Distilling Effective Supervision for Robust Medical Image Segmentation with Noisy Labels

Despite the success of deep learning methods in medical image segmentation tasks, human-level performance relies on massive training data with high-quality annotations, which are expensive and time-consuming to collect. The fact is that there exist low-quality annotations label noise, leads suboptimal learned models. Two prominent directions for noisy labels include pixel-wise noise robust image-level training. In this work, we propose a novel framework address segmenting by distilling effective supervision information from both pixel levels. particular, explicitly estimate uncertainty every as estimation, using original pseudo labels. Furthermore, present an method accommodate more complements pixel-level learning. We conduct extensive experiments simulated real-world datasets. results demonstrate advantageous our compared state-of-the-art baselines