A Lightweight CNN Architecture for Efficient Brain Tumor Detection in MRI Scans

The intricate morphology of brain tumors poses significant diagnostic challenges in MRI interpretation. While AI-driven systems offer potential for automation, balancing accuracy with computational efficiency remains critical for clinical adoption. This work introduces a lightweight convolutional neural network optimized for brain tumor detection and classification in MRI scans. The architecture’s design emphasizes a systematic exploration of layer-ordering strategies, with experiments revealing that batch normalization in post-activation mode (Post-BN) outperforms Pre-BN in training stability and classification accuracy. Contrary to expectations, integrating shortcut connections for residual learning demonstrated negligible performance gains. Evaluated on the Figshare (multi-class) and Br35H (binary) datasets, the model achieves state-of-the-art accuracy while maintaining resource efficiency through minimized parameters and FLOPs. These findings highlight the importance of strategic layer ordering over architectural complexity in deep learning for medical imaging, offering a framework for efficient and reliable tumor detection that could generalize to other vision-based diagnostic tasks.