Efficient deep learning-based automated diagnosis from echocardiography with contrastive self-supervised learning.

Area of Science:

• Artificial Intelligence
• Cardiology
• Medical Imaging

Background:

• Self-supervised learning (SSL) advances automated medical image diagnosis using limited labeled data.
• Existing SSL methods are not optimized for video-based modalities like echocardiography.
• Expert labeling for medical imaging is challenging and time-consuming.

Purpose of the Study:

• Develop a novel SSL approach for echocardiogram videos.
• Learn robust representations for efficient fine-tuning in cardiac disease diagnosis.
• Improve label efficiency in automated medical image analysis.

Main Methods:

• Introduced EchoCLR, a self-supervised contrastive learning framework for echocardiogram videos.
• Utilized contrastive learning to identify distinct patient videos.
• Incorporated frame reordering to predict correct video sequence.

Main Results:

• EchoCLR pretraining significantly improved classification performance for left ventricular hypertrophy (LVH) and aortic stenosis (AS).
• Achieved 0.72 AUROC for LVH using 10% labeled data, outperforming standard transfer learning (0.61 AUROC).
• Achieved 0.82 AUROC for severe AS using 1% labeled data, outperforming transfer learning (0.61 AUROC).

Conclusions:

• EchoCLR uniquely learns representations from echocardiogram videos.
• Demonstrates the efficacy of SSL for label-efficient disease classification.
• Highlights the potential of EchoCLR for automated cardiac diagnostics.