Integrating chemotherapy, radiotherapy, and O6-methylguanine-DNA methyltransferase (MGMT) status with deep-learning cellular tumor volumetry sharpens prediction of glioblastoma recurrence on postoperative MRI.
Authors
Affiliations (3)
Affiliations (3)
- Department of Physics, Engineering Physics and Optics, Laval University, Quebec City, QC, Canada.
- Department of Medical Imaging, Faculty of Medicine, University of Toronto, Toronto, ON, Canada.
- Department of Radiology and Nuclear Medicine, Laval University, Quebec City, QC, Canada.
Abstract
Distinguishing glioblastoma recurrence from posttreatment effects on magnetic resonance imaging (MRI) remains a major diagnostic challenge. While recent models consider O6-methylguanine-DNA methyltransferase (MGMT) status, most machine learning models addressing this problem assume uniform chemotherapy and radiotherapy exposure, overlooking real-world treatment variability. This study aimed to determine whether integrating chemotherapy, radiotherapy, and MGMT status together with automated cellular tumor volume (VolCT) improves predictive accuracy in posttreatment glioblastoma assessment. We retrospectively analyzed 135 postsurgical MRI examinations from 97 glioblastoma patients treated between January 2008 and December 2022. The dataset included 105 confirmed recurrences and 30 cases of treatment-related change. A total of 8465 radiomic features were extracted from 5 MRI sequences (T1-weighted pre/postcontrast, T2-weighted, FLAIR, and restricted spectrum imaging cellularity maps). Cellular tumor volumes were automatically segmented using nnU-Net and combined with chemotherapy radiotherapy and MGMT status. Models were trained using Extremely Randomized Trees with nested Monte Carlo cross-validation. The baseline Top10 radiomic model achieved an area under the curve (AUC) of 0.765. Incorporation of nnU-Net-derived VolCT significantly improved performance to 0.809 (<i>P</i> = 2.84 × 10<sup>-7</sup>). Adding chemotherapy, radiotherapy, or MGMT status to the Top10+VolCT model yielded further gains, with AUCs ranging from 0.828 to 0.830 (all <i>P</i> ≤ .032). The full model combining VolCT with chemotherapy, radiotherapy, and MGMT achieved the best performance, with an AUC of 0.853 (<i>P</i> = .0006 vs Top10+VolCT). Incorporating chemotherapy, radiotherapy, and MGMT status improves posttreatment glioblastoma classification. Deep learning-derived cellular tumor volumetry further enhances radiomics-based performance, highlighting the value of combining clinical context with advanced computational imaging.