A multi-level contrastive learning-based correction model for semi-supervised fetal echocardiography segmentation

Abstract

Fetal echocardiography is an important test for diagnosing fetal congenital heart disease (CHD). The apical four-chamber view (A4C) is the most basic and essential view for fetal heart examination. Segmentation of key anatomical structures based on the A4C view provides a basis for functional assessment and disease diagnosis. However, current fetal heart segmentation methods do not fully leverage unlabeled intermediate frames in ultrasound videos. Furthermore, the complex and variable structure of the fetal heart often causes category confusion, significantly affecting the precision and reliability of segmentation. In this paper, we propose a multi-level contrastive learning-based correction model (MLCL-CM) for semi-supervised segmentation of five key cardiac structures in fetal echocardiography. MLCL-CM improves segmentation performance on labeled keyframes by exploiting the large number of unlabeled intermediate frames. To alleviate misclassification problems, we construct positive and negative pairs in category- and pixel-level contrastive learning and introduce two uncertainty quantization techniques to guide the learning of more robust and fine-grained discriminative feature representations in global and local spaces. In addition, we integrate an adaptive correction model to achieve more stable and consistent predictions by dynamically weighting the inconsistent regions. Experimental results on actual A4C views with 100 and 300 fetuses show that the proposed MLCL-CM outperforms the state-of-the-art methods.