AlzheimerViT: harnessing lightweight vision transformer architecture for proactive Alzheimer's screening

Alzheimer's is a disease in the human brain characterized by gradual memory loss, confusion, and alterations in behavior. It is a complex and continuously degenerative disorder of the nervous system, which still has early detection and diagnosis as challenges to overcome. The disease causes significant damage in individuals suffering from the disorder as they progressively lose cognitive ability. Its diagnosis and management depend primarily on the ability to diagnose early to initiate proper intervention. Unfortunately, this remains a difficult feat, given the resemblance of early signs of the disease with symptoms associated with normal aging and other disorders involving cognition. While clinical tests have their limitations, brain imaging such as MRI can provide detailed insights into changes in the brain. Deep learning techniques, mainly when applied to MRI data, have proven helpful in the early detection of Alzheimer's Disease.

In the proposed study, a lightweight, self-attention-based vision transformer (ViT) is employed to predict Alzheimer's disease using MRI images from the OASIS-3 dataset. Data pre-processing and augmentation techniques have been added to strengthen the model's generalization ability and enhance model performance, which is visualized using Grad-Cam.

The proposed model achieves exceptional results with an accuracy of 98.57%, approximate precision of 98.7%, Recall of about 98.47%, and specificity of 98.67%. It also achieves a Kappa Score of 97.2% and an AUC ROC Score of 99%.

This paper, along with comprehensive data pre-processing and augmentation, represents one of the major steps toward achieving more robust and clinically applicable models for Alzheimer's disease prediction. The proposed study indicates that deep learning models have the potential to enhance the diagnosis of Alzheimer's disease. By integrating Deep learning techniques with careful data processing, more reliable early detection models can be developed, which in turn leads to better treatment outcomes.