GAN-BASED MEDICAL IMAGE GENERATION FOR RARE PATHOLOGIES USING TRANSFER LEARNING AND RADIMAGENET WEIGHTS

Abstract

Incorporating Artificial Intelligence into medical imaging has opened new avenues for addressing the longstanding challenge of class imbalance within diagnostic datasets, most specifically in mammography, where malignant examples are underrepresented. This work presents a GAN-based framework specifically designed for generating high-fidelity, low-population pathological classes of synthetic mammography images, thereby enhancing the availability of data for improving diagnostic model learning and generalization. The presented method comprises a dual-branch discriminator system, with one branch utilizing a DenseNet121 network pretrained on RadImageNet dataset to extract domain-relevant features. A Wasserstein GAN with Gradient Penalty (WGAN-GP) is utilized throughout the entire framework to provide a stable mode for adversarial learning and address issues such as mode collapse. The CBIS-DDSM dataset served as the basis for all experiments carried out, and images were preprocessed for standardized dimensions and further subjected to data augmentation methods for enhancing generalization. Realism and diversity were evaluated for the synthetic images using quantitative measures like the Kernel Inception Distance (KID), Fréchet Inception Distance (FID), Learned Perceptual Image Patch Similarity (LPIPS), and Multi-Scale Structural Similarity (MS-SSIM). The results confirmed that the optimal balancing between realism and diversity was realized using the value of the gradient penalty weight of λ = 3.0 and was the optimum across the remainder for the majority of the measures, with the KID attaining 0.1765 and FID attaining 179.35 upon convergence. These results establish the value of incorporating radiology-focused pretrained models within GAN structures and indicate how adjusting the gradient penalties facilitates balancing the realism and diversity trade-off in synthetic medical imaging.