TA-LSDiff:Topology-Aware Diffusion Guided by a Level Set Energy for Pancreas Segmentation

TL;DR

TA-LSDiff introduces a topology-aware diffusion framework for pancreas segmentation by guiding a diffusion-based segmentation model with a four-term level-set energy (region, length, area, distance) and a pixel-adaptive refinement module. The energy terms inject explicit geometric and topological priors into the reverse diffusion process, enabling topology-consistent and boundary-smooth segmentations without explicit PDE evolution. The approach unifies diffusion priors with classical level-set theory via a energy-gradient view, supported by proofs linking Chan-Vese gradient flow with topological derivatives. Empirical results across four public datasets show state-of-the-art Dice scores and robust generalization, with ablations confirming the contributions of each energy term and the PAR module. This framework offers a practical, accurate solution for pancreas segmentation and a blueprint for energy-guided diffusion in other topologically sensitive medical-image tasks.

Abstract

Pancreas segmentation in medical image processing is a persistent challenge due to its small size, low contrast against adjacent tissues, and significant topological variations. Traditional level set methods drive boundary evolution using gradient flows, often ignoring pointwise topological effects. Conversely, deep learning-based segmentation networks extract rich semantic features but frequently sacrifice structural details. To bridge this gap, we propose a novel model named TA-LSDiff, which combined topology-aware diffusion probabilistic model and level set energy, achieving segmentation without explicit geometric evolution. This energy function guides implicit curve evolution by integrating the input image and deep features through four complementary terms. To further enhance boundary precision, we introduce a pixel-adaptive refinement module that locally modulates the energy function using affinity weighting from neighboring evidence. Ablation studies systematically quantify the contribution of each proposed component. Evaluations on four public pancreas datasets demonstrate that TA-LSDiff achieves state-of-the-art accuracy, outperforming existing methods. These results establish TA-LSDiff as a practical and accurate solution for pancreas segmentation.

Figures (7)

• Figure 1: The blue box represents the processing procedure of the diffusion probability model, and we use the ResNet encoder and U-Net decoder to achieve this process. The encoder consists of a group of conditional encoders and a segmentation encoder with an attention mechanism on the feature fusion path. The green box represents the processing procedure of the level set energy and pixel adaptive module.
• Figure 2: Segmentation results on the AbdomenCT-1K dataset (first row) and the NIH dataset (second row). Contours: Red = Gold Standard, Green = Predicted Result.
• Figure 3: Segmentation results on the MSD dataset (first row) and the WORD dataset (second row). Contours: Red = Gold Standard, Green = Predicted Result.
• Figure 4: Comparison of segmentation results on different instances without (first row) and with (second row) the level set evolution. Contours: Red = Gold Standard, Green = Predicted Result.
• Figure 5: Conceptual illustration of Pixel--adaptive Refinement (PAR). An affinity kernel is computed for each pixel to measure its proximity to its neighbors. This kernel is then iteratively applied to the predicted mask (not the level set evolution) via adaptive convolution to obtain the refined segmentation result.

Theorems & Definitions (5)

• Proposition 1
• proof
• Proposition 2
• proof
• proof