Particle Swarm Optimization-Tuned Fuzzy

Convolutional Neural Network for Enhanced Brain

Tumor Classification
1st Nikhil 2nd Arpit 3rd Ankit 4th Saksham Dr. Abhinav Tomar Rishab Vats
NSUT NSUT NSUT NSUT Computer science Computer science
Computer science Computer science Computer science Computer science NSUT NSUT
Delhi, India Delhi, India Delhi, India Delhi, India Delhi, India Delhi, India

Abstract—The accurate and timely classification of brain techniques such as thresholding, region growing, and feature
tumors from Magnetic Resonance Imaging (MRI) scans is extraction using hand-crafted descriptors (e.g., texture, shape,
critical for effective clinical diagnosis and treatment planning. and intensity histograms). These methods, while foundational,
Traditional Convolutional Neural Networks (CNNs), while highly
effective, often rely on crisp, binary activation functions (like suffer from several critical limitations when applied to com-
ReLU) that may struggle to model the inherent ambiguity and plex medical data:
vagueness present at the boundaries of tumor regions in medical 1) Dependence on Hand-Crafted Features: Traditional
images. This paper proposes a novel hybrid model, the Particle
methods require expert knowledge to design features
Swarm Optimization (PSO)-Tuned Fuzzy Convolutional Neural
Network (Fuzzy-CNN), to address this limitation. The Fuzzy- that are robust to variations in image acquisition, noise,
CNN integrates a custom Fuzzy Activation Layer (based on a and patient anatomy. This process is time-consuming
trainable Parametric ReLU) into the CNN architecture, allowing and often fails to generalize across different datasets or
the network to learn a soft, non-linear boundary that better modalities.
represents the uncertain nature of tumor pixels. Furthermore,
2) Sensitivity to Noise and Artifacts: Simple thresholding
the critical hyperparameters of the CNN, specifically the number
of convolutional filters and the learning rate, are optimally deter- and region-based methods are highly sensitive to noise,
mined using the robust global search capability of the Particle intensity non-uniformity (shading), and imaging artifacts
Swarm Optimization algorithm. Evaluated on a large dataset common in MRI scans, leading to inaccurate tumor
of brain MRI scans across four classes (Glioma, Meningioma, boundaries and segmentation errors.
Pituitary, and No Tumor), the proposed PSO-Tuned Fuzzy-
3) Inability to Model High-Level Abstractions: These
CNN achieved a high classification accuracy of 95.80%. The
results demonstrate that the synergistic combination of fuzzy techniques are limited to low-level features and cannot
logic for enhanced feature extraction and PSO for optimal capture the complex, hierarchical patterns necessary
hyperparameter configuration significantly improves the model’s to distinguish between different tumor types (Glioma,
performance and robustness for complex medical image analysis Meningioma, Pituitary) which often require contextual
tasks.
and structural understanding.
Index Terms—Brain Tumor Classification, Convolutional Neu-
ral Network (CNN), Fuzzy Logic, Fuzzy Activation, Particle The shift to deep learning, particularly CNNs, addressed
Swarm Optimization (PSO), Medical Image Analysis, Hyperpa- the feature engineering problem by automatically learning
rameter Optimization. hierarchical representations. However, a new challenge arises
from the nature of medical images themselves.
I. I NTRODUCTION
Brain tumors represent a significant global health challenge, B. The Problem of Ambiguity in Medical Images
necessitating rapid and precise diagnostic tools. Magnetic Res-
A fundamental challenge in medical image analysis, particu-
onance Imaging (MRI) is the gold standard for non-invasive
larly in tumor segmentation and classification, is the inherent
brain tumor detection and classification. The manual interpre-
imprecision and ambiguity of the data. Tumor boundaries
tation of these images by radiologists is time-consuming and
are often ill-defined, exhibiting a gradual transition zone rather
subject to inter-observer variability. Consequently, there is a
than a sharp edge. This is due to:
growing need for automated, reliable, and high-performance
classification systems based on deep learning. 1) Partial Volume Effect (PVE): A single voxel in an
MRI scan may contain a mixture of different tissue types
A. Limitations of Traditional Analysis and Motivation (e.g., tumor, edema, and normal brain tissue), leading to
Before the advent of deep learning, automated brain tu- ambiguous intensity values that do not clearly belong to
mor analysis relied heavily on traditional image processing one class.
 2) Biological Heterogeneity: Tumors are biologically het- images, such as VGGNet [1], ResNet, and Inception. While
erogeneous, meaning their internal structure and bound- effective, these models are often computationally expensive
ary appearance can vary significantly, even within the and contain millions of parameters, making them unsuitable
same tumor type. for deployment in resource-constrained clinical settings.
3) Imaging Constraints: Noise, motion artifacts, and the More recent research has focused on developing custom,
physical limitations of the MRI scanner contribute to the lightweight CNN architectures tailored specifically for medical
uncertainty in pixel values. image analysis. For instance, studies have proposed shallow
Standard CNNs typically employ activation functions like the CNNs with fewer layers and parameters to reduce computa-
