CathAI: fully automated coronary angiography interpretation and stenosis estimation

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Coronary artery disease (CAD) diagnosis relies heavily on coronary angiography.
• Visual assessment of angiograms for stenosis is subjective and variable.
• There is a need for objective and reproducible methods in CAD assessment.

Purpose of the Study:

• To develop and validate a fully automated approach for interpreting coronary artery stenosis from standard coronary angiograms.
• To assess the performance of the automated system (CathAI) in localization, estimation, and prediction of obstructive CAD.
• To evaluate the system's generalizability through internal and external validation.

Main Methods:

• Utilized 13,843 angiographic studies from UCSF for neural network training.
• Developed sequential neural networks for stenosis localization and estimation.
• Performed internal validation on hold-out test datasets.
• Conducted external validation using data from UOHI and MHI.
• Retrained models using quantitative coronary angiography (QCA) data.

Main Results:

• CathAI achieved state-of-the-art performance on real-world angiograms.
• High accuracy (≥90% PPV, sensitivity, F1 score) for projection angle identification and coronary artery detection.
• AUC of 0.862 for predicting obstructive CAD (≥70% stenosis) in the training set.
• External validation at UOHI showed an AUC of 0.869 for obstructive CAD prediction.
• Retrained model on MHI QCA data achieved an AUC of 0.775.

Conclusions:

• Multiple neural networks can be sequenced for automated analysis of real-world angiograms.
• The CathAI system demonstrates potential to enhance standardization and reproducibility in coronary stenosis assessment.
• Automated interpretation offers a promising tool for objective CAD diagnosis and management.