Mutual Evidential Deep Learning for Medical Image Segmentation

TL;DR

This work addresses pseudo-label quality in semi-supervised medical image segmentation by proposing Mutual Evidential Deep Learning (MEDL), which employs two heterogeneous networks to generate complementary evidence and fuses them with class-aware evidential fusion. An uncertainty-driven, asymptotic Fisher information-based learning strategy then guides curriculum-like training from confident to uncertain voxels for both labeled and unlabeled data, leveraging voxel-wise belief masses and reliability masking. Key contributions include CAEF for robust pseudo-label synthesis, a Fisher-information-based evidential loss for unlabeled data, and a voxel-level weighting scheme that optimizes learning order; extensive experiments across five datasets demonstrate state-of-the-art performance. The approach provides a principled, uncertainty-aware framework for semi-supervised MIS with strong potential for robust deployment in clinical settings.

Abstract

Existing semi-supervised medical segmentation co-learning frameworks have realized that model performance can be diminished by the biases in model recognition caused by low-quality pseudo-labels. Due to the averaging nature of their pseudo-label integration strategy, they fail to explore the reliability of pseudo-labels from different sources. In this paper, we propose a mutual evidential deep learning (MEDL) framework that offers a potentially viable solution for pseudo-label generation in semi-supervised learning from two perspectives. First, we introduce networks with different architectures to generate complementary evidence for unlabeled samples and adopt an improved class-aware evidential fusion to guide the confident synthesis of evidential predictions sourced from diverse architectural networks. Second, utilizing the uncertainty in the fused evidence, we design an asymptotic Fisher information-based evidential learning strategy. This strategy enables the model to initially focus on unlabeled samples with more reliable pseudo-labels, gradually shifting attention to samples with lower-quality pseudo-labels while avoiding over-penalization of mislabeled classes in high data uncertainty samples. Additionally, for labeled data, we continue to adopt an uncertainty-driven asymptotic learning strategy, gradually guiding the model to focus on challenging voxels. Extensive experiments on five mainstream datasets have demonstrated that MEDL achieves state-of-the-art performance.

Figures (4)

• Figure 1: The overview of Mutual Evidential Deep Learning framework. We use two different segmentation models to predict evidence for labeled and unlabeled data. For unlabeled data, class-aware evidential fusion (CAEF) combines complementary evidence from the two models to generate pseudo-labels and reliability measures. Reliability measures mask the original evidential predictions from each network, and evidential deep learning (EDL) uncertainty with weighted average estimates rank each voxel, applying asymptotic Fisher information-based EDL (FIE). For labeled data, a similar strategy guides learning without the need for reliability-based masking and FIE ($\odot$ represents the fusion operation).
• Figure 2: Visualization results on the LA, Pancreas, and TBAD datasets. The first, second, and third rows show the results of the comparison method, the proposed method, and the ground truth (GT).
• Figure 3: Visualization results on the ACDC dataset. The first, second, and third rows are the results of the comparison method, the proposed method, and the ground truth (GT).
• Figure 4: 10% labeled ratio performance comparison on the usage of EF and CAEF.