Potential of artificial intelligence to differentiate between non-obstructive and obstructive coronary artery disease on coronary CT angiography

Abstract Background Coronary computed tomography angiography (CCTA) plays an important role in the modern assessment of coronary artery disease (CAD). With the increasing number of CCTA examinations, technologies such as Artificial Intelligence (AI) that reduce reader workload and improve inter-reader agreement are of great interest. The Coronary Artery Disease Reporting and Data System (CAD-RADS) is a structured reporting system to categorise the severity of CAD and guide patient management. Patients with low CAD-RADS categories (1-2) usually need preventive medical therapies but no downstream testing. However, patients in more severe categories (CAD-RADS 3-5) may require further testing, such as stress tests or invasive coronary angiography. The aim of this study was to assess the potential of an on-site AI to differentiate between non-obstructive and potentially obstructive CAD. Methods In this study, 395 patients (65.3% male, 64.9 ± 10.0 years) who underwent a clinically indicated CCTA examination were retrospectively included. Cases with significant artefacts and coronary artery segments 2 mm in diameter were not included into the analysis. The severity of CAD was graded as non-obstructive (CAD-RADS ≤2) or potentially obstructive (CAD-RADS ≥3) by consensus of two highly experienced readers. The final study population consisted of 50.6% non-obstructive and 49.4% obstructive CAD patients. The CCTA reconstruction with the best image quality was selected, transferred to the on-site AI prototype and fully automatically processed to yield the CAD-RADS category. Cohen's kappa and intraclass correlation were used to compare the results of the readers and the AI prototype. Results In the subgroup with non-obstructive CAD (n=200, 58.0% male), there was high agreement between the AI and the readers, with 96.0% of patients being correctly classified by the AI. Accordingly, 92.8% of patients with obstructive CAD (n=195, 72.8% male) were correctly classified by the AI. The results showed very good agreement between human and AI results in the overall population, with a kappa of 0.89 (95% CI 0.84-0.94) and an overall ICC of 0.89 (95% CI 0.84-0.93). Conclusions AI can reliably differentiate between non-obstructive and obstructive CAD in cases with good image quality. This could help readers to identify low-risk patients and expedite CCTA analysis. However, AI still has some limitations, such as the identification of artefacts in cases with impaired image quality or the detection of totally occluded coronary arteries. Further development is needed to improve its diagnostic accuracy.