Bridging the Diagnostic Gap: Classification of MRI-Based Brain Tumors Using a CNN and Transformer-Based Hybrid Deep Learning Method

Accurate classification of brain tumors in Magnetic Resonance Imaging (MRI) scans is pivotal for timely clinical intervention, yet remains challenging due to intricate anatomical variations and overlapping imaging features. This research addresses this gap by proposing a hybrid deep learning architecture that synergizes 2D Convolutional Neural Networks (CNNs), 3D CNNs, and Transformer-Based attention mechanisms. The framework leverages pre-trained 2D CNN models—MobileNetV2 and EfficientNetV2B0—to extract spatial hierarchies from axial MRI slices, while a 3D CNN captures volumetric dependencies inherent in multi-dimensional MRI data. A Transformer block, equipped with multi-head self-attention, further refines feature representation by modeling long-range contextual relationships across image regions. The model is trained on a publicly available MRI dataset comprising four tumor classes (glioma, meningioma, pituitary, and non-tumorous), with preprocessing steps including intensity normalization, resizing, and augmentation (rotation, flipping, shear) to bolster generalization. Training optimization integrates ReduceLROnPlateau (learning rate reduction upon stagnated validation loss) and EarlyStopping (halting training at performance plateau) strategies to ensure efficient convergence. The hybrid architecture demonstrates robust diagnostic capability, achieving state-of-the-art classification performance, thereby underscoring its potential as an automated, reliable tool for multi-class brain tumor identification in clinical oncology workflows.