Hybrid GAN-Augmented Multi-Modal Medical Imaging Framework with Squeeze-and-Excitation CNN for Robust Classification

Abstract

A hybrid GAN architecture was designed to solve multi-modal medical image synthesis and disease classification by combining key architectural principles from DC-GAN, Conditional GAN, and SR-GAN, thereby enhancing training stability, providing label-conditioned image synthesis, and improving perceptual image quality. The proposed framework was extensively tested on large, diverse medical imaging datasets, including chest X-ray images to identify pneumonia, retinal fundus images to evaluate diabetic retinopathy, brain MRI images to detect tumors, microscopic images of leukemia white blood cells, and dermoscopic images to analyze skin cancer. Quantitative experimentation revealed a steady convergent behavior, the values of the generator loss and discriminator loss continued to decline throughout each of the datasets and the lowest values were found in diabetic retinopathy cases (generator loss of 0.522 and discriminator loss of 0.425) and leukemia cases (generator loss of 0.285 and discriminator loss of 0.224), maintaining the presence of diagnostically significant pathological features. Computational efficiency was also high, with about 0.75 hours of training time and a relatively small number of 0.67 million parameters, compared to SR-GAN-based models that require more than 10 hours of training time and more than 2.3 million parameters. The success of the framework was also confirmed by the quality of the generated images, attaining a high signal-to-noise ratio of 36.742, structural similarity index of 0.93, and Fréchet inception distance of 30.402, which is better than several other recent state-of-the-art methods, such as DRForecastGAN, GAN-VSP, IFGAN, and Pix2Pix GAN. Also, incorporating a Squeeze-and-Excitation convolutional neural network classifier not only led to a significant boost in disease classification performance but also improved the accuracy of diabetic retinopathy and pneumonia to 0.90 and 0.98, respectively. In general, the suggested hybrid GAN model has great potential as a low-cost, high-quality solution for medical image generation, data augmentation, and automated disease detection in clinical decision-support systems.