Consensus-guided evaluation of self-supervised learning in echocardiographic segmentation

Left ventricle segmentation is a fundamental task in echocardiography, essential for assessing cardiac function. However, deep learning models for segmentation rely on large labelled datasets, which are expensive and time-consuming to annotate. Self-supervised learning has emerged as a promising approach to leverage unlabelled data, but its effectiveness for left ventricle segmentation remains underexplored.

This study investigates self-supervised learning for echocardiographic segmentation, comparing various pretext tasks. The impact of dataset size and distribution on pre-training is examined, revealing that excessive unlabelled data can degrade performance due to redundancy and low variability. A novel multi-expert labelled dataset is introduced to enhance segmentation evaluation, using consensus-based annotations to reduce annotation noise and improve reliability.

Among the self-supervised learning methods evaluated, contrastive learning consistently outperforms other approaches, particularly in low-label settings. The study demonstrates that AI models pre-trained using self-supervised learning and fine-tuned with only 15% of labelled data achieve stronger alignment with multi-expert consensus than any individual expert.

The findings suggest that AI models can generalise well across expert annotations, providing more reliable and reproducible assessments.