Introduction: Mitral valve prolapse (MVP) affects 2–3% of the general population and is a leading cause of mitral regurgitation and sudden cardiac death in young adults. Mitral annulus disjunction (MAD), a structural variant frequently co-occurring with MVP, is increasingly recognised as an independent arrhythmic risk factor. Despite their clinical significance, both conditions remain underdiagnosed at the population scale due to reliance on manual image review. Automated detection in large imaging cohorts could enable prevalence estimation and downstream risk stratification.
Methods: We developed a two-stage deep learning pipeline using the nnU-Net framework, applied to the three-chamber long-axis end-systolic CMR view. Stage one detected four anatomical landmarks to define a region of interest around the mitral valve. Stage two segmented the region bounded by the anterior and posterior mitral leaflets and the annular plane. A case was classified as screen-positive if either leaflet displaced beyond the annular plane by more than 2 mm in at least one of four ensemble predictions. The pipeline was trained on 440 cases (75 confirmed MVP from a clinical dataset; 365 UK Biobank controls) and applied to 60,895 UK Biobank participants across two batches.
Results: The screen-positive rate was 9.2% in the entire cohort. All 5,582 algorithmically positive cases underwent expert adjudication, alongside 1,991 sampled negatives. The false-negative rate in sampled negatives was 2.0%. Against expert adjudication (n=6,966), the pipeline achieved sensitivity of 0.999 and specificity of 0.575 for MVP or MAD detection, and sensitivity of 1.000 with positive predictive value of 0.461 for MVP alone. Adjudicated prevalence was 3.77% for MVP and 1.99% for isolated MAD.
Conclusion: This pipeline enables reliable, large-scale automated screening for MVP and MAD in population CMR cohorts, supporting both epidemiological research and cardiovascular risk stratification.