Classification of Brain MRI Images using Deep Learning: The DeiT3 Model and the Use of Feature Fusion Methods

Brain tumors are among the diseases that can seriously threaten human life and can be fatal. Early diagnosis of brain tumors plays a crucial role in the treatment process of the disease. However, accurately and quickly diagnosing this disease remains one of the significant challenges of modern medical technologies. Currently, advanced imaging techniques such as magnetic resonance imaging (MRI) are generally used for detecting brain tumors. This study proposes an artificial intelligence-based diagnostic approach using MRI images that include brain tumor types and consist of four classes. The proposed approach includes preprocessing, model training, feature fusion, and selection as final steps. In the preprocessing step, Grad-CAM and LBP techniques are applied to the original dataset, resulting in a total of three datasets, including the original one. These datasets are then trained with the DeiT3 model to obtain three separate feature sets (original, Grad-CAM-based, LBP-based). The feature sets are fused using a feature fusion technique, and the performance of the combined sets is evaluated using SVM methods. Feature selection methods (Chi2, Relief) are applied to the best-performing Grad-CAM & LBP-based feature set to highlight the most efficient features. Experimental analysis results show that a success rate of 99.5% was achieved using the SVM method.