Received 6 November 2024, accepted 18 November 2024, date of publication 21 November 2024, date of current version 5 December 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3504266

Enhancing Medical Image Classification Through

PSO-Optimized Dual Deterministic Approach
and Robust Transfer Learning
ASIF RAZA 1,2 , SHAHRULNIZA BIN MUSA 1,3 , AHMAD SHAHRAFIDZ BIN KHALID1 ,
MUHAMMAD MANSOOR ALAM4 , MAZLIHAM MOHD SU’UD 4 , AND FOUZIA NOOR5
1 Universiti Kuala Lumpur—Malaysian Institute of Information Technology (UniKL MIIT), Kuala Lumpur 50250, Malaysia
2 Department of CS and IT, Sir Syed University of Engineering and Technology, Karachi 75300, Pakistan
3 UniKL–LR University Joint ICT Laboratory (KLR-JIL), Universiti Kuala Lumpur–La Rochelle University, 17000 La Rochelle, France
4 Faculty of Computing and Informatics, Multimedia University, Cyberjaya 63100, Malaysia
5 Department of Diagnostic and Radiology, Combined Military Hospital, Malir Cantonment, Karachi 75070, Pakistan.

Corresponding author: Shahrulniza Bin Musa ([email protected])

ABSTRACT Effective transfer learning, within medical image classification, is probably one of the most
critical areas of research due to the associated complexities with the nature of medical images. These involve
variations in the acquisition techniques of images, resolutions, modalities, and patient demographics. This
study aims to develop a robust transfer learning model by combining Mobile-Net with Particle Swarm
Optimization, called MOB-CFPSO that classifies efficiently variety of datasets across heterogenic dataset
modalities, including colored, black, and white datasets. A high-performing machine learning model needs
enough data to enable robust feature extraction that can identify patterns in each class, allowing the model
to differentiate between different classes accurately. The current statistical algorithm is less effective in
the case of colored and black image classification and they also produce errors in the classification of
findings of obvious disease in heterogenic image datasets e.g. color, Black, and white. There is a need
to design and implement a robust image classification model based on heterogenic datasets that will
reduce the error rate and optimize classification accuracy by using the Transfer Learning technique. Model
robustness in deep learning refers to the capacity to sustain performance and deliver accurate predictions,
even in the face of uncertainties, disturbances, or adversarial conditions. Two separate experiments are
conducted on a pre-trained Mobile-Net model along with Modified PSO with constriction factor for model
optimization, accuracy enhancement, and reduction of error rate for disease classification on two distinct
datasets. DR Dataset comprised 10 distinct, whereas the MRI brain tumor included 8 various classes. The
results showed that the proposed Model MOB-CFPSO exceeds the state-of-the-art algorithm in datasets,
achieving high accuracy, robustness, precision, recall, and F1-score, with acceptable and favorable validation
loss. A comprehensive MRI study showed that the MOB-CFPSO achieved 100% and 99.86% of training
and validation accuracy respectively, with 0.0071% and 0.0071% of validation and training loss. However,
Precision, Recall, and F1 scores were 100%, 100%, and 100%, respectively. DR Dataset achieved 95.09%
and 97.66% of training and validation accuracy respectively whereas validation loss was 0.1902%, and
training loss was 0.2469%. Precision, Recall, and F1 score were 98.3%, 92.4%, and 95.3% respectively.

INDEX TERMS PSO optimization, transfer learning, diabetic retinopathy, mobile net, robust model, medical
image classification.

I. INTRODUCTION
The associate editor coordinating the review of this manuscript and Transfer learning leverages prior knowledge to address
approving it for publication was Jeon Gwanggil . new, related challenges more effectively, using data from
2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
177144 For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 13, 2025
A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

adjacent fields. Computational intelligence has enhanced and classification from MRI scans remain a significant chal-
this process through network-based, Bayes, fuzzy, and lenge. CNN model is implemented along with a U-Net-based
intelligence-based transfer learning [1]. Diabetic retinopathy, model for segmentation for automatic brain image classifica-
a vision-threatening diabetes complication [2], is difficult tion into four categories. Authors [9] have implemented six
to diagnose early. The author proposed an ensemble vot- well-established datasets to test the classification model and
ing architecture to detect its severity, comparing it with train the segmentation model, allowing a direct comparison of
CNN models like InceptionV3, VGG16, DenseNet121, and how segmentation affects tumor classification in brain MRI
ResNet50. scans.
Deep learning techniques for DR grading often fail to The author introduced a deep neural network architec-
accurately detect all phases, with metrics like AUC ROC ture named NeuroNet19, built on the VGG19 framework.
and F1-score proving unreliable. A modified EfficientNetB3 It integrates a unique component called the Inverted Pyra-
model, using Quadratic Weighted Kappa (QWK) as the mid Pooling Module (iPPM), which captures multi-scale
primary metric [3], was developed to address this particu- feature maps to extract both local and global image con-
lar issue. AI [4] models are categorized into classification, texts. This approach enhances the feature maps generated
regression, clustering [5], and dimensionality reduction using by the VGG19 backbone, regardless of the tumor’s size or
PCA [6] as shown in Figure 1. location. To promote transparency and accountability in the
Author [7] proposed a Stacking-based technique that model, the author implemented Explainable AI techniques.
significantly improved medical image classification. Aug- Specifically, they used Local Interpretable Model-Agnostic
menting consistently enhanced models without over fitting, Explanations (LIME), which highlight the features or regions
applicable even in single model channels. Cross-validation the model focuses on when making predictions for individual
with bagging also improved performance but relied on images [10].
adequate feature representation. The study highlighted the The Prime objective of the research ensure accurate
effectiveness of arithmetic pooling functions like Mean or diagnosis with minimal errors by optimizing the deep learn-
Majority Voting and demonstrated that ensemble learning is ing model’s performance in Heterogeneous MRI and DR
powerful. Datasets. The key goals are to reduce velocity through a
constriction factor, minimize error rates, and improve clas-
sification accuracy for diagnosing classes within a dataset
comprised of MRI and DR data.
This paper shows its significance and contribution in the
following manners.
• In the first Integration deep learning model Mobile Net
combined and optimized by Meta heuristic PSO, aimed
to improve Model accuracy, reduced error rate and mis-
classification as well.
• Furthermore, PSO is mathematically modified by the
use of the Constriction factor (CF) which controls explo-
ration and exploitation in Search space. Constriction
factor (discriminant) based approach controlled by ∅,
This approach leads to individuals converging over
time. In contrast to other evolutionary computation tech-
niques, the constriction factor method guarantees con-
vergence of the search process through a mathematical
FIGURE 1. Machine learning paradigm [7].
model.
• The proposed research was conducted on the novel
Researchers have developed automated diagnosis tech- dataset with the Ethical approval of the Univer-
niques [8], requiring deep learning models trained on large sity of Kuala Lumpur, vide Ethical approval Numbe
annotated datasets. Transfer learning used pre-trained net- (UNIKL REC/2024/PG/APV/01).MRI contains 08 vari-
works like VGG-Net, and addressed the lack of annotated ous classes related to brain tumors and 10 classes can be
data. found in OCT (Optical Coherence Tomography) scans
Early detection of brain tumors is crucial for improv- usually called DR (diabetic retinopathy).
ing patient outcomes and treatment choices, while precise MOB-CFPSO hyper parameter settings are mentioned as
classification and segmentation of these tumors are key for follows in Table 1.
crafting personalized treatment plans. Although Magnetic
Resonance Imaging (MRI) is widely used for brain scans II. RELATED WORK
and AI-driven detection methods have advanced, creating This study [11] focuses on using CNN with PSO to effi-
a reliable, Robust, and efficient model for tumor detection ciently detect and classify diabetic retinopathy (DR) from

