Multi-site retrospective analysis of diffusion and perfusion magnetic resonance imaging correlates to glioma characteristics derived from radio-pathomic maps

Abstract

Background: This study determines the relationship between diffusion and perfusion-based magnetic resonance imaging signatures and radio-pathomic maps of tumor pathology in a large, multi-site cohort.

Methods: This study included perfusion imaging from presurgical relative cerebral blood volume (rCBV) images from the UPenn-GBM dataset and presurgical arterial spin labeling (ASL) imaging from the UCSF-PDGM dataset. Diffusion imaging included fractional anisotropy (FA) estimates derived from diffusion tensor imaging for each subject from each institution. A previously validated autopsy-based model was applied to the structural images from each patient to generate quantitative radio-pathomic maps of cell density and extracellular fluid (ECF). Mean cell density, ECF density, FA, rCBV calculated from dynamic susceptibility contrast imaging, and rCBF calculated from ASL were computed for each patient and statistically compared within contrast-enhancement (CE) and the non-enhancing peritumor FLAIR hyperintensity (FH).

Results: Both rCBV and ASL showed a positive correlation with cell density within CE (rCBV: R = 0.280, P < .001; ASL: R = 0.117, P = .023). However, both perfusion metrics also showed no association with cell density within the FH region at the group level (rCBV: R = 0.0162, P = .731; ASL: R = -0.020, P = .652). Negative correlations were observed between FA and ECF density across both CE and FH in both the UPenn-GBM (CE: r = -.204, P < .001, FH: r = -.332, P < .001) and the UCSF-PDGM (CE:r = -.179, P < .001, FH:-0.355, P < .001). Additionally, a positive ASL-cell density association per subject within FH was associated with a worse survival prognosis (HR = 5.58, P = .022).

Conclusions: These results suggest that radio-pathomic maps of tumor pathology provide complementary information to other MR signatures that reveal prognostically valuable signatures of the non-enhancing tumor margin.

Figure 1.

Overview of study methodology. Radio-pathomic maps of cell density and extracellular fluid (ECF) density are generated from T1, T1C, FLAIR, and ADC images that have been aligned to the FLAIR and intensity normalized (aside from ADC). A 5 × 5 sliding frame is used as input to the trained radio-pathomic model for generating maps, which uses a bagging random forest architecture to predict cell/ECF density from ground truth autopsy pathology. These maps are then compared to fractional anisotropy and perfusion-based metrics within the contrast-enhancing region (orange, area surrounding tumor core) and the non-enhancing FLAIR hyperintense region (blue, area surrounding contrast enhancement).

Figure 2.

Associations between perfusion-based rCBV values and predicted cell density within the UPenn-GBM dataset demonstrated a positive association within contrast enhancement (R = 0.28, P < .0001) but not beyond the enhancing margin (R = 0.0162, P = .731). Examples from individual subjects indicate areas of hypercellularity occurring beyond the contrast-enhancing margin that do not show signs of hypervascularity, despite good concordance within the enhancing margin. CPM = cellularity prediction map.

Figure 3.

Associations between arterial spin labeling values and cell density within the UCSF-PDGM dataset, show positive associations within enhancement (R = 0.110, P = .023) and no relationship beyond enhancement (R = −0.020, P = .652), similar to the contrast-based perfusion results seen in the UPenn-GBM dataset. Individual examples also show similar trends of hypercellularity existing in and beyond the FLAIR hyperintense margin that do not show distinguishing perfusion signatures. CPM = cellularity prediction map.

Figure 4.

Examples of perfusion-based metrics acquired near death compared to predicted cellularity maps (CPM), as well as actual cellularity from autopsy pathology (Hist.). Both the cellularity predictive maps and the aligned ground truth histology highlight areas of hypercellularity beyond the contrast-enhancing region with non-elevated perfusion in the region, potentially indicating areas of non-angiogenic tumor infiltration. For example, patients have been included in the training of the CPM model and thus do not necessarily represent prediction behavior on held-out data.

Figure 5.

Correlations between fractional anisotropy (FA) and extracellular fluid (ECF) density indicate a consistent negative association between the 2 variables both within contrast-enhancement and in non-enhancing FLAIR hyperintensity. Examples from individual subjects show reduced FA within areas of high predicted ECF, particularly near the primary tumor mass.

Figure 6.

Kaplan–Meier curve showing different survival outcomes for patients with positive (R > 0.1), negative (R < −0.1), or no (−0.1 < R < 0.1) within-FLAIR perfusion-cellularity correlation (PCC) that had undergone gross total resection in the UCSF-PDGM dataset, indicating that patients with negative PCC survive longer than patients with high PCC. Examples from individual subjects highlight this trend, where a short-term survival subject shows regions of non-enhancing tumor highlighted by both the arterial spin labeling and cell density map, whereas a long-term survivor shows reduced perfusion in an area with moderately increased cellularity.