UCAD: Uncertainty-guided Contour-aware Displacement for semi-supervised medical image segmentation

TL;DR

The UCAD leverages superpixels to generate anatomically coherent regions aligned with anatomy boundaries, and an uncertainty-guided selection mechanism to selectively displace challenging regions for better consistency learning, and proposes a dynamic uncertainty-weighted consistency loss.

Abstract

Existing displacement strategies in semi-supervised segmentation only operate on rectangular regions, ignoring anatomical structures and resulting in boundary distortions and semantic inconsistency. To address these issues, we propose UCAD, an Uncertainty-Guided Contour-Aware Displacement framework for semi-supervised medical image segmentation that preserves contour-aware semantics while enhancing consistency learning. Our UCAD leverages superpixels to generate anatomically coherent regions aligned with anatomy boundaries, and an uncertainty-guided selection mechanism to selectively displace challenging regions for better consistency learning. We further propose a dynamic uncertainty-weighted consistency loss, which adaptively stabilizes training and effectively regularizes the model on unlabeled regions. Extensive experiments demonstrate that UCAD consistently outperforms state-of-the-art semi-supervised segmentation methods, achieving superior segmentation accuracy under limited annotation. The code is available at:https://github.com/dcb937/UCAD.

Figures (3)

• Figure 1: Illustration of different displacement strategies: (a) Random displacement: randomly exchanges rectangular regions between labeled and unlabeled images, ignoring anatomical boundaries; (b) Uncertainty-guided displacement: divides images into grid patches and performs uncertainty-guided region exchange, but still lacks structural alignment; (c) Uncertainty-guided contour-aware displacement (ours): generates superpixel-based, anatomically coherent regions and selectively displaces uncertain superpixel regions; (d) Comparing displacement results, we observe that conventional methods often impair structural and spatial priors, while our method maintains this information.
• Figure 2: Overview of our framework. (a) The teacher network takes the unlabeled image $X^u$ as input to generate a pseudo label $\tilde{Y}^u$ and a pixel-wise uncertainty score $U(X^u)$. (b) The unlabeled image $X^u$ is partitioned into superpixel region using the SLIC achanta2012slic algorithm, and the uncertainty $U(X^u)$ guides the selection of $N$ uncertain regions to form the mask $\mathcal{M}$. The masked regions are then displaced with the labeled image $X^l$, resulting in a mixed image $X_{mix}$ and its corresponding mixed label $\tilde{Y}_{mix}$. The student network is trained on $X_{mix}$ to predict $P_{mix}$, which is supervised by $\tilde{Y}_{mix}$ and an uncertainty-weighted consistency loss $\mathcal{L}_{unc}$. For simplicity, the reverse displacement process is omitted.
• Figure 3: Visual comparison on the 10% Synapse dataset: spleen, right kidney, left kidney, gallbladder, pancreas, liver, stomach, and aorta.