VOLUME 13, 2025 177145

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

TABLE 1. MOB-CFPSO hyper parameter. image processing faster author created a network by apply-
ing a filter called anisotropic diffusion moreover they have
applied techniques and thresholds to separate the cancer-
ous tissue. Furthermore, the model analyzed patterns and
histogram-oriented gradients within the cancerous regions.
To improve the effectiveness of these learned features author
used both error rate analysis and relief methods. Finally,
classified features that received the scores from both Particle
Swarm Optimization (PSO) and regression neural networks.
retinal images, addressing the time and resource challenges In years there has been an increase, in the number of deaths
of manual evaluation. The FE_mines [12] model improves driving extensive research on tumor detection. The use of
feature extraction in medical images without extra data, image processing and computer vision (CV) techniques has
boosting class discrimination by 1-13% over traditional meth- turned out to be the most popular technique for identifying
ods. A study [13] combining deep neural networks with tumors in MRI. Detection brings benefits such as improved
nature-inspired optimizations exceeded in accuracy by using speed and efficiency as compared to normal diagnosis. This
pre-trained models in DR diagnosis using the IDRiD retinal is especially important because brain tumors can have vary-
scan database. Methods like PSO, ACO, and BEE enhanced ing shapes making it challenging for doctors to locate them
accuracy, while KNN [14] classifiers and Tamura features, accurately. The author Incorporates data filtration techniques
optimized by PSO, improved DR recognition. PSO also opti- like noise reduction; image refining and skull removal before
mizes CNN [15] hyper parameters, enhancing accuracy and applying the PSO-based segmentation method for the clas-
reducing loss in brain disease detection, validated on ADNI, sification by CNN along with Alex Net and Inception V3.
Kaggle Alzheimer’s, and brain tumor datasets. Furthermore, the author extracted relevant features from the
Fuzzy logic methods effectively detect microaneurysms image and then combined these features with the final classi-
in fundus images. PB-PSO clustering was used on the fication result [22].
DIARETDB0 dataset, achieving 99.9% accuracy for early
DR detection [16]. A new feature descriptor [17] improved
age-invariant face ID accuracy on FGNET and MORPH A. CRITICAL ANALYSIS AND CHALLENGES IN LITERATURE
datasets. A four-phase model using FFT, 2D-SWT, a chaotic REVIEW
wrapper, and RNN-LSTMs handled diabetic retinopathy The literature review reveals and persists with some chal-
diagnosis efficiently. An optimized ML method using SURF, lenges within the field of computer vision, such as effectively
KAZE, and PSO enhanced brain tumor detection in MRI extracting hidden features from colored images in heteroge-
images. neous datasets. Optimization of a deep learning model by
The proposed [18] model has four phases: FFT-based setting hyper parameters and weights within fully connected
fragmentation to detect patterns, 2D-SWT for feature layers. Researchers have diligently addressed the complex-
extraction, a chaotic wrapper for attribute selection, and ity of age-related variations in facial recognition especially
RNN-LSTM for classification. This method efficiently when the images have incredibly low resolution. Network
addresses non-linearity in fundus images with low compu- generalization and scalability are also major challenges in
tational load while ensuring accurate diabetic retinopathy medical image classification, demanding improvements in
diagnosis. Another [19] study optimized brain tumor detec- accuracy and processing time To address these challenges,
tion in MRI images using SURF, KAZE, and k-means researchers have explored and applied numerous versions
clustering with PSO to reduce redundant features, improving of Particle Swarm Optimization (PSO), such as Relief PSO
accuracy. The hybrid feature vectors were tested with various and Chaotic Particle Swarm Optimization (CPSO), which are
ML classifiers. well-known for their enhanced ability to explore and their
The author [20] suggests the use of CNNs and PSO potential to achieve better results, especially in complicated
for diagnosing diabetic retinopathy. Various pre-processing problem areas.
schemes for result comparison in association with PSO that is Furthermore, researchers’ emphasis has mostly shifted
used to lower classification error rates with the trained CNN. towards CPSO since it can boost exploration and improve out-
The design of a CNN involves optimizing certain parameters comes in grid-based solutions while tackling complex issues.
such as the number of convolutional layers, filter character- Moreover, there is a significant focus on using CPSO to
istics including their size, pool size, learning algorithm, the improve problems in feature selection, which might result in
quantity of fully connected layers with their corresponding better solutions. However, considering this prevailing trend,
neurons, batch size, and training epochs. This contains feature this research deviates by focusing explicitly on the develop-
extraction from the retinal and the usage of histogram equal- ment of a dual deterministic model rooted in deep learning,
ization within all RGB channels as a method for processing. incorporating a constriction factor-based PSO approach. This
The author [21] is to use the characteristics of MRI novel approach not only strengthens convergence speed but
images to classify types of brain cancer. To make the also mitigates issues related to premature convergence. Thus,

177146 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

techniques like image zooming, panning, and noise reduc-

tion are used to enhance data utility [23]. Figure 2 shows
the brain tumor classification dataset, and an additional dia-
betic retinopathy dataset with 10 classes was included to
expand [24] the study which evaluates model robustness and
performance [25].

