Exploring interpretable echo analysis using self-supervised parcels.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Cardiovascular Diagnostics

Background:

• Fully supervised deep learning for AI in echocardiography requires extensive labeled data, a significant bottleneck due to expert annotation needs.
• Existing methods struggle with the scarcity of labeled medical imaging datasets for training robust AI models.

Purpose of the Study:

• To explore self-supervised learning (SSL) for cardiac imaging analysis, focusing on interpretability, robustness, and safety.
• To develop an AI approach that reduces reliance on labeled data for predicting heart failure endpoints.

Main Methods:

• Utilized a modified Self-supervised Transformer with Energy-based Graph Optimisation (STEGO) network with a self-DIstillation with NO labels (DINO) backbone.
• Pre-trained the model on diverse medical and non-medical data to generate self-segmented outputs ('parcels') identifying cardiac substructures.
• Employed SSL on a large unlabeled dataset to discover features for downstream tasks and train smaller models.

Main Results:

• Self-learned 'parcels' demonstrated robustness across different patient profiles and cardiac cycle phases.
• These parcels provided high interpretability and effectively captured clinically relevant cardiac substructures.
• The approach showed adaptability across various requirements when evaluated on public datasets.

Conclusions:

• Self-supervised learning effectively addresses the challenge of labeled data scarcity in medical imaging.
• The proposed method enhances the efficiency and interpretability of cardiac imaging analysis and diagnostic procedures.
• This AI strategy holds significant potential for improving patient care and clinical decision-making in cardiology.