Abstract Despite prior success in classifying recurrent glioma noninvasively with multi-parametric MRI and AI, clinical applicability has yet to be demonstrated due to a lack of robust model evaluation and spatial preservation of tumor characteristics. This study develops, robustly evaluates, and clinically validates an interpretable model for predicting recurrent tumors from spatially varying, histopathologically-confirmed tissue samples. Machine learning models were developed using 254 pre-surgical multi-parametric MRI patches surrounding coordinates of tissue samples taken during recurrent surgery. A test AUROC of 0.74 ± 0.08 for distinguishing recurrent tumors, and 0.99 ± 0.01 for normal-appearing brain, demonstrated the feasibility of spatially mapping heterogeneity. Important features were consistent with current literature, and uncertainty was correlated with model failures ( p ≤ 0.05). Volumetrics derived from prediction maps of recurrent tumors generated using a separate cohort of 56 patients with recurrent high-grade gliomas were significantly associated with survival. These results demonstrate a step towards clinical applicability of spatially mapping glioma recurrence.
Ellison et al. (Thu,) studied this question.