EVIL: Evidential Inference Learning for Trustworthy Semi-supervised Medical Image Segmentation

Recently, uncertainty-aware methods have attracted increasing attention in semi-supervised medical image segmentation. However, since these methods rely heavily on the prediction However, current methods usually suffer from of pseudo label, false predictions will severely degrade the drawback that it is difficult to balance the computational the segmentation performance. To improve the quality of cost, estimation accuracy, and theoretical support in a unified pseudo labels, some uncertainty-aware methods have been framework. To alleviate this problem, we introduce proposed, including Monte Carlo dropout (MC-dropout)- the Dempster-Shafer Theory of Evidence (DST) into semisupervised based [9], Information-Entropy-based [10], and Prediction medical image segmentation, dubbed EVidential Variance-based [11] methods. However, these methods suffer Inference Learning (EVIL). EVIL provides a theoretically from some problems: (1) Although MC-dropout is mathematically guaranteed solution to infer accurate uncertainty quantification guaranteed by Bayesian theory, its training process in a single forward pass. Trustworthy pseudo labels on is costly due to the multiple sampling operations; (2) Due unlabeled data are generated after uncertainty estimation. The to the limited sampling times, MC-dropout can't obtain accurate recently proposed consistency regularization-based training uncertainty quantification; (3) Other two uncertainty paradigm is adopted in our framework, which enforces the estimation methods have advantages in computational cost, consistency on the perturbed predictions to enhance the generalization but they lack theoretical support, leading to unstable pseudo with few labeled data. Experimental results show label generation.