Self-supervised learning for label-free segmentation in cardiac ultrasound

Segmentation and measurement of cardiac chambers from ultrasound is critical, but laborious and poorly reproducible. Neural networks can assist, but supervised approaches require the same problematic manual annotations. We build a pipeline for self-supervised segmentation combining computer vision, clinical knowledge, and deep learning. We train on 450 echocardiograms and test on 18,423 echocardiograms (including external data), using the resulting segmentations to calculate measurements. Coefficient of determination (r²) between clinically measured and pipeline-predicted measurements (0.55-0.84) are comparable to inter-clinician variation and to supervised learning. Average accuracy for detecting abnormal chambers is 0.85 (0.71-0.97). A subset of test echocardiograms (n = 553) have corresponding cardiac MRIs (the gold standard). Correlation between pipeline and MRI measurements is similar to that of clinical echocardiogram. Finally, the pipeline segments the left ventricle with an average Dice score of 0.89 (95% CI [0.89]). Our results demonstrate a manual-label free, clinically valid, and scalable method for segmentation from ultrasound.