Glioma grade assessment by using histogram analysis of diffusion tensor imaging-derived maps

Abstract

Introduction

Current endeavors in neuro-oncology include morphological validation of imaging methods by histology, including molecular and immunohistochemical techniques. Diffusion tensor imaging (DTI) is an up-to-date methodology of intracranial diagnostics that has gained importance in studies of neoplasia. Our aim was to assess the feasibility of discriminant analysis applied to histograms of preoperative diffusion tensor imaging-derived images for the prediction of glioma grade validated by histomorphology.

Methods

Tumors of 40 consecutive patients included 13 grade II astrocytomas, seven oligoastrocytomas, six grade II oligodendrogliomas, three grade III oligoastrocytomas, and 11 glioblastoma multiformes. Preoperative DTI data comprised: unweighted (B ₀) images, fractional anisotropy, longitudinal and radial diffusivity maps, directionally averaged diffusion-weighted imaging, and trace images. Sampling consisted of generating histograms for gross tumor volumes; 25 histogram bins per scalar map were calculated. The histogram bins that allowed the most precise determination of low-grade (LG) or high-grade (HG) classification were selected by multivariate discriminant analysis. Accuracy of the model was defined by the success rate of the leave-one-out cross-validation.

Results

Statistical descriptors of voxel value distribution did not differ between LG and HG tumors and did not allow classification. The histogram model had 88.5% specificity and 85.7% sensitivity in the separation of LG and HG gliomas; specificity was improved when cases with oligodendroglial components were omitted.

Conclusion

Constructing histograms of preoperative radiological images over the tumor volume allows representation of the grade and enables discrimination of LG and HG gliomas which has been confirmed by histopathology.

Appendix

Mathematical equations used for generating scalar maps

$$ {\hbox{FA}} = \sqrt {{\frac{2}{3}}} \; \times \;\sqrt {{\frac{{{{\left( {{\lambda_1} - \overline \lambda } \right)}^2} + {{\left( {{\lambda_2} - \overline \lambda } \right)}^2} + {{\left( {{\lambda_3} - \overline \lambda } \right)}^2}}}{{\lambda_1^2 + \lambda_2^2 + \lambda_3^2}}}} $$

(1)

$$ {\lambda_\parallel } = {\lambda_1} $$

(2)

$$ \lambda \bot = {{{\left( {{\lambda_2} + {\lambda_3}} \right)}} \left/ {2} \right.} $$

(3)

$$ {\hbox{Trace}} = {\lambda_1} + {\lambda_2} + {\lambda_3} $$

(4)

where λ ₁, λ ₂, and λ ₃ are the three eigenvalues of the diffusion tensor and $\overline \lambda$ stands for the mean of the three eigenvalues.

$$ {\hbox{S}}{{\hbox{I}}_{\rm{DWI}}} = {\hbox{S}}{{\hbox{I}}_{\rm{T2}}}\; \times \;{{\hbox{e}}^{ - \left( { - b\; \times \;{\rm{ADC}}} \right)}} $$

(5)

where SI_DWI is the signal intensity of the voxels on the DWI images, SI_T2 means the signal intensity on T2 acquisitions (b ₀ images). ADC is the apparent diffusion coefficient while b is a factor reflecting the strength of diffusion weighting.