ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


ERS2313. Feasibility and Diagnostic Performance of a New CCTA-Derived and AI-Based Fully Automated System for Detection of Coronary Artery Disease
Authors * Denotes Presenting Author
  1. Verena Brandt; Medical University of South Carolina; Robert-Bosch-Hospital
  2. Akos Varga-Szemes; Medical University of South Carolina
  3. U. Joseph Schoepf; Medical University of South Carolina
  4. Raffi Bekeredjian; Robert-Bosch-Hospital
  5. Tilman Emrich; Medical University of South Carolina
  6. Josua Decker; Medical University of South Carolina; University Hospital Augsburg
  7. Gilberto Aquino *; Medical University of South Carolina
Objective:
To evaluate a novel coronary CT angiography (CCTA)-derived fully automated artificial intelligence (AI)-based software solution for automated coronary artery segmentation and stenosis assessment using the Coronary Artery Disease Reporting & Data System (CAD-RADS).

Materials and Methods:
Image datasets of 100 consecutive patients (48% male, 48.3±10.8 years) who underwent clinically indicated CCTA were retrospectively analyzed. Two readers independently evaluated CCTAs for the degree of coronary artery stenosis on a per-segment level using the 18-coronary artery segment model with subsequent CAD-RADS classification according to SCCT guidelines. A fully automated investigational AI-based software prototype by Siemens was designed and tested on the CCTA datasets and compared to human reading. Inter-reader agreement was assessed using Cohen’s kappa. Subsequently, the diagnostic performance of the software prototype for detection of diseased coronary artery segments were assessed.

Results:
Forty-one patients had coronary artery disease (CAD) with stenosis in at least one segment. Agreement between expert readers was 0.83 for CAD-RADS and 0.89 for the identification of diseased segments. The software prototype yielded a sensitivity of 97.6% (92.8-100), and a negative predictive value of 96.9% (90.8-100) for the detection of diseased segments, respectively. The software prototype reliably detected 40 out of 41 patients with CAD. One patient who was not correctly identified had a small, calcified plaque without associated coronary artery stenosis (CAD-RADS 1). The average computational time of the software prototype was 240s per case.

Conclusion:
The fully automated investigational AI-based software prototype demonstrated fast and reliable identification of patients with coronary artery stenosis on CCTA with high diagnostic accuracy.