Leveraging Capsule Networks for Robust Brain Tumor Classification

and Detection in MRI Scans

Sandeep Shiraskar, Simon Vellandurai and Dominick Rizk

The Department of Electrical Engineering and Computer Science, The Catholic University of America,

Washington, D.C. 20064, U.S.A.

Keywords: Brain Tumor Detection, Capsule Networks, MRI Classification, Deep Learning, Medical Image Analysis,

Tumor Classification.

Abstract: Brain tumors are life-threatening conditions where early detection and accurate classification are critical for

timely and effective treatment. Misclassification or delayed identification of tumors can result in fatal

consequences. Current deep learning techniques, predominantly based on Convolutional Neural Networks

(CNNs), have demonstrated success in tumor detection but face limitations due to their inability to handle

diverse and extensive datasets effectively. Moreover, CNNs suffer from information loss in pooling layers,

leading to suboptimal performance in capturing global dependencies in MRI tumor images. To overcome

these challenges, we propose the use of a modified Capsule Network to address the limitations of CNNs.

Capsule Networks retain spatial hierarchies and dependencies, enabling improved performance in tumor

detection and classification tasks. Our approach achieves near-perfect classification accuracy across four

classes—pituitary, glioma, meningioma, and no tumor—using a diverse and augmented dataset. The dataset

comprises publicly available MRI images from Figshare, Sartaj, and Br35 collections, providing a robust

platform for evaluating model performance. Experimental results demonstrate that our method not only

achieves superior accuracy compared to existing techniques but also maintains its performance across a

broader range of data. These findings highlight the potential of Capsule Networks as a reliable and effective

solution for brain tumor classification tasks, paving the way for advancements in medical imaging and

diagnostic technologies.

1 INTRODUCTION

Brain tumors are among the most life-threatening

diseases, with early detection being critical for

effective treatment and improved survival rates.

Magnetic Resonance Imaging (MRI) is commonly

used for brain tumor detection due to its high-

resolution imaging capabilities. However, the

complexity of tumor classification from MRI scans

poses significant challenges. The vast amount of data

generated from MRI scans, combined with the

intricate and varied nature of brain tumors, makes

manual classification a time-consuming and error-

prone task. An efficient, automated system is

essential to aid medical professionals in diagnosing

and classifying brain tumors accurately and timely.

Convolutional Neural Networks (CNNs) have

become the leading deep learning architecture for

medical image classification, including brain tumor

detection. CNNs excel at extracting hierarchical

features from images and have achieved impressive

results in various computer vision tasks. However,

their application to medical image classification,

particularly in the context of brain tumors, is not

without limitations. One of the primary challenges of

CNNs is their inability to handle spatial relationships

between features effectively. Brain tumors, which

vary in shape, size, and location, require a model that

can retain and understand these spatial characteristics.

CNNs struggle to generalize across different

transformations or rotations of images, as they are

reliant on large datasets to account for such

variations. Unfortunately, in medical imaging, such

extensive datasets are often not available, and data

augmentation alone cannot overcome this issue.

Furthermore, the use of pooling layers in CNNs leads

to a loss of spatial resolution, which is detrimental

when it comes to tasks that require accurate location-

based classification, such as tumor detection.

1288

Shiraskar, S., Vellandurai, S. and Rizk, D.

Leveraging Capsule Networks for Robust Brain Tumor Classiﬁcation and Detection in MRI Scans.

DOI: 10.5220/0013323000003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1288-1296

ISBN: 978-989-758-737-5; ISSN: 2184-433X

While CNNs have been the go-to solution for

medical image classification, these limitations

highlight the need for more advanced techniques

capable of addressing the challenges in medical

imaging. Capsule Networks (CapsNets) have been

introduced as a potential solution. CapsNets are

designed to overcome the shortcomings of CNNs by

encoding both the presence and spatial orientation of

features, thus preserving important geometric

relationships. This ability to maintain spatial

hierarchies and relationships allows CapsNets to

perform better on tasks that involve complex image

structures, such as brain tumor classification. By

preserving the location and orientation of features,

CapsNets offer the potential to improve the accuracy

of tumor classification and overcome the

shortcomings of CNNs.

Despite the promising results of CapsNets,

challenges remain in their application to brain tumor

detection. Current CapsNet-based methods have

shown improvements over traditional CNN

approaches, but they still face issues related to

computational complexity and suboptimal

segmentation accuracy. Additionally, training

CapsNets on smaller, limited datasets can hinder their

ability to generalize to unseen variations in tumor

characteristics. These gaps underscore the need for

further research and refinement of CapsNet

architectures, along with the development of more

diverse and augmented medical image datasets, to

fully realize their potential in brain tumor

classification.

In this study, we present a modified Capsule

Network (CapsNet) model tailored for brain tumor

classification. Capsule Networks are designed to

address the limitations of traditional Convolutional

Neural Networks (CNNs) by leveraging capsules—

groups of neurons that output vectors representing

both the probability and spatial properties (pose) of

features. A key advantage of CapsNet is its ability to

recognize spatial relationships and part-whole

hierarchies, which enhances generalization across

transformed data.

Our model begins with standard convolutional

layers to extract lower-level features from the input

images. These features are then processed by a

custom Capsule Layer, which performs feature

detection by utilizing a weight matrix and

encapsulating these features as vectors. The Capsule

Layer uses a routing mechanism (such as dynamic

routing, though simplified in this implementation) to

route outputs from lower-level capsules to higher-

level ones, ensuring that the spatial relationships

between detected features are preserved.

The model’s output layer uses softmax activation

to classify images based on the output from the

capsule layer, enabling the network to learn complex

feature hierarchies and improve accuracy. The

network is trained using standard backpropagation,

with the training process monitored using validation

data over multiple epochs.

2 LITERATURE REVIEW

Recent studies have extensively compared popular

deep learning architectures such as CNN, VGG, and

ResNet for brain tumor classification, highlighting

both their strengths and limitations. For instance,

(Anwar, 2024) explored the use of CNNs for brain

tumor detection and segmentation, demonstrating the

model's strong capability for image classification.

However, the study also highlighted issues such as

feature loss during downsampling and the need for

more efficient feature representations (Anwar, 2024).

VGG and ResNet, although effective for image

classification tasks, face challenges in accurately

capturing fine-grained details necessary for precise

tumor segmentation. In particular, VGG, known for

its depth and simplicity, and ResNet, which utilizes

residual connections to avoid the vanishing gradient

problem, often struggle to handle complex spatial

relationships in medical images, such as in the case of

brain tumor segmentation (Ibrahim, 2023). These

findings underscore the need for improved models

that can better preserve the spatial hierarchies of

features in medical image data.

The drawbacks of CNNs and traditional

architectures have led to the development of alternative

models, notably Capsule Networks (CapsNet).

CapsNet, introduced by (Hinton, 2018), addresses

some of the shortcomings of CNNs, particularly in

terms of capturing spatial hierarchies and rotation

invariance. Capsule Networks preserve spatial

relationships between features by using "capsules,"

which are groups of neurons encoding both the

presence and orientation of objects. This approach

improves the model's robustness in recognizing

complex patterns and spatial features, making it

particularly suited for medical image analysis,

including brain tumor detection (Sabour, 2017).

Mathematically, the basic operation of a Capsule

Network is described by dynamic routing, where

capsules use a dynamic algorithm to route

information between layers. This allows capsules to

better maintain the spatial relationships between

features, overcoming the problem of information loss

seen in CNNs during max-pooling operations. Max-

Leveraging Capsule Networks for Robust Brain Tumor Classiﬁcation and Detection in MRI Scans

1289

pooling, commonly used in CNNs for dimensionality

reduction, can discard crucial spatial information,

which is a critical issue for tasks like tumor detection.

In contrast, CapsNet's dynamic routing algorithms

help preserve this information, leading to more

accurate representations of tumors in medical images

(Hinton et al., 2018).

For a better understanding of how CNNs, VGG,

and ResNet operate, we outline their basic

architectures and key operations below:

2.1 CNN (Convolutional Neural

Network)

A CNN consists of several layers: convolutional

layers for feature extraction, pooling layers for

dimensionality reduction, and fully connected layers

for classification. The basic operation of a CNN can

be described mathematically as:

𝑦=𝑅𝑒𝐿𝑈(𝑊⋅𝑥+𝑏)

Where 𝑥 is the input image, 𝑊 are the learned

weights, and 𝑏 is the bias term. The ReLU activation

function is applied elementwise to introduce non-

linearity.

2.2 VGG Network

VGG is a deep CNN architecture known for its simple

and consistent design. It uses small 3×3 convolutional

filters stacked on top of each other, followed by max-

pooling layers. While VGG is effective in feature

extraction, its deeper networks are prone to

overfitting on smaller datasets (Simonyan, 2014).

Figure 1: VGGNet proposed by (Simonyan, 2014).

2.3 ResNet (Residual Networks)

ResNet addresses the vanishing gradient problem

through the introduction of residual connections,

allowing gradients to flow more easily through deeper

networks. A basic ResNet block can be expressed as:

𝑦=𝐹(𝑥,{𝑊𝑖})+𝑥

Where 𝐹(𝑥, {𝑊𝑖}) is the residual function learned

by the network, and 𝑥 is the input to the block. This

formulation allows for more efficient training of deep

networks (He, 2016).

2.4 CapsNet (Capsule Networks)

Capsule Networks, while effective in overcoming

some of the challenges faced by traditional models,

come with their own set of challenges, particularly

computational complexity. The dynamic routing

algorithm, which is central to CapsNets, is

computationally expensive, making it less feasible for

real-time clinical applications with large datasets

(Chen, 2022). Nevertheless, recent research has

shown that CapsNet-based models outperform

traditional architectures in terms of segmentation

accuracy, especially in medical imaging tasks ( (Shi,

2020); (Zhang, 2021).

Traditional CNNs, while effective in feature

extraction, struggle with loss of spatial information

due to pooling layers, sensitivity to transformations,

and dependency on large datasets, which are often

unavailable in medical imaging (M. Sharma, 2024).

CapsNets overcome these drawbacks by encoding

features as vectors and utilizing dynamic routing,

enabling robust classification even with limited data

and preserving spatial hierarchies critical for medical

diagnosis (Afshar, 2019); (Raythatha,

2023).Innovations like the InceptionCapsule model,

integrating self-attention and Inception-ResNet

architectures, promise enhanced accuracy, but

challenges such as overfitting, computational

inefficiency, and limited transfer learning persist

(Sadeghnezhad, 2024)

Current studies emphasize the importance of using

whole-brain images rather than segmented regions to

retain locational context, an area where CapsNets

demonstrate superiority over state-of-the-art CNN

models like ResNet and DenseNet (Raythatha & V.,

2023). This trajectory positions CapsNets as a

promising solution for precise classification of brain

tumor subtypes,

The hybrid approach, combining CNNs and

CapsNets, has also shown promise. For example, (Hu,

2023) proposed a hybrid model that integrates CNNs

for initial feature extraction and CapsNets for

enhanced spatial representation, leading to state-of-

the-art results in brain tumor detection.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1290

Figure 2: Number of images per tumor type.

Taking reference from (Nitish Srivastava, 2014).

on dropout in deep neural networks our contribution

builds on these advancements by enhancing accuracy

through the integration of denser layers and

introducing capsule dropout, a novel regularization

method that refines the dropout process by

systematically dropping capsule vectors rather than

individual neurons. This approach ensures better

feature coherence and robustness, improving multi-

scale feature extraction capabilities. Furthermore, our

model expands the application of CapsNets to a

larger, more diverse dataset, achieving high accuracy

and demonstrating their scalability for complex

medical imaging tasks. This refined framework not

only overcomes limitations of traditional

architectures but also establishes a robust pathway for

integrating CapsNets into real-world diagnostic tools.

3 METHODOLOGY

The methodology implemented in this study involves

leveraging a modified Capsule Network for brain

tumor classification. The network is designed to

process MRI images by extracting multi-scale

features while preserving spatial hierarchies crucial

for accurate classification. Data preprocessing steps

include resizing, normalization, and augmentation

(e.g., rotation and intensity variation) to enhance

diversity and improve model robustness. A novel

capsule dropout mechanism was introduced to

selectively deactivate capsule vectors, improving

regularization and feature extraction. The model

employs a cross-entropy loss function and is

optimized using the Adam optimizer with an adaptive

learning rate. Performance evaluation is conducted

using metrics such as accuracy, precision, recall, and

F1-score. Training and testing were performed on

high-performance hardware using TensorFlow and

Keras frameworks. This approach emphasizes

enhanced feature extraction and robust learning for

improved classification performance.

3.1 Dataset

The dataset used for brain tumor classification in this

study is sourced from publicly available datasets from

Sartaj, br35, and Figshare. These datasets include

brain MRI images categorized into four classes:

pituitary, glioma, meningioma, and no tumor.

The data preparation process involves splitting the

dataset into three subsets: training, validation, and

testing, in a 70:15:15 ratio. Approximately 4,944

images are allocated to the training set, 1,059 images

to the validation set, and 1,311 images to the testing

set.

Image dimensions across the datasets vary, with

the training set having image widths ranging from

150 to 1920 pixels and heights from 168 to 1446

pixels. The validation set includes images with widths

between 150 and 1275 pixels and heights from 168 to

1427 pixels, while the testing set features images with

widths from 150 to 1149 pixels and heights between

168 and 1019 pixels.

3.2 Pre-Processing

During the preprocessing step, all images are resized

to a target size of 150x150 pixels and normalized to a

range of [0, 1]. This ensures that the images have

consistent dimensions and standardized pixel values,

which are suitable for input into a Capsule Network

model. The resizing and normalization are performed

using the resize_and_normalize_image()

function, which converts images to RGB mode,

resizes them using ImageOps.fit(), and

normalizes the pixel values by dividing by 255.

For data augmentation, only the training dataset

undergoes transformations such as random rotations,

width and height shifts, zoom, and horizontal flips.

This augmentation is done using the Keras

ImageDataGenerator class, which helps

increase the diversity of training data, allowing the

model to learn more robust features. The training

images are passed through the generator with the

flow() method, which outputs a 4D array, where the

first dimension represents the batch size (in this case,

1), followed by the image dimensions (150x150) and

the number of color channels (3 for RGB). The

augmentation is applied dynamically on-the-fly as the

model trains.

In terms of image array shape:

• After resizing and normalization, the images

have a shape of (150, 150, 3) for each image.

Leveraging Capsule Networks for Robust Brain Tumor Classiﬁcation and Detection in MRI Scans

1291

• When using augmentation, the images are

expanded to a 4D shape of (1, 150, 150, 3)

before being passed through the

augmentation process. This is necessary

because Keras expects a batch of images,

and the augmentation operation requires this

additional batch dimension. After

augmentation, the resulting image is

flattened back to a 3D format (150, 150, 3).

Thus, the training images are ready for model

input, with the augmentation applied only to the

training data, while validation and testing datasets are

resized and normalized without augmentation to

ensure unbiased performance evaluation.

Figure 3: Pre processed Image Samples from the

meningioma tumors.

Figure 4: Image Augmentation Samples from the

meningioma tumors.

3.3 System Model

Dynamic Routing: The core feature of CapsNet is

dynamic routing, where capsules in a lower layer

dynamically choose which capsules in the higher

layer they should send their outputs to. This routing

process helps in capturing complex spatial

relationships between different objects in an image.

Figure 5: NN after Dropout.

Figure 6: Neural Network.

Primary Capsules: In a CapsNet, primary capsules

are created from convolutional layers. These capsules

output a vector instead of a scalar, where each vector

component represents different properties (like

position or orientation) of the object.

Higher-Level Capsules: Higher-level capsules take

input from the primary capsules and are responsible

for grouping together the information about specific

object classes or parts of objects.

This is the workflow followed:

• Input Layer: The input image is passed

through an initial convolutional layer to

extract features.

• Primary Capsules: These features are then

passed to the primary capsule layer, which

uses a series of convolutional capsules to

generate feature vectors that describe

various parts of the object.

• Routing by Agreement: Capsules in the

primary layer are routed to higher-level

capsules based on their agreement. This

helps capture the spatial relationships

between objects and parts in the image.

• Dropout in Capsules: Dropout is applied in

this routing process, as well as within the

capsule outputs, to prevent co-adaptation of

neurons. This increases the model’s ability

to generalize.

• Output Layer: The final capsule layer

produces the final classification of the

image, typically using the length of the

capsule vector to indicate the probability of

the object being in that class.

• Loss Function: The loss function in

CapsNet is typically a margin loss, which

encourages the network to assign high

probability to the correct class while

penalizing the activation of incorrect

classes.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1292

3.3.1 Vector Inputs and Outputs

Unlike traditional neurons, capsules use vectors to

encode information. The vector length represents the

probability of the presence of a specific feature or

object, while the vector direction encodes specific

properties like orientation, position, or size of the

feature.

3.3.2 Capsule Computation Process

Each capsule in a lower layer generates a prediction

vector for every capsule in the next higher layer.

𝑢

∣

=𝑊



𝑢



𝑢



: The output of capsule iii.

𝑊



: A weight matrix that transforms the lower-

layer output to align with the expected input of the

higher-layer capsule.

𝑢

∣

: The prediction vector for capsule j.

The total input 𝑆



to capsule j is a weighted sum of

prediction vectors, calculated as:

𝑆



=∑



𝑐



𝑢

∣

Where 𝑐



is the coupling coefficient that

determines how much influence capsule i's prediction

has on capsule j. Coupling coefficients are updated

dynamically during the routing process.

After calculating the total input, the output vector

𝑣



is obtained using the squashing function:

𝑣



∣∣ 𝑠



∣∣



1+∣∣𝑠



∣∣



𝑠



∣∣ 𝑠



∣∣

This non-linear function ensures that the vector

length ∣∣ 𝑣



∣∣ is bounded between 0 and 1. Shorter

input vectors shrink toward 0, while longer vectors

shrink slightly below 1.

Figure 7: Capsule Network Architecture used in our method.

Figure 8: Accuracy plot.

Figure 9: Loss over epochs.

Coupling coefficients 𝑐



are computed using a

softmax function over the logits 𝑏



, were

𝑐



 (



)

∑

 (



)

Logits 𝑏



are initialized to 0 and iteratively

updated during the routing process. The softmax

ensures that the coefficients 𝑐



for each lower-layer

capsule i sum to 1, distributing its output among

higher-layer capsules.

3.3.3 Dynamic Routing Mechanism

Dynamic routing is an iterative process that refines

the coupling coefficients based on the agreement

between prediction vectors 𝑢

∣

and the actual

output𝑣



of the higher-layer capsule.

1: 𝑢



← inputs

2: 𝑊



← weights

3: 𝑢



← 𝑊



* 𝑢



4: 𝑏



← 0

5: for n iterations do

6: 𝑐



← P exp(𝑏



) / k exp(𝑏



)

7: 𝑠



← Σ k 𝑐



* 𝑢



8: 𝑣



← Σ k 𝑠



* 𝑠



1 + 2 * 𝑠



9: 𝑏



← 𝑏



+ 𝑢



· 𝑣



10: return 𝑣



Algorithm 1: The routing-by-agreement algorithm

(CapsNet).

Leveraging Capsule Networks for Robust Brain Tumor Classiﬁcation and Detection in MRI Scans

1293

The process begins by initializing the logits 𝑏



=0.

Coupling coefficients 𝑐



are computed using a

softmax function over the logits 𝑏



The total input 𝑆



and output 𝑣



for higher-layer

capsules are calculated, and logits 𝑏



are updated

based on the agreement:

𝑏



←𝑏



+𝑢

∣

⋅𝑣



Agreement increases if the prediction vector aligns

with the higher-layer output.

