Complex & Intelligent Systems

https://doi.org/10.1007/s40747-024-01406-2

ORIGINAL ARTICLE

Malaria parasitic detection using a new Deep Boosted and Ensemble

Learning framework
Hafiz M. Asif1 · Saddam Hussain Khan2 · Tahani Jaser Alahmadi3 · Tariq Alsahfi4 · Amena Mahmoud5

Received: 30 July 2023 / Accepted: 22 February 2024

© The Author(s) 2024

Abstract
Malaria is a potentially fatal plasmodium parasite injected by female anopheles mosquitoes that infect red blood cells and
cause millions of lifelong disability worldwide yearly. However, specialists’ manual screening in clinical practice is laborious
and prone to error. Therefore, a novel Deep Boosted and Ensemble Learning (DBEL) framework, comprising the stacking of
new Boosted-BR-STM convolutional neural networks (CNN) and the ensemble ML classifiers, is developed to screen malaria
parasite images. The proposed Boosted-BR-STM is based on a new dilated-convolutional block-based Split Transform
Merge (STM) and feature-map Squeezing–Boosting (SB) ideas. Moreover, the new STM block uses regional and boundary
operations to learn the malaria parasite’s homogeneity, heterogeneity, and boundary with patterns. Furthermore, the diverse
boosted channels are attained by employing Transfer Learning-based new feature-map SB in STM blocks at the abstract,
medium, and conclusion levels to learn minute intensity and texture variation of the parasitic pattern. Additionally, to enhance
the learning capacity of Boosted-BR-STM and foster a more diverse representation of features, boosting at the final stage
is achieved through TL by utilizing multipath residual learning. The proposed DBEL framework implicates the stacking of
prominent and diverse boosted channels and provides the generated discriminative features of the developed Boosted-BR-
STM to the ensemble of ML classifiers. The proposed framework improves the discrimination ability and generalization of
ensemble learning. Moreover, the deep feature spaces of the developed Boosted-BR-STM and customized CNNs are fed into
ML classifiers for comparative analysis. The proposed DBEL framework outperforms the existing techniques on the NIH
malaria dataset that are enhanced using discrete wavelet transform to enrich feature space. The proposed DBEL framework
achieved Accuracy (98.50%), Sensitivity (0.9920), F-score (0.9850), and AUC (0.9960), which suggests it to be utilized for
malaria parasite screening.

Keywords Screening · Squeezing · Boosting · Split-transform and merge · Transfer learning · Malaria · Parasite · Cognitive ·
Disabilities

Introduction
B Saddam Hussain Khan
saddamhkhan@ueas.edu.pk Malaria is a life-threatening illness transmitted by female
B Tahani Jaser Alahmadi Anopheles mosquitoes that inject plasmodium parasites with
tjalahmadi@pnu.edu.sa one nasty bite. In most cases, plasmodium parasites target
Hafiz M. Asif healthy red blood cells (RBC) around 1 or 2 weeks after their
h.asif@squ.edu.om
3 Department of Information Systems, College of Computer
Tariq Alsahfi
tmalsahfi@uj.edu.sa and Information Sciences, Princess Nourah bint Abdulrahman
University, P.O. Box 84428, 11671 Riyadh, Saudi Arabia
Amena Mahmoud
4 Department of Information Systems and Technology, College
Amena_mahmoud@fci.kfs.edu.eg
of Computer Science and Engineering, University of Jeddah,
1 Department of Electrical and Computer Engineering, Sultan Jeddah, Saudi Arabia
Qaboos University, Muscat, Oman 5 Faculty of Computers and Information, Department of
2 Department of Computer Systems Engineering, University of Computer Science, KafrElSkeikh University, Kafr el-Sheikh,
Engineering and Applied Sciences, Swat 19060, Pakistan Egypt

123
 Complex & Intelligent Systems

emergence in the human body [1]. This bacterial infection Transform Merge (STM) block. Moreover, the developed
is hazardous to kids, persons with impaired immune sys- STM block of deep Boosted-BR-STM uses the concept of
tems, pregnant women, and the elderly at risk [2]. Falciparum homogenous and heterogeneous. The significant contribu-
malaria is particularly dangerous to pregnant women as it tions are as follows:
increases stillbirth, maternal death, miscarriage, and new-
born [3]. The World Health Organization reported roughly 1. A new Deep Boosted and Ensemble Learning (DBEL)
241 million malaria suspects and 627,000 fatalities in 2021. framework is proposed comprising a new residual learn-
The African continent is perhaps the most afflicted, account- ing based Boosted-BR-STM CNN and ensemble learning
ing for 95% of 90% of deaths and 80% of child disability for detecting RBCs infected with the plasmodium falci-
caused by acute malaria [4, 5]. parum using blood smear images. The dataset is initially
RBCs were microscopically examined in a thick, thin enhanced and reduced in dimension using a discrete
blood smear frequently used to identify malaria [6]. The thick wavelet transform (DWT) to improve computational
and thin-smear test aids in identifying the density of parasites complexity.
in a person’s body and malaria species, respectively [7, 8]. 2. The proposed deep Boosted-BR-STM exploits the chan-
Expert pathologists manually analyze blood smear films to nel SB technique with a novel STM block. Moreover, the
get a microscopic diagnosis which is time-consuming, labo- developed STM block uses the concept of homogenous
rious, and unreliable [9]. Malaria patients are usually seen and heterogeneous.
in emerging countries, where diagnostic lab facilities and 3. The innovative SB concept is carefully integrated into
tools are unavailable. In addition, a global shortage of trained the new STM block at abstract, medium, and conclusion
professionals significantly impacts the healthcare systems levels effectively capturing the diverse pattern of homo-
of developing countries [10]. Therefore, a computer-based geneous, heterogeneous, contrast, and textural variations
screening tool is essential for speedy and reliable malaria of the parasitic cell. SB notion is utilized by merging
analysis [11]. reduced prominent channels with TL-based extracted
Artificial intelligence (AI) and machine learning (ML) additional feature maps to improve Boosted-BR-STM
have aided in the development of malaria diagnostic meth- performances. In addition, to improve the learning capac-
ods that are effective and precise in processing large amounts ity of Boosted-BR-STM and promote a more diverse
of parasite-contaminated RBC samples [12–14]. Computer- representation of features, residual learning-based fea-
based software will support clinicians in diagnosis and ture map boosting is achieved at the final stage through
therapy, facilitating established lab practices [15–17]. How- TL.
ever, this convention performed poorly on massive data 4. The proposed framework grants the inherent properties
and was incapable of learning complex patterns. There- of diverse prominent and boosted channels to the dis-
fore, deep learning (DL) algorithms emerged and inspired criminative feature level and fed to the ensemble of ML
researchers’ interest in dealing with enormous amounts of classifiers, improving the capability of discrimination
data and learning complicated patterns [18, 19]. They have and generalization of ensemble learning. The ML clas-
significant growth for medical imaging infection diagnosis sifiers’ ensemble effectively reduces feature dimension
[20–22]. Malaria parasite analysis is crucial for diagnosing and improves diverse decision space. Ultimately, early
infected cells. In this regard, a deep CNN-based identification detection reduces the possibility of permanent disabili-
helps quickly and accurately analyze malaria parasite images. ties.
Several CNN-based classification frameworks and experi- 5. The proposed hybrid DBEL framework performance is
mental models have extensively been employed on the NIH compared to existing techniques utilized in NIH-malaria
malaria dataset to improve detection [23, 24]. CNN extracts original and enhanced datasets.
deep features automatically by avoiding the time-consuming
hand-crafted feature extraction and reduces computational The rest of the manuscript is arranged in the follow-
power [23, 25]. ing ways. “Related work” and “Malaria parasite detection
To our knowledge, this study is the first to introduce a scheme”, the remainder of the article discusses related work
new deep hybrid framework comprised of novel Deep resid- and the proposed malaria detection framework. Materials
ual and spatial blocks CNN and ensemble learning to analyze and implementation details are provided in “Experimental
malaria parasite-afflicted patients accurately and efficiently. configuration”. The performance analysis and outcomes are
The proposed deep Boosted-BR-STM exploits the channel presented in “Results and discussion”. Finally, the article’s
Squeezing and Boosting (SB) technique with a novel Split conclusion is in “Conclusion”.

