, this material is limited, prohibiting auxiliary techniques for
specific subtyping, such as immunocytochemistry. Only a limited number of studies have
addressed the issue of utilizing cytological images for training neural networks for lung
cancer diagnosis and subtyping (Table 6). The first study for the classification of lung cancer
cytological images (ADC, SCC, SCLC) achieved a classification accuracy of 71% after the
data augmentation process [97]. In addition to this study, Teramoto et al. further extended
their work for the classification of lung cytological images (real and synthesized) into
benign and malignant with a generative adversarial network (GAN) [98]. The proposed
method achieved an AUC of 0.901. Similarly, the classification of benign and malignant
cells from cytological pleural effusions WSIs, by a weakly supervised model achieved an
AUC of 0.9526 [99]. The model had a significantly strong correlation with the histological
diagnosis gold standard as well as with senior cytopathologists’ diagnosis. Misclassifi-
cation was observed when poor adhesion of tumor cells or clusters of mesothelial cells
were present. Diagnosis between benign and malignant cells from cytological specimens
was performed in the studies of Lin and Teramoto et al., including 499 and 322 images,
respectively [100,101]. Distinct morphological features (size of cells, nuclei, and nucleoli) of
cytological specimens of lung cancer were recognizable by four different fine-tuned deep
CNNs (DCNNs) [102]. Three out of four DL models resulted in a classification accuracy
Cancers 2023, 15, 3981 18 of 33
of more than 73% for lung cancer subtyping into ADC, SCC, and SCLC; however, some
cases of poorly differentiated NSCLC were misclassified. Furthermore, the distinction
between LCNEC and SCLC showed promising results in the study of Gonzalez et al. [103].
Three classifiers were developed with three distinct datasets of Diff-Quik® -, Papanicolaou-
and H&E-stained cytological WSIs and achieved an AUC of 1, 1 and 0.875, respectively.
Lastly, endobronchial ultrasound (EBUS)-guided transbronchial needle aspiration (TBNA)
cytological images were employed for diagnosing mediastinal metastatic lesions [104].
The study by Wang et al. was the first to include EBUS-TBNA cytological images for
automatic segmentation of enlarged mediastinal lymph nodes metastasis, outperforming
three state-of-the-art baseline models.
Table 6. Characteristics of studies developing models for the cytological interpretation of lung cancer.

AlexNet
Acc: 73.70%
82 images from 36 ADC
AlexNet GoogLeNet/InceptionV3
cases, 125 images from
Tsukamoto, 2022 GoogLeNet/InceptionV3 3-class: Acc: 66.80%
14 SCC cases and
[102] VGG-16 ADC/SCC/SCLC VGG16
91 images from 5 SCLC
ResNet-50 Acc: 76.80%
cases
ResNet50
Acc: 74.00%
Novel DL model
Pr: 93.40% in 1st and 91.80%
122 WSIs from in 2nd experiment
Custom Architecture Binary: EBUS-guided TBNA Se: 89.80% in 1st and 96.30%
Wang, 2022
with 8 Convolution and positive/negative samples from in 2nd experiment
[104]
1 Deconvolution layers lymph node metastasis Tri-Service General DSC: 82.20% in 1st and
Hospital 94.00% in 2nd experiment
IoU: 83.20% in 1st and
88.70% in 2nd experiment
Acc: 91.67%
Xie, 2022 Binary: 404 WSIs from Shangai Sp: 94.44%
ResNet-18
[99] benign/malignant Pulmonary Hospital Se: 87.50%
AUC: 0.95 (95% CI: 0.90–0.99)
Acc: 98.80% for patch-based
499 images from classification, 95.50% for
97 patients from image-based classification
National Taiwan and 92.90% for patient-based
Lin, 2021 Binary:
ResNet-101 University Cancer classification
[100] benign/malignant
Center and National Se: 98.80% for patch-based
Taiwan University classification
Hsin-Chu Hospital Sp: 98.80% for patch-based
classification
MIL approach with
attention mechanism
Teramoto, 2021 Binary: Images from
and several CNN Acc: 91.60%
[101] benign/malignant 322 patients
architectures as
backbone
Combination of
Teramoto, 2020 progressive growing Binary:
Images from 60 patients Acc: 85.30%
[98] GAN and benign/malignant
VGG-16 architecture
Cancers 2023, 15, 3981 19 of 33
Table 6. Cont.

Diff-Quik® -stained model
AUC: 1.00 with a threshold at
Se: 100.00% and Sp: 87.50%
114 cytological and Pap-stained model
Gonzalez, 2020 A deep learning model Binary:
histological slides from AUC: 1.00 with a threshold at
[103] based on InceptionV3 LCNEC/SCLC
40 cases Se: 100.00% and Sp: 85.70%
H&E-stained model
AUC: 0.88 with a threshold at
Se: 100.00% and Sp: 87.50%
Original images
Acc: 73.20% for ADC, 44.80%
Custom architecture for SCC, 75.80% for SCLC
Teramoto, 2017 consisting of 3 3-class: and 62.10% overall
76 cases
[97] convolutions and 3 ADC/SCC/SCLC Augmented images
Fully Connected layers Acc: 89.00% for ADC, 60.00%
for SCC, 70.30% for SCLC
and 71.10% overall
CNN, Convolutional Neural Network; DSC, Dice Similarity Coefficient; EBUS, Endobronchial Ultrasound; GAN,
Generative Adversarial Network; H&E, Hematoxylin & Eosin; IoU, Intersection over Union; LCNEC, Large Cell
Neuroendocrine Carcinoma; MIL, Multiple Instance Learning; Pr, Precision; SCC, Squamous Cell Carcinoma;
SCLC, Small Cell Lung Cancer; Sp, Specificity; Se, Sensitivity; TBNA, Transbronchial Needle Aspiration; WSI,
Whole Slide Image.
3.3. PD-L1 Expression Status

