融合边缘信息和双分支注意力的息肉分割算法  

Polyp segmentation algorithm combining boundary knowledge and dual-branch attention

作　　者：赵依林 郭逸 刘雨烟 张晴 ZHAO Yilin;GUO Yi;LIU Yuyan;ZHANG Qing(School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China)

机构地区：[1]上海应用技术大学计算机科学与信息工程学院,上海201418

出　　处：《应用技术学报》2024年第3期367-375,共9页Journal of Technology

基　　金：上海市自然科学基金项目(21ZR1462600)资助。

摘　　要：针对息肉与其周围黏膜之间边界不清晰导致的息肉分割精度低这一问题,提出一种融合边缘信息和双分支注意力的息肉分割模型(BDANet),用于在结肠镜检查图像中准确分割息肉。使用边缘提取感知模块以同时利用低级细节信息和高级语义信息生成边缘细节特征,将其与全局特征相融合,生成具有边界感知意识的全局信息,采用残差学习结构在不同层级强调边界学习,由深至浅整合不同层级的侧输出特征,设计双分支注意力模块同时从正向和反向学习目标物体的边界,降低模型对边界区域的预测不确定性。针对4个息肉分割常用数据集的7个指标的定量和定性评估表明,所提模型BDANet能够有效提高分割精度。Aiming at the problem of poor precision of polyp segmentation due to unclear boundaries between polyps and the surrounding mucosa,a polyp segmentation model(BDANet)combining boundary knowledge and dual-branch attention is proposed for accurate polyp segmentation in colonoscopy images.A boundary extraction and awareness module is designed to simultaneously utilize low-level detail information and high-level semantic information for generating edge detail features.These features are fused with global features to generate globally aware information with boundary perception.A residual learning structure is adopted to emphasize boundary learning at different levels.Side output features from various levels are integrated from deep to shallow.A dual-branch attention module is designed to learn boundaries of target objects both in forward and reverse directions,reducing the uncertainty of model in predicting boundary regions.Quantitative and qualitative evaluations on 7 metrics of 4 commonly used polyp segmentation datasets demonstrate that the proposed BDANet can effectively improve segmentation accuracy.

关 键 词：息肉分割 注意力机制 卷积神经网络 医学图像分割 结直肠癌 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]