Automated Agatston Score Computation in non-ECG Gated CT Scans Using Deep Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• The Agatston score is a key metric for cardiovascular disease risk, traditionally calculated from CT scans by analyzing atherosclerotic plaques.
• Current methods involve complex segmentation of Coronary Artery Calcifications (CACs).

Purpose of the Study:

• To develop a convolutional neural network (CNN) capable of directly predicting the Agatston score from non-contrast chest CT scans.
• To eliminate the requirement for prior CAC segmentation in Agatston score calculation.

Main Methods:

• Utilized a dataset of 5973 non-contrast chest CT scans with pre-computed Agatston scores.
• Employed an object detector for automatic heart cropping, followed by training a 3D deep CNN.
• The network was trained to regress the Agatston score directly from the cropped heart images.

Main Results:

• Achieved a high Pearson correlation coefficient (r = 0.93, p ≤ 0.0001) against manual Agatston scores in a test set of 1000 cases.
• Successfully stratified 72.6% of cases into standard cardiovascular risk groups.
• Performance is comparable to state-of-the-art methods that require CAC segmentation.

Conclusions:

• A CNN can accurately estimate the Agatston score directly from cardiac images without CAC segmentation.
• This represents a simplified and novel approach to Agatston score computation.
• The method shows comparable results to existing, more complex techniques.