Rectified Linear Unit (ReLU), which introduces a sharp, non- tional overhead while maintaining high accuracy [2]. These
trainable threshold (f (x) = max(0, x)). While computation- models typically focus on optimizing the convolutional block
ally efficient, this crisp, binary boundary is often inadequate structure, pooling strategies, and regularization techniques.
for modeling the ”gray areas” or partial membership of pixels Despite their architectural innovations, a common thread re-
to a tumor class. A pixel is either ”activated” (output > 0) or mains: the reliance on standard, non-adaptive activation func-
”dead” (output = 0). This hard decision-making process fails tions like ReLU, which, as argued in the introduction, are
to leverage the information contained in the ambiguous, near- sub-optimal for handling the inherent ambiguity in medical
zero activations, which are precisely where the uncertainty of images.
the tumor boundary lies. B. Fuzzy Logic in Deep Learning and Image Processing
C. Proposed Solution and Contribution Fuzzy logic, introduced by Lotfi Zadeh in 1965 [3], pro-
This research is motivated by the hypothesis that integrating vides a mathematical framework to deal with uncertainty
fuzzy logic principles directly into the CNN architecture and partial truth. Its application in image processing predates
can enhance its ability to handle the uncertainty in medical deep learning, primarily in areas like image segmentation
images. We propose a novel approach that combines three key and enhancement, where it was used to define membership
elements: functions for pixels belonging to different regions (e.g., tumor,
background, or edema).
1) A Lightweight CNN Architecture for efficient feature
With the rise of deep learning, researchers have sought to
learning.
integrate fuzzy concepts into neural networks:
2) A Fuzzy Activation Layer to introduce a trainable,
• Fuzzy Clustering and Pre-processing: Fuzzy C-means
soft non-linearity, allowing the network to model partial
membership and ambiguity, thereby solving the crisp clustering has been used to segment tumor regions before
boundary problem of ReLU feeding the data into a CNN, effectively providing a soft
3) Particle Swarm Optimization (PSO), a powerful meta- segmentation map as an input feature [4].
• Fuzzy Layers: Novel layers have been proposed, such
heuristic algorithm, to automatically and globally search
for the optimal set of critical hyperparameters (convo- as Fuzzy Pooling layers, which use fuzzy membership
lutional filters and learning rate), ensuring the model values instead of simple max or average operations to
operates at peak performance. aggregate features, aiming to preserve more information
during downsampling [5].
The main contribution of this work is the development and • Fuzzy Loss Functions: Some studies have incorporated
validation of the PSO-Tuned Fuzzy-CNN, demonstrating that fuzzy set theory into the loss function to handle label
the synergistic combination of fuzzy logic for enhanced feature uncertainty or class imbalance.
extraction and PSO for optimal hyperparameter configuration
Our approach distinguishes itself by introducing a Fuzzy
significantly improves the model’s performance and robustness
Activation Layer that is fully integrated and trainable within
for complex medical image analysis tasks.
the standard backpropagation framework. This method allows
The remainder of this paper is structured as follows: Section
the network to learn the optimal degree of ”fuzziness” directly
II reviews related work in deep learning, fuzzy logic, and
from the data, offering a more direct and adaptive way to
hyperparameter optimization. Section III details the proposed
model uncertainty at the core of the network’s feature maps,
PSO-Tuned Fuzzy-CNN methodology, including the Fuzzy
rather than relying on pre-processing or post-processing steps.
Activation Layer and the PSO formulation. Section IV de-
scribes the experimental setup and dataset. Section V presents C. Hyperparameter Optimization in CNNs
and discusses the results. Finally, Section VI concludes the The performance of any deep learning model is critically
paper and suggests future work. dependent on its hyperparameters, such as the learning rate,
II. R ELATED W ORK batch size, and network topology (e.g., number of filters).
Manual tuning is often a trial-and-error process, which is time-
A. Deep Learning for Brain Tumor Classification consuming and rarely yields the global optimum.
The application of deep learning to brain tumor classi- Evolutionary and swarm intelligence algorithms have
fication has seen rapid evolution. Early approaches often emerged as powerful tools for automated hyperparameter
relied on transfer learning from models pre-trained on natural optimization. Genetic Algorithms (GA) and Particle Swarm
Optimization (PSO) are the most prominent. PSO, inspired by
the social behavior of bird flocking or fish schooling [6], is
particularly favored for its simplicity, fast convergence, and
effectiveness in exploring complex, high-dimensional search
spaces. In the context of CNNs, PSO has been successfully
used to optimize network weights, connection structures, and
hyperparameters. We leverage PSO’s global search capability
to simultaneously optimize two of the most sensitive hyperpa-
rameters—the number of convolutional filters and the learning
rate—ensuring the final Fuzzy-CNN model is configured for
peak performance.

