Diffusion Probabilistic Multi-cue Level Set for Reducing Edge Uncertainty in Pancreas Segmentation

TL;DR

This work addresses the challenging task of pancreas segmentation in CT images, where small organ size, variable morphology, and low contrast compromise edge accuracy. It introduces Diff-mcs, a coarse-to-fine framework that uses a diffusion probabilistic model to generate a probabilistic prior for the pancreas and then refines edges with a multi-cue level set that fuses grayscale, texture, and prior information. The method achieves state-of-the-art Dice scores across AbdomenCT-1K, NIH, and MSD datasets (e.g., Stage1 Dice ~89.4, Stage2 Dice ~92.3 on AbdomenCT-1K), with ablation studies confirming the importance of rough position, texture cues, and the diffusion-based prior. Uncertainty analyses show improved edge robustness and reduced variability after the refinement, highlighting practical benefits for reliable pancreas delineation in clinical settings. The approach offers a principled integration of probabilistic priors and variational edge refinement, with code available at the provided GitHub repository.

Abstract

Accurately segmenting the pancreas remains a huge challenge. Traditional methods encounter difficulties in semantic localization due to the small volume and distorted structure of the pancreas, while deep learning methods encounter challenges in obtaining accurate edges because of low contrast and organ overlapping. To overcome these issues, we propose a multi-cue level set method based on the diffusion probabilistic model, namely Diff-mcs. Our method adopts a coarse-to-fine segmentation strategy. We use the diffusion probabilistic model in the coarse segmentation stage, with the obtained probability distribution serving as both the initial localization and prior cues for the level set method. In the fine segmentation stage, we combine the prior cues with grayscale cues and texture cues to refine the edge by maximizing the difference between probability distributions of the cues inside and outside the level set curve. The method is validated on three public datasets and achieves state-of-the-art performance, which can obtain more accurate segmentation results with lower uncertainty segmentation edges. In addition, we conduct ablation studies and uncertainty analysis to verify that the diffusion probability model provides a more appropriate initialization for the level set method. Furthermore, when combined with multiple cues, the level set method can better obtain edges and improve the overall accuracy. Our code is available at https://github.com/GOUYUEE/Diff-mcs.

Figures (9)

• Figure 1: According to different threshold values ($\beta=0.6$, $\beta=0.4$, $\beta=0.2$, $\beta=0$) of the prior probability map, we obtain pancreas segmentation ground truth and inferred results respectively. Among these results, the red mask represents the pancreas segmentation ground truth, the green mask represents the inferred segmentation result, and the yellow area indicates accurately predicted regions.
• Figure 2: Based on the diffusion probabilistic model, the pancreas segmentation results show that areas in the probability map closer to red indicate higher confidence of belonging to the pancreas, while areas closer to blue indicate lower confidence.
• Figure 3: The multi-cue level set method is guided by a diffusion probabilistic model. Where the blue box is the process of the diffusion probabilistic model, and the green box represents the multi-cue level set method. In the blue box, we implement it with a ResNet encoder following a U-Net decoder. The encoder consists of a set of condition encoders and segmentation encoders with an FF-parser on the feature fusion path to constrain noise and connection. In the green box, The multi-cue level set method combines grayscale cues, texture cues, and prior cues to segment the pancreas by determining the initial level set positions using the prior probability distributions obtained from the network.
• Figure 4: The first row shows the segmentation results of the diffusion probabilistic model, and the second row shows the segmentation results of Diff-mcs. The red contour is the gold standard, and the green contour is the predicted segmentation results.
• Figure 5: The results on the NIH dataset, the red edge line represents the real mask, the green mask represents the first stage coarse segmentation results, the light red mask represents the Diff-mcs results, the overlap is indicated by yellow, and the upper right corner is the Dice of the two stages respectively.