A Hybrid MLP-CNN Generator for Medical Image Synthesis with WGAN-GP: Application to Breast Cancer Imaging

Abstract

Medical imaging serves as a powerful instrument for diagnosing and treating various diseases; however, the development of robust deep-learning models has been stifled by the problems of scarcity as well as imbalance in the availability of good-quality annotated datasets. Classic data augmentation techniques, such as transformation techniques and intensity manipulations, are also rather limited and do not reflect the complex structural variations inherent in medical images. In the proposed study, an attempt was made to fill this gap by exploring the use of image augmentation via Generative Adversarial Networks (GANs), particularly focusing on the Wasserstein GAN with Gradient Penalty (WGAN-GP). Herein, the authors propose a framework based on WGAN-GP, which supports training stability while avoiding analog GAN disadvantages like mode collapse or gradient vanishing. A model will even be able to access medical images as input data, as well as qualitative and quantitative evaluations about the likeness and clinical relevance of images generated from it. The validation process involves Fre´chet Inception Distance (FID), Structural Similarity Index (SSIM), Mean Squared Error (MSE), and texture analysis to compare synthetic with real ones. There is also an expert-based visual assessment of generated images to evaluate the applicability of image generation for end-use within the medical context. Results indicate that GAN-based augmentation is effective in increasing data diversity and creates images almost indistinguishable from real medical scans yet retaining anatomical components. These implications emphasize the potentiality of GANs addressing data scarcity challenges for deep learning applications in health. It also addresses possible ethical and regulatory issues raised in relation to synthetic medical data, such as data privacy, fairness, and clinical validation in the broader context of regulatory and ethical challenges. This research contributes to medical image synthesis by realizing that GAN-based augmentation is likely to be an effective means for improving deep learning performance in medical diagnostics. The methodology opens the door for future studies interested in the integration of generative models into clinical workflows to improve diagnostic accuracy and model generalizability in real-world applications.