2021/04/27 by Simon Reiß, Constantin Seibold, Reiß, Simon +7 · 1 citation
Computer Science · Medicine · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #I.2.6 #I.4.6 #I.5.1 #I.5.4 #Medical Image Segmentation Techniques #Retinal Imaging and Analysis
paper · pdf · doi:10.48550/arxiv.2104.13243
openalex publication_date 2021/04/27 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Pixel-wise segmentation is one of the most data and annotation hungry tasks\nin our field. Providing representative and accurate annotations is often\nmission-critical especially for challenging medical applications. In this\npaper, we propose a semi-weakly supervised segmentation algorithm to overcome\nthis barrier. Our approach is based on a new formulation of deep supervision\nand student-teacher model and allows for easy integration of different\nsupervision signals. In contrast to previous work, we show that care has to be\ntaken how deep supervision is integrated in lower layers and we present\nmulti-label deep supervision as the most important secret ingredient for\nsuccess. With our novel training regime for segmentation that flexibly makes\nuse of images that are either fully labeled, marked with bounding boxes, just\nglobal labels, or not at all, we are able to cut the requirement for expensive\nlabels by 94.22% - narrowing the gap to the best fully supervised baseline to\nonly 5% mean IoU. Our approach is validated by extensive experiments on retinal\nfluid segmentation and we provide an in-depth analysis of the anticipated\neffect each annotation type can have in boosting segmentation performance.\n