Detecting glioblastoma infiltration beyond conventional imaging tumour margins using MTE-NODDI Open Access

Glioblastoma (GBM) is the most common and aggressive primary brain tumour in adults (Omuro & DeAngelis, 2013). The current standard of care includes maximal safe surgical resection followed by radiotherapy and concomitant chemotherapy with temozolomide (Stupp et al., 2005). Despite such active treatment strategies, the prognosis remains extremely poor and the outcome for nearly all patients is dismal, with a median survival of <15 months and an average 5-year survival rate of <5% (Miller et al., 2021). The key cause of resistance to available treatments is diffuse infiltration of tumour cells into the normal brain parenchyma, a hallmark of GBM (Cuddapah et al., 2014; Vehlow & Cordes, 2013). Infiltration precludes curative surgery and causes recurrence from low numbers of cells that have invaded past the margin of original resection. Radiotherapy target volumes are planned to account for an unseen population of invading tumour cells into healthy tissues that can extend beyond the gross tumour volume (Niyazi et al., 2023), yet tumour recurrence occurs, possibly indicating migrating tumour cells outside the planned treatment area. As cell invasion may not be uniform from the tumour bulk, an image-guided approach may provide a method to more closely map migrating tumour cells and thus allowing a more precisely tailored therapy plan. Invasion in GBM is thus a major clinical challenge because traditional imaging methods, which include post-gadolinium contrast-enhanced T1-weighted, T2-weighted and fluid attenuated inversion recovery (FLAIR) MRI, cannot detect these infiltrating cells, making it impossible to accurately determine tumour margins (Lasocki & Gaillard, 2019; Price & Gillard, 2011). Other methods to aid tumour delineation include 5‑aminolevulinic acid (5‑ALA)-guided surgery as an adjunct to maximise resection (Hadjipanayis et al., 2015; Schatlo et al., 2015; Stepp & Stummer, 2018; Stummer et al., 2006) and intraoperative MRI (Kubben et al., 2011; Schatlo et al., 2015). However, these imaging techniques offer limited identification of the invading margin, which inevitably leads to tumour recurrence within 2 cm of the original tumour site in 80–90% of cases (Chamberlain, 2011; Claes et al., 2007; De Bonis et al., 2013; Petrecca et al., 2013; Vehlow & Cordes, 2013). Recent studies utilising magnetic resonance spectroscopy (MRS) (Laack et al., 2021), ¹⁸F-DOPA positron emission tomography (PET) (Laack et al., 2021), and advanced MRI (Kim et al., 2021) for dose-escalated radiation therapy have shown early promise, suggesting that improving tumour volume definition can lead to better treatment planning in GBM.

Diffusion weighted imaging (DWI) is sensitive to diffusion of water molecules in biological tissues. DWI techniques provide non-invasive indices associated with the pattern of water diffusion, which reflect tissue microstructure (Basser et al., 1994; Le Bihan, 2014). As GBM cells infiltrate away from the tumour bulk into adjacent brain tissues, they confront new and heterogenous microenvironments, resulting in tumour-affected changes to the underlying tissue. As such, DWI may provide improved sensitivity to microstructural changes caused by the invading tumour, in turn delineating margins outside the traditionally identified tumour bulk (Maier et al., 2010; Sugahara et al., 1999; Thoeny & Ross, 2010). Neurite orientation dispersion and density imaging (NODDI) is an advanced method of DWI, which uses a biophysical compartment-based model for studying microstructural changes of brain tissue (Zhang et al., 2012). NODDI assumes that the signal measured from each tissue voxel originates from a combination of three types of microstructural compartments: intra-neurite, extra-neurite, and free water volume fractions, each of which have a unique effect on water diffusion within the environment. This allows for compartment-specific markers of microstructural properties to be determined, such as the neurite density index (NDI) (the fraction of tissue that comprises neurites; axons or dendrites), the extra-neurite volume fraction (f_en) (the fraction of tissue other than the neurites including microglia, astrocytes, oligodendrocytes, soma, ependymal cells, extra-cellular matrices, and vasculature), and orientation dispersion index (ODI), which reflects the spatial configuration of the neurite structures (Zhang et al., 2012). NODDI-derived measures are sensitive to altered tissue microstructure in brain development, maturation, and aging (Cox et al., 2016; Kamiya et al., 2020), as well as numerous neuropathological conditions (Churchill et al., 2019; Colgan et al., 2016; Fu et al., 2020; Kamagata et al., 2017; Kamiya et al., 2020; Kraguljac et al., 2021; Mastropietro et al., 2019; Mitchell et al., 2019; Nazeri et al., 2017; Palacios et al., 2020; Parker et al., 2018; Schneider et al., 2017; Sone et al., 2018; Spano et al., 2018; Z. Wang et al., 2019; Winston et al., 2014). Previous studies have reported the application of NODDI in investigating brain tumour bulk, differentiating primary and metastatic tumours (Caverzasi et al., 2016; Kadota et al., 2020; Mao et al., 2020), as well as regions of oedema within tumours (Masjoodi et al., 2018; Okita et al., 2023), and tumour grading (Figini et al., 2018; Maximov et al., 2017; Wen et al., 2015; Zhao et al., 2018).