123
Complex & Intelligent Systems

Related work performance. The prior research primarily focused on the

accuracy of the validation dataset. Mostly, the previous stud-
Several ML techniques were automatically determined in ies are evaluated using accuracy; however, the results of a
stained RBC samples employing methods over time [26]. single-measure evaluation of the detecting system could be
ML algorithms such as Extreme-Learning Machine methods, more reliable. Therefore, the detection rate and F-score are
Global White Balance, Logistic Regression, SVM, Adaptive essential indicators of parasite detection. Moreover, testing
Non-linear Grayscale, and KNN have enhanced classifica- evaluation measures on a vast, previously unseen test dataset
tion performance [27, 28]. The malaria parasite-infected are necessary for determining the detection model’s robust-
features like RBCs, color, size, form, and statistical data ness. Furthermore, several models have been evaluated on a
were extracted and provided to ML algorithms [29, 30]. limited dataset of images of malaria parasites.
Moreover, the deep belief network technique was used to
classify the parasites, which yielded a 90.21% accuracy [31].
The approaches mentioned above, on the other hand, indi- Malaria parasite detection scheme
cated prediction accuracy in the range of 84–94%. Several
researchers used deep CNN-based detection that automati- The proposed DBEL framework and subsequent prompt
cally and quickly identifies cells infected with the malaria effective treatment in sufficient time can detect permanent
parasite, achieving a 91.50% detection rate [31–33]. Med- disabilities from the malaria parasite. The microscope ana-
ical datasets with labels are generally limited in size and lyzes malaria parasites from blood films stained with various
thus unsuitable for practical application. Therefore, TL has chemical stains [42]. Various scanners, chemical stains,
been utilized in a limited labeled malaria dataset to forecast concentrations, and methods can cause color differences.
malaria parasites in RBC [32]. TL-based existing CNNs have Therefore, analyzing parasite and artifact patterns becomes
been employed to improve and have attained accuracies of challenging and causes substantial margin errors. These chal-
75–94% [33–35]. Furthermore, R-CNN for parasite identi- lenges necessitate designing a new deep CNN and ensemble
fication has been employed with considerable performance learning framework to capture complicated patterns and iden-
[36]. tify malaria infectious microscopic thin smear samples. The
Previous studies employed VGG16 to identify malaria developed framework comprised a new Boosted-BR-STM,
using the common NIH dataset and achieved 95.96% accu- ensemble learning, and contemporary customized CNNs to
racy [37]. The balanced NIH dataset included 27,556 images screen parasite-infected cells from healthy images. The pro-
and was resized to 224 by 224 pixels. A sequential-tailored posed malaria detection scheme utilizes three experimental
CNN was obtained with a 95.90% F1 score, 94.7% sensi- arrangements: (1) the proposed DBEL, (2) the proposed
tivity, and 92.70% accuracy [25]. The existing ResNet50 DBML framework, and (3) the evaluation of existing CNNs.
yielded a 95.40% accuracy using the malaria dataset [38]. These existing CNNs learned from scratch and fine-tuned
Moreover, the custom-made CNN model demonstrated using TL on the malaria NIH dataset. Moreover, these CNNs
96.82% accuracy, 96.33% using five convolutional and pool- are evaluated as (i) The deep features are extracted from
ing layers, and 96.82% F1-score [23]. Eventually, a hybrid existing CNNs and provided to the ML classifier (ii) The
platform that systematically reduces structural and empiri- Softmax-based evaluations. Additionally, the DWT and data
cal risk has been recently adopted and achieved a sensitivity augmentation as pre-processing techniques for improving the
of 93.44% and an accuracy of 93.13% [25, 32]. In another detection CNNs performance to screen stain impurities or
study, a modified Capsule Network (CapsNet) hybrid screen- artifacts from parasite images and analyze different parasitic
ing algorithm has been reported for automated identification patterns. The overall workflow is arranged in Fig. 1.
and pixel label based classification (segmentation) of malaria
parasite-infected RBCs, achieving an accuracy of 98.70% Data enhancement
[39]. Moreover, modified YOLOV3 and YOLOV4 mod-
els have been employed on a publicly available malaria Discrete wavelet transformations (DWT) effectively produce
dataset and achieved an accuracy of 95.46% and 96.14%, salient feature maps and reduce noise [43]. DWT divides the
respectively [40]. A recent study employed YOLOv5 and image into four sub-parts: high, low, and diagonal-resolution
DarKnet-53 Plasmodium Falciparum life stages detection channels. The next level of DWT, the low and diagonal-
and achieved an accuracy of 95.20% and 96.02%, respec- resolution channels, is stored and altered to provide the
tively [41]. However, certain limitations to the prior work necessary DWT coefficients for parasite detection. Moreover,
are: previous CNN models have been used to detect par- a 75% dimension reduction at each level minimizes computa-
asite images designed specifically for natural samples. In tional complexities. The low- and high-resolution images are
contrast, the medical images have distinct patterns and tex- combined and reconstructed using IDWT to produce highly
tures of the parasite malaria infection, limiting the model’s informative (low-resolution and diagonal features) enhanced

123
 Complex & Intelligent Systems

A Malaria Parasite Detection Frameworks (Overview)

Parasitic
NIH Malaria
Pre-processing Proposed Frameworks
Dataset
Non-Parasitic

Malaria Parasite Detection Frameworks (Detailed)

Proposed Deep Boosted Feature Spaces and Ensemble Learning
Proposed Boosted- zDeepBoosted Ensemble of
B BR-STM Feature Space Classifiers

Proposed Deep Boosted Feature Spaces and ML Classifiers

SVM
Proposed Boosted- Deep Boosted
AdaBoostM1
BR-STM Feature Space
MLP Parasitic
Pre-Processing
Malaria Dataset
(DWT)
Non-Parasitic
Trained from Scratch and TL (Normal, Artifact)
SVM
Implementation of
Customized CNNs
Softmax

C
Validation Malaria Parasite Detection Frameworks
Dataset Validation Loss

Parameter Optimization
Model Validation
Proposed Deep Boosted Feature Spaces and Ensemble Learning

Proposed Boosted- Deep Boosted Ensemble of

BR-STM Feature Space Classifiers
Data Augmentation

70% data

Proposed Deep Boosted Feature Spaces and ML Classifiers

Pre- Processing Training
Dataset SVM
Proposed Boosted- Deep Boosted
Enhanced Dataset (DWT) BR-STM
AdaBoostM1
Feature Space
MLP

Softmax and ML Based Detection of Customized CNNs

Dataset Implementation of
well-established and TL-based SVM
CNNs
VGG
GoogleNet
InceptionV3
ResNet Softmax
DenseNet-201
ShuffleNet
Xception

Training Phase

Trained Models
Ensemble of
Classifiers
Proposed Boosted-BR- Parasitic
STM SVM
NIH Malaria Dataset AdaBoostM1
MLP
Testing Data Trained from Scratch
30% data
and TL-based Non-Parasitic
Softmax (Normal,
Artifact)
Enhanced Dataset
Testing Phase

