Abstract
Alzheimer’s disease (AD) diagnosis remains challenging due to the scarcity of annotated medical imaging data, pronounced class imbalance, and the subtlety of early-stage structural brain changes. Traditional deep learning approaches often overfit to majority classes and lack interpretability, limiting their clinical adoption. We propose a multimodal deep learning framework that fuses MRI-derived spatial features with standardized clinical parameters to enhance Alzheimer’s stage classification. The pipeline incorporates a GAN-based augmentation module to synthetically expand minority class samples, validated using MS-SSIM and expert review, thereby mitigating class imbalance. MRI features are extracted using a fine-tuned ResNet50, where lower convolutional layers preserve general spatial feature detectors while higher layers adapt to AD-specific morphological patterns. Clinical features, normalized via z-score scaling, are concatenated with MRI-derived embeddings to form a unified multimodal representation. This joint vector is processed through fully connected layers for four-stage classification Non-Demented, Very Mild Demented, Mild Demented and Moderate Demented. Model interpretability is ensured via Grad-CAM for spatial saliency mapping and clinical feature attribution, offering transparent decision support for healthcare practitioners. Experimentally, the framework achieved 97.56% validation accuracy, 0.98 macro precision, 0.97 macro recall and 0.98 macro F1-score, outperforming a baseline CNN (85.00% accuracy) and a GAN-augmented CNN (95.00% accuracy). Minority class recognition, such as for Moderate Demented improved from near-zero representation to 99% precision and recall due to balanced augmentation. This approach offers a scalable, interpretable, and high-accuracy diagnostic solution with the potential to transform automated neurodegenerative disease detection in clinical and telemedicine settings.

This publication has 31 references indexed in Scilit: