Preliminary Radiomic Analysis on Quantitative Parameters’ maps derived from DTI Images of Glioma patients

Gliomas are the most common type of Central Nervous System (CNS) tumours, making up to 60% of all brain tumours. They are classified and graded according to the 2021 WHO CNS5 classification: the combined histological and molecular grading for gliomas ranges from grade 1 to grade 4, with very different survival probabilities. The diagnosis is made through histopathology evaluation of tissue samples collected during surgical tumour resection, which is the primary therapeutic approach for glioma patients. The aim of this thesis is to perform a preliminary radiomics analysis on Diffusion Tensor Imaging (DTI) parameters’ maps, in particular on Fractional Anysotropy (FA), Mean Diffusivity (MD), Mode of the anisotropy (MO), eigenvalues (L1, L2, L3) and no-diffusion weighting (S0) maps. The dataset consisted of 34 patients’ data, including conventional Magnetic Resonance (MR) images, DTI images and biopsy data. Radiomics analysis consisted in data preprocessing, tumour segmentation, features’ extraction, features’ selection, model implementation and model validation. Data preprocessing included brain extraction, image coregistration and DTI parameters’ evaluation using FSL software. Tumour Regions Of Interest (ROIs) segmentation was performed on conventional MR images, using semi-automatic segmentation tools on 3D Slicer software, and the ROIs were validated by an experienced neuroradiologist. Radiomic features (RFs) were extracted from tumour’s ROIs on all DTI parameters’ maps, using a home-made Python program. RFs’ selection consisted in: constant RFs removal, features’ clustering based on Pearson’s standard correlation coefficient, dendrogram evaluation and significant RFs’ selection. Significant RFs were selected by performing a T-test between classification groups’ mean RF values, and the feature with lower p-value was considered to be the most significant. Three types of classification were tested: (1) primary vs secondary lesion; (2) glioma vs other type of CNS primary tumour; (3) grade 4 glioblastoma vs other lower-grade glioma. For each significant RF, the Receiver Operating Characteristic (ROC) curve was evaluated and the quality of the classifier was estimated with the Area Under the Curve (AUC) value. The predictive model was implemented by assembling the results from all significant RF’s classifications, choosing the best thresholds from the ROC curves, and model performance was evaluated with internal and external validation. A total of 742 RFs were extracted, 106 for each DTI parameter’s map, and of these 46 were constant. Features’ clustering and selection resulted in 13 final features: 3 RFs significant for classification (1), 5 RFs significant for classification (2) and 5 RFs significant for classification (3). ROC curves showed moderate and high classification ability of the RFs, with AUC values ranging from 0.8 to 0.96. For the three final classifications, made by combining all RFs’ classifiers, ROC curves showed high performance, with AUC values of 1.0, 0.98 and 0.96, respectively. The model showed good performance in the internal validation, correcly classifying 93% of glioblastomas, 67% of lower-grade gliomas, 100% of other primary tumours and 100% of metastasis. External validation showed good performance on glioblastomas (all glioblastomas were correctly classified) but bad performance on lower-grade gliomas (the one lower-grade patient was misclassified as glioblastoma). In conclusion, the preliminary radiomics analysis highlighted 13 significant RFs for glioma patients’ classification. A model on these RFs was implemented and internally validated, while external validation showed uncertain results. Overall, the analysis showed promising results and paved the way to further investigate the predictivity of DTI radiomics on glioma patients.