The Adaptive Medical Image Compression Based On A Hybrid Neural Network With Built-In ROI Detection

Abstract

This study addresses the critical challenge of efficiently compressing the rapidly growing volume of medical images while preserving essential diagnostic details, particularly within the Regions of Interest (ROI). Traditional compression techniques, whether lossless or lossy, often struggle to balance high compression efficiency with image quality lossless methods offer limited data reduction, while lossy techniques risk removing vital clinical information. To overcome these limitations, a comprehensive hybrid compression framework is developed, integrating segmentation and compression within a single deep neural network. The system employs Convolutional Neural Networks (CNNs) to accurately segment medical images and identify ROIs, while an autoencoder-based compression module performs selective encoding applying near-lossless compression for ROI regions to maintain diagnostic fidelity and lossy compression for non-ROI (NROI) areas to maximize storage savings. This unified design eliminates the need for separate processing stages, reduces computational complexity, and enhances compression performance. The proposed framework was validated using the CLEF MED X-ray and BRATS MRI datasets, demonstrating high effectiveness and adaptability across different modalities. Experimental results achieved a Peak Signal-to-Noise Ratio (PSNR) of 56.07 dB for ROI and 45.12 dB for NROI, with an overall compression ratio of 6.73, confirming its strong balance between data reduction and image quality.