A MRI-based radiomics combined prediction model for radiation-induced temporal lobe injury in nasopharyngeal carcinoma patients

Abstract

PurposeRadiation-induced temporal lobe injury (RTLI) is a common radiation-induced injury in patients with nasopharyngeal carcinoma (NPC). Our aim is to evaluate the potential factors associated with RTLI and to develop a predictive model to assess RTLI. MethodA total of 393 NPC patients who had received full course radiotherapy at two institutions were included. Patients from institution I were divided into a training cohort and a testing cohort according to the time of first-line treatment. Patients from institution II were used for external validation. The pre-treatment T1-weighted, T2-weighted and contrast-enhanced T1-weighted (T1c-weighted) magnetic resonance imaging (MRI) images before treatment were retrieved for radiomics feature extraction. The left and right temporal regions were separately defined and delineated as regions of interest (ROIs) by the MIM segmentation system. The relative contributions of different MRI sequences to the radiomics (Rad) score evaluation were compared. The least absolute shrinkage and selection operator (LASSO) regression algorithm was used to calculate Rad scores and develop comprehensive models. ResultsThe radiomics model consisting of all MRI features (combined T1-weighted, T2-weighted and T1c-weighted) was most associated with RTLI. The AUCs were 0.77 (95% CI: 0.69–0.85) in the testing cohort and 0.74 (95% CI: 0.70–0.79) in the external validation cohort. The AUC for the clinical-only model (including clinical stage and D0.5 cc) was 0.52 (95% CI: 0.31–0.74), whereas models including clinical factors (clinical stage), dosimetric factors (D0.5 cc) and Rad scores from all MRI sequences were more optimal. The AUC was 0.76 (95% CI: 0.67–0.84). ConclusionWe generated a feasible RTLI prediction model with radiomics, clinical and dosimetric factors.