华侨大学学报（自然科学版）2025，Vol.46Issue(5)：513-527,15.DOI:10.11830/ISSN.1000-5013.202507008
全局特征信息共享交互的无标记弱边缘细胞实例分割网络
Label-Free Weak-Edge Cell Instance Segmentation Network for Global Feature Information Sharing and Interaction
摘要
Abstract
Aiming at the problem of insufficient instance segmentation accuracy caused by low contrast,weak cell edges,and diverse irregular shapes in label-free microscopic images,a global feature information sharing and interaction network based on Protonet theory is proposed.First,a cross-stage multi-branch edge self-awareness module is designed to enhance feature representation of low contrast images and improve adaptability to irregular cell geometries.Second,a multi-branch aggregated downsampling architecture is constructed to al-leviate feature loss problem in both the backbone network and feature pyramid.Finally,a feature information sharing decoupled segmentation head is proposed to achieve cross-level joint optimization of classification and regression tasks,while a Shape-IoU loss function is introduced to improve cell contour localization accuracy.Experimental results show that the proposed method achieves mAP@0.50 of 86.73％,56.63％and 45.13％on easy,medium and difficult segmentation tasks,respectively.These results demonstrate that the proposed method can effectively improve the segmentation accuracy of weak cell edges instances in label-free microscopic images,confirming its effectiveness and practicality.关键词
无标记细胞/实例分割/跨阶段多分支边缘自感知模块/多分支聚合下采样/损失函数Key words
label-free cells/instance segmentation/cross-phase multi-branch edge self-awareness module/multi-branch aggregated downsampling/loss function分类
计算机与自动化引用本文复制引用
雷昊翔,张富举,孙浩,洪岚,傅玉青,杜永兆..全局特征信息共享交互的无标记弱边缘细胞实例分割网络[J].华侨大学学报（自然科学版）,2025,46(5):513-527,15.基金项目
福建省财政拨款项目(5032501) （5032501）
