基于部分标注数据集的血管内超声图像深度学习分割  

Segmentation of Intravascular Ultrasound by Deep Learning on Partially Labelled Datasets

作　　者：余炜 吴鹏 涂圣贤[1] YU Wei;WU Peng;TU Shengxian(Biomedical Instrument Institute,School of Biomedical Engineering,Shanghai Jiao Tong University,Shanghai 200030,China)

机构地区：[1]上海交通大学生物医学工程学院生物医学仪器研究所,上海200030

出　　处：《复旦学报（自然科学版）》2023年第4期457-466,共10页Journal of Fudan University：Natural Science

摘　　要：由于各医学图像分割任务的差异性,通常在单独的数据集上进行神经网络的训练,而跨数据集的共享信息可提高各项任务的表现。本文旨在提出一种部分监督的语义分割方法在由两个数据集合并的部分标注数据集上进行血管内超声影像(IVUS)的冠脉结构分割,即:使用多标签语义分割来解决数据集标注不一致的问题,并提出了一种具有类再平衡策略的非对称双分支网络来提高分割性能。本文提出的方法取得了比全监督方法更好地分割性能;并且分割结果与参考标准在管腔面积(r=0.99;P<0.001)、中膜面积(r=0.99;P<0.001)和斑块负荷(r=0.95;P<0.001)也取得了强相关性及优秀的一致性;此外,本文方法有效地利用了标注不一致的部分标注数据集,缓解了IVUS影像分割中“数据饥饿”的问题。Typically,neural networks are trained on separate datasets in medical image segmentation tasks due to the difference between each individual task.However,training with shared information can improve the performance of each task and save inference time using one network.This paper aims to propose a partially supervised semantic segmentation method to segment artery structures on intravascular ultrasound(IVUS)using a union of two datasets.Multi-label semantic segmentation was used on the final partially labelled dataset to tackle the annotation inconsistency,and an asymmetric dual-branch network was proposed with class-rebalancing strategy to improve the segmentation performance.The method achieves better performance than the fully supervised methods.Strong correlation and good agreement of lumen area(r=0.99;P<0.001),media-adventitia area(r=0.99;P<0.001),and plaque burden(r=0.95;P<0.001)between the ground truth and the prediction are observed.The proposed method effectively utilizes a partially labelled IVUS dataset with annotation inconsistency and alleviates the problem of“data hunger”in IVUS segmentation.

关 键 词：冠状动脉疾病 深度学习 图像分割 血管内超声 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] R445.1[医药卫生—影像医学与核医学]