NODDI is still an emerging technology, akin to contemporary compartmental models of DWI. Currently, the existing NODDI model (Zhang et al., 2012) does not account for differences in T2 relaxation between the model compartments. Therefore, when differences in compartmental T2 exist, it is difficult to disentangle differences in diffusion from relaxation. This results in suboptimal estimation of the NODDI parameters, with overestimation of the free-water volume fraction leading to inaccuracies in measurements of NDI and ODI (Bouyagoub et al., 2016), potentially impacting the application of NODDI to conditions with T2 changes, such as in tumour tissue (Lampinen et al., 2019; Wen et al., 2015). Gong et al. (2020) recently proposed the multiple echo time NODDI (MTE-NODDI) technique, which provides robust estimates of the non-T2-weighted NODDI parameters and compartment-specific T2 values from diffusion data acquired at multiple echo times. With its ability to provide the same NODDI measures with improved parameter precision, MTE-NODDI may enhance the conspicuity of neuropathology, including imaging of brain tumours and the invading margin.

Therefore, in this study, we investigate the effectiveness of MTE-NODDI as a non-invasive in vivo imaging technique to improve detection of the tumour margin and demonstrate sensitivity of the technique to infiltrating tumour cells in a well-characterised patient-derived xenograft mouse model of GBM. We further report correlations of MTE-NODDI parameters with fluorescence intensity from green fluorescent protein (GFP) positive tumour cells and evaluate the technique’s specificity to tumour-induced microstructural pathology.

All procedures were performed in compliance with the Animal Scientific Procedures Act, 1986 and approved by the UCL Animal Welfare and Ethical Review Body (AWERB) in accordance with the international guidelines of the Home Office (UK).

A patient-derived xenograft mouse model of GBM was used in the current study. The patient (G144) was diagnosed with malignant astrocytoma (GBM), with increased copy numbers of chromosome 7 (EGFR genes map to this chromosome) (Pollard et al., 2009). Xenografts were performed on CD-1 nude mice using the G144 cell line which shows propensity for invasion through the white matter into the contralateral hemisphere. The tumour model utilised in this study was originally established in female mice (Brooks et al., 2021; Pollard et al., 2009) and, as such, only female mice were used. 8–12 week-old immunocompromised mice (n = 7) (purchased from Charles River Laboratories) underwent stereotactic implantation of 1 x 10⁵ GFP-labelled G144 cells (anteroposterior 0, mediolateral -2.5, dorsoventral -3). The tumour model used in this study shows high latency with all the mice presenting tumours and they were imaged at three timepoints post injection; 4, 6, and 8 weeks, following which they were sacrificed. A separate group of age- and strain-matched naïve mice (n=3) were used in a control experiment and underwent the same imaging protocol as above.

Prior to commencing the MRI acquisitions, the mice were placed in an induction chamber and anesthetised with inhaled isoflurane (2% isoflurane at 1 l/min O₂) until withdrawal reflex was lost. They were then transferred into the MRI mouse cradle with bite bar, nose cone, and ear bars to ensure a well-secured head position to minimise motion artifacts. Eye ointment was applied to prevent drying. Temperature and breathing rate were monitored throughout all the experiments using a rectal probe and a respiration pad (SA Instruments). Core body temperature was maintained at 37 ± 0.5°C via regulation of an adjustable temperature water bath. Isoflurane level was manually adjusted throughout the duration of the scan to maintain a consistent respiration rate of ~100 bpm.

Images were acquired on a 9.4 T Bruker imaging system (BioSpec 94/20 USR) with a horizontal bore and 440 mT/m gradient set with an outer/ inner diameter of 205 mm/116 mm, respectively (BioSpec B-GA 12S2), 86 mm volume coil, and a four-channel array receiver-surface coil (RAPID Biomedical GmbH) for the transmission and reception of RF signal. Tumours were localised using a structural T2-TurboRARE sequence (fast-spin echo, Paravision v6.0.1). The olfactory bulbs were used as an anatomical landmark to maintain consistency in slice positioning between subjects, and the slices covered the cortex and all subcortical structures up to the cerebellum. Imaging parameters for the T2-weighted acquisition were as follows: TR = 4000 ms, TE = 45 ms, FOV = 21 x 16 mm², data matrix 256 x 196, and 14 x 600 µm coronal slices.