PD-L1 is an immune checkpoint protein expressed on tumor cells and activated im-
mune cells [105]. In NSCLC patients, assessment of PD-L1 expression is pivotal for guiding
patients’ treatment selection with immune checkpoint inhibitors (ICIs). IHC is the currently
accepted diagnostic assay performed on formalin-fixed paraffin-embedded (FFPE) lung
tissue or cytological specimens [106]. There are different platforms for IHC interpretation,
PD-L1 antibodies, guidelines for evaluation and scoring, as well as positivity cut-offs for
immunotherapy selection. Currently, four IHC assays (28-8 and 22C3 from DAKO, SP263
and SP142 from Ventana) have been approved for use by the Food and Drug Administra-
tion (FDA). The 22C3 and 28-8 pharmDx (DAKO) IHC assays are companion diagnostics
for selecting patients for pembrolizumab and nivolumab, respectively [107,108]. SP142
and SP263 (Ventana) IHC assays are also FDA-approved for companion diagnostic to
atezolizumab. Evaluation of PD-L1 expression with the 22C3 and 28-8 pharmDx, as well
as SP263 (Ventana) assays, only refers to the PD-L1 expression on tumor cells, while, on
the other hand, the SP142 (Ventana) assay refers to tumor and immune cells staining [109].
As PD-L1 scoring algorithms determine the therapeutic choice and interobserver discor-
dance is common, it is conceivable that quantitative validation of PD-L1 expression by
DL algorithms may assist pathologists in their assessment. In a recent study, Hondelink
et al. developed a fully supervised DL model for PD-L1 TPS assessment in NSCLC WSIs
according to three cut-off points (<1%, 1–50%, and 50–100%) [110]. TPS prediction was in
concordance with the mean score of three pathologists in 79% of the cases. Misclassification
of some cases was noted when positive PD-L1 immune cells were present around the tumor
site, the intensity of PD-L1 positive neoplastic cells was weak, or when non-membranous
staining was detected. In a similar framework, Liu et al. performed tumor region seg-
mentation and nuclei detection for PD-L1 TPS prediction on SCC WSIs according to three
cut-off points (<1%, 1–49%, and ≥50%) [111]. Their proposed model’s predictions were
compared to the pathologist’s prediction with different experience levels. The model’s clas-
sification accuracy was 74.51%, higher than trainees (71.55%) but lower than subspecialist
and non-subspecialist pathologists (97.06% and 84.03%, respectively). In another study,
TPS assessment reached high performance in terms of sensitivity and specificity in both
Cancers 2023, 15, 3981 20 of 33
1% and 50% cut-off points [112]. The classification was performed on slides stained with
22C3 antibody, and the proposed patch-based dual-scale categorization method based on
VGG16 architecture achieved higher performance compared to VGG16. The study of Sha
et al. resulted in an AUC of 0.80 on a balanced testing cohort in classifying positive and
negative PD-L1 tumor cells [113]. In SCC-separated cases, the model achieved a lower
AUC compared to ADC cases (0.64 and 0.83, respectively), maybe due to an imbalance in
the training cohort. In the studies of Kapil et al., TPS was estimated by dividing the pixel
number of positive tumor cells by the total pixel number of positive and negative tumor
cells [114,115]. Of all the included studies estimating PD-L1 TPS, these two were the only
ones using slides stained with SP263 antibody with the cut-off point defined at 25%. In their
first study, fully- and semi-supervised network architectures were used for estimating TPS
in NSCLC specimens, with results agreeing with pathologists’ evaluation, while, in their
subsequent study, TPS estimation was performed with a GAN. Two classification problems
were addressed, namely, a binary task for epithelial and non-epithelial region segmentation
as well as TPS estimation. An additional dataset of WSIs stained with the epithelial marker
Pan-Cytokeratin was used for the binary segmentation task of the epithelial benign and
malignant regions. In the study of Wu et al., PD-L1 IHC slides stained with 22C3 assay
were used for training U-Net to perform tumor area detection and TPS calculation [116].
The model was highly consistent with trained pathologists and achieved high performance
when further tested in SP263 (Ventana) stained slides (accuracy of 0.9326 and 0.9624 for
22C3- and SP263-stained slides, respectively). Furthermore, the authors demonstrated that
the AI-based model could help untrained pathologists with TPS assessment by reducing
the time of microscopic examination. In the same framework, three automated workflows
based on DL, including both 22C3 (DAKO) and SP263 (Ventana) IHC assays, and two
cut-off points (<1%, ≥50%), achieved better performance in the <1% cut-off point [117]. The
model by Choi et al. achieved an area under the receiver operating characteristic (AUROC)
of 0.889 in detecting PD-L1 positive and negative tumor cells and estimating TPS value,
while it significantly increased the concordance of pathologists after a disagreement (ini-
tial/baseline concordance of 81.4% versus revised concordance of 90.2%) [118]. Aitrox’s AI
performance for PD-L1 expression by Huang et al. was comparable to those of experienced
pathologists, while it surpassed inexperienced ones (Table 7) [119].
Table 7. Characteristics of studies developing models for the assessment of programmed cell death
ligand 1 expression in lung cancer using histological data.
1st Author,
Technical Method Classification IHC Assay Dataset Performance Metrics
Year
3-class: Best model
MobileNetV2 for PD-L1+ tumor 22C3 pharmDx 1288 samples from LCC 95% CI: 0.86–0.89
Cheng, 2022
classification and cells/PD-L1+ immune (DAKO) and Zhejiang Cancer with PD-L1 (22C3) assay
[117]
YOLO for detection cells/PD-L1− tumor SP263 (Ventana) Hospital and 0.81–0.91 with PD-L1
cells (SP263) assay
348 slides from AUROC: 0.89 for PD-L1+
Samsung Medical cells and 0.81 for PD-L1−
Binary:
Choi, 2022 22C3 pharmDx Center and 131 slides cells
Faster R-CNN PD-L1+/PD-L1−
[118] (DAKO) from Seoul National F1-score: 72.30% for
tumor cells
University Bundang PD-L1+ cells and 72.20%
Hospital for PD-L1− cells
rs : 0.87
3-class: negative PD-L1
Acc: 79.13% for all
expression (TPS:
222 WSIs from Fudan subsets, 85.29% for
Huang, 2022 U-Net based <1%)/low PD-L1 22C3 pharmDx
University Shanghai negative TPS subset,
[119] architecture expression (TPS: (DAKO)
Cancer Center 77.79% for low TPS
1–49%)/high PD-L1
subset, and 72.73% for
expression (TPS: ≥50%)
high TPS subset
Cancers 2023, 15, 3981 21 of 33
Table 7. Cont.
1st Author,
Technical Method Classification IHC Assay Dataset Performance Metrics
Year
199 stage IV NSCLC
A novel supervised
WSIs stained with
Hondelink, deep learning model 3-class: ICC: 0.96 (95% CI:
22C3 pharmDx PD-L1 22C3 antibody
2022 based on AIFORIA TPS < 0.94–0.97)
(DAKO) from Leiden
[110] CREATE software 1%/1–49%/50–100% Cohen’s kappa: 0.68
University Medical
(v4.6)
Centre
Binary
Acc: 93.26%
501 NSCLC WSIs Sp: 96.41%
Binary:
A novel supervised from Peking Pr: 92.48%
tumor/non-tumor 22C3 pharmDx
Wu, 2022 deep learning University Cancer Re: 86.09%
3-class: (DAKO) and
[116] algorithm based on Hospital and Tianjin F1-score: 88.71%
TPS < SP263 (Ventana)
U-Net Medical University IoU: 80.51%
1%/1–49%/50–100%
Cancer Hospital 3-class
r: 0.94–0.95 in 22C3 assay
and 0.98 in SP263 assay
DASGAN network: Binary: epithelial/non-
an extension of epithelial
56 WSIs stained with
CycleGAN 3-class: tumor PD-L1+ Binary
Pan-Cytokeratin and
Kapil, 2021 architecture epithelial region/tumor F1-score: 88.60%
SP263 (Ventana) 122 WSIs stained
[114] An extension of the PD-L1− epithelial 3-class
with PD-L1 SP263
deep survival region/other regions F1-score: 85.00%
antibody
learning (immune, stromal,
methodology necrotic)
TPS prediction
4-class: PD-L1+ tumor F1-score: 90.24% with 1%
cells/PD-L1− tumor and 81.82% with 50%
cells/PD-L1+ immune cut-off
cells/other region 300 NSCLC slides AUC: 0.97 with 1% and
DSC-VGG-16: a
3-class: maximum stained with PD-L1 0.99 with 50% cut-off
novel dual-scale
Wang, 2021 counts of PD-L1+ 22C3 pharmDx 22C3 antibody from Se: 88.10% with 1% and
categorization-based
[112] tumor cell (TP1)/50% (DAKO) Changhai and 75.00% with 50% cut-off
deep learning model
PD-L1+ tumor cell of Changzheng Sp: 95.59% with 1% and
based on VGG-16
TP1 (TP2)/25% PD-L1+ hospitals 98.98% with 50% cut-off
tumor cell of TP1 (TP3) Cohen’s kappa: 0.79 (95%
3-class: TPS < CI: 0.68–0.90)
1%/1–49%/50–100% Lcc: 0.88 (95% CI:
0.83–0.92)
Automated Tumor
Proportion Scoring
System: a novel deep 96 SCC WSIs stained
Acc: 74.51%
learning model using 3-class: with PD-L1 22C3
Liu, 2021 22C3 pharmDx MAE: 8.65 (95% CI:
Res50UNet for tumor TPS < antibody from Fudan
[111] (DAKO) 6.42–10.90)
region segmentation 1%/1–49%/50–100% University Shanghai
r: 0.94
and MicroNet for Cancer Center
tumor nuclei
detection
3-class: tumor AUC: 0.80 for all cases,
Sha, 2019 22C3 pharmDx
Modified ResNet-18 PD-L1+/tumor 130 NSCLC samples 0.83 for ADC cases and
[113] (DAKO)
PD-L1−/other 0.64 for SCC cases
Lcc: 0.94
270 NSCLC slides r: 0.95
Binary:
Kapil, 2018 Auxiliary Classifier from NCT01693562 MAE: 8.00
PD-L1+/PD-L1− SP263 (Ventana)
[115] GAN and NCT02000947 OPA: 0.88
tumor regions
clinical trials NPA: 0.90
PPA: 0.85
Abbreviations: Acc, Accuracy; ADC, Adenocarcinoma; AUC, Area Under the Curve; AUROC, Area Under
the Receiver Operating Characteristic; CI, Confidence Interval; CNN, Convolutional Neural Network; GAN,
Generative Adversarial Network; ICC, Interclass Correlation Coefficient; IHC, Immunohistochemistry; IoU,
Intersection over Union; LCC, Linear Correlation Coefficient; Lcc, Lin’s concordance coefficient; MAE, Mean
Absolute Error; NPA, Negative Percent Agreement; NSCLC, Non-Small Cell Lung Cancer; OPA, Overall Percent
Agreement; PD-L1, Programmed cell Death Ligand 1; PPA, Positive Percent Agreement; Pr, Precision; r, Pearson’s
correlation coefficient; Re, Recall; rs , Spearman’s rank correlation coefficient; SCC, Squamous Cell Carcinoma; Se,
Sensitivity; Sp, Specificity; TPS, Tumor Proportion Score; WSI, Whole Slide Image.
Cancers 2023, 15, 3981 22 of 33
3.4. Deep Learning Approaches

From a clinical point of view, the main challenges in Digital Pathology are (i) the
extremely large size of the images produced by whole-slide scanning and the requirement
for pathologists to evaluate the entire specimen; (ii) the digitization of annotated findings of
interest, which is a very demanding and time-consuming process. The latter, combined with
the fact that DL techniques require a large amount of training data, intensifies the problem
of the provision of reliable results. Many studies presented in the literature attempt to
overcome the lack of annotations by using weakly supervised or semi-supervised learning
techniques instead of fully supervised approaches. These approaches interact with known
CNN architectures to classify patches of images or to detect tissue alterations and/or
morphological features of cancer. Weakly supervised learning is a branch of machine
learning (ML) that aims to use less or lower quality labels for training predictive models.
It works by leveraging the unlabeled data or refining the labels to improve the model
performance. In terms of Digital Pathology, weakly supervised methods use a small
number of annotations by selecting informative patches to classify the WSIs [25]. General
approaches of weakly supervised learning in histopathological images have been proposed,
employing VGG-16 [25], EM-CNN [52], EfficientNet-B3 [20], and ResNet [80]. Furthermore,
most of the presented studies in this category employ MIL [34,85], which is a weakly
supervised learning technique that groups data points into bags. Each bag is labeled with
the class by the instance count of that particular class. This technique is well-suited for
histology slide classification because it is designed to operate on weakly-labeled images [65].
For example, clustering-constrained-attention MIL (CLAM), developed by Lu et al. [47], is
a weakly supervised method that uses attention-based learning to automatically identify
subregions of high diagnostic value and, thus, accurately classify the whole slide. Other
works combine the MIL approaches with well-known architectures of CNNs, such as
ResNet [48,65,95], EfficientNetB1 [22], and SimCLR [59]. Moreover, Teramoto et al. [101]
compared several CNNs as backbones (LeNet, AlexNet, ResNet, Inception, DenseNet) using
MIL and an attention mechanism, while Hou et al. [57] presented 14 different combinations
of expectation maximization (EM)-based MIL approach with Logistic Regression and
Support Vector Machine (SVM). Finally, another work that attempted to overcome the lack
of labeled data employs a semi-supervised approach inspired by YOLOv5 for the detection
of micropapillary lung ADC. This method implements a teacher model, which is directly
trained by the ground truth data, and a student model, which indirectly learns from the
teacher model [68].
From a technical perspective, the extremely large size of images and the complexity
of classification or detection problems in these images as well generate a very demanding
process in terms of computational resources and training time of supervision. Typically,
researchers can follow two main different approaches: (i) to develop a custom architecture,
implementing all the components of both convolutional and fully connected layers and
defining all the super parameters of the network or (ii) to use already pre-trained archi-
tectures and take advantage of transfer learning from other datasets (i.e., IMAGENET).
Custom architectures can be more accurate than pre-trained CNNs with transfer learning
if they are designed well for a specific problem and trained on an adequate set of images.
However, they require more time and resources to develop and train. For these reasons,
custom CNNs are mostly less deep than the pre-trained models to overcome the limitations
of the demanding implementation and the computational requirements. Thus, most of the
presented custom architectures for lung cancer consist of up to three convolution layers as
well as up to three fully connected layers [21,49,61,74,82,97]. One of these works utilizes
two different color spaces developing two same feature extractors, one for RGB and one
based on HLS [82]. More extended architectures schemas have also been presented, devel-
oping six convolutions and two dense layers [84] or more than five convolution layers along
with one devolution for upscaling [40,104]. Finally, the deeper CNN in this category, called
Deep Hipo, operates on both magnifications (20× and 5×), and it is based on CAT-NET
developing 19 layers in total [28].
Cancers 2023, 15, 3981 23 of 33
As a result of the limited implementation effort needed, the vast majority of DL

