Feature Fusion Hybrid Deep Learning Model for Brain Tumor Classification from MRI Scans

—Brain tumors pose a critical health threat, requiringaccurate classification from magnetic resonance imaging (MRI)scans for effective diagnosis and treatment. Manual methodsare error-prone and inefficient, while deep learning modelsoften rely on large, augmented datasets and lack interpretability. This study proposes a Feature Fusion Hy- bred DeepLearning Model, fusing InceptionV3 (multi-scale extraction) andXception (efficient convolutions) features at the convolutionallayer, surpassing traditional transfer learning and decision-levelensembles. Evaluated on the PMRAM Bangladeshi Brain CancerMRI Dataset (1505 images across four classes: pituitary tumor(373), glioma (373), meningioma (363), normal brain (396)),using preprocessing (resizing to 299×299, normalization, trainingtime augmentation), focal loss, and Test-Time Augmentation, themodel achieves 99.34% accuracy and 0.99 macro/weighted precision/recall/F1, out- performing single models (92.23–98.27%)and decision-level ensembles (98.00–98.14%), with class-wise F1scores of 0.96–1.00. Distinct from prior output-level hybridmodels, the proposed feature-map-level fusion, leveraging CUDAenabled GPUs, demonstrates robust performance on a small,unaugmented dataset, reflecting clinical data scarcity in resourceconstrained settings. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations enhance interpretability, andfive-fold cross-validation confirms low overfitting, with minimalerrors (e.g., 9 glioma-meningioma misclassifications).