부산대학교 도서관

To move toward personalized therapy, the overall survival of poor prognosis cancers, such as head and neck (H&N) cancer, needs to be modeled with robust algorithms. The recently emerged radiomics approach uses quantitative features for predicting therapeutic responses. For cancer patients, radiotherapy provides 3D dose distribution. Encoding the embedded spatial information of dose distributions (called Dosiomics) can produce valuable features helpful for modeling prognostications. This study investigates the role of Radiomic and Dosiomic features in predicting the overall survival of H&N cancer patients after radiotherapy. We included 240 H&N cancer patients from five different institutions whose data was collected on the TCIA database. In total, 228 radiomic and Dosiomic features were retrieved from the GTVs including 79 first-order features and 136 three-dimensional texture features. Due to the significant role of clinical features in the prognosis of H&N cancer patients, we used 13 clinical features (gender, age, histology, smoking status, HPV status, T-staging, N-staging, TNM staging, primary tumor site, treatment modalities) alongside the radiomic and Dosiomic features. The random survival forest (RSF) machine learning (ML) model was employed in conjunction with five feature selection (FS) methods including C-Index, Variable hunting (VH), Variable hunting Variable Importance (VH. VIMP), Minimal Depth (MD), and Mutual Information (MI). The hyperparameters were optimized by grid search. Then concordance indices (C-Indices) were reported as the model performance quantification measured by 5-fold cross-validation for OS prediction. The best performance was achieved by combining RSF with MD (C-index=0.71) for the Dosiomics strategy. Moreover, the RSF, in a combination of VH. VIMP feature selection method achieved a C-index of 0.70 for the Radiomics strategy. This study showed that RSF algorithms trained by Dosiomic and radiomic features could significantly predict the survival and prognostication of H&N cancer patients. At the same time, the Dosiomics strategy have a stronger role in predicting overall survival.