Extending the compression range of biomedical images for machine vision analysis

The growing adoption of biomedical machine vision algorithms to perform detection, segmentation, and classification tasks, is driving a shift in compression paradigms, progressively replacing perceptual quality by performance of machine vision tasks as the target encoding optimization. Therefore, improving task performance rather than image quality has become a new research problem in biomedical image compression. This paper presents a contribution to extend the useful compression range from lossless to lossy while keeping the performance of biomedical machine vision algorithms. Automatic detection of mitochondrias in electron microscopy images, using a learning-based network (YOLO), is the case-study investigated in this work. Two types of new results are presented in regard to detection performance. In the first one, it is shown that compression ratios up to 15 can be used, for a maximum of 3% of detection loss. Then in the second one, by using compressed images for training, it is shown that the compression range can be increased up to 135 times, while missing less than 5% of the detections.