ESTRO 2024 - Abstract Book

S5058

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

was further employed to choose the most relevant and low-redundancy features. Three single models and a hybrid model were developed from radiomics, dosiomics, and clinical features. The models were evaluated by 10-fold cross validation. The single models were trained by LASSO, Random Forest, and Extreme Gradient Boosting (XGBoost). Both the individual and combined feature categories were analyzed for RIHT prediction performance. The hybrid model was constructed by a combination of decision tree and linear regression where the extreme-high (mean dose > 6300cGy) and low (volume receiving more than 4000cGy < 80%) dose[2,3] patients were directly assigned as positive and negative prediction respectively, and the rest of the patients were modeled by LASSO. The mean testing area under curve (AUC) averaged across the 10 cross-validation was calculated as an unbiased estimation of the modeling performance. T-test was applied to compare the prediction ability of the hybrid model and other models.

Results:

There are 115 patients who developed RIHT(26.9%) by the last follow-up time, with a median follow-up duration of 4.36 years. 369 (86.2%) patients received mean radiation dose between 4000cGy~6300cGy in Thyroid. The RIHT prediction model was constructed with 20 radiomics features, 20 dosiomics features, and seven clinical features as candidate features after the feature selection process. The mean testing AUC of multi-modalities models developed by LASSO, Random Forest and XGBoost were 0.73, 0.66. 0.72, respectively. The comparison of individual modality was based on the LASSO model, the mean testing AUC of radiomics, dosiomics, and clinical features were 0.68, 0.66, 0.61, respectively. The hybrid model that contains two radiomics, two dosiomics, and one clinical features achieved the highest performance with mean testing AUC of 0.83. The hybrid model had significantly better prediction ability than the LASSO model (p = 0.02) and the XGBoost model (p = 0.005), respectively.

Conclusion:

The hybrid model used both radiomics, dosiomics and clinical features showed best performance in predicting RIHT after IMRT for NPC patients. The 10-fold average testing AUC is 0.83, which is significantly better than the individual modality models.

Keywords: Hypothyroidism, Radiomics, Dosiomics

References:

1. Teng, X., Zhang, J., Zwanenburg, A., Sun, J., Huang, Y., Lam, S., Zhang, Y., Li, B., Zhou, T., Xiao, H., Liu, C., Li, W., Han, X., Ma, Z., Li, T., & Cai, J. (2022). Building reliable radiomic models using image perturbation. Scientific reports, 12(1), 10035. https://doi.org/10.1038/s41598-022-14178-x 2.Sommat, K., Ong, W. S., Hussain, A., Soong, Y. L., Tan, T., Wee, J., & Fong, K. W. (2017). Thyroid V40 Predicts Primary Hypothyroidism After Intensity Modulated Radiation Therapy for Nasopharyngeal Carcinoma. International Journal of Radiation Oncology*Biology*Physics, 98(3), 574–580. https://doi.org/10.1016/j.ijrobp.2017.03.007 3.Xu, Y., Peng, H., Su, G., Cheng, Y., Guo, Q., Guo, L., Peng, X.-E., & Ke, J. (2023). Thyroid V40 is a good predictor for subclinical hypothyroidism in patients with nasopharyngeal carcinoma after intensity modulated radiation therapy: A randomized clinical trial. Radiation Oncology, 18(1), 141. https://doi.org/10.1186/s13014-023-02329-x