Motivation: Non-invasively distinguishing recurrent tumor (rTumor) from treatment-induced effects (TxE) is an ongoing challenge for treating glioma. Goal(s): We build off previous work to spatially map probabilities of rTumor and TxE by expanding the maps beyond the visually identifiable lesion, include robust evaluation criteria, and associate prediction maps to survival. Approach: We predict abnormal brain regions to generate maps of TxE without relying exclusively on the T2-lesion, and associate spatial prediction features to survival. Results: The model discriminates TxE from rTumor with an AUROC=0.74 and abnormal brain with 0.99, which enables expansion of inference. Spatial prediction features were significantly (p<.05) associated with survival. Impact: Spatial maps of recurrent glioma and treatment-effects that account for normal brain demonstrate reliable performance for mapping glioma beyond the visually identifiable lesion with anatomical MRI. Spatial prediction features are associated with survival and may enhance treatment decisions.
Ellison et al. (Tue,) studied this question.