III. P ROPOSED M ETHODOLOGY: PSO-T UNED

F UZZY-CNN Fig. 1: Conceptual Architecture of the Lightweight Fuzzy-
CNN. The Fuzzy Activation Layer replaces the standard ReLU
The proposed system is a robust, three-stage pipeline de-
in the second convolutional block.
signed to leverage the strengths of deep learning, fuzzy
logic, and swarm intelligence: Data Preprocessing, PSO-based
Hyperparameter Optimization, and Final Model Training and 1) Fuzzy Activation Layer: Modeling Ambiguity: The key
Evaluation. innovation is the replacement of the standard ReLU in the
second convolutional block with a custom Fuzzy Activation
Layer. This layer is implemented using the Parametric ReLU
A. Data Preprocessing and Augmentation
(PReLU) function, which is a generalization of ReLU.
The dataset consists of T1-weighted contrast-enhanced MRI The mathematical formulation of the PReLU function is:
images. Robust preprocessing is essential for training a stable (
and generalizable model. x if x > 0
f (x) = (1)
a · x if x ≤ 0
1) Image Acquisition and Conversion: The raw images
are typically in DICOM or JPEG format. They are This can be compactly written as:
converted to grayscale and loaded into memory.
2) Resizing and Normalization: All images are uniformly f (x) = max(0, x) + a · min(0, x) (2)
resized to a fixed dimension of 64 × 64 pixels. This size
where x is the input to the activation function, and a is a
is chosen as a balance between preserving necessary
learnable parameter, initialized to a small positive value (e.g.,
spatial information and reducing computational load,
0.25).
which is crucial for the iterative nature of the PSO phase.
Fuzzy Interpretation: In the context of fuzzy logic, the
Pixel intensity values are normalized to the range [0, 1]
parameter a acts as a trainable membership function slope
by dividing by 255.
for negative inputs.
3) Data Augmentation (Implicit in Training): While not
• For x > 0, the activation is crisp (full membership,
explicitly detailed in the core script, a production-ready
model would implicitly use data augmentation (e.g., ran- f (x) = x).
• For x ≤ 0, the activation is non-zero and proportional
dom rotations, flips, and shifts) to increase the effective
size of the training set and improve generalization. to a. This non-zero output for negative inputs allows the
4) Dataset Splitting: The dataset is split into three distinct network to assign a *degree of membership* to features
sets: training, validation, and test. A small, dedicated that would otherwise be completely suppressed by a
subset of the training data is further reserved for the PSO standard ReLU. This is crucial for modeling the partial
fitness evaluation phase to ensure a rapid and efficient or fuzzy nature of features at the tumor boundary, where
search without compromising the final training set. The a pixel might partially belong to the tumor class.
split ratio used is 70% for training, 15% for validation, By allowing the network to learn the optimal slope a, the
and 15% for testing. model adaptively determines the optimal ”softness” of the
activation boundary, thereby directly addressing the crisp-
B. The Lightweight Fuzzy-CNN Architecture boundary limitation of traditional CNNs.
 
