Efficient Medical Image Analysis in the JPEG-AI Compressed Domain
Year: 2026
Type: article
Abstract: Standard medical image analysis pipelines usually require full image decompression before applying computer vision models, imposing considerable computational and memory burdens, particularly with modern learning-based codecs such as JPEG-AI. This requirement restricts the deployment of advanced diagnostic tools in settings where resources are limited, and limits the effective reuse of the rich latent features generated by these codecs. This article introduces ProMIC, a framework for performing medical image analysis directly in the JPEG-AI compressed domain, removing the need for computationally expensive image decompression. It exploits the codec’s latent representations through two key components: the Domain Translator, that maps compressed features into a space compatible with standard computer vision models, and the Mirror Training Strategy, that aligns compressed-domain and pixel-domain knowledge via guided translator learning and end-to-end refinement. Experiments across three datasets, Lucchi++, Kasthuri++, and CVC-ClinicDB, show that ProMIC delivers accuracy on par with or better than pixel-domain baselines in detection and segmentation, while substantially lowering computational and memory demands. Notably, on CVC-ClinicDB, it reduces inference complexity by 65% with a minor drop of 1.5 percentage points (pp) in AP50. On Lucchi++, it improved AP75 performance by nearly 28 pp, when compared to a state-of-the-art approach. In segmentation, it improves mIoU by 2 pp while lowering decoding and feature extraction complexity by 70%. ProMIC thus establishes an efficient, scalable route to high-quality medical image understanding directly in the JPEG-AI compressed domain. The code and model weights will be available at https://github.com/nicoVascon/ProMIC once the paper is accepted.
Source: IEEE Access
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7213996527