Deep Learning Facilitates Automation of Wall Layer Quantification in Heart Transplant Coronary OCT

Purpose Measurements of coronary intima and media thickness in 3D OCT images following heart transplantation (HTx) is used for assessing early cardiac allograft vasculopathy (CAV) status and progression. OCT layers become invisible with excessive CAV or focal atherosclerosis. We report a novel deep-learning (DL) method for automated identification of layered coronary wall regions in OCT and assess whether automation can replace tedious manual analysis. Methods Our DL approach is trained using expert-defined layered/non-layered OCT regions of coronary wall. OCT image data in all frames of the OCT pullback are processed by a fully convolutional neural network (CNN), in which strided CNN layers and sigmoid activation and thresholding of the output determine angular regions of layered/non-layered appearance for each OCT image frame. Using a leave-20%-out approach, DL CNN was trained and its performance evaluated in OCT pullbacks acquired at 1 and 12 months after HTx in 116 patients (232 OCT pullbacks, 85580 OCT frames). Results DL identification of layered regions reached 86.3% correctness when compared with a manually-determined independent standard. In regions with layered wall appearance, intimal and medial surfaces were segmented using our previously reported LOGISMOS segmentation method and used as a basis for morphologic quantifications. Lumen areas, intimal, medial, and intimal-medial thicknesses, and their changes between baseline OCT imaging at 1 month and follow-up at 12 months obtained using manual and DL analyses were compared. Table shows that no statistically significant differences were found between manual and automated DL analysis. Conclusion Automated DL determination of layered regions of coronary wall required 10s of computer time compared to 1 hour of manual expert effort while achieving statistically identical results. Automated DL-based analysis of coronary wall can replace manual analysis in future quantitative studies of CAV progression post-HTx. Measurements of coronary intima and media thickness in 3D OCT images following heart transplantation (HTx) is used for assessing early cardiac allograft vasculopathy (CAV) status and progression. OCT layers become invisible with excessive CAV or focal atherosclerosis. We report a novel deep-learning (DL) method for automated identification of layered coronary wall regions in OCT and assess whether automation can replace tedious manual analysis. Our DL approach is trained using expert-defined layered/non-layered OCT regions of coronary wall. OCT image data in all frames of the OCT pullback are processed by a fully convolutional neural network (CNN), in which strided CNN layers and sigmoid activation and thresholding of the output determine angular regions of layered/non-layered appearance for each OCT image frame. Using a leave-20%-out approach, DL CNN was trained and its performance evaluated in OCT pullbacks acquired at 1 and 12 months after HTx in 116 patients (232 OCT pullbacks, 85580 OCT frames). DL identification of layered regions reached 86.3% correctness when compared with a manually-determined independent standard. In regions with layered wall appearance, intimal and medial surfaces were segmented using our previously reported LOGISMOS segmentation method and used as a basis for morphologic quantifications. Lumen areas, intimal, medial, and intimal-medial thicknesses, and their changes between baseline OCT imaging at 1 month and follow-up at 12 months obtained using manual and DL analyses were compared. Table shows that no statistically significant differences were found between manual and automated DL analysis. Automated DL determination of layered regions of coronary wall required 10s of computer time compared to 1 hour of manual expert effort while achieving statistically identical results. Automated DL-based analysis of coronary wall can replace manual analysis in future quantitative studies of CAV progression post-HTx.