Weakly/Semi-supervised Left Ventricle Segmentation in 2D Echocardiography with Uncertain Region-Aware Contrastive Learning

Segmentation of the left ventricle in 2D echocardiography is essential for cardiac function measures, such as ejection fraction. Fully-supervised algorithms have been used in the past to segment the left ventricle and then offline estimate ejection fraction, but it is costly and time-consuming to obtain annotated segmentation ground truths for training. To solve this issue, we propose a weakly/semi-supervised framework with multi-level geometric regularization of vertex (boundary points), boundary and region predictions. The framework benefits from learning discriminative features and neglecting boundary uncertainty via the proposed contrastive learning. Firstly, we propose a multi-level regularized semi-supervised paradigm, where the regional and boundary regularization is favourable owing to the intrinsic geometric coherence of vertex, boundary and region predictions. Secondly, we propose an uncertain region-aware contrastive learning mechanism along the boundary via a hard negative sampling strategy for labeled and unlabeled data at the pixel level. The uncertain region along the boundary is omitted, enabling generalized semi-supervised learning with reliable boundary prediction. Thirdly, for the first time, we proposed a differentiable ejection fraction estimation module along with 2D echocardiographic left ventricle segmentation for end-to-end training without offline post-processing. Supervision on ejection fraction can also serve as weak supervision for the left ventricular vertex, region and boundary predictions without further annotations. Experiments on the EchoNet-Dynamic datasets demonstrate that our method outperforms state-of-the-art semi-supervised approaches for segmenting left ventricle and estimating ejection fraction.