BrainGuard: Unified Approach for MRI-Based Detection of Brain Diseases Using Mixture of Experts

Abstract

Accurate and efficient diagnosis of neurological disorders remains a significant challenge due to the complexity and similarity of imaging patterns across conditions such as brain tumors, stroke, and dementia. Existing deep learning systems typically rely on disease-specific models, require extensive datasets, and often depend heavily on data augmentation, limiting their suitability for real-world clinical deployment. This research introduces BrainGuard, a unified deep learning framework designed to perform multi-disease classification from 224×224 MRI images without the need for auxiliary datasets or augmentation. The proposed system employs a Mixture of Experts (MoE) architecture integrating EfficientNet, DenseNet, and ResNet backbones to enhance feature extraction through specialist networks. A gating mechanism selectively combines expert outputs, prioritizing the most relevant representations to improve generalization across diverse neurological conditions. Additionally, refined attention mechanisms further strengthen the model’s ability to capture critical diagnostic features while maintaining computational efficiency.

Initial experimental results demonstrate the effectiveness of this approach, with the model achieving 96.11% accuracy, 96.15% precision, 96.11% recall, and a 96.09% F1-score across multiple disease classes using a balanced dataset of 1,000 images per category. Notably, this performance was attained without augmentation, indicating strong intrinsic generalization capabilities. These findings highlight the potential of BrainGuard as a scalable, clinically relevant diagnostic tool capable of reducing computational overhead and supporting rapid, multi-condition medical imaging analysis. The framework represents a step toward practical, unified AI-assisted diagnosis in resource-constrained healthcare environments.