Fig. 1 The flow diagram of the developed malaria parasite detection scheme

123
Complex & Intelligent Systems

Non-Parasitic (Healthy / Stain Impurities) Parasitic

c
a

d
b

Fig. 2 Parasite and non-parasite (healthy, stain artifact) samples from the NIH malaria dataset are shown in panels (a, c). While (b, d) displays their
DWT-enhanced examples, respectively

Table 1 Data augmentation detail The proposed channel Boosted-BR-STM CNN

Parameter Values
This study introduces a novel Boosted-BR-STM CNN for the
Rotation [± 30] ° detection of parasitic malaria in patients. The CNN incor-
Shearing [0 30]
porates a unique STM-based convolution block to analyze
both homogenous and boundary patterns in parasitic regions.
X–Y reflection [± 1]
Additionally, we employ novel SB concepts within STM to
produce condensed and diverse feature maps. The channel
SB technique is applied at the initial, intermediate, and final
stages to capture textural differences between parasites and
images [44]. The original and improved DWT data are shown
typical artifacts [46], as illustrated in Fig. 3.
in Fig. 2.
The Boosted-BR-STM effectively learns abstract and tar-
get features by organizing three STM blocks with a consistent
structure. Each block consists of four dilated-convolutional
Data augmentation units, employing max and average pooling to explore border
determination, regional consistency, and textural variation
Data augmentation is widely applied to boost the robust- [47, 48], as shown in the Eqs. (1–3). This systematic
ness of the model and lessen the risk of overfitting. Data implementation aids in efficiently exploring border or edge
instances are increased in runtime (on-the-fly augmentation) determination, regional homogenous, and textural variation
during the model training using different data augmenta- of parasitic and artifact images.
tion techniques [45]. This work uses rotation, reflection, and
shearing, as shown in Table 1. Moreover, scaling and trans- 
m 
n
xk, l  xk+i−1, l+n−1 fi, j , (1)
lation parameters are kept by default [1 1] and translation [±
i1 j1
5], respectively. Data augmentation strategies are employed
to learn different variations and improve the model”s gen- xmax k, l  maxi1, ..., w, j1, ..., w xk+i−1, l+ j−1 , (2)
eralization, especially, shearing learn distortions, real-world
w w
variations, and occlusions. 1 
xavg k, l  xk+i−1, l+ j−1 . (3)
w2
i1 j1

The proposed deep boosted and ensemble learning In the equation, ‘x’ represents the input feature map, size
framework is shown with the symbols ‘k’ x ‘l’ and ‘I’ x ‘j’, and ‘f’
stands for the filter (1) in Eqs. (2–3). As indicated in the
This research proposes a new deep DBEL framework that is equation, every one of the four convolutional blocks (B, C,
comprised of the developed Boosted-BR-STM BRNet and D, and E) uses channel SB differently to learn distinct par-
ensemble techniques for malaria parasite detection. More- asitic feature sets (4). While learning blocks D and E from
over, the developed Boosted-BR-STM penultimate layers are scratch, TL builds additional channels in blocks B, C, M, and
used for deep feature extraction. The workflow of the malaria N to provide different feature maps. Each STM convolutional
parasite detection framework is illustrated in Fig. 3. block has 32, 64, 128, and 128, 256, 512 channel dimensions

123
 Complex & Intelligent Systems

Proposed Boosted-BR-STM Detection CNN

TL-based Fine-Tuned 256
32,32 64,64 128,128 256, 256
Auxiliary
Learner
Blocks Blocks Blocks Blocks Feature
Maps
N N M N M N M N

64,64, 32 128, 128, 64 256,256, 128

A Deep Feature
Fully Connected Layers
Extraction
Detection
Block B Block B Block B Dropout
50%

Parasitic
DWT Image
128 256 768 768, 256 SVM
64,64 Block C Block C Block C
AdaBoost Ensemble of
Block A + Boosted + Boosted
Channels
+ Boosted Block F
M1 Classifiers
Channels

(Non-Parasitic)
Channels
Block D Block D Block D

Healthy
MLP
256
1024
Block E Block E Block E Ensemble Learning

B Block M Block N

Evaluation
Accuracy

Model
F-score
Block A Block B Block C Block D Block E Block F
Sensitivity
Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 1×1 Convol. 3×3
Convol Specificity
Precision
RelU RelU RelU RelU RelU RelU RelU RelU
Rel PR-AUC
ROC-AUC
Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 3×3 Convol. 1×1 Convol. 3×3 Convol. 3×3
Convol

RelU RelU RelU RelU RelU RelU + + SVM: Support Vector Machine
MLP: Multi Layer Perceptron
Average-Pooling Convol. 1×1 Convol. 1×1 Average-Pooling Max-Pooling Max- Pooling RelU
l RelU
Re

RelU RelU Convol. 1×1 Convol. 1×1

Average-Pooling Max-Pooling RelU RelU

Max-Pooling Average-Pooling

Fig. 3 The proposed malaria parasite detection framework comprised the developed boosted-BR-STM and ensemble of classifiers

when squeezed and boosted, correspondingly [49]. Finally, and preserve target-specific features and minimize overfit-
a systematic approach has been adopted, which involves ting. va serve as an example of the number of neurons in
the stacking of TL-based residual learning-based M and N Eq. (5). Lastly, Eq. (6), where c stands for the number of
blocks, culminating in their concatenation at the final stage classes, represents softmax, an activation function.
to effectively learn diverse feature spaces. Three residual
Significance of using auxiliary channels and Squeezing—
blocks are sequentially arranged to facilitate the acquisition
Boosting (SB) Ideas
of diverse features, wherein the number of channels progres-
The hybrid framework’s representational capacity is
sively escalates from 32 to 256.
enriched by introducing multiple additional channels into the
TL’s core role is to gain information from the trained
advanced deep Boosted-BR-STM through ensemble learn-
source-domain and solve issues in the target-domain while
ing. The Squeezing–Boosting (SB) ideas employed in the
pursuing a high level of performance. Block A also uses
proposed deep CNNs, initially squeezed channels to get
region smoothing techniques to minimize the distortion and
salient and informative feature space. Then, combined with
outlier acquired while capturing the input images [50]. The
each STM block at abstract, mid, and high levels to boost and
boosted channel is handled in block F to lessen connection
achieve a diverse feature map using TL-based auxiliary chan-
intensity and obtain ideal attributes.
nels. Concatenating the prominent and notable information
from various deep CNNs using distinct channels enhances the
xBoosted  b(xB ||xC ||xD ||xE ), (4) malaria infection depiction. Moreover, feature-map enhance-
ment is achieved by incorporating auxiliary channels from
pre-trained residual models. These supplementary channels

A 
B
xD B F  va xBoosted , (5) are subsequently employed in conjunction with the proposed
a b CNN at the final stage. The SB-based deep CNN effec-
tively learns intricate local and global patterns, enabling the
exi discrimination of textural variations between parasitic and
σ(x)  c xc
. (6)
i1 e healthy samples.

