Comparative analysis of diffusion models for glioma malignancy grade differentiation
Ivan I. Maximov1, Aram Tonoyan2, Daniel Edelhoff1, Igor Pronin2, and Dieter Suter1

1TU Dortmund University, Dortmund, Germany, 2Burdenko Neurosurgery Institute, Moscow, Russian Federation


Diffusion weighted imaging is very powerful technique allows one to diagnose and differentiate glioma malignancy in vivo. Recently, the non-Gaussian diffusion models exhibited their efficacy in a detection of microsctructure changes for low and high grade gliomas. In this study we perform a comparative analysis of four diffusion models (DTI, DKI, two- and three-compartment models) in ability to provide a quantitative measure of the gliomas malignancy in vivo.


We provide a quantitative comparison of characteristic diffusion metrics obtained by four diffusion models (DTI, DKI, two- and three-compartment models of diffusion) in ability to distinguish the glioma malignancy grades II, III, and IV from in vivo datasets.


A group of 15 patients with different supratentorial gliomas: 5 subjects with grade II, 5 subjects with grade III, and 5 subjects with grade IV, underwent the diffusion weighted measurements (b = 1000, 2500 s/mm2, 60 diffusion directions, 3 mm3 isotropic resolution). The used workflow is represented in Fig.1. Initially, all datasets were corrected for subject motion and eddy current distortions using FSL. Appropriate b-matrix rotations were performed. The noise was corrected using a χ2 distribution function1. All raw datasets were interpolated using the cubic spline method and anisotropic diffusion filtering up to 1.5 mm3 isotropic resolution2. The DTI, DKI, and two-compartment model metrics such as fractional anisotropy (FA), kurtosis anisotropy3 (KA), apparent water fraction4 (AWF), and tortuosity4 (TORT) were evaluated using ExploreDTI5 and the linear weighted estimator. Neurite orientation dispersion and density imaging (NODDI) metrics such as orientation dispersion index (ODI) and intra-axonal volume fraction (TD) were estimated in two stage: at the beginning the whole brain maps were evaluated6. Next the new masks (see Fig.1) were created in regions with the suspicious results of an optimisation (the values equal to 0.704833 in ODI maps), and a two-step optimisation was rerun for these regions. This improvement allows us to obtain corrected values of ODI/TD in the area of tumour and edema. For the statistical analysis we used only highest glioma grade regions in each subject chosen by hand by a trained radiologist.


In Figure 2 we present an example of three patients with different gliomas and corresponding masks. One can see histograms of estimated metrics (normalised and averaged over the groups) FA, KA, AWF, TORT, TD, and ODI. In Figure 3 we presented the scatter plots of different diffusion metrics.


Diffusion-weighted imaging provides useful information about microstructure difference between healthy and diseased brain tissue. At the moment, a clinical “gold standard” such as DTI and the recently popular, DKI metrics are used as the biomarkers in a differentiation of glioma grades7. However, as we have shown (see Figs. 2,3) the diffusion models based on two- and three-compartment models allow one to get better contrast and richer information about the structure changes in tumour tissue depending on the glioma grade. In particular, the NODDI metrics which originally were developed for a healthy tissue exhibit a growing contribution of extra compartment based on cancer cells to the diffusion signal. It can be treated as a very sensitive biomarker of glioma malignancy and provides a good signaling metrics in glioma grade analysis.


In conclusion we demonstrated the feasibility of using diffusion models based on compartmentalization of brain tissue7.8. We have shown that ODI, TD, and AWF maps provided unique contrasts within regions with highest glioma malignancy and allow one to make a separation within low- and high glioma grades, including splitting inside of high-grade gliomas.


IIM and DS thank DFG (SU 192/32-1) for a partial support.


1. Aja-Fernandez et al., Magn. Reson. Med. 65 (2011) 1195.

2. Yushkevich et al., NeuroImage 31 (2006) 1116.

3. Poot et al., IEEE Trans. MI 29 (2010) 819.

4. Fieremans et al., NeuroImage 58 (2011) 177.

5. Leemans et al., Proc ISMRM 17 (2009) 3537.

6. Zhang et al., NeuroImage 61 (2012) 1000.

7. Van Cauter et al., Radiology 263 (2012) 492.

8. Wen et al., NeuroImage: Clinical 9 (2015) 291.


Schematic workflow. The raw datasets (3 mm3) were noise/motion corrected, resampled (1.5 mm3), and normalised using the anisotropic diffusion filter. The conventional diffusion metrics such as fractional anisotropy (FA), kurtosis anisotropy (KA), and two compartment model metrics: aparent water fraction (AWF), and tortuosity (TORT) were evaluated by ExploreDTI. The NODDI parameters such as orientation dispersion index (ODI) and tract density (TD) were evaluated in two step optimisation.

Example of highest malignancy grade masking using the T2 weighted images with corresponding location of the contralateral normally appearing white matter. The histograms presented the frequency distribution of normalised values for the different diffusion metrics with corresponding standard deviations.

Scatter plots of the different diffusion metrics localised by masking over highest glioma grades in each subject.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)