基于最优传输特征选择的医学图像分割迁移学习

Transfer learning of medical image segmentation based on optimal transport feature selection

在无监督领域自适应迁移学习过程中,域无关特征导致模型分割性能下降,而目前并没有针对迁移学习分割模型有效的特征选择方法。为解决该问题,提出了一个基于最优传输的迁移学习通用特征选择模块,可以应用到多种无监督领域自适应图像分割模型中。该模块利用分割准确性加权最优传输选择两个域的最优样本子集,再将样本子集特征进行熵正则化最优传输,得到两个域特征相似性降序列表来去掉域无关特征。将通用特征选择模块应用到三种无监督领域自适应模型中解决新冠肺炎图像分割问题,均在一定程度上提升了模型性能。

In the unsupervised domain adaptive transfer learning process,domain-independent features lead to the degradation of model segmentation performance,but there is no effective feature selection method for transfer learning segmentation model at present.To solve this problem,a general feature selection module for transfer learning was proposed based on optimal transport,which can be applied to various unsupervised domain adaptive image segmentation models.In this module,the optimal sample subsets of two domains are selected by weighted optimal transport of segmentation accuracy,and then the features of sample subsets are subjected to entropy regularized optimal transport,so as to obtain a descending list of similarity between two domains to remove domain-independent features.The universal feature selection module is applied to three unsupervised domain adaptive models to solve the problem of Covid-19 image segmentation,which improves the model performance to a certain extent.