methods presented in the literature for lung cancer leverage well-known Convolutional
Networks architectures, which are often pre-trained in different datasets. Several of
these use architectures included in cloud-based platforms or frameworks, such as HALO-
AI [30,78,81,89], HEAL Platform [50], AIFORIA [110], and Caffe [53,64,78]. In these cases,
many well-known CNNs have been employed. For example, Wang et al. [50] used Incep-
tionV3, ResNet50, VGG19, MobileNetV2, ShuffleNetV2 and MNASNET, while Yu et al. [53]
employed AlexNet, GoogLeNet (InceptionV3), VGG-16 and ResNet-50. For classifica-
tion problems, the most employed architectures seem to be the ResNet-based models,
such as ResNet-18 [63,93,99], ResNet-50 [69,83], and ResNet-101 [36,100], as well as the
Inception-based models [55,56,86]. Apart from these, U-Net [116,119], Xception [58,94,96],
Hover-Net [76,83] and InceptionV3 [41] have also been used in several studies. In compara-
tive studies or studies which use multiple architectures of CNNs, several other models have
been presented, such as NASNetLarge [27], EfficientNet [26], SqueezeNet [39], etc. On the
other hand, few approaches utilized known DL detectors for segmentation or quantitation
purposes. Choi et al. [118] detected PD-L1 positive and PD-L1 negative tumor cells using
Faster R-CNN, while Cheng et al. [117] for the same problem employed YOLO. Finally,
Wang et al. [75] detected six different classes of cells segmenting the images with Mask
R-CNN.
More sophisticated DL methods have been proposed, either modifying known archi-
tectures of CNNs or combining two differing CNN architectures and CNN architectures
with classic ML techniques. Most of the modified architectures are based on ResNet.
DeepRePath [72] is a novel CNN model based on ResNet-50 that operates on different
magnifications building two CNNs, while a similar approach proposed by Sha et al. [113]
developed two branches for the processing of small and large field-of-view features of
PD-L1 classes. On the other hand, SE-ResNet-50 [38] focuses on the improvement of the ac-
tivation function introducing CroRELU. Other novel modifications of known architectures
are the KimiaNet22 based on DenseNet [44], the MR-EM-CNN, which extracts hierarchical
multiscale features on an EM-CNN model [43], the DSC-VGG16, which provides a dual
scale categorization of PD-L1 classes based on VGG16 [112], the WIFPS model [31] based
on EfficientNet-B5, and the novel architecture proposed by Gonzales et al. [103], which
utilizes three different stains. Finally, Raczkowski
˛ et al. [79] developed a novel architecture
called ARA-CNN, which is inspired by both ResNet and DarkNet models.
By combining different CNN models or CNNs with classic ML techniques, researchers
attempt to provide better performance in several categories of lung cancer problems.
Combinations of different CNN models presented in the literature are (i) ResNet-50
with U-Net [111], (ii) EfficientNet with U-Net [77], and (iii) DeepLadV3 with Incepetion-
ResNetV2 [71]. By combining DL and ML approaches, Wang et al. [42] introduced the
LungDIG architecture, which employs an Inception-V3 model along with a classic mul-
tilayer perceptron. Two other approaches extract deep features utilizing the convolution
layers of CNNs and then provide predictions using logistic regression [19,62]. SVM has
also been used in cooperation with CNN models. Perez et al. [46] merged information
from ResNet-18 from the processing of WSIs along with SVM from RNA-sequencing data,
while Toğaçar et al. [45] and Hu et al. [88] combined SVM with DarkNet and Xception
models, respectively. Finally, principal component analysis (PCA) techniques have been
used along with CNNs architectures for dimensionality reduction of the extracted fea-
tures [18,88]. The contribution of DL in lung cancer presents several other methods that
employ Graph-based CNNs, GANs, and autoencoders. Graph CNNs have been used to
identify regions or cell structural features that are highly associated with the class label. In
this category, three approaches have been proposed, where Graph-based modules are com-
bined with AlexNet [54], VGG16 [66], and ResNet. [37]. GANs are mostly used to generate
informative synthetic sets of images in order to increase the training set and, thus, avoid
overtraining issues. DASGAN, which is an extension of the CycleGAN architecture, has
been introduced [114], merging two stains and leading deep survival learning methodology.
Cancers 2023, 15, 3981 24 of 33
Teramoto et al. introduced a progressive growing approach of GANs (PGGAN) combined

with the VGG-16 model [98], while Mayer et al. [92] combined GANs with semi-supervised
learning. Another auxiliary classifier GANs (AC-GANs) approach has been proposed by
Kapil et al. [115] to generate classifier models and detect ALK and ROS1 fusions directly
from H&E images. Finally, an unsupervised DL model that employs stacked autoencoders
has been developed by Sheikh et al. [67].
4. Discussion
DL is progressively embraced in Pathology, especially for breast, colorectal, prostate,
and lung cancer diagnosis, transforming the current landscape of medicine [120–131]. AI
could play a pivotal role in the multidisciplinary approach to diagnosis and patient man-
agement. As already underlined above, in lung cancer, classification, accurate diagnosis
and subtyping depend on distinct morphological features among cancer cells combined
with staining patterns, tumor biological characteristics, and molecular data of mutations.
Lung cancer histology is characterized by cellular heterogeneity, challenging the diag-
nostic process [132]. Several histological features can be defined by examining a single
H&E-stained slide, such as glandular differentiation in lung ADC, the presence of kera-
tinization and intercellular bridges in SCC, as well as scant cytoplasm and poorly defined
cell borders in SCLC. However, for differential diagnosis, special immunohistochemical
staining is required for accuracy. According to the WHO guidelines, the terminology
for lung cancer classification in small biopsies or cytology and resection specimens must
follow the proposed criteria [2]. For example, in resection specimens, lung ADC cases must
be morphologically determined by the predominant histological pattern (lepidic, acinar,
papillary, micropapillary, solid). The distinction of lung neuroendocrine tumors (NETs)
directly from the H&E slide can also be challenging, whereas NETs are further classified as
typical carcinoids, atypical carcinoids, SCLC, and LCNEC. Given that small biopsies and
cytology specimens are encountered for diagnosis in about 70% of the patients, the available
diagnostic material is often limited and thus, every effort should be employed to preserve
sufficient material for molecular analysis. Therefore, it is strongly recommended to use only
a limited panel of biomarkers, including the most representative ones for immunostaining
for differential diagnosis. However, this approach can hamper accurate diagnosis. Here,
AI could be of great help to the pathologist by guiding with high accuracy the prevailing
diagnosis from an H&E-stained slide.
Data extraction of our systematic review demonstrated that DL-based methodologies
for lung cancer diagnosis are mainly performed on histological H&E WSIs, with ADC
versus SCC being the predominant classification task, as shown in Table 2. All the studies
were performed with high classification accuracy for identifying ADC and SCC. Secondly,
many studies utilized different CNN architectures for classifying ADC, SCC, and SCLC
in small biopsies. The higher performance was in the study of Kanavati et al. [24] (AUC
of 0.94–0.99), which included a large number of images. Only two studies designed a
classification task for identifying ADC, SCC, SCLC, and LCNEC on WSIs [27,28]. This
4-class task represents the realistic daily practice of a pathologist. In both studies, the
AUC was over 0.90, encouraging the fact that such DL models could be employed and of
great value in a pathology laboratory. The third most common approach in histological
slides was the employment of DL-based models for lung ADC histological subtyping. The
studies of Sheikh [67] and DiPalma et al. [65] achieved the highest classification accuracy
performing a 5-class problem (lepidic, acinar, papillary, micropapillary, solid). Albeit
limited in number, eight noteworthy studies utilized cytological slides for lung cancer
diagnosis or classification. Four of them performed a binary classification task for benign
and malignant cell detection [98–101]. All studies showed good classification accuracy;
however, compared to the classification problems performed on histological data, the
dataset was limited in the majority of the studies. In addition, in the cytology section,
the most common classification task for lung cancer (ADC, SCC, and SCLC) resulted in
modest classification accuracies, including state-of-the-art architectures (66–77% and ~71%),
Cancers 2023, 15, 3981 25 of 33
with the main limitation being the small number of images included for training (55 and
76 cytological slides, respectively) [97,102]. Prediction of OS and risk of recurrence as well
as identification of prognostic features, were also the aim of many research papers, in which
the predicted output emerged after nuclei segmentation, TILs quantification, identification
of gene expression, or clinical data. The highest AUC (0.917) for ALK rearrangements
Cancers 2023, 15, 3981 26 of 35
prediction was in the study by Chen et al. [31], while EGFR mutations were predicted
with an AUC of 0.824, 0.84, and 0.83 in the studies by Wang, Yang, and Coudray et al.,
respectively [31,50,56,87]. In the most recent study by Pao et al., the prediction of EGFR
mutational status in 2099 lung ADC tissue specimens reached an AUC of 0.87 [95]. As far
as PD-L1 quantification is concerned, the majority of studies included datasets consisting
of WSIs stained with the 22C3 antibody. The remaining studies included slides stained
with the SP263 antibody or a combination of 22C3 and SP263 antibodies. For quantitative
problems, such as TPS estimation for PD-L1 expression, labeling ground truth must be as
consistent as possible to avoid misclassification concerningconcerning the specific cut-off
cut-off points for
evaluationDL-based
PD-L1 evaluation DL-basedmodels
modelsforforPD-L1
PD-L1TPS TPSestimation
estimation offer
offer several
several advantages
advantages to
to pathologists
pathologists as TPS
as TPS quantification
quantification is a is a time-consuming
time-consuming process
process prone prone to subjective
to subjective esti-
estimation.
mation. Despite
Despite the the extensive
extensive research
research andand progress
progress on histological
on histological images,
images, further
further re-
research on cytological material, including a larger dataset, is considered
search on cytological material, including a larger dataset, is considered essential for opti-essential for
optimizing
mizing classification
classification performance.
performance.
According to the technical
According to the technical pointpoint of
of view,
view, summarizing
summarizing the the methods
methods presented
presented in in the
the
literature, most of them (78 studies) developed supervised learning methodologies,
literature, most of them (78 studies) developed supervised learning methodologies, basi- basi-
cally dealing
cally dealing with
with classification
classification problems
problems of of the
the medical
medical question.
question. Specifically,
Specifically, 11
11 studies
studies
implemented custom CNN architectures, 36 studies employed
implemented custom CNN architectures, 36 studies employed known models with or known models with or
without transfer learning, 11 studies modified known architectures, and, finally,
without transfer learning, 11 studies modified known architectures, and, finally, 14 stud- 14 studies
combined
ies combinedCNNsCNNs either with
either each
with eachother
other ororwith
withMLMLtechniques.
techniques.Apart
Apart from
from the
the above
above
crisp categories of supervised learning, the category named “other methods”
crisp categories of supervised learning, the category named “other methods” contained contained six
supervised, one weakly supervised, and one unsupervised method (eight
six supervised, one weakly supervised, and one unsupervised method (eight studies in studies in total).
Weakly supervised methods are 13 in total, while there are one semi-supervised and one
total). Weakly supervised methods are 13 in total, while there are one semi-supervised
unsupervised method (Figure 2).
and one unsupervised method (Figure 2).
Figure 2. Summary of DL methods for lung cancer.