For MTE-NODDI, DWI images were acquired using a 4-shot spin echo-planar imaging (EPI) sequence. Imaging parameters were: 14 slices, slice thickness = 600 µm; FOV = 20 x 16 mm²; data matrix 110 x 85; and TR= 2500 ms. To implement MTE-NODDI, the DWI images were acquired at three different echo times of 30, 45, and 60 ms. At each echo time, MRI protocol consisted of two shells, detailed as follows:

1. Shell One: 30 directions, five b = 0 s/mm² images, and diffusion weighting of b = 2000 s/mm²

2. Shell Two: 15 directions, five b = 0 s/mm² images, and diffusion weighting of b = 700 s/mm²

with gradient duration and separation δ∕Δ = 4.5∕11 ms for all b-values and TE’s. The overall acquisition time for each mouse per imaging session was ~120 minutes.

The effects of noise and imaging artifacts on the DWI images were reduced by applying a denoising method based on the random matrix theory (MRtrix3) (Veraart et al., 2016), correction of B0 inhomogeneity, and motion with TOPUP (Andersson & Sotiropoulos, 2016) in FMRIB Software Library (FSL, University of Oxford, UK). The T2-weighted and DWI acquisitions were then co-registered to a reference b=0 image. T2-weighted images were resampled to match the DWI images in AFNI (Analysis of Functional NeuroImages, National Institutes of Health, USA), and brain masks were created manually on T2-weighted images using ITK-SNAP. Mean diffusivity (MD) measures were generated from diffusion data obtained at TE = 30 ms using dtifit in FSL, which fits a diffusion tensor model at each voxel of the data that has been pre-processed. NODDI parameters were estimated for each TE session separately with the NODDI MATLAB Toolbox, and the estimated parameters were aligned to the first TE session to extract TE-independent MTE-NODDI parameters (Gong et al., 2020).

For quantitative analysis, tumour regions of interest (ROIs) were manually defined on T2-weighted images (T2WI) using ITK-SNAP by the two authors (SK and DG) who were blinded to the other’s results. The images and ROIs were presented to a consulting neurosurgeon (CSH) for approval. In our experience with the G144 model, tumour cells exhibit a spread beyond the tumour bulk, infiltrating into the peritumoural regions of corpus callosum, striatum, and cortex regions (Brooks et al., 2021). Therefore, the following ROIs were selected:

Regions with hyperintense signal in the ipsilateral hemisphere on the coronal T2WI were identified as tumour bulk.

Regions outside the tumour bulk, where there are no T2WI changes, were defined in the ipsilateral hemisphere in three distinct anatomical regions—corpus callosum, striatum, and cortex. Given that the G144 model replicates the GBM feature of invasion along the corpus callosum, a margin region in the corpus callosum adjacent to the tumour bulk was always selected. ROIs were adjusted to exclude any confounding anatomical structures (e.g., the lateral ventricle in the striatal ROI, grey matter structures in the corpus callosum ROI, etc.). These exclusion criteria resulted in ROIs of the following sizes: corpus callosum—1 pixel (~0.182 mm) in the xy direction; striatum and cortex—up to 3 pixels (~0.550 mm). As can be seen in the representative images presented in this manuscript (Fig. 2), confounding anatomical structures such as external capsule may contaminate striatal ROIs (or corpus callosum in the cortical ROI) if extended beyond 3 pixels. In the contralateral hemisphere, ROIs equivalent to the peritumoural margins were defined as controls for comparison. Subsequently, the same ROIs were copied onto diffusion MRI images and mean values for NDI, f_en, ODI, and MD were extracted for quantitative analysis.

Due to fixation and shrinkage of tissue during histology, a pixel-by-pixel comparison was not possible across the MRI and histology data sets. As such, ROIs encompassing the cortex across both hemispheres of the brain were chosen to enable an unbiased comparison between the two sets of images (Supp. Fig. 1). In animals where tumour bulk was identified in the cortex, these pixels were excluded from both the MRI and GFP images to ensure accurate correlation analysis. Opting for an ROI covering the entire cortex simplifies delineation, while simultaneously mitigating any artifacts and inaccuracies caused due to tissue fixation.

