Experimental Evaluation of Synthetic GAN based Data Augmentation Scheme to Predict Improved Rare Diseases from Medical Imaging

Medical imaging can pose a vital challenge to rare disease detection because of the scarcity of medical imaging data and imbalance in the classes. This paper presents a synthetic augmentation framework, Conditional Generative Adversarial Network (cGAN), which can be used to solve these problems and enhance the work of diagnostic systems in various categories of rare diseases. Open datasets of MRI, CT, and X-ray modalities that were obtained and published in Kaggle were curated, and annotated classes of muscular dystrophies, rare brain lesions, and pediatric skeletal anomalies. It implemented a piping system that consists of preprocessing, synthetic image generation via GANs and the downstream CNN classifier (ResNet50, DenseNet121, EfficientNet). The outcomes prove that GAN-enhanced datasets outperform in classification and recall significantly. In particular, the ResNet50 has an accuracy of 91.4 that is higher compared to real-only (84.1) and classical augmented (86.9) models. Optimal performance was also observed at a synthetic to real ratio of 5: 1 as found out in ablation studies. The FID (36.3), Inception Score (3.78), and SSIM (0.76) quantitative quality measurements were used to determine the realism and diversity of synthetic samples. The article demonstrates the potential of synthetic augmentation to enhance the ability to detect underrepresented pathologies and offers a repeatable design of improving rare disease classifiers. The approach can be applied in the medical fields with limited labeled data but early diagnosis is a critical factor.