Artificial intelligence-based analysis of stent malapposition and expansion from intracoronary optical coherence tomography

Background

Intracoronary optical coherence tomography (OCT) can be used to assess stent malapposition and its expansion rates, which are critical measurements of percutaneous coronary intervention (PCI) success. Recent advances in artificial intelligence (AI) have demonstrated the potential to provide automated, rapid, and accurate post-PCI stent evaluation, potentially improving workflow efficiency and diagnostic confidence.

Purpose

To develop a rapid, automated AI system for stent strut detection and perform subsequent stent expansion and malapposition analysis simultaneously from OCT images after PCI procedures.

Methods

This study developed a deep learning model consists of two convolutional neural networks. The first neural network was trained (N=8640 frames from 16 pullbacks) to segment individual stent struts within OCT frames, while the second one was trained (N=4300 frames from 290 pullbacks) to delineate the vessel lumen afterwards. The segmented stent struts were then used to calculate frame-wise stent expansion, defined as the difference in spline area between the reference and predicted strut expansion projection. Malapposition was identified by detecting struts in the lumen with a distance larger than 0.2 mm from the vessel wall. Manually annotated stents struts and malappositions combining with spline interpolated stent expansion areas were used as reference.

Results

The system was evaluated on a held-out test set comprising 2160 OCT frames from 4 patients. The stent strut detection model achieved an overall precision of 0.913 and a recall of 0.828 across all frames, indicating high accuracy in identifying stent struts. The stent expansion area error was 0.019 mm2 [0.007-0.86] (median [Q1-Q3]) across all analysed frames, providing a quantitative measure of stent placement success. Malapposition detection achieved an overall precision of 0.926 and a recall of 0.910 on stent strut level.

The entire pipeline, including image preprocessing, strut segmentation, lumen contour delineation, expansion calculation, and malapposition assessment, was completed in under 20 seconds for pullbacks of 540 frames.

Conclusion

This study demonstrates that AI-based stent detection and evaluation, based on OCT imaging, can be performed rapidly and with high accuracy to identify struts, calculate frame-wise stent expansion and detect malapposition. Unlike commercial OCT analysis software, which is typically specialised to single case online study during the procedure, our system also enables retrospective analysis of large numbers of OCT pullbacks, providing comprehensive research and quality improvement initiatives. While this evaluation was performed on a limited number of patients, the results suggest that AI-based OCT analysis could reduce workload for interventional cardiologists. Future research will focus on expanding this system to larger, diverse patient populations to evaluate its impact on downstream decisionsFor image description, please refer to the figure legend and surrounding text.