The feature-maps of blocks D and E are denoted in Eq. (4)

by the variables x D and x E , respectively. Likewise, blocks Ensemble learning
B and C are based on additional channels created using TL
and are shown as x B and xC , respectively. b(.) represents the We employ deep boosted feature-maps and classifier ensem-
boosting operation b(.). Additionally, the developed CNN ble strategies in the proposed hybrid learning. The essence
employs fully connected layers with dropout layers to collect of this boosted hybrid learning lies in enhancing the feature

123
Complex & Intelligent Systems

space used for combining ML classifiers. Ensemble clas- Table 2 Benchmarked NIH malaria detail
sifiers, guided by a voting strategy, amalgamate decisions
Characteristics Overview
from various classifiers. This hybrid learning strategy har-
nesses the strengths of the developed deep Boosted-BR-STM Total 27.5 k Samples
and the discriminating power of ML classifiers by isolating Non-parasite (healthy/impurities) 13.8 k Samples
key characteristics. We extract deep feature spaces from the
Parasite 13.8 k Samples
final layers of boosted deep CNNs and feed them to com-
Train, validation (70%) (9.6 k, 9.6 k)
petitive ML classifiers, including SVM [48], MLP [51], and
Test (30%) (4.1 k, 4.1 k)
AdaBoostM1 [52]. The activation functions are represented
in f SV M (.), f M L P (.), f Adaboost (.) as shown in Eqs. (7–10). Input image 164 × 164 × 3
The proposed DBEL framework extracts deep boosted Improved image size 82 × 82 × 1
channels from the proposed Boosted-BR-STM to achieve
diverse channels and provides to ensemble classifier. The
ensemble method combines the outputs of individual models, Utilization of customized CNNs
leveraging their diverse strengths. The ensemble aggre-
gates predictions from multiple models, reducing the risk Several current CNNs like VGG, ResNet, GoogleNet,
of overfitting and enhancing the generalization ability of the DenseNet, and Inception have been adapted to classify par-
proposed framework. Moreover, the DBEL benefits from the asite malaria images for comparative analysis [23, 25, 32,
generalized detection model and provides deep rich infor- 35, 37, 38]. The existing CNN’s abstract and final classifica-
mation feature space and ensemble learning [53]. The deep tion layers are customized using additional layers according
feature maps are generated, the 2nd last fully connected to the input and target-class dimension of the dataset. CNN
(FC) layer of developed Boosted-BR-STM and customized models were primarily backpropagation strategies learned
CNNs and fed to the ML classifier. Ultimately, integrating from scratch where initial weights were randomly selected.
deep boosting and ensemble classifiers improves the DBEL The convolutional layers’ initial weights were borrowed from
framework generalization ability. In Eq. (10), f Ensemble (.) an pre-train models using TL to improve the model conver-
ensemble of ML classifiers makes the final decision from the gence. In this regard, we used TL to adopt effective model
boosted feature-maps. parameters derived from the modified prior CNNs designed
to precisely capture the target-domain specific parasite char-
y M L P  f M L P (x D B F ), (7) acteristics mostly on the malaria dataset employing improved
filter weights acquired from ImageNet [55].

y SV M  f SV M (x D B F ), (8)
Experimental configuration
yAdaboost  f Adaboost (x D B F ), (9)
Dataset
yFinal  f Ensemble ( f M L P (x D B F ) ,
The NIH dataset separated parasitic cells from a thin blood
f SV M (x D B F ) , f Adaboost (x D B F )). (10) smear slide for research on monitoring and diagnosis [25,
56]. The dataset includes samples of falciparum patients
Significance of hybrid learning collected from the Mahidol Oxford Tropical Medicine Cen-
The training of the proposed deep Boosted-BR-STM and cus- ter and Bangladesh’s Chittagong Medical College [57]. The
tomized CNNs may sometimes cause overfitting. Therefore, institutes mentioned above are proficient slide readers who
the proposed hybrid framework learns the effective discrim- differentiate between parasitic images and healthy individ-
ination features and improves generalization. Additionally, uals. Plasmodium is present mainly in positive (parasitic)
SVM [48], MLP [51], and AdaBoostM1 [52] represent three specimens, while the artifact effect is seen in negative (non-
distinct ML classifiers that minimize the structural risk. Deep parasitic or healthy) samples with staining and contaminants.
CNNs cause empirical risk by reducing training error using The malaria dataset distribution for experimental setup is
the optimal hyper-parameter selection [48]. The proposed shown in Table 2 and Fig. 4, pictorially depicting the para-
framework significantly reduces both training and test errors site and non-parasitic or healthy RBC samples.
and thus improves the generalization. Moreover, ensemble
learning aims to increase performance and promotes integrat-
ing numerous feature spaces into a single rich information
feature vector [54].

123
 Complex & Intelligent Systems

Precison
Classified Parasitic
 ,
a Classified Parasitic + Incorrectly Classified Parasitic
(12)
c
Classified Parasitic
b Sensitivity  , (13)
Total Parasitic Samples
Classified Healty
Specificity  , (14)
Total Healty Individuals
Fig. 4 Panel (a, b) DWT enhanced parasitic and normal samples,
respectively (Sen × Pre)
F − Score  2 , (15)
Sen + Pre

Table 3 Hold-out based optimal hyper-parameter selection

Hyper-parameter Values
Results and discussion
Learning-rate (α) 10–3
This study uses the original and DWT-improved datasets
Optimizer SGD
to establish a novel detection DBEL framework system
Epoch 10
for malaria and distinguish parasite malaria sufferers from
Momentum 0.90
healthy individuals. Moreover, a novel Boosted-BR-STM
Loss Cross entropy is developed for testing the usefulness of boosting and
boundary-region based STM to forecast affected parasite
cells in RBC-thin-smears microscopy samples, and its effec-
Model training tiveness is contrasted with modified existing CNNs. Further-
more, the existing CNNs are upgraded, deployed in the TL,
The training and testing parts of the dataset are divided into and trained from scratch. Table 4 assesses the developed
a 70:30% ratio. Additionally, cross-validation is executed in malaria detection employing defined performance metrics.
the training phase, which is partitioned 80:20 into training
and validation sets during model training. The training set Performance analysis of detection results
includes the data validation set when optimum parameters are
picked via hold-out cross-validation. Table 3 lists the selected Enhanced dataset evaluation
optimal parameters in detail. The MATLAB-2022a tool was
used to create the modified CNNs. The tests were carried out Impurities, staining, and noise artifacts in the original malaria
on an NVIDIA-GeForce GTX-T-Dell PC with 32 GB RAM dataset generated a striking likeness between the normal
and CUDA support. CNNs required nearly 12–24 h during and parasite samples. In this regard, an enhancement pro-
training or 1–2 h for each epoch. cess is essential to remove stains/impurities and impulsive
noise from blood smear images. In this case, DWT reduces
noise effects, stain impurity, and computing time, attaining
Performance evaluation streamlined, improved feature channels. Moreover, DWT
coefficients preserve important low-resolution and diag-
Standard performance criteria are used to evaluate the effi- onal features that aid in differentiating malaria-infected
ciency of the detection CNNs and the developed hybrid learn- RBCs from healthy samples. Employing CNN techniques
ing framework. The detection measures, including Accuracy, on improved DWT image outperformed the standard dataset
Precision, Sensitivity, Specificity, and F-score, are used as in terms of Accuracy (1.86–3.1%), F-Score (1.8–2.9%),
optimization metrics for evaluating the technique’s perfor- Sensitivity (0.60–2.2%), Precision (2.40–5.50%), Specificity
mance. These measurements, along with a mathematical (2.6–6.1%), as shown in Table 4 and Fig. 5. The minimum,
explanation and abbreviation, are explained in Eqs. 11. average, and maximum performance for each parameter
is demonstrated in Fig. 5. Moreover, DWT is frequently
employed as an enhancement approach for data cleaning,
Classified Correctly significant feature-map generation, and improved classifica-
Accuracy  × 100, (11) tion framework performance.
Total Samples

123
Complex & Intelligent Systems

Table 4 Performance of
Boosted-BR-STM and current Models Trained-scratch (original-data)
CNNs on unseen dataset
Accuracy% F-score Sensitivity Precision Specificity

ShuffleNet 92.85 0.931 0.965 0.899 0.892

ResNet-50 94.96 0.951 0.980 0.924 0.919
SqueezeNet 95.16 0.953 0.976 0.931 0.928
Inceptionv3 95.40 0.955 0.980 0.931 0.928
ResNet-18 95.66 0.958 0.986 0.931 0.927
Xception 95.22 0.953 0.969 0.938 0.936
DenseNet-201 95.73 0.958 0.978 0.939 0.936
VGG-19 95.74 0.958 0.966 0.950 0.949
VGG-16 95.79 0.958 0.965 0.951 0.95
GoogleNet 95.23 0.953 0.960 0.946 0.945
Proposed boosted-BR-STM 96.67 0.967 0.986 0.949 0.901

