Development and validation of pericoronary adipose tissue attenuation analysis system

Coronary artery disease remains a leading cause of death worldwide, and early detection of coronary inflammation is essential for effective risk assessment. Pericoronary adipose tissue (PCAT) attenuation on coronary computed tomography angiography (CCTA) has emerged as a promising imaging biomarker of coronary inflammation. This study developed and evaluated an automated analysis pipeline for PCAT quantification based on coronary artery segmentation using five deep learning models: 3D U-Net, Attention U-Net, Dynamic U-Net, UNETR, and MedSAM2. Using 600 CCTA cases from an open-source database, the system automatically performed coronary artery segmentation, centerline extraction, and PCAT attenuation measurement with stepwise visual verification. Among the models, the Dynamic U-Net achieved the highest segmentation accuracy, with Dice coefficients of 0.91, 0.84, and 0.82 for the Right Coronary Artery (RCA), Left Anterior Descending (LAD), and Left Circumflex (LCX), respectively. Similar trends were observed for the PCAT regions (Dice: 0.89, 0.84, 0.82). The surface Dice was above 0.96 for arteries and 0.97 for PCAT at a 2-mm tolerance, and the Dynamic U-Net showed the most accurate boundaries, with HD95 values below 2 mm for all arteries. For PCAT attenuation, the Dynamic U-Net achieved mean errors within ± 1 HU across arteries (RCA: −0.06 HU; LAD: −0.06 HU; LCX: −0.24 HU), confirming high reproducibility. The RCA consistently showed the best accuracy, while the LCX exhibited larger errors due to its complex anatomy. The entire analysis was completed within approximately one minute per case, demonstrating the feasibility of the proposed automated pipeline for reliable, clinically applicable assessment of coronary inflammation.