Synthetic MRI as an Alternative Training Resource for Brain Tumor Classification: A GAN-Based Study

Accurate classification of brain tumors from magnetic resonance (MR) imaging data very important for clinical decision support systems. However, the privacy issues encountered in MR data and the high cost of data collection processes make it difficult to develop artificial intelligence-based detection systems. Hence, synthetic data generation has emerged as a promising solution to mitigate data scarcity while preserving patient privacy. This study evaluates the feasibility of using synthetic MRI data as an alternative training resource and compares real, synthetic, and combined data for training scenarios under a unified experimental framework. Moreover, synthetic data generation is considered not only as a mechanism for performance enhancement but also as a potential approach for supporting privacy-sensitive medical imaging applications where access to real patient data may be restricted. In this paper, three experimental scenarios were considered: real MR images only, GAN-generated synthetic MR images only, and a hybrid dataset combining both. Transfer learning models were used and their classification success was evaluated using various metrics. The findings show that models perform better with synthetic images alone, while in some models, using original and synthetic data together can have a positive impact on performance. These results demonstrate that GAN-generated synthetic images can provide meaningful data diversity and may serve as an alternative training resource in medical image classification tasks, particularly in scenarios where access to real data is limited.