Table 5 Performance comparison

of Boosted-BR-STM and current Models Trained-scratch on enhanced data (DWT)
CNNs on unseen dataset
Accuracy% F-score Sensitivity Precision Specificity

ShuffleNet 94.71 0.949 0.976 0.923 0.918

SqueezeNet 95.34 0.954 0.969 0.940 0.938
Inceptionv3 95.43 0.955 0.966 0.944 0.942
ResNet-50 95.55 0.956 0.969 0.944 0.942
Xception 95.67 0.957 0.980 0.936 0.933
ResNet-18 95.71 0.958 0.969 0.947 0.945
VGG-16 95.80 0.958 0.971 0.946 0.945
DenseNet-201 95.84 0.959 0.973 0.945 0.944
GoogleNet 95.85 0.959 0.982 0.938 0.935
VGG-19 95.95 0.960 0.966 0.954 0.953
Proposed Boosted-BR-STM 97.51 0.975 0.982 0.968 0.968

The proposed Boosted-BR-STM by decreasing false positives, which reduces the med-
ical staff’s workload significantly. The tables demon-
The developed Boosted-BR-STM improved the generaliz- strate that the Boosted-BR-STM fared better than reported
ability of the proposed DBEL scheme compared to early approaches using the NIH dataset. The Boosted-BR-STM
techniques and yielded better results on a DWT-enhanced test performs better than other methods; Accuracy values range
dataset. Table 5 and Fig. 5 show a significant enhancement from (1.16–7.44%), F-score (1.2–7.5%), and Sensitivity
in the developed Boosted-BR-STM to forecast plasmodium (2.5–4.8%) (Fig. 5). The developed Boosted-BR-STM lower
falciparum-infected patients. The Boosted-BR-STM exploits False Negative (FN  71) and improved sensitivity relative
homogenous and boundary-driven parasite patterns in the to the finest, most reputable TL-based DenseNet CNN (FN
STM block, boosting through STM and residual learning,  131) on unseen DWT enhanced dataset, as demonstrated
and TL to improve the Sensitivity and F1-score. Region in Tables 5 and 6.
homogenous and boundary characteristics aid in learning the
distorted parasitic samples. Moreover, TL-based generated
The proposed hybrid learning
feature maps and SB can capture subtle contrast and texture
variation across artifact and parasite samples.
The malaria parasite samples are diagnosed using a new
The data augmentation strategies employed in the training
deep-boosted feature-map and ensemble learning (DBEL).
portion yielded performance improvement and model gen-
In this regard, three competing ML classifiers are ensem-
eralization. Moreover, Accuracy and Precision are attained
bled with feature vectors of the proposed Boosted-BR-STM

123
 Complex & Intelligent Systems

Fig. 5 The proposed Boosted-BR-STM CNN and DBEL framework performance gain over the existing detection. Moreover, an enhanced DWT
samples considerably improved performance

that contribute as feature extractors. Moreover, the malaria proposed DBEL outperformed the customized CNNs by
parasitic images are detected by extracting the deep features providing a deep-boosted feature map to the majority voting-
from existing CNNs and providing them to ML classifiers based ensembled classifiers. Combining boosted deep feature
in the DHML scheme. The significance of exploiting deep maps and utilizing three classifiers creates a hybrid of diverse
features is identified and compared with Softmax-based feature spaces and ensemble learning. Boosting learning aims
evaluation. The DHML using TL-based fine-tuned exist- to increase performance and promotes integrating numerous
ing CNNs scheme outperforms Softmax-based evaluation in feature spaces into a single rich information feature vector.
terms of Accuracy (0.47–1.7%), F-Score (0.50–1.6%), Pre- Moreover, the ensemble classifiers’ improve the proposed
cision (0.50–3.6%), Specificity (0.5–4%), as illustrated in framework’s differentiation capacity by merging deep fea-
Table 7. ture maps to construct the boosted and diverse feature space.
The proposed DBEL framework The individual models within the ensemble like SVM,
MLP, and AdaBoostM1 are chosen because they have distinct
The hybrid learning approach evaluates the performance characteristics and strengths. SVM learns optimal decision
of deep-boosted feature maps and ensemble learning. The boundaries, MLP captures complex non-linear relationships,

123
Complex & Intelligent Systems

Table 6 TL-based existing CNNs

performance analysis Models TL-based CNNs

Accuracy% F-score Sensitivity Precision Specificity

SqueezeNet 95.08 0.952 0.978 0.927 0.923

ShuffleNet 95.71 0.957 0.971 0.944 0.943
VGG-19 95.79 0.958 0.969 0.948 0.947
Inceptionv3 95.91 0.959 0.971 0.948 0.947
GoogleNet 96.03 0.962 0.983 0.941 0.938
Xception 96.07 0.961 0.971 0.951 0.95
ResNet-50 96.17 0.962 0.974 0.951 0.949
VGG-16 96.2 0.962 0.975 0.95 0.949
ResNet-18 96.21 0.963 0.985 0.942 0.94
DenseNet-201 96.31 0.963 0.968 0.958 0.958

Table 7 Performance of
TL-based current CNNs deep Models TL-based DHML scheme
feature and ML
Accuracy% F-score Sensitivity Precision Specificity

SqueezeNet 95.96 0.961 0.985 0.938 0.934

ShuffleNet 96.26 0.963 0.969 0.957 0.956
VGG-19 96.34 0.964 0.974 0.954 0.953
Inceptionv3 96.48 0.965 0.975 0.956 0.955
GoogleNet 96.57 0.966 0.975 0.957 0.956
Xception 96.65 0.967 0.980 0.955 0.953
ResNet-50 96.65 0.967 0.983 0.952 0.950
VGG-16 96.69 0.967 0.971 0.963 0.963
ResNet-18 96.70 0.968 0.982 0.954 0.952
DenseNet-201 96.78 0.968 0.978 0.958 0.957
Comparative analysis with the existing work on NIH dataset
Dense-Net [23] 90.54 0.9050 0.9400 0.9080 0.8770
Inception [23] 93.06 0.9306 0.9300 0.9306 0.9310
Xception [23] 94.94 0.9490 0.9260 0.9510 0.9750
Customized ResNet-50 [38] 95.40 – – – –
Seq-CNN [25] 95.90 0.9590 0.9470 – 0.9720
VGG-16 [37] 95.96 0.9560 – 0.9680 –
Various malaria dataset
Modified YOLOv3 [40] 95.46 – – – –
Modified YOLOv4 [40] 96.14 – – – –
YOLOv5 [41] 95.20 – – – –
DarkNet-53 [41] 96.02 – – – –

and AdaBoostM1 enhances the classification of challeng- Specificity (1.5–8.6%), as shown in Table 8 and Fig. 5.
ing instances. The selection of these classifiers is based Finally, the proposed DBEL framework further reduced
on their inherent ability to minimize structural risk, and the FN (34), as contrasted to the best and most reputable
this contributes to improved performance. The proposed DenseNet-201. However, a few samples were missed due
DBEL framework outperformed the existing techniques for to a similarity between malaria-infected and healthy people
classifying malaria parasite samples in terms of Accuracy owing to impurity, stained, or noise anomalies in non-
(1.72–5.65%), F-Score (1.70–5.4%), Precision (1.60–8%), parasitic instances, as seen in Fig. 6. In the summarized

