Evaluation of a deep learning-driven centerline extraction algorithm for optimizing the diagnosis of the"gray zone"in noninvasive coronary fractional flow reserve
10.3969/j.issn.1004-8812.2025.06.003
- VernacularTitle:深度学习驱动的中心线提取算法对无创冠状动脉血流储备分数"灰区"诊断的优化价值研究
- Author:
Zi-qiang GUO
1
;
Xi WANG
;
Zi-nuan LIU
;
Yi-pu DING
;
Ran XIN
;
Dong-kai SHAN
;
Jun GUO
;
Yun-dai CHEN
;
Jun-jie YANG
Author Information
1. 中国人民解放军总医院第六医学中心心血管病医学部,北京 100048;中国人民解放军医学院,北京 100853
- Publication Type:Journal Article
- Keywords:
Deep learning;
Coronary artery disease;
Fractional flow reserve;
Centerline extraction;
Fully convolutional network
- From:
Chinese Journal of Interventional Cardiology
2025;33(6):312-318
- CountryChina
- Language:Chinese
-
Abstract:
Objective To evaluate the diagnostic performance of the minimum-cost-path-based CT angiography-derived fractional flow reserve(MCP-FFR)and the deep learning-driven CT angiography-derived fractional flow reserve(DeepCL-FFR),and to particularly explore the potential value of the DeepCL algorithm in improving diagnostic accuracy within the"gray zone."Methods A retrospective analysis was conducted on 151 coronary vessels from 109 patients with coronary artery disease,who were hospitalized at the General Hospital of the People's Liberation Army between January 2020 and June 2021.Pearson correlation and Bland-Altman plots were employed to assess the correlation and agreement of the two CT-FFR methods with invasive FFR.A CT-FFR range of 0.70-0.80 was defined as the diagnostic"gray zone."The accuracy,sensitivity,specificity,positive predictive value,and negative predictive value for detecting hemodynamic abnormalities were calculated and analyzed.The DeLong test was used to compare the areas under the receiver operating characteristic curves(AUC)between the two CT-FFR calculation methods.Results Both CT-FFR methods exhibited a positive correlation with invasive FFR(MCP-FFR:r=0.75,P＜0.001;DeepCL-FFR:r=0.86,P＜0.001)and showed good agreement(MCP-FFR:mean difference=0.010,P=0.351;DeepCL-FFR:mean difference=-0.003,P=0.772).Both DeepCL-FFR(AUC 0.97,95％CI 0.94-0.99)and MCP-FFR(AUC 0.92,95％CI 0.88-0.97)demonstrated favorable diagnostic performance for detecting hemodynamic abnormalities(P=0.122).In the"gray zone"for hemodynamic abnormality,the diagnostic accuracy of MCP-FFR was 68.8％,whereas DeepCL-FFR increased it to 89.7％.DeepCL-FFR also exhibited superior diagnostic performance(AUC 0.89,95％CI 0.73-0.99)within the"gray zone,"which was significantly higher than that of MCP-FFR(AUC 0.71,95％CI 0.54-0.87)(P＜0.001).Conclusions The deep learning-driven coronary centerline extraction algorithm,DeepCL,demonstrates superior diagnostic performance in CT-FFR for detecting hemodynamic abnormalities,particularly by significantly improving diagnostic accuracy in the"gray zone."
