DPUSegDiff: A Dual-Path U-Net Segmentation Diffusion model for medical image segmentation
Yazhuo Fan, Jianhua Song, Yang Lu, Xinrong Fu, Xinying Huang, Lei Yuan · Electronic Research Archive · 2025
Denoising diffusion probabilistic models (DDPM) have had remarkable success in image generation. Inspired by this, recent medical image segmentation tasks have started to use diffusion-based methods. These methods leverage iterations and sampling to generate smoother and more representative implicit integration. However, current diffusion-based segmentation models mainly rely on traditional neural networks and seldom focus on effectively interacting semantic and noise information. Moreover, they usually use a single network architecture instead of a hybrid one combining CNN and Transformer. To address limitations, we propose a dual-path U-Net segmentation diffusion (DPUSegDiff) model. It comprises two U-shaped networks based on the edge augmented local encoder (EALE) and the mixed transformer global encoder (MTGE). EALE uses the Sobel operator for local feature extraction, and MTGE has a cross-attention mechanism to facilitate information interaction. To integrate information from both paths selectively and adaptively, we design a bilateral gated transformer module (BGTM) to combine deep semantic information effectively. Experiments on three segmentation tasks—skin lesions, polyps, and brain tumors—show that the proposed DPUSegDiff outperforms other state-of-the-art (SOTA) methods in segmentation performance and generalization ability. The code has been released on GitHub (https://github.com/Fanyyz/DPUSegDiff).