Uncertainty-guided model learning for trustworthy medical image segmentation

Although recent AI algorithms have greatly improved accuracy, there is still skepticism among clinicians about the reliability and trustworthiness of these algorithms when used in real-life situations. To address this, our research proposes a new method for trustworthy medical image segmentation, which aims to produce reliable segmentation results and reliable uncertainty estimations without imposing an excessive computational burden. One key feature of our approach is the extraction of feature maps from each decoder and their fusion based on voxel-level uncertainty to leverage multi-scale semantic information fully. We also model the probability and uncertainty of medical image segmentation problems using subjective logic theory, which quantifies the uncertainty of the backbone by modeling class probabilities as a Dirichlet distribution and calculating the distribution strength. Moreover, our framework learns to gather reliable evidence from features, leading to the final segmentation results. To further improve model performance, an uncertainty-based adaptive threshold strategy is designed based on voxel-level uncertainty. This strategy dynamically adjusts the threshold based on the model's learning state, guiding the model to focus on regions with high uncertainty to optimize segmentation results. Extensive experiments on multiple public datasets have validated the effectiveness of our approach, the code is released at https://github.com/SCUT-Ye/UGML .