Optimized Lightweight Convolutional Neural Network Architecture for Multi-Class Brain Tumor Classification from Magnetic Resonance Imaging

Abstract

Brain tumor classification from Magnetic Resonance Imaging (MRI) is a vital clinical task in modern neuroradiology that requires high diagnostic accuracy to enable appropriate treatment and prognosis. We propose an optimized deep learning architecture with three architectural novelties—the Micro Adaptive Feature Extractor, Micro Multi-Scale Processor, and Context-Aware Intelligent Pooling—and present the world-lightest ultra-efficient brain tumor classifier. Our optimized model achieves 99.22% classification accuracy on the largest brain tumor MRI dataset to date (Mendeley), comprising 12,064 high-resolution images across 4 tumor classes (glioma, meningioma, pituitary adenoma, no tumor) using only 659,228 trainable parameters. Our system is thoroughly ablated and systemically validated with every architectural novelty validated individually and in combination, intelligent pooling providing the greatest 2.18% accuracy uplift. Our optimized architecture is extensively validated on cross-dataset independent test sets, surpassing domain adaptation baselines and achieving state-of-the-art generalization (Kaggle test: 99.24%, Br35H test: 98.33%). Our system uses 47x fewer parameters than state-of-the-art U-Net approaches while achieving top classification accuracy using only 12ms inference latency and 2.6 MB model size, making our solution perfect for deployment on underpowered mobile medical edge devices. These findings demonstrate that architectural innovation substantially outweighs brute-force parameter accumulation in medical image classification tasks.