Predicting peritumoral glioblastoma infiltration and subsequent recurrence using deep-learning-based analysis of multi-parametric magnetic resonance imaging

Abstract

Purpose: Glioblastoma (GBM) is the most common and aggressive primary adult brain tumor. The standard treatment approach is surgical resection to target the enhancing tumor mass, followed by adjuvant chemoradiotherapy. However, malignant cells often extend beyond the enhancing tumor boundaries and infiltrate the peritumoral edema. Traditional supervised machine learning techniques hold potential in predicting tumor infiltration extent but are hindered by the extensive resources needed to generate expertly delineated regions of interest (ROIs) for training models on tissue most and least likely to be infiltrated.

Approach: We developed a method combining expert knowledge and training-based data augmentation to automatically generate numerous training examples, enhancing the accuracy of our model for predicting tumor infiltration through predictive maps. Such maps can be used for targeted supra-total surgical resection and other therapies that might benefit from intensive yet well-targeted treatment of infiltrated tissue. We apply our method to preoperative multi-parametric magnetic resonance imaging (mpMRI) scans from a subset of 229 patients of a multi-institutional consortium (Radiomics Signatures for Precision Diagnostics) and test the model on subsequent scans with pathology-proven recurrence.

Results: Leave-one-site-out cross-validation was used to train and evaluate the tumor infiltration prediction model using initial pre-surgical scans, comparing the generated prediction maps with follow-up mpMRI scans confirming recurrence through post-resection tissue analysis. Performance was measured by voxel-wised odds ratios (ORs) across six institutions: University of Pennsylvania (OR: 9.97), Ohio State University (OR: 14.03), Case Western Reserve University (OR: 8.13), New York University (OR: 16.43), Thomas Jefferson University (OR: 8.22), and Rio Hortega (OR: 19.48).

Conclusions: The proposed model demonstrates that mpMRI analysis using deep learning can predict infiltration in the peri-tumoral brain region for GBM patients without needing to train a model using expert ROI drawings. Results for each institution demonstrate the model's generalizability and reproducibility.

Keywords: deep learning; glioblastoma; infiltration; multi-parametric MRI; recurrence.

Fig. 1

Manual region of interest: Near-ROI and Far-ROI example on the T2-FLAIR and T1CE MRI scans.

Fig. 2

Workflow of automated ROI definition. To create infiltrative ROI, manually drawn Far and Near ROI trained patch-based deep learning model to create a prediction map that was further processed. Non-infiltrative ROI is obtained by following expert-defined voxel selection steps.

Fig. 3

Top row shows the detailed steps to obtain an automatically created Infiltrative ROI. (a) Train a patch-based deep learning network with manually drawn Far and Near ROIs. (b) Binarize the heatmap created in panel (a) with a threshold of 0.8 and exclude any voxels of non-infiltrative ROI. (c) Any voxels 1 mm around the tumor core were excluded based on the segmentation map. (d) $n$ voxels were randomly selected from the defined infiltrative ROI and collected $5\times 5\times 5$ patches for further training. The bottom row shows the steps to automatically create non-infiltrative ROI. (e) Obtain a 7 mm boundary of the entire tumor based on brain tumor segmentation. (f) Voxels within 7 mm from the tumor core and recurrence ROIs were excluded. (g) Any voxels in the bottom 20% of T2-FLAIR signal intensity were excluded. (h) Randomly select $n$ voxels from defined non-infiltrative ROI and collect $5\times 5\times 5$ patches for further training.

Fig. 4

Overall workflow of proposed patch-based deep learning training. Patches were collected from each sequence of mpMRI scans using two different classes of ROIs. Collected patches were fed into a deep learning network to train the model. To create probability maps, patches created for target voxels in the edema region were collected and evaluated with the trained deep learning model.

Fig. 5

Network structure of the proposed deep learning model. $5\times 5\times 5\text{voxel}$ patches feed into the network, which output is a value between 0 and 1 to indicate the likelihood of infiltration of the patch’s center voxel.

Fig. 6

Representation of regions used in the proposed evaluation method. R-ROI: Red zone marked by radiologists augmented with an extra 3 mm margin. White sections denote areas excluded from the assessment due to immediate proximity to the tumor core or recurrence. NR-ROI: Encompasses regions distant from the tumor core, R-ROI, and those adjacent zones exempted from the evaluation.

Fig. 7

Example of peritumoral infiltration prediction heatmap. Panel (a) shows pre-operative MRIs including T1, T1CE, T2, and T2-FLAIR. Panel (b) shows pre-operative T1CE scans overlaid with an infiltrative prediction map contrasted with post-operative T1CE scans with recurrence pointed. Panel (c) displays MRI scans at recurrence time points.