Abstract
BACKGROUND
Radiomic texture analysis (TA) from standard MRI imaging may be able to discriminate between responders versus non-responders in glioblastoma patients treated with pembrolizumab immunotherapy.
METHODS
14 patients (5 males; mean age 58 years; range: 32–72 years), with pathologically-proven recurrent GBM, enrolled in a pembrolizumab clinical trial, were retrospectively evaluated. Immunotherapy Response Assessment in Neuro-Oncology(iRANO) were performed. Patients were categorized based on: 1) best response or 2) overall response (OR) using the iRANO status at the last scan time in the trial. Patients with progressive disease (PD) were classified as non-responders, while patients with partial response (PR) or stable disease (SD) were classified as responders. T2-FLAIR (edema/invasion) and post-contrast T1WI (enhancing tumor) of baseline scans were co-registered and segmented (3D Slicer, v.4.3.1) to create a volume of interest for Radiomic TA. A total of 4880 texture features were extracted. Feature selection was performed using Lasso regularization. For classification and predictive model building, gbtree booster of XGboost with Leave-One-Out Cross-Validation (LOOCV) was used on the selected texture features to build a binary logistic regression model and classify the patients into respective groups<.strong>
RESULTS
Using the best response classification, 10 patients were classified as non-responders and four patients classified as responders (1 SD; 3 PR). Using 13 radiomic features, these patients could be classified into their respective responding groups with a sensitivity, specificity and accuracy of 100%, p-value=0.0089. Based on OR, 12 patients were classified as non-responders and two as responders (2 SD). Seven features were able to differentiate the responding patients with a sensitivity, specificity and accuracy of 100%, p-value=0.0089.
CONCLUSION
Radiomic TA was able to discriminate and predict those GBM patients that are responders versus non-responders to pembrolizumab with high robustness. Of note, given the small number of patients in this cohort, a larger cohort of patients is needed to minimize overfitting.