G144 tumour cells were transduced to constitutively express GFP using a lentiviral approach. Lentivirus carrying the GFP transgene was produced by cotransfecting with the HIV-1 packaging vector Delta8.9 and the VSVG envelope glycoprotein into 293 T cells using polyethylenimine. Virus was concentrated by ultracentrifugation (3 h, 50,000 × g, 4°C) (Brooks & Parrinello, 2017). For quantification of the fluorescence intensity from GFP+ tumour cells, mice were perfused with 4% PFA and the brain post-fixed in 4% paraformaldehyde overnight at 4°C. Two explanted brains from the MRI-imaged mice were assigned for downstream mechanistic analyses and, as such, were not included in this study, rendering them unavailable for immunofluorescence. From the remaining mice (n = 5), 50 μm vibratome sections were cover-slipped, dried, and imaged using a Leica DMi8 Inverted microscope (Leica Microsystems, Wetzlar, Germany). Immunofluorescence images from each mouse (n = 5; 1 slice per mouse) were individually imported into ImageJ, and ROIs encompassing the entire cortex were drawn. Images were then thresholded by intensity to eliminate any signal from background, and signal intensity from GFP-positive tumour cells within the ROI was quantified as % Area under fluorescence and exported into GraphPad Prism for correlation with MRI measurements.

Statistical analysis of the data was performed using GraphPad Prism 9.0. All data are expressed as mean ± SD and were tested for normality using Shapiro-Wilk test. For the intra-animal changes between contralateral and peritumoural margin regions, paired t-tests were used to test for statistically significant difference. For the longitudinal study, we reported the data as mean ± SD of paired differences in all animals at each timepoint. In addition, Pearson correlation analysis was used to determine the correlation between MRI parameters and GFP+ fluorescence intensity from tumour cells. For all statistical tests, the significance level was p < 0.05.

We probed the invading margin of GBM using MTE-NODDI (Gong et al., 2020), a recent improvement of the conventional NODDI model (Zhang et al., 2012), which enables us to estimate non-T2-weighted NODDI parameters such as NDI, f_en, and ODI. A well-characterised G144 patient-derived xenograft mouse model, which recapitulates the hallmark feature of GBM to invade to the contralateral hemisphere along the white matter tracts of the corpus callosum (Brooks et al., 2021; Pollard et al., 2009), was used in combination with MTE-NODDI.

To assess the sensitivity of MTE-NODDI to the spread of tumour cells, GFP-labelled G144 cells were stereotactically injected into the striatum of immunocompromised mice and imaged longitudinally on the 9.4T MRI system at three timepoints after injection (4, 6, and 8 weeks). Previous studies suggest that the invasive cell fates and invasion patterns of G144 cells were different in different microenvironments of the brain (Brooks et al., 2021). In line with these observations, at the three timepoints, we used T2WI images to define hyperintense tumour bulk and peritumoural margin regions extending into the corpus callosum as a region of white matter, as well as cortex and striatum as regions of grey matter for quantitative analysis using MTE-NODDI.

Additionally, to ascertain whether MTE-NODDI provides improved sensitivity to infiltrating GBM cells, we also applied a standard diffusion tensor imaging (DTI) model to the diffusion data to calculate maps of MD. Figure 1 presents images of naïve and G144 tumour-bearing mice from a bespokeT2WI sequence, together with concomitant maps of NDI, f_en, ODI, and MD. The tumour bulk is readily identified as a hyperintense area on the T2WI image. The peritumoural margin regions (areas without T2 changes) surrounding the hyperintense bulk (Fig. 2) are located in the cortex, corpus callosum, and striatum.

A focus of this study was to investigate the conspicuity of the MTE-NODDI parameters to peritumoural margin regions beyond the tumour bulk, where no discernable changes were present in the T2WI images. At the 8-week timepoint, when the tumour is most developed, comparison of the MTE-NODDI parameters between the peritumoural margin and equivalent contralateral regions (cortex, corpus callosum, and striatum) revealed marked changes in NDI, f_en, and ODI values outside the tumour bulk (Fig. 3, Supp. Fig. 2).

In the cortex, all three MTE-NODDI parameters NDI (p < 0.05), f_en (p < 0.05) and ODI (p < 0.0005) demonstrated clear differences in the peritumoural margin; most notably in the ODI (Fig. 3A, Supp. Fig. 2A). Similarly, in the corpus callosum margin, NDI (p < 0.005), ODI (p < 0.01), and f_en (p < 0.0001) values demonstrated excellent sensitivity of MTE-NODDI to detect tumour margin in both grey and white matter regions outside the bulk (Fig. 3B, Supp. Fig. 2B). As in the corpus callosum and cortex, peritumoural margin ROIs in the striatum had a lower ODI (p < 0.0001), NDI (p < 0.005) and higher f_en (p < 0.001) values (Fig. 3C, Supp. Fig. 2C).