123
 Complex & Intelligent Systems

Table 8 Performance analysis of the proposed frameworks

Framework The proposed DBEL framework work

Accuracy% F-score Sensitivity Precision Specificity

Deep boosted machine learning

SVM 97.73 0.977 0.985 0.970 0.969
AdaboostM1 97.92 0.979 0.988 0.971 0.971
MLP 98.08 0.981 0.988 0.974 0.973
Deep boosted ensemble learning
SVM + AdaboostM1 98.21 0.982 0.991 0.974 0.973
AdaboostM1 + MLP 98.26 0.982 0.99 0.975 0.975
MLP + SVM 98.34 0.983 0.991 0.976 0.976
Proposed DBEL (SVM + AdaboostM1 + MLP) 98.50 0.985 0.992 0.979 0.978

Customized CNNs

a TL is an essential DL technique to perform satisfactorily

instead of training CNN from scratch. TL adopted custom
CNNs utilizing DWT-enriched data outperformed learning
from scratch, as shown in Table 6. The TL-based fine-
b tuned existing CNNs perform better than those learned
from scratch and existing techniques: accuracy (0.36–1.6%),
F1-Score (0.30–1.4%), sensitivity (0.30–1.9%), precision
(0.40–3.5%), and specificity (0.5–4%). Tables 5, 6 and 8
Fig. 6 The enhanced stain artifacts and parasitic samples are shown in show the comparative study with customized deep-CNN
panels (a) and (b). The parasite images have a high degree of resem-
blance to artifacts, resulting in a high rate of miss detection
models, along with Fig. 5. TL enhanced the CNN learn-
ing convergence through pattern initialization thus, lowering
the miss-classification rate while maintaining precision. This
way, we have introduced several innovative concepts to demonstrates TL’s ability to analyze medical image fields,
systematically enhance the framework’s performance and especially malaria detection.
generalization.
The proposed Deep Boosted-BR-STM CNN novel con- Feature space visualization
cepts
The distinct class samples enhanced the ability of model
• Boundary and regional feature extraction and STM Block learning and robustness. Therefore, the principal components
integration analysis (PCA) aims to depict the significant discriminative
• Optimal feature map extraction and concatenation of features. The top-3 (PC1 vs. PC2 and PC1 vs. PC3) dis-
diverse maps to achieve boosted channels through TL- tinct components of the test samples are displayed in Fig. 8,
based auxiliary channel generation with distinctive coloring for every feature point to distin-
• Channel boosting through TL-based auxiliary channel guish parasite-containing samples from non-parasitic ones
generation [58, 59]. The feature space is examined to better comprehend
the proposed Boosted-STM decision-making BR’s behavior
in contrast to the existing DenseNet-201 model.
Deep Hybrid and Ensemble Learning
The developed Boosted-BR-STM and DBEL-generated
feature space depict the discrimination among parasite and
• Extracting deep features from the proposed Boosted-BR- healthy samples, as illustrated in Fig. 7. The Boosted-BR-
STM CNN and utilizing ML Classifiers STM has the highest evaluating effectiveness, according to
• Exploiting proposed deep features and Ensemble of ML the 2D PC plots. The current CNN DenseNet-201 has dis-
Classifiers for the final malaria parasite detection. covered that feature spaces are cluttered and inaccurately
unable to distinguish between two classes using the original

123
Complex & Intelligent Systems

Fig. 7 The visualization for the generated of PC1, PC2, and PC3 distinct features for the proposed DBEL, Boosted-BR-STM, and DenseNet on
DWT-enriched datasets

123
 Complex & Intelligent Systems

Fig. 7 continued

data. Additionally, compared to DenseNet-201 trained from are drawn for the developed DBEL framework, Boosted-BR-
scratch using improved DWT data, TL-based DenseNet-201 STM, and customized CNNs on test data. These graphs depict
better indicates the instance class. the capability of the classifier to differentiate across various
potential ranges. The detection cut-off of the best positive
class classifier can be considerably accessed via ROC and
Graphical analysis PR curves; results are shown in Fig. 8 [60]. The Boosted-
BR-STM reduced false-negative or miss-classified parasitic
The Sensitivity and Precision metrics for assessing the mod- compared to existing CNNs on an enhanced dataset with a PR
els are critical for malaria parasite detection. Therefore, Pre-
cision–Recall (PR) and Receiver–Operating Curve (ROC)

123
Complex & Intelligent Systems

Fig. 8 The developed Boosted-BR-STM with TL-based fine-tuned CNN’s ROC and PR curves on original and DWT enriched data

AUC  99.20% [61]. Moreover, the developed DBEL frame- that our system’s testing time for each parasitic sample aver-
work maintains minimal false-positives by detecting malarial aged 5–7 s. In the future, the proposed DBEL is intended
infections with strong sensitivity due to its high ROC-AUC to help healthcare practitioners by developing an application
value. to automatically screen, identify, and stage parasitic malaria
patients for clinical examination, as well as impurity, stained,
or noise anomalies in non-parasitic instances. Moreover, the
Conclusion developed framework can be employed in diagnosing mon-
keypox, brain tumors, lung cancer, and breast cancer using
Early detection of malaria can be treated properly and medical images.
quickly enough to prevent irregular disability. The proposed
Acknowledgements The authors extend their appreciation to the
hybrid DBEL framework stack deep boosted features of the Princess Nourah bint Abdulrahman University Researchers Supporting
developed residual Boosted-BR-STM CNN and ensemble Project number (PNURSP2024R513), Princess Nourah bint Abdulrah-
learning to detect RBC thin smears microscope images of man University, Riyadh, Saudi Arabia. We thank the Department of
Computer Systems Engineering, University of Engineering and Applied
malaria patients. The proposed DBEL framework benefits
Sciences (UEAS), Swat, Pakistan for providing the necessary facilities
from DWT enhancements, data augmentation, incorporating to carry out this research work.
TL, inherent discriminative deep features from the devel-
oped Boosted-BR-STM, and ensemble learning to improve Data availability The standard benchmark NIH Malaria dataset has
been collected from the standard medical centers and made available in
detection performance. Moreover, the developed Boosted- the standard open-access Kaggle and other repositories. https://www.
BR-STM employs TL for the diverse rich information kaggle.com/datasets/iarunava/cell-images-for-detecting-malaria.
feature-maps generation, residual learning, and SB ideas in
the STM block, enhancing the ability to learn homogeneity, Declarations
and parasitic structural patterns. Moreover, residual learning
systematically extracts features, starting with basic image- Conflict of interest The authors declare that they have no known com-
peting financial interests or personal relationships that could have
level features to more intricate texture-based differences.
appeared to influence the work reported in this paper.
This novel approach allows for the acquisition of boosted
features at varying levels of granularity. The proposed frame- Informed consent Not applicable.
work achieved an Accuracy of 98.50%, an AUC of 0.996, an
F-score of 0.985, and a Sensitivity of 0.992. Furthermore, the Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing, adap-
proposed Boosted-BR-STM CNN and framework excelled tation, distribution and reproduction in any medium or format, as
in existing techniques, experimentally in accurately detect- long as you give appropriate credit to the original author(s) and the
ing malaria-infected patients. We have assessed our proposed source, provide a link to the Creative Commons licence, and indi-
framework’s runtime compared to manual analysis, gauging cate if changes were made. The images or other third party material
in this article are included in the article’s Creative Commons licence,
its efficiency and time-saving potential. It is worth noting unless indicated otherwise in a credit line to the material. If material