B. DATA CONVERSION AND IMAGE DIMENSIONS

Each MRI was collected originally in the form of DICOM for-
mat which is usually the default image extension of the MRI
machine but all images were eventually converted into. jpeg
by the software called DICOM viewer. Each of the images
contains 2560∗ 2560 pixels resolution. MRI dataset contains
various modalities e.g. SCANO S/C/A, SCANO S/C/A-2, T1
AX 6mm, T2 AX 6mm, Edge SCANO S/C/A-2, Edge T2 AX
6mm, FLAIR COR 6mm, Edge T1 AX 6mm, T2 COR 6mm,
Edge FLAIR COR 6mm, Edge T2 COR 6mm. Every image
contains a horizontal and vertical resolution of 96 dpi with a
depth of 24 bits as shown in Figure 3.
When preparing the retinopathy OCT scans for processing,
their original resolution of 2976∗ 2976 pixels was reduced
to 224∗ 224∗ 3 pixels. With the use of image augmentation
techniques, all classes generate a minimum of 8242 training
images including 2180 images for validation and 2173 testing
images. Additionally, Figure 4 displays the training samples
for each class.
Total of 8242 DR images were classified into 10 dis-
tinct classes that are diabetic papillopathy (994), mild npdr
(990),mild npdr & csme (1553), moderate npdr (1049), mod-
erate npdr & csme (488), pdr (1465), proliferative diabetic
retinopathy+ csme (192), severe npdr (144), severe npdr &
csme (702), very severe npdr (638) as mentioned in Figure 5.
To facilitate reliable findings despite the limited dataset size,
FIGURE 2. Architectural flow. several steps were taken to enhance its suitability.
OCT classes contained diabetic papillopathy (994), mild
the study contributes a unique perspective to the ongoing dis- npdr (990), mild npdr & csme (1553), moderate npdr (1049),
course, offering insights into the synergistic potential of deep moderate npdr & csme (488), pdr (1465), proliferative dia-
learning and PSO methodologies in addressing contemporary betic retinopathy + csme (192), severe npdr (144), severe
challenges within computer vision. npdr & csme (702), very severe npdr (638) as mentioned in
Figure 6.
Table 2 provides a summary of the dataset’s distribution,
III. METHODOLOGY
image dimensions, and resolution. It highlights the diver-
The base neural network was fine-tuned by freezing all but the
sity of classes, image expansion, and resolution details. All
first two layers and adding custom layers to extend Mobile
images are in.jpeg format.
Net while retaining pre-trained weights. Abundant data is
crucial for accurate image classification, particularly with
contrast and color variations. The Figure 2 Architectural Flow C. MAGNETIC RESONANCE IMAGE
used a 70%/15%/15% data split for training, validation, and A dataset of 3,334 brain MRI images was categorized into
testing. A PSO constriction factor was applied to optimize eight types as mentioned in Table 3. Images were resized from
Mobile Net improving learning rate, epochs, convergence, 224 × 224 × 3.
and validation accuracy.
D. CONVOLUTIONAL NEURAL NETWORKS (CNN) WITH
A. DATASET AUGMENTATION MODERN TRANSFER LEARNING
The Model success is largely shaped by its training process A widely recognized DL technique designed for the identi-
and learning framework. A common issue is the mismatch fication and classification of features in computer-generated
between the dataset and annotation criteria. Pre-processing images is the Convolutional Neural Network (CNN) [25].

VOLUME 13, 2025 177147

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

FIGURE 3. (Continued.) MRI modalities and classes abscess.

FIGURE 4. Number of training images per class (MRI).

FIGURE 5. Number of training images per class DR.

comprehends images [26]. Just like the visual cortex, CNNs

FIGURE 3. MRI modalities and classes abscess. meticulously analyze a substantial amount of information
when presented with visual stimuli. To capture comprehen-
sive knowledge across the entire perceptual region, neurons
CNNs were initially developed to handle visual data and were within the CNN possess specialized abilities inside their
inspired by the way the human visual cortex processes and receptive grounds and engage in communication with one

177148 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

TABLE 2. Dataset distribution for OCT (Diabetic retinopathy).

TABLE 3. Distribution of datasets for MRI training and testing.

FIGURE 6. Class and types of DR.

another [27], [28]. Figure 7 shows how convolutional neural

network (CNN) neurons are similar to individual receptive
fields that exist in actual nerve cells. A unique kernel matrix
is used by each network layer to aid in the detection of cer-
tain patterns including peaks, borders, bends, and more [29].
A CNN model comprises multiple pooling and convolutional
layers arranged in a specific sequence. The outputs from each
layer are transmitted toward the subsequent layer within the FIGURE 7. CNN architecture.
network, effectively serving as a hidden layer [30].
Adam optimizer [31] can determine different learning rates
for each parameter within a neural network. Adam is a learn- change the learning rates associated with each weight in the
ing rate-adjustable optimizer. Adam is an acronym that stands neural network as mentioned in Equation 1.
for adaptive moment estimation. Adam utilizes projections of
the gradient’s initial and subsequent moments to dynamically Mn = E(x)2 (1)

VOLUME 13, 2025 177149

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

Where, input to the sigmoid function, which may include a vector or a

M Is the Moments scalar. Using a smooth S-shaped curve, the sigmoid function
x Is Random variables transforms any real-valued input into a range of 0 to 1.
E Projected value to the power of n
Yi,j = max k−1 k−1
m=0 max n=0 Xs.i+m,s.j+n (4)
n Is the number of iterations of random variables
1
sigmoid function = σ (x) = (5)
The final step in utilizing the moving averages is to adjust the 1 + e−x
learning rate on a per-parameter basis. In the Adam optimiza-
The input is flattened initially, and subsequently passed
tion method, this process is quite straightforward. To update
through a dense layer with ReLU activation, a dropout layer
the model weights, the following procedure is employed.
for regularization, and a final dense layer with softmax acti-
Here, ‘w’ represents the model’s weights, (resembling the let-
vation for classification. This structure can be used in deep
ter ‘n’) and signifies the step size, which can potentially vary
learning tasks in image classification. The technique requires
with each iteration and essentially encapsulates the update
a pre-trained base model which adds custom layers at the top
rule employed by Adam mentioned in Equation 2.
of the mobile-Net model to fine-tune for a specific task.
m∧ t
wt = wt − 1 − n √ (2)
vt + E E. ENHANCING MODEL PERFORMANCE BY THE
Where, INTEGRATION OF PRE-TRAINED CLASSIFIERS
In recent times, we have observed a substantial enhancement
w is the model weight
in the effectiveness of image classification tasks, primar-
n is the number of iterations (random variables)
ily attributable to the widespread adoption of deep learning
m and v are first and second moment estimations.
techniques. However, the worth of these techniques is often
E is the expected value of n
contingent on having access to large sets of annotated data,
The classification operation within the model is carried out posing a challenge when dealing with smaller datasets.
by using the Fully Connected Layer for the ultimate last To address this limitation, transfer learning strategies have
layer [32]. Convolutional neural networks (CNNs) consist emerged as a promising solution. This study seeks to equate
of four layers: input, convolution, classification, and output. the performance of various pre-trained neural networks.
The most popular activation layer, the ReLU layer, chooses Conducted evaluations across four distinct image datasets,
weight values between zero and infinity. The ReLU layer considering 16 different pre-trained models. Our findings
performs a threshold operation on each of the inputs, and reveal that CLIP-ViT-B and ViT-H-14 consistently exhibited
values less than zero are replaced by zero. The output of the best overall performance across the datasets. Notably,
this convolutional operation is an activation map that shows the CLIP-ResNet50 model demonstrated comparable per-
the kernel’s calculated response for every image point. The formance with lower variability. Consequently, the Authors
method for learning the Dot Product in the convolutional part provided compelling evidence to guide the selection of mod-
is shown in Equation 3. els for feature extraction in image classification [33].DR
stands as a persistent diabetic retinal disorder, posing a sig-
KM 
X  nificant threat to the vision of a substantial portion of the
Conv2Dtn
(a,b) =
tn
crf(a,b) ∗ k(a,b) (3) working-age population and potentially leading to visual
b=1 impairment. The classification of DR severity levels has
Where, been a formidable challenge, given the difficulty in analyz-
KM: Kernel Matrix ing lesion features. An effective detection method is crucial
(a,b): Image rows and columns for screening, enabling the classification of subtle retinal
crf: Current Frame pathologies. Deep neural architectures have proven instru-
N: Total iterations mental in diagnosing eye diseases, aiding ophthalmologists
tn: Total numbers and Kernel parameters in delivering timely treatment [34].

