Evaluation Of Deep Learning In Coronary Optical Coherence Tomography Lumen And Strut Auto-Analysis

Background:Optical coherence tomography(OCT),as an emerging imaging technology which has been widely used in coronary artery disease related studies,with its high resolution(approximately 10-20?m),provides sequenced coronary artery cross-section images and detailed microstructure morphology information inside coronary lumens^[1],hence shows its great potential in assisting clinical strategies.However,now the common methods used to analyze OCT images are mainly manual or semi-automatic ways,which are time consuming and tedious,blocks its range in clinical uses in some extent.In recent years,deep learning as a novel algorithm of artificial intelligence(AI),shed lights on the fields of medical imaging process,and already appeared in some studies of coronary OCT image auto-analysis.Compared to traditional machine learning,deep learning has advantages in simple pre-processing^[2,3]and accurate characteristic extraction etc.^[4].For now,deep learning in post percutaneous coronary intervention(PCI)follow-up OCT images,including lumen and strut etc.auto-analysis,have not been systematically studied yet.Objectives:This study aims at validating the accuracies of deep learning auto-analysis in post-PCI follow-up OCT images,and mainly focus on lumen contour,stent strut and strut coverage.Methods:61 patients who underwent post-PCI follow-up coronary angiography(CAG)and OCT scans of the previous stent segment from September 2012 to September 2014,were enrolled and divided into training group and test group randomly.A total of 7918 OCT cross-section images which contain struts were extracted from the total OCT data of the two groups.After the OCT images in training group were used for the deep learning network learning process and help form a sophisticated network,the images of the test group were then put into the network and produce the predicted output images,which were compared with the manual analysis standards to verify the accuracies of lumen segmentation,strut detection and strut coverage classification.The image features which affecting the prediction results were also analyzed.Results:The first part:The overall Dice coefficient of the deep learning network in the coronary OCT lumen segmentation reached 0.970,with an accuracy of 97.6%.Automatic segmentation and manual segmentation of lumen area is highly correlated(Pearson correlation coefficient of 0.992,p<0.001),and clinical measurements such as average lumen area,maximum lumen area,minimal lumen area and lumen volume are all keep strong correlation with manual segmentation results(Pearson correlation coefficient>0.95,p<0.001),with no statistical difference between the methods(p>0.05).In cross-sections with side branches or bifurcation,the precision of the automatic lumen segmentation was slightly worse than that without side branches or bifurcation(96.83%vs.97.52%,p<0.001),but the Dice coefficient was slightly better(0.971 vs.0.969,p<0.001).Furthermore,the features of 100 cross-sections with the worst Dice coefficient were analyzed,and it was found that the guidewire artifact greatly affected the segmentation results of the deep learning network.The second part:The deep learning network automatically detects coronary OCT struts with accuracy up to 92.8%and sensitivity up to 92.4%.The strut characteristics that affect the accuracy of automatic detection are related to two factors,include the struts located at the side branches or bifurcation(accuracy 67.1%)and the struts with non-obvious signal attenuation(accuracy 38.6%)(p<0.001).The sensitivity of automatic detection was better in covered struts(84.7%vs 93.1%,p<0.001),and decreased as the thickness of neointima increased(p<0.001).The features that affect the sensitivity of automatic detection are mainly the strut overlap detection(accounting for 88.5%of the total false positive struts),and the guidewire artifact and reverberation artifact behind the struts are also important factors affecting the sensitivity.Based on manual segmented lumen contour,the deep learning network measurements on average neointimal thickness are with poor correlation with manual analysis results(Pearson correlation coefficient of0.805,p<0.001).Furthermore,based on the distance between automatically detected struts and lumen contour,strut coverage classification was performed and found that the differences in the proportions of coverage between automatic and manual classification is significant(p<0.001).Conclusion:The deep learning network in automatic analysis of post-PCI follow-up coronary OCT images can reach a relatively high accuracy in lumen contour segmentation,and it is highly correlated with manual results in lumen area,lumen volume and other clinical measurements.However,the segmentation of some guidewire artifacts and bifurcation is not satisfactory.It can achieve high accuracy and sensitivity in strut detection,but it has low sensitivity in the detection of struts with non-characteristic features and cannot meet the requirements for clinical measurements such as the average neointima thickness and the strut coverage classification.Therefore,it needs to be further optimized and improved to reach the level for clinical application.