diagnostics

Review
Application of Artificial Intelligence in Pathology: Trends
and Challenges
Inho Kim 1 , Kyungmin Kang 1 , Youngjae Song 1 and Tae-Jung Kim 2, *

1 College of Medicine, The Catholic University of Korea, 222 Banpo-daero, Seocho-gu,

Seoul 06591, Republic of Korea
2 Department of Hospital Pathology, Yeouido St. Mary’s Hospital, College of Medicine, The Catholic University
of Korea, 10, 63-ro, Yeongdeungpo-gu, Seoul 07345, Republic of Korea
* Correspondence: kimecho@catholic.ac.kr; Tel.: +82-2-3779-2157

Abstract: Given the recent success of artificial intelligence (AI) in computer vision applications, many
pathologists anticipate that AI will be able to assist them in a variety of digital pathology tasks.
Simultaneously, tremendous advancements in deep learning have enabled a synergy with artificial
intelligence (AI), allowing for image-based diagnosis on the background of digital pathology. There
are efforts for developing AI-based tools to save pathologists time and eliminate errors. Here, we
describe the elements in the development of computational pathology (CPATH), its applicability
to AI development, and the challenges it faces, such as algorithm validation and interpretability,
computing systems, reimbursement, ethics, and regulations. Furthermore, we present an overview of
novel AI-based approaches that could be integrated into pathology laboratory workflows.

Keywords: artificial intelligence; computational pathology; digital pathology; histopathology image

analysis; deep learning






Citation: Kim, I.; Kang, K.; Song, Y.; 1. Introduction
Kim, T.-J. Application of Artificial
Pathologists examine pathology slides under a microscope. To diagnose diseases
Intelligence in Pathology: Trends and
with these glass slides, many traditional technologies, such as hematoxylin and eosin
Challenges. Diagnostics 2022, 12, 2794.
(H&E) staining and special staining, have been used. However, even for experienced
https://doi.org/10.3390/
pathologists, intra- and interobserver disagreement cannot be avoided through visual
diagnostics12112794
observation and subjective interpretation [1]. This limited agreement has resulted in the
Academic Editor: Masayuki Tsuneki necessity of computational methods for pathological diagnosis [2–4]. Because automated
Received: 26 September 2022
approaches can achieve reliable results, digital imaging is the first step in computer-aided
Accepted: 11 November 2022
analysis [5]. When compared to traditional digital imaging technologies that process static
Published: 15 November 2022
images through cameras, whole-slide imaging (WSI) is a more advanced and widely used
technology in pathology [6].
Publisher’s Note: MDPI stays neutral
Digital pathology refers to the environment that includes tools and systems for dig-
with regard to jurisdictional claims in
itizing pathology slides and associated metadata, in addition their storage, evaluation,
published maps and institutional affil-
and analysis, as well as supporting infrastructure. WSI has been proven in multiple studies
iations.
to have an excellent correlation with traditional light microscopy diagnosis [7] and to
be a reliable tool for routine surgical pathology diagnosis [8,9]. Indeed, WSI technology
provides a number of advantages over traditional microscopy, including portability, ease
Copyright: © 2022 by the authors.
of sharing and retrieving images, and task balance [10]. The establishment of the digital
Licensee MDPI, Basel, Switzerland. pathology environment contributed to the development of a new branch of pathology
This article is an open access article known as computational pathology (CPATH) [11]. Novel terminology and definitions
distributed under the terms and have resulted from advances in computational pathology (Table 1) [12]. The computational
conditions of the Creative Commons analysis of pathology slide images has made direct disease investigation possible rather
Attribution (CC BY) license (https:// than relying on a pathologist analyzing images on a screen [13]. AI approaches aided by
creativecommons.org/licenses/by/ deep learning results are frequently used to combine information from digitized pathology
4.0/). images with their associated metadata. Using AI approaches that computationally evaluate

Diagnostics 2022, 12, 2794. https://doi.org/10.3390/diagnostics12112794 https://www.mdpi.com/journal/diagnostics

Diagnostics 2022, 12, 2794 2 of 20

the entire slide image, researchers can detect features that are difficult to detect by eye
alone, which is now the state-of-the-art in digital pathology [14].

Table 1. Computational pathology definitions.

Terms Definition

Artificial intelligence (AI) The broadest definition of computer science dealing with the ability of a computer to simulate
human intelligence and perform complicated tasks.
Computational pathology A branch of pathology that involves computational analysis of a broad array of methods to
(CPATH) analyze patient specimens for the study of disease. In this paper, we focus on the extraction of
information from digitized pathology images in combination with their associated metadata,
typically using AI methods such as deep learning.
Convolutional neural A form of deep neural networks with one or more convolutional layers and various different
networks (CNN) layers that can be trained using the backpropagation algorithm and which is suitable for
learning 2D data such as images.
Deep learning A subclassification of machine learning that imitates a logical structure similar to how people
conclude using a layered algorithm structure called an artificial neural network.
Digital pathology An environment in which traditional pathology analysis utilizing slides made of cells or tissues
is converted to a digital environment using a high-resolution scanner.
End-to-end training An opposite concept of feature-crafted methods in a machine learning model, a method which
learns the ideal value simultaneously rather than sequentially using only one pipeline. It works
smoothly when the dataset is large enough.
Ground truth A concept of a dataset’s ‘true’ category, quantity, or label that serves as direction to an
algorithm in the training step. The ground truth varies from the patient- or slide-level to
objects or areas within the picture, depending on the objective.
Image segmentation A technique for classifying each region into a semantic category by decomposing an image to
the pixel level.
Machine learning An artificial intelligence that parses data, learns from it, and makes intelligent judgments based
on what it has learned.
Metadata A type of data that explains other data. A single histopathology slide image in CPATH may
include patient disease, demographic information, previous treatment records and medical
history, slide dyeing information, and scanner information as metadata.
Whole-slide image (WSI) An whole histopathological glass slide digitized at microscopic resolution as a digital
representation. Slide scanners are commonly used to create these complete slide scans. A slide
scan viewing platform allows for image examination similar to that of a regular microscope.

The conventional pathological digital image machine learning method requires par-
ticularly educated pathologists to manually categorize abnormal picture attributes before
incorporating them into algorithms. Manually extracting and analyzing features from
pathological images was a time-consuming, labor-intensive, and costly method that led
to many disagreements among pathologists on whether features are typical [15]. Human-
extracted visual characteristics must be translated into numerical forms for computer
algorithms, but identifying patterns and expressing them with a finite number of feature
markers was nearly impossible in some complex diseases. Diverse and popular studies
to ‘well’ learn handmade features became the basis for a commercially available medical
image analysis system. After all the algorithm development steps, its performance often
had a high false-positive rate, and generalization in even typical pathological images was
limited [16]. Deep learning, however, has enabled computers to automatically extract
feature vectors from pathology image example data and learn to build optimal algorithms
on their own [17,18], even outperforming physicians in some cases, and has now emerged
as a cutting-edge machine learning method in medical clinical practice [19]. Diverse deep
architectures trained with huge image datasets provide biological informatics discoveries
and outstanding object recognition [20].
Diagnostics 2022, 12, 2794 3 of 20

The purpose of this review is to enhance the understanding of the reader with an up-
date on the implementation of artificial intelligence in the pathology department regarding
requirements, work process and clinical application development.

2. Deveopment of AI Aided Computational Pathology

Integrating artificial intelligence into the workflow of the pathology department can
perform quality control of the pre-analytic, analytic, and post-analytic phases of the pathol-
ogy department’s work process, allowing quality control of scan images and formalin-fixed
paraffin-embedded tissue blocks, integrated diagnosis with joining clinical information,
ordering necessary pathology studies including immunohistochemistry and molecular
studies, automating repetitive tasks, on-demand consultation, and cloud server manage-
ment (Figure 1), which, finally allow precision medicine by enabling us to use a wide range
of patient data, including pathological images, to develop disease-preventive and treatment
methods tailored to individual patient features. To achieve the above-mentioned goals,
there are crucial elements required for CPATH. A simple summary of the required steps for
the application of an AI with CPATH is demonstrated in Figure 2.

Figure 1. Embedding AI into pathology department workflow. The digital pathology supplies
whole-slide images to artificial intelligence, which performs quality control of pre-analytic phase,
analytic phase and post-analytic phase of pathology laboratory process.

Figure 2. Requirement for clinical applications of artificial intelligence with CPATH.

2.1. Equipment
Transitioning from glass to digital workflows in AP requires new digital pathology
equipment, image management systems, improved data management and storage capac-
ities, and additional trained technicians [21]. While the use of advanced high-resolution
Diagnostics 2022, 12, 2794 4 of 20

hardware with multiple graphical processing units can speed up training, it can become
prohibitively expensive. Pathologists must agree to changes to a century-old workflow.
Given that change takes time, pathologist end-users should anticipate change-management
challenges independent of technological and financial hurdles. AI deployment in the
pathology department requires digital pathology. Digital pathology has many proven uses,
including primary and secondary clinical diagnosis, telepathology, slide sharing, research
data set development, and pathology education or teaching [22]. Digital pathology systems
provide time- and cost-saving improvements over the traditional microscopy technique and
improve inter-observer variation with adequate slide image management software, inte-
grated reporting systems, improved scanning speeds, and high-quality images. Significant
barriers include the introduction of technologies without regulatory-driven, evidence-based
validation, the resistance of developers (academic and industrial), and the requirement for
commercial integration and open-source data formats.

2.2. Whole Slide Image

In the field of radiology, picture archiving and communication systems (PACS) were
successfully introduced owing to infrastructure such as stable servers and high-performance
processing devices, and they are now widely used in deep learning sources [23,24]. Sim-
ilarly, in the pathology field, a digital pathology system was developed that scans tradi-
tional glass slides using a slide scanner to produce a WSI; it then stores and transmits
it to servers [13]. Because WSI which has an average of 1.6 billion pixels and occupies
4600 megabytes (MB) per unit, thus taking up much more space than a DICOM (digital
imaging and communications in medicine) format, this technique took place later in pathol-
ogy than in radiography [25]. However, in recent years, scanners, servers, and technology
that can quickly process WSI have made this possible, allowing pathologists to inspect
images on a PC screen [6].

