Interpretable machine learning model based on 18F-FDG PET/CT radiomics for prognostic evaluation of diffuse large B-cell lymphoma
10.3760/cma.j.cn321828-20240131-00049
- VernacularTitle:18F-FDG PET/CT影像组学的可解释性机器学习模型对弥漫性大B细胞淋巴瘤的预后评估
- Author:
Caozhe CUI
1
;
Ning MA
1
;
Qiannan WANG
1
;
Xiaomeng LI
1
;
Yayuan LI
1
;
Zhifang WU
1
Author Information
1. 山西医科大学第一医院核医学科、分子影像精准诊疗省部共建协同创新中心，太原 030001
- Publication Type:Journal Article
- Keywords:
Lymphoma, large B-cell, diffuse;
Radiomics;
Positron-emission tomography;
Tomography, X-ray computed;
Fluorodeoxyglucose F18
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2025;45(1):1-6
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop radiomics score (RS) based on 18F-FDG PET/CT, and construct the machine learning model combining clinical and other relevant factors for personalized prediction of 2-year event-free survival (2-EFS) in patients with diffuse large B-cell lymphoma (DLBCL), and to perform interpretability analysis of the model. Methods:A total of 91 patients (49 males, 42 females; age (57.8±12.8) years) with pathologically confirmed DLBCL from December 2017 to December 2020 at the First Hospital of Shanxi Medical University were retrospectively analyzed. According to the ratio of 7∶3, patients were randomly divided into training set ( n=63) and test set ( n=28), and divided into non-progression group and progression group according to the follow-up results. The whole-body PET semi-quantitative parameters were calculated from the PET/CT images before treatment, and 328 radiomics features were extracted from the largest target lesions of patients. The least absolute shrinkage and selection operator (LASSO) was used to develop the RS. Clinical and PET characteristic difference analysis was performed through χ2 test and Mann-Whitney U test. Extreme gradient boosting (XGBoost) models were constructed based on clinical, PET radiomics features and RS, and the prediction efficiency of each model was evaluated by ROC AUC. The model interpretability was analyzed by using shapely additive explanation (SHAP). Results:Of all patients, 32 had disease progression and 59 did not. There were no significant differences in baseline characteristics between the training set and the test set ( χ2 values: 0.06-1.84, U values: 665.00-763.00, all P>0.05). The comparison between the progression group and non-progression group in the training set showed statistical differences in the international prognostic index (IPI) score ( χ2=4.87, P=0.027), myelocytomatosis viral oncogene (MYC) protein expression ( χ2=4.29, P=0.038), and metabolic tumor volume (MTV; U=307.00, P=0.038). Seven radiomics features were screened by LASSO. Among XGBoost models with different feature combinations, IPI score, MYC protein expression, MTV combined with RS had the highest predictive efficiency (training set: AUC=0.73; test set: AUC=0.70). Through SHAP analysis, RS was the most predictive feature in the optimal model. Conclusion:The machine learning integrated model of IPI score, MYC protein expression and MTV combined with RS can effectively predict the prognosis of DLBCL patients, and baseline 18F-FDG PET/CT radiomics can be used as a potential means to evaluate the prognosis of DLBCL patients.