Figure 2. Summary of DL methods for lung cancer.
To
To conclude
conclude about
aboutthe
themost
mostcommonly
commonlyused known
used knownarchitectures, thethe
architectures, employed ar-
employed
chitectures have
architectures been
have counted
been countedforfor
each
eachstudy,
study,and
andthe
theresults
resultsare
arepresented
presentedininFigure
Figure 3.
3.
Note that several studies have not used known architectures (for example the studies that
develop custom CNN architectures), while several studies employ more than one.
Cancers 2023, 15, 3981 26 of 33
Cancers 2023, 15, 3981 27 of 35

Note that several studies have not used known architectures (for example the studies that
develop custom CNN architectures), while several studies employ more than one.
12%
3% ResNet Inception
3% 31%
4%
VGG EfficientNet
AlexNet MobileNet
4%
U-Net Xception
4%
YOLO DenseNet
5% Other
7% 15%
12%
Distributionofofknown
Figure3.3.Distribution
Figure knownDL
DLarchitectures
architecturesfor
forlung
lungcancer.
cancer.
Ourreview
Our reviewshows
shows that
that many
many of of
thethe employed
employed DLDL methods
methods in lung
in lung cancer
cancer are par-
are partic-
ticularly extensive and sophisticated, as well as scalarly evolving into
ularly extensive and sophisticated, as well as scalarly evolving into new techniques fol- new techniques
following the development of AI. According to the comparative studies presented in this
lowing the development of AI. According to the comparative studies presented in this
review, DL methods overall outperform traditional ML techniques. This superiority of DL
review, DL methods overall outperform traditional ML techniques. This superiority of DL
could partially be explained by the quality of the features feeding the fully connected layers.
could partially be explained by the quality of the features feeding the fully connected lay-
The features in CNNs are not selected subjectively by the specialists but are automatically
ers. The features in CNNs are not selected subjectively by the specialists but are automat-
extracted from the convolutional layers, maximizing the carried information.
ically extracted from the convolutional layers, maximizing the carried information.
Comparing the reviewed architectures, it is evident from the results of the review
Comparing the reviewed architectures, it is evident from the results of the review
that ResNet-based and Inception-based architectures have been used in about half of
that ResNet-based and Inception-based architectures have been used in about half of the
the methods presented in the literature, showing high performances compared to other
methods presented in the literature, showing high performances compared to other archi-
architectures. The existence of residual blocks in most of these architectures (all ResNet
tectures. The existence of residual blocks in most of these architectures (all ResNet and
and InceptionV4 models) seems to operate efficiently and effectively in biopsy image
InceptionV4 models) seems to operate efficiently and effectively in biopsy image pro-
processing. Jumping features directly from a convolutional layer to many subsequent
cessing. Jumping like
layers operates features
mergingdirectly from from
features a convolutional layer to
different digital many subsequent
magnifications layers
of scanning.
operates like merging features from different digital magnifications of
Such a procedure seems to make sense for biopsy imaging, where different magnificationsscanning. Such a
procedure
of scanning seems to make
provide sense
different for biopsy
knowledge imaging,
about where different magnifications
the microenvironment of the cells. of
scanning provide different knowledge about the microenvironment of the cells.
It is also meaningful to summarize the limitations of the DL techniques in lung cancer.
It 8isemphasizes
Table also meaningful
several tolimitations
summarizeofthe thelimitations
applicationofofthe
theDL techniques
proposed in lung can-
DL methodologies
cer. Table 8 emphasizes several limitations of the application of the
in lung cancer diagnosis that we were able to identify based on our systematic proposed DL method-
review.
ologies in lung cancer diagnosis that we were able to identify based on
Some of them are generally well-known constraints, while some others are related to the our systematic
review.
imaging Some of them
problem are generally
of lung biopsies.well-known constraints, while some others are related
to the Our
imaging problem
findings of lung biopsies.
demonstrate that the field of Digital Pathology for lung cancer diagnosis
has evolved rapidly in the last 5 years. However, at least for most laboratories, the use of
Table
these8.capabilities
Deep learning
in limitations for practice
daily clinical lung cancer applications.左对齐
is still in its early stages. Adopting a fully digital
workflow can be challenging,
Limitation and limitations must be overcome for implementation in the
Property
clinical setting. Digital slide generation
Lack of interpretability According to the is the first step
review, onlyinamoving from traditional
few approaches focus on toper-
Digital
Pathology. WSI scanners provide high-quality images of histological and cytological slides.
and explainability forming tasks that require common sense reasoning, such as
These images can be uploaded and remotely reviewed by pathologists and cytologists on
understanding the physical characteristics of the cells. More
a computer, while they can be available for review by multiple pathologists. However,
the organization and storageexplainable
of largeartificial
amounts intelligence
of digitizedapproaches could
data require highbecomputing
pro-
posed ininfrastructure,
power, storage space, technical the future. and backup capability. Furthermore, as a
Training limitations with Deep learning
consequence of digitized data, ethical issues algorithms require
are arising massivethe
concerning amounts
sharingofoflabeled
sensitive
inadequate samples
personal data. DL modelsdata to achieve
require good performance,
large amounts and thus,
of data for training, thousands
testing, of an-
and validation,
which are retrieved from notations must beTherefore,
hospital archives. performeda by pathologists.
regulatory framework is essential to
protect
Less patient’s
powerful rights and
in prob- ensure
Deep the security
learning of sensitive
algorithms medical
are mainly data and
designed for confidentiality.
classifica-
lems beyond classification problems, such as image recognition and natural language
tion processing. They are less effective for other types of problems,
such as regression, clustering, etc.
Cancers 2023, 15, 3981 27 of 33
Table 8. Deep learning limitations for lung cancer applications.
Limitation Property
Lack of interpretability and According to the review, only a few approaches focus on
explainability performing tasks that require common sense reasoning, such
as understanding the physical characteristics of the cells.
More explainable artificial intelligence approaches could be
proposed in the future.
Training limitations with Deep learning algorithms require massive amounts of labeled
inadequate samples data to achieve good performance, and thus, thousands of
annotations must be performed by pathologists.
Less powerful in problems Deep learning algorithms are mainly designed for
beyond classification classification problems, such as image recognition and natural
language processing. They are less effective for other types of
problems, such as regression, clustering, etc.
Lack of global generalization Deep learning algorithms often overfit the training data and
fail to generalize to new or unlabeled data. For example, a
deep learning model may perform well on images from a
specific microscopic scanner but poorly on images from a
different microscope.
High memory and The training of deep models using extremely large size of
computational cost images, such as biopsies, constitutes a very demanding
requirements process in terms of computational resources and training time
of the supervision.
5. Conclusions
The field of Digital Pathology is evolving rapidly and, in the following years, is ex-
pected to be an inextricable part of a pathology laboratory. As highlighted above, AI-based
approaches in Pathology are accompanied by several advantages, yet many challenges
remain to be considered. Research for lung cancer diagnosis, prognosis, and prediction
using DL methods is constantly improving to provide more accurate and reliable results.
Moreover, for quantitative tasks, such as PD-L1 TPS estimation, the need for AI-based
models is underlined because of their ability to provide reliable and objective assessment,
eliminating subjective estimations that lead to intra- and inter-observer variability. The
ongoing research and the efforts being made are at the forefront of transforming cancer
diagnosis and treatment.
Author Contributions: Conceptualization, N.G., A.T.T. and A.B.; data curation, A.D., E.B. and T.K.;
funding acquisition, N.G. and A.T.T.; methodology, G.N. and A.B.; supervision, N.G., A.T.T. and A.B.;
visualization, N.G.; writing—original draft preparation, A.D., E.B., T.K. and N.G.; writing—review
and editing, A.D., E.B., T.K., G.N. and A.B. All authors have read and agreed to the published version
of the manuscript.
Funding: This research has been co-financed by the European Union and Greek national funds
through the Operational Program Competitiveness, Entrepreneurship, and Innovation, under the call
RESEARCH-CREATE-INNOVATE: T2EDK-03660 (Project: Deep in Biopsies).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: All data that support the findings of this study are available from the
corresponding author upon reasonable request.
Acknowledgments: Athena Davri is supported by Greece and the European Union—European
Regional Development Fund (ERDF) under the Operational Program “Competitiveness Entrepreneur-
ship Innovation” (EPAnEK), NSRF 2014-2020 (MIS 5047236).
Conflicts of Interest: The authors declare no conflict of interest.
Cancers 2023, 15, 3981 28 of 33
References
1. Sung, H.; Ferlay, J.; Siegel, R.L.; Laversanne, M.; Soerjomataram, I.; Jemal, A.; Bray, F. Global Cancer Statistics 2020: GLOBOCAN
Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J. Clin. 2021, 71, 209–249. [CrossRef]
2. WHO Classification of Tumours Editorial Board (Ed.) WHO Classification of Tumours. In Thoracic Tumours, 5th ed.; International
Agency for Research on Cancer: Lyon, France, 2021; ISBN 978-92-832-4506-3.
3. Anand, K.; Phung, T.L.; Bernicker, E.H.; Cagle, P.T.; Olsen, R.J.; Thomas, J.S. Clinical Utility of Reflex Ordered Testing for
Molecular Biomarkers in Lung Adenocarcinoma. Clin. Lung Cancer 2020, 21, 437–442. [CrossRef]
4. Rodriguez-Canales, J.; Parra-Cuentas, E.; Wistuba, I.I. Diagnosis and Molecular Classification of Lung Cancer; Springer: Cham,
Switzerland, 2016; pp. 25–46. [CrossRef]
5. Malapelle, U.; Leprieur, E.G.; Kamga, P.T.; Chiasseu, M.T.; Rolfo, C. Editorial: Emerging Biomarkers for NSCLC: Recent Advances
in Diagnosis and Therapy. Front. Oncol. 2021, 11, 694578. [CrossRef] [PubMed]