Gadolinium-enhanced T1 images, alongside T2 images, are used clinically to define the area of tumour bulk (Price & Gillard, 2011). Due to the limited time during the MRI acquisition protocol, we did not acquire post-contrast T1 images. Therefore, in a separate group of tumour-bearing animals (n = 2), we compared the measured tumour volumes in our model of GBM using both gadolinium-enhanced T1-weighted and the T2WI used to identify tumour bulk. The tumour volumes were comparable (mean percentage difference = 6.7%), supporting the use of T2WI to define tumour bulk in our model (Supp. Fig. 3).

When defining the equivalent control ROIs (bulk, peritumoural margins in the cortex, striatum, and corpus callosum) in the contralateral hemisphere, there could be slight variations in ROI placement due to the tumour mass in the ipsilateral hemisphere. As such, additional controls were performed in a group of age- and strain-matched naïve mice to assess the effect of the small variation in ipsi- and contralateral ROIs (n=3) placement, due to the tumour mass in the ipsilateral striatum, on the MRI data. The MRI images from the naïve mice were co-registered to the images from the tumour-bearing mice using FMRIB Software Library (FSL, University of Oxford, UK). Subsequently, the ROIs drawn on tumour-bearing mice were superimposed onto images from the naïve mice and mean values for NDI, f_en, ODI, and MD were extracted for quantitative analysis. We did not note any differences in NDI, f_en, and ODI values between ROIs due to the tumour (Supp. Fig. 4). Therefore, there is limited contribution to the peritumoural margin data from the regional background variation in the selected ROIs.

We next assessed the sensitivity of MTE-NODDI parameters to tumour-induced microstructure changes (NDI, f_en, and ODI) at all three imaging timepoints (4, 6, and 8 weeks) across three anatomical regions (Fig. 4A). MTE-NODDI parameters measured from ROIs drawn at the respective timepoints were clearly sensitive to tumour evolution over the 4-week imaging period. These peritumoural margin changes were most notable from the 6-week time point onwards. For instance, beginning at the 6-week timepoint, all MTE-NODDI parameters in the striatum exhibited sensitivity to tumour-induced pathology, with decreased NDI (p < 0.005), and ODI (p < 0.05), while the f_en (p < 0.005) was increased (Fig. 4Ai-iii).

Our subsequent aim was to evaluate whether MTE-NODDI was more sensitive to tumour-induced changes than the established DWI measure of MD. We observed that differences in NDI, f_en, and ODI values were consistently higher than MD values, indicating that MTE-NODDI offers increased sensitivity to microstructural changes in the peritumoural margin as the tumour invades normal brain tissue. More specifically, we detected increases in striatal ODI as early as week 4 without any associated changes in MD (Fig. 4B). This heightened sensitivity was also observed at 6 and 8 weeks in the cortex, as well as at 8 weeks in the corpus callosum, thereby providing strong evidence for the increased conspicuity of MTE-NODDI to tumour-induced microstructural changes in regions without MD or T2WI changes. Importantly we did not observe any changes (NDI, f_en, ODI, and MD) in control age-matched and strain-matched naïve animals (n=3) over the 3 timepoints (Supp. Fig. 5). Taken together, this provides evidence that MTE-NODDI parameters are sensitive to tumour-induced microstructural pathology as it invades into regions outside the tumour bulk.

Finally, we sought to examine the effectiveness of MTE-NODDI as an in vivo imaging marker to detect invading tumour cells in the tumour margin, by comparing fluorescence (% area) from GFP+ tumour cells measured using immunohistological images with the MRI parameters (Fig. 5). Importantly, the GFP+ cell distribution observed in the immunohistological images confirmed tumour cell infiltration into all three anatomical brain regions tested (cortex, corpus collosum, and striatum) (Fig. 5; top panel). MTE-NODDI values in the cortex show that both f_en (R² = 0.81 and p < 0.05) and NDI (R² = 0.76 and p < 0.05) demonstrated strong correlations with GFP+ fluorescence, with MD displaying no such correlation (R² = 0.21 and p = 0.44). As such, we observe improved specificity of the technique over MD for detecting tumour-induced pathology correlating with migrating tumour cells. To expand upon these findings, we analysed an additional 9 GFP+ immunohistological sections from the rostral edge of the tumour bulk. We did not find any correlations between GFP+ fluorescence and the MRI parameters. However, in this anterior region, the cell count was eight times lower (0.28%) compared to the central, more cellularly dense portion of the tumour (1.56%) (p = 0.05) (Supp. Fig. 6). We believe the lack of correlation is due to the scarcity of cells thus limiting sensitivity, as most likely beyond the limits of detection in the regions at the anterior boundary of the tumour.

