An explainable vision transformer approach for automated neurological disorder classification from brain MRI scans

Diagnosing Alzheimer’s Disease (AD), Multiple Sclerosis (MS), and brain tumors early enough to matter still depends heavily on brain MRI, and brain MRI still depends on radiologists who are in short supply almost everywhere. Convolutional neural networks have made real progress on this problem, but a single convolution only sees a small patch of the image, which limits how well a purely convolutional model can tie together pathology that is spread across the brain. This study addresses that limitation from two directions at once: a controlled, same-protocol comparison of established CNN backbones, and a lightweight attention mechanism grafted onto one of them. We classify the full eight-class Multi-Class Neurological Disorder (MCND) dataset (AD MildDemented, AD ModerateDemented, AD VeryMildDemented, Multiple Sclerosis, Normal, and three brain-tumor subtypes: glioma, meningioma, and pituitary), comprising 9,564 MRI images split 80:20 into 7,651 training and 1,913 test images. Four established CNNs, VGG-16, ResNet-50, DenseNet121, and EfficientNet-B0, were fine-tuned under one shared protocol to serve as controlled baselines. We then built VGG-16 + CBAM: a VGG16 backbone fitted with a Convolutional Block Attention Module and a much smaller classification head in place of VGG-16’s original fully connected layers, trained for 15 epochs under the same optimizer and augmentation settings used for the baselines. It reached 98.90% accuracy, a weighted F1-score of 0.9890, a macro F1 of 0.9893, a macro AUC-ROC of 0.9997, and an MCC of 0.9874 on the held-out test set, beating all four baselines (98.12%–98.85% accuracy) while using just 14.88 million parameters, roughly 9% of VGG-16’s 134.29 million, for a 56.8 MB model. We also report parameter counts, model size, FLOPs, GPU inference latency, and training time for every baseline, so the accuracy gain can be weighed against actual computational cost rather than accuracy alone. Grad-CAM maps for all eight classes show the model focusing on plausible anatomical regions: periventricular and hippocampal areas for the AD stages, white matter for MS, and the tumor core for each tumor subtype. A t-SNE projection of the learned features shows the eight classes forming distinct, mostly non-overlapping clusters. Taken together, the results suggest a small, attention-augmented CNN can match or beat larger backbones on this task at a fraction of the cost, without losing the interpretability a clinician would need to trust it.