Noninvasive Autopsy-Validated Tumor Probability Maps Identify Glioma Invasion Beyond Contrast Enhancement

Abstract

Background and objectives: This study identified a clinically significant subset of patients with glioma with tumor outside of contrast enhancement present at autopsy and subsequently developed a method for detecting nonenhancing tumor using radio-pathomic mapping. We tested the hypothesis that autopsy-based radio-pathomic tumor probability maps would be able to noninvasively identify areas of infiltrative tumor beyond traditional imaging signatures.

Methods: A total of 159 tissue samples from 65 subjects were aligned to MRI acquired nearest to death for this retrospective study. Demographic and survival characteristics for patients with and without tumor beyond the contrast-enhancing margin were computed. An ensemble algorithm was used to predict pixelwise tumor presence from pathological annotations using segmented cellularity (Cell), extracellular fluid, and cytoplasm density as input (6 train/3 test subjects). A second level of ensemble algorithms was used to predict voxelwise Cell, extracellular fluid, and cytoplasm on the full data set (43 train/22 test subjects) using 5-by-5 voxel tiles from T1, T1 + C, fluid-attenuated inversion recovery, and apparent diffusion coefficient as input. The models were then combined to generate noninvasive whole brain maps of tumor probability.

Results: Tumor outside of contrast was identified in 41.5% of patients, who showed worse survival outcomes (hazard ratio = 3.90, P < .001). Tumor probability maps reliably tracked nonenhancing tumor on a range of local and external unseen data, identifying tumor outside of contrast in 69% of presurgical cases that also showed reduced survival outcomes (hazard ratio = 1.67, P = .027).

Conclusion: This study developed a multistage model for mapping gliomas using autopsy tissue samples as ground truth, which was able to identify regions of tumor beyond traditional imaging signatures.

FIGURE 1.

Imaging examples, clinical characteristics, and survival analyses for patients with tumor outside contrast enhancement. Analyses were performed separately for the full data set, patients with MRI less than 90 days before death, and patients with a primary GBM. Survival analyses were conducted using Cox proportional hazards regression and only include patients who have received Rad + TMZ treatment. Results indicate increased TOC frequency among GBMs and patients who have received Rad + TMZ or Bev treatment. Patients with TOC also show reduced survival compared with patients without TOC. *P < .05, **P < .01. GBM, glioblastoma; TMZ, temozolomide; TOC, tumor outside of contrast enhancement.

FIGURE 2.

Pathological tumor prediction. A, The pathological tumor prediction model uses cell density, ECF, and Cyt segmentations to distinguish tumor vs nontumor using the pathological annotations as ground truth. B, The RUS Tree algorithm was the highest-performing tumor prediction model (AUC = 0.857) and was used in the final multistage TPM model. C, Example segmentations and D, TPMs from the RUS Tree model show accurate tumor prediction in both infiltrative tumor and PN areas, while avoiding both normal tissue and areas of necrosis. AUC, area under the curve; Cyt, cytoplasm; ECF, extracellular fluid; PN, pseudopalisading necrosis; RUS, random undersampled; TPM, tumor probability map.

FIGURE 3.

Radio-pathomic maps of tissue segmentations. A, 5 by 5 voxel tiles from T1, T1 + C, FLAIR, and ADC were used to predict voxelwise cell density, ECF, and Cyt using bagging random forests. B, Test set performance indicates an average subject-level root mean squared error within a standard deviation of the tissue ground truth for each tissue class, indicating satisfactory model performance for most subjects. C, Example tissue predictions show areas of accurately predicted cellularity beyond contrast enhancement, distinguishing between vasogenic edema and hypercellular areas within the FLAIR hyperintense region. The subject on the left shows a portion of hypercellularity extending posterior to the contrast-enhancing margin, while highlighting a reduction in cell density within the FLAIR hyperintensity anterior to the contrast enhancement. The subject on the right shows an area of increased cellularity in the absence of contrast enhancement and diffusion restriction. ADC, apparent diffusion coefficient; Cyt, cytoplasm; ECF, extracellular fluid; FLAIR, fluid-attenuated inversion recovery; GBM, glioblastoma.

FIGURE 4.

Example TPM for a test set subject (GBM, Female, 80yo), along with corresponding MRI and segmentation predictions. The TPM for this patient accurately highlights an area of tumor outside contrast enhancement and in the absence of diffusion restriction. In addition, the cell density map highlights that the entire tumor area is hypercellular, whereas the ECF map highlights a high-ECF core to the hypercellular area, which correctly suggests an area of PN (high cellularity, high ECF). This demonstrates the ability for TPMs to distinguish between pathologically distinct regions of tumor. Cyt, cytoplasm; ECF, extracellular fluid; FLAIR, fluid-attenuated inversion recovery; PN, pseudopalisading necrosis; TPM, tumor probability map.

FIGURE 5.

Example TPMs for 3 individuals. The first case presents an area of high tumor probability outside of contrast enhancement, avoiding a central necrotic area and identifying tumor extent that goes beyond the FLAIR hyperintense signal with pathological conformation of tumor presence. The second case shows high tumor probability well circumscribed within the contrast-enhancing lesion, correctly identifying a lack of tumor outside of contrast in this area while highlighting increased cellularity in the posterior edge of the tumor relative to the anterior. The third case not only shows an area of high tumor probability in an area with some tumor invasion but also includes reactive gliosis in the high tumor probability area, indicating a representative false positive. FLAIR, fluid-attenuated inversion recovery; PN, pseudopalisading necrosis; TPM, tumor probability map.

FIGURE 6.

External TPMs and survival analyses for predicted tumor outside of contrast enhancement presurgery. A, Example TPMs applied to externally collected data show similar-quality TPMs to those included in our study and show areas of predicted tumor beyond contrast enhancement. External Data 3 show additional histology from biopsy-confirmed tumor in a nonenhancing area of predicted high tumor probability presurgery. B, Examples of pTOC and no pTOC, showing patients with predicted nonenhancing tumor presence compared with patients with no predicted tumor presence outside of contrast enhancement. C, Kaplan–Meier survival curves showing longer overall survival in patients with pTOC presurgery (HR = 1.67 (1.07-2.62), P = .027), similar to the pattern observed with autopsy-confirmed tumor outside contrast before death seen in Figure 1. FLAIR, fluid-attenuated inversion recovery; GBM, glioblastoma; HR, hazard ratio; PN, pseudopalisading necrosis; pTOC, predicted tumor outside of contrast enhancement; TPM, tumor probability map.