2.3. Quality Control Using Artificial Intelligence

AI tools can be embedded within a pathology laboratory workflow before or after
the diagnosis of the pathologist. Before cases are sent to pathologists for review, an AI
tool can be used to triage them (for example, cancer priority or improper tissue section)
or to help with screening for unexpected events (e.g., tissue contamination or microorgan-
isms). After reviewing a case, pathologists can also use AI tools to execute certain tasks
(e.g., counting mitotic figures for tumor grading or measuring nucleic acid quantification).
AI software can also run in the background and execute tasks such as quality control and
other tasks all the time (e.g., correlation with clinical or surgical information). The ability
of AI, digital pathology, and laboratory information systems to work together is the key
to making a successful AI workflow that fits the needs of a pathology department. Fur-
thermore, pre-analytic AI implementation can affect the process of molecular pathology.
Personalized medicine and accurate quantification of tumor and biomarker expression
have emerged as critical components of cancer diagnostics. Quality control (QC) of clinical
tissue samples is required to confirm the adequacy of tumor tissue to proceed with further
molecular analysis [26]. The digitization of stained tissue slides provides a valuable way to
archive, preserve, and retrieve important information when needed.

2.4. Diagnosis and Quantitation

A combination of deep learning methods in CPATH has been developed to excavate
unique and remarkable biomarkers for clinical applications. Tumor-infiltrating lympho-
cytes (TILs) are a prime illustration, as their spatial distributions have been demonstrated to
be useful for cancer diagnosis and prognosis in the field of oncology [27]. TILs are the prin-
cipal activator of anticancer immunity in theory, and if TILs could be objectively measured
across the tumor microenvironment (TME), they could be a reliable biomarker [20]. TILs
have been shown to be associated with recurrence and genetic mutations in non-small cell
lung cancer (NSCLC) [28], and lymphocytes, which have been actively made immune, have
Diagnostics 2022, 12, 2794 5 of 20

proved to have a better response, leading to a longer progression-free survival than the
ones that did not show much immunity [29]. Because manual quantification necessitates a
tremendous amount of work and is easily influenced by interobserver heterogeneity [30,31],
many approaches are being tested in order to overcome these hurdles and determine a
clinically meaningful TIL cutoff threshold [32]. Recently, a spatial molecular imaging tech-
nique obtaining spatial lymphocytic patterns linked to the rich genomic characterization of
TCGA samples has exemplified one application of the TCGA image archives, providing
insights into the tumor-immune microenvironment [20].
On a cellular level, spatial organization analysis of TME containing multiple cell types,
rather than only TILs, has been explored, and it is expected to yield information on tumor
progression, metastasis, and treatment outcomes [33]. Tissue segmentation is done using the
comprehensive immunolabeling of specific cell types or spatial transcriptomics to identify
a link between tissue content and clinical features, such as survival and recurrence [34,35].
In a similar approach, assessing image analysis on tissue components, particularly focusing
on the relative amount of area of tumor and intratumoral stroma, such as the tumor-
stroma ratio (TSR), is a widely studied prognostic factor in several cancers, including breast
cancer [36,37], colorectal cancer [38,39], and lung cancer [40]. Other studies in CPATH
include an attempt to predict the origin of a tumor in cancers of unknown primary source
using only a histopathology image of the metastatic site [41].
One of the advantages of CPATH is that it allows the simultaneous inspection of
histopathology images along with patient metadata, such as demographic, gene sequenc-
ing or expression data, and progression and treatment outcomes. Several attempts are
being made to integrate patient pathological tissue images and one or more metadata to
obtain novel information that may be used for diagnosis and prediction, as it was discov-
ered that predicting survival using merely pathologic tissue images was challenging and
inaccurate [42]. Mobadersany et al. used a Cox proportional hazards model integrated
with a CNN to predict the overall survival of patients with gliomas using tissue biopsy
images and genetic biomarkers such as chromosome deletion and gene mutation [43].
He et al. used H&E histopathology images and spatial transcriptomics, which analyzes
RNA to assess gene activity and allocate cell types to their locations in histology sections
to construct a deep learning algorithm to predict genomic expression in patients with breast
cancer [44]. Furthermore, Wang et al. employed a technique known as ‘transcriptome-wide
expression-morphology’ analysis, which allows for the prediction of mRNA expression and
proliferation markers using conventional histopathology WSIs from patients with breast
cancer [45]. It is also highly promising in that, as deep learning algorithms progress in
CPATH, it can be a helpful tool for pathologists and doctors making decisions. Studies have
been undertaken to see how significant an impact assisting diagnosis can have. Wang et al.
showed that pathologists employing a predictive deep learning model to diagnose the
metastasis of breast cancer from WSIs of sentinel lymph nodes reduced the human error
rate by nearly 85% [46]. In a similar approach, Steiner et al. looked at the influence of
AI in the histological evaluation of breast cancer with lymph node metastasis, comparing
pathologist performance supported by AI with pathologist performance unassisted by AI to
see whether supplementation may help. It was discovered that algorithm-assisted pathol-
ogists outperformed unassisted pathologists in terms of accuracy, sensitivity, and time
effectiveness [47].

3. Deep Learning from Computational Pathology

3.1. International Competitions
The exponential development in scanner performance making producing WSI easier
and faster than previously, along with sophisticated viewing devices, major advancements
in both computer technology and AI, as well as the accordance to regulatory requirements of
the complete infrastructure within the clinical context, have fueled CPATH’s rapid growth
in recent years [15]. Following the initial application of CNNs in histopathology at ICPR
2012 [48], several studies have been conducted to assess the performance of automated deep
Diagnostics 2022, 12, 2794 6 of 20

learning algorithms analyzing histopathology images in a variety of diseases, primarily

cancer. CPATH challenges are being promoted in the same way that competitions and
challenges are held in the field of computer engineering to develop technologies and
discover talented rookies. CAMELYON16 was the first grand challenge ever held, with the
goal of developing CPATH solutions for the detection of breast cancer metastases in H&E-
stained slides of sentinel lymph nodes and to assess the accuracy of the deep learning
algorithms developed by competition participants, medical students and experienced
professional pathologists [49]. The dataset from the CAMELYON16 challenge, which took
a great deal of work, was used in several other studies and provided motive for other
challenges [50–52], attracting major machine learning companies such as Google to the
medical artificial intelligence field [53], and is said to have influenced US government
policy [54]. Since then, new challenges have been proposed in many more cancer areas
using other deep learning architectures with greater datasets, providing the driving force
behind the growth of CPATH (Table 2). Histopathology deep learning challenges can attract
non-medical engineers and medical personnel, provide prospects for businesses, and make
the competition’s dataset publicly available, benefiting future studies. Stronger deep
learning algorithms are expected to emerge, speeding the clinical use of new algorithms
in digital image analysis. Traditional digital image analysis works on three major types
of measures: image object localization, classification, and quantification [12], and deep
learning in CPATH focuses on those metrics similarly. CPATH applications include tumor
detection and classification, invasive or metastatic foci detection, primarily lymph nodes,
image segmentation and analysis of spatial information, including ratio and density, cell
and nuclei classification, mitosis counting, gene mutation prediction, and histological
scoring. Two or more of these categories are often researched together, and deep learning
architectures like convolutional neural networks (CNN) and recurrent neural networks are
utilized for training and applications.

Table 2. Examples of grand challenges held in CPATH.

Challenge Year Staining Challenge Goal Dataset

Private set—
GlaS Segmentation of colon glands of stage T3 and T4 colorectal
2015 H&E 165 images from
challenge [55] adenocarcinoma
16 WSIs
Evaluation of new and current algorithms for automatic
Private set—
CAMELYON16 [56] 2016 H&E identification of metastases in WSIs from H&E-stained lymph
221 images
node sections
TUPAC Prediction of tumor proliferation scores and gene expression of
2016 H&E 821 TCGA WSIs
challenge [57] breast cancer using histopathology WSIs
Development of quantitative biomarkers to determinate cancer
BreastPathQ [58] 2018 H&E Private set—96 WSIs
cellularity of breast cancer from H&E-stained WSIs
BACH Classification of H&E-stained breast histopathology images Private set—40 WSIs
2018 H&E
challenge [59] and performing pixel-wise labeling of WSIs and 500 images
Provision of a dataset as well as an evolution platform for current LYON19 test set
LYON19 [60] 2019 IHC
lymphocyte detection algorithms in IHC-stained images containing 441 ROIs
Evaluation of algorithms for detecting signet ring cells and
DigestPath [61] 2019 H&E screening colonoscopy tissue from histopathology images of the Private set—127 WSIs
digestive system
Evaluation of algorithms to discriminate HER2-positive breast
HEROHE cancer specimens from HER2-negative breast cancer specimens
2020 H&E Private set—359 WSIs
ECDP [62] with high sensitivity and specificity only using H&E-stained
slides
Detection of mitotic figures from breast cancer histopathology
MIDOG
2021 H&E images scanned by different scanners to overcome the Private set—200 cases
challenge [63]
‘domain-shift’ problem and improve generalization
Diagnostics 2022, 12, 2794 7 of 20

Table 2. Cont.

Challenge Year Staining Challenge Goal Dataset

Evaluation of algorithms for nuclear segmentation and
CoNIC
2022 H&E classification into six types, along with cellular composition 4981 patches
challenge [64]
prediction
Private dataset—
Development of WSI registration algorithms that can align WSIs
H&E, 750 cases consisting of
ACROBAT [65] 2022 of IHC-stained breast cancer tissue sections with corresponding
IHC 1 H&E and 1–4
H&E-stained tissue regions
matched IHC

3.2. Dataset and Deep Learning Model

Since public datasets for machine learning learning in CPATH, such as the Cancer
Genome Atlas (TCGA), the Cancer Image Archive (TCIA), and public datasets created by
several challenges, such as the CAMELYON16 challenge dataset, are freely accessible to
anyone, researchers who do not have their own private data can conduct research and
can also use the same dataset as a standard benchmark by several researchers comparing
the performance of each algorithm [15]. Coudray et al. [66], Using the inception-v3 model
as a deep learning architecture, assessed the performance of algorithms in classification
and genomics mutation prediction of NSCLC histopathology pictures from TCGA and
a portion of an independent private dataset, which was a noteworthy study that could
detect genetic mutations using WSIs such as STK11 (AUC 0.85), KRAS (AUC 0.81), and
EGFR (AUC 0.75). Guo et al. used the Inception-v3 model to classify the tumor region of a
breast cancer [67]. Bulten et al. used 1243 WSIs of private prostate biopsies, segmenting
individual glands to determine Gleason growth patterns using UNet, followed by cancer
grading, and achieved performance comparable to pathologists [68]. Table 3 contains
additional published examples utilizing various deep learning architectures and diverse
datasets. A complete and extensive understanding of deep learning concepts and existing
architectures can be found [17,69], while a specific application of deep learning in medical
image analysis can be read [70–72]. To avoid bias in algorithm development, datasets should
be truly representative, encompassing the range of data that would be expected in the real
world [19], including both the expected range of tissue features (normal and pathological)
and the expected variation in tissue and slide preparation between laboratories.