Precise tumour margin delineation in GBM remains a profound challenge due to its highly invasive nature and diffuse pattern of tumour cell migration into healthy brain regions. Here, we report the first application of MTE-NODDI to non-invasively detect infiltrating cells in a patient-derived xenograft mouse model of GBM in regions outside the tumour bulk. We demonstrate that MTE-NODDI detected tumour-induced pathology in margin regions without T2 or MRI-diffusion changes. Furthermore, MTE-NODDI parameters correlated with immunohistological measure of tumour cell infiltration. These findings highlight the sensitivity of MTE-NODDI to invading tumour cells and associated pathology, suggesting its utility in imaging peritumoural margins of GBM.

Traditionally, NODDI was developed to provide microstructural information that is more specific than DTI by separating the diffusion signal into intra-neurite, extra-neurite, and free water fractions (Zhang et al., 2012). The NODDI model, generates two key variables, neurite density index (NDI) and orientation dispersion index (ODI). Additionally, free water fraction (f_iso) and extra-neurite volume fraction (f_en) can be derived when studying the influence of CSF and extra-neurite structures (like soma, glial cells, invading tumour cells etc.) respectively. The technique is suitable for both grey and white matter (Winston et al., 2014) and has been valuable in evaluating axonal/dendritic degeneration and fiber orientation in various neurological diseases, such as Alzheimer’s disease (Colgan et al., 2016; Fu et al., 2020; Parker et al., 2018), Parkinson’s disease (Kamagata et al., 2017; Mitchell et al., 2019), MS (Schneider et al., 2017; Spano et al., 2018), ischemic stroke (Mastropietro et al., 2019; Z. Wang et al., 2019), and epilepsy (Sone et al., 2018; Winston et al., 2014). Importantly, NODDI has also been successfully applied to brain tumours (Kamiya et al., 2020; Sanvito et al., 2021), primarily in grading (Figini et al., 2018; Zhao et al., 2018) but also to differentiate primary GBM tumour bulk from brain metastases (Kadota et al., 2020; Mao et al., 2020; Wen et al., 2015), as well as vasogenic oedema from tumour infiltration (Caverzasi et al., 2016; Chong et al., 2021; Masjoodi et al., 2018). However, there are no reports investigating the role of NODDI in detecting infiltrating tumour margin beyond regions with T1 or T2 abnormalities.

Outside the bulk of the tumour, in the hard-to-detect peritumoural margin regions, we observe marked changes in the MTE-NODDI parameters. Several studies have examined the association between mean diffusivity (MD), also known as the apparent diffusion coefficient (ADC), and tumour cellularity in GBM. Some studies have reported a decrease in MD with tumour progression and a negative correlation between MD and tumour cellularity in these regions (Chang et al., 2017; Chenevert et al., 2000; Kono et al., 2001; Sugahara et al., 1999), while others have found MD maps less useful or reported inverse findings (Berro et al., 2019; Price et al., 2006; Sadeghi et al., 2008). When we compared our measurements of MTE-NODDI changes to the more established MD, we observed that MTE-NODDI has a greater sensitivity to infiltrating tumour cells. In our mouse model, we observed an increase in MD values in the peritumoural margins, suggesting the influence of other factors such as peritumoural oedema (Lemercier et al., 2014). Given that the hyperintense regions in T2WI and contrast-enhancing regions in T1-weighted imaging were comparable, this finding suggests that the model does not exhibit oedema. The absence of oedema is reinforced by the lack of hyperintense signals on T2WI outside the boundaries of the contrast-enhancing region. Our findings suggest that infiltrating tumour cells, perhaps coupled with microstructural changes induced by cell invasion, can be detected by MTE-NODDI. As such, MTE-NODDI with its improved sensitivity may provide a tool for detecting tumour invasion into the peritumoural margin regions in GBM.