6. Lindeman, N.I.; Cagle, P.T.; Aisner, D.L.; Arcila, M.E.; Beasley, M.B.; Bernicker, E.H.; Colasacco, C.; Dacic, S.; Hirsch, F.R.; Kerr,
K.; et al. Updated Molecular Testing Guideline for the Selection of Lung Cancer Patients for Treatment with Targeted Tyrosine
Kinase Inhibitors: Guideline from the College of American Pathologists, the International Association for the Study of Lung
Cancer, and the Association for Molecular Pathology. Arch. Pathol. Lab. Med. 2018, 142, 321–346. [CrossRef] [PubMed]
7. Osmani, L.; Askin, F.; Gabrielson, E.; Li, Q.K. Current WHO Guidelines and the Critical Role of Immunohistochemical Markers in
the Subclassification of Non-Small Cell Lung Carcinoma (NSCLC): Moving from Targeted Therapy to Immunotherapy. Semin.
Cancer Biol. 2018, 52, 103–109. [CrossRef] [PubMed]
8. Udall, M.; Rizzo, M.; Kenny, J.; Doherty, J.; Dahm, S.; Robbins, P.; Faulkner, E. PD-L1 Diagnostic Tests: A Systematic Literature
Review of Scoring Algorithms and Test-Validation Metrics. Diagn. Pathol. 2018, 13, 12. [CrossRef]
9. Kim, H.; Kwon, H.J.; Park, S.Y.; Park, E.; Chung, J.-H. PD-L1 Immunohistochemical Assays for Assessment of Therapeutic
Strategies Involving Immune Checkpoint Inhibitors in Non-Small Cell Lung Cancer: A Comparative Study. Oncotarget 2017, 8,
98524–98532. [CrossRef]
10. Wynes, M.W.; Sholl, L.M.; Dietel, M.; Schuuring, E.; Tsao, M.S.; Yatabe, Y.; Tubbs, R.R.; Hirsch, F.R. An International Interpretation
Study Using the ALK IHC Antibody D5F3 and a Sensitive Detection Kit Demonstrates High Concordance between ALK IHC and
ALK FISH and between Evaluators. J. Thorac. Oncol. 2014, 9, 631–638. [CrossRef]
11. Zacharias, M.; Absenger, G.; Kashofer, K.; Wurm, R.; Lindenmann, J.; Terbuch, A.; Konjic, S.; Sauer, S.; Gollowitsch, F.; Gorkiewicz,
G.; et al. Reflex Testing in Non-Small Cell Lung Carcinoma Using DNA- and RNA-Based next-Generation Sequencing—A
Single-Center Experience. Transl. Lung Cancer Res. 2021, 10, 4221–4234. [CrossRef]
12. Wang, S.; Yang, D.M.; Rong, R.; Zhan, X.; Fujimoto, J.; Liu, H.; Minna, J.; Wistuba, I.I.; Xie, Y.; Xiao, G. Artificial Intelligence in
Lung Cancer Pathology Image Analysis. Cancers 2019, 11, 1673. [CrossRef]
13. Baxi, V.; Edwards, R.; Montalto, M.; Saha, S. Digital Pathology and Artificial Intelligence in Translational Medicine and Clinical
Practice. Mod. Pathol. 2022, 35, 23–32. [CrossRef] [PubMed]
14. Travis, W.D.; Brambilla, E.; Noguchi, M.; Nicholson, A.G.; Geisinger, K.R.; Yatabe, Y.; Beer, D.G.; Powell, C.A.; Riely, G.J.; Van
Schil, P.E.; et al. International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory
Society International Multidisciplinary Classification of Lung Adenocarcinoma. J. Thorac. Oncol. 2011, 6, 244–285. [CrossRef]
[PubMed]
15. Bubendorf, L.; Lantuejoul, S.; de Langen, A.J.; Thunnissen, E. Nonsmall Cell Lung Carcinoma: Diagnostic Difficulties in Small
Biopsies and Cytological Specimens. Eur. Respir. Rev. 2017, 26, 170007. [CrossRef] [PubMed]
16. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.;
Brennan, S.E.; et al. The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews. Int. J. Surg. 2021, 88,
105906. [CrossRef]
17. Ouzzani, M.; Hammady, H.; Fedorowicz, Z.; Elmagarmid, A. Rayyan—A Web and Mobile App for Systematic Reviews. Syst. Rev.
2016, 5, 210. [CrossRef]
18. Jain, D.K.; Lakshmi, K.M.; Varma, K.P.; Ramachandran, M.; Bharati, S. Lung Cancer Detection Based on Kernel PCA-Convolution
Neural Network Feature Extraction and Classification by Fast Deep Belief Neural Network in Disease Management Using
Multimedia Data Sources. Comput. Intell. Neurosci. 2022, 2022, 3149406. [CrossRef]
19. Jiao, Y.; Yuan, J.; Qiang, Y.; Fei, S. Deep Embeddings and Logistic Regression for Rapid Active Learning in Histopathological
Images. Comput. Methods Programs Biomed. 2021, 212, 106464. [CrossRef]
20. Kanavati, F.; Toyokawa, G.; Momosaki, S.; Rambeau, M.; Kozuma, Y.; Shoji, F.; Yamazaki, K.; Takeo, S.; Iizuka, O.; Tsuneki, M.
Weakly-Supervised Learning for Lung Carcinoma Classification Using Deep Learning. Sci. Rep. 2020, 10, 9297. [CrossRef]
21. Moranguinho, J.; Pereira, T.; Ramos, B.; Morgado, J.; Costa, J.L.; Oliveira, H.P. Attention Based Deep Multiple Instance Learning
Approach for Lung Cancer Prediction Using Histopathological Images. In Proceedings of the 2021 43rd Annual International
Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Guadalajara, Jalisco, Mexico, 1–5 November 2021;
pp. 2852–2855. [CrossRef]
22. Tsuneki, M.; Kanavati, F. Weakly Supervised Learning for Multi-Organ Adenocarcinoma Classification in Whole Slide Images.
PLoS ONE 2022, 17, e0275378. [CrossRef]
Cancers 2023, 15, 3981 29 of 33
23. Civit-Masot, J.; Bañuls-Beaterio, A.; Domínguez-Morales, M.; Rivas-Pérez, M.; Muñoz-Saavedra, L.; Rodríguez Corral, J.M.
Non-Small Cell Lung Cancer Diagnosis Aid with Histopathological Images Using Explainable Deep Learning Techniques. Comput.
Methods Programs Biomed. 2022, 226, 107108. [CrossRef]
24. Kanavati, F.; Toyokawa, G.; Momosaki, S.; Takeoka, H.; Okamoto, M.; Yamazaki, K.; Takeo, S.; Iizuka, O.; Tsuneki, M. A Deep
Learning Model for the Classification of Indeterminate Lung Carcinoma in Biopsy Whole Slide Images. Sci. Rep. 2021, 11, 8110.
[CrossRef]
25. Wang, X.; Chen, H.; Gan, C.; Lin, H.; Dou, Q.; Tsougenis, E.; Huang, Q.; Cai, M.; Heng, P.-A. Weakly Supervised Deep Learning
for Whole Slide Lung Cancer Image Analysis. IEEE Trans. Cybern. 2020, 50, 3950–3962. [CrossRef]
26. Yang, H.; Chen, L.; Cheng, Z.; Yang, M.; Wang, J.; Lin, C.; Wang, Y.; Huang, L.; Chen, Y.; Peng, S.; et al. Deep Learning-Based
Six-Type Classifier for Lung Cancer and Mimics from Histopathological Whole Slide Images: A Retrospective Study. BMC Med.
2021, 19, 80. [CrossRef] [PubMed]
27. Yang, J.W.; Song, D.H.; An, H.J.; Seo, S.B. Classification of Subtypes Including LCNEC in Lung Cancer Biopsy Slides Using
Convolutional Neural Network from Scratch. Sci. Rep. 2022, 12, 1830. [CrossRef] [PubMed]
28. Kosaraju, S.; Park, J.; Lee, H.; Yang, J.W.; Kang, M. Deep Learning-Based Framework for Slide-Based Histopathological Image
Analysis. Sci. Rep. 2022, 12, 19075. [CrossRef]
29. Kriegsmann, M.; Haag, C.; Weis, C.A.; Steinbuss, G.; Warth, A.; Zgorzelski, C.; Muley, T.; Winter, H.; Eichhorn, M.E.; Eichhorn, F.;
et al. Deep Learning for the Classification of Small-Cell and Non-Small-Cell Lung Cancer. Cancers 2020, 12, 1604. [CrossRef]
[PubMed]
30. Ilié, M.; Benzaquen, J.; Tourniaire, P.; Heeke, S.; Ayache, N.; Delingette, H.; Long-Mira, E.; Lassalle, S.; Hamila, M.; Fayada, J.; et al.
Deep Learning Facilitates Distinguishing Histologic Subtypes of Pulmonary Neuroendocrine Tumors on Digital Whole-Slide
Images. Cancers 2022, 14, 1740. [CrossRef]
31. Chen, Y.; Yang, H.; Cheng, Z.; Chen, L.; Peng, S.; Wang, J.; Yang, M.; Lin, C.; Chen, Y.; Wang, Y.; et al. A Whole-Slide Image
(WSI)-Based Immunohistochemical Feature Prediction System Improves the Subtyping of Lung Cancer. Lung Cancer 2022, 165,
18–27. [CrossRef]
32. Mengash, H.A.; Alamgeer, M.; Maashi, M.; Othman, M.; Hamza, M.A.; Ibrahim, S.S.; Zamani, A.S.; Yaseen, I. Leveraging Marine
Predators Algorithm with Deep Learning for Lung and Colon Cancer Diagnosis. Cancers 2023, 15, 1591. [CrossRef]
33. Al-Jabbar, M.; Alshahrani, M.; Senan, E.M.; Ahmed, I.A. Histopathological Analysis for Detecting Lung and Colon Cancer
Malignancies Using Hybrid Systems with Fused Features. Bioengineering 2023, 10, 383. [CrossRef]
34. Wang, Z.; Bi, Y.; Pan, T.; Wang, X.; Bain, C.; Bassed, R.; Imoto, S.; Yao, J.; Daly, R.J.; Song, J. Targeting Tumor Heterogeneity:
Multiplex-Detection-Based Multiple Instance Learning for Whole Slide Image Classification. Bioinformatics 2023, 39, btad114.
[CrossRef]
35. Patil, A.; Diwakar, H.; Sawant, J.; Kurian, N.C.; Yadav, S.; Rane, S.; Bameta, T.; Sethi, A. Efficient Quality Control of Whole Slide
Pathology Images with Human-in-the-Loop Training. J. Pathol. Inform. 2023, 14, 100306. [CrossRef]
36. El-Ghany, S.A.; Azad, M.; Elmogy, M. Robustness Fine-Tuning Deep Learning Model for Cancers Diagnosis Based on Histopathol-
ogy Image Analysis. Diagnostics 2023, 13, 699. [CrossRef] [PubMed]
37. Zheng, Y.; Gindra, R.H.; Green, E.J.; Burks, E.J.; Betke, M.; Beane, J.E.; Kolachalama, V.B. A Graph-Transformer for Whole Slide
Image Classification. IEEE Trans. Med. Imaging 2022, 41, 3003–3015. [CrossRef] [PubMed]
38. Liu, Y.; Wang, H.; Song, K.; Sun, M.; Shao, Y.; Xue, S.; Li, L.; Li, Y.; Cai, H.; Jiao, Y.; et al. CroReLU: Cross-Crossing Space-Based
Visual Activation Function for Lung Cancer Pathology Image Recognition. Cancers 2022, 14, 5181. [CrossRef]
39. Attallah, O.; Aslan, M.F.; Sabanci, K. A Framework for Lung and Colon Cancer Diagnosis via Lightweight Deep Learning Models
and Transformation Methods. Diagnostics 2022, 12, 2926. [CrossRef] [PubMed]
40. Wang, Z.; Xu, Y.; Tian, L.; Chi, Q.; Zhao, F.; Xu, R.; Jin, G.; Liu, Y.; Zhen, J.; Zhang, S. A Multi-Task Convolutional Neural Network
for Lesion Region Segmentation and Classification of Non-Small Cell Lung Carcinoma. Diagnostics 2022, 12, 1849. [CrossRef]
41. Le Page, A.L.; Ballot, E.; Truntzer, C.; Derangère, V.; Ilie, A.; Rageot, D.; Bibeau, F.; Ghiringhelli, F. Using a Convolutional Neural
Network for Classification of Squamous and Non-Squamous Non-Small Cell Lung Cancer Based on Diagnostic Histopathology
HES Images. Sci. Rep. 2021, 11, 23912. [CrossRef] [PubMed]
42. Wang, X.; Yu, G.; Yan, Z.; Wan, L.; Wang, W.; Cui, L. Lung Cancer Subtype Diagnosis by Fusing Image-Genomics Data and Hybrid
Deep Networks. IEEE/ACM Trans. Comput. Biol. Bioinform. 2023, 20, 512–523. [CrossRef]
43. Zhao, L.; Xu, X.; Hou, R.; Zhao, W.; Zhong, H.; Teng, H.; Han, Y.; Fu, X.; Sun, J.; Zhao, J. Lung Cancer Subtype Classification Using
Histopathological Images Based on Weakly Supervised Multi-Instance Learning. Phys. Med. Biol. 2021, 66, 235013. [CrossRef]
44. Dehkharghanian, T.; Rahnamayan, S.; Riasatian, A.; Bidgoli, A.A.; Kalra, S.; Zaveri, M.; Babaie, M.; Seyed Sajadi, M.S.; Gonzalelz,
R.; Diamandis, P.; et al. Selection, Visualization, and Interpretation of Deep Features in Lung Adenocarcinoma and Squamous
Cell Carcinoma. Am. J. Pathol. 2021, 191, 2172–2183. [CrossRef] [PubMed]
45. Toğaçar, M. Disease Type Detection in Lung and Colon Cancer Images Using the Complement Approach of Inefficient Sets.
Comput. Biol. Med. 2021, 137, 104827. [CrossRef] [PubMed]
46. Carrillo-Perez, F.; Morales, J.C.; Castillo-Secilla, D.; Molina-Castro, Y.; Guillén, A.; Rojas, I.; Herrera, L.J. Non-Small-Cell Lung
Cancer Classification via RNA-Seq and Histology Imaging Probability Fusion. BMC Bioinform. 2021, 22, 454. [CrossRef] [PubMed]
47. Lu, M.Y.; Williamson, D.F.K.; Chen, T.Y.; Chen, R.J.; Barbieri, M.; Mahmood, F. Data-Efficient and Weakly Supervised Computa-
tional Pathology on Whole-Slide Images. Nat. Biomed. Eng. 2021, 5, 555–570. [CrossRef]
Cancers 2023, 15, 3981 30 of 33
48. Chen, C.L.; Chen, C.C.; Yu, W.H.; Chen, S.H.; Chang, Y.C.; Hsu, T.I.; Hsiao, M.; Yeh, C.Y.; Chen, C.Y. An Annotation-Free
Whole-Slide Training Approach to Pathological Classification of Lung Cancer Types Using Deep Learning. Nat. Commun. 2021,
12, 1193. [CrossRef]
