Semi-supervised domain adaptation incorporating three-way decision for multi-view echocardiographic sequence segmentation

Abstract

Multi-view echocardiographic sequence segmentation is essential for the diagnosis of cardiac diseases in clinical practice. However, the variation in cardiac structures in different views and the lack of manual annotations make it challenging to establish a generalized segmentation model. In this paper, we propose a Bidirectional semi-supervised domain adaptation (BSDA) method based on the three-way decision to learn a generalized segmentation model for different views. Specifically, the two-branch structure of BSDA regards echocardiographic sequence data of different views as different domains. The source-pretrained model first roughly predicted the segmentation results for the target domain. Then, BSDA provides the segmentation model with reliable probabilistic supervision and feature-based pseudo-labels to make secondary decisions. Besides, the proposed stage-wise training strategy can better cope with the varied appearance of the cardiac structures in echocardiographic sequence. We evaluate our BSDA on three publicly available datasets, corroborating the superiority of BSDA to segment cardiac structures of multi-view echocardiographic sequences.