Hybrid Deep Feature Extraction and mRMR-Based Optimization-Assisted Machine Learning Approaches in MRI-Based Brain Tumor Classification

Abstract

Brain tumors are among the important neurological diseases that require early and accurate diagnosis due to their high mortality rates and serious clinical outcomes. Although Magnetic Resonance Imaging (MRI) is widely used in the evaluation of brain tumors, structural and tissue similarities between different tumor types can complicate the diagnostic process. This increases the need for computer-aided decision systems. This study conducted a comparative analysis of machine learning approaches supported by mRMR-based feature selection and hyperparameter optimization for determining the optimal number of features in MRI-based brain tumor classification. Features extracted from MRI images belonging to glioma, meningioma, no tumor, and pituitary tumor classes were used in the study. The baseline and optimized performances of KNN, SVM, RF, XGBoost, and LightGBM algorithms were evaluated based on accuracy, precision, recall, F1-Score, and runtime criteria. After optimization, the accuracy of the KNN algorithm increased from 81.10% to 88.13%, while the accuracy of the SVM algorithm increased from 88.57% to 92.75%. The SVM classifier exhibited greater performance relative to other approaches, attaining the best accuracy of 92.75% and F1-Score of 92.73% on the test set.