Interpretable deep learning prediction of 3d assessment of cardiac function

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning is increasingly used in medical decision-making, necessitating explainability for regulatory compliance and user trust.
• Accurate assessment of left-ventricular function in cardiac imaging is vital for patient risk stratification and cardiovascular disease diagnosis.
• Existing video-based methods offer high accuracy but lack explainability and clinical workflow integration.
• Explainable 2D semantic segmentation methods improve transparency but compromise accuracy.

Purpose of the Study:

• To develop a deep learning approach for cardiac imaging that is both highly accurate and explainable.
• To create a method that integrates seamlessly with the standard clinical workflow for evaluating left-ventricular function.
• To enhance trust and transparency in AI-driven medical decision-making.

Main Methods:

• A frame-by-frame 3D depth-map approach was developed for predicting left-ventricular ejection fraction.
• The method utilizes the conventional clinical workflow, including the method of discs for left ventricular volume evaluation.
• The approach ensures that predictions are interpretable by clinicians.

Main Results:

• The proposed 3D depth-map approach achieved high accuracy with a mean absolute error of 6.5%.
• The method demonstrated superior reproducibility compared to human evaluation.
• Generated volume predictions are interpretable, allowing for clinical intervention and adjustment.

Conclusions:

• The frame-by-frame 3D depth-map approach successfully balances accuracy and explainability in cardiac imaging.
• This method aligns with clinical workflows, offering a more trustworthy and transparent alternative to previous deep learning techniques.
• The approach provides reproducible and interpretable results, facilitating better clinical decision-making in cardiovascular disease management.