Development of a deep learning-based algorithm for the automatic detection and quantification of aortic valve calcium

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Radiology
• Medical Image Analysis

Background:

• Aortic valve calcium (AVC) quantification is crucial for assessing cardiovascular risk.
• Manual quantification of AVC from CT scans can be time-consuming and subject to inter-observer variability.
• Automated methods using deep learning (DL) offer a potential solution for objective and efficient AVC assessment.

Purpose of the Study:

• To develop and validate a DL-based algorithm for automated quantification of AVC volume and Agatston score.
• To compare the performance of DL-derived metrics against manual measurements by radiologists for AVC severity classification.

Main Methods:

• Retrospective analysis of 589 non-enhanced, ECG-gated cardiac CT scans.
• Development of a DL algorithm for AVC segmentation and volume quantification.
• Calculation of Agatston score from attenuation values.
• Comparison of DL performance with four radiologists' visual gradings in two reading rounds.

Main Results:

• The DL algorithm achieved a Dice coefficient of 0.807 for AVC segmentation.
• DL-based AVC volume quantification showed 97.0% accuracy (AUC 0.964) for severe AVC classification, outperforming radiologists (accuracy 69.7%-91.9%, AUC 0.762-0.923).
• DL-based Agatston score achieved 92.9% accuracy (AUC 0.933), also superior to radiologists (accuracy 77.8-89.9%, AUC 0.791-0.903).

Conclusions:

• Deep learning-based automated quantification of AVC is comparable to manual measurements.
• The DL algorithm demonstrates superior diagnostic performance in classifying severe AVC compared to radiologist readers.