Abstract
BACKGROUND: Intravascular optical coherence tomography (IVOCT) adoption has been limited by the complexity of image interpretation. The interpretation of histologic subtypes beyond lipid, calcium, and fibrous is challenging to human readers. To assist and standardize IVOCT image analysis, we demonstrate an artificial intelligence algorithm based on a histology data set that identifies lipid pools, fibrofatty, calcified lipid, and calcified fibrous in human coronary arteries for the first time. METHODS: Sixty-seven human coronary arteries were imaged with IVOCT within 24 hours after death and then underwent histologic examination. IVOCT images were coregistered and segmented into histologic subtypes: lipid pools, fibrofatty tissue, calcified lipid, and calcified fibrous tissue. Experiments regarding lipidic plaque included fibrofatty tissue, lipid pools, and calcified lipids. Experiments regarding calcium plaque included calcified fibrous and calcified lipid plaques. Optical coherence tomography images were lumen justified and cropped to a depth of 200 pixels (1 mm) to account for limited optical coherence tomography penetration depth. IVOCT segmentations from expert readers guided by histology were used to train segmentation neural networks. RESULTS: For each data set, in addition to testing each of these subtypes individually, we trained and tested the model on the combined grouping of subtypes. Combined lipid subtypes achieved validation and test Dice (Sørensen-Dice coefficient) of 0.63 and 0.40, respectively, whereas combined calcium subtypes achieved validation and test Dice of 0.66 and 0.62, respectively. CONCLUSIONS: This histology-validated artificial intelligence algorithm driven by histologic subtypes can identify plaque subtypes not evident to a human reader. The reported algorithm can provide a fast solution to IVOCT image interpretation.