Equation 4, Y (i,j) represents the output value at point (i,j) F. MOBILE NET
within the pooled feature map. Here, the input feature map The deep learning plays a vital role in object detection and
is denoted by X, the pooling kernel size is denoted by k classification in computer vision, especially in the domains
(particularly in a square window), and the stride is denoted of computer, mobile, and embedded vision. The Mobile-Net
by s, in this equation. Each pooling space is defined by kernel framework, which is both efficient and lightweight, is often
size and stride, this equation finds the maximum value that used for these kinds of applications. Due to limitations in
reduces the spatial dimensions of the feature map without memory, energy, and power, such architectures are difficult
changing its significant attributes. A well-liked activation to deploy on microcontroller units with limited resources.
function in neural networks and machine learning, the sig- In addition, pruning or deep compression typically results in
moid function is shown in Equation 5. Here, x represents the a decline in overall model accuracy when reducing the size

177150 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

and total number of parameters. In comparison to the base-

line mobile net, the three hybrid simulations by Mobile-Net
models shown in this study enhanced accuracy while also
being smaller in size, having fewer layers, requiring less
average computation time, and exhibiting little over fitting
problems. The lightweight mobile Net is an acronym given
to the smallest model, which measures in at 9.9 MB. Its
objective is to boost accuracy by using the Random Erasing
regularization approach in place of dropout and by substitut-
ing drop activation for the conventional non-linear activation
function ReLU. Instead of applying depth-wise separable FIGURE 8. Mobile net [36].
convolutions, which are employed in the baseline mobile
Net [35], Separable Convolutions are used to reduce the
layered volume and size of the model as well. The feature H. CONSTRICTION FACTOR
vector is evaluated in Equation 06. The constriction factor in particle swarm optimization is man-
aged through the search space known as explore and exploit.
F = f (X ; W , b) (6) The inclusion of a constriction factor strengthened and opti-
mized particle movement and convergence in solution space
Where, as shown in Equation 10.
f = is the feature vector 2
X = input image CF = √ (10)
W = weight tensor |2 − φ − φ2 − 4φ|
b = bias tensor Combining equations 9 and 10 we get Equation 11.
υi (t + 1) = CF[ω.υi (t) + c1 .r1 . pbesti − xi (t)

G. PARTICLE SWARM OPTIMIZATION (PSO)
+ c2 .r2 . gbesti − xi (t) ]

(11)
The Eberhart and Kennedy introduced Particle Swarm Opti-
mization (PSO) [37], A meta-heuristic algorithm, PSO uses
I. PSO OBJECTIVE FUNCTION
particles to explore the solution space, guided by individual
The objective function has lower and upper limits for each
and collective best positions. Each iteration enhances conver-
bound, which constitute iteration and learning rate for opti-
gence towards the optimal solution, efficiently tuning hyper
mization benchmark. The lower bounds for the learning
parameters for model optimization, as shown by velocity and
rate and the number of epochs and swarm have been set
location update equations. Equations 07, 08, and 09 show
at 0.001 and 2, respectively, indicating the lowest values
velocity and location updates, respectively.
these parameters can take. Conversely, the upper bounds for
υi (t + 1) = υi (t) + .r1 . pbest i − xi (t)
 the learning rate and the number of epochs are established
at 0.005 and 3, representing the maximum values they can
+ c2 .r2 . gbest i − xi (t)

(7)
attain. Furthermore, the optimization process is constrained
Where, to run for a maximum of one iteration, making it a highly
limited search space. This configuration provides a basis
υi (t) = inertia. for fine-tuning and controlling the parameters of the swarm
c1 c2 = coefficients optimization algorithm to achieve specific objectives within
r1 r2 = random values these defined bounds and iterations. Image resolution of
pbest = particle best position (224*224*3) pixels with RGB color contrast was included
gbest = Global best position of the particle in the original input images. These concerns are impor-
tant in identifying the models’ behaviors and performance
ω max − ωmin throughout the evaluation and training processes. The PSO
ω = ωmax − × Iter (8)
Iter max optimization is performed and the best solution along with
its fitness value is obtained. Hyper parameters of objection
Where,
functions are mentioned in Table 4.
ω = weights for (min, max) parameters
Iter max = max iteration J. FINE-TUNING AND LAYER OPTIMIZATION
Iter = current iterations Layer optimization and customization have been made to the
pre-trained Mobile Net model. It begins by freezing all the
υi (t + 1) = ω.υi (t) + c1 .r1 . pbest i − xi (t)
 layers in the base model to retain their pre-trained weights.
However, it makes an exception for the first two layers, setting
+ c2 .r2 . gbest i − xi (t)

(9)
them as trainable. Subsequently, custom layers are added

VOLUME 13, 2025 177151

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

TABLE 4. PSO objective function parameters. TABLE 5. Diabetic retinopathy accuracy metrics.

to the model mobile Net model. The output from the base
model is flattened, followed by a fully connected layer with
1024 units and a ReLU activation function. A dropout layer
with a dropout rate of 0.2 is introduced for regularization
and finally, a dense layer with 8 classes of Brain tumor and
10 classes of retinopathy with softmax activation function is
appended to produce the model’s output. This modification
extends the original model while preserving the pre-trained
weights in the majority of its layers.

IV. RESULTS AND DISCUSSION

Robustness in deep learning refers to the model that performs
well across various conditions including various datasets,
shapes, size, dimensions, and extracted features. I’s a desir-
able characteristics that indicate the model stability that
may have varied during the training and validation pro-
cess [38].To check the robustness of the MOB-CFPSO model,
we have performed three consecutive simulations on volu-
metric images on both of the datasets i.e. MRI and DR to
check the reliability and accuracy of the model shown in
Tables 5 and 6. Accuracy Metrics of the Dual deterministic
model MOB-CFPSO (Mobile-Net Constriction Factor Parti-
cle Swarm Optimization) with state-of-the-art performance
measures can be seen extensively in the following tables.
Diabetic papillopathy shows high precision, recall, and
F1-score values, indicating the actual and correctly identi-
fied positive cases with minimized false positive instances.
Specifically, precision values ranging from 89% to 98% sug-
gest a low false positive rate, meaning that when the models
predict positive cases, they are corrected the majority of
the time. Similarly, recall values ranging from 92% to 96%
indicate a high rate of correctly identifying true positive cases
out of all actual positive cases. Mild npdr exhibited strong
performance with slightly higher precision values compared
to Diabetic Papillopathy. Precision values ranging from 92%
and ending at 92% imply a low false positive rate. Addi-
tionally, recall values ranging from 95% to 98% indicate a
high rate of correctly identifying true positive cases out of all
actual positive cases, along with corresponding true positive
and false positive counts. Mild npdr + csme across all simu-
lations, showed credible performance, with precision values
ranging from 92% to 98%, indicating a relatively low false
positive rate. Similarly, recall values ranging from 94% to
97% suggest a high rate of correctly identifying true positive
cases. The F1 scores, ranging from 94% to 97%, reflect the between the two metrics. Moderate npdr class established
harmonic mean of precision and recall, representing a balance robust performance, with precision values ranging from 94%

177152 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

TABLE 5. (Continued.) Diabetic retinopathy accuracy metrics. TABLE 6. MRI accuracy metrics.

