DiSCNet: Directional Split Convolution for compute-efficient brain tumor diagnosis

Brain tumor classification from magnetic resonance imaging (MRI) remains challenging because tumor appearance varies substantially across patients, while scanner- and protocol-related differences can alter image intensity distributions and weaken model generalization. This study aims to develop a compact yet high-performing deep learning framework that can classify heterogeneous brain MRI more reliably without relying on large-capacity backbones. To this end, we propose DiSCNet, a lightweight architecture built on an InceptionNeXt-inspired hierarchical design and centered on a novel Directional Split Convolution (DiSC) block. The proposed block diversifies receptive fields through complementary local and directional depthwise branches, while Global Response Normalization and Efficient Channel Attention are incorporated to improve feature stability and channel selectivity under acquisition variability. The model was evaluated on a unified benchmark constructed from five publicly available MRI repositories, comprising 17,888 images across four classes: glioma, meningioma, pituitary and non-tumor. Under a single training and evaluation protocol, DiSCNet was compared against 71 contemporary convolutional, transformer-based, and hybrid architectures. DiSCNet achieved the best overall performance, with 0.9922 accuracy, 0.9916 precision, 0.9930 recall, and 0.9923 F1-score, while using only 2.78 million parameters. Class-wise analysis further showed strong and balanced recognition across all diagnostic categories, and Grad-CAM visualizations indicated that the model predominantly focused on lesion-relevant regions. These findings demonstrate that a carefully designed lightweight architecture can outperform substantially larger models and provide an efficient, robust, and clinically relevant solution for four-class brain tumor MRI classification.