Glioma Grade Classification Using Machine Learning and MRI Radiomics: A Single-Center Prospective Study Comparing Original and Wavelet-Transformed Features From Anatomical, Diffusion-Weighted, and Post-Contrast Imaging

Abstract

Background and Aims Accurate glioma grade classification is critical for prognostic assessment and clinical decision-making. This study aimed to evaluate the impact of wavelet-based radiomics feature analysis on the performance of machine learning (ML) models for glioma grade classification using diffusion-weighted imaging (DWI), structural MRI sequences, and contrast-enhanced T1-weighted (T1Gd) images.Methods This prospective study included 92 patients with histopathologically confirmed gliomas. MRI examinations were performed using a 1.5 T scanner and included apparent diffusion coefficient (ADC) and exponential ADC (eADC) maps, T1-weighted, T2-FLAIR, and T1Gd sequences. Image preprocessing consisted of deep learning-based autoencoder denoising and intensity normalization using the Removal of Artificial Voxel Effect by Linear regression (RAVEL) method. Tumor volumes of interest were manually segmented by two experienced neuroradiologists. Radiomics features were extracted from original images and wavelet-transformed images by Pyradiomics, followed by feature selection using the least absolute shrinkage and selection operator (LASSO). Glioma grade classification was performed using fine-tuned Decision Tree, Logistic Regression, K-nearest neighbors, Na & iuml;ve Bayes, Support Vector Machine, Multi-layer Perceptron, and Random Forest models, with hyperparameters tuned via fivefold cross-validation on an 80 training subset and performance assessed on an independent 20 held-out test set.Results The addition of wavelet-based radiomics features significantly improved classification performance across all ML models and MRI sequences. The highest performance was achieved by a fine-tuned Random Forest model trained with original and wavelet-transformed images' radiomics features extracted from eADC maps, with an average AUC of 0.96, accuracy of 0.97, sensitivity of 0.94, and specificity of 0.91.Conclusion Adding wavelet-transformed radiomics feature to original image radiomics feature analysis of eADC maps combined with fine-tuned Random Forest classification provides an accurate, non-invasive, and efficient approach for glioma grade classification with strong potential for clinical application.