FIGURE 9. Training & validation accuracy.

to 98%, specifying a relatively low false positive rate. Recall

values ranged from 93% to 100%, implying a high rate of
correctly identifying true positives. F1 scores, ranging from
94% to 99%, reflect a balance between precision and recall.
Moderate npdr + csme class reached high precision, recall,
and F1-score values, with the precision range from 95% to
100%, recall 86% to 99%, and F1-score 91% to 100% across
all three simulations. PDR class achieved high precision
(95% to 100%), recall (99% to 100%), and F1-score (97%
to 100%). pdr+csme achieved precision (100%), meaning no
false positives were reported. However, there is variability
in recall, with values ranging from 78% to 100%. F1 score
shows continuous improvement from 88% to 100%. Severe
npdr recall result shows some continuous improvement from
(0.70 to 0.89) moreover severe npdr + csme class showed showing a low occurrence of false positives. Nevertheless,
that npdr with csme accuracy scores range from 96% to 98%, there is a range of variability in recall, covering from 79%

VOLUME 13, 2025 177153

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

TABLE 7. Model optimization of modified results of MOBPSO.

to 96%,as well as in F1-score, spanned from 88% to 96%.

Very severe npdr class shows perfect precision, recall, and
F1-score values of 100%, the models accurately identified all
positive cases without any false positives or false negatives
across all simulations, showcasing their exceptional accuracy
and reliability in diagnosing very severe npdr.
FIGURE 10. Training and validation loss (MRI).

A. MAGNETIC RESONANCE IMAGE

Three different experiments have been conducted to check the
model robustness that shows extremely substantial results in
MRI tumor diagnosis as shown in Table 6. In both classes,
Abscess and Arachnoid Cyst conditions validate high preci-
sion, recall, and F1-score values, with precision ranging from
98% to 100%, recall from 99% to 100%, and F1-score from
98 to 100% across all simulations. Low-grade astrocytoma
and meningioma precision ranged from 94% to 100%, r ecall
started and ended up at 100%, and F1-score from 97% to
100% across simulations.
Meningioma, precision consistently achieved 100%, while
recall ranged from 83% to 98% and F1-score from 91% to
99%. In Metastases (METS) and Tuberculomas, as well as
Neurocysticercosis and Meningitis the precision ranged from
97% to 100%, recall from 88% to 100%, and F1-score from
92% to 100% across simulations. Neurocysticercosis and
Meningitis, precision varied from 90% to 100%, recall from
97% to 100%, and F1-score from 95% to 100%. Porocephalic
Cyst, precision ranged from 94% to 100%, recall from 89 to
100%, and F1-score from 97% to 100% across simulations
FIGURE 11. Training & validation loss (DR).
and thalamic glioma, precision consistently achieved 100%,
while recall and F1 score started and ended up at 100%. These
results indicate the model’s effectiveness in accurately iden- Validation process of images which ensured the training accu-
tifying cases of both porocephalic cyst and thalamic glioma, racy of 93.91, while second (70%,15%, 15% → 3868, 1873,
with generally high precision, recall, and F1-score values. 1879), and third simulation of model (70%,15%,15% →
MRI model demonstrated flawless training and validation 3889, 2180, 2173) reached the Training and validation accu-
accuracy at 100% and 99.86% respectively as shown in racy of 95.09 and 97.06% respectively. In the same way,
Figure 09. Robustness is considered to be the most vital part the third MRI model simulation contained 70%, 15%, and
of the dual deterministic model therefor; we have conducted 15% → 1863,732,739 Training, Testing, and Validation pro-
three different simulations to split the model into volumetric cesses of the images that showed training and validation
data (number of images) and further applied image augmen- accuracy of 100 % and 99.86% respectively.
tation to check model behavior for un-seen data as shown in Moreover, the training loss is impressively low at 0.0091,
Table 7. while the validation low loss at 0.071 still reflects the reli-
In the First simulation of DR dataset that contained ability of MOB-PSO in terms of robustness and model
(70%,15%,15% → 3778, 1515, 1499) Training, Testing, optimization as shown in Figure 10.

177154 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

FIGURE 12. Training & validation accuracy (DR).

FIGURE 14. Confusion matrix validation (DR).

FIGURE 13. Confusion matrix training (DR).

DR achieved a commendable training accuracy of approx-

imately 95.09% and 97.66% training and validation accuracy
as shown in Figure 12, Training and validation loss of
0.2469% and 0.1902% respectively however model loss can FIGURE 15. Confusion matrix training (MRI).
be seen in Figure 11.
Figures 13 and 14 present the confusion matrix for
both the training and validation of the diabetic retinopathy npdr+csme correctly identified 388 with 05 misclassified
dataset, illustrating the performance of a machine-learning instances. The accompanying heat map visually conveys
model across all classes. The majority of the classes have these values through color representation, enhancing the ease
correctly identified and classified all instances with zero of interpretation and analysis of the model on both the train-
misclassification which are mild npdr, moderate npdr+csme, ing and validation sets.
pdr, pdr+csme, and very severe npdr, however, the confu- Figures 15 and 16 present the confusion matrix of MRI,
sion matrix displays very few and minor misclassification enlightening the performance of a machine learning model
instances while Diabetic papliopathy.In the same way, mild across all eight classes. The notable number of correctly

VOLUME 13, 2025 177155

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

FIGURE 18. ROC/AUC with DR classes.

FIGURE 16. Confusion matrix validation (MRI).

TABLE 8. Comparative analysis.

FIGURE 17. ROC/AUC with MRI classes.

identified instances and the number of true predictions for

each class. Considering both the confusion matrix there are
no single misclassified instances (absess, archnoid cyst, low
grade astrocytoma, miningioma, mets and tuberculomas, neu-
rocustechrrosi, and meningitus, poroncephalic cyst, thalmic
glioma) at all, which proves the model robustness.
ROC and AUC of the MRI dataset showed that there are
no error rates or misclassifications in any of these classes.
Despite this fact, every class managed got a perfect score,
except meningioma, which managed to reach a 98% accuracy
rate with no error rate as well, which is quite satisfactory as Figure 17 shows the ROC graph with an accuracy of 100%
shown in Figure 17 even in the DR dataset only class mild among all of the MRI classes. This performance indicates the
npdr produced the highest error rate of 0.08% as shown in model reaches faultless sensitivity and efficiency with zero
Figure 18. percent error rates as well.

