联合影像数据集腹部多器官分割方法研究  

Research of Joint-Dataset Abdominal Multi-Organ Segmentation Method

作　　者：吴泽静 陈春晓[1] 陈志颖 徐俊琪 傅雪 Wu Zejing;Chen Chunxiao;Chen Zhiying;Xu Junqi;Fu Xue(Department of Biomedical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学生物医学工程系,南京211106

出　　处：《中国生物医学工程学报》2023年第2期129-138,共10页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(12071215);南京航空航天大学研究生创新项目(xcxjh20210324)。

摘　　要：医学影像多器官分割对于手术治疗规划和辅助诊断等临床应用至关重要。目前大多数公开的医学影像数据集都仅对部分器官进行标注,由此建立的分割模型泛化性较差,很难同时满足对多器官的精确自动分割。本研究针对腹部数据集部分器官标注及分割精度较低的问题,构建了基于联合影像数据集的腹部多器官自动分割网络C2F-MSNet。C2F-MSNet分割网络由粗分割和细分割两个阶段构成。在粗分割阶段,利用条件控制模块显式地控制神经网络在多个部分标注的数据集上进行联合训练,并引入注意力模块和深监督策略;在细分割阶段,通过粗分割结果索引细分割区域、引导细分割并构建多尺度细分割网络。在KiTS、Decathlon-liver、Decathlon-spleen和Decathlon-pancreas等4个数据集中的663例CT数据上进行实验,以Dice相关系数(DSC)和豪斯多夫距离(HD)作为分割结果的评判标准,肾脏、肝脏、脾脏和胰腺分割后的DSC分别为0.967、0.964、0.956、0.838,HD分别为12.51、25.02、6.68、12.58。实验结果表明,C2F-MSNet分割网络可以有效解决多标签部分标注的问题,实现联合数据集多器官的精准自动分割。Abdominal multi-organ segmentation of medical images is essential for clinical applications such as surgical treatment planning and assisted diagnosis.Most published medical image datasets label partial organs only,which is difficult for accurate multiple organs segmentation of medical images,thus the segmentation model developed by this approach is not generally applicable.In this paper,we proposed a joint-dataset-based multi-organ segmentation abdominal network:C2F-MSNet,which contained coarse segment and fine segment.During coarse segmentation,the explicit conditional control module was employed for the training of the network on multiple partially labeled datasets,while the self-attention module and the deep supervision strategy were implemented.During the fine segmentation,the fine segmentation area was indexed by rough segmentation results,and the fine segmentation was guided and the multi-scale fine segmentation network was constructed.Experiments were performed on 663 CT data obtained from KiTS,Decathlon-liver,Decathlon-spleen and Decathlon-pancreas datasets,evaluated by dice similarity coefficient(DSC)and Hausdorff distance(HD).The results of DSC reached 0.967,0.964,0.956,and 0.838 for kidney,liver,spleen,and pancreas,respectively,and that of HD reached 12.51,25.02,6.68,and 12.58,respectively.Experiment results showed that the C2F-MSNet effectively solved the training problem of multiple partially labeled datasets and achieved accurate multi-organ segmentation in the joint datasets.

关 键 词：多器官分割 两阶段分割 联合数据集 条件控制模块 

分 类 号：R318[医药卫生—生物医学工程]