Lightweight Vision Transformer with transfer learning for interpretable Alzheimer's disease severity assessment

Early and reliable diagnostic tools are critical for slowing the progression of Alzheimer's disease (AD), a neurodegenerative disorder characterized by memory loss and cognitive decline. This study introduces, ViTTL, lightweight deep learning framework for assessing the severity of AD using MRI data. ViTTL integrates Vision Transformers (ViT) with pre-trained convolutional neural networks utilized in transfer learning mode, to extract informative features from 2D MRI slices. Among the evaluated combinations, the ViT-DenseNet201 model integrated with an artificial neural network (ANN) classifier achieved the highest classification accuracy (99.89%) on the OASIS dataset. To ensure interpretability, we incorporated LIME and GRAD-CAM method, which consistently focus on cortical and hippocampal regions known to be associated with Alzheimer's pathology. The average Dice similarity coefficient across runs was 0.85 with a standard deviation of 0.03, indicating high consistency in the model's focus regions against ground truth annotations by expert radiologists. ViTTL also achieved a substantial reduction in model size from 83.0 MB to 6.47 MB enabling deployment in resource-limited environments without compromising performance. Validation on an independent dataset (Kaggle) and comparative performance analysis against state-of-the-art methods further support the robustness and generalizability. These findings demonstrate that ViTTL is a promising tool for accurate, interpretable, and resource-efficient AD diagnosis, with strong potential for clinical translation and patient outcome improvement. The related codes are available at https://github.com/RuhikaSharma/enhanced-alzheimer-risk-assessment .