3.4 Training Configuration

The model was trained on a Linux Ubuntu system

running on WSL2. TensorFlow GPU was used with

CUDA 12.6, leveraging an NVIDIA RTX 4060 GPU

for accelerated computation. The training involved

500 epochs with a batch size of 64, optimizing the

model using the Adam optimizer. Input data consisted

of training and validation datasets, passed as

train_images and train_labels for training and

val_images and val_labels for validation.

Figure 10: Confusion Matrix.

Table 1: Training Configurations.

Parameter Value

Operating System Linux (Ubuntu)

GPU

NVIDIARTX 4060

CUDA Version 12.6

TensorFlow Version 2.17

Epochs 300

Batch Size 64

Optimizer Adam

4 RESULTS

The performance of the modified capsule network

model was evaluated on a set of MRI brain images,

and several key metrics were computed to assess its

effectiveness in tumor classification. The model

achieved a training accuracy of 95.3% and a

validation accuracy of 94.7% and a test accuracy of

95.66% over 300 epochs, demonstrating its strong

generalization capability and converging around 264

epochs. To further evaluate the model's performance,

we also computed precision and specificity for each

class of tumor, including glioma, meningioma,

pituitary tumors, and non-tumor cases. Precision

values for the glioma, meningioma, pituitary, and no

tumor categories were:

Table 2: Precision Specificity per Class.

Class Precision Specificity

Glioma 0.905 0.921

Meningioma 0.857 0.914

Pituitary 0.912 0.94

Notumor 0.945 0.953

90.5%, 85.7%, 91.2%, and 94.5%, respectively,

indicating the model's ability to correctly identify

positive cases with minimal false positives.

Specificity values for these classes were similarly

high, with the model achieving 92.1%, 91.4%, 94.0%,

and 95.3%, respectively, highlighting its

effectiveness in correctly classifying non-tumor cases

and avoiding false positives. Additionally, heatmaps

generated for the tumor images show that the model

was able to focus its attention on the relevant regions,

such as the central brain region for glioma, and

maximum for meningioma further confirming its

accuracy and reliability in identifying tumor

locations. The confusion matrix heatmap reveals the

performance of the model in classifying tumor types.

The diagonal elements indicate correct

classifications, with higher values signifying better

accuracy. Off-diagonal elements represent

misclassifications, suggesting areas where the model

struggles. While the model shows decent overall

performance, there's room for improvement in

distinguishing between similar tumor types, such as

meningioma and pituitary, and notumor and glioma.

Strategies like data augmentation, feature

engineering, model selection, and ensemble methods

can potentially enhance the model's accuracy.

Additionally, addressing class imbalance and

ensuring data quality are crucial for further

optimization. The following table summarizes

comparison with previous works:

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1294

Table 3: Comparison with previous works.

Method Accuracy Remarks

(Havaei, 2017) 90% Two-pathway architecture for

rain tumor MRI Images

(Adu) 95.5% CapsNet with a new activation

function for enhanced

accuracy

(Raythatha,

2023)

93.55% Capsule Networks on whole

rain MRIs

(Goceri, 2020) 92.8% Capsule Networks

(Afshar, 2019) 90.89% Capsule Networks

Our Method 95.66% CapsNet with Regularized

Dropout

5 CONCLUSIONS

In this study, we demonstrated the effectiveness of a

modified capsule network (CapsNet) for brain tumor

classification using MRI images. Our approach

achieved an accuracy of 92.8%, outperforming several

traditional and advanced methodologies. The results

emphasize the ability of CapsNet to capture spatial

hierarchies and maintain the integrity of geometric

features crucial for accurate tumor classification.

When compared to other works in the field, such

as those by Raythatha and V. M. (2023), Goceri

(2020), and Afshar et al. (2019), our model shows

competitive accuracy levels, demonstrating the

promise of CapsNet in medical image processing. For

instance, Goceri (2020) achieved an accuracy of

92.65%, while Afshar et al. (2019) reported 90.89%.

These results suggest that CapsNet offers a reliable

and efficient classification approach, similar to or

better than traditional CNN models.

While the current findings are promising, there is

potential for further improvements. Future work will

include extending this approach by incorporating

segmentation techniques to enhance the model's

ability to delineate tumor boundaries and improve

overall classification accuracy. Additionally, we plan

to explore the integration of hybrid models that use

self-attention mechanisms, which have demonstrated

significant potential in other domains.