In terms of tumour growth, MTE-NODDI parameters readily detected tumour evolution (4–8 weeks) as the tumour cells infiltrate out from the bulk into the peritumoural margins. NODDI has been previously used to investigate the relationship between disease progression and WM-changes in traumatic brain injury (Palacios et al., 2020) stroke (Mastropietro et al., 2019), athletic concussion (Churchill et al., 2019), and schizophrenia (Kraguljac et al., 2021). Taken together with our data, we anticipate that MTE-NODDI could be valuable in evaluating longitudinal microstructural changes in the brain induced by infiltrating tumour cells.

Previous studies have noted increased ODI in the tumour bulk, likely due to disruption of neuronal structural integrity (Masjoodi et al., 2018; Onishi et al., 2022). Conversely, we observe a decrease in ODI in the margin regions, which may suggest increased microstructural directionality due to invading GBM tumour cells. Recently, invasion in GBM has been described as having multiple traits of neuronal development, including formation of neurogliomal synapses that simulate ‘neurite-like’ tumour microtubes (Venkataramani et al., 2019), alongside increases in solid stress and compression (Seano et al., 2019), which may lead to a decrease in dispersion of neurites and consequently lower ODI. Thus, it is conceivable, that in the peritumoural margin regions, tumour cell infiltration may contribute to the observed decrease in ODI values.

Our observed increase in f_en values may be driven by infiltrating tumour cells within the extra-neurite spaces in the peritumoural margin. When patient-derived glioma cells are implanted into rodent brains, a substantial amount of these cells migrate along the vascular surfaces to invade the healthy brain (Brooks & Parrinello, 2017; Cuddapah et al., 2014). These infiltrating tumour cells can be associated with the NODDI extra-neurite compartment (f_en), in the same fashion as microglia, glial cells, and ependymal cells (Caverzasi et al., 2016; Kadota et al., 2020). Previously, it has been demonstrated that microglial density is a key contributor to measures of NODDI extra-neurite compartment (f_en) (Yi et al., 2019). Similarly, as GBM cells infiltrate into the peritumoural margins with tumour development, there is a corresponding increase in f_en values supporting the contribution of GBM cells to an increase in the measured extra-neurite compartment index (f_en).

We also observed lower NDI in the peritumoural regions, which may be due to a relative decrease in intra-neurite volume fraction caused by infiltrating tumour cells in the extra-neurite spaces, causing a shift in diffusion property of the tissue from a restricted intra-neurite compartment towards a hindered extra-neurite compartment (Kadota et al., 2020; Yi et al., 2019). Therefore, as tumour cells invade into the margin, this may reduce the intra-neurite volume fraction leading to a decrease NDI. Previous studies featuring use of NODDI in the glioma bulk (Jin et al., 2020; Kadota et al., 2020; Onishi et al., 2022; Zhao et al., 2018) have shown a decreased NDI, suggesting possible neuronal loss and fibre disruption, yet this may also reflect the increased cellular density in the tumour bulk, as observed in the margin of our study.

These observations were further explored by investigating the relationship between MTE-NODDI parameters and infiltrating tumour cells (GFP+ fluorescence). In our tumour model, f_en and NDI correlated well with GFP+ fluorescence. NODDI has been previously used to investigate microstructural alterations in an Alzheimer’s mouse model of tau pathology, where NDI values correlated with disease burden (Colgan et al., 2016). Our results indicate that infiltrating tumour cells may account for the changes in MTE-NODDI parameters, although this does not exclude additional tumour-induced microstructural changes, such as pathological demyelination (Brooks et al., 2021; J. Wang et al., 2018), and water retention in the surrounding tissue (Onishi et al., 2022; Wen et al., 2015).

As this was a proof-of-concept study, there is an opportunity for building on the experimental design. The potential presence of tumour cells within the contralateral control hemisphere may result in an underestimation of the observed differences between the peritumoural and contralateral control ROIs. Incorporating a baseline scan to establish control values before the onset of disease may contribute to the ability of MTE-NODDI to detect peritumoural invasion. There is also an opportunity to further disentangle the observed differences in MTE-NODDI metrics resulting from infiltrating tumour cells and those caused due to pathological microstructural changes such as demyelination, angiogenesis, alterations to the extra cellular matrix, and other deformations in the brain due to tumour-growth (Brooks et al., 2022; Lathia et al., 2011; Seano et al., 2019). Furthermore, a multi-timepoint acquisition of MRI data with matched multi-slice histology at each timepoint would enable a comprehensive understanding of the specificity of MTE-NODDI metrics to infiltrating tumour cells across the entire brain. Investigating the histological images from the anterior portion of the tumour did not yield significant findings, which we believe is likely due to the limited number of invading tumour cells in that region (8 times lower compared to the central core of the tumour). In addition, the relatively small size of the tumour compared to the MRI slice thickness may result in small or subtle areas of tumour infiltration to be missed or averaged out in the imaging data, reducing the ability to precisely correlate the MRI findings with the histological data. Future studies should consider using reduced imaging slice thickness and multiple histological sections spanning the entire tumour length to further establish specificity of the technique for detecting migrating tumour cells. Additionally, establishing direct co-registration between MRI images and histological sections could facilitate a pixel-by-pixel analysis of the MRI signal, which was not possible in this study. Moreover, in the present study, we define the tumour bulk and peritumoural margin regions based on T2-weighted images. To better align with imaging protocols commonly employed in clinical settings, a multi-parametric acquisition that includes contrast-enhanced T1-weighted images and FLAIR imaging for ROI delineation, along with T2 mapping for quantitative analysis, would be more suitable. In terms of other methods, the ongoing development of imaging techniques (Laack et al., 2021; Laprie et al., 2024; Ramesh et al., 2022) suggests that a multiparametric imaging approach may contribute to understanding this complex and challenging brain tumour type.

