Deep Learning-Based Prediction of Segmental Fractional Flow Reserve from Optical Coherence-Derived Anatomical and Compositional Features

초록

BACKGROUND Intracoronary imaging-derived physiologic allow per-vessel assessment of flow obstruction by capturing the cumulative effects of multifocal lesions. As hyperemic flow dynamics are influenced by the complex interplay of luminal geometry and vessel composition, we aimed to develop a deep learning algorithm to predict segmental changes in fractional flow reserve (AFFR). METHODS We designed a novel deep learning model based on a Gated Recurrent Unit (GRU) architecture to generate a virtual AFFR curve along the coronary artery. The model integrates structural and compositional features extracted from optical coherence tomography (OCT) imaging and was trained and validated using 157 OCT pullbacks from 86 patients to provide AFFROCT. Model performance was evaluated using Pearson's correlation coefficient and the area under the pullback curve (AUPC). RESULTS This model, which incorporates both lumen geometry, vascular calcification, and side branches to provide AFFROCT, outperformed models using partial features or conventional feature-based machine learning algorithms. It showed a strong correlation with invasively measured wire-based AFFR values (R=0.932, p<0.001; Figure A). At the lesion-level, the model accurately localized pressure drop regions, with an AUPC correlation coefficient of 0.937 (p<0.001; Figure B). CONCLUSION This GRU-based deep learning approach enables accurate, segmental prediction of physiologic significance from OCT imaging and facilitates precise localization of flow-limiting stenoses, supporting improved diagnostic decision-making in obstructive coronary artery disease.