Challenges and Strategies for Deep Learning in Cardiovascular Imaging: Ejection Fraction and Heart Failure...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning (DL) shows promise for automated cardiac measurements.
• Clinical translation of DL methods faces implementation challenges.

Purpose of the Study:

• Evaluate challenges in DL-based cardiac measurements using left ventricular ejection fraction (LVEF) for heart failure (HF) management.
• Discuss mitigation strategies for identified DL challenges.

Main Methods:

• Utilized supervised end-to-end learning for automated LVEF measurements across three diverse populations (N=3,538).
• Investigated challenges related to evaluation metrics, training data, and model generalization.
• Analyzed DL model performance in the context of HF management.

Main Results:

• Evaluation metrics like AUC showed unreliability for dichotomized HF diagnosis due to population variations.
• Model performance improved even with a 40% reduction in training data.
• Generalization challenge observed with performance degradation on external data; domain knowledge integration improved performance.

Conclusions:

• Training data characteristics and model generalization pose significant challenges for DL in cardiac imaging.
• Evaluation metrics can obscure underperforming DL algorithms.
• Integrating imaging domain knowledge and careful consideration of data enhance DL model design, evaluation, and clinical applicability.