49. Masud, M.; Sikder, N.; Nahid, A.-A.; Bairagi, A.K.; AlZain, M.A. A Machine Learning Approach to Diagnosing Lung and Colon
Cancer Using a Deep Learning-Based Classification Framework. Sensors 2021, 21, 748. [CrossRef]
50. Wang, Y.; Coudray, N.; Zhao, Y.; Li, F.; Hu, C.; Zhang, Y.-Z.; Imoto, S.; Tsirigos, A.; Webb, G.I.; Daly, R.J.; et al. HEAL: An
Automated Deep Learning Framework for Cancer Histopathology Image Analysis. Bioinformatics 2021, 37, 4291–4295. [CrossRef]
51. Kobayashi, K.; Bolatkan, A.; Shiina, S.; Hamamoto, R. Fully-Connected Neural Networks with Reduced Parameterization for
Predicting Histological Types of Lung Cancer from Somatic Mutations. Biomolecules 2020, 10, 1249. [CrossRef]
52. Xu, X.; Hou, R.; Zhao, W.; Teng, H.; Sun, J.; Zhao, J. A Weak Supervision-Based Framework for Automatic Lung Cancer
Classification on Whole Slide Image. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering
in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020; pp. 1372–1375. [CrossRef]
53. Yu, K.H.; Wang, F.; Berry, G.J.; Ré, C.; Altman, R.B.; Snyder, M.; Kohane, I.S. Classifying Non-Small Cell Lung Cancer Types and
Transcriptomic Subtypes Using Convolutional Neural Networks. J. Am. Med. Inform. Assoc. 2020, 27, 757–769. [CrossRef]
54. Shi, X.; Su, H.; Xing, F.; Liang, Y.; Qu, G.; Yang, L. Graph Temporal Ensembling Based Semi-Supervised Convolutional Neural
Network with Noisy Labels for Histopathology Image Analysis. Med. Image Anal. 2020, 60, 101624. [CrossRef]
55. Khosravi, P.; Kazemi, E.; Imielinski, M.; Elemento, O.; Hajirasouliha, I. Deep Convolutional Neural Networks Enable Discrimina-
tion of Heterogeneous Digital Pathology Images. EBioMedicine 2018, 27, 317–328. [CrossRef] [PubMed]
56. Coudray, N.; Ocampo, P.S.; Sakellaropoulos, T.; Narula, N.; Snuderl, M.; Fenyö, D.; Moreira, A.L.; Razavian, N.; Tsirigos, A.
Classification and Mutation Prediction from Non–Small Cell Lung Cancer Histopathology Images Using Deep Learning. Nat.
Med. 2018, 24, 1559–1567. [CrossRef] [PubMed]
57. Hou, L.; Samaras, D.; Kurc, T.M.; Gao, Y.; Davis, J.E.; Saltz, J.H. Patch-Based Convolutional Neural Network for Whole Slide
Tissue Image Classification. In Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
Las Vegas, NV, USA, 27–30 June 2016; Volume 2016, pp. 2424–2433. [CrossRef]
58. Dolezal, J.M.; Srisuwananukorn, A.; Karpeyev, D.; Ramesh, S.; Kochanny, S.; Cody, B.; Mansfield, A.S.; Rakshit, S.; Bansal, R.; Bois,
M.C.; et al. Uncertainty-Informed Deep Learning Models Enable High-Confidence Predictions for Digital Histopathology. Nat.
Commun. 2022, 13, 6572. [CrossRef]
59. Tavolara, T.E.; Gurcan, M.N.; Niazi, M.K.K. Contrastive Multiple Instance Learning: An Unsupervised Framework for Learning
Slide-Level Representations of Whole Slide Histopathology Images without Labels. Cancers 2022, 14, 5778. [CrossRef] [PubMed]
60. Liu, M.; Li, L.; Wang, H.; Guo, X.; Liu, Y.; Li, Y.; Song, K.; Shao, Y.; Wu, F.; Zhang, J.; et al. A Multilayer Perceptron-Based Model
Applied to Histopathology Image Classification of Lung Adenocarcinoma Subtypes. Front. Oncol. 2023, 13, 1172234. [CrossRef]
[PubMed]
61. Antonio, V.A.A.; Ono, N.; Saito, A.; Sato, T.; Altaf-Ul-Amin, M.; Kanaya, S. Classification of Lung Adenocarcinoma Transcriptome
Subtypes from Pathological Images Using Deep Convolutional Networks. Int. J. Comput. Assist. Radiol. Surg. 2018, 13, 1905–1913.
[CrossRef]
62. Sadhwani, A.; Chang, H.-W.; Behrooz, A.; Brown, T.; Auvigne-Flament, I.; Patel, H.; Findlater, R.; Velez, V.; Tan, F.; Tekiela, K.;
et al. Comparative Analysis of Machine Learning Approaches to Classify Tumor Mutation Burden in Lung Adenocarcinoma
Using Histopathology Images. Sci. Rep. 2021, 11, 16605. [CrossRef]
63. Wei, J.W.; Tafe, L.J.; Linnik, Y.A.; Vaickus, L.J.; Tomita, N.; Hassanpour, S. Pathologist-Level Classification of Histologic Patterns
on Resected Lung Adenocarcinoma Slides with Deep Neural Networks. Sci. Rep. 2019, 9, 3358. [CrossRef]
64. Gertych, A.; Swiderska-Chadaj, Z.; Ma, Z.; Ing, N.; Markiewicz, T.; Cierniak, S.; Salemi, H.; Guzman, S.; Walts, A.E.; Knudsen,
B.S. Convolutional Neural Networks Can Accurately Distinguish Four Histologic Growth Patterns of Lung Adenocarcinoma in
Digital Slides. Sci. Rep. 2019, 9, 1483. [CrossRef]
65. DiPalma, J.; Suriawinata, A.A.; Tafe, L.J.; Torresani, L.; Hassanpour, S. Resolution-Based Distillation for Efficient Histology Image
Classification. Artif. Intell. Med. 2021, 119, 102136. [CrossRef]
66. Xiao, W.; Jiang, Y.; Yao, Z.; Zhou, X.; Sui, X.; Zheng, Y. LAD-GCN: Automatic Diagnostic Framework for Quantitative Estimation
of Growth Patterns during Clinical Evaluation of Lung Adenocarcinoma. Front. Physiol. 2022, 13, 946099. [CrossRef]
67. Sheikh, T.S.; Kim, J.Y.; Shim, J.; Cho, M. Unsupervised Learning Based on Multiple Descriptors for WSIs Diagnosis. Diagnostics
2022, 12, 1480. [CrossRef] [PubMed]
68. Gao, Y.; Ding, Y.; Xiao, W.; Yao, Z.; Zhou, X.; Sui, X.; Zhao, Y.; Zheng, Y. A Semi-Supervised Learning Framework for Micropapillary
Adenocarcinoma Detection. Int. J. Comput. Assist. Radiol. Surg. 2022, 17, 639–648. [CrossRef]
69. Maleki, F.; Ovens, K.; Gupta, R.; Reinhold, C.; Spatz, A.; Forghani, R. Generalizability of Machine Learning Models: Quantitative
Evaluation of Three Methodological Pitfalls. Radiol. Artif. Intell. 2023, 5, e220028. [CrossRef]
70. Tan, Z.; Xue, H.; Sun, Y.; Zhang, C.; Song, Y.; Qi, Y. The Role of Tumor Inflammatory Microenvironment in Lung Cancer. Front.
Pharmacol. 2021, 12, 688625. [CrossRef] [PubMed]
71. Barmpoutis, P.; Di Capite, M.; Kayhanian, H.; Waddingham, W.; Alexander, D.C.; Jansen, M.; Kwong, F.N.K. Tertiary Lymphoid
Structures (TLS) Identification and Density Assessment on H&E-Stained Digital Slides of Lung Cancer. PLoS ONE 2021, 16,
e0256907. [CrossRef]
Cancers 2023, 15, 3981 31 of 33
72. Shim, W.S.; Yim, K.; Kim, T.-J.; Sung, Y.E.; Lee, G.; Hong, J.H.; Chun, S.H.; Kim, S.; An, H.J.; Na, S.J.; et al. DeepRePath: Identifying
the Prognostic Features of Early-Stage Lung Adenocarcinoma Using Multi-Scale Pathology Images and Deep Convolutional
Neural Networks. Cancers 2021, 13, 308. [CrossRef]
73. Wu, Z.; Wang, L.; Li, C.; Cai, Y.; Liang, Y.; Mo, X.; Lu, Q.; Dong, L.; Liu, Y. DeepLRHE: A Deep Convolutional Neural Network
Framework to Evaluate the Risk of Lung Cancer Recurrence and Metastasis from Histopathology Images. Front. Genet. 2020, 11,
768. [CrossRef] [PubMed]
74. Wang, S.; Wang, T.; Yang, L.; Yang, D.M.; Fujimoto, J.; Yi, F.; Luo, X.; Yang, Y.; Yao, B.; Lin, S.Y.; et al. ConvPath: A Software Tool
for Lung Adenocarcinoma Digital Pathological Image Analysis Aided by a Convolutional Neural Network. EBioMedicine 2019,
50, 103–110. [CrossRef]
75. Wang, S.; Rong, R.; Yang, D.M.; Fujimoto, J.; Yan, S.; Cai, L.; Yang, L.; Luo, D.; Behrens, C.; Parra, E.R.; et al. Computational
Staining of Pathology Images to Study the Tumor Microenvironment in Lung Cancer. Cancer Res. 2020, 80, 2056–2066. [CrossRef]
76. Shvetsov, N.; Grønnesby, M.; Pedersen, E.; Møllersen, K.; Busund, L.-T.R.; Schwienbacher, R.; Bongo, L.A.; Kilvaer, T.K. A
Pragmatic Machine Learning Approach to Quantify Tumor-Infiltrating Lymphocytes in Whole Slide Images. Cancers 2022, 14,
2974. [CrossRef] [PubMed]
77. Guo, H.; Diao, L.; Zhou, X.; Chen, J.N.; Zhou, Y.; Fang, Q.; He, Y.; Dziadziuszko, R.; Zhou, C.; Hirsch, F.R. Artificial Intelligence-
Based Analysis for Immunohistochemistry Staining of Immune Checkpoints to Predict Resected Non-Small Cell Lung Cancer
Survival and Relapse. Transl. Lung Cancer Res. 2021, 10, 2452–2474. [CrossRef] [PubMed]
78. Pham, H.H.N.; Futakuchi, M.; Bychkov, A.; Furukawa, T.; Kuroda, K.; Fukuoka, J. Detection of Lung Cancer Lymph Node
Metastases from Whole-Slide Histopathologic Images Using a Two-Step Deep Learning Approach. Am. J. Pathol. 2019, 189,
2428–2439. [CrossRef]
79. Raczkowski,
˛ Ł.; Paśnik, I.; Kukiełka, M.; Nicoś, M.; Budzinska, M.A.; Kucharczyk, T.; Szumiło, J.; Krawczyk, P.; Crosetto, N.;
Szczurek, E. Deep Learning-Based Tumor Microenvironment Segmentation Is Predictive of Tumor Mutations and Patient Survival
in Non-Small-Cell Lung Cancer. BMC Cancer 2022, 22, 1001. [CrossRef]
80. Qaiser, T.; Lee, C.-Y.; Vandenberghe, M.; Yeh, J.; Gavrielides, M.A.; Hipp, J.; Scott, M.; Reischl, J. Usability of Deep Learning
and H&E Images Predict Disease Outcome-Emerging Tool to Optimize Clinical Trials. NPJ Precis. Oncol. 2022, 6, 37. [CrossRef]
[PubMed]
81. Sakamoto, T.; Furukawa, T.; Pham, H.H.N.; Kuroda, K.; Tabata, K.; Kashima, Y.; Okoshi, E.N.; Morimoto, S.; Bychkov, A.;
Fukuoka, J. A Collaborative Workflow between Pathologists and Deep Learning for the Evaluation of Tumour Cellularity in Lung
Adenocarcinoma. Histopathology 2022, 81, 758–769. [CrossRef]
82. Hattori, H.; Sakashita, S.; Tsuboi, M.; Ishii, G.; Tanaka, T. Tumor-Identification Method for Predicting Recurrence of Early-Stage
Lung Adenocarcinoma Using Digital Pathology Images by Machine Learning. J. Pathol. Inform. 2023, 14, 100175. [CrossRef]
83. Pan, X.; Lin, H.; Han, C.; Feng, Z.; Wang, Y.; Lin, J.; Qiu, B.; Yan, L.; Li, B.; Xu, Z.; et al. Computerized Tumor-Infiltrating
Lymphocytes Density Score Predicts Survival of Patients with Resectable Lung Adenocarcinoma. iScience 2022, 25, 105605.
[CrossRef]
84. Aprupe, L.; Litjens, G.; Brinker, T.J.; Van Der Laak, J.; Grabe, N. Robust and Accurate Quantification of Biomarkers of Immune
Cells in Lung Cancer Micro-Environment Using Deep Convolutional Neural Networks. PeerJ 2019, 7, e6335. [CrossRef]
85. Yu, F.; Wang, X.; Sali, R.; Li, R. Single-Cell Heterogeneity-Aware Transformer-Guided Multiple Instance Learning for Cancer
Aneuploidy Prediction from Whole Slide Histopathology Images. IEEE J. Biomed. Heal. Inform. 2023, 1–11. [CrossRef]
86. Levy-Jurgenson, A.; Tekpli, X.; Kristensen, V.N.; Yakhini, Z. Spatial Transcriptomics Inferred from Pathology Whole-Slide Images
Links Tumor Heterogeneity to Survival in Breast and Lung Cancer. Sci. Rep. 2020, 10, 18802. [CrossRef] [PubMed]
87. Yang, Y.; Yang, J.; Liang, Y.; Liao, B.; Zhu, W.; Mo, X.; Huang, K. Identification and Validation of Efficacy of Immunological
Therapy for Lung Cancer From Histopathological Images Based on Deep Learning. Front. Genet. 2021, 12, 642981. [CrossRef]
[PubMed]
88. Hu, J.; Cui, C.; Yang, W.; Huang, L.; Yu, R.; Liu, S.; Kong, Y. Using Deep Learning to Predict Anti-PD-1 Response in Melanoma
and Lung Cancer Patients from Histopathology Images. Transl. Oncol. 2021, 14, 100921. [CrossRef] [PubMed]
89. Terada, Y.; Takahashi, T.; Hayakawa, T.; Ono, A.; Kawata, T.; Isaka, M.; Muramatsu, K.; Tone, K.; Kodama, H.; Imai, T.; et al.
Artificial Intelligence–Powered Prediction of ALK Gene Rearrangement in Patients with Non–Small-Cell Lung Cancer. JCO Clin.
Cancer Inform. 2022, 6, e2200070. [CrossRef]
90. Tomita, N.; Tafe, L.J.; Suriawinata, A.A.; Tsongalis, G.J.; Nasir-Moin, M.; Dragnev, K.; Hassanpour, S. Predicting Oncogene
Mutations of Lung Cancer Using Deep Learning and Histopathologic Features on Whole-Slide Images. Transl. Oncol. 2022, 24,
101494. [CrossRef]
91. Huang, K.; Mo, Z.; Zhu, W.; Liao, B.; Yang, Y.; Wu, F.-X. Prediction of Target-Drug Therapy by Identifying Gene Mutations in
