Deep Learning Model Automates Coronary Calcium Scoring with High Accuracy

A newly published open-access study introduces a multi-task deep-learning model for automated segment-level coronary artery calcium (CAC) scoring on non-contrast CT. The work, led by Marc Dewey and co-authored by Bernhard Föllmer, Sotirios Tsogias, Federico Biavati, and others, is based on data from over 1,500 patients in the multicenter DISCHARGE trial.

The proposed neural network performed both calcification segmentation and segment-level artery identification, achieving strong results:

  • Sensitivity: 0.732
  • Specificity: 0.978
  • F1-score: 0.717
  • Cohen’s κ (segment-level agreement): 0.808 – nearly equivalent to interobserver agreement (0.809)

These findings confirm that deep learning can reliably automate complex CAC scoring tasks, reducing manual effort and observer variability.

Congratulations to Bernhard Föllmer for this important contribution to clinical AI and cardiovascular imaging.

The full article is available in Volume 15 of Insights into Imaging.