Table 3. Summary of recent convolutional neural network models in pathology image analysis.

Publication Deep Learning Input Training Goal Dataset

Zhang et al. [73] CNN WSI Diagnosis of bladder cancer TCGA and private—913 WSIs
Shim et al. [74] CNN WSI Prognosis of lung cancer Private—393 WSIs
Im et al. [75] CNN WSI Diagnosis of brain tumor subtype private—468 WSIs
Mi et al. [76] CNN WSI Diagnosis of breast cancer private dataset—540 WSIs
Hu et al. [77] CNN WSI Diagnosis of gastric cancer private—921 WSIs
Pei et al. [78] CNN WSI Diagnosis of brain tumor classification TCGA—549 WSIs
Salvi et al. [79] CNN WSI Segmentation of normal prostate gland Private—150 WSIs
Lu et al. [80] CNN WSI Genomic correlation of breast cancer TCGA and private—1157 WSIs
Cheng et al. [81] CNN WSI Screening of cervical cancer Private—3545 WSIs
Kers et al. [82] CNN WSI Classification of transplant kidney Private—5844 WSIs
Zhou et al. [83] CNN WSI Classification of colon cancer TCGA—1346 WSIs
Hohn et al. [84] CNN WSI Classification of skin cancer Private—431 WSIs
Wang et al. [45] CNN WSI Prognosis of gastric cancer Private—700 WSIs
Shin et al. [85] CNN, GAN WSI Diagnosis of ovarian cancer TCGA—142 WSIs
Abbreviation: CNN, convolutional neural network; WSI, whole-slide image; TCGA, The Cancer Genome Atlas.

CNNs are difficult to train end-to-end because gigapixel WSIs are too large to fit in
GPU memory, unlike many natural pictures evaluated in computer vision applications.
A single WSI requires over terabytes of memory, yet high-end GPUs only give tens of
Diagnostics 2022, 12, 2794 8 of 20

gigabytes. Researchers have suggested alternatives such as partitioning the WSI into small
sections (Figure 3) using only a subset or the full WSI compressed with semantic infor-
mation preserved. Breaking WSI into little patches and placing them all in the GPU to
learn everything takes too long; thus, picking patches to represent WSI is critical. For these
reasons, randodmizing paches [86], selecting patches from region of interests [42], and ran-
domly selecting patches among image clustering [87] were proposed. The multi-instance
learning (MIL) method is then mostly employed in the patch aggregation step, which
involves collecting several patches from a single WSI and learning information about the
WSI as a result. Traditional MILs treat a single WSI as a basket, assuming that all patches
contained within it have the same WSI properties. All patches from a cancer WSI, for exam-
ple, are considered cancer patches. This method appears to be very simple, yet it is quite
beneficial for cancer detection, and representation can be ensured if the learning dataset
is large enough [88], which also provides a reason why various large datasets should be
produced. If the learning size is insufficient, predicted patch scores are averaged, or classes
that account for the majority of patch class predictions are estimated and used to represent
the WSI. A more typical way is to learn patch weights using a self-attention mechanism,
which uses patch encoding to calculate weighed sum of patch embeddings [89], with a
higher weight for the patch that is closest to the ideal patch for performing a certain task
for each model. Techniques such as max or mean pooling and certainty pooling, which
are commonly utilized in CNNs, are sometimes applied here. There is an advantage to
giving interpretability to pathologists using the algorithm because approaches such as
self-attention can be presented in the form of a heatmap on a WSI based on patch weights.

Figure 3. Images are divided into small patches obtained from tissue of WSI, which are subsequently
prepared to have semantic features extracted from each patch. Green tiles indicate tumor region; red
tiles indicate non-tumor region. Images from Yeouido St. Mary’s hospital.

3.3. Overview of Deep Learning Workflows

WSIs are flooding out of clinical pathology facilities around the world as a result of the
development of CPATH, including publicly available datasets, which can be considered a
desirable cornerstone for the development of deep learning because it means more data are
available for research. However, as shown in some of the previous studies, the accuracy of
performance, such as classification and segmentation by algorithms commonly expressed
in the area under the curve (AUC), must be compared to pathological images manually
annotated by humans in order to calculate the accuracy of the performance. In this way, su-
pervised learning is a machine learning model that uses labeled learning data for algorithm
learning and learns functions based on it, and it is the machine learning model most utilized
in CPATH so far. According to the amount and type of data, object and purpose (whether
the target is cancer tissue or substrate tissue and calculating the number of lymphocytes), it
can be divided into qualitative and distinct or quantitative and continuous representations,
expressed as ‘classification’ [90] and ‘regression’ [91], respectively. Because the model is
constructed by simulating learning data, labeled data are crucial, and the machine learning
model’s performance may vary. Unsupervised learning uses unlabeled images, unlike
previous scenarios. This technology is closer to an AI since it helps humans collect infor-
mation and build knowledge about the world. Except for the most basic learning, such
as language character acquisition, we can identify commonalities by looking at applied
Diagnostics 2022, 12, 2794 9 of 20

situations and extending them to other objects. To teach young children to recognize dogs
and cats, it is not required to exhibit all breeds. ‘Unsupervised learning’ can find and assess
patterns in unlabeled data, divide them into groups, or perform data visualization in which
specific qualities are compacted to two or three if there are multiple data characteristics or
variables that are hard to see. A study built a complex tissue classifier for CNS tumours
based on histopathologic patterns without manual annotation. It provided a framework
comparable to the WHO [92], which was based on microscopic traits, molecular character-
istics, and well-understood biology [93]. This study demonstrated that the computer can
optimize and use some of the same histopathologic features used by pathologists to assist
grouping on its own.
In CPATH, it is very important to figure out how accurate a newly made algorithm
is, so there is still a lot of supervised learning. Unsupervised learning still makes it hard
to keep up with user-defined tasks, but it has the benefit of being a very flexible way to
build data patterns that are not predictable. It also lets us deal with changes we did not
expect and allows us to learn more outside of the limits of traditional learning. It helps us
understand histopathology images and acts as a guide for precision medicine [94].
Nonetheless, unsupervised learning is still underdeveloped in CPATH, and even after
unsupervised learning, it is sometimes compared with labeled data to verify performance,
making the purpose a little ambiguous. Bulten et al. classified prostate cancer and non-
cancer pathology using clustering, but still had to verify the algorithm’s ability using
manually annotated images, for example [95].
Currently, efforts are made to make different learning datasets by combining the
best parts of supervised and unsupervised learning. This is done by manually labeling
large groups of pathological images. Instead of manually labeling images, such as in the
2016 TUPAC Challenge, which was an attempt to build standard references for mitosis
detection [96], “weakly supervised learning” means figuring out only a small part of an
image and then using machine learning to fill in the rest. Several studies have shown that
combining sophisticated learning strategies with weakly supervised learning methods can
produce results that are similar to those of a fully supervised model. Since then, many
more studies have been done on the role of detection and segmentation in histopathology
images. “NuClick”, a CNN-based algorithm that won the LYON19 Challenge in 2019,
showed that structures such as nuclei, cells, and glands in pathological images can be
labeled quickly, consistently, and reliably [97], whereas ‘CAMEL’, developed in another
study, only uses sparse image-level labels to produce pixel-level labels for creating datasets
to train segmentation models for fully supervised learning [98].

4. Current Limitations and Challenges

Despite considerable technical advancements in CPATH in recent years, the deploy-
ment of deep learning algorithms in real clinical settings is still far from adequate. This
is because, in order to be implemented into existing or future workflows, the CPATH
algorithm must be scientifically validated, have considerable clinical benefit, and not cause
harm or confuse people at the same time [99]. In this section, we will review the roadblocks
to full clinical adoption of the CPATH algorithm, as well as what efforts are currently
being made.

4.1. Acquiring Quality Data

It is critical that CPATH algorithms be trained with high-quality data so that they
can deal with the diverse datasets encountered in real-world clinical practice. Even in
deep learning, the ground truth should be manually incorporated into the dataset in
order to train appropriate diagnostic contexts in supervised learning to classify, segment,
and predict images based on it [100]. The ground truth can be derived from pathology
reports grading patient outcomes or tumors, as well as scores assessed by molecular
experiments, depending on the study’s goals, which are still determined by human experts
and need a significant amount of manual labor to obtain a ‘correct’ dataset [12]. Despite the
Diagnostics 2022, 12, 2794 10 of 20

fact that datasets created by professional pathologists are of excellent quality, vast quantities
are difficult to obtain due to the time, cost, and repetitive and arduous tasks required. As a
result, publicly available datasets have been continuously created, such as the ones from
TCGA or grand challenges, with the help of weakly supervised learning. Alternative efforts
have recently been made to gather massive scales of annotated images by crowdsourcing
online. Hughes et al. used a crowdsourced image presentation platform to demonstrate
deep learning performance comparable to that of a single professional pathologist [101],
while López-Pérez et al. used a crowdsourced deep learning algorithm to help a group of
doctors or medical students who were not pathologists make annotations comparable to an
expert in breast cancer images [102]. Crowdsourcing may generate some noise, but it shows
that non-professionals of various skill levels could assist with pathological annotation and
dataset generation. Obtaining quality data entails more than just obtaining a sufficient
raw pathological image slide of a single disease from a patient or hospital; it also includes
preparing materials to analyze and process the image in order to extract useful data for
deep learning model training. By using strategies such as selecting patches with cells while
excluding patches without cells from raw pictures, as demonstrated in Figure 4, collecting
quality data may be made easier.

