Key result
Self-supervised AI fine-tuned with 15% labelled data outperforms individual experts in LV segmentation consensus.
Why the study?
Deep learning models for left ventricle segmentation require expensive, time-consuming labelled datasets, and the effectiveness of self-supervised learning using unlabelled data remains underexplored.
Does self-supervised learning pre-training improve echocardiographic left ventricle segmentation accuracy and ejection fraction estimation when annotated training data is limited?
Does self-supervised learning pre-training improve echocardiographic left ventricle segmentation accuracy and ejection fraction estimation when annotated training data is limited?
Self-supervised learning significantly reduces the manual annotation burden required to develop robust echocardiographic segmentation models, achieving high clinical agreement even in data-scarce settings.
May enable label-efficient LV segmentation models; hypothesis-generating pending prospective validation and outcome studies.
BACKGROUND: 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. METHODS: 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. RESULTS: 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. CONCLUSION: The findings suggest that AI models can generalise well across expert annotations, providing more reliable and reproducible assessments.
No takes yet. Share an insight, caveat, or question.
Naidoo et al. (2025) studied Echocardiographic left ventricle segmentation (n=1,824). Self-supervised learning (Contrastive learning/SimCLR) vs. Randomly initialized baseline model (purely supervised) was evaluated on Ejection fraction (EF) error at 5% labelled training data. AI models pre-trained using self-supervised learning and fine-tuned with 15% of labelled data achieve stronger alignment with multi-expert consensus for left ventricle segmentation than individual experts.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: