Combined transformer encoder-CNN architecture with texture features for MRI-based Alzheimer's disease detection

Alzheimer's disease (AD) develops long before clinical symptoms appear, which highlights the need for early diagnostic tools. Current diagnostic techniques often miss the subtle structural brain changes in the early stages of AD, creating a critical gap in timely medical intervention. This paper proposes a combined modified Transformer