As deep learning techniques in medical imaging

continue to evolve, particularly through advanced

architecture such as hybrid self-attention models and

refined segmentation methods, we anticipate that

these improvements will pave the way for more

accurate and robust tumor detection systems.

ACKNOWLEDGEMENTS

The authors would like to express sincere

appreciation to Burns Faculty Fellowship and the

ORAU Ralph E. Powe Junior Faculty Enhancement

Award, both of which provided crucial resources and

opportunities to advance this work.

REFERENCES

Adu, A. e. (n.d.). Brain Tumor Classification Using Capsule

Networks with Enhanced Activation Functions. MDPI.

Afshar, P. P. (2019). Capsule networks for brain tumor

classification based on MRI images and coarse tumor

boundaries. ICASSP 2019 – 2019 IEEE International

Conference on Acoustics, Speech and Signal

Processing (ICASSP), (pp. 1368-1372).

Anwar, F. e. (2024). CNN-based brain tumor detection and

segmentation. IEEE Xplore.

Chen, X. e. (2022). Challenges of Capsule Networks in

medical image analysis. Journal of Healthcare

Engineering, 1-9.

Fatihah, B. A. (2023). Comparison review on brain tumor

classification and segmentation using convolutional

neural network (CNN) and capsule network.

International Journal of Advanced Computer Science

and Applications,, 14(4).

Goceri, E. (2020, April). CapsNet topology to classify

tumours from brain images and comparative evaluation.

IET Image Processing: Volume 14, Issue 5, pp. 882-

889.

Havaei, M. e. (2017). Brain Tumor Segmentation with

Deep Neural Networks. Medical Image Analysis, 35,

123-134.

Hayat, M. A. (2023). Early detection of brain tumor using

capsule networK. International Journal for

Multidisciplinary Research (IJFMR), 8(2), 1-7.

He, K. e. (2016). Deep Residual Learning for Image

Recognition.

Hinton, G. e. (2018). Capsule Networks. Nature, 529(7587),

1-19.

Hu, Y. e. (2023). Hybrid CNN and CapsNet model for brain

tumor detection. IEEE Transactions on Biomedical

Engineering,, 70(7), 2234-2243.

Ibrahim, M. &. (2023). Challenges of VGG and ResNet in

brain tumor segmentation. BMC Medical Informatics.

Kumar, M. R. (2023). An improved segmentation method

for brain cancer using capsule neural networks.

ICTACT Journal on Image & Video Processing, 13(4),

2577-2584.

M. Sharma, S. B. (2024). A Review of Deep Learning

Models for Brain Tumor Prediction Using MRI Images.

Asia Pacific Conference on Innovation in Technology

(APCIT), (pp. 1-6). Mysore.

Nitish Srivastava, G. H. (2014). Dropout: A Simple Way to

Prevent Neural Networks from Overfitting. JMLR,

15(56):1929−1958.

Raythatha, Y. &. (2023). A paradigm shift in brain tumor

classification: Harnessing the potential of capsule

networks. IEEE 2nd International Conference on Data,

Decision and Systems (ICDDS), (pp. 1-6).

Leveraging Capsule Networks for Robust Brain Tumor Classiﬁcation and Detection in MRI Scans

1295

Sabour, S. e. (2017). Dynamic Routing Between Capsules.

NeurIPS.

Sadeghnezhad, E. &. (2024). InceptionCapsule: Inception-

Resnet and CapsuleNet with self-attention for medical

image classification. arXiv. https://arxiv.org/abs/2402.

02274.

Shi, L. e. (2020). CapsNet-based brain tumor segmentation.

IEEE Transactions on Medical Imaging, 39(2), 541-

550.

Simonyan, K. &. (2014). Very Deep Convolutional

Networks for Large-Scale Image Recognition.

arXiv:1409.1556.

Zhang, X. e. (2021). Attention-enhanced Capsule Networks

for brain tumor segmentation. BMC Medical Imaging,

21(1), 1-10.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1296