ProtoASNet: Comprehensive evaluation and enhanced performance with uncertainty estimation for aortic stenosis...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Aortic stenosis (AS) is a common heart valve disease requiring precise diagnosis.
• Current automated AS classification uses black-box deep learning, limiting clinical trust.
• Need for interpretable AI in diagnosing AS severity from echocardiography.

Purpose of the Study:

• Introduce ProtoASNet, a prototype-based neural network for interpretable AS severity classification.
• Enhance trustworthiness and clinical adoption of AI in echocardiography analysis.
• Incorporate uncertainty estimation for improved diagnostic reliability.

Main Methods:

• Developed ProtoASNet, a neural network using learned spatio-temporal prototypes for AS classification.
• Predictions based on similarity scores between input videos and prototypes.
• Integrated abstention loss for aleatoric uncertainty estimation.

Main Results:

• ProtoASNet achieved 80.0% balanced accuracy on a private dataset and 79.7% on the TMED-2 dataset.
• Prototypes highlighted clinically relevant markers like calcification and leaflet movement.
• Discarding uncertain cases improved accuracy to 82.4% on the private dataset.

Conclusions:

• ProtoASNet offers an interpretable and trustworthy AI solution for AS severity classification.
• The model provides visual evidence and uncertainty quantification, aiding clinical decision-making.
• This approach facilitates interactive use of deep learning in echocardiography.