Machine-learning based MRI radiomics models for early detection of radiation-induced brain injury in nasopharyngeal carcinoma

Abstract

Background: Early radiation-induced temporal lobe injury (RTLI) diagnosis in nasopharyngeal carcinoma (NPC) is clinically challenging, and prediction models of RTLI are lacking. Hence, we aimed to develop radiomic models for early detection of RTLI. Methods: We retrospectively included a total of 242 NPC patients who underwent regular follow-up magnetic resonance imaging (MRI) examinations, including contrast-enhanced T1-weighted and T2-weighted imaging. For each MRI sequence, four non-texture and 10,320 texture features were extracted from medial temporal lobe, gray matter, and white matter, respectively. The relief and 0.632 + bootstrap algorithms were applied for initial and subsequent feature selection, respectively. Random forest method was used to construct the prediction model. Three models, 1, 2 and 3, were developed for predicting the results of the last three follow-up MRI scans at different times before RTLI onset, respectively. The area under the curve (AUC) was used to evaluate the performance of models. Results: Of the 242 patients, 171 (70.7%) were men, and the mean age of all the patients was 48.5 ± 10.4 years. The median follow-up and latency from radiotherapy until RTLI were 46 and 41 months, respectively. In the testing cohort, models 1, 2, and 3, with 20 texture features derived from the medial temporal lobe, yielded mean AUCs of 0.830 (95% CI: 0.823-0.837), 0.773 (95% CI: 0.763-0.782), and 0.716 (95% CI: 0.699-0.733), respectively. Conclusion: The three developed radiomic models can dynamically predict RTLI in advance, enabling early detection and allowing clinicians to take preventive measures to stop or slow down the deterioration of RTLI.