Abstract
Recent investigations have shown that diffusion tensor magnetic resonance imaging (DT-MRI) and magnetic resonance spectroscopic imaging (MRSI) provide useful information in diagnosis brain tumor glioma. Therefore, various software have been developed and patented which enable users to overlay multimodal MRI such as DTI-MRI and MRSI for more accurate localization of abnormal lesions. The purpose of this work was to further investigate the potential of these modalities in diagnosis of brain glioma tumors. This paper integrates DTI and MRSI features for differentiating high grade from low grade brain glioma tumors. Three DTI and two MRSI features are extracted from twelve histologically proven brain glioma patients in pre-treatment status. Artificial Neural Network (ANN) classifiers are developed to estimate tumor grades by optimizing Receiver Operating Characteristic (ROC) curves. Our study shows that all DTI and MRSI features are statistically significantly different (P < 0.05) in high grade versus low grade tumors. Low grade tumors have higher mean diffusivity (MD) (1.43±0.21) and lower fractional anisotropy (FA) (0.14±0.04) and relative anisotropy (RA) (0.11±0.03) compared with high grade tumors (1.29±0.47, 0.18±0.12, 0.15±0.10, respectively). Also, Cho/Cr and Cho/NAA ratios of high grade tumors (2.27±1.24, 2.16±1.96) are higher than those of low grade tumors (1.29±0.44, 0.88±0.31). The proposed ANN classifier using FA, RA, Cho/Cr, and Cho/NAA features generates the largest volume under the ROC (the highest accuracy), illustrating that the proposed integration of DTI and MRSI features improves accuracy of tumor grade estimation.
Keywords: Artificial neural networks, brain glioma, diffusion tensor magnetic resonance imaging, magnetic resonance spectroscopic imaging, pattern recognition, tumor grade.
20