Study On The Automated Segmentation Of Bioresorbable Vascular Scaffold In Intravascular Optical Coherence Tomography Images

Coronary artery disease(CAD)has become one of the diseases with the highest mortality rate globally in the last few years.The number of death due to CAD in china ranks secondly in the word.Percutaneous coronary intervention(PCI)is one of the principal treatment options for CAD,which relieves the narrowing of coronary artery by coronary catheterization and keeps the blood vessel open by stent implantation.Absorb Bioresorbable vascular scaffold(BVS)is considered to be the most promising stents since BVS could offer temporary radial strength and be fully absorbed at a later stage.However,stent thrombosis may result from malapposition when BVS is implanted improperly and stent thrombosis is potential problem.Therefore,struts apposition analysis in real time is important to ensure the safety of the implanted BVS and has great value in clinical application.Intracavitary imaging is required for assistant diagnosis in percutaneous coronary intervention(PCI).Intravascular optical coherence tomography(IVOCT),due to its high axial resolution,is currently the most advanced imaging modality that enables BVS malapposition analysis.However,it is time-consuming and labor-intensive for expert to analysis one pullback containing thousands of stents.Computer-aided analysis can provide automatic and effective struts analysis in real time.In this article,automatic BVS struts segmentation algorithms are proposed.Struts malapposition analysis can be obtained based on the struts contour and coronary artery lumen contour.Traditional machine learning algorithm and deep learning algorithm are used to segment BVS automatically.In traditional machine learning method,a novel framework of automatic struts segmentation based on four corners is proposed in consideration of the prior knowledge that strut has an evident feature of box-shape.Firstly,Adaboost algorithm based on the extended Haar-like feature is used to train a corner classifier.Secondly,candidate corners detection is carried out in the region of interest of the BVS strut.Finally,four optimal corners are selected from the candidate corners by post-processing and BVS strut contour can obtained based on these four corners.In deep learning method,a BVS contour segmentation model is trained based on the Mask R-CNN model.Using the trained model,the instance segmentation of BVS strut can be obtained directly.BVS struts malapposition analysis can be automatically conducted based on the segmentation results.In order to evaluate the performance of the proposed algorithms,five pullbacks are used as test set.The segmentation results obtained by our proposed algorithm are compared with those golden standard BVS contour marked by the expert manually,and the segmentation results obtained by the existing dynamic programming method.Experimental results show that the corner-based method reaches an average dice of 0.89,which is improved by 0.08 compared to dynamic programming method.It costs about 240 seconds for one pullback and can't satisfy the real-time requirement in clinical.The Mask R-CNN method reaches an average recall rate of 0.994,precision of 0.996 and dice of 0.91.It takes about 20 seconds for one pullback.It is the state-of-the-art among the proposed methods and can satisfy the real-time clinical requirement.It concludes that our proposed method is effective and can segment struts accurately and robustly.Furthermore,automatic struts malapposition analysis in real time is feasible based on the strut contour in clinical practice.