Deep Convolutional Radiomic Features on Diffusion Tensor Images for Classification of Glioma Grades

Area of Science:

• Neuroimaging
• Radiology
• Oncology

Background:

• Glioma grading is crucial for treatment and prognosis.
• Accurate grading aids in determining optimal patient management strategies.

Purpose of the Study:

• To investigate novel radiomic features from fractional anisotropy (FA) and mean diffusivity (MD) maps of brain diffusion tensor imaging (DTI) for computer-aided glioma grading.
• To assess the efficacy of these features in classifying low-grade gliomas (LGGs) versus high-grade gliomas (HGGs) and differentiating between grade III and IV gliomas.

Main Methods:

• Retrospective analysis of 108 patients with pathologically confirmed gliomas who underwent DTI.
• Extraction of radiomic features (texture, morphological, deep features from CNNs) from manually delineated tumor regions on FA and MD maps.
• Classification using support vector machines with leave-one-out cross-validation.

Main Results:

• Combined FA+MD features achieved high performance: AUC=0.93, accuracy=0.94 for LGG vs HGG; AUC=0.99, accuracy=0.98 for grade III vs IV.
• Deep radiomic features demonstrated superior prediction ability compared to traditional features.
• Performance metrics remained robust when considering different tumor subregions.

Conclusions:

• Radiomic features from FA and MD maps in DTI are effective for noninvasive glioma classification and grading.
• This approach offers a valuable tool for distinguishing between LGGs and HGGs, and between high-grade gliomas.
• Deep learning-based radiomics shows promise for improving diagnostic accuracy in glioma grading.