Automated deep learning-based comprehensive assessment of stent in intravascular optical coherence tomography

Abstract

The process of manually identifying and quantifying coronary stents in intravascular optical coherence tomography (IVOCT) images is time-consuming and subjective. In this study, a deep learning-based target recognition model, SMA-YOLO, was developed for automatic identification and quantitative assessment of coronary stents. Based on YOLOv8, the model combines adjacent-frame information, a customized feature-extraction backbone, a multi-scale dynamic feature pyramid network, and an adaptive task decomposition and alignment detection head. A retrospective multicenter dataset comprising 118 unique patients, 181 pullbacks, and 7,732 frames from three hospitals was used for model development and evaluation. SMA-YOLO achieved precision and recall of 98.0% and 97.4% on the internal test set and 97.0% and 96.0% on the independent external test set, respectively. The model maintained high performance across the evaluated stent materials, postoperative stages, complex clinical scenarios, and two OCT systems. Detected struts were further used to measure stent area, malapposition distance, and neointimal thickness and to generate two- and three-dimensional visualizations. These automatic measurements agreed closely with those of two senior experts (all r > 0.98). Model-assisted annotation also reduced expert labeling time and improved the assessment accuracy of junior physicians. The results show that SMA-YOLO supports automatic stent identification, quantitative measurement, and spatial visualization in IVOCT.