Diagnosis-oriented Medical Image Compression with Efficient Transfer Learning

• DOI: 10.48550/arxiv.2310.13250
• Publication Date: 2023-01-01
• Authors: Guangqi Xie; Xin Li; Xiaohan Pan; Zhibo Chen
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV); Computer Science - Computer Vision and Pattern Recognition; FOS: Electrical engineering, electronic engineering, information engineering; Electrical Engineering and Systems Science - Image and Video Processing; 3. Good health

Remote medical diagnosis has emerged as a critical and indispensable technique in practical systems, where data are required to be efficiently compressed transmitted for by either professional doctors or intelligent devices. In this process, large amount of redundant content irrelevant the is subjected high-fidelity coding, leading unnecessary transmission costs. To mitigate this, we propose diagnosis-oriented image compression, special semantic compression task designed scenarios, targeting reduce cost without compromising accuracy. However, collecting sufficient optimize such system significantly expensive challenging due privacy issues lack annotation. study, DMIC, first efficient transfer learning-based codec, which can effectively optimized with only few-shot annotated examples, reusing knowledge existing reinforcement task-driven coding framework, i.e., HRLVSC [1]. Concretely, focus on tuning partial parameters policy network bit allocation within HRLVSC, enables it adapt images. work, validate our DMIC typical task, Coronary Artery Segmentation. Extensive experiments have demonstrated that achieve 47.594%BD-Rate savings compared HEVC anchor, A2C module (2.7% parameters) 1 sample.