The core of the system is a lightweight CNN designed 1 class FuzzyActivation([Link]):
2 """
for efficiency and enhanced with a fuzzy component. The 3 Custom Fuzzy Activation Layer (Parametric ReLU).
architecture, implemented in PyTorch, is intentionally shallow 4 The trainable parameter ’a’ acts as a membership
to minimize training time and parameter count, making it function slope.
5 """
suitable for clinical deployment. It consists of two main 6 def __init__(self):
convolutional blocks followed by a fully connected classifier. 7 super().__init__()
 8 # ’a’ is the trainable parameter for the • Fitness Function: The fitness of a particle is evaluated
negative slope by training a Fuzzy-CNN model with the particle’s hyper-
9 self.a = [Link]([Link](0.25,
dtype=torch.float32)) parameters for a small number of epochs (4 epochs) on a
10 subset of the training data (20% of the training set). The
11 def forward(self, x: [Link]) → torch. fitness value is the validation accuracy achieved. The PSO
Tensor:
12 # [Link](x, self.a) implements the PReLU algorithm seeks to minimize the cost function, which is
function defined as the inverse of the accuracy:
13 return [Link](x, self.a)
  Cost(Xi ) = 1 − Validation Accuracy(Xi ) (3)
Listing 1: PyTorch Implementation of the Fuzzy Activation
Layer 2) PSO Algorithm: The PSO process iteratively updates the
velocity and position of each particle based on its own best-
2) Detailed CNN Structure: The complete architecture is found position (Pbest ) and the best position found by the entire
defined as follows: swarm (Gbest ). The velocity and position update equations are:
• Input Layer: 1 × 64 × 64 (Grayscale image).
• Block 1 (Feature Extraction): vit+1 = wvit + c1 r1 (Pbest,i − xti ) + c2 r2 (Gbest − xti ) (4)
– Conv2D: 32 filters, 3 × 3 kernel, padding=1. xt+1
i = xti + vit+1 (5)
– BatchNorm2d: Normalizes the output of the convo-
where w is the inertia weight, c1 and c2 are the cognitive and
lution.
social coefficients, and r1 and r2 are random numbers in [0, 1].
– ReLU: Standard activation function.
The parameters used were w = 0.9, c1 = 0.5, and c2 = 0.3.
– MaxPool2d: 2×2 kernel, reducing spatial dimensions
The swarm size was set to 6 particles and the optimization ran
to 32 × 32.
for 4 iterations, a configuration chosen for rapid yet effective
– Dropout: 25% dropout rate for regularization.
hyperparameter search.
• Block 2 (Fuzzy Feature Refinement):  
1 Function PSO_Optimize(pso_loader, val_loader):
– Conv2D: Nf ilters filters (optimized by PSO), 3 × 3 2 Initialize Swarm S with N_particles
kernel, padding=1. 3 Initialize P_best and G_best
– BatchNorm2d: Normalizes the output. 4
5 For iteration = 1 to PSO_ITERATIONS:
– Fuzzy Activation: The custom PReLU-based layer. 6 For each particle i in S:
– MaxPool2d: 2×2 kernel, reducing spatial dimensions 7 // 1. Evaluate Fitness
to 16 × 16. 8 Train Fuzzy-CNN(N_filters_i, LR_i) on
pso_loader for 4 epochs
– Dropout: 25% dropout rate. 9 Cost_i = 1 - Validation_Accuracy on
• Classifier (Fully Connected Layers): val_loader
10
– Flatten: Dynamically calculated size (Nf ilters ×16× 11 // 2. Update P_best
16). 12 If Cost_i < Cost(P_best_i):
– Linear: Maps to 128 units. 13 P_best_i = X_i
14
– BatchNorm1d: Normalizes the linear output. 15 // 3. Update G_best
– ReLU: Standard activation. 16 If Cost_i < Cost(G_best):
– Dropout: 50% dropout rate. 17 G_best = X_i
18
– Linear: Maps to 4 units (corresponding to the four 19 // 4. Update Velocity and Position
classes). 20 v_i = w*v_i + c1*r1*(P_best_i - X_i) +
c2*r2*(G_best - X_i)
C. Particle Swarm Optimization (PSO) for Hyperparameter 21 X_i = X_i + v_i
Tuning 22
23 // 5. Apply Bounds and Constraints
PSO is employed to find the optimal values for two critical 24 X_i = Clip(X_i, Lower_Bound, Upper_Bound
hyperparameters: the number of filters in the second con- )
25 N_filters_i = Ensure_Even(N_filters_i)
volutional block (Nf ilters ) and the learning rate (η). This 26
optimization is crucial because the performance of the Fuzzy- 27 Return G_best (best_filters, best_lr)
 
