Stroke Classification with Two-Dimensional Convolutional Neural Networks: Traditional vs. Generative Adversarial Network-based Data Augmentation for Uneven Class Distribution

Estine, kumar and Deshant, Singh and Anuraganand, Sharma and Surya, Prakash (2025) Stroke Classification with Two-Dimensional Convolutional Neural Networks: Traditional vs. Generative Adversarial Network-based Data Augmentation for Uneven Class Distribution. Stroke Classification with Two-Dimensional Convolutional Neural Networks: Traditional vs. Generative Adversarial Network-based Data Augmentation for Uneven Class Distribution, 36 (1685). pp. 1-13. ISSN 2576-9898

Text - Accepted Version Download (1MB)

Abstract

Uneven class distribution in medical data sets such as stroke data, presents a substantial challenge to the classification model, resulting in suboptimal performance and a biased model toward the majority class. This research analyzes the potential of GAN-based data augmentation over traditional data augmentation methods on a 2-dimensional Convolutional Neural Network to improve sampling. GAN-based data augmentation is used to effectively explore the solution space, and these methods generate synthetic samples to balance the class distribution. The integration of these approaches aims to improve the model’s robustness and predictability. Experimental analysis shows promising results in classification accuracy, showing the potential of the GAN-based data Augmentation techniques on the proposed CNN model for dealing with uneven class distribution. Deep Convolutional - Generative Adversarial Network (DCGAN), along with Image-Based ISMOTified-GAN (iSMOTified GAN) and Image-Based Markov Chain Monte Carlo (iMCMC-GAN), is compared to classic data augmentation techniques. The positive results demonstrate the potential of GAN- based data augmentation methods as viable approaches to resolving the issues associated with imbalanced datasets. While all augmentation techniques showed improvements in classification accuracy, DCGAN performed the best, achieving an accuracy of 98.01%. The results of this study provide valuable insights into the potential of GAN-generated synthetic data in enhancing classification, addressing both data limitations and privacy concerns.

Item Type:	Journal Article
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Divisions:	Faculty of Science, Technology and Environment (FSTE) > School of Computing, Information and Mathematical Sciences
Depositing User:	Anuraganand Sharma
Date Deposited:	15 Dec 2025 00:18
Last Modified:	15 Dec 2025 00:18
URI:	https://repository.usp.ac.fj/id/eprint/15179

Actions (login required)

View Item