Predictive value of 18F-FDG PET/CT habitat radiomics combining stacking ensemble learning for prognosis in patients with hepatocellular carcinoma
10.3760/cma.j.cn321828-20240613-00205
- VernacularTitle:18F－FDG PET/CT生境影像组学结合堆叠集成学习对肝细胞肝癌患者生存预后的预测价值
- Author:
Chunxiao SUI
1
;
Kun CHEN
;
Qian SU
;
Rui TAN
;
Wengui XU
;
Xiaofeng LI
Author Information
1. 天津医科大学肿瘤医院分子影像及核医学诊疗科、国家恶性肿瘤临床医学研究中心、天津市肿瘤防治重点实验室、天津市恶性肿瘤临床研究中心，天津 300060
- Publication Type:Journal Article
- Keywords:
Carcinoma, hepatocellular;
Radiomics;
Positron-emission tomography;
Tomography, X-ray computed;
Fluorodeoxyglucose F18
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2025;45(5):263-268
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the prognostic value of 18F-FDG PET/CT-based habitat radiomics combined with stacking ensemble learning model in overall survival (OS) of patients with hepatocellular carcinoma (HCC). Methods:A total of 136 HCC patients (114 males, 22 females, age (55.3±10.4) years) who underwent 18F-FDG PET/CT before treatment between January 2018 and January 2023 were retrospectively analyzed. Eighty-five cases from Tianjin First Central Hospital and 51 cases from Tianjin Medical University Cancer Institute and Hospital were used as a training cohort and an external validation cohort, respectively. The tumor volume of interest (VOI) was delineated on PET and CT images, and a total of 4 habitats were segmented by using the Otsu algorithm, including PET high ∩ CT low, PET low ∩ CT low, PET high ∩ CT high, and PET low ∩ CT high. After the feature selection, a total of 36 stacking ensemble learning models were established, and the optimal model was selected based on the calculated concordance index (C-index). Moreover, a combined model was developed by integrating the optimal model with clinical information. The predictive efficacy of those models was assessed by time-dependent ROC curves. Results:The model based on PET high ∩ CT high habitat radiomics features with multilayer perceptron (MLP) classifier had the highest C-index (0.770) in the external validation cohort, and it was regarded as the optimal radiomics model. The combined model incorporating this model with clinical information achieved an improved C-index of 0.815 in the external validation cohort. The combined model outperformed the other models for OS prediction, with a time-dependent AUC of 0.919, 0.900, and 0.862 in predicting the 1-year, 2-year, and 3-year OS, respectively. Conclusions:18F-FDG PET/CT-based habitat analysis outperforms traditional radiomics in OS prediction for HCC patients. By integrating the optimal habitat model with the clinical model, the combined model is able to improve the predictive efficacy.