In conclusion, this study establishes the clear value of MTE-NODDI in detecting infiltrating GBM cells in the peritumoural margin outside regions of T2WI and MD abnormalities. MTE-NODDI parameters NDI, f_en, and ODI all change markedly in areas with tumour cell infiltration and associated microstructural alterations. Moreover, the correlations between immunohistological measure of tumour cells and MTE-NODDI parameters provide evidence of sensitivity of MTE-NODDI to low numbers of infiltrating tumour cells. Mounting evidence shows improved survival with extending resection into the non-contrast enhancing regions (De Leeuw & Vogelbaum, 2019; Jackson et al., 2020). Therefore, findings from this study suggest that MTE-NODDI could be used to improve tumour margin detection and aid neurosurgeons in maximising the extent of surgical resection to achieve better survival and quality-of-life outcomes.

The data generated in this study are available upon reasonable request from the corresponding author.

S.R.K.: Data curation, Formal analysis, Investigation, Methodology, Visualisation, and Writing—original draft. D.G.: Data curation, Formal analysis, Investigation, Visualisation, and Writing—review & editing. D.T.: Data curation, Methodology, Software, and Writing—review & editing. L.J.B.: Methodology, Writing—review & editing. T.G.: Methodology, Validation, and Software. A.C.S.J.: Formal analysis, Visualisation, and Writing—review & editing. C.S.H.: Formal analysis, Writing—review & editing. M.C.: Formal analysis, Writing—review & editing. T.L.K.: Formal analysis, Writing—review & editing. J.A.W.: Methodology, Writing—review & editing. D.J.S.: Formal analysis, Writing—review & editing. L. T.: Formal analysis, Writing—review & editing. V.S.: Formal analysis, Supervision, Funding acquisition, and Writing—review & editing. S.M.: Formal analysis, Supervision, Funding acquisition, and Writing—review & editing. H.Z.: Metholodology, Validation, and Software. S.P.: Conceptualisation, Formal analysis, Supervision, Funding acquisition, and Writing—review & editing. M.F.L.: Conceptualisation, Formal analysis, Supervision, Funding acquisition, and Writing—original draft.

The authors declare that they have no competing interests.

S.R.K. was supported by The Edinburgh-UCL Cancer Research UK (CRUK) Brain Tumour Centre of Excellence award [C7893/A27590]. This work was supported in part by The Oli Hilsdon Foundation through The Brain Tumour Charity, grant number (GN-000595), in connection with the program “Mapping the spatio-temporal heterogeneity of glioblastoma invasion” (D.G., M.C., A.C.S.J.). D.T.’s salary was supported by CRUK City of London (CoL) Centre. A.C.S.J. is supported by the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, a Schmidt Futures program. L.J.B. was funded from Cancer Research UK (CRUK) RadNet in the form of an early career fellowship. C.S.H. was supported by a CRUK Pioneer Award (C70568/A29787), an AMS Starter Grant (SGL021\1034), a National Brain Appeal Innovation Award (NBA/NSG/BTB), and UCLH BRC NIHR funding. J.A.W. is supported by the Wellcome Trust (225345/Z/22/Z). VS was supported by the Engineering and Physical Sciences Research Council (EP/N014529/1). S.M. was supported by the UK Medical Research Council. S.P. was supported by CRUK (awards C55501/A21203 and 7550844). M.F.L. was supported by the Rosetrees Trust and the John Black Charitable Foundation (grant no. A2200). The authors would like to thank Shereen Nizari for their help in data acquisition.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00472.