Multi-scale interactive pyramid network for polyp segmentation

Shuo Dai

doi:10.1016/j.bspc.2025.108749

Multi-scale interactive pyramid network for polyp segmentation

2025-10-11

0

PRE

AI

DOI：10.1016/j.bspc.2025.108749

Original

Original request for help

Save

Abstract

En 中文

In colonoscopy image analysis, accurate identification and segmentation of polyps are crucial for the diagnosis of colorectal cancer. While significant progress has been made using deep learning methods, traditional UNet models suffer from inadequate multi-scale feature fusion, making it challenging to effectively differentiate polyp edges from surrounding tissues. To address this issue, we propose P-UNet, a novel multi-scale interactive network structure. P-UNet integrates varying numbers of multi-scale feature maps at each layer of the encoder through a pyramid structure to comprehensively fuse global and local information, thereby resolving the issue of information loss. Furthermore, we introduce the Edge-Aware Adaptive Feature Adjustment Convolutional Block (EAB) to enhance the model’s focus on edge regions. Validation on two public datasets demonstrates the superior performance of our model across most evaluation metrics, close to or surpassing current state-of-the-art methods. For related code, please refer to: https://github.com/LCUDai/PUNet.git .

Multi-scale interactive pyramid network for polyp segmentation

Abstract

Journal

Organization

Cited Papers