Background: Posterior systolic curling (PSC) is a dynamic morphofunctional abnormality involving the posterior mitral annulus and adjacent inferobasal left ventricular wall, characterised by late-systolic regional hypermobility and associated with malignant ventricular arrhythmias and sudden cardiac death. Its diagnosis still relies on clinicians' qualitative visual assessment, and no validated, standardised quantitative approach is currently available. This study introduces a fully automated self-supervised learning framework for patient-level PSC detection from localised echocardiographic motion patterns.
Methods: Transthoracic echocardiographic parasternal long-axis cine loops from 100 patients (40 PSC, 60 noPSC; 44 females; median age 58 years) were retrospectively analysed. PSC diagnosis was determined by three independent clinicians based on dynamic visual assessment. The posterior annulus–inferobasal wall region was automatically localised using a convolutional neural network. Regional one-cycle motion was then encoded using optical flow, and fixed-length motion clips were extracted from each patient. A convolutional encoder combined with temporal attention was pretrained in a self-supervised manner using contrastive learning, masked modelling, and temporal ordering to learn localised motion dynamics without clip-level annotations. For downstream diagnosis, multiple clips from each patient were aggregated through top-k attention multiple-instance learning for patient-level prediction. Performance was assessed using nested stratified group cross-validation with patient-level separation.
Results: The proposed framework achieved a mean AUC of 0.913 ± 0.061, with mean accuracy, sensitivity, specificity, precision, and F1-score of 0.83 ± 0.05, 0.80 ± 0.06, 0.85 ± 0.06, 0.79 ± 0.08, and 0.79 ± 0.06, respectively. Temporal attention profiles identified the most informative portions of the cardiac cycle for classification.
Conclusions: Self-supervised learning of localised cardiac motion, combined with patient-level multiple-instance aggregation, provides a promising strategy for quantitative PSC detection from echocardiography. This framework may support earlier, more reproducible, and less operator-dependent identification of PSC, while illustrating the broader value of motion-based representation learning for limited-data cardiovascular imaging.