改进协同训练的肺部CT影像COVID-19病灶分割方法

Approach of COVID-19 lesions segmentation on lung CT image based on improved co-training

在新型冠状病毒肺炎的肺部病灶分割任务中,基于半监督学习进行病灶分割可以利用大量未标记数据。针对半监督学习中伪标签置信度不足问题,采用UNet和DeepLabV3+作为基础网络搭建协同训练框架,以集成方法获取高质量伪标签;引入JS距离度量伪标签的不确定性,给予伪标签监督损失一个正则项,减轻低质量伪标签对分割性能的影响。在公开数据集中进行实验,获得Dice系数76.06%、IOU分数65.1%、敏感度分数77.22%和精确率分数81.46%。

In the task of lung lesion segmentation for COVID-19,semi-supervised learning can utilize a large amount of unlabeled data to perform the lesions segmentation task.Aiming at the problem of insufficient confidence in pseudo-labels in semi-supervised learning,UNet and DeepLabV3+were used as the basic network to build a co-training framework to obtain high-quality pseudo-labels by ensemble learning.The uncertainty of pseudo-labels in the JS distance measurement was introduced,a regular term was given to reduce the impact of low quality pseudo-labels on segmentation performance.After experimented on the public datasets,the proposed method obtains 76.06%Dice scores,65.1%IOU scores,77.22%sensitivity scores,and 81.46%precision scores.