177156 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

Below is Table 8 (comparative analysis) of various mod- [3] S. V. Chilukoti, L. Shan, A. S. Maida, and X. Hei, ‘‘A reliable diabetic
els with optimum accuracy by using a heterogenic dataset. retinopathy grading via transfer learning with quadratic weighted Kappa
metric,’’ Res. Sq., pp. 1–18, Feb. 2023, doi: 10.21203/rs.3.rs-2383368/v1.
Researchers proposed various PSO variants for classification [Online]. Available: https://bmcmedinformdecismak.biomedcentral.com/
and feature extraction. This analysis is based on the related articles/10.1186/s12911-024-02446-x
work section contained in this paper. [4] Z. Rudnicka, J. Szczepanski, and A. Pregowska, ‘‘Artificial intelligence-
based algorithms in medical image scan segmentation and intelligent visual
content generation—A concise overview,’’ Electronics, vol. 13, no. 4,
B. SIMULATION ENVIRONMENT p. 746, Feb. 2024, doi: 10.3390/electronics13040746.
Python (Anaconda 6.5.2) was run on an Intel i7-2720QM [5] Y. Ding, L. Li, W. Wang, and Y. Yang, ‘‘Clustering propagation for uni-
versal medical image segmentation,’’ presented at the IEEE/CVF Conf.
CPU at 2.2 GHz with 32 GB RAM, This hardware setup Comput. Vis. Pattern Recognit. (CVPR), Jun. 2024.
enabled the effective execution of the models and allowed [6] P. Arora, R. Mehta, and R. Ahuja, ‘‘An integration of meta-heuristic
for the assessment of their performance. Python libraries and approach utilizing kernel principal component analysis for multi-
modal medical image registration,’’ Cluster Comput., vol. 27, no. 5,
modules used for pre-trained transfer learning tasks include, pp. 6223–6246, Aug. 2024, doi: 10.1007/s10586-024-04281-1.
scikit learn, os, shutil, random, numpy, TensorFlow, math, [7] J.-G. Lee, S. Jun, Y.-W. Cho, H. Lee, G. B. Kim, J. B. Seo, and
and matplotlib. N. Kim, ‘‘Deep learning in medical imaging: General overview,’’ Korean
J. Radiol., vol. 18, no. 4, p. 570, Aug. 2017, doi: 10.3348/kjr.2017.18.
4.570.
V. CONCLUSION AND FUTURE WORK [8] M. K. Jabbar, J. Yan, H. Xu, Z. U. Rehman, and A. Jabbar, ‘‘Trans-
In this research, we have developed a dual determinis- fer learning-based model for diabetic retinopathy diagnosis using retinal
tic model using the MOB-CFPSO Mobile-Net framework. images,’’ Brain Sci., vol. 12, no. 5, p. 535, Apr. 2022, doi: 10.3390/brain-
sci12050535.
A modified particle swarm optimization approach is used, [9] A. Akter, N. Nosheen, S. Ahmed, M. Hossain, M. A. Yousuf,
with a constriction factor to reduce velocity, minimize error M. A. A. Almoyad, K. F. Hasan, and M. A. Moni, ‘‘Robust clinical
rate, and enhance classification accuracy. The model per- applicable CNN and U-Net based algorithm for MRI classification and
segmentation for brain tumor,’’ Expert Syst. Appl., vol. 238, Mar. 2024,
formed training and evaluation using a dataset that was Art. no. 122347, doi: 10.1016/j.eswa.2023.122347.
obtained and constituted of MRI and DR data. Both datasets [10] R. Haque, M. M. Hassan, A. K. Bairagi, and S. M. Shariful Islam,
included a wide range of features and classes. Ensured that ‘‘NeuroNet19: An explainable deep neural network model for the clas-
sification of brain tumors using magnetic resonance imaging data,’’
the machine learning model accurately diagnoses classes with Sci. Rep., vol. 14, no. 1, p. 1524, Jan. 2024, doi: 10.1038/s41598-024-
a minimum error rate. MOB-CFPSO has achieved unprece- 51867-1.
dented 100% accuracy for the MRI dataset and 97.6% for the [11] B. Revathi, S. K. K. Elizabeth, P. Nagaraj, S. S. Birunda, and D. Nithya,
DR dataset, as compared to the baseline model. ‘‘Particle swarm optimization based detection of diabetic retinopa-
thy using a novel deep CNN,’’ in Proc. 3rd Int. Conf. Artif. Intell.
In the future, PSO may be used to alleviate the issue of Smart Energy (ICAIS), Coimbatore, India, Feb. 2023, pp. 998–1003, doi:
mode collapse in Generative Adversarial Networks (GANs). 10.1109/ICAIS56108.2023.10073926.
Mode collapse is the phenomenon when the generator only [12] F. H. Kuwil, ‘‘A new feature extraction approach of medical image based
on data distribution skew,’’ Neurosci. Inf., vol. 2, no. 3, Sep. 2022,
generates a restricted range of samples. Particle Swarm Art. no. 100097, doi: 10.1016/j.neuri.2022.100097.
Optimization (PSO) may enhance the GAN by dynamically [13] K. S. Gorde and A. A. Gurjar, ‘‘A comparative analysis on diabetic
modifying parameters during the training phase, hence facil- retinopathy using deep learning and nature based optimization algo-
rithms,’’ in Proc. Int. Conf. Intell. Syst., Adv. Comput. Commun. (ISACC),
itating the exploration of a broader spectrum of produced Silchar, India, Feb. 2023, pp. 1–9, doi: 10.1109/ISACC56298.2023.
samples. 10083804.
[14] E. Barges and E. Thabet, ‘‘GLDM and Tamura features based KNN and
particle swarm optimization for automatic diabetic retinopathy recognition
DATA COLLECTION STATEMENT
system,’’ Multimedia Tools Appl., vol. 82, no. 1, pp. 271–295, Jan. 2023,
This research was carried out with ethical approval from the doi: 10.1007/s11042-022-13282-4.
University of Kuala Lumpur, Malaysia, under the approval [15] R. Ibrahim, R. Ghnemat, and Q. A. Al-Haija, ‘‘Improving Alzheimer’s
number (UNIKL REC/2024/PG/APV/01). disease and brain tumor detection using deep learning with particle
swarm optimization,’’ AI, vol. 4, no. 3, pp. 551–573, Jul. 2023, doi:
10.3390/ai4030030.
ACKNOWLEDGMENT [16] U. Bhimavarapu and G. Battineni, ‘‘Automatic microaneurysms detection
I would like to express my heartfelt gratitude to for early diagnosis of diabetic retinopathy using improved discrete par-
ticle swarm optimization,’’ J. Personalized Med., vol. 12, no. 2, p. 317,
Dr. Fouzia Noor [48], a Medical Specialist and Radiologist, Feb. 2022, doi: 10.3390/jpm12020317.
who played a significant role in this study by actively partic- [17] R. K. Tripathi and A. S. Jalal, ‘‘Novel local feature extraction for age
ipating and performing the data labeling and annotation. invariant face recognition,’’ Expert Syst. Appl., vol. 175, Aug. 2021,
Art. no. 114786, doi: 10.1016/j.eswa.2021.114786.
[18] Y. B. Özçelik and A. Altan, ‘‘Overcoming nonlinear dynamics in dia-
REFERENCES betic retinopathy classification: A robust AI-based model with chaotic
[1] J. Lu, V. Behbood, P. Hao, H. Zuo, S. Xue, and G. Zhang, ‘‘Transfer swarm intelligence optimization and recurrent long short-term memory,’’
learning using computational intelligence: A survey,’’ Knowl.-Based Syst., Fractal Fractional, vol. 7, no. 8, p. 598, Aug. 2023, doi: 10.3390/fractal-
vol. 80, pp. 14–23, May 2015, doi: 10.1016/j.knosys.2015.01.010. fract7080598.
[2] M. Oulhadj, J. Riffi, C. Khodriss, A. M. Mahraz, A. Bennis, A. Yahyaouy, [19] M. U. Ali, K. D. Kallu, H. Masood, S. J. Hussain, S. Ullah, J. H. Byun,
F. Chraibi, M. Abdellaoui, I. B. Andsaloussi, and H. Tairi, ‘‘Diabetic A. Zafar, and K. S. Kim, ‘‘A robust computer-aided automated brain
retinopathy prediction based on transfer learning and ensemble voting,’’ tumor diagnosis approach using PSO-ReliefF optimized Gaussian and
in Digital Technologies and Applications, vol. 668. Cham, Switzerland: non-linear feature space,’’ Life, vol. 12, no. 12, p. 2036, Dec. 2022, doi:
Springer, 2023, pp. 929–937, doi: 10.1007/978-3-031-29857-8_92. 10.3390/life12122036.

