结合多尺度信息和特征增强的息肉分割方法  

Polyp Segmentation Method Combining Multiscale Information and Feature Enhancement

作　　者：汪琴韵 于瓅[1] WANG Qin-yun;YU Li(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan,Anhui,232001,China)

机构地区：[1]安徽理工大学计算机科学与工程系,安徽淮南232001

出　　处：《新疆师范大学学报（自然科学版）》2025年第1期1-10,共10页Journal of Xinjiang Normal University(Natural Sciences Edition)

基　　金：2021年安徽省重点研究与开发计划项目(202104d07020010)。

摘　　要：计算机辅助诊断技术在结肠息肉图像分割中具有十分重要的作用。本研究针对结肠息肉分割存在的边缘检测分割比较复杂、息肉与正常组织对比度较低等问题,使用DeepLabv3+模型进行息肉分割并对其进行改进。该模型将残差网络作为主干网络进行特征信息提取,引入特征强化模块对低层次特征进行处理,并加入混合注意力机制对高层次特征进行关键信息的捕捉。得到改进模型在Kvasir-SEG与CVC-ClinicDB两个数据集上的mIoU为82.19%和90.40%.实验结果表明,改进的模型息肉分割效果优于初始DeepLabv3+模型,在息肉分割的精度以及边缘分割上有一定提升。Computer-aided diagnosis technology plays a very important role in colon polyp image segmentation.In view of the problems of colon polyp segmentation such as complicated edge detection and segmentation and low contrast between polyps and normal tissue,an improved segmentation model based on DeepLabv3+is proposed.This model uses the residual network as the backbone network to extract feature information,introduces a feature enhancement module to process low-level features,and adds a hybrid attention mechanism to capture key information for high-level features.The average intersection and union ratios of this model reached 82.19%and 90.40%respectively in experiments on the Kvasir-SEG data set and the CVC-ClinicDB data set.Experimental results show that the polyp segmentation effect of the improved model is better than that of the initial DeepLabv3+model,and there has been a certain improvement in the accuracy of polyp segmentation and edge segmentation.

关 键 词：DeepLabv3+ 息肉分割 混合注意力 特征增强 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]