基于对比学习的非对称编解码结构的心脏MRI分割研究

Research on Cardiac MRI Segmentation Based on Asymmetric Encoding and Decoding Structure of Contrastive Learning

全监督方法在心脏磁共振成像(MRI)分割任务中的成功依赖于大规模标记数据集,然而由于患者隐私及人工标注困难等问题,心脏MRI标注数据规模较小,使全监督方法面临挑战。基于半监督的对比学习方法,设计双分支编码与单分支解码的心脏MRI分割网络CPCL-Net,引入图像和像素的联合对比损失,提升了模型对数据样本的特征表达能力。为了增强CPCL-Net对Hard负样本的分割精度,设计动态自适应加权模块(DAWM),利用生成的α和β权重因子评估样本级别和像素级别的训练贡献度,使模型分割精度得到大幅提升。基于自动心脏诊断挑战赛(ACDC)数据集的实验结果表明,该网络模型仅利用少量标注样本即可获得较高的分割精度,缓解了心脏MRI高质量标注样本不足导致的分割精度低下问题,并且在标注样本规模相同的情况下,对心脏左心室、右心室、心肌等部位的分割精度分别为86.17%、85.52%、84.55%,优于现有的4种典型半监督分割模型以及经典的对比学习框架Sim-CLR,有效缓解了全监督分割模型对样本规模的依赖及过拟合问题。

The success of a fully supervised method in cardiac Magnetic Resonance Imaging(MRI)segmentation tasks relies on large-scale labeled datasets.However,owing to problems associated with patient privacy and difficulty in manual labeling,the small scale of cardiac MRI annotation data makes the fully supervised method challenging.Based on the semisupervised contrastive learning method,a cardiac MRI segmentation network,CPCL-Net,with double-branch coding and single-branch decoding is developed,and the joint contrastive loss of images and pixels is introduced to improve the ability of the model to express the features of the data samples.To enhance the segmentation accuracy of CPCL-Net for Hard negative samples,a Dynamic Adaptive Weighting Module(DAWM)is developed,and the training contribution at the sample and pixel levels is evaluated using the generated and weight factors,which considerably improves the segmentation accuracy of the model.Experimental results based on the Automated Cardiac Diagnosis Challenge(ACDC)dataset indicate that the network model can obtain high segmentation accuracy with only a small number of annotated samples,which alleviates the problem of low segmentation accuracy caused by insufficient high-quality annotated samples of cardiac MRI.The segmentation accuracies for the left ventricle,right ventricle,and myocardium of the heart are 86.17%,85.52%,and 84.55%,respectively,under the same annotation sample size.The accuracy values are greater than those of the four existing semi-supervised segmentation models and the classical contrastive learning framework,Sim-CLR,which effectively alleviates the dependence and overfitting of the fully supervised model on the sample size.