基于双模型交互学习的半监督医学图像分割  

Interactive Dual-model Learning for Semi-supervised Medical Image Segmentation

作　　者：方超伟 李雪 李钟毓[3] 焦李成 张鼎文[4,5] FANG Chao-Wei;LI Xue;LI Zhong-Yu;JIAO Li-Cheng;ZHANG Ding-Wen(School of Artificial Intelligence,Xidian University,Xi'an 710071;School of Mechano-Electronic Engineering,Xidian University,Xi'an 710071;School of Software Engineering,Xi'an Jiaotong University,Xi'an 710049;Brain and Artificial Intelligence Laboratory,Northwestern Polytechnical University,Xi'an 710072;Institute of Artificial Intelligence,Hefei Comprehensive National Science Center,Hefei 230088)

机构地区：[1]西安电子科技大学人工智能学院,西安710071 [2]西安电子科技大学机电工程学院,西安710071 [3]西安交通大学软件学院,西安710049 [4]西北工业大学脑与人工智能实验室,西安710072 [5]合肥综合性国家科学中心人工智能研究院,合肥230088

出　　处：《自动化学报》2023年第4期805-819,共15页Acta Automatica Sinica

基　　金：国家自然科学基金(62003256,61876140,U21B2048)资助。

摘　　要：在医学图像中,器官或病变区域的精准分割对疾病诊断等临床应用有着至关重要的作用,然而分割模型的训练依赖于大量标注数据.为减少对标注数据的需求,本文主要研究针对医学图像分割的半监督学习任务.现有半监督学习方法广泛采用平均教师模型,其缺点在于,基于指数移动平均(Exponential moving average,EMA)的参数更新方式使得老师模型累积学生模型的错误知识.为避免上述问题,提出一种双模型交互学习方法,引入像素稳定性判断机制,利用一个模型中预测结果更稳定的像素监督另一个模型的学习,从而缓解了单个模型的错误经验的累积和传播.提出的方法在心脏结构分割、肝脏肿瘤分割和脑肿瘤分割三个数据集中取得优于前沿半监督方法的结果.在仅采用30%的标注比例时,该方法在三个数据集上的戴斯相似指标(Dice similarity coefficient,DSC)分别达到89.13%,94.15%,87.02%.Accurate segmentation of organs or lesions in medical images plays a significant role in clinical applications such as clinical diagnosis.However,learning segmentation models require a large number of annotated samples.This paper focuses on the semi-supervised medical image segmentation to relieve the dependence on labeled samples.A widely used semi-supervised learning method is temporally averaging a student model as the teacher model.However,it accumulates the incorrect knowledge of the student model as well.To address the above issue,we propose an interactive dual-model learning algorithm.Aiming to prevent the propagation and accumulation of error knowledge,we devise a specific mechanism for judging and measuring the instability of network predictions.Only pixels with relatively more stable predictions in one model are employed to supervise the other model.Extensive experiments on three datasets including cardiac structure segmentation,liver tumor segmentation,and brain tumor segmentation,demonstrate that the proposed method outperforms the state-of-the-art semi-supervised methods.When 30%of annotations are available,the Dice similarity coefficient(DSC)metric of our method reaches 89.13%,94.15%and 87.02%respectively on the above three datasets.

关 键 词：半监督学习 医学图像分割 双模型交互学习 平均教师 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] R318[自动化与计算机技术—计算机科学与技术]