Efficient Compression of Medical Images Using Generative Adversarial Networks

Abstract

The exponential growth of medical imaging data necessitates efficient compression techniques that balance storage reduction and diagnostic fidelity. Traditional methods (e.g., JPEG and DICOM) often compromise clinically relevant features. This study evaluates deep-learning-based compression using Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) for radiology applications. A dataset of 1,200 MRI, CT, and X-ray images was collected from the Tikur Anbessa Specialized Hospital, Ethiopia, with ethical approval (AAU-IRB-2023-001). Images were preprocessed (normalized to [0,1], resized to 256×256 pixels) and augmented (rotation, flipping). GANs and VAEs were trained to compress images, and their performances were evaluated using the Inception Score (IS), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). Diagnostic fidelity was assessed by two radiologists using 100 reconstructed images. GANs achieved superior compression ratios (1:12 vs. 1:10 for VAEs) while preserving diagnostic features, evidenced by higher IS (8.5 ± 0.3 vs. 7.8 ± 0.4) and PSNR (32.5 dB vs. 30.1 dB). Radiologists identified 98% of pathologies in GAN-compressed images versus 95% for VAEs. SSIM scores exceeded 0.92 for both models, confirming structural integrity. GAN-based compression outperforms traditional methods and VAEs in radiological imaging, offering a scalable solution for PACS and telemedicine. Future work will validate this robustness across diverse modalities and low-resource settings.