Radiomic spatial maps for identifying viable tumor extent on multi-parametric MRI for Glioblastoma

ABSTRACT: Glioblastoma (GBM) has a complex infiltrating tumor microenvironment which extends well beyond the visible enhancing tumor margins and plays a substantial role in GBM recurrence and poor outcomes.

Unfortunately, in the absence of a precise spatial map of tumor extent, it is often difficult to differentiate infiltrating tumor from vasogenic edema on clinical MRI during radiation/surgical planning. The untreated infiltrating tumor ultimately contributes to over 90% of GBM recurrences. An equally pressing challenge is the difficulty in

distinguishing recurrent tumor from treatment-effects following chemoradiation. Due to the histologically diverse landscape of post-treated lesions, treatment-effects often co-exist with tumor recurrence, and mimic appearance on imaging. In the absence of reliable tools, 15-20% of patients with GBM recurrence are incorrectly diagnosed

due to sampling error associated with intracranial biopsy. Thus, developing a non-invasive spatial map of GBM tumor extent that can reliably identify infiltrating/recurrent tumor from confounding pathologies (treatment- effects/edema), will have significant implications in radiation/surgical-planning and post-treatment management.

Recently, we developed a Radiomic-Image (Rad-I) map of tumor extent that uses computational features corresponding to the micro-architectural image measurements of disorder in the local intensity gradients (i.e., gradient entropy). The initial version of the Rad-I map has been evaluated to distinguish recurrent tumors versus

treatment-effects on post-treatment Gd-T1w MRI with an 85% accuracy on n=75 studies, and to distinguish infiltrating tumor versus vasogenic edema on pre-treatment MRI scans with a 94% accuracy on n=42 studies. In this R01 project, we propose to improve on our initial version of Rad-I map by incorporating (1) additional

anatomical (T2w, FLAIR) and functional MR sequences (perfusion) and (2) a novel “lesion complexity” feature, which captures organizational changes in the tissue composition via graph-theoretic approaches on MRI scans. Overcoming limitations pertaining to small cohorts and lack of spatially mapped ex-vivo histology for validation,

Rad-I maps will be extensively validated on (1) a large multi-institutional MRI cohort with co-localized histopathology and (2) the PRESERVE clinical trial designed to capture GBM heterogeneity via multiple co- localized tissue samples/lesion. These cohorts will also allow for establishing associations of our new radiomic

features with underlying histological/molecular tumor characteristics- a prerequisite for clinical adoption. Lastly, Rad-I maps will be evaluated within a tumor board survey to address the clinically challenging problem of distinguishing recurrent tumors versus treatment effects. Criteria for success for Rad-I maps are that they are

at least non-inferior to the accuracy of stereotactic biopsies (85-90%) in identifying tumor niches corresponding

to (a) viable/infiltrating tumor vs. edema and (b) recurrent tumor vs. treatment effects. Multi-institutional validation and end-user (tumor board) feedback will further confirm the utility of Rad-I maps as a noninvasive alternative to surgical biopsies; thereby paving the way for radiation/surgical and post-treatment management in GBM tumors.