Fusion of 3D Convolutional Neural Network and Multifractal Features for High- and Low-grade Glioma Classification using Diffusion Tensor Imaging

The classification of gliomas is essential for the accurate diagnosis and effective treatment planning of brain tumors, particularly in distinguishing between low-grade gliomas (LGG) and high-grade gliomas (HGG). Conventional magnetic resonance imaging (MRI) based classification, although common, often relies on radiological expertise and can miss microstructural changes in the white matter due to the presence of a tumor. This study proposes a novel approach by integrating the potential of 3D Convolutional Neural Network (CNN) with multifractal features derived from Diffusion Tensor Imaging (DTI) to improve the classification accuracy of LGG and HGG patients. MRI scans from 86 patients (n= 39 LGG and n=47 HGG) were processed to generate Fractional Anisotropy (FA), Axial Diffusivity (AD), Radial Diffusivity (RD), and Mean Diffusivity (MD) 3D Nifti images of tumor and normal-appearing white matter regions. Multifractal features were extracted and fused with 3D CNN-derived features from the DTI maps of these regions. The 3D CNN model achieved notable improvements in classification accuracy, sensitivity, specificity and area under the receiving operating characteristics score after integrating multifractal features. 3D CNN-derived features from FA and MD images combined with its multifractal features demonstrated the highest classification accuracy (94 % for FA and 92% for MD of tumor region). Additionally, a novel approach using principal component analyses was developed to explain the separability of LGG and HGG classes before and after integrating multifractal features with 3D CNN-derived features. Our findings suggest that incorporating multifractal features of DTI in a 3D CNN framework offers a robust approach for glioma classification, providing a detailed characterization of tumor impact on brain white matter tissue.