CathAI: Fully Automated Interpretation of Coronary Angiograms using AI

Read the full paper at arxiv


Robert Avram, Jeffrey E. Olgin, Alvin Wan, Zeeshan Ahmed, Louis Verreault-Julien, Sean Abreau, Derek Wan, Joseph E Gonzalez, Derek Y So, Krishan Soni, Geoffrey H Tison


Coronary heart disease (CHD) is the leading cause of adult death worldwide. The minimally-invasive procedure called coronary angiography provides the primary gateway for nearly all CHD diagnosis and clinical management decisions, as well as stent-based therapy for CHD. 




We developed CathAI—a pipeline of multiple deep neural network algorithms—to accomplish automated interpretation of coronary angiograms.


Our work demonstrates that AI-based, fully-automated interpretation of real-world coronary angiograms can accurately estimate the severity of coronary artery narrowing, also called "stenosis".







CHD results from atherosclerotic (cholesterol-filled) plaques that can narrow the coronary arteries that feed the heart. This can limit blood flow to cardiac tissues and ultimately cause heart attacks. The decision to treat CHD with coronary stents, bypass surgery, or CHD medications alone, relies upon angiography to identify coronary artery stenoses greater than 70% in severity.

The current standard-of-care is for physicians to interpret angiograms using ad-hoc visual estimation of coronary stenosis severity. This has been unchanged for over 70 years and suffers from high inter-observer variability, operator bias and poor reproducibility. Variability in visual stenosis assessment ranges from 15 to 45%. 

A more standardized, reproducible approach to angiogram interpretation would have critical clinical importance.




To automate coronary angiogram interpretation, we developed a pipeline of multiple deep neural network algorithms of specific architectures to accomplish the necessary tasks for automated coronary stenosis assessment. Angiograms flow from one algorithm to the next.

The sequence of tasks includes:

   ►Classification of angiographic projection angle

   ►Classification of primary anatomic structure

   ►Localization of objects, i.e. artery segments and stenoses

   ►Estimation of severity of coronary stenosis




We assembled a dataset of ~200,000 angiogram videos from ~12,000 distinct individuals to develop and validate CathAI. 


CathAI algorithms demonstrated state-of-the-art performance for each of the required tasks necessary for automated angiogram interpretation.




To identify the overall projection angle, CathAI (Algorithm 1) demonstrated positive predictive value, sensitivity and F1 score of ≥90% for each. To identify left or right coronary artery (the primary task of Algorithm 2), the above performance metrics, respectively, were ≥93% for each.


CathAI demonstrated strong performance to identify clinically significant "obstructive" coronary artery stenosis, defined as ≥70% stenosis.

Algorithm 4 exhibited an area under the receiver operating characteristic curve (AUC) of 0.862 to predict obstructive coronary stenosis compared to expert clinical diagnosis. CathAI performance was strongest when predictions were averaged across videos visualizing the same artery segment.


To examine external generalizability, we validated CathAI in unselected real-world angiograms from the University of Ottawa Heart Institute, a separate healthcare institution in a different country. 


Two expert physician adjudicators independently interpreted coronary artery stenosis severity in this external dataset. Compared to the average of these stenosis estimates, CathAI achieved an AUC of 0.869 (95% CI: 0.830-0.907) for obstructive stenosis without any additional training. 


We used several algorithm explainability techqniues to better understand how the CathAI algorithms make their predictions.

CathAI algorithms tended to use well-understood regions of angiographic images to achieve intended tasks. For example, for Algorithm 2 to determine primary anatomic structure as being the left coronary artery, regions around the left anterior descending artery were most important.   →

Similarly, Algorithm 4 primarily used pixels around the areas of coronary artery stenosis to achieve its task of predicting stenosis severity.


This work demonstrates that multiple purpose-built neural networks can function in sequence to accomplish the complex series of tasks required for automated analysis of real-world angiograms, and achieve state-of-the-art results.

♦ Our validation of CathAI in both a hold-out UCSF test dataset and in a high-volume Canadian health system demonstrates that CathAI's strong overall performance generalizes to real-world external data.

♦ Deployment of CathAI may serve to increase standardization and reproducibility in coronary artery stenosis assessment—the lynchpin of CHD diagnosis and treatment.

♦ This pipeline provides an algorithmic foundation for many future tasks in automated angiogram interpretation, such real-time recommendations for additional procedures or testing.



Automated CHD assessment from angiograms by CathAI may serve to increase standardization in coronary stenosis assessment, one of the most critical junctures in CHD clinical decision making. 

CathAI achieves state-of-the-art performance for each task required for interpretation of real-world angiograms, providing a foundation to achieve a wide range of future tasks in automated angiogram interpretation.

The entire CathAI pipeline analyzes an angiogram within seconds on commonly-available GPU hardware, enabling real-time analysis during the angiogram procedure itself to assist physician interpretation.


Read the full paper at arxiv

Email questions or inquiries here.