VOLUME 13, 2025 177157

 A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

[20] P. Melin, D. Sánchez, and R. Cordero-Martínez, ‘‘Particle swarm opti- [38] T. Freiesleben and T. Grote, ‘‘Beyond generalization: A theory of robust-
mization of convolutional neural networks for diabetic retinopathy clas- ness in machine learning,’’ Synthese, vol. 202, no. 4, pp. 1–28, Sep. 2023,
sification,’’ in Fuzzy Logic and Neural Networks for Hybrid Intelligent doi: 10.1007/s11229-023-04334-9.
System Design. Cham, Switzerland: Springer, 2023, pp. 237–252, doi: [39] P. Porwal. (Apr. 24, 2018). Indian Diabetic Retinopathy Image
10.1007/978-3-031-22042-5_14. Dataset (IDRiD). [Online]. Available: https://ieee-dataport.org/open-
[21] B. Shyamala and S. H. D. Brahmananda, ‘‘Brain tumor classification access/indian-diabetic-retinopathy-image-dataset-idrid
using optimized and relief-based feature reduction and regression neu- [40] A. A. Abdulsahib, M. A. Mahmoud, H. Aris, S. S. Gunasekaran, and
ral network,’’ Biomed. Signal Process. Control, vol. 86, Sep. 2023, M. A. Mohammed, ‘‘An automated image segmentation and useful feature
Art. no. 105279, doi: 10.1016/j.bspc.2023.105279. extraction algorithm for retinal blood vessels in fundus images,’’ Electron-
[22] M. Ali, J. H. Shah, M. A. Khan, M. Alhaisoni, U. Tariq, T. Akram, ics, vol. 11, no. 9, p. 1295, Apr. 2022, doi: 10.3390/electronics11091295.
Y. J. Kim, and B. Chang, ‘‘Brain tumor detection and classification using [41] Y. Yin, D. Chang, G. Song, S. Sang, T. Zhi, J. Liu, L. Luo, and
PSO and convolutional neural network,’’ Comput., Mater. Continua, M. Soleymani, ‘‘FG-net: Facial action unit detection with generalizable
vol. 73, no. 3, pp. 4501–4518, 2022, doi: 10.32604/cmc.2022.030392. pyramidal features,’’ presented at the IEEE/CVF Winter Conf. Appl. Com-
put. Vis. (WACV), Jan. 2024.
[23] L. Wang, H. Lu, and P. Zhang, ‘‘Online visual tracking via cross-similarity-
[42] N. Di Domenico, G. Borghi, A. Franco, and D. Maltoni, ‘‘Dealing
based Siamese network,’’ Concurrency Comput., Pract. Exper., vol. 33,
with subject similarity in differential morphing attack detection,’’ 2024,
no. 15, p. e5617, Aug. 2021, doi: 10.1002/cpe.5617.
arXiv:2404.07667.
[24] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, ‘‘Rethinking [43] H. Shakibania, S. Raoufi, B. Pourafkham, H. Khotanlou, and
the inception architecture for computer vision,’’ in Proc. IEEE Conf. M. Mansoorizadeh, ‘‘Dual branch deep learning network for detection and
Comput., Vis. Pattern Recognit. (CVPR), 2016, pp. 2818–2826. stage grading of diabetic retinopathy,’’ Biomed. Signal Process. Control,
[25] C. Srinivas, M. Zakariah, Y. A. Alothaibi, K. Shaukat, B. Partibane, and vol. 93, Jul. 2024, Art. no. 106168, doi: 10.1016/j.bspc.2024.106168.
H. Awal, ‘‘Deep transfer learning approaches in performance analysis [44] C. Im, C.-B. Song, J. Lee, D. Kim, and H. Seo, ‘‘Investigating the effect of
of brain tumor classification using MRI images,’’ J. Healthcare Eng., brain atrophy on transcranial direct current stimulation: A computational
vol. 2022, pp. 1–17, Mar. 2022, doi: 10.1155/2022/3264367. study using ADNI dataset,’’ Comput. Methods Programs Biomed., vol. 257,
[26] Y.-D. Zhang, S. C. Satapathy, D. S. Guttery, J. M. Górriz, and Dec. 2024, Art. no. 108429, doi: 10.1016/j.cmpb.2024.108429.
S.-H. Wang, ‘‘Improved breast cancer classification through combin- [45] T. Rahman, M. S. Islam, and J. Uddin, ‘‘MRI-based brain tumor classifica-
ing graph convolutional network and convolutional neural network,’’ tion using a dilated parallel deep convolutional neural network,’’ Digital,
Inf. Process. Manage., vol. 58, no. 2, Mar. 2021, Art. no. 102439, doi: vol. 4, no. 3, pp. 529–554, Jun. 2024, doi: 10.3390/digital4030027.
10.1016/j.ipm.2020.102439. [46] S. Metlek and H. Çetiner, ‘‘ResUNet+: A new convolutional and atten-
[27] T. Goel, R. Murugan, S. Mirjalili, and D. K. Chakrabartty, ‘‘OptCoNet: tion block-based approach for brain tumor segmentation,’’ IEEE Access,
An optimized convolutional neural network for an automatic diagnosis vol. 11, pp. 69884–69902, 2023, doi: 10.1109/ACCESS.2023.3294179.
of COVID-19,’’ Int. J. Speech Technol., vol. 51, no. 3, pp. 1351–1366, [47] M. I. Sharif, J. P. Li, M. A. Khan, S. Kadry, and U. Tariq, ‘‘M3BTCNet:
Mar. 2021, doi: 10.1007/s10489-020-01904-z. Multi model brain tumor classification using metaheuristic deep neural
[28] Y. Kurata, M. Nishio, Y. Moribata, A. Kido, Y. Himoto, S. Otani, network features optimization,’’ Neural Comput. Appl., vol. 36, no. 1,
K. Fujimoto, M. Yakami, S. Minamiguchi, M. Mandai, and Y. Nakamoto, pp. 95–110, Jan. 2024, doi: 10.1007/s00521-022-07204-6.
‘‘Automatic segmentation of uterine endometrial cancer on multi-sequence [48] F. Noor. (MBBS, FCPS II) (MBBS, FCPS II). Diagnostic Radiol. Combined
MRI using a convolutional neural network,’’ Sci. Rep., vol. 11, no. 1, Mil. Hospital, Malir Cantt, Karachi, Pakistan. Accessed: Aug. 18, 2024.
p. 14440, Jul. 2021, doi: 10.1038/s41598-021-93792-7. [Online]. Available: [email protected]
[29] Z. Wang, W. Zhao, W. Du, N. Li, and J. Wang, ‘‘Data-driven fault diagnosis
method based on the conversion of erosion operation signals into images
and convolutional neural network,’’ Process Saf. Environ. Protection,
vol. 149, pp. 591–601, May 2021, doi: 10.1016/j.psep.2021.03.016.
[30] C. L. Choudhury, C. Mahanty, R. Kumar, and B. K. Mishra, ‘‘Brain
tumor detection and classification using convolutional neural network ASIF RAZA received the B.S. and M.S. degrees
and deep neural network,’’ in Proc. Int. Conf. Comput. Sci., Eng. specializing in computer networks, in 2001 and
Appl. (ICCSEA), Gunupur, India, Mar. 2020, pp. 1–4, doi: 10.1109/ICC- 2006, respectively. Currently, he is pursuing the
SEA49143.2020.9132874. Ph.D. degree in information technology with Uni-
[31] Z. Zhang, ‘‘Improved Adam optimizer for deep neural networks,’’ in Proc. versity of Kuala Lumpur, MIIT Campus, Malaysia,
IEEE/ACM 26th Int. Symp. Qual. Service (IWQoS), Banff, AB, Canada, with a focus on artificial intelligence and machine
Jun. 2018, pp. 1–2, doi: 10.1109/IWQOS.2018.8624183. learning in medical imaging. He is a Senior Lec-
[32] J. M. Sahan, E. I. Abbas, and Z. M. Abood, ‘‘A facial recogni- turer with the Sir Syed University of Engineering
tion using a combination of a novel one dimension deep CNN and and Technology.
LDA,’’ Mater. Today, Proc., vol. 80, pp. 3594–3599, Jul. 2023, doi:
10.1016/j.matpr.2021.07.325.
[33] E. da Silva Puls, M. V. Todescato, and J. L. Carbonera, ‘‘An evaluation of
pre-trained models for feature extraction in image classification,’’ 2023,
arXiv:2310.02037.
[34] A. M. Dayana and W. R. S. Emmanuel, ‘‘An enhanced swarm
optimization-based deep neural network for diabetic retinopathy clas- SHAHRULNIZA BIN MUSA received the mas-
sification in fundus images,’’ Multimedia Tools Appl., vol. 81, no. 15, ter’s degree in integrated research study and the
pp. 20611–20642, Jun. 2022, doi: 10.1007/s11042-022-12492-0. Ph.D. degree in communication network secu-
[35] D. Sinha and M. El-Sharkawy, ‘‘Thin MobileNet: An enhanced mobilenet rity from the Faculty of Electrical and Electronic
architecture,’’ in Proc. IEEE 10th Annu. Ubiquitous Comput., Electron. Engineering, Loughborough University, U.K., in
Mobile Commun. Conf. (UEMCON), New York, NY, USA, Oct. 2019, 2005 and 2008, respectively. He started his career
pp. 280–285, doi: 10.1109/UEMCON47517.2019.8993089. with UniKL, in 2002. Apart from teaching and
[36] ResearchGate. FIGURE 7. The Mobile Net Architecture [15]: Mobile Net postgraduate supervision. He is currently the
is Built on. Accessed: Aug. 18, 2024. [Online]. Available: https://www. Deputy President in charge of the academic and
researchgate.net/figure/The-MobileNet-architecture-15-MobileNet-is- technology of Universiti Kuala Lumpur (UniKL).
built-on-depthwise-separable-convolutions_fig5_352621942 He is active in software project consultation and development in business
[37] J. Kennedy, ‘‘Swarm intelligence,’’ in Handbook of Nature-Inspired and applications, enterprise resource planning (ERP), and customer relation
Innovative Computing, A. Y. Zomaya, Ed., Boston, MA, USA: Kluwer, management (CRM). His research interests include cyber security, the IoT
2006, pp. 187–219, doi: 10.1007/0-387-27705-6_6. application, the IoT security, big data analytics, and SDN.

