Reference-Guided Pseudo-Label Generation for Medical Semantic Segmentation

Producing densely annotated data is a difficult and tedious task for medical imaging applications. To address this problem, we propose novel approach to generate supervision semi-supervised semantic segmentation. We argue that visually similar regions between labeled unlabeled images likely contain the same semantics therefore should share their label. Following thought, use small number of as reference material match pixels in an image best fitting pixel set. This way, avoid pitfalls such confirmation bias, common purely prediction-based pseudo-labeling. Since our method does not require any architectural changes or accompanying networks, one can easily insert it into existing frameworks. achieve performance standard fully supervised model on X-ray anatomy segmentation, albeit using 95% fewer images. Aside from in-depth analysis different aspects proposed method, further demonstrate effectiveness reference-guided learning paradigm by comparing against methods retinal fluid segmentation with competitive improve upon recent work up 15% mean IoU.