Machine Learning-Based Comparison of Non-Contrast ASL and DSC-MRI Perfusion for Differentiating Recurrent High-Grade Glioma from Treatment Effects

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background/Objectives: Differentiating high-grade glioma recurrence from treatment-related changes remains challenging on conventional MRI. This study evaluated non-contrast arterial spin labeling (ASL) perfusion MRI for this distinction and compared its performance with dynamic susceptibility contrast (DSC) perfusion MRI. Methods: Postoperative follow-up MRI examinations obtained between November 2019 and May 2021 were retrospectively reviewed. The cohort included 63 MRI examinations from 36 adults treated for high-grade glioma. ASL, routine MRI, and DSC images were independently assessed by two neuroradiologists. Final diagnosis was based on histopathology or longitudinal clinical and imaging follow-up. Reader agreement, diagnostic performance, and an exploratory patient-level grouped machine learning analysis using ASL-only, DSC-only, and combined ASL–DSC features were evaluated. Results: ASL- and DSC-derived perfusion parameters were significantly higher in tumor recurrence than in treatment-related changes (all p < 0.001). Both techniques showed high diagnostic performance; DSC achieved the highest accuracy, whereas ASL provided high specificity across readers. Inter-reader agreement ranged from substantial to almost perfect. In grouped cross-validation, ASL-only, DSC-only, and combined models achieved mean AUCs of 0.946, 0.997, and 0.997, respectively. Permutation testing confirmed that combined-model performance exceeded chance expectations (empirical p = 0.002). Conclusions: Non-contrast ASL perfusion showed diagnostic performance comparable to DSC for differentiating high-grade glioma recurrence from treatment effects. ASL may provide a reliable non-invasive alternative for longitudinal surveillance, particularly when gadolinium administration is undesirable or contraindicated.