Impact of Deep Learning-based Artificial Intelligence Assistance on Reader Agreement in Coronary CT Angiography Interpretation.

Abstract

Purpose To evaluate the impact of a fully automated, multitask deep learning (DL) algorithm on interreader agreement of coronary artery disease (CAD) detection and stenosis classification using coronary CT angiography (CCTA). Materials and Methods This retrospective study included CCTA examinations (n = 623 patients) performed for clinical indications on CT systems from multiple vendors between January 2010 and December 2019. An expert reader (reader 1) analyzed all CCTA scans manually and with artificial intelligence (AI)-assisted reading at the lesion, coronary segment, and patient levels using the CAD Reporting and Data System (CAD-RADS). The AI algorithm detected, quantified, and classified coronary lesions. Interreader agreement was evaluated using a second expert reader (reader 2), who analyzed a randomly selected subset of 274 patients. CAD-RADS scores from radiologist reports (reader 3) were available for 362 patients. In a subgroup of 30 patients with disagreements, R2 also interpreted the cases using AI assistance. Agreement between readings, with and without AI, was assessed using Spearman correlation, and logistic regression and mixed models evaluated the impact of AI-assisted reading on CAD-RADS classification. Results The final study sample included 11 214 coronary segments analyzed from 623 patients (mean age ± SD, 54.8 years ± 15.7; 341 male). Of these patients, 295 (47.9%) had no CAD (CAD-RADS 0), 213 (33.6%) had low risk of coronary obstruction (CAD-RADS < 3), and 115 (18.5%) had high risk of obstructive disease (CAD-RADS ≥ 3). With AI assistance, reader 1 demonstrated improved agreement with reader 2 (ρ = 0.899-0.949; P < .001) and reader 3 (ρ = 0.889-0.938; P < .001). In the subgroup with reader 1-AI disagreement, agreement between reader 1 and reader 2 was low with manual readings (ρ = 0.688) but increased substantially when both readers used AI-assisted reading (ρ = 0.975; P < .001). Conclusion AI-assisted reading using a DL algorithm significantly improved interreader agreement for CAD-RADS classification at CCTA. Keywords: Applications - CT, CT-Coronary Angiography, Deep Learning Supplemental material is available for this article. © RSNA, 2025.