Figure 4. (a) Random sampling of 100 patches selected arbitrarily from an WSI image. (b) Random
sampling of 100 patches after application of Laplace filter (which highlights areas with great changes)
from WSI image. Images from Yeouido St. Mary’s Hospital

4.2. Data Variation

Platform diversity, integration, and interoperability represent yet another significant
hurdle for the creation and use of AI tools [103]. Recent findings show that current AI
models, when trained on insufficient datasets, even when utilizing precise and pixel-by-
pixel labelling, can exhibit a 20% decline in performance when evaluated on independent
datasets [88]. Deep learning-based algorithms have produced outstanding outcomes in
image analysis applications, including digitized slide analysis. Deep learning-based sys-
tems face several technological problems, including huge WSI data, picture heterogeneity,
and feature complexity. To achieve successful generalization properties, the training data
Diagnostics 2022, 12, 2794 11 of 20

must include a diverse and representative sample of the disease’s biological and mor-
phological variability, as well as the technical variables introduced in the pre-analytical
and analytical processes in the pathology department, as well as the image acquisition
process [104]. A generic deep learning-based system for histopathology tissue analysis.
The previously introduced framework is a series of strategies in the preprocessing-training-
inference pipeline that showed improved efficiency and generalizability. Such strategies
include an ensemble segmentation model, dividing the WSI into smaller overlapping
patches, efficient inference algorithms, and a patch-based uncertainty estimation methodol-
ogy [105,106]. Technical variability challenges can also be addressed by standardizing and
preparing CPATH data to limit the effects of technical variability or to make the models
robust to technical variability. Training the deep learning model on large and diverse
datasets may lower the generalization error to some extent [107].
The amount and quality of input data determine the performance of the deep learning
algorithm [108,109]. Although the size of datasets has been growing over the years with the
development in CPATH, even if algorithms trained using learning datasets perform well on
test sets, it is difficult to be certain that algorithms perform well on actual clinical encounters
because clinical data come from significantly more diverse sources than studies. Similarly,
when evaluating the performance of deep learning algorithms with a specific validation
set for each grand challenge, it is also difficult to predict whether they will perform well
in actual clinical practice. Color variation is a representative example of the variation of
data. Color variation is caused by differences in raw materials, staining techniques used
across different pathology labs, patient intervariability, and different slide scanners, which
affect not just color but also overall data variation [110]. As a result, color standardization
as an image preparation method has long been devised to overcome this problem in WSI.
Because predefined template images were used for color normalization in the past, it was
difficult to style transformation between different image datasets, but recent advances in
generative adversarial networks (GAN) among deep learning artificial neural networks
have allowed patches to be standardized without organizational changes. For example,
using the cycle-GAN technique, Swiderska-Chadaj et al. reported an AUC of 0.98 and
0.97 for two different datasets constructed from prostate cancer WSIs [72,111]. While
efforts are being made to reduce variation and create well-defined standardized data,
such as color standardization and attempts to establish global standards for pathological
tissue processing, staining, scanning, and digital image processing, data augmentation
techniques are also being used to create learning datasets with as many variations as
possible in order to learn the many variations encountered in real life. Not only the
performance of the CPATH algorithm but also many considerations such as cost and
explainability should be thoroughly addressed when deciding which is more effective for
actual clinical introduction.

4.3. Algorithm Validation

Several steps of validation are conducted during the lengthy process of developing
a CPATH algorithm in order to test its performance and safety. To train models and
evaluate performance, CPATH studies on typical supervised algorithms separate annotated
data into individual learning datasets and test datasets, the majority of which employ
datasets with features fairly similar to those of learning datasets in the so-called ‘internal
verification’ stage. Afterwards, through so-called ‘external validation’, which uses data for
tests that have not been used for training, it is feasible to roughly evaluate if the algorithm
performs well with the data it would encounter in real clinical practice [15]. However,
simply because the CPATH algorithm performed well at this phase, it is hard to ascertain
whether it will function equally well in practical practice [112]. While many studies on
the CPATH algorithm are being conducted, most studies use autonomous standards due
to a lack of established clinical verification standards and institutional validation. Even
if deep learning algorithms perform well and are employed with provisional permission,
it is difficult to confirm that their performance exhibits the same confirmed effect when
Diagnostics 2022, 12, 2794 12 of 20

the algorithm is upgraded in the subsequent operation process. Efforts are being made
to comprehend and compare diverse algorithms regardless of research techniques, such
as the construction of a complete and transparent information reporting system called
TRIPOD-AI in the prediction model [113].
Finally, it should be noted that the developed algorithm does not result in a single
performance but rather continues within the patient’s disease progress and play an auxiliary
role in decision-making; thus, relying solely on performance as a ratification metric is not
ideal. This suggests that, in cases where quality measure for CPATH algorithm performance
is generally deemed superior to or comparable to pathologists, it should be defined by
examining the role of algorithms in the whole scope of disease progression in a patient
in practice [114]. This is also linked to the solution of the gold-standard paradox [14].
This is a paradox which may ariase during the segmentation model’s quality control,
where pathologists are thought to be the most competent in pathological picture analysis,
but algorithmic data are likely to be superior in accuracy and reproducibility. This paradox
may alternatively be overcome by implementing the algorithm as part of a larger system
that tracks the patient’s progress and outcomes [12].

4.4. Regulatory Considerations

One of the most crucial aspects for deep learning algorithms to be approved by regula-
tory agencies in order to use AI in clinical practice is to understand how it works, as AI is
sometimes referred to be a “black box” because it is difficult for humans to comprehend
exactly what it does [114]. Given the difficulty of opening up deep learning artificial neu-
ral networks and their limited explainability due to the difficulty of understanding how
countless parameters interact at the same time, more reliable and explainable models for
complex and responsible behaviors for diagnosis and treatment decisions and prediction
are required [115]. As a result, attempts have been made to turn deep learning algorithms
into “glass boxes” by clarifying the input and calculating the output in a way that humans
can understand and analyze [116–118].
The existing regulatory paradigm is less adequate for AI since it requires rather small
infrastructure and little human interaction, and the level of progress or results are opaque
to outsiders, so potential dangers are usually difficult to identify [119]. Thus far, the White
House has issued a memorandum on high-level regulatory principles for AI in all fields in
November 2020 [120], the European Commission issued a similar white paper in February
2020 [121], and UNESCO made a global guideline on AI ethics in November 2021 [122],
but these documents unfortunately do not provide a very detailed method to operate
artificial intelligence in the context of operations. Because artificial intelligence is generally
developed in confined computer systems, progress has been made outside of regulatory
environments thus far, and regulatory uncertainty can accelerate development while also
fueling systemic dangers at the same time. Successful AI regulations, as with many new
technologies, are expected to be continuously problematic in the future, as regulations
and legal rules will still lag behind developing technological breakthroughs [123]. Self-
regulation in industrial settings can be theoretically beneficial and is already in use [124],
but it has limitations in practice because it is not enforced. Ultimately, a significant degree
of regulatory innovation is required to develop a stable AI environment. The most crucial
issue to consider in this regard is that, in domains such as health care, where even a slight
change can have a serious influence, regulations of AI should be built with the consideration
of the overall impact on humans rather than making arbitrary decisions alone.

5. Novel Trends in CPATH

5.1. Explainable AI
Because most AI algorithms have unclear properties due to their complexity and often
lacking robustness, there are substantial issues with AI trust [125]. Furthermore, there
is no agreement on how pathologists should include computational pathology systems
into their workflow [126]. Building computational pathology systems with explainable
Diagnostics 2022, 12, 2794 13 of 20

artificial intelligence (xAI) methods is a strong substitute for opaque AI models to address
these issues [127]. Four categories of needs exist for the usage of xAI techniques and
their application possibilities [128]: (1) Model justification: to explain why a decision was
made, particularly when a significant or unexpected decision is created, all with the goal of
developing trust in the model’s operation; (2) Model controlling and debugging: to avoid
dangerous outcomes. A better understanding of the system raises the visibility of unknown
defects and aids in the rapid identification and correction of problems; (3) Model improving:
When a user understands why and how a system achieved a specific result, he can readily
modify and improve it, making it wiser and possibly faster. Understanding the judgments
created by the AI model, in addition to strengthening the explanation-generating model,
can improve the overall work process; (4) Knowledge discovery: One can discover new
rules by seeing the appearance of some invisible model results and understanding why and
how they appeared. Furthermore, because AI entities are frequently smarter than humans,
it is possible to learn new abilities by understanding their behavior.
Recent studies in breast pathology xAI quickly presented the important diagnostic
areas in an interactive and understandable manner by automatically previewing tissue
WSIs and identifying the regions of interest, which can serve pathologists as an interactive
computational guide for computer-assisted primary diagnosis [127,129]. An ongoing
study is being done to determine which explanations are best for artificial intelligence
development, application, and quality control [130], which explanations are appropriate
for situations with high stakes [115], and which explanations are true to the explained
model [131].
With the increasing popularity of graph neural networks (GNNs), their application in a
variety of disciplines requires explanations for scientific or ethical reasons in medicine [132].
This makes it difficult to define generalized explanation methods, which are further compli-
cated by heterogeneous data domains and graphs. Most explanations are therefore model-
and domain-specific. GNN models can be used for node labeling, link prediction, and graph
classification [133]. While most models can be used for any of the above tasks, defining
and generating explanations can affect how a GNN xAI model is structured. However,
the power of these GNN models is limited by their complexity and the underlying data
complexity, although most, if not all, of the models can be grouped under the augmented
paradigm [134]. Popular deep learning algorithms and explainability techniques based on
pixel-wise processing ignore biological elements, limiting pathologists’ comprehension.
Using biological entity-based graph processing and graph explainers, pathologists can now
access explanations.