Lung Cancer with Histopathological Stained Image and Deep Learning Techniques. Front. Oncol. 2021, 11, 642945. [CrossRef]
92. Mayer, C.; Ofek, E.; Fridrich, D.E.; Molchanov, Y.; Yacobi, R.; Gazy, I.; Hayun, I.; Zalach, J.; Paz-Yaacov, N.; Barshack, I. Direct
Identification of ALK and ROS1 Fusions in Non-Small Cell Lung Cancer from Hematoxylin and Eosin-Stained Slides Using Deep
Learning Algorithms. Mod. Pathol. 2022, 35, 1882–1887. [CrossRef]
93. Niu, Y.; Wang, L.; Zhang, X.; Han, Y.; Yang, C.; Bai, H.; Huang, K.; Ren, C.; Tian, G.; Yin, S.; et al. Predicting Tumor Mutational
Burden from Lung Adenocarcinoma Histopathological Images Using Deep Learning. Front. Oncol. 2022, 12, 927426. [CrossRef]
Cancers 2023, 15, 3981 32 of 33
94. Dammak, S.; Cecchini, M.J.; Breadner, D.; Ward, A.D. Using Deep Learning to Predict Tumor Mutational Burden from Scans of
H&E-Stained Multicenter Slides of Lung Squamous Cell Carcinoma. J. Med. Imaging 2023, 10, 017502. [CrossRef]
95. Pao, J.J.; Biggs, M.; Duncan, D.; Lin, D.I.; Davis, R.; Huang, R.S.P.; Ferguson, D.; Janovitz, T.; Hiemenz, M.C.; Eddy, N.R.; et al.
Predicting EGFR Mutational Status from Pathology Images Using a Real-World Dataset. Sci. Rep. 2023, 13, 4404. [CrossRef]
96. Li, X.; Cen, M.; Xu, J.; Zhang, H.; Xu, X.S. Improving Feature Extraction from Histopathological Images through a Fine-Tuning
ImageNet Model. J. Pathol. Inform. 2022, 13, 100115. [CrossRef]
97. Teramoto, A.; Tsukamoto, T.; Kiriyama, Y.; Fujita, H. Automated Classification of Lung Cancer Types from Cytological Images
Using Deep Convolutional Neural Networks. Biomed Res. Int. 2017, 2017, 4067832. [CrossRef] [PubMed]
98. Teramoto, A.; Tsukamoto, T.; Yamada, A.; Kiriyama, Y.; Imaizumi, K.; Saito, K.; Fujita, H. Deep Learning Approach to Classification
of Lung Cytological Images: Two-Step Training Using Actual and Synthesized Images by Progressive Growing of Generative
Adversarial Networks. PLoS ONE 2020, 15, 0229951. [CrossRef] [PubMed]
99. Xie, X.; Fu, C.C.; Lv, L.; Ye, Q.; Yu, Y.; Fang, Q.; Zhang, L.; Hou, L.; Wu, C. Deep Convolutional Neural Network-Based
Classification of Cancer Cells on Cytological Pleural Effusion Images. Mod. Pathol. 2022, 35, 609–614. [CrossRef] [PubMed]
100. Lin, C.-K.; Chang, J.; Huang, C.-C.; Wen, Y.-F.; Ho, C.-C.; Cheng, Y.-C. Effectiveness of Convolutional Neural Networks in
the Interpretation of Pulmonary Cytologic Images in Endobronchial Ultrasound Procedures. Cancer Med. 2021, 10, 9047–9057.
[CrossRef]
101. Teramoto, A.; Kiriyama, Y.; Tsukamoto, T.; Sakurai, E.; Michiba, A.; Imaizumi, K.; Saito, K.; Fujita, H. Weakly Supervised Learning
for Classification of Lung Cytological Images Using Attention-Based Multiple Instance Learning. Sci. Rep. 2021, 11, 20317.
[CrossRef]
102. Tsukamoto, T.; Teramoto, A.; Yamada, A.; Kiriyama, Y.; Sakurai, E.; Michiba, A.; Imaizumi, K.; Fujita, H. Comparison of Fine-
Tuned Deep Convolutional Neural Networks for the Automated Classification of Lung Cancer Cytology Images with Integration
of Additional Classifiers. Asian Pac. J. Cancer Prev. 2022, 23, 1315–1324. [CrossRef]
103. Gonzalez, D.; Dietz, R.L.; Pantanowitz, L. Feasibility of a Deep Learning Algorithm to Distinguish Large Cell Neuroendocrine
from Small Cell Lung Carcinoma in Cytology Specimens. Cytopathology 2020, 31, 426–431. [CrossRef]
104. Wang, C.-W.; Khalil, M.-A.; Lin, Y.-J.; Lee, Y.-C.; Huang, T.-W.; Chao, T.-K. Deep Learning Using Endobronchial-Ultrasound-
Guided Transbronchial Needle Aspiration Image to Improve the Overall Diagnostic Yield of Sampling Mediastinal Lym-
phadenopathy. Diagnostics 2022, 12, 2234. [CrossRef]
105. Doroshow, D.B.; Bhalla, S.; Beasley, M.B.; Sholl, L.M.; Kerr, K.M.; Gnjatic, S.; Wistuba, I.I.; Rimm, D.L.; Tsao, M.S.; Hirsch, F.R.
PD-L1 as a Biomarker of Response to Immune-Checkpoint Inhibitors. Nat. Rev. Clin. Oncol. 2021, 18, 345–362. [CrossRef]
106. Hopkins, E.; Moffat, D.; Parkinson, I.; Robinson, P.; Jersmann, H.; Dougherty, B.; Birader, M.I.; Francis, K.; Nguyen, P. Cell Block
Samples from Endobronchial Ultrasound Transbronchial Needle Aspiration Provide Sufficient Material for Ancillary Testing in
Lung Cancer—A Quaternary Referral Centre Experience. J. Thorac. Dis. 2016, 8, 2544–2550. [CrossRef] [PubMed]
107. Herbst, R.S.; Baas, P.; Kim, D.-W.; Felip, E.; Pérez-Gracia, J.L.; Han, J.-Y.; Molina, J.; Kim, J.-H.; Arvis, C.D.; Ahn, M.-J.; et al.
Pembrolizumab versus Docetaxel for Previously Treated, PD-L1-Positive, Advanced Non-Small-Cell Lung Cancer (KEYNOTE-
010): A Randomised Controlled Trial. Lancet 2016, 387, 1540–1550. [CrossRef] [PubMed]
108. Reck, M.; Rodríguez-Abreu, D.; Robinson, A.G.; Hui, R.; Csőszi, T.; Fülöp, A.; Gottfried, M.; Peled, N.; Tafreshi, A.; Cuffe, S.; et al.
Pembrolizumab versus Chemotherapy for PD-L1–Positive Non–Small-Cell Lung Cancer. N. Engl. J. Med. 2016, 375, 1823–1833.
[CrossRef] [PubMed]
109. Lantuejoul, S.; Damotte, D.; Hofman, V.; Adam, J. Programmed Death Ligand 1 Immunohistochemistry in Non-Small Cell Lung
Carcinoma. J. Thorac. Dis. 2019, 11, S89–S101. [CrossRef] [PubMed]
110. Hondelink, L.M.; Hüyük, M.; Postmus, P.E.; Smit, V.T.H.B.M.; Blom, S.; von der Thüsen, J.H.; Cohen, D. Development and Valida-
tion of a Supervised Deep Learning Algorithm for Automated Whole-Slide Programmed Death-Ligand 1 Tumour Proportion
Score Assessment in Non-Small Cell Lung Cancer. Histopathology 2022, 80, 635–647. [CrossRef]
111. Liu, J.; Zheng, Q.; Mu, X.; Zuo, Y.; Xu, B.; Jin, Y.; Wang, Y.; Tian, H.; Yang, Y.; Xue, Q.; et al. Automated Tumor Proportion Score
Analysis for PD-L1 (22C3) Expression in Lung Squamous Cell Carcinoma. Sci. Rep. 2021, 11, 15907. [CrossRef]
112. Wang, X.; Chen, P.; Ding, G.; Xing, Y.; Tang, R.; Peng, C.; Ye, Y.; Fu, Q. Dual-Scale Categorization Based Deep Learning to Evaluate
Programmed Cell Death Ligand 1 Expression in Non-Small Cell Lung Cancer. Medicine 2021, 100, E25994. [CrossRef]
113. Sha, L.; Osinski, B.L.; Ho, I.Y.; Tan, T.L.; Willis, C.; Weiss, H.; Beaubier, N.; Mahon, B.M.; Taxter, T.J.; Yip, S.S.F. Multi-Field-of-View
Deep Learning Model Predicts Nonsmall Cell Lung Cancer Programmed Death-Ligand 1 Status from Whole-Slide Hematoxylin
and Eosin Images. J. Pathol. Inform. 2019, 10, 24. [CrossRef]
114. Kapil, A.; Meier, A.; Steele, K.; Rebelatto, M.; Nekolla, K.; Haragan, A.; Silva, A.; Zuraw, A.; Barker, C.; Scott, M.L.; et al. Domain
Adaptation-Based Deep Learning for Automated Tumor Cell (TC) Scoring and Survival Analysis on PD-L1 Stained Tissue Images.
IEEE Trans. Med. Imaging 2021, 40, 2513–2523. [CrossRef]
115. Kapil, A.; Meier, A.; Zuraw, A.; Steele, K.E.; Rebelatto, M.C.; Schmidt, G.; Brieu, N. Deep Semi Supervised Generative Learning
for Automated Tumor Proportion Scoring on NSCLC Tissue Needle Biopsies. Sci. Rep. 2018, 8, 17343. [CrossRef]
116. Wu, J.; Liu, C.; Liu, X.; Sun, W.; Li, L.; Gao, N.; Zhang, Y.; Yang, X.; Zhang, J.; Wang, H.; et al. Artificial Intelligence-Assisted
System for Precision Diagnosis of PD-L1 Expression in Non-Small Cell Lung Cancer. Mod. Pathol. 2022, 35, 403–411. [CrossRef]
[PubMed]
Cancers 2023, 15, 3981 33 of 33
117. Cheng, G.; Zhang, F.; Xing, Y.; Hu, X.; Zhang, H.; Chen, S.; Li, M.; Peng, C.; Ding, G.; Zhang, D.; et al. Artificial Intelligence-
Assisted Score Analysis for Predicting the Expression of the Immunotherapy Biomarker PD-L1 in Lung Cancer. Front. Immunol.
2022, 13, 893198. [CrossRef] [PubMed]
118. Choi, S.; Cho, S.I.; Ma, M.; Park, S.; Pereira, S.; Aum, B.J.; Shin, S.; Paeng, K.; Yoo, D.; Jung, W.; et al. Artificial Intelligence–Powered
Programmed Death Ligand 1 Analyser Reduces Interobserver Variation in Tumour Proportion Score for Non–Small Cell Lung
Cancer with Better Prediction of Immunotherapy Response. Eur. J. Cancer 2022, 170, 17–26. [CrossRef] [PubMed]
119. Huang, Z.; Chen, L.; Lv, L.; Fu, C.C.; Jin, Y.; Zheng, Q.; Wang, B.; Ye, Q.; Fang, Q.; Li, Y. A New AI-Assisted Scoring System for
PD-L1 Expression in NSCLC. Comput. Methods Programs Biomed. 2022, 221, 106829. [CrossRef]
120. Sadoughi, F.; Kazemy, Z.; Hamedan, F.; Owji, L.; Rahmanikatigari, M.; Talebi Azadboni, T. Artificial Intelligence Methods for the
Diagnosis of Breast Cancer by Image Processing: A Review. Breast Cancer Targets Ther. 2018, 10, 219–230. [CrossRef]
121. Robertson, S.; Azizpour, H.; Smith, K.; Hartman, J. Digital Image Analysis in Breast Pathology—From Image Processing
Techniques to Artificial Intelligence. Transl. Res. 2018, 194, 19–35. [CrossRef]
122. Naik, N.; Tokas, T.; Shetty, D.K.; Hameed, B.M.Z.; Shastri, S.; Shah, M.J.; Ibrahim, S.; Rai, B.P.; Chłosta, P.; Somani, B.K. Role of
Deep Learning in Prostate Cancer Management: Past, Present and Future Based on a Comprehensive Literature Review. J. Clin.
Med. 2022, 11, 3575. [CrossRef]
123. Reddy, B.S.K.; Balakrishna, R. A Review on Detection of Prostate Cancer Using Deep Learning Techniques. AIP Conf. Proc. 2022,
2463, 020004. [CrossRef]
124. Davri, A.; Birbas, E.; Kanavos, T.; Ntritsos, G.; Giannakeas, N.; Tzallas, A.T.; Batistatou, A. Deep Learning on Histopathological
Images for Colorectal Cancer Diagnosis: A Systematic Review. Diagnostics 2022, 12, 837. [CrossRef]
125. Rompianesi, G.; Pegoraro, F.; Ceresa, C.D.; Montalti, R.; Troisi, R.I. Artificial Intelligence in the Diagnosis and Management of
Colorectal Cancer Liver Metastases. World J. Gastroenterol. 2022, 28, 108–122. [CrossRef]
126. Thakur, N.; Yoon, H.; Chong, Y. Current Trends of Artificial Intelligence for Colorectal Cancer Pathology Image Analysis: A
Systematic Review. Cancers 2020, 12, 1884. [CrossRef]
127. Li, Y.; Chen, D.; Wu, X.; Yang, W.; Chen, Y. A Narrative Review of Artificial Intelligence-Assisted Histopathologic Diagnosis and
Decision-Making for Non-Small Cell Lung Cancer: Achievements and Limitations. J. Thorac. Dis. 2021, 13, 7006–7020. [CrossRef]
[PubMed]
128. Tanaka, I.; Furukawa, T.; Morise, M. The Current Issues and Future Perspective of Artificial Intelligence for Developing New
Treatment Strategy in Non-Small Cell Lung Cancer: Harmonization of Molecular Cancer Biology and Artificial Intelligence.
Cancer Cell Int. 2021, 21, 454. [CrossRef] [PubMed]
129. Singhal, N.; Soni, S.; Bonthu, S.; Chattopadhyay, N.; Samanta, P.; Joshi, U.; Jojera, A.; Chharchhodawala, T.; Agarwal, A.; Desai,
M.; et al. A Deep Learning System for Prostate Cancer Diagnosis and Grading in Whole Slide Images of Core Needle Biopsies.
Sci. Rep. 2022, 12, 3383. [CrossRef] [PubMed]
130. Li, D.; Han, X.; Gao, J.; Zhang, Q.; Yang, H.; Liao, S.; Guo, H.; Zhang, B. Deep Learning in Prostate Cancer Diagnosis Using
Multiparametric Magnetic Resonance Imaging with Whole-Mount Histopathology Referenced Delineations. Front. Med. 2022, 8,
810995. [CrossRef]
131. Tătaru, O.S.; Vartolomei, M.D.; Rassweiler, J.J.; Virgil, O.; Lucarelli, G.; Porpiglia, F.; Amparore, D.; Manfredi, M.; Carrieri, G.;
Falagario, U.; et al. Artificial Intelligence and Machine Learning in Prostate Cancer Patient Management—Current Trends and
Future Perspectives. Diagnostics 2021, 11, 354. [CrossRef]
132. de Sousa, V.M.L.; Carvalho, L. Heterogeneity in Lung Cancer. Pathobiology 2018, 85, 96–107. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

