Artificial intelligence-based locoregional markers of brain peritumoral microenvironment

Abstract

In malignant primary brain tumors, cancer cells infiltrate into the peritumoral brain structures which results in inevitable recurrence. Quantitative assessment of infiltrative heterogeneity in the peritumoral region, the area where biopsy or resection can be hazardous, is important for clinical decision making. Here, we derive a novel set of Artificial intelligence (AI)-based markers capturing the heterogeneity of tumor infiltration, by characterizing free water movement restriction in the peritumoral region using Diffusion Tensor Imaging (DTI)-based free water volume fraction maps. We leverage the differences in the peritumoral region of metastasis and glioblastomas, the former consisting of vasogenic versus the latter containing infiltrative edema, to extract a voxel-wise deep learning-based peritumoral microenvironment index (PMI). Descriptive characteristics of locoregional hubs of uniformly high PMI values are then extracted as AI-based markers to capture distinct aspects of infiltrative heterogeneity. The proposed markers are utilized to stratify patients' survival and IDH1 mutation status on a population of 275 adult-type diffuse gliomas (CNS WHO grade 4). Our results show significant differences in the proposed markers between patients with different overall survival and IDH1 mutation status (t test, Wilcoxon rank sum test, linear regression; p < 0.01). Clustering of patients using the proposed markers reveals distinct survival groups (logrank; p < 10^-5, Cox hazard ratio = 1.82; p < 0.005). Our findings provide a panel of markers as surrogates of infiltration that might capture novel insight about underlying biology of peritumoral microstructural heterogeneity, providing potential biomarkers of prognosis pertaining to survival and molecular stratification, with applicability in clinical decision making.

Figure 1

Different steps of the processing pipeline. The pipeline consists of three steps. (1) creation of free water map and masks of tumor and edema, (2) creation of the PMI map, (3) extracting locoregional hubs and AI-based markers. PMI: Peritumoral microenvironment index, CNN: Convolutional neural networks, FERNET: Freewater estimatoR using Interpolated Initialization, MRI: Magnetic resonance imaging.

Figure 2

Datasets: train, validation, test (survival and IDH1-mutation analysis). IDH1: Isocitrate-Dehydrogenase. NOS/NEC: Not elsewhere classified/Not otherwise specified.

Figure 3

Generation of the PMI map and locoregional hubs. The inputs to the CNN are patches (boxes) extracted from the free water volume fraction map in the peritumoral region from both glioblastomas (red) and metastases (blue) labeled as low free water and high free water which are used to train the CNN. Locoregional hubs are extracted from PMI. Descriptive characteristics of the locoregional hubs are extracted as AI-based markers. PMI: Peritumoral microenvironment index, CNN: Convolutional neural networks.

Figure 4

A schematic view of locoregional hubs descriptive characteristics: number of hubs, size of hubs, shape heterogeneity, directional heterogeneity, and spatial heterogeneity. The purple circle displays the tumor core and blue hubs are located in the peritumoral edema. Arrows are directed toward increasing the values of the descriptive characteristics. Figure in the top-left is the reference figure for the hubs. (i) Moving from left to right, to the top-middle: the spatial heterogeneity of hubs increases; to the top-right: the shape heterogeneity increases. (ii) Moving from top to the bottom, left: the number of hubs increases; center: size of hubs increases; right: directional heterogeneity increases.

Figure 5

The PMI map and AI-based markers for IDH1-wildtype grade 4 glioma patients with different duration of survival. (a) AI-based markers (Descriptive characteristics of PMI locoregional hubs) for long and short survival groups, p value < 0.05 (*), p value < 0.005 (**), Linear regression was used with age and sex as covariates. (b) representative samples of the PMI map with T1CE, FA, MD, and T2-FLAIR images for long and short survival patients. (c) Kaplan–Meier estimates of the two clusters generated by K-means clustering. PMI: Peritumoral Microenvironment Index, IDH1: Isocitrate-Dehydrogenase 1, T1CE: T1 post-contrast, FA: Fractional anisotropy, MD: Mean diffusivity, T2-FLAIR: T2 weighted fluid attenuated inversion recovery.

Figure 6

The PMI map and AI-based markers for grade 4 glioma patients with different IDH1 mutation status. (a) AI-based markers (Descriptive characteristics of PMI locoregional hubs) for IDH1-mutant and IDH1-wildtype groups, p value < 0.05 (*), p value < 0.005 (**), Linear regression was used with age and sex as covariates. (b) representative samples of the PMI map with T1CE, FA, MD, and T2-FLAIR images for IDH1-mutant and IDH1-wildtype patients. IDH1: Isocitrate-Dehydrogenase 1, T1CE: T1- and post-contrast, FA: Fractional anisotropy, MD: Mean diffusivity, T2-FLAIR: T2 weighted fluid attenuated inversion recovery.