Dosiomics Risk Model for Predicting Radiation Induced Temporal Lobe Injury and Guiding Individual Intensity-Modulated Radiation Therapy

Abstract

Purpose: We aimed to assess the value of dose distribution-based dosiomics and planning computed tomography-based radiomics to predict radiation-induced temporal lobe injury (TLI) and guide individualized intensity modulated radiation therapy.

Methods and materials: A total of 5599 nasopharyngeal carcinoma patients were enrolled, including 2503, 1072, 988, and 1036 patients in the training, validation, prospective test, and external test cohorts, respectively. The concordance index (C-index) was used to compare the performance of the radiomics and dosiomics models with that of the quantitative analyses of normal tissue effects in the clinic and Wen's models. The predicted TLI-free survival rates of redesigned simulated plans with the same dose-volume histogram but different dose distributions for same patient in a cohort of 30 randomly selected patients were compared by the Wilcoxon matched-pairs signed-rank test.

Results: The radiomics and dosiomics signatures were constructed based on 30 selected computed tomography features and 10 selected dose distribution features, respectively, which were important predictors of TLI-free survival (all P <.001). However, the radiomics signature had a low C-index. The dosiomics risk model combining the dosiomics signature, D_1cc, and age had favorable performance, with C-index values of 0.776, 0.811, 0.805, and 0.794 in the training, validation, prospective test, and external test cohorts, respectively, which were better than those of the quantitative analyses of normal tissue effects in the clinic model and Wen's model (all P <.001). The dosiomics risk model can further distinguish patients in a same risk category divided by other models (all P <.05). Conversely, the other models were unable to separate populations classified by the dosiomics risk model (all P > .05). Two simulated plans with the same dose-volume histogram but different dose distributions had different TLI-free survival rates predicted by dosiomics risk model (all P ≤ .002).

Conclusions: The dosiomics risk model was superior to traditional models in predicting the risk of TLI. This is a promising approach to precisely predict radiation-induced toxicities and guide individualized intensity modulated radiation therapy.