5.2. Ethics and Security

AI tool creation must take into account the requirement for research and ethics ap-
proval, which is typically necessary during the research and clinical trial stages. Developers
must follow the ethics of using patient data for research and commercial advantages.
Recognizing the usefulness of patient data for research and the difficulties in obtaining
agreement for its use, the corresponding institution should establish a proper scheme to
provide individual patients some influence over how their data are used [103]. Individual
institutional review boards may have additional local protocols for permitting one to opt
out of data use for research, and it is critical that all of these elements are understood and
followed throughout the design stage of AI tool creation [104]. There are many parallels to
be found with the AI development pipeline; while successful items will most likely transit
through the full pathway, supported by various resources, many products will, however,
fail at some point. Each stage of the pipeline, including the justification of the tool for
review and being recommended for usage in clinical guidelines, can benefit from mea-
surable outcomes of success in order to make informed judgments about which products
should be promoted [135]. This usually calls for proof of cost or resource savings, quality
improvements, and patient impact and is thus frequently challenging to demonstrate,
especially when the solution entails major transformation and process redesign.
Diagnostics 2022, 12, 2794 14 of 20

Whether one uses a cloud-based AI solution for pathology diagnostics depends on a

number of things, such as the preferred workflow, frequency of instrument use, software
and hardware costs, and whether or not the IT security risk group is willing to allow the
use of cloud-based solutions. Cloud-based systems must include a business associate’s
agreement, end-to-end encryption, and unambiguous data-use agreements to prevent data
breaches and inappropriate use of patient data [21].

6. Conclusions and Future Directions

AI currently has enormous potential to improve pathology practice by reducing
errors, improving reproducibility, and facilitating expert communication, all of which
were previously difficult with microscopic glass slides. Recent trends of AI applicaion
should be affordable, practical, interoperable, explainable, generalizable, manageable,
and reimbursable [21]. Many researchers are convinced that AI in general and deep learning
in particular could help with many repetitive tasks using digital pathology because of
recent successes in image recognition. However, there are currently only a few AI-driven
software tools in this field. As a result, we believe pathologists should be involved from the
start, even when developing algorithms, to ensure that these eagerly anticipated software
packages are improved or even replaced by AI algorithms. Despite popular belief, AI will
be difficult to implement in pathology. AI tools are likely to be approved by regulators
such as the Food and Drug Administration.
The quantitative nature of CPATH has the potential to transform pathology laboratory
and clinical practices. Case stratification, expedited review and annotation, and the output
of meaningful models to guide treatment decisions and predict patterns in medical fields
are all possibilities. The pathology community needs more research to develop safe and
reliable AI. As clinical AI’s requirements become clearer, this gap will close. AI in pathology
is young and will continue to mature as researchers, doctors, industry, regulatory agencies,
and patient advocacy groups innovate and bring new technology to health care practi-
tioners. To accomplish its successful application, robust and standardized computational,
clinical, and laboratory practices must be established concurrently and validated across
multiple partnering sites.

Author Contributions: Conceptualization, T.-J.K.; data curation, K.K. and Y.S.; writing—original
draft preparation, I.K.; writing—review and editing, I.K. and T.-J.K.; supervision, T.-J.K.; funding
acquisition, T.-J.K. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the National Research Foundation of Korea (NRF) through
a grant funded by the Korean government (MSIT) (grant number 2017R1E1A1A01078335 and
2022R1A2C1092956) and the Institute of Clinical Medicine Research in the Yeouido St. Mary’s Hospital.
Institutional Review Board Statement: The study was conducted according to the guidelines of
the Declaration of Helsinki and approved by the Institutional Review Board of Yeouido St. Mary’s
hospital (SC18RNSI0005 approved on 22 January 2018).
Informed Consent Statement: Not applicable.
Data Availability Statement: Data sharing is not applicable to this article.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Gurcan, M.N.; Boucheron, L.E.; Can, A.; Madabhushi, A.; Rajpoot, N.M.; Yener, B. Histopathological image analysis: A review.
IEEE Rev. Biomed. Eng. 2009, 2, 147–171. [CrossRef] [PubMed]
2. Wolberg, W.H.; Street, W.N.; Heisey, D.M.; Mangasarian, O.L. Computer-derived nuclear features distinguish malignant from
benign breast cytology. Hum. Pathol. 1995, 26, 792–796. [CrossRef]
3. Choi, H.K.; Jarkrans, T.; Bengtsson, E.; Vasko, J.; Wester, K.; Malmström, P.U.; Busch, C. Image analysis based grading of bladder
carcinoma. Comparison of object, texture and graph based methods and their reproducibility. Anal. Cell. Pathol. 1997, 15, 1–18.
[CrossRef] [PubMed]
Diagnostics 2022, 12, 2794 15 of 20

4. Keenan, S.J.; Diamond, J.; Glenn McCluggage, W.; Bharucha, H.; Thompson, D.; Bartels, P.H.; Hamilton, P.W. An automated
machine vision system for the histological grading of cervical intraepithelial neoplasia (CIN). J. Pathol. 2000, 192, 351–362.
[CrossRef]
5. Weinstein, R.S.; Graham, A.R.; Richter, L.C.; Barker, G.P.; Krupinski, E.A.; Lopez, A.M.; Erps, K.A.; Bhattacharyya, A.K.; Yagi, Y.;
Gilbertson, J.R. Overview of telepathology, virtual microscopy, and whole slide imaging: Prospects for the future. Hum. Pathol.
2009, 40, 1057–1069. [CrossRef] [PubMed]
6. Farahani, N.; Parwani, A.V.; Pantanowitz, L. Whole slide imaging in pathology: Advantages, limitations, and emerging
perspectives. Pathol. Lab. Med. Int. 2015, 7, 4321.
7. Saco, A.; Ramírez, J.; Rakislova, N.; Mira, A.; Ordi, J. Validation of whole-slide imaging for histolopathogical diagnosis: Current
state. Pathobiology 2016, 83, 89–98. [CrossRef]
8. Al-Janabi, S.; Huisman, A.; Vink, A.; Leguit, R.J.; Offerhaus, G.J.A.; Ten Kate, F.J.; Van Diest, P.J. Whole slide images for primary
diagnostics of gastrointestinal tract pathology: A feasibility study. Hum. Pathol. 2012, 43, 702–707. [CrossRef]
9. Snead, D.R.; Tsang, Y.W.; Meskiri, A.; Kimani, P.K.; Crossman, R.; Rajpoot, N.M.; Blessing, E.; Chen, K.; Gopalakrishnan, K.;
Matthews, P.; et al. Validation of digital pathology imaging for primary histopathological diagnosis. Histopathology 2016, 68,
1063–1072. [CrossRef]
10. Williams, B.J.; Bottoms, D.; Treanor, D. Future-proofing pathology: The case for clinical adoption of digital pathology. J. Clin.
Pathol. 2017, 70, 1010–1018. [CrossRef]
11. Astrachan, O.; Hambrusch, S.; Peckham, J.; Settle, A. The present and future of computational thinking. ACM SIGCSE Bulletin.
2009, 41, 549-550. [CrossRef]
12. Abels, E.; Pantanowitz, L.; Aeffner, F.; Zarella, M.D.; van der Laak, J.; Bui, M.M.; Vemuri, V.N.; Parwani, A.V.; Gibbs, J.; Agosto-
Arroyo, E.; et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: A white
paper from the Digital Pathology Association. J. Pathol. 2019, 249, 286–294. [CrossRef]
13. Zarella, M.D.; Bowman, D.; Aeffner, F.; Farahani, N.; Xthona, A.; Absar, S.F.; Parwani, A.; Bui, M.; Hartman, D.J. A practical
guide to whole slide imaging: A white paper from the digital pathology association. Arch. Pathol. Lab. Med. 2019, 143, 222–234.
[CrossRef] [PubMed]
14. Aeffner, F.; Wilson, K.; Martin, N.T.; Black, J.C.; Hendriks, C.L.L.; Bolon, B.; Rudmann, D.G.; Gianani, R.; Koegler, S.R.; Krueger, J.;
et al. The gold standard paradox in digital image analysis: Manual versus automated scoring as ground truth. Arch. Pathol. Lab.
Med. 2017, 141, 1267–1275. [CrossRef] [PubMed]
15. Van der Laak, J.; Litjens, G.; Ciompi, F. Deep learning in histopathology: The path to the clinic. Nat. Med. 2021, 27, 775–784.
[CrossRef]
16. Chan, H.P.; Samala, R.K.; Hadjiiski, L.M.; Zhou, C. Deep learning in medical image analysis. Annu. Rev. Biomed. Eng. 2020, 19,
221–248.
17. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef]
18. Schmidhuber, J. Deep learning in neural networks: An overview. Neural Netw. 2015, 61, 85–117. [CrossRef]
19. Verma, P.; Tiwari, R.; Hong, W.C.; Upadhyay, S.; Yeh, Y.H. FETCH: A Deep Learning-Based Fog Computing and IoT Integrated
Environment for Healthcare Monitoring and Diagnosis. IEEE Access 2022, 10, 12548–12563. [CrossRef]
20. Saltz, J.; Gupta, R.; Hou, L.; Kurc, T.; Singh, P.; Nguyen, V.; Samaras, D.; Shroyer, K.R.; Zhao, T.; Batiste, R.; et al. Spatial
organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Rep.
2018, 23, 181–193. [CrossRef]
21. Cheng, J.Y.; Abel, J.T.; Balis, U.G.; McClintock, D.S.; Pantanowitz, L. Challenges in the development, deployment, and regulation
of artificial intelligence in anatomic pathology. Am. J. Pathol. 2021, 191, 1684–1692. [CrossRef] [PubMed]
22. Volynskaya, Z.; Evans, A.J.; Asa, S.L. Clinical applications of whole-slide imaging in anatomic pathology. Adv. Anat. Pathol. 2017,
24, 215–221. [CrossRef] [PubMed]
23. Mansoori, B.; Erhard, K.K.; Sunshine, J.L. Picture Archiving and Communication System (PACS) implementation, integration &
benefits in an integrated health system. Acad. Radiol. 2012, 19, 229–235. [PubMed]
24. Lee, J.G.; Jun, S.; Cho, Y.W.; Lee, H.; Kim, G.B.; Seo, J.B.; Kim, N. Deep learning in medical imaging: General overview. Korean J.
Radiol. 2017, 18, 570–584. [CrossRef] [PubMed]
25. Kohli, M.D.; Summers, R.M.; Geis, J.R. Medical image data and datasets in the era of machine learning—Whitepaper from the
2016 C-MIMI meeting dataset session. J. Digit. Imaging 2017, 30, 392–399. [CrossRef] [PubMed]
26. Chung, M.; Lin, W.; Dong, L.; Li, X. Tissue requirements and DNA quality control for clinical targeted next-generation sequencing
of formalin-fixed, paraffin-embedded samples: A mini-review of practical issues. J. Mol. Genet. Med. 2017, 11, 1747–0862.
27. Salgado, R.; Denkert, C.; Demaria, S.; Sirtaine, N.; Klauschen, F.; Pruneri, G.; Wienert, S.; Van den Eynden, G.; Baehner, F.L.;
Pénault-Llorca, F.; et al. The evaluation of tumor-infiltrating lymphocytes (TILs) in breast cancer: Recommendations by an
International TILs Working Group 2014. Ann. Oncol. 2015, 26, 259–271. [CrossRef]
28. AbdulJabbar, K.; Raza, S.E.A.; Rosenthal, R.; Jamal-Hanjani, M.; Veeriah, S.; Akarca, A.; Lund, T.; Moore, D.A.; Salgado, R.;
Al Bakir, M.; et al. Geospatial immune variability illuminates differential evolution of lung adenocarcinoma. Nat. Med. 2020,
26, 1054–1062. [CrossRef]
Diagnostics 2022, 12, 2794 16 of 20

