Explainable DSC Perfusion Radiomics for Post-Treatment Glioblastoma Assessment Using a Soft-Voting Machine-Learning Ensemble

Background/Objectives: After chemoradiation for glioblastoma, residual or recurrent tumor (tumor-positive disease, TP) and treatment-related change (TRC), which includes pseudoprogression and radiation necrosis, look alike on conventional MRI. We tested whether radiomic features from dynamic susceptibility contrast (DSC) perfusion MRI alone can separate them. Methods: We analyzed 224 post-treatment MRI examinations from 168 patients in the public UCSD-PTGBM collection. A machine-learning ensemble was trained on 1223 radiomic features from the leakage-corrected relative cerebral blood volume (rCBV) map, with training (109 patients, 143 examinations) and test (59 patients, 81 examinations) data separated at the patient level and the model fixed before testing. The analysis was repeated on the cerebral blood flow (CBF) map from the same scans, and DSC was compared with arterial spin labeling (ASL) in patients who had both. Results: On the independent test set (patient-level analysis), the model distinguished TP from TRC with an area under the receiver operating characteristic curve (AUC) of 0.79 (95% CI 0.67–0.89). It detected 31 of 49 patients with tumor (sensitivity 63%) and correctly identified all 10 patients with TRC (specificity 100%); at a predefined higher-sensitivity cut-off, sensitivity was 69% and specificity 80%. With the CBF map, the AUC did not differ significantly (0.78; p = 0.89), but there was a different balance of sensitivity and specificity. In an exploratory subset of 36 patients, DSC and ASL did not differ significantly (AUC 0.82 vs. 0.71; p = 0.52). Conclusions: DSC perfusion radiomics alone identified tumor with high specificity but missed about one-third of tumors: a positive result may support a diagnosis of tumor, whereas a negative result cannot exclude it. External, multicenter validation is needed before clinical use.