Data Augmentation for MRI Brain Tumor Classification Using Fine-Tuned Stable Diffusion and a Discriminator

Abstract

Artificial intelligence significantly enhances the analysis of complex medical images, par- ticularly MRI scans, by enabling accurate classification and early diagnosis. However, the scarcity and imbalance of medical imaging datasets, often due to rare diseases or lim- ited access to high-quality equipment, lead to biased models and reduced classification accuracy. This study introduces DreamDiffGAN, a hybrid framework that integrates a fine-tuned Stable Diffusion model with DreamBooth and a GAN-inspired discrimina- tor to generate anatomically realistic synthetic MRI images. By combining the gener- ative capabilities of Stable Diffusion with a discriminator to ensure image fidelity, the model produces high-quality images that closely mimic real MRI scans, even with limited training data. The proposed approach enhances dataset diversity, mitigates overfitting, and improves classification performance for brain tumor detection. Experimental results demonstrate that DreamDiffGAN outperforms traditional augmentation and standalone DreamBooth methods, achieving a classification accuracy of 96.67% and a recall of 93.33% on a ResNet-18 classifier, significantly reducing false negatives critical for clinical applica- tions. This framework offers a scalable solution for data-scarce medical imaging contexts, with potential applications beyond MRI to other modalities.