Cancers 15 03981 v2

Uploaded by

Copyright:

Available Formats

Cancers 15 03981 v2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cancers 15 03981 v2

Uploaded by

Copyright:

Available Formats

cancers

1 Department of Pathology, Faculty of Medicine, School of Health Sciences, University of Ioannina,

Cancers 2023, 15, 3981. https://doi.org/10.3390/cancers15153981 https://www.mdpi.com/journal/cancers

2. Materials and Methods

2.1. Search Strategy

2.2. Study Eligibility Criteria

2.3. Study Selection

2.4. Data Extraction

3.1. Histology 3.1. Histology

1st Author, Year Technical Method Classification Dataset Performance Metrics

3.1.2. Lung Cancer Classification

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

3.1.3. NSCLC Subtypes Classification

3.1.4. Lung ADC Predominant Architectural Patterns Classification

1st Author, Year Technical Method Classification Dataset Performance Metrics

3.1.5. Prediction of Prognosis and Survival

3.1.6. Prediction of Significant Molecular Alterations

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

1st Author, Year Technical Method Classification Dataset Performance Metrics

3.3. PD-L1 Expression Status

3.4. Deep Learning Approaches

As a result of the limited implementation effort needed, the vast majority of DL

Teramoto et al. introduced a progressive growing approach of GANs (PGGAN) combined

Figure 2. Summary of DL methods for lung cancer.

Cancers 2023, 15, 3981 27 of 35

Table 8. Deep learning limitations for lung cancer applications.

You might also like