177158 VOLUME 13, 2025

A. Raza et al.: Enhancing Medical Image Classification Through PSO-Optimized Dual Deterministic Approach

AHMAD SHAHRAFIDZ BIN KHALID received MAZLIHAM MOHD SU’UD received the mas-
the Ph.D. degree in the security and privacy of ter’s degree in electrical and electronics engineer-
IoT from ENSEEIHT, Toulouse, France. He is ing from the University of Montpellier, in 1993,
currently an Associate Professor with the Uni- and the Ph.D. degree in computer engineer-
versity of Kuala Lumpur, Malaysia. His research ing from Université de La Rochelle, in 2007.
interests include the Internet of Things (IoT), From 2013 to 2020, he was the President/CEO of
ambient assisted living (AAL), and radio fre- Universiti Kuala Lumpur, Malaysia. Since 2020,
quency identification (RFID). he has been the President/CEO of Multimedia
University, Malaysia. He has vast experience in
publishing papers in high-quality international sci-
entific journals and conference proceedings. He has numerous years of
experience in the industrial and academic fields.
MUHAMMAD MANSOOR ALAM received the
M.S. degree in system engineering, the M.Sc.
degree in computer science, the Ph.D. degree in
computer engineering, and the Ph.D. degree in
electrical and electronics engineering in France,
U.K., and Malaysia, respectively, and the Trè s
Honorable degree (Hons.) from Université de La
Rochelle. He was the Associate Dean of CCSIS
and the Head of the Department of Mathematics, FOUZIA NOOR received the M.B.B.S. and R.M.P. degrees from the
the Department of Statistics, and the Computer Combined Military Hospital, Karachi. She is currently a Registered Med-
Science Department, IoBM, Pakistan. He is currently associated with Riphah ical Specialist and Radiologist with Combined Military Hospital. She has
International University, Islamabad. He is a Professor of computer science. authored numerous research papers in the fields of medical sciences, with a
He is also working as an Adjunct Professor with Multimedia University, focus on both invasive and non-invasive diagnostic techniques.
Cyberjaya, Malaysia.

VOLUME 13, 2025 177159