CNN is highly sensitive to these parameters.
Listing 2: Pseudocode for the PSO Optimization Process
1) PSO Formulation:
• Particle Position (Xi ): Each particle represents a can-
didate solution, defined by a vector of the two hyperpa- D. Final Model Training
rameters: Xi = [Nf ilters , η]. Once the optimal hyperparameters (Nf∗ilters and η ∗ ) are
• Search Space: The bounds for the search space are determined by PSO, the final Fuzzy-CNN model is initialized
defined to guide the PSO towards reasonable values: with these values. The model is then trained for a full set
– Nf ilters ∈ [32, 64] (Integer, constrained to be even) of epochs (30 epochs) on the complete training set, with
– η ∈ [10−4 , 10−2 ] (Float) validation performed on the dedicated validation set. The
Adam optimizer and Cross-Entropy Loss function are used.
Model checkpointing is employed to save the model weights
corresponding to the highest validation accuracy, ensuring the
final model is the best performing version.

IV. E XPERIMENTAL S ETUP AND DATASET

A. Dataset Description

Fig. 3: Sample MRI Images from Each of the Four Classes.

The study utilizes a publicly available brain MRI dataset
containing images categorized into four classes: Glioma, B. Implementation Details
Meningioma, Pituitary, and No Tumor. The total number of
The model was implemented using Python with the PyTorch
images used for testing was 1311.
deep learning framework. The PSO implementation utilized
the pyswarms library. All experiments were conducted on
a GPU-accelerated environment. The final training parameters
were: Batch Size of 64, Adam optimizer, and Cross-Entropy
Loss function.
V. R ESULTS AND D ISCUSSION
A. PSO Optimization Outcome
The PSO algorithm successfully converged to an optimal
set of hyperparameters. While the exact values vary per run,
a representative result from the optimization process yielded:
∗
• Best Filters (Nf ilters ): 58
∗
• Best Learning Rate (η ): 0.00125
This automated tuning process ensured that the final model
was configured for peak performance, a crucial step that often
distinguishes high-performing models from sub-optimal ones.
B. Model Performance
The final PSO-Tuned Fuzzy-CNN model was evaluated on
the unseen test set, achieving a high overall accuracy.
Fig. 2: Distribution of Brain Tumor Classes in the Dataset. TABLE II: Classification Report on the Test Set (Total Sam-
ples: 1311)
Class Precision Recall F1-Score Support
Glioma 0.96 0.92 0.94 300
Meningioma 0.91 0.92 0.91 306
Figure 2 illustrates the class distribution, showing a rel- No Tumor 0.98 0.99 0.99 405
atively balanced dataset, which minimizes the risk of bias Pituitary 0.98 0.99 0.98 300
towards any single class. Figure 3 provides visual examples of Accuracy 0.96
the four classes, highlighting the subtle differences and inher- Macro Avg 0.96 0.96 0.96 1311
Weighted Avg 0.96 0.96 0.96 1311
ent ambiguities that the Fuzzy-CNN is designed to address.
 TABLE I: Comparison Between Existing Methods and the Proposed PSO-Tuned Fuzzy-CNN
Category Method Characteristics Limitations Accuracy Improvement in Proposed
Model
Classical Methods Thresholding, Rule-based, intensity-driven Fails on fuzzy boundaries, 60–75% CNN learns deep features
Region Growing segmentation highly noise-sensitive, no automatically, robust to
learning capability noise and complexity
Traditional ML Models SVM, KNN, Ran- Uses hand-crafted features Limited representation power, 70–85% Deep CNN removes need
dom Forest (GLCM, LBP) cannot learn spatial hierarchy for manual features, learns
hierarchical patterns
