DeepNeo: Deep Learning for neointimal tissue characterization using optical coherence tomography

AimsThis study aimed to develop a deep-learning algorithm to enable a fully-automated analysis and interpretation of optical coherence tomography (OCT) pull-backs from patients after percutaneous coronary intervention (PCI). Methods and resultsIn 1148 frames from 92 OCTs, neointima was manually classified as homogeneous, heterogenous, neoatherosclerosis, or not analyzable at quadrant level by an experienced expert. Additionally, stent and lumen contours were annotated in 90 frames to enable segmentation of lumen, stent struts and neointima. Annotated frames were used to train "DeepNeo", a deep learning tool for prediction of neointimal tissue characteristics. Performance of DeepNeo was additionally evaluated in an animal model of neoatherosclerosis, using co-registered histopathology images as the gold-standard. DeepNeo demonstrated excellent classification performance of neointimal tissue with an overall accuracy of 75%, comparable to manual classification accuracy of two clinical experts (75%, 71%). The accurate performance of DeepNeo was confirmed in an animal model of neoatherosclerosis, where an overall accuracy of 87% was achieved. Segmentation of lumen, stent struts and neointima in human pullbacks yielded very good performance with mean Dice overlap scores of 0.99, 0.66 and 0.86. ConclusionDeepNeo is the first deep learning algorithm allowing fully automated segmentation and classification of neointimal tissue, with a performance comparable to human experts. DeepNeo might ultimately help assess vascular healing after percutaneous coronary intervention in a standardized, reliable and time-efficient manner, support therapeutic decisions and improve the detection of patients at risk of future cardiac events.