Deep learning model to quantify left atrium volume on routine non-contrast chest CT and predict adverse outcomes

Area of Science:

• Cardiology
• Radiology
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) is standard for lung cancer screening, generating substantial unanalyzed non-pulmonary data.
• Left atrial (LA) volume is a key cardiovascular indicator, but its measurement from LDCT scans is not routine.

Purpose of the Study:

• To validate a deep learning (AI) model for automated segmentation and measurement of LA volumes from LDCT scans.
• To assess the predictive capability of AI-derived LA volume index (AI-LAVi) for cardiovascular outcomes.

Main Methods:

• Retrospective analysis of 273 patients undergoing LDCT for lung cancer screening.
• Quantification of LA volumes by expert radiologists and a prototype AI algorithm.
• Comparison of AI and expert LA volume index (LAVi) and evaluation of their association with 5-year cardiovascular outcomes using logistic regression.

Main Results:

• Excellent correlation between AI and expert LA volume measurements (ICC=0.950).
• AI-LAVi demonstrated significant associations with new-onset atrial fibrillation (OR 1.12), heart failure hospitalization (OR 1.07), and major adverse cardiac and cerebrovascular events (MACCE) (OR 1.04).

Conclusions:

• A novel deep learning algorithm accurately quantifies LA volume from lung cancer screening LDCT scans.
• AI-LAVi is a significant predictor of future cardiovascular events, including atrial fibrillation, HF hospitalization, and MACCE.