End-to-End, Pixel-Wise Vessel-Specific Coronary and Aortic Calcium Detection and Scoring Using Deep Learning

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Cardiovascular Disease Assessment

Background:

• Conventional methods for quantifying coronary artery calcium (CAC) and aortic calcium (AC) can be time-consuming and lead to information loss from medical images.
• There is a need for more efficient and accurate automated methods for CAC and AC assessment in patients without a history of coronary artery disease (CAD).

Purpose of the Study:

• To develop and validate an end-to-end deep learning model as an alternative to conventional CAC and AC identification methods.
• To assess the model's accuracy in segmenting vessel-specific CAC and AC and predicting Agatston scores.

Main Methods:

• A deep learning model was trained, tested, and validated on CT scans from 377 patients without a history of CAD using a 60:20:20 data split.
• The model's performance was evaluated using the Dice score for segmentation accuracy and correlation coefficients (rho) for Agatston score prediction.
• Automated segmentation time was recorded.

Main Results:

• The deep learning model achieved a high overall Dice score of 0.952, indicating robust segmentation performance.
• No significant differences in performance were observed between male and female patients, or between age groups (<65 and ≥65 years).
• The model demonstrated strong correlation and agreement for both CAC (rho = 0.876) and AC (rho = 0.947) prediction, with automated segmentation taking approximately 4 seconds per patient.

Conclusions:

• The developed deep learning model accurately and robustly identifies vessel-specific CAC and AC.
• The model's ability to predict Agatston scores that correlate well with manual annotations makes it a valuable tool.
• This automated approach offers a faster and potentially more informative alternative to conventional methods, with significant implications for research and clinical practice.