123
 Complex & Intelligent Systems

is not included in the article’s Creative Commons licence and your 15. Krishnadas P, Sampathila N (2021) Automated Detection of
intended use is not permitted by statutory regulation or exceeds the malaria implemented by deep learning in Pytorch. 2021 IEEE
permitted use, you will need to obtain permission directly from the copy- Int. Conf. Electron. Comput. Commun. Technol., IEEE; pp
right holder. To view a copy of this licence, visit http://creativecomm 01–5.https://doi.org/10.1109/CONECCT52877.2021.9622608
ons.org/licenses/by/4.0/. 16. Kalkan SC, Sahingoz OK (2019) Deep learning based classifica-
tion of malaria from slide images. 2019 Sci Meet Electr Biomed
Eng Comput Sci EBBT 2019. https://doi.org/10.1109/EBBT.2019.
8741702
References 17. Baroni L, Salles R, Salles S, Guedes G, Porto F, Bezerra E et al
(2020) An analysis of malaria in the Brazilian Legal Amazon using
1. Keleta Y, Ramelow J, Cui L, Li J (2021) Molecular interactions divergent association rules. J Biomed Inform 108:103512. https://
between parasite and mosquito during midgut invasion as targets to doi.org/10.1016/j.jbi.2020.103512
block malaria transmission. Npj Vaccines. https://doi.org/10.1038/ 18. Asam M, Khan SH, Akbar A, Bibi S, Jamal T, Khan A et al (2022)
s41541-021-00401-9 IoT malware detection architecture using a novel channel boosted
2. Gupta S, Gazendam N, Farina JM, Saldarriaga C, Mendoza I, and squeezed CNN. Sci Rep 12:15498. https://doi.org/10.1038/
López-Santi R et al (2021) Malaria and the heart: JACC state-of- s41598-022-18936-9
the-art review. J Am Coll Cardiol 77:1110–1121. https://doi.org/ 19. Zahoora U, Khan A, Rajarajan M, Khan SH, Asam M, Jamal T
10.1016/j.jacc.2020.12.042 (2022) Ransomware detection using deep learning based unsu-
3. Yimam Y, Nateghpour M, Mohebali M, Afshar MJA (2021) A pervised feature extraction and a cost sensitive Pareto Ensemble
systematic review and meta-analysis of asymptomatic malaria classifier. Sci Rep 12:15647. https://doi.org/10.1038/s41598-022-
infection in pregnant women in Sub-Saharan Africa: a challenge for 19443-7
malaria elimination efforts. PLoS ONE. https://doi.org/10.1371/ 20. Khan A, Khan SH, Saif M, Batool A, Sohail A, Khan MW (2022)
journal.pone.0248245 A survey of deep learning techniques for the analysis of COVID-19
4. Tegegne Y, Worede A, Derso A, Ambachew S (2021) The preva- and their usability for detecting omicron
lence of malaria among children in Ethiopia: a systematic review 21. Du X, Wang X, Xu F, Zhang J, Huo Y, Ni G et al (2022) Morpholog-
and meta-analysis. J Parasitol Res. https://doi.org/10.1155/2021/ ical components detection for super-depth-of-field bio-micrograph
6697294 based on deep learning. Microscopy 71:50–59. https://doi.org/10.
5. World Health Organization (WHO). WHO Malaria Policy Advi- 1093/jmicro/dfab033
sory Group ( MPAG ) meeting 2021:13–4. 22. Zafar MM, Rauf Z, Sohail A, Khan AR, Obaidullah M, Khan SH
6. Abbas N, Saba T, Rehman A, Mehmood Z, Javaid N, Tahir M et al et al (2022) Detection of tumour infiltrating lymphocytes in CD3
(2019) Plasmodium species aware based quantification of malaria and CD8 stained histopathological images using a two-phase deep
parasitemia in light microscopy thin blood smear. Microsc Res CNN. Photodiagnosis Photodyn Ther 37:102676. https://doi.org/
Tech 82:1198–1214. https://doi.org/10.1002/jemt.23269 10.1016/j.pdpdt.2021.102676
7. Yoon J, Jang WS, Nam J, Mihn DC, Lim CS (2021) 23. Maqsood A, Farid MS, Khan MH, Grzegorzek M (2021) Deep
An automated microscopic malaria parasite detection system malaria parasite detection in thin blood smear microscopic images.
using digital image analysis. Diagnostics. https://doi.org/10.3390/ Appl Sci 11:1–19. https://doi.org/10.3390/app11052284
diagnostics11030527 24. Lin M, Huang C, Chen R, Fujita H, Wang X (2021) Direc-
8. Barber BE, William T, Grigg MJ, Yeo TW, Anstey NM (2013) tional correlation coefficient measures for Pythagorean fuzzy sets:
Limitations of microscopy to differentiate Plasmodium species in a their applications to medical diagnosis and cluster analysis. Com-
region co-endemic for Plasmodium falciparum, Plasmodium vivax plex Intell Syst 7:1025–1043. https://doi.org/10.1007/s40747-020-
and Plasmodium knowlesi. Malar J 12:8. https://doi.org/10.1186/ 00261-1
1475-2875-12-8 25. Rajaraman S, Antani SK, Poostchi M, Silamut K, Hossain MA,
9. Mukry SN, Saud M, Sufaida G, Shaikh K, Naz A, Shamsi TS Maude RJ et al (2018) Pre-trained convolutional neural networks
(2017) Laboratory diagnosis of malaria: Comparison of manual as feature extractors toward improved malaria parasite detection in
and automated diagnostic tests. Can J Infect Dis Med Microbiol. thin blood smear images. PeerJ. https://doi.org/10.7717/peerj.4568
https://doi.org/10.1155/2017/9286392 26. Das DK, Ghosh M, Pal M, Maiti AK, Chakraborty C (2013)
10. Maity M, Gantait K, Mukherjee A, Chatterjee J. Visible spectrum- Machine learning approach for automated screening of malaria par-
based classification of malaria blood samples on handheld spec- asite using light microscopic images. Micron 45:97–106. https://
trometer. I2MTC 2019 - 2019 IEEE Int Instrum Meas Technol doi.org/10.1016/j.micron.2012.11.002
Conf Proc 2019. https://doi.org/10.1109/I2MTC.2019.8826860. 27. Sarkar RP, Maiti A (2019) Investigation of dataset from dia-
11. Microwave A, Engineering E. Software / Diagnostic Manual betic retinopathy through discernibility-based k-NN algorithm.
n.d.:1–8. Adv Intell Syst Comput 812:93–100. https://doi.org/10.1007/978-
12. Somasekar J, Sharma A, Madhusudhana Reddy N, Padmanabha 981-13-1540-4_10
Reddy YCA (2020) Image analysis for automatic enumeration of 28. Colwell T, WBMCS (2016) Automated detection of P. falci-
rbc infected with plasmodium parasites-implications for malaria parum using machine learning algorithms with quantitative phase
diagnosis. Adv Math Sci J 9:1229–1237. https://doi.org/10.37418/ images of unstained cells. PLoS One 11:e0163045. https://doi.org/
amsj.9.3.48 10.1371/journal.pone.0163045
13. Molina A, Rodellar J, Boldú L, Acevedo A, Alférez S, Merino A 29. Go T, Kim JH, Byeon H, Lee SJ (2018) Machine learning-based
(2021) Automatic identification of malaria and other red blood cell in-line holographic sensing of unstained malaria-infected red blood
inclusions using convolutional neural networks. Comput Biol Med cells. J Biophotonics 11:e201800101
136:104680. https://doi.org/10.1016/j.compbiomed.2021.104680 30. Mehanian C, Jaiswal M, Delahunt C, Thompson C, Horning M, Hu
14. Leckenby J, Li H, Negus K, Pickering M, Adorno T, Horkheimer L et al (2017) Computer-Automated Malaria Diagnosis and Quanti-
M et al (2009) A semi-automatic method for quantification and tation Using Convolutional Neural Networks. Proc - 2017 IEEE Int
classification of erythrocytes infected with malaria parasites in Conf Comput Vis Work ICCVW 2017;2018-Janua:116–25. https://
microscopic images. J Biomed Inform 42:296–307. https://doi.org/ doi.org/10.1109/ICCVW.2017.22
10.1016/j.jbi.2008.11.005

