基于跨任务一致性的半监督肝脏CT图像分割

Semi-supervised Liver CT Image Segmentation Based on Cross-task Consistency

目前基于深度学习的医学图像分割方法往往需要大量带标记数据训练网络模型,然而医学影像的标记数据获取通常非常昂贵,半监督学习能使模型利用大量未标记数据和少量标记数据学习。该文提出了一种基于跨任务一致性的半监督学习框架来降低神经网络模型训练时需要的标记数据成本。该方法利用V-Net网络作为主干框架并添加两个辅助解码器,同时在解码器中引入一个辅助回归任务,提高模型分割性能,并在主副解码器的分割任务和回归任务之间施加正则化约束的跨任务一致性损失,该框架能够学习到大量未标记数据的几何先验信息。在LiTS2017 Challenges数据集上验证了该方法的有效性。在使用20%标记数据的实验中,该方法的Dice系数和Jaccard指数分别达到了93.95%和88.87%,相比全监督V-Net网络模型训练下的Dice系数和Jaccard系数分别提高了3.60百分点和5.78百分点。实验结果表明,该方法在使用少量带标记数据情况下达到接近100%带标记数据训练分割肝脏的精度,与其他的半监督方法相比分割精度更优。

Current deep learning-based medical image segmentation methods often require large amounts of labeled data to train network models.However,labeled data acquisition for medical images is usually quite expensive,and semi-supervised learning enables models to learn using large amounts of unlabeled data and small amounts of labeled data.In this paper,a semi-supervised learning framework based on cross-task consistency is proposed to reduce the cost of labeled data required for neural network model training.The method utilizes the V-Net network as the backbone framework and adds two auxiliary decoders,while introducing an auxiliary regression task in the decoder to improve the model segmentation performance and imposing a regularization-constrained cross-task consistency loss between the segmentation and regression tasks of the primary and secondary decoders,which is able to learn a large amount of unlabeled data geometric prior information.We validate the effectiveness of the proposed method on the LiTS2017 Challenges dataset.The Dice coefficient and Jaccard index of the proposed method reached 93.95%and 88.87%,respectively,in the experiments using 20%labeled data,which increased by 3.60 percentage points and 5.78 percentage points compared with the fully supervised V-Net network model training.The experimental results show that the proposed method achieves the accuracy of segmenting the liver with a small amount of labeled data close to 100%labeled data training,and the segmentation accuracy is better compared with other semi-supervised methods.