.png){kind=icon}
.png){kind=icon}
Return
RBML-Diff: Diffusion model with region-boundary mutual learning for polyp segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2026.110145.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Although diffusion models have shined brightly in numerous computer vision tasks, these models are usually unavailable to effectively capture the correlation of boundary and non-boundary information, leading to inaccurate segmentation results on challenged areas in polyp segmentation. To address this issue, we propose a diffusion model with region-boundary mutual learning for polyp segmentation, namely RBML-Diff, which is a flexible dual-branch feature-guided diffusion framework. First, we design a region-aware module that can achieve both intra-layer feature interaction and inter-layer feature collaboration. The module thoroughly explores the correlation of multi-scale region features and improves the generalization capacity of RBML-Diff for polyps with morphological changes. Second, we design a boundary-aware denoising network, which can boost boundary information extraction using Laplacian operator and learn mutually region and boundary features during the denoising process to obtain more accurate segmentation results for polyps with ambiguous boundaries. We conducted extensive experiments on five publicly available polyp segmentation datasets. The experimental results show that the proposed RBML-Diff is superior to the state-of-the-arts methods in terms of segmentation accuracy and inference time, and demonstrates robustness in handling challenging cases such as small and flat polyp segmentation. The source code will be available at https://github.com/M0ck4ry/RBML-Diff .
Keywords:
polyp segmentation
diffusion model
region-boundary mutual learning
boundary-aware denoising
feature-guided diffusion
Journal
IF:
4.9
Papers:
9.7K
Citations:
2.4W