A deep learning-based algorithm for detecting significant in-stent restenosis in coronary CT angiography.
Authors
Affiliations (9)
Affiliations (9)
- Department of Clinical Medical Sciences, Seoul National University College of Medicine, Seoul, Korea.
- Institute of Convergence Medicine with Innovative Technology, Seoul, Korea.
- Department of Radiology, Seoul National University Hospital, 101 Daehak-ro, Jongno-gu, Seoul, 03080, Korea.
- Department of Radiology, Seoul National University College of Medicine, Seoul, Korea.
- Department of Radiology, Seoul National University Hospital, 101 Daehak-ro, Jongno-gu, Seoul, 03080, Korea. [email protected].
- Department of Radiology, Seoul National University College of Medicine, Seoul, Korea. [email protected].
- Institute of Radiation Medicine, Seoul National University Medical Research Center, Seoul, Korea. [email protected].
- Department of Clinical Medical Sciences, Seoul National University College of Medicine, Seoul, Korea. [email protected].
- Institute of Convergence Medicine with Innovative Technology, Seoul, Korea. [email protected].
Abstract
Coronary CT angiography (CCTA) is widely used for monitoring patients after stent implantation, but beam-hardening and blooming artifacts make assessment of in-stent restenosis (ISR) challenging. This study developed and evaluated a deep learning algorithm for automated detection of significant ISR in CCTA images. A retrospective analysis of 257 patients with 408 stents was conducted, divided into a development set (n = 126 patients, 186 stents) and an internal validation set (n = 131 patients, 222 stents), with invasive coronary angiography serving as the reference standard. The algorithm's diagnostic performance was compared to two expert radiologists using standard diagnostic metrics. The deep learning algorithm achieved 87.8% accuracy (95% confidence interval [CI], 82.9-91.5%), with sensitivity of 75.4% (95% CI, 62.9-84.8%), specificity of 92.1% (95% CI, 87.0-95.3%), positive predictive value (PPV) of 76.8% (95% CI, 64.2-85.9%), and negative predictive value (NPV) of 91.6% (95% CI, 86.3-94.9%). McNemar's test revealed no significant differences in overall diagnostic accuracy between the algorithm and Reader1 (p = 0.133) or Reader2 (p = 0.150). Notably, the algorithm demonstrated significantly higher sensitivity compared to both readers (43.9% [95% CI, 31.8-56.7%] and 56.1% [955% CI, 43.3-68.2%], p < 0.05) and higher NPV compared to Reader1 (p = 0.026). Subgroup analyses confirmed higher sensitivity across different scanner platforms and imaging conditions, including GE systems (p = 0.016) and 80 kilovoltage peak settings (p = 0.006). The deep learning algorithm demonstrated higher sensitivity for detecting significant ISR while maintaining comparable overall diagnostic accuracy to expert radiologists, although this was accompanied by numerically slightly lower specificity, supporting its potential as a decision-support tool.