The early detection of Alzheimer’s disease (AD)
is crucial for successful treatment and care of the patient as
AD gradually reduces memory and cognitive abilities. This
study presents a deep learning system capable of distinguishing
between AD and healthy controls in the OASIS structural brain
MRI data set but designed with the ability of decisions being
checked rather than trusted. The key idea of the method is
using a 3D convolutional neural network (CNN) as well as a
3D Swin Transformer (Transformer): The CNN captures the
local anatomical detail, and the Transformer captures the global
context over the entire volumetric scan. Each MRI is first
preprocessed such as skull stripping, intensity normalization,
resizing, denoising to clean the signal prior to reaching the model.
All of this is implemented via PyTorch. To elucidate the region of
the brain that was responsible for each classification, we selected
Grad-CAM and Attention Rollout to trace the learning pathways
used by the network. The results on the test set were 92.31% of
accuracy, with 89.00% sensitivity to AD, 80.91% of precision. The
findings indicate the proposed hybrid CNN–Swin Transformer
model is a feasible solution towards the AD classification problem
without compromising interpretability and accuracy.

Alzheimer’s Disease,,Brain MRI,3D Convolu- tional Neural Network (3D-CNN),3D Swin Transformer,Ex- plainable Artificial Intelligence (XAI),Grad-CAM,Attention Rollout,Deep Learning