Integrative model combining deep learning,clinical and radiomic features enhances EGFR mutation prediction in non-small cell lung cancer
10.16016/j.2097-0927.202509040
- VernacularTitle:融合深度学习、临床与影像组学特征的综合模型提升非小细胞肺癌EGFR突变预测效能
- Author:
Taotao YANG
1
;
Wei CHEN
;
Cancan CHEN
;
Wanying YAN
;
Dawei WANG
;
Kunlin XIONG
;
Zhiyuan SUN
;
Xianqi WANG
Author Information
1. 陆军军医大学(第三军医大学)第一附属医院放射科
- Keywords:
non-small cell lung cancer;
epidermal growth factor receptor mutation;
deep learning;
radiomics
- From:
Journal of Army Medical University
2025;47(23):2991-3001
- CountryChina
- Language:Chinese
-
Abstract:
Objective To evaluate the predictive value of deep learning features from chest CT images combined with clinical and radiomics features for epidermal growth factor receptor(EGFR)mutations in non-small cell lung cancer(NSCLC).Methods This case-control study retrospectively analyzed clinical and imaging data of 1 070 NSCLC patients from radiology departments at three hospitals(January 2013 to October 2023).Patients were divided into:a training set(n=502)and internal validation set(n=217)via 7∶3 randomization of 719 cases from the First Affiliated Hospital of Army Medical University;external validation set 1(n=173)from General Hospital of Eastern Theater Command;external validation set 2(n=178)from Daping Hospital of Army Medical University.Deep learning features were extracted using a 2.5D convolutional neural network(CNN)with ResNet101 backbone,radiomics features were derived from CT images,and clinical risk factors were identified to construct models.An integrated model combined deep learning,clinical,and radiomics features.All four models were developed using random forest(RF)classifiers.Calibration curves assessed goodness-of-fit,and decision curve analysis(DCA)evaluated clinical utility.Results The deep learning model achieved AUCs of 0.833 7(95%CI:0.770 6～0.884 7),0.815 1(0.741 6～0.882 8),and 0.810 1(0.745 2～0.873 6)in the internal and two external validation sets,respectively.Clinical models yielded AUCs of 0.731 0(0.660 2～0.802 1),0.746 0(0.666 4～0.824 9),and 0.813 4(0.743 1～0.883 6);radiomics models showed AUCs of 0.762 4(0.692 4～0.825 1),0.745 4(0.671 1～0.814 3),and 0.724 7(0.639 7～0.801 6).The integrated model demonstrated optimal performance with AUCs of 0.905 5(0.857 0～0.945 4),0.832 7(0.763 3～0.896 4),and 0.889 0(0.834 4～0.934 3).DCA indicated significant net benefit for EGFR prediction at threshold probabilities of 0.15～0.85 using the integrated model.Conclusion Deep learning features from CT images effectively predict EGFR mutation status in NSCLC.The integrated model combining deep learning,clinical,and radiomics features further enhances predictive performance.
