Brain Tumor Classification from MRI Using InceptionV3 with Squeeze-and-Excitation Attention and Multi-Scale Pooling

Abstract

Brain tumor classification from Magnetic Resonance Imaging (MRI) is a clinically critical yet challenging task due to substantial intra-class variability and morphological similarity between tumor types. This paper presents IV3-SE-BTCNet, a novel hybrid deep learning architecture that integrates InceptionV3 transfer learning with Squeeze-and-Excitation (SE) attention blocks and multi-scale pooling for robust four-class brain tumor classification (Glioma, Meningioma, Pituitary, and No Tumor). A comprehensive 10-step domain-specific preprocessing pipeline — including ROI cropping, skull stripping, CLAHE contrast enhancement, hybrid denoising, and histogram equalization — is applied to a balanced 7,200-image MRI dataset. The model is trained using a two-phase strategy: Phase 1 (frozen backbone feature extraction, 80 epochs) followed by Phase 2 (selective fine-tuning, 80 epochs). The proposed model achieves a test accuracy of 96.76%, precision of 96.80%, recall of 96.76%, and F1-score of 96.77%, substantially outperforming CNN baselines reported in comparative literature: Custom CNN (92.72%), MobileNetV2 (89.12%), VGG16 (82.28%), and EfficientNetB0 (27.45%). AUC scores exceed 0.993 across all four tumor classes on the test set. These results demonstrate that attention-augmented transfer learning combined with domain-specific preprocessing provides a highly effective and reliable decision-support tool for radiological brain tumor diagnosis.