Attention-Guided Swin Transformer with CNN Feature Extraction for Brain Tumor Detection in MRI

Abstract

Brain tumors represent one of the most lethal forms of intracranial malignancy, necessitating precise and timely diagnosis for effective clinical intervention. Magnetic Resonance Imaging (MRI) serves as the primary neuroimaging modality for tumor identification; however, manual interpretation remains susceptible to inter-observer variability and diagnostic delays. This paper proposes an Attention-Guided Swin Transformer framework integrated with CNN-based feature extraction for automated, multi-class brain tumor classification from MRI scans. The proposed methodology leverages EfficientNet-B3 as a backbone for local feature extraction, augmented by the Convolutional Block Attention Module (CBAM) for channel and spatial attention refinement. The refined feature maps are subsequently partitioned into non-overlapping patches and processed through a hierarchical Swin Transformer employing shifted window self-attention, enabling multi-scale global context modeling. The classification head comprises global average pooling, fully connected dense layers, and dropout regularization, culminating in a Softmax output layer for four-class classification: No Tumor, Glioma, Meningioma, and Pituitary. Experiments are conducted on the BraTS MRI dataset with standard preprocessing including skull stripping, normalization, resizing to 224×224, and data augmentation. The proposed model achieves 97.8% accuracy, outperforming baseline methods including standard CNN (88.4%), ResNet50 (91.2%), EfficientNet-B3 (93.7%), ViT (94.5%), Swin Transformer (95.9%), and CNN+CBAM (94.8%). Ablation studies confirm the complementary contributions of each architectural component. These results demonstrate the clinical applicability of the proposed hybrid framework for early and reliable brain tumor diagnosis.