Automated Segmentation of Metal and BVS Stent Struts from OCT Images Using U-Net

Abstract

AbstractPercutaneous Coronary Intervention (PCI) is an effective treatment for coronary artery diseases. PCI treatment is usually carried out with stent implantation to provide structural support to balloon dilated blood vessel, reducing risk of re-narrowing. Intravascular Optical Coherence Tomography (OCT) can provide a series of cross-section images depicting the internal structure of the artery and residing stent during PCI treatment. Stent struts segmentation for OCT images is necessary to provide quantitative data regarding quality of stent deployment during PCI and severity of restenosis during follow-up examination. Manual segmentation of stent struts is not efficient and infeasible due to large number of stent struts presented in each pullback of OCT images. Thus, automated stent struts segmentation is necessary to help clinicians in getting quantified data from OCT images within intraoperative time frame. In this paper, an automated stent strut segmentation algorithm was developed, utilizing 3D information of stent structure and state-of-the-art U-Net. The implementation of the algorithm preserves the spatial resolution of the full-size OCT images without down-sampling. The algorithm was trained and tested on both Bioresorbable Vascular Scaffold (BVS) and metal stent images. It achieved Dice’s coefficient of 0.82 for BVS images, precision of 0.90 and recall of 0.85 for metal stent images. This algorithm works for both BVS and metal stents OCT images and adapts to different stent conditions.KeywordsStent strut segmentationOptical coherence tomographyU-NetDeep learningPercutaneous coronary intervention