290 - Intratumoral-Peritumoral CT Radiomics with Machine Learning for Predicting Local Tumor Response after Radiotherapy in Non-Small Cell Lung Cancer
Presenter(s)
C. Tan1,2, Z. Yang1, Z. G. Guo1, M. Fan3, and Z. Chen2; 1The Second Affiliated Hospital of Chongqing Medical University, Chongqing, Chongqing, China, 2West China Longquan Hospital, Sichuan University; The First People’s Hospital of Longquanyi District, Chengdu, SiChuan, China, 3Sichuan Cancer Hospital and Research Institute, University of Electronic Science and Technology of China, Chengdu, China
Purpose/Objective(s): The hypothesis is that radiomics features derived from combined intratumoral and peritumoral CT regions provide superior prediction of local tumor response after radiotherapy for non–small cell lung cancer compared with single-region or image-level fusion strategies. The primary objective was to compare the predictive performance of radiomics-based machine learning models across different tumor region definitions and to develop a clinically applicable nomogram integrating radiomics and clinical factors.
Materials/Methods: This retrospective study included 500 patients treated with definitive thoracic radiotherapy, comprising 350 patients as the training cohort and 150 patients from an independent center for external validation. Pretreatment CT images were used to delineate the gross tumor volume and a 3-mm peritumoral region. Radiomics features were extracted using four strategies: intratumoral, peritumoral, intratumoral–peritumoral feature concatenation, and image-level fusion. Machine learning models were constructed using XGBoost, LightGBM, logistic regression, and support vector machine algorithms. Tumor response was assessed according to RECIST, with objective response rate (ORR; CR+PR) evaluated by imaging at 3, 6, and 12 months after radiotherapy. Model performance was assessed using the area under the receiver operating characteristic curve, calibration, and decision curve analyses. The optimal radiomics signature was combined with independent clinical predictors to construct a nomogram. Radiotherapy was predominantly delivered using conventional fractionation, with SBRT included, and prescription doses ranged from 45 to 66 Gy.
Results: Among all radiomics strategies and classifiers, the INTRAPERI-based XGBoost model achieved the best radiomics-only performance, with an AUC of 0.880 (95%CI 0.844–0.913) in the training cohort and 0.740 (95%CI 0.654–0.816) in external validation. A clinical model alone demonstrated moderate discrimination (training AUC 0.812; validation AUC 0.638).After integrating the optimal radiomics signature with clinical predictors, the nomogram achieved AUCs of 0.924 (95%CI 0.897–0.948) in the training cohort and 0.753 (95%CI 0.674–0.831) in the external validation cohort, significantly outperforming the clinical model (P<0.001 in training; P=0.012 in validation). Overall, 261 of 500 patients (52.2%) achieved an objective response.
Conclusion: A machine learning model based on combined intratumoral and peritumoral CT radiomics features enables robust prediction of local tumor response after radiotherapy for non–small cell lung cancer. Integration of radiomics with clinical factors into a nomogram further improves discrimination and supports its potential role in individualized radiotherapy decision-making.