Progressive Uncertainty-Guided Evidential U-KAN for Trustworthy Medical Image Segmentation

TL;DR

This work tackles trustworthy medical image segmentation by addressing the underutilization of uncertainty maps for guiding feature learning. It introduces Evidential U-KAN, which combines Progressive Evidential Uncertainty Guided Attention (PEUA) with Semantic-Preserving Evidence Learning (SAEL) to refine segmentation in hard regions while preserving semantic integrity. The method represents outputs with a Dirichlet-based evidential framework, employs a low-rank designed attention mechanism on uncertainty maps, and iteratively updates uncertainty guidance until convergence. Empirical results on multiple datasets show improved Dice, IoU, and edge accuracy, along with stronger uncertainty estimation, suggesting practical benefits for robust clinical decision-making.

Abstract

Trustworthy medical image segmentation aims at deliver accurate and reliable results for clinical decision-making. Most existing methods adopt the evidence deep learning (EDL) paradigm due to its computational efficiency and theoretical robustness. However, the EDL-based methods often neglect leveraging uncertainty maps rich in attention cues to refine ambiguous boundary segmentation. To address this, we propose a progressive evidence uncertainty guided attention (PEUA) mechanism to guide the model to focus on the feature representation learning of hard regions. Unlike conventional approaches, PEUA progressively refines attention using uncertainty maps while employing low-rank learning to denoise attention weights, enhancing feature learning for challenging regions. Concurrently, standard EDL methods suppress evidence of incorrect class indiscriminately via Kullback-Leibler (KL) regularization, impairing the uncertainty assessment in ambiguous areas and consequently distorts the corresponding attention guidance. We thus introduce a semantic-preserving evidence learning (SAEL) strategy, integrating a semantic-smooth evidence generator and a fidelity-enhancing regularization term to retain critical semantics. Finally, by embedding PEUA and SAEL with the state-of-the-art U-KAN, we proposes Evidential U-KAN, a novel solution for trustworthy medical image segmentation. Extensive experiments on 4 datasets demonstrate superior accuracy and reliability over the competing methods. The code is available at \href{https://anonymous.4open.science/r/Evidence-U-KAN-BBE8}{github}.

Figures (4)

• Figure 1: Compared to traditional evidential deep learning methods, Semantic-preserving evidence learning strategy is capable of generating more precise uncertainty estimates and evidence distributions in ill-defined margins. Not only does it achieve superior guidance effects through the progressive evidence uncertainty guided attention, but it also significantly enhances the accuracy and robustness of uncertainty quantification. Particularly in scenarios with ambiguous class boundaries and uneven sample distributions, our method demonstrates higher reliability and stability.
• Figure 2: The overview of the proposed framework. Evidential U-KAN is an iterative network based on a U-KAN with five down-sampling stages. In the shallow layers, we propose the Evidence Enhancement Block (EEB) to integrate multi-scale information and enhance the model's capability for evidence extraction. In the deeper layers, the Tokenized KAN Block (TKB) replaces the MLP to enhance the model's learning capacity. The network replaces the skip connection of the shallowest layer with the Evidential Uncertainty Guided Attention module (EUGA), using the uncertainty map from the previous iteration (initialized as an all-ones matrix) as a condition to guide the model's segmentation. The iteration process stops once the convergence condition is satisfied.
• Figure 3: The overview of our Evidential Uncertainty Guided Attention module.We use a self-attention-like mechanism to extract an attention weight matrix from the uncertainty map. This matrix is then applied to weight the extracted input image features, enhancing key features and suppressing redundant information. To preserve the original input features, a residual connection is added at the end, combining the weighted features with the original input through element-wise addition.
• Figure 4: Visualization display of segmentation results of different models on the CVC-linicDB dataset, where Input represents the input image, GT represents the Ground truth, Pred represents the model segmentation result, UM represents the Uncertainty Map, and Diff represents the error between the segmentation result and the Ground truth