A consistency study on lipid plaque quantification by optical coherence tomography:TransDeepLab model versus manual analysis

YE Hao-yi; HAN Ze-jun; CUI Si-mei; WU Xiao-shuang; GAO Lei

doi:10.3969/j.issn.1672-5301.2026.02.007

Abstract：Objective To investigate the agreement between TransDeepLab model and manual measurements for detecting lipid plaques in optical coherence tomography(OCT)images.Methods We prospectively enrolled 143 patients with coronary heart disease who underwent elective percutaneous coronary intervention(PCI)and OCT examination at the Sixth Medical Center of Chinese PLA General Hospital from January to December 2024,forming the training set.An additional 67 patients(70 lesions)constituted the test set,while 27 patients(30 lesions)from the First Medical Center of Chinese PLA General Hospital during the same period were included as the external validation set.Two physicians conducted manual analysis using OCT offline image analysis software to measure the maximum lipid arc,mean lipid arc,lipid length,and minimum fibrous cap thickness,with final adjudication performed by a senior physician.The artificial intelligence(AI)model was developed based on the TransDeepLab architecture,which integrates DeepLab and the Swin Transformer framework.Results Based on OCT-derived lipid plaque data from 237 patients across two centers of Chinese PLA General Hospital,using manual analyses by senior physicians as the reference standard,this study systematically evaluated the agreement between the TransDeepLab model and manual interpretation.In the internal validation,the AI model and manual analysis showed good agreement in maximum lipid arc(r=0.842,ICC=0.811),mean lipid arc(r=0.721,ICC=0.715),and lipid length(r=0.913,ICC=0.874),and moderate agreement in fibrous cap thickness(r=0.652,ICC=0.650).T he Bland-Altman plot indicated that the differences for most parameters fell within the acceptable limits.In the external validation,the agreement for all four-measurement metrics declined.Nevertheless,this AI model demonstrated good agreement with manual analysis in the quantification of OCT-derived lipid plaque length and lipid arc,whereas agreement for fibrous cap thickness was suboptimal.Conclusion The TransDeepLab model showed good and reproducible measurement agreement in the quantification of OCT-derived lipid plaques.Its assessments of lipid arc and length were in close concordance with the manual reference standard,while agreement for fibrous cap thickness was moderate yet statistically significant.This model holds promise for clinical application in the assessment of lipid plaques.