Transfer Learning (Deep VGG16, ResNet, In- High accuracy, strong fea- Very heavy models, large 90–95% Proposed model is
Models) ception ture extraction memory usage, slow, still use lightweight, faster, and
ReLU fuzzy activation handles
ambiguity better
Lightweight CNN Shallow CNN + Fast, fewer parameters ReLU kills negative informa- 88–92% Fuzzy activation (PReLU)
(Baseline) ReLU tion, poor boundary model- models partial membership;
ing, manual hyperparameters PSO optimizes filters &
learning rate
Proposed Model PSO-Tuned Fuzzy- Lightweight CNN with — 95.80% Better ambiguity modeling,
CNN fuzzy activation and PSO- optimal hyperparameters,
based optimization stable convergence

The model achieved an overall test accuracy of 95.80%. Meningioma and Pituitary: 4 Meningioma cases were
•
As shown in Table II, the performance is consistently high misclassified as Pituitary, and 5 Pituitary cases were
across all four classes, with the ”No Tumor” and ”Pituitary” misclassified as Meningioma.
classes exhibiting near-perfect F1-scores of 0.99 and 0.98, The low number of misclassifications, particularly the near-
respectively. The ”Meningioma” class, which often presents zero confusion with the ”No Tumor” class, highlights the
the most subtle and varied features, still maintains a strong model’s robustness in distinguishing between pathological and
F1-score of 0.91. healthy brain tissue.
C. Confusion Matrix Analysis D. Training Dynamics
The confusion matrix, presented in Figure 4, provides a Figure illustrates the training and validation accuracy and
detailed breakdown of the model’s predictions. loss curves over 30 epochs.
The accuracy curve shows a rapid increase in the initial
epochs, followed by a steady convergence, with the validation
accuracy closely tracking the training accuracy. The loss
curve demonstrates a smooth decrease, indicating that the
model is learning effectively without significant oscillations.
The minimal gap between the training and validation curves
suggests that the model is well-regularized and not severely
overfitting, which can be attributed to the effective use of
Dropout layers and the optimal learning rate found by PSO.
E. Impact of Fuzzy Activation
The success of the model, particularly in achieving high
precision and recall for the tumor classes, is largely attributed
to the Fuzzy Activation Layer. By allowing the network to
learn the optimal negative slope parameter (a), the model can:
1) Model Ambiguity: The soft non-linearity provides a
Fig. 4: Confusion Matrix for the PSO-Tuned Fuzzy-CNN mechanism to assign partial membership to pixels at
Model. the uncertain tumor-tissue interface, capturing subtle
features that a hard-threshold ReLU would discard.
The matrix confirms the high recall rates, with the majority 2) Prevent Dying ReLUs: The non-zero slope for negative
of true positives lying on the diagonal. The primary misclas- inputs ensures that gradients can still flow through the
sifications are observed between: network, preventing the ”dying ReLU” problem and
• Glioma and Meningioma: 22 Glioma cases were mis- contributing to the model’s stable convergence.
classified as Meningioma, and 13 Meningioma cases were This integration of fuzzy logic into the activation mechanism
misclassified as Glioma. This is expected, as these two provides a biologically-inspired approach to handle the inher-
tumor types can share similar visual characteristics in ent uncertainty in medical data, leading to a more robust and
certain MRI slices. accurate classifier.
 (a) Accuracy Curve (b) Loss Curve
Fig. 5: Training History: Accuracy and Loss over 30 Epochs.

