Abstract: Brain tumor diagnosis remains a major challenge due to the complex and heterogeneous nature of tumor structures. Accurate and early classification of tumor types is essential for guiding treatment strategies and improving patient outcomes. Although deep learning has shown great potential in medical image analysis, many existing models either require substantial computational resources or fail to generalize effectively across diverse tumor categories. In this study, we propose LIMRI-Net, a lightweight and high-performing Convolutional Neural Network (CNN) for brain tumor classification using Magnetic Resonance Imaging (MRI) scans. The framework integrates standard convolutional layers with multiscale atrous convolutions to capture spatial details across different receptive fields. An inception module further enhances multiscale feature extraction, while a Squeeze-and-Excitation (SE) block adaptively recalibrates channel responses to strengthen discriminative capability without increasing computational overhead. LIMRI-Net was evaluated on a publicly available four-class brain MRI dataset, where it achieved an accuracy of 98.63%. To assess robustness, additional experiments were performed on the Bangladesh MRI dataset, containing 6,056 images across three classes, where the model reached 98.18% accuracy. Remarkably, LIMRI-Net achieved the lowest parameter count (0.42 M), smallest model size, and minimal FLOPs compared to state-of-the-art methods, highlighting its suitability for deployment in resource-constrained clinical environments. Furthermore, Grad-CAM visualizations were employed to highlight the salient regions driving predictions, thereby improving interpretability. These findings demonstrate that LIMRI-Net offers an effective balance of accuracy, efficiency, and transparency, making it a reliable diagnostic support tool for advancing precision in brain tumor classification.

Key words: Brain tumors')">Brain tumors, Deep learning, Lightweight, Multiscale atrous convolutions, Feature extraction, Computational overhead