High-Fidelity Medical Image and Video Compression via Implicit Neural Representations Enhanced with Model-Agnostic Meta-Learning

Abstract

This work presents a novel framework for medical image and video compression, targeting MRI, X-ray, and dynamic imaging sequences. The method leverages Implicit Neural Representations (INRs) enhanced with Model-Agnostic Meta-Learning (MAML) to optimize preprocessing, feature extraction, and model architecture for both static and temporal data. Conventional compression techniques often degrade image quality at high compression ratios, risking the loss of critical diagnostic information. The proposed approach integrates three key components: (1) Z-score normalization for improved representation efficiency, (2) hybrid feature extraction using Histogram of Oriented Gradients (HOG), Discrete Wavelet Transform (DWT), and Local Binary Patterns (LBP) to preserve structural edges, textures, and contrast, and (3) temporal coherence modeling with 3D convolutions and motion-compensated prediction for video sequences. At its core, a multi-scale INR architecture encodes spatial and temporal information as continuous coordinate functions, while MAML initialization accelerates convergence and improves generalization across modalities. Hierarchical positional encodings, sparsity-aware loss, and temporal consistency regularization further enhance compression fidelity. Evaluations on NIH Chest X-ray, fastMRI, and cardiac cine-MRI datasets show up to 1.8 dB PSNR and 0.019 MS-SSIM improvements over standard codecs, with 28–35% bitrate reduction for images and 40–45% for videos. The framework offers a real-time, clinically viable solution for telemedicine and resource-limited healthcare settings.