Bayes-Optimized EfficientViT: A Hybrid Deep Learning Framework for Multi-Modal Brain Stroke and Alzheimer’s Classification

Alzheimer’s disease (AD) and brain stroke (BS) are two common neurological conditions that result in significant cognitive and physical deficits; prompt and accurate image-based diagnosis is crucial for early management. However, most automated approaches now in use focus on single-disease classification and often have limited applicability on heterogeneous imaging modalities, significant processing cost, and poor clinical interpretability. To address these limitations, this study proposes EffViTNet, a Bayesian-optimized hybrid deep learning framework that combines EfficientNet-B0 and Vision Transformer (ViT-B16) for dual-disease multi-class classification utilizing MRI (AD) and CT (BS) data. The ViT branch uses transformer encoding and patch-wise tokenization to capture long-range global contextual links, whereas the EfficientNet branch learns fine-grained local structural patterns. For the final prediction, the compact hybrid representation produced by concatenating the two feature streams is submitted to a fully connected Softmax classifier. To ensure efficient and stable convergence, key training hyperparameters (learning rate and weight decay) are adjusted using Bayesian Optimization by maximizing validation performance. Grad-CAM is also used to enhance clinical trust by highlighting disease-relevant areas and providing visual explanations. Experimental results indicate that the proposed model achieves superior performance compared with existing approaches, reaching an accuracy of 96.7%, 96.06% precision, 96.88% recall, and an F1-score of 0.9645. The suggested approach performs well for automated neurological disease diagnosis across different modalities in terms of accuracy, robustness, and explainability.