Abstract / Summary
Coronary artery disease (CAD) leads to considerable morbidity and mortality worldwide.Intravascular ultrasound (IVUS) is an imaging modality to guide percutaneous coronary intervention; however, it requires time-consuming, expert-dependent manual analysis.Artificial intelligence (AI) has been increasingly used for IVUS analysis, but there is limited evidence on its diagnostic accuracy and clinical utility.This review aimed to evaluate the diagnostic performance and clinical usefulness of AI for automated IVUS analysis of coronary arteries.The research protocol was prospectively registered (PROSPERO, CRD420261454806) and followed the PRISMA 2020 guidelines for systematic reviews.PubMed/MEDLINE and Cochrane CENTRAL were searched from inception through July 23, 2026, and July 26, 2026, respectively, for publications assessing AI, machine learning, or deep learning algorithms for the automated analysis of coronary IVUS, including segmentation of structures, identification of plaque types, and quantitative measurements.Two reviewers independently screened and selected eligible studies and extracted the data.The risk of bias was assessed using QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) and visualized using robvis, and the GRADE approach was used to rate the certainty of evidence.Of the 802 records identified (PubMed, n = 216; Cochrane CENTRAL, n = 586), eight studies met the inclusion criteria and were included in this review.They reported on the performance of convolutional neural networks, U-Net/U-Net++, Efficient-UNet, DeepLabv3+-based models, and gradient boosting in segmenting lumen, vessel, plaque, and stent in coronary IVUS.The Dice similarity coefficient ranged from 0.67 to 0.97, and the intersection over union for segmentation (including Jaccard similarity) ranged from 0.66 to 0.97, with the best performance for lumen and vessel segmentation and the poorest for stent and calcium.The sensitivity and specificity ranged from 62% to 90% and 88% to 99%, respectively, with the correlation to expert assessment generally above 0.9.Overall, the risk of bias was mainly related to patient selection (five out of eight studies) and the reference standard (three out of eight studies), and none of the studies was rated as having a high risk of bias.Due to between-study heterogeneity in the tasks, reference standards, and outcome measures, a meta-analysis was not performed, and the certainty of evidence was rated as low.The current evidence suggests that AI-based methods have good potential for analyzing coronary IVUS with acceptable accuracy, which could be helpful for high-throughput IVUS interpretation in clinical practice.They appear to perform best at segmenting lumen and vessel structures and worst at segmenting stent and calcium.However, due to the overall low certainty of evidence from these mostly retrospective studies, their use should be limited to complementing the work of experts in cardiovascular medicine.