A Hybrid Deep and Broad Learning Approach for Automatic Brain Tumor Classification

Abstract

Accurate MRI-based brain-tumor classification is vital for timely treatment planning; yet conventional CNN pipelines can struggle to generalize across heterogeneous tumor phenotypes. We present RIE50, a hybrid framework that fuses pretrained CNN feature extractors with a Broad Learning System (BLS) enhancement layer and lightweight classification head. RIE50 aggregates multi-level convolutional descriptors, passes them through a BLS layer to expand and refine representations with nonlinear feature mapping, and then classifies four categories (glioma, meningioma, pituitary, and no-tumor). The model integrates diverse activation functions (SELU, PReLU, Mish) within the CNN blocks to improve feature diversity and gradient flow, and is optimized with standard first-order optimizers. Evaluated on multi-source brain MRI benchmarks (4-class setup), RIE50 delivers strong accuracy and balanced sensitivity/specificity across tumor types, outperforming single-backbone baselines while remaining computationally lightweight due to the shallow, wide BLS head. These results indicate that widening the representation via BLS complements deep features and can yield robust, data-efficient tumor classification for clinical decision support.