2421 - Radiomics Analysis in SBRT for Pulmonary Oligometastases: The Potential of Peritumoral Features in Predicting Treatment Response
Presenter(s)
Y. Lu1, Y. Wang2, Y. Ding3, W. He4, J. Yang2,5, H. Yu2,5, S. Yan2,5, G. Ren6, and F. Zhao2,5; 1Department of Radiation Oncology, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China, 2Department of Radiation Oncology, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang, China, 3School of Basic Medical Sciences and Forensic Medicine, Hangzhou Medical College, Hangzhou, Zhejiang, China, 4Department of Radiology, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang, China, 5Cancer Center, Zhejiang University, Hangzhou, Zhejiang, China, 6Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong
Purpose/Objective(s): Predicting treatment outcomes is a crucial consideration in stereotactic body radiotherapy (SBRT) for pulmonary oligo-metastases. Developing a tool with more accurate predictive performance for pulmonary oligo-metastases treatment outcome is essential for optimizing radiotherapy strategies. This retrospective study aims to take advantage of radiomics to develop and validate radiomic models using a multi-center dataset.
Materials/Methods: This retrospective study included 223 tumors from 146 patients, divided into a training set (n=165) and an external validation set (n=58). Radiomic features were extracted from the gross tumor volume (GTV) and peritumoral regions (pGTV), representing the tumor microenvironment (TME), and integrated with clinical data. Multilayer Perceptron (MLP) models were constructed to assess how radiomic features from different regions and clinical data influence the models’ predictive performance. Meanwhile, SHapley Additive exPlanations (SHAP) were used to interpret feature contributions, providing insights into the integration of clinical and radiomic features in the final predictive models.
Results: The four MLP models, constructed using different combinations of data, demonstrated varying performances in predicting treatment outcomes. Model-G (with GTV features) achieved a validation AUC of 0.8024, with Model-P (with pGTV features) performing slightly lower at 0.7079. Meanwhile, Model-GP (with GTV and pGTV features) demonstrated the improved performance with a validation AUC of 0.8509, reflecting the added value of peritumoral features. The best performance was observed with Model-GPC (with GTV, pGTV, and clinical features). This model achieved a validation AUC of 0.9030, demonstrating its ability to integrate radiomic and clinical information for a robust classification capability.
Conclusion: Our study shows that the MLP model (Model-GPC) based on clinical-radiomic features accurately predicts SBRT treatment response in pulmonary oligometastases. By capturing tumor and peritumoral heterogeneity, the model identifies effectively respond patients and guides adjustments for non-respond patients. The inclusion of peritumoral features and SHAP analysis enhances prediction and provides insights for optimizing SBRT strategies. This study has been supported by the National Natural Science Foundation of China (82171890). Corresponding authors: F. Zhao, R. Ge, S. Yan.