IMG-05. Novel AI model for automated prediction of radiation necrosis versus tumor recurrence in Glioblastoma patients using brain MRI

Abstract INTRODUCTION Distinguishing radiation necrosis (RN) from tumor progression (TP) after chemoradiation for Glioblastoma is essential for timely therapy adaptation. Misclassification can delay escalation to second-line systemic agents or expose patients to unnecessary reoperation and prolonged steroid treatment. Conventional MRI lacks specificity, and biopsy carries morbidity and extends time-to-treatment. We developed a rapid, AI-based classifier that analyzes routine post-contrast T1-weighted MRI and provides an immediate probability of RN versus TP, aiming to shorten decision-making pathways. METHODS We retrospectively screened 597 Glioblastoma patients treated with the Stupp protocol (2010–2024). 288 who underwent re-craniotomy for suspected recurrence and had complete pre-operative MRI constituted the dataset. A Vision Transformer (ViT) was trained on 230 patients and validated on 58, utilizing axial slices centered on the enhancing lesion. Histopathology served as the reference standard. Accuracy, sensitivity, specificity, precision, and the area under the receiver-operating-characteristic curve (AUC) were calculated, and significance was assessed using McNemar’s test. Inference time per study was recorded on a standard GPU. RESULTS The mean age was 59 ± 7 years; 46% were female. The lesion distribution was frontal 38%, temporal 17%, and other 45%. The ViT achieved 86% accuracy (95% CI 80-91%), sensitivity 85%, specificity 88%, and AUC 0.91, outperforming majority-class prediction (p=0.002). The median inference time was 2 s. Grad-CAM heat maps highlighted lesion margins and perilesional vessels, mirroring radiological cues for RN. CONCLUSIONS A ViT classifier distinguished RN from TP using single-sequence MRI with an AUC of 0.91 and near-instantaneous inference, which supports earlier treatment decisions and may spare patients from invasive biopsies. A prospective, multi-center evaluation integrating multi-parametric imaging will assess its impact on workflow speed, healthcare costs, and outcomes.