Self-Supervised Multi-Task Learning for the Detection and Classification of RHD-Induced Valvular Pathology.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Rheumatic heart disease (RHD) is a major global health concern requiring advanced diagnostic methods.
• Automated analysis of echocardiograms can significantly improve RHD diagnosis and management.

Purpose of the Study:

• To evaluate self-supervised multi-task learning (SSL) for automated echocardiographic analysis in RHD.
• To compare the performance of DINOv2 and SimCLR for predicting echocardiographic views, diagnosing RHD, and assessing severity.

Main Methods:

• Utilized two SSL methods: DINOv2 (vision-transformer) and SimCLR (ResNet-based).
• Pre-trained models on a large unlabelled echocardiogram dataset, followed by fine-tuning on a labelled subset.
• Employed UMAP and t-SNE for visualizing learned feature representations.

Main Results:

• DINOv2 achieved high accuracies: 92% for view classification, 98% for condition detection, 99% for severity assessment.
• SimCLR also showed strong performance: 99% for view classification, 92% for condition detection, 96% for severity assessment.
• Both models demonstrated effective feature capture, with distinct clusters observed in visualizations.

Conclusions:

• Self-supervised multi-task learning shows significant potential for automated echocardiogram analysis in RHD.
• This approach offers a scalable and efficient method for improving RHD diagnosis, particularly in resource-limited settings.