123
Complex & Intelligent Systems

31. Bibin D, Nair MS, Punitha P (2017) Malaria parasite detec- 48. Khan SH, Sohail A, Khan A, Hassan M, Lee YS, Alam J et al
tion from peripheral blood smear images using deep belief (2021) COVID-19 detection in chest X-ray images using deep
networks. IEEE Access 5:9099–9108. https://doi.org/10.1109/ boosted hybrid learning. Comput Biol Med. https://doi.org/10.
ACCESS.2017.2705642 1016/j.compbiomed.2021.104816
32. Var E, Boray Tek F (2018) Malaria parasite detection with deep 49. Aziz A, Sohail A, Fahad L, Burhan M, Wahab N, Khan A (2020)
transfer learning. UBMK 2018 - 3rd Int Conf Comput Sci Eng pp Channel boosted convolutional neural network for classification of
298–302. https://doi.org/10.1109/UBMK.2018.8566549 mitotic nuclei using histopathological images. In: Proc. 2020 17th
33. Dong Y, Jiang Z, Shen H, David Pan W, Williams LA, Reddy Int. Bhurban Conf. Appl. Sci. Technol. IBCAST 2020.https://doi.
VVB et al (2017) Evaluations of deep convolutional neural net- org/10.1109/IBCAST47879.2020.9044583
works for automatic identification of malaria infected cells. 2017 50. Khan SH (2022) COVID-19 detection and analysis from lung CT
IEEE EMBS Int Conf Biomed Heal Informatics, BHI 2017, pp images using novel channel boosted CNNs. 2209.10963
101–104https://doi.org/10.1109/BHI.2017.7897215 51. Gardner M, Dorling S (1998) Artificial neural networks (the mul-
34. Lenet-5, convolutional neural networks 2015. tilayer perceptron)—a review of applications in the atmospheric
35. Szegedy C, Wei Liu, Yangqing Jia, Sermanet P, Reed S, Anguelov sciences. Atmos Environ 32:2627–2636. https://doi.org/10.1016/
D et al (2015) Going deeper with convolutions. 2015 IEEE S1352-2310(97)00447-0
Conf. Comput. Vis. Pattern Recognit., vol. 07-12- June, IEEE; pp 52. Cao Y, Miao Q-G, Liu J-C, Gao L (2013) Advance and prospects
1–9.https://doi.org/10.1109/CVPR.2015.7298594 of AdaBoost algorithm. Acta Autom Sin 39:745–758. https://doi.
36. Hung J, Goodman A, Lopes S, Carpenter A (2013) Applying faster org/10.1016/S1874-1029(13)60052-X
R-CNN for object detection on malaria images. J R Stat Soc Ser A 53. Zahoor MM, Qureshi SA, Bibi S, Khan SH, Khan A, Ghafoor U
Stat Soc 175:417–433 et al (2022) A new deep hybrid boosted and ensemble learning-
37. Huq A, Pervin MT (2020) Robust deep neural network model for based brain tumor analysis using MRI. Sensors 22:2726. https://
identification of malaria parasites in cell images. In: 2020 IEEE Reg doi.org/10.3390/s22072726
10 Symp TENSYMP 2020:1456–1459. https://doi.org/10.1109/ 54. Ganaie MA, Hu M, Tanveer M, Suganthan PN (2021) Ensemble
TENSYMP50017.2020.9230832. deep learning: a review.
38. Reddy ASB, Juliet DS (2019) Transfer Learning with ResNet-50 55. Khan SH, Khan A, Lee YS, Hassan M, Jeong WK (2022) Seg-
for Malaria Cell-Image Classification. 2019 Int. Conf. Commun. mentation of shoulder muscle MRI using a new Region and Edge
Signal Process., Boston, MA: IEEE, pp 0945–9. https://doi.org/10. based Deep Auto-Encoder. Multimed Tools Appl. https://doi.org/
1109/ICCSP.2019.8697909. 10.1007/s11042-022-14061-x
39. Maity M, Jaiswal A, Gantait K, Chatterjee J, Mukherjee A (2020) 56. ARUNAVA. Malaria Cell Images Dataset | Kaggle n.d. https://
Quantification of malaria parasitaemia using trainable seman- www.kaggle.com/datasets/iarunava/cell-images-for-detecting-
tic segmentation and capsnet. Pattern Recognit Lett 138:88–94. malaria (accessed Dec 20, 2022).
https://doi.org/10.1016/j.patrec.2020.07.002 57. Maude RJ, Hasan MU, Hossain MA, Sayeed AA, Kanti Paul S,
40. Abdurahman F, Fante KA, Aliy M (2021) Malaria parasite detec- Rahman W et al (2012) Temporal trends in severe malaria in Chit-
tion in thick blood smear microscopic images using modified tagong, Bangladesh. Malar J. https://doi.org/10.1186/1475-2875-
YOLOV3 and YOLOV4 models. BMC Bioinformatics 22:112. 11-323
https://doi.org/10.1186/s12859-021-04036-4 58. Lobo SA, Siswadi, Bakhtiar T (2015) Visualization of classified
41. Zedda L, Loddo A, Di Ruberto C (2022) A deep learning based data with kernel principal component analysis. Glob J Pure Appl
framework for malaria diagnosis on high variation data set. Ann Math 11:2347–2356. https://doi.org/10.31227/osf.io/cbfxu
Tour Res 3:358–370. https://doi.org/10.1007/978-3-031-06430- 59. Barshan E, Ghodsi A, Azimifar Z, Zolghadri JM (2011) Super-
2_30 vised principal component analysis: visualization, classification
42. Houwen B (2002) Blood film preparation and staining pro- and regression on subspaces and submanifolds. Pattern Recognit
cedures. Clin Lab Med 22:1–14. https://doi.org/10.1016/S0272- 44:1357–1371. https://doi.org/10.1016/j.patcog.2010.12.015
2712(03)00064-7 60. Hajian-Tilaki K (2013) Receiver operating characteristic (ROC)
43. Sakthidasan alias Sankaran K, Nagarajan V (2021) Noise removal curve analysis for medical diagnostic test evaluation. Casp J Intern
through the exploration of subjective and apparent denoised patches Med 4:627–635
using discrete wavelet transform. IETE J Res 67:843–852. https:// 61. Boyd K, Eng KH, Page CD (2013) Area under the precision-recall
doi.org/10.1080/03772063.2019.1569483 curve: Point estimates and confidence intervals. Lect Notes Comput
44. Pandit P, Anand A (2019) Diagnosis of malaria using wavelet Sci (Including Subser Lect Notes Artif Intell Lect Notes Bioinfor-
coefficients and dynamic time warping. Int J Appl Comput Math. matics). 8190 LNAI:451–66. https://doi.org/10.1007/978-3-642-
https://doi.org/10.1007/s40819-019-0614-2 40994-3_29
45. Shorten C, Khoshgoftaar TM (2019) A survey on image data aug-
mentation for deep learning. J Big Data. https://doi.org/10.1186/
s40537-019-0197-0
Publisher’s Note Springer Nature remains neutral with regard to juris-
46. Khan SH, Shah NS, Nuzhat R, Majid A, Alquhayz H, Khan A
dictional claims in published maps and institutional affiliations.
(2022) Malaria parasite classification framework using a novel
channel squeezed and boosted CNN. Microscopy. https://doi.org/
10.1093/jmicro/dfac027
47. Khan SH, Sohail A, Zafar MM, Khan A (2021) Coronavirus disease
analysis using chest X-ray images and a novel deep convolutional
neural network. Photodiagnosis Photodyn Ther 35:102473. https://
doi.org/10.1016/j.pdpdt.2021.102473

123