Discovering pericoronary adipose tissue differences between normal and atherosclerotic coronary segments with radiomics analysis

Abstract

Abstract Introduction Pericoronary adipose tissue (PCAT) extracted from coronary CT angiography (CCTA) is a novel marker for coronary inflammation. Establishing correlation between PCAT changes and the vascular lesion itself is essential for using PCAT to monitor clinical and subclinical disease status. Past studies focused on patient-level and vessel-level PCAT differences. However, lesion-level PCAT differences beyond mean attenuation values are not well understood. Purpose To investigate the differences between PCAT adjacent to atherosclerotic plaques and PCAT adjacent to normal coronary segments using radiomics analysis. Material and Methods This study enrolled 100 consecutive patients who underwent CCTA. A total of 446 plaques were successfully annotated in 89 patients. The PCAT of these 446 plaques and also their normal segment counterparts were extracted. Following prior literature, adipose tissue was defined as voxels with attenuation ranging -190 to -30 HU. PCAT was extracted within 3 mm distance from the vessel wall, along the middle 10 mm of each plaque. The PCAT of normal coronary artery segments were extracted both proximal and distal with distance D mm to the plaque with 10 mm length. We used 3 different distances to define the normal segments (D={10, 20, 30}). PyRadiomics was used to generate radiomics features and the random forest classifier from Scikit-Learn was used to distinguish PCAT extracted adjacent to atherosclerotic plaques and PCAT adjacent to normal coronary segments. To minimize bias due to vessel diameter or shape and focus on the texture of PCAT, we included only the First Order, GLCM, GLDM, GLRLM, GLSZM and NGTDM features (total 93 features) from PyRadiomics. We used 5-fold cross validation splitted at patient-level to develop the classifier. Result The radiomics model achieved acceptable discrimination for D=30 mm (mean AUROC = 0.70), borderline discrimination for D=20 mm (mean AUROC=0.68) and D=10 mm (mean AUROC = 0.60). These results suggest a trend toward greater differences between PCAT further away from the lesion compared with PCAT adjacent to the lesion. The feature "Gray Level Size Zone Matrix (GLSZM): Low Gray Level Zone Emphasis" had the highest importance in 4 out of 5 folds in our D = 30 mm experiment. The mean value of this feature is significantly different for the PCAT adjacent to the lesions compared to PCAT extracted from the normal segments (0.152 vs 0.166, p=0.001). Conclusion We investigated the difference between PCAT adjacent to atherosclerotic plaques and PCAT adjacent to normal coronary segments using radiomics analysis. We found greater discrimination when the normal segments were extracted further away from the lesion, suggesting PCAT changes due to atherosclerosis may exist beyond the length of the plaque. Moreover, "Gray Level Size Zone Matrix (GLSZM): Low Gray Level Zone Emphasis" could serve as a novel radiomic biomarker to correlate PCAT changes to atherosclerosis.