Investigating How to Optimally Combine Multimodal MRI Data to Better Identify Glioblastoma Infiltration.
Haitham Al-Mubarak1, Antoine Vallatos2, Joanna Birch3, Lindsay Gallagher1, Lesley Glmour4, Anthony Chalmers5, and William Holmes4

1Glasgow Experimental MRI Center, University of Glasgow, Glasgow, United Kingdom, 2Centre of Clinical Brain Science, Edinburgh, United Kingdom, 3Institute of Cancer Sciences, University of Glasgow, Glasgow, United Kingdom, 4University of Glasgow, Glasgow, United Kingdom, 5Glasgow university, Glasgow, United Kingdom


The infiltration of glioblastoma tumour cells into normal tissue presents a major obstacle to effective treatment, may then be responsible for tumour recurrence after surgery. Clinical MRI failed to detect the invasion of tumour cells. The purpose of this study is to investigate how the information contained in the individual MR images and multi-regression analysis can be used to probe of invasion, applying a mouse model of an infiltrative brain tumour.


Glioblastoma (GBM) is the most common and aggressive primary brain tumour, with an average survival of 6-12 months after diagnosis. GBM cells can progressively infiltrate neighbouring normal brain regions. For surgical resection and radiotherapy planning, it is important to be able to identify the outer boundary of this infiltration, however conventional MRI fails to do this. Hence, it is crucial to develop methods that better enable delineation of marginal regions with low tumour cell density. In a previous work, we investigated the ability of individual MRI contrasts (T2, CE-T1, DWI, ADC, DTI, ASL) to identify the boundary of infiltration. This was quantitatively assessed by using stacks of in-plane histological sections to generate tumour cell density maps, which were then co-registered to the MR images. In this study, we investigate how the information contained in the individual MR contrast images can best be combined, to achieve this we performed a voxel-by-voxel multi-regression analysis of the co-registered MRI/histology dataset.

Material and methods

Nine nude CD1 mice were injected intracranially with infiltrative human glioblastoma cells (G7). Multimodal MRI was acquired, T2w, DE-T1, DTI, multi boli Arterial Spin label (mbASL) [1]. Following MRI, mice were sacrificed and the brains were frozen. Staining for Human Leukocyte Antigen (HLA) allowed determining the cellular burden. Five histological slices were cut in the plane of the MRI slice, registered and stacked to account for MRI slice thickness. A Tumour Cell Density maps (TCD) was produced and the outer boundary of a tumour manually delineated (see fig.1). A multiple quadratic regression method (equation 1) was used to find the best combination of MR data, i.e. the coefficients bi. Finally, these coefficients were used to generate a Quadratic Regression Map (QRM), with which to compare with the actual TCD (see fig.2).

y=b °+b1x1+b2x2+b3x1x2+b4x2+b5x2 (equation 1)


Figure 3a,b shows two QRM maps created from FA, ADC , mbASL and T2w MR images. These QRM maps exhibit high linear correlation (r>0.8) the corresponding TCD map generate with histology (HLA).


QRM maps were generated using a multiple quadratic regression method and compared to TCD maps generated from stacked in-plane histology (HLA). This approach shows much promise in probing the boundary of tumour cell infiltration in glioblastoma.


H. Al-Mubarak would like to thanks the Ministry of Higher Education and Scientific Research in Iraq for financial support.


[1] Vallatos and et al, Magnetic Resonance in Medicine, 2017.


Figure.1: Example of creating a TCD from three histology (HLA) sections (20 μm thickness) after co-registration.

Figure2: Combining of several MR images to create a QRM by using multiple quadratic regression method.

Figure 3: Multimodal QRM created from Mr Images (a) QRM generated from (T2w, ADC and FA) (b) QRM generated from (T2w, mbASL and FA).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)