Temporal-Consistent Segmentation of Echocardiography with Co-learning from Appearance and Shape

Accurate and temporal-consistent segmentation of echocardiography is important for diagnosing cardiovascular disease. Existing methods often ignore consistency among the segmentation sequences, leading to poor ejection fraction (EF) estimation. In this paper, we propose to enhance temporal consistency of the segmentation sequences with two co-learning strategies of segmentation and tracking from ultrasonic cardiac sequences where only end diastole and end systole frames are labeled. First, we design an appearance-level co-learning (CLA) strategy to make the segmentation and tracking benefit each other and provide an eligible estimation of cardiac shapes and motion fields. Second, we design another shape-level co-learning (CLS) strategy to further improve segmentation with pseudo labels propagated from the labeled frames and to enforce the temporal consistency by shape tracking across the whole sequence. Experimental results on the largest publicly-available echocardiographic dataset (CAMUS) show the proposed method, denoted as CLAS, outperforms existing methods for segmentation and EF estimation. In particular, CLAS can give segmentations of the whole sequences with high temporal consistency, thus achieves excellent estimation of EF, with Pearson correlation coefficient 0.926 and bias of 0.1%, which is even better than the intra-observer agreement.