Motion-corrected coronary calcium scores by a convolutional neural network: a robotic simulating study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• Coronary calcium scoring on non-triggered CT is susceptible to motion artifacts, leading to inaccurate Agatston scores.
• Accurate assessment of coronary calcifications is crucial for cardiovascular risk stratification.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) for classifying motion-induced blurred images of coronary plaques.
• To correct coronary calcium scores obtained from non-triggered chest CT scans using the developed CNN.

Main Methods:

• An Inception v3 CNN was trained using CT images of artificial coronary arteries with calcified plaques subjected to motion artifacts (0-90 mm/s).
• The CNN classified blurred images into nine categories based on plaque density and size.
• Agatston scores derived from CNN predictions were used as corrected scores, with variations calculated against scores from static scans.

Main Results:

• The CNN achieved an overall classification accuracy of 79.2% for nine plaque categories, with higher accuracy for high-density plaques (88.3%).
• CNN correction reduced the median Agatston score variation in moving plaques from 37.8% to 3.7% (p < 0.001).
• Sensitivity for detecting coronary calcifications improved from 65% to 85% post-correction.

Conclusions:

• A deep CNN effectively classifies motion-induced blurred coronary plaque images and corrects coronary calcium scores on non-triggered CT.
• CNN-based correction significantly reduces Agatston score variability and enhances the sensitivity of calcification detection.
• This approach represents a significant step towards improving the clinical accuracy of coronary calcium scoring.