基于多监督注意力机制神经网络的脑胶质瘤循环肿瘤细胞分割算法  

Neural network-based multi-level supervision and attention mechanism algorithm for brain glioma CTC segmentation

作　　者：袁红杰 杨艳[1] 张东[1] 杨双 YUAN Hongjie;YANG Yan;ZHANG Dong;YANG Shuang(School of Physics and Technology,Wuhan University,Wuhan 430072,China;School of Electronic Information and Automation,Guilin University of Aerospace Technology,Guilin 541004,China)

机构地区：[1]武汉大学物理科学与技术学院,湖北武汉430072 [2]桂林航天工业学院电子信息与自动化学院,广西桂林541004

出　　处：《中国医学物理学杂志》2022年第7期828-833,共6页Chinese Journal of Medical Physics

基　　金：国家重点研发计划973项目(2011CB707900)。

摘　　要：为了提升脑胶质瘤循环肿瘤细胞的分割准确率,解决人工分割中肉眼分辨边界困难、目标占比小和操作流程繁琐等问题,提出一种端到端的像素级分割算法。针对数据特征,提出一种基于多监督机制的混合损失函数用以提升预测区域与目标区域的交并比,同时训练网络向预测正确目标个数的方向收敛;其次,在网络中逐层加入卷积块注意力机制模块,使得网络能在空间、通道层面重点学习数据特征,进一步提升预测准确率;最后,通过采用混合训练的方式,只需一个网络模型就能直接分割出细胞核、细胞质区域,缩减训练流程。实验结果表明,此分割算法对比U-Net网络在召回率、精确率以及Dice系数方面均有显著提升,在细胞核分割方面,分别达到92.20%、86.56%、88.27%;在细胞质分割方面,分别达到89.33%、85.31%、86.33%。In order to improve the segmentation accuracy of glioma circulating tumor cells(CTC)and solve the problems of difficulties in distinguishing boundaries with the naked eye,small target ratio and cumbersome operation process in manual segmentation,an end-to-end pixel-level segmentation algorithm is proposed.In view of data features,the proposed algorithm utilizes a hybrid loss function based on a multi-level supervision mechanism to improve the intersection-over-union between the prediction area and the ground truth area,and iterate the network converging in the direction of predicting the right numbers of targets.Then,the convolutional block attention module is put in every layer of the proposed network enables the network to focus on learning data features at the spatial and channel dimensions,thereby further improving the prediction accuracy.Finally,the proposed algorithm can segment the nucleus and cytoplasm by one network model through hybrid training,which simples the process of training.The experimental results showed that compared with U-Net network,the proposed segmentation algorithm has improved in terms of recall rate,precision and Dice coefficient.The above-mentioned indexes are 92.20%,86.56%,88.27%for cell nucleus segmentation,and 89.33%,85.31%,86.33%for cytoplasm segmentation.

关 键 词：脑胶质瘤 循环肿瘤细胞 多监督 卷积块注意力机制模块 小目标分割 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]