Performance Comparison of Lightweight Neural Networks for Brain Tumor Classification on Magnetic Resonance Imaging (MRI)

Brain tumors remain a major cause of mortality worldwide, where early and accurate diagnosis is critical for effective treatment planning. This study compares the performance of lightweight neural networks for brain tumor classification on the “Brain Tumor MRI Dataset” from Kaggle. The Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology guided the experimental workflow, including data understanding, preprocessing, modelling, and evaluation. Two lightweight architectures, a Multilayer Perceptron (MLP) and LeNet, were trained and evaluated for comparison. Results show that LeNet outperformed the MLP, achieving test accuracy of 92.67% with a loss of 1.65, while the MLP attained 84.38% accuracy with a loss of 0.87. LeNet required fewer parameters which were 914,384 than the MLP which were 2,329,754. Experiments were conducted on an NVIDIA RTX 3000 Ada Generation Laptop GPU with 8 GB dedicated memory. Under this setting, LeNet completed training in 1 h 16 min 45.26 s, compared with 1 h 30 min 26.79 s for the MLP. While recent transfer-learning and deeper CNN studies on brain tumor MRI classification commonly report accuracies above 98%, they often rely on substantially heavier architectures; in contrast, the proposed LeNet achieved 93.33% accuracy with only 914,384 parameters and shorter training time, highlighting a practical trade-off between accuracy and efficiency. These findings demonstrate that lightweight convolutional architectures can achieve superior classification performance with lower computational cost, highlighting their suitability for resource-constrained medical imaging applications and edge-based healthcare systems.