Sensitivity of Generative Adversarial Network Augmentation on Brain Tumor Detection Using Convolutional Neural Network
Students & Supervisors
Student Authors
Supervisors
Abstract
Medical imaging datasets for brain tumor detection are often limited, which can hinder the performance of deep learning models. Generative Adversarial Networks (GANs) have emerged as a promising approach to augment such datasets. In this study, we investigate the effect of varying proportions of GAN-generated and real MRI images on the performance of a Convolutional Neural Network (CNN) for classifying healthy and tumorous scans. Synthetic images were generated using a DCGAN and combined with real images at different ratios to train a custom CNN. The model was evaluated on an independent real-world test set. Results demonstrate that adding a small fraction of GAN-generated images (e.g., 10% GAN, 90% real) enhances model performance, achieving 95.2% test accuracy with precision, recall, and F1-score exceeding 95%. However, higher proportions of synthetic data led to a gradual decline in performance, indicating potential overfitting to GAN artifacts. These findings highlight the utility of GAN-based augmentation for small datasets, while emphasizing the importance of maintaining a balanced mix of real and synthetic data to ensure robust generalization.
Keywords
Publication Details
- Type of Publication:
- Conference Name: IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health 2025 (BECITHCON 2025)
- Date of Conference: 29/11/2025 - 29/11/2025
- Venue: Eastern University, Dhaka, Bangladesh
- Organizer: IEEE Bangladesh section and IEEE Engineering in Medicine and Biology Society Bangladesh