Uncertainty-aware Evidential Fusion-based Learning for Semi-supervised Medical Image Segmentation

TL;DR

This work tackles semi-supervised medical image segmentation by introducing an uncertainty-aware evidential fusion framework. It combines cross-region evidential predictions from mixed and original samples through IPAF and guides fine-grained voxel learning with voxel-wise asymptotic learning (VWAL), which fuses information entropy into the evidential uncertainty. The approach yields state-of-the-art results on four benchmark datasets (LA, Pancreas-CT, ACDC, TBAD) across low-label regimes, demonstrating improved Dice/Jaccard and boundary metrics while reducing reliance on labeled data. By explicitly modeling and leveraging multi-source uncertainty, the method enhances calibration and focuses learning on challenging voxel regions, with potential impact on clinical segmentation tasks.

Abstract

Although the existing uncertainty-based semi-supervised medical segmentation methods have achieved excellent performance, they usually only consider a single uncertainty evaluation, which often fails to solve the problem related to credibility completely. Therefore, based on the framework of evidential deep learning, this paper integrates the evidential predictive results in the cross-region of mixed and original samples to reallocate the confidence degree and uncertainty measure of each voxel, which is realized by emphasizing uncertain information of probability assignments fusion rule of traditional evidence theory. Furthermore, we design a voxel-level asymptotic learning strategy by introducing information entropy to combine with the fused uncertainty measure to estimate voxel prediction more precisely. The model will gradually pay attention to the prediction results with high uncertainty in the learning process, to learn the features that are difficult to master. The experimental results on LA, Pancreas-CT, ACDC and TBAD datasets demonstrate the superior performance of our proposed method in comparison with the existing state of the arts.

Figures (5)

• Figure 1: The diagram illustrates the fusion process of the model's predictions for the voxels at the same location in the original samples and the samples after mixed area restoration. It demonstrates the redistribution and adjustment of the voxel's confidence level in belonging to a category after interacting with the overall uncertainty of the prediction.
• Figure 2: In the evidential fusion-based framework, the pre-training phase involves the student network processing original and mixed data to generate separate predictions. It then restores mixed data to match original samples, employing an enhanced probabilistic strategy for fusing voxel predictions from two sources, considering prediction uncertainty. This process adjusts confidence levels and updates voxel uncertainty, facilitating balanced loss calculation through fused voxel predictions and dynamic weighting. During self-training, the addition of unlabeled data, with labels from the teacher network, is the key change. The teacher network's parameters are updated via the Exponential Moving Average (EMA) method.
• Figure 3: Visualization of experimental results on Left Atrium (LA) and Pancreas-CT dataset. The first line in the visualized image represents the results of the comparative methods (A&D, BCP). The second line shows the visualized results of the proposed method. The third line indicates the segmentation ground truth corresponding to each image.
• Figure 4: Visualization of experimental results on ACDC dataset. The first line in the visualized image represents the results of the comparative method (A&D). The second line shows the visualized results of the proposed method. The third line indicates the segmentation ground truth corresponding to each image.
• Figure 5: Visualization of experimental results on TBAD dataset. The first line in the visualized image represents the results of the comparative method (UPCoL). The second line shows the visualized results of the proposed method. The third line indicates the segmentation ground truth corresponding to each image.