ABSTRACT
Purpose To use T1-, T2-weighted and diffusion tensor MR images to portray glioma grade by employing a voxel-wise supervised machine learning approach, and to assess the feasibility of this tool in preoperative tumor characterization.
Materials and Methods Conventional MRI, DTI datasets and histopathological evaluations of 40 patients with WHO grade II-IV gliomas were retrospectively analyzed. Databases were construed incorporating preoperative images, tumor delineation and grades. This data was used to train a multilayer perceptron based artificial neural network that performed voxel-by-voxel correlation of tumor grade and the feature vector. Results were mapped to grayscale images, whereas grade map was defined as a composite image that depicts grade assignments for intra-tumoral regions. The voxel-wise probability for high grade tumor classification was calculated for the entire tumor volumes, defined as the grade index.
Results The color hue on glioma grade maps allowed the discrimination of low and high grade cases. This method revealed connection between the heterogeneous appearance of tumors and the histopathological findings. Classification by the grade index had 92.31% specificity, 85.71% sensitivity.
Conclusion Glioma grade maps are advantageous in the visualization of the heterogeneous nature of intra-tumoral diffusion and relaxivity and can further enhance the characterization of tumors by providing a preoperative modality that expands information available for clinicians.
- ADC
- apparent diffusion coefficient;
- ANN
- artificial neural networks;
- DWI
- diffusion weighted imaging;
- DTI
- diffusion tensor imaging;
- FA
- fractional anisotropy;
- HGPM
- high grade probability map;
- LGPM
- low grade probability map;
- TPM
- tumor probability map;
- WHO
- World Health Organization