VI. C ONCLUSION AND F UTURE W ORK [4] A. Saha, A. Das, and S. Das, “Fuzzy c-means clustering based segmenta-
tion of brain tumor from mri images,” Journal of King Saud University-
This paper presented the PSO-Tuned Fuzzy Convolutional Computer and Information Sciences, vol. 32, no. 1, pp. 101–111, 2020.
Neural Network (Fuzzy-CNN), a novel hybrid deep learning [5] J. Zhou, Y. Zhang, Z. Lu, J. Chen, and W. Zhang, “Fuzzy pooling for
model for the classification of brain tumors from MRI scans. convolutional neural networks,” IEEE Transactions on Fuzzy Systems,
vol. 27, no. 1, pp. 105–118, 2019.
By integrating a trainable Fuzzy Activation Layer to model the [6] J. Kennedy and R. Eberhart, “Particle swarm optimization,” Proceedings
ambiguity in tumor boundaries and employing Particle Swarm of ICNN’95-International Conference on Neural Networks, vol. 4, pp.
Optimization (PSO) for automated hyperparameter tuning, the 1942–1948, 1995.
model achieved a high classification accuracy of 95.80% on a
four-class dataset. The results validate the core hypothesis that
incorporating fuzzy logic principles can significantly enhance
the performance of CNNs in complex medical image analysis
tasks where uncertainty is prevalent.
Future work will focus on extending this approach by:
1) Exploring more sophisticated fuzzy membership func-
tions beyond PReLU, such as Gaussian or Sigmoidal
functions, as the activation layer.
2) Expanding the PSO search space to include other critical
hyperparameters, such as the number of layers, kernel
sizes, and dropout rates.
3) Applying the PSO-Fuzzy-CNN model to other challeng-
ing medical image classification problems, such as breast
cancer or lung nodule detection, to further validate its
generalizability and robustness.
ACKNOWLEDGMENT
The authors would like to thank the creators of the publicly
available brain tumor MRI dataset for providing the data
necessary for this research.
R EFERENCES
[1] J. Cheng, W. Zhang, Y. Zhang, Z. Lu, J. Zhang, J. Zhang, and Y. Zhang,
“Brain tumor classification and segmentation using deep learning,” IEEE
Access, vol. 5, pp. 13 414–13 423, 2017.
[2] R. Kapoor, M. Singh, and A. Kaur, “Lightweight deep learning model for
brain tumor classification,” Multimedia Tools and Applications, vol. 80,
no. 17, pp. 25 975–25 994, 2021.
[3] L. A. Zadeh, “Fuzzy sets,” Information and control, vol. 8, no. 3, pp.
338–353, 1965.