29. Park, S.; Ock, C.Y.; Kim, H.; Pereira, S.; Park, S.; Ma, M.; Choi, S.; Kim, S.; Shin, S.; Aum, B.J.; et al. Artificial Intelligence-
Powered Spatial Analysis of Tumor-Infiltrating Lymphocytes as Complementary Biomarker for Immune Checkpoint Inhibition in
Non-Small-Cell Lung Cancer. J. Clin. Oncol. 2022, 40, 1916–1928. [CrossRef]
30. Khoury, T.; Peng, X.; Yan, L.; Wang, D.; Nagrale, V. Tumor-infiltrating lymphocytes in breast cancer: Evaluating interobserver
variability, heterogeneity, and fidelity of scoring core biopsies. Am. J. Clin. Pathol. 2018, 150, 441–450. [CrossRef]
31. Swisher, S.K.; Wu, Y.; Castaneda, C.A.; Lyons, G.R.; Yang, F.; Tapia, C.; Wang, X.; Casavilca, S.A.; Bassett, R.; Castillo, M.;
et al. Interobserver agreement between pathologists assessing tumor-infiltrating lymphocytes (TILs) in breast cancer using
methodology proposed by the International TILs Working Group. Ann. Surg. Oncol. 2016, 23, 2242–2248. [CrossRef] [PubMed]
32. Gao, G.; Wang, Z.; Qu, X.; Zhang, Z. Prognostic value of tumor-infiltrating lymphocytes in patients with triple-negative breast
cancer: A systematic review and meta-analysis. BMC Cancer 2020, 20, 179. [CrossRef] [PubMed]
33. Lee, K.; Lockhart, J.H.; Xie, M.; Chaudhary, R.; Slebos, R.J.; Flores, E.R.; Chung, C.H.; Tan, A.C. Deep Learning of Histopathology
Images at the Single Cell Level. Front. Artif. Intell. 2021, 4, 754641. [CrossRef] [PubMed]
34. Jiao, Y.; Li, J.; Qian, C.; Fei, S. Deep learning-based tumor microenvironment analysis in colon adenocarcinoma histopathological
whole-slide images. Comput. Methods Programs Biomed. 2021, 204, 106047. [CrossRef]
35. Failmezger, H.; Muralidhar, S.; Rullan, A.; de Andrea, C.E.; Sahai, E.; Yuan, Y. Topological tumor graphs: A graph-based spatial
model to infer stromal recruitment for immunosuppression in melanoma histology. Cancer Res. 2020, 80, 1199–1209. [CrossRef]
36. Moorman, A.; Vink, R.; Heijmans, H.; Van Der Palen, J.; Kouwenhoven, E. The prognostic value of tumour-stroma ratio in
triple-negative breast cancer. Eur. J. Surg. Oncol. 2012, 38, 307–313. [CrossRef]
37. Roeke, T.; Sobral-Leite, M.; Dekker, T.J.; Wesseling, J.; Smit, V.T.; Tollenaar, R.A.; Schmidt, M.K.; Mesker, W.E. The prognostic
value of the tumour-stroma ratio in primary operable invasive cancer of the breast: A validation study. Breast Cancer Res. Treat.
2017, 166, 435–445. [CrossRef]
38. Kather, J.N.; Krisam, J.; Charoentong, P.; Luedde, T.; Herpel, E.; Weis, C.A.; Gaiser, T.; Marx, A.; Valous, N.A.; Ferber, D.; et al.
Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Med.
2019, 16, e1002730. [CrossRef]
39. Geessink, O.G.; Baidoshvili, A.; Klaase, J.M.; Ehteshami Bejnordi, B.; Litjens, G.J.; van Pelt, G.W.; Mesker, W.E.; Nagtegaal, I.D.;
Ciompi, F.; van der Laak, J.A. Computer aided quantification of intratumoral stroma yields an independent prognosticator in
rectal cancer. Cell. Oncol. 2019, 42, 331–341. [CrossRef]
40. Zhang, T.; Xu, J.; Shen, H.; Dong, W.; Ni, Y.; Du, J. Tumor-stroma ratio is an independent predictor for survival in NSCLC. Int. J.
Clin. Exp. Pathol. 2015, 8, 11348.
41. Lu, M.Y.; Chen, T.Y.; Williamson, D.F.; Zhao, M.; Shady, M.; Lipkova, J.; Mahmood, F. AI-based pathology predicts origins for
cancers of unknown primary. Nature 2021, 594, 106–110. [CrossRef] [PubMed]
42. Zhu, X.; Yao, J.; Huang, J. Deep convolutional neural network for survival analysis with pathological images. In Proceedings of
the 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Shenzhen, China, 15–18 December 2016;
pp. 544–547.
43. Mobadersany, P.; Yousefi, S.; Amgad, M.; Gutman, D.A.; Barnholtz-Sloan, J.S.; Vega, J.E.V.; Brat, D.J.; Cooper, L.A. Predicting
cancer outcomes from histology and genomics using convolutional networks. Proc. Natl. Acad. Sci. USA 2018, 115, E2970–E2979.
[CrossRef] [PubMed]
44. He, B.; Bergenstråhle, L.; Stenbeck, L.; Abid, A.; Andersson, A.; Borg, Å.; Maaskola, J.; Lundeberg, J.; Zou, J. Integrating spatial
gene expression and breast tumour morphology via deep learning. Nat. Biomed. Eng. 2020, 4, 827–834. [CrossRef]
45. Wang, Y.; Kartasalo, K.; Weitz, P.; Acs, B.; Valkonen, M.; Larsson, C.; Ruusuvuori, P.; Hartman, J.; Rantalainen, M. Predicting
molecular phenotypes from histopathology images: A transcriptome-wide expression-morphology analysis in breast cancer.
Cancer Res. 2021, 81, 5115–5126. [CrossRef] [PubMed]
46. Wang, D.; Khosla, A.; Gargeya, R.; Irshad, H.; Beck, A.H. Deep learning for identifying metastatic breast cancer. arXiv 2016,
arXiv:1606.05718.
47. Steiner, D.F.; MacDonald, R.; Liu, Y.; Truszkowski, P.; Hipp, J.D.; Gammage, C.; Thng, F.; Peng, L.; Stumpe, M.C. Impact of deep
learning assistance on the histopathologic review of lymph nodes for metastatic breast cancer. Am. J. Surg. Pathol. 2018, 42, 1636.
[CrossRef]
48. Roux, L.; Racoceanu, D.; Loménie, N.; Kulikova, M.; Irshad, H.; Klossa, J.; Capron, F.; Genestie, C.; Le Naour, G.; Gurcan, M.N.
Mitosis detection in breast cancer histological images An ICPR 2012 contest. J. Pathol. Inform. 2013, 4, 8.
49. Bejnordi, B.E.; Veta, M.; Van Diest, P.J.; Van Ginneken, B.; Karssemeijer, N.; Litjens, G.; Van Der Laak, J.A.; Hermsen, M.; Manson,
Q.F.; Balkenhol, M.; et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women
with breast cancer. JAMA 2017, 318, 2199–2210. [CrossRef]
50. Kong, B.; Wang, X.; Li, Z.; Song, Q.; Zhang, S. Cancer metastasis detection via spatially structured deep network. In Information
Processing in Medical Imaging. IPMI 2017; Lecture Notes in Computer Science; Springer: Cham, Switzerland, 2017; pp. 236–248.
51. Li, Y.; Ping, W. Cancer metastasis detection with neural conditional random field. arXiv 2018, arXiv:1806.07064.
52. Lin, H.; Chen, H.; Graham, S.; Dou, Q.; Rajpoot, N.; Heng, P.A. Fast scannet: Fast and dense analysis of multi-gigapixel
whole-slide images for cancer metastasis detection. IEEE Trans. Med. Imaging 2019, 38, 1948–1958. [CrossRef]
53. Liu, Y.; Gadepalli, K.; Norouzi, M.; Dahl, G.E.; Kohlberger, T.; Boyko, A.; Venugopalan, S.; Timofeev, A.; Nelson, P.Q.; Corrado,
G.S.; et al. Detecting cancer metastases on gigapixel pathology images. arXiv 2017, arXiv:1703.02442.
Diagnostics 2022, 12, 2794 17 of 20

54. Bundy, A. Preparing for the Future of Artificial Intelligence; Executive Office of the President National Science and Technology
Council: Washington, DC, USA, 2017.
55. Sirinukunwattana, K.; Pluim, J.P.; Chen, H.; Qi, X.; Heng, P.A.; Guo, Y.B.; Wang, L.Y.; Matuszewski, B.J.; Bruni, E.; Sanchez, U.;
et al. Gland segmentation in colon histology images: The glas challenge contest. Med. Image Anal. 2017, 35, 489–502. [CrossRef]
[PubMed]
56. Fan, K.; Wen, S.; Deng, Z. Deep learning for detecting breast cancer metastases on WSI. In Innovation in Medicine and Healthcare
Systems, and Multimedia; Springer: Berlin/Heidelberg, Germany, 2019; pp. 137–145.
57. Zerhouni, E.; Lányi, D.; Viana, M.; Gabrani, M. Wide residual networks for mitosis detection. In Proceedings of the 2017 IEEE
14th International Symposium on Biomedical Imaging (ISBI 2017), Melbourne, Australia, 18–21 April 2017; pp. 924–928.
58. Rakhlin, A.; Tiulpin, A.; Shvets, A.A.; Kalinin, A.A.; Iglovikov, V.I.; Nikolenko, S. Breast tumor cellularity assessment using deep
neural networks. In Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision Workshops, Seoul, Korea,
27–28 October 2019.
59. Aresta, G.; Araújo, T.; Kwok, S.; Chennamsetty, S.S.; Safwan, M.; Alex, V.; Marami, B.; Prastawa, M.; Chan, M.; Donovan, M.; et al.
Bach: Grand challenge on breast cancer histology images. Med. Image Anal. 2019, 56, 122–139. [CrossRef] [PubMed]
60. Jahanifar, M.; Koohbanani, N.A.; Rajpoot, N. Nuclick: From clicks in the nuclei to nuclear boundaries. arXiv 2019,
arXiv:1909.03253.
61. Da, Q.; Huang, X.; Li, Z.; Zuo, Y.; Zhang, C.; Liu, J.; Chen, W.; Li, J.; Xu, D.; Hu, Z.; et al. DigestPath: A Benchmark Dataset with
Challenge Review for the Pathological Detection and Segmentation of Digestive-System. Med. Image Anal. 2022, 80, 102485.
[CrossRef] [PubMed]
62. Eloy, C.; Zerbe, N.; Fraggetta, F. Europe unites for the digital transformation of pathology: The role of the new ESDIP. J. Pathol.
Inform. 2021, 12, 10. [CrossRef]
63. Aubreville, M.; Stathonikos, N.; Bertram, C.A.; Klopleisch, R.; ter Hoeve, N.; Ciompi, F.; Wilm, F.; Marzahl, C.; Donovan, T.A.;
Maier, A.; et al. Mitosis domain generalization in histopathology images–The MIDOG challenge. arXiv 2022, arXiv:2204.03742.
64. Graham, S.; Jahanifar, M.; Vu, Q.D.; Hadjigeorghiou, G.; Leech, T.; Snead, D.; Raza, S.E.A.; Minhas, F.; Rajpoot, N. Conic: Colon
nuclei identification and counting challenge 2022. arXiv 2021, arXiv:2111.14485.
65. Weitz, P.; Valkonen, M.; Solorzano, L.; Hartman, J.; Ruusuvuori, P.; Rantalainen, M. ACROBAT—Automatic Registration of
Breast Cancer Tissue. In Proceedings of the 25th International Conference on Medical Image Computing and Computer Assisted
Intervention (MICCAI 2022), Singapore, 18–22 September 2022.
66. 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]
67. Guo, Z.; Liu, H.; Ni, H.; Wang, X.; Su, M.; Guo, W.; Wang, K.; Jiang, T.; Qian, Y. A fast and refined cancer regions segmentation
framework in whole-slide breast pathological images. Sci. Rep. 2019, 9, 882. [CrossRef]
68. Bulten, W.; Pinckaers, H.; van Boven, H.; Vink, R.; de Bel, T.; van Ginneken, B.; van der Laak, J.; Hulsbergen-van de Kaa, C.;
Litjens, G. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: A diagnostic study. Lancet
Oncol. 2020, 21, 233–241. [CrossRef]
69. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
70. Litjens, G.; Kooi, T.; Bejnordi, B.E.; Setio, A.A.A.; Ciompi, F.; Ghafoorian, M.; Van Der Laak, J.A.; Van Ginneken, B.; Sánchez, C.I.
A survey on deep learning in medical image analysis. Med. Image Anal. 2017, 42, 60–88. [CrossRef] [PubMed]
71. Shen, D.; Wu, G.; Suk, H.I. Deep learning in medical image analysis. Annu. Rev. Biomed. Eng. 2017, 19, 221–248. [CrossRef]
72. Yi, X.; Walia, E.; Babyn, P. Generative adversarial network in medical imaging: A review. Med. Image Anal. 2019, 58, 101552.
[CrossRef] [PubMed]
73. Zhang, Z.; Chen, P.; McGough, M.; Xing, F.; Wang, C.; Bui, M.; Xie, Y.; Sapkota, M.; Cui, L.; Dhillon, J.; et al. Pathologist-level
interpretable whole-slide cancer diagnosis with deep learning. Nat. Mach. Intell. 2019, 1, 236–245. [CrossRef]
74. 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, 3308. [CrossRef]
75. Im, S.; Hyeon, J.; Rha, E.; Lee, J.; Choi, H.J.; Jung, Y.; Kim, T.J. Classification of diffuse glioma subtype from clinical-grade
pathological images using deep transfer learning. Sensors 2021, 21, 3500. [CrossRef]
76. Mi, W.; Li, J.; Guo, Y.; Ren, X.; Liang, Z.; Zhang, T.; Zou, H. Deep learning-based multi-class classification of breast digital
pathology images. Cancer Manag. Res. 2021, 13, 4605. [CrossRef]
77. Hu, Y.; Su, F.; Dong, K.; Wang, X.; Zhao, X.; Jiang, Y.; Li, J.; Ji, J.; Sun, Y. Deep learning system for lymph node quantification and
metastatic cancer identification from whole-slide pathology images. Gastric Cancer 2021, 24, 868–877. [CrossRef]
78. Pei, L.; Jones, K.A.; Shboul, Z.A.; Chen, J.Y.; Iftekharuddin, K.M. Deep neural network analysis of pathology images with
integrated molecular data for enhanced glioma classification and grading. Front. Oncol. 2021, 11, 2572. [CrossRef]
79. Salvi, M.; Bosco, M.; Molinaro, L.; Gambella, A.; Papotti, M.; Acharya, U.R.; Molinari, F. A hybrid deep learning approach for
gland segmentation in prostate histopathological images. Artif. Intell. Med. 2021, 115, 102076. [CrossRef] [PubMed]
Diagnostics 2022, 12, 2794 18 of 20

80. Lu, Z.; Zhan, X.; Wu, Y.; Cheng, J.; Shao, W.; Ni, D.; Han, Z.; Zhang, J.; Feng, Q.; Huang, K. BrcaSeg: A deep learning approach
for tissue quantification and genomic correlations of histopathological images. Genom. Proteom. Bioinform. 2021, 19, 1032–1042.
[CrossRef] [PubMed]
81. Cheng, S.; Liu, S.; Yu, J.; Rao, G.; Xiao, Y.; Han, W.; Zhu, W.; Lv, X.; Li, N.; Cai, J.; et al. Robust whole slide image analysis for
cervical cancer screening using deep learning. Nat. Commun. 2021, 12, 5639. [CrossRef] [PubMed]
82. Kers, J.; Bülow, R.D.; Klinkhammer, B.M.; Breimer, G.E.; Fontana, F.; Abiola, A.A.; Hofstraat, R.; Corthals, G.L.; Peters-Sengers, H.;
Djudjaj, S.; et al. Deep learning-based classification of kidney transplant pathology: A retrospective, multicentre, proof-of-concept
study. Lancet Digit. Health 2022, 4, e18–e26. [CrossRef]
83. Zhou, C.; Jin, Y.; Chen, Y.; Huang, S.; Huang, R.; Wang, Y.; Zhao, Y.; Chen, Y.; Guo, L.; Liao, J. Histopathology classification and
localization of colorectal cancer using global labels by weakly supervised deep learning. Comput. Med. Imaging Graph. 2021,
88, 101861. [CrossRef]
84. Höhn, J.; Krieghoff-Henning, E.; Jutzi, T.B.; von Kalle, C.; Utikal, J.S.; Meier, F.; Gellrich, F.F.; Hobelsberger, S.; Hauschild, A.;
Schlager, J.G.; et al. Combining CNN-based histologic whole slide image analysis and patient data to improve skin cancer
classification. Eur. J. Cancer 2021, 149, 94–101. [CrossRef]
85. Shin, S.J.; You, S.C.; Jeon, H.; Jung, J.W.; An, M.H.; Park, R.W.; Roh, J. Style transfer strategy for developing a generalizable deep
learning application in digital pathology. Comput. Methods Programs Biomed. 2021, 198, 105815. [CrossRef]
86. Naik, N.; Madani, A.; Esteva, A.; Keskar, N.S.; Press, M.F.; Ruderman, D.; Agus, D.B.; Socher, R. Deep learning-enabled breast
cancer hormonal receptor status determination from base-level H&E stains. Nat. Commun. 2020, 11, 5727.
87. Yao, J.; Zhu, X.; Jonnagaddala, J.; Hawkins, N.; Huang, J. Whole slide images based cancer survival prediction using attention
guided deep multiple instance learning networks. Med. Image Anal. 2020, 65, 101789. [CrossRef]
88. Campanella, G.; Hanna, M.G.; Geneslaw, L.; Miraflor, A.; Werneck Krauss Silva, V.; Busam, K.J.; Brogi, E.; Reuter, V.E.; Klimstra,
D.S.; Fuchs, T.J. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat.
Med. 2019, 25, 1301–1309. [CrossRef]
89. Lu, M.Y.; Williamson, D.F.; Chen, T.Y.; Chen, R.J.; Barbieri, M.; Mahmood, F. Data-efficient and weakly supervised computational
pathology on whole-slide images. Nat. Biomed. Eng. 2021, 5, 555–570. [CrossRef] [PubMed]
90. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the 2016 IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
91. Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmentation. In Proceedings of the 2015 IEEE
Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 3431–3440.
92. Louis, D.N.; Perry, A.; Reifenberger, G.; Von Deimling, A.; Figarella-Branger, D.; Cavenee, W.K.; Ohgaki, H.; Wiestler, O.D.;
Kleihues, P.; Ellison, D.W. The 2016 World Health Organization classification of tumors of the central nervous system: A summary.
Acta Neuropathol. 2016, 131, 803–820. [CrossRef] [PubMed]
93. Faust, K.; Bala, S.; Van Ommeren, R.; Portante, A.; Al Qawahmed, R.; Djuric, U.; Diamandis, P. Intelligent feature engineering and
ontological mapping of brain tumour histomorphologies by deep learning. Nat. Mach. Intell. 2019, 1, 316–321. [CrossRef]
94. Roohi, A.; Faust, K.; Djuric, U.; Diamandis, P. Unsupervised machine learning in pathology: The next frontier. Surg. Pathol. Clin.
2020, 13, 349–358. [CrossRef] [PubMed]
95. Bulten, W.; Litjens, G. Unsupervised prostate cancer detection on H&E using convolutional adversarial autoencoders. arXiv 2018,
arXiv:1804.07098.
96. Veta, M.; Heng, Y.J.; Stathonikos, N.; Bejnordi, B.E.; Beca, F.; Wollmann, T.; Rohr, K.; Shah, M.A.; Wang, D.; Rousson, M.; et al.
Predicting breast tumor proliferation from whole-slide images: The TUPAC16 challenge. Med. Image Anal. 2019, 54, 111–121.
[CrossRef]
97. Koohbanani, N.A.; Jahanifar, M.; Tajadin, N.Z.; Rajpoot, N. NuClick: A deep learning framework for interactive segmentation of
microscopic images. Med. Image Anal. 2020, 65, 101771. [CrossRef]
98. Xu, G.; Song, Z.; Sun, Z.; Ku, C.; Yang, Z.; Liu, C.; Wang, S.; Ma, J.; Xu, W. Camel: A weakly supervised learning framework for
histopathology image segmentation. In Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision, Seoul,
Korea, 27 October–2 November 2019; pp. 10682–10691.
99. Huss, R.; Coupland, S.E. Software-assisted decision support in digital histopathology. J. Pathol. 2020, 250, 685–692. [CrossRef]
100. Madabhushi, A.; Lee, G. Image analysis and machine learning in digital pathology: Challenges and opportunities. Med. Image
Anal. 2016, 33, 170–175. [CrossRef]
101. Hughes, A.J.; Mornin, J.D.; Biswas, S.K.; Beck, L.E.; Bauer, D.P.; Raj, A.; Bianco, S.; Gartner, Z.J. Quanti.us: A tool for rapid,
flexible, crowd-based annotation of images. Nat. Methods 2018, 15, 587–590. [CrossRef]
102. López-Pérez, M.; Amgad, M.; Morales-Álvarez, P.; Ruiz, P.; Cooper, L.A.; Molina, R.; Katsaggelos, A.K. Learning from crowds in
digital pathology using scalable variational Gaussian processes. Sci. Rep. 2021, 11, 11612. [CrossRef] [PubMed]
103. Colling, R.; Pitman, H.; Oien, K.; Rajpoot, N.; Macklin, P.; CM-Path AI in Histopathology Working Group; Bachtiar, V.; Booth, R.;
Bryant, A.; Bull, J.; et al. Artificial intelligence in digital pathology: A roadmap to routine use in clinical practice. J. Pathol. 2019,
249, 143–150. [CrossRef] [PubMed]
104. Acs, B.; Rantalainen, M.; Hartman, J. Artificial intelligence as the next step towards precision pathology. J. Intern. Med. 2020,
288, 62–81. [CrossRef] [PubMed]
Diagnostics 2022, 12, 2794 19 of 20

105. Khened, M.; Kori, A.; Rajkumar, H.; Krishnamurthi, G.; Srinivasan, B. A generalized deep learning framework for whole-slide
image segmentation and analysis. Sci. Rep. 2021, 11, 11579. [CrossRef] [PubMed]
106. Pradhan, K.S.; Chawla, P.; Tiwari, R. HRDEL: High Ranking Deep Ensemble Learning-based Lung Cancer Diagnosis Model.
Expert Syst. Appl. 2022, 213, 118956. [CrossRef]
107. Narla, A.; Kuprel, B.; Sarin, K.; Novoa, R.; Ko, J. Automated classification of skin lesions: From pixels to practice. J. Investig.
Dermatol. 2018, 138, 2108–2110. [CrossRef]
108. Bera, K.; Schalper, K.A.; Rimm, D.L.; Velcheti, V.; Madabhushi, A. Artificial intelligence in digital pathology—New tools for
diagnosis and precision oncology. Nat. Rev. Clin. Oncol. 2019, 16, 703–715. [CrossRef]
109. Niazi, M.K.K.; Parwani, A.V.; Gurcan, M.N. Digital pathology and artificial intelligence. Lancet Oncol. 2019, 20, e253–e261.
[CrossRef]
110. Wu, Y.; Cheng, M.; Huang, S.; Pei, Z.; Zuo, Y.; Liu, J.; Yang, K.; Zhu, Q.; Zhang, J.; Hong, H.; et al. Recent Advances of Deep
Learning for Computational Histopathology: Principles and Applications. Cancers 2022, 14, 1199. [CrossRef]
111. Swiderska-Chadaj, Z.; de Bel, T.; Blanchet, L.; Baidoshvili, A.; Vossen, D.; van der Laak, J.; Litjens, G. Impact of rescanning and
normalization on convolutional neural network performance in multi-center, whole-slide classification of prostate cancer. Sci.
Rep. 2020, 10, 14398. [CrossRef]
112. Kleppe, A.; Skrede, O.J.; De Raedt, S.; Liestøl, K.; Kerr, D.J.; Danielsen, H.E. Designing deep learning studies in cancer diagnostics.
Nat. Rev. Cancer 2021, 21, 199–211. [CrossRef] [PubMed]
113. Collins, G.S.; Moons, K.G. Reporting of artificial intelligence prediction models. Lancet 2019, 393, 1577–1579. [CrossRef]
114. Kelly, C.J.; Karthikesalingam, A.; Suleyman, M.; Corrado, G.; King, D. Key challenges for delivering clinical impact with artificial
intelligence. BMC Med. 2019, 17, 195. [CrossRef]
115. Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.
Nat. Mach. Intell. 2019, 1, 206–215. [CrossRef] [PubMed]
116. Shrikumar, A.; Greenside, P.; Kundaje, A. Learning important features through propagating activation differences. In Proceedings
of the 34th International Conference on Machine Learning, Sydney, Australia, 6–11 August 2017; pp. 3145–3153.
117. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks
via Gradient-Based Localization. In Proceedings of the 2017 IEEE International Conference on Computer Vision, Venice, Italy,
22–29 October 2017; pp. 618–626.
118. Guidotti, R.; Monreale, A.; Ruggieri, S.; Turini, F.; Giannotti, F.; Pedreschi, D. A survey of methods for explaining black box
models. ACM Comput. Surv. 2018, 51, 93. [CrossRef]
119. Scherer, M.U. Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies. Harv. JL Tech. 2015,
29, 353. [CrossRef]
120. House, W. Guidance for regulation of artificial intelligence applications. Memo. Heads Exec. Dep. Agencies 2020.
121. Commission, E. On Artificial Intelligence—A European Approach to Excellence and Trust; European Commission Luxembourg:
Luxembourg, 2020.
122. Pita, E.V. UNESCO and the Governance of Artificial Intelligence in a Globalized World. The Need for a New Legal Architecture.
Ann. Fac. Der. U. Extremad. 2021, 37, 273.
123. Allen, T.C. Regulating Artificial Intelligence for a Successful Pathology Future. Arch. Pathol. Lab. Med. 2019, 143, 1175–1179.
[CrossRef]
124. Guihot, M.; Matthew, A.F.; Suzor, N.P. Nudging robots: Innovative solutions to regulate artificial intelligence. Vand. J. Ent. Tech.
L. 2017, 20, 385.
125. von Eschenbach, W.J. Transparency and the black box problem: Why we do not trust AI. Philos. Technol. 2021, 34, 1607–1622.
[CrossRef]
126. Parwani, A.V. Next generation diagnostic pathology: Use of digital pathology and artificial intelligence tools to augment a
pathological diagnosis. Diagn. Pathol. 2019, 14, 138. [CrossRef] [PubMed]
127. Tosun, A.B.; Pullara, F.; Becich, M.J.; Taylor, D.; Fine, J.L.; Chennubhotla, S.C. Explainable AI (xAI) for anatomic pathology. Adv.
Anat. Pathol. 2020, 27, 241–250. [CrossRef] [PubMed]
128. Krajna, A.; Kovac, M.; Brcic, M.; Šarčević, A. Explainable Artificial Intelligence: An Updated Perspective. In Proceedings of
the 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO), Opatija,
Croatia, 23–27 May 2022; pp. 859–864.
129. Binder, A.; Bockmayr, M.; Hägele, M.; Wienert, S.; Heim, D.; Hellweg, K.; Ishii, M.; Stenzinger, A.; Hocke, A.; Denkert, C.; et al.
Morphological and molecular breast cancer profiling through explainable machine learning. Nat. Mach. Intell. 2021, 3, 355–366.
[CrossRef]
130. Lage, I.; Chen, E.; He, J.; Narayanan, M.; Kim, B.; Gershman, S.; Doshi-Velez, F. An evaluation of the human-interpretability of
explanation. arXiv 2019, arXiv:1902.00006.
131. Sixt, L.; Granz, M.; Landgraf, T. When explanations lie: Why modified BP attribution fails. arXiv 2019, arXiv:1912.09818.
132. Holzinger, A.; Malle, B.; Saranti, A.; Pfeifer, B. Towards multi-modal causability with graph neural networks enabling information
fusion for explainable AI. Inf. Fusion 2021, 71, 28–37. [CrossRef]
133. Zhang, J.; Chen, J.; Xuan, Q. Link Prediction Based on Hyper-Substructure Network. In Graph Data Mining; Springer:
Berlin/Heidelberg, Germany, 2021; pp. 29–48.
Diagnostics 2022, 12, 2794 20 of 20

134. Adnan, M.; Kalra, S.; Tizhoosh, H.R. Representation learning of histopathology images using graph neural networks. In
Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, Seattle, WA, USA,
14–19 June 2020; pp. 988–989.
135. Koromina, M.; Pandi, M.T.; Patrinos, G.P. Rethinking drug repositioning and development with artificial intelligence, machine
learning, and omics. OMICS J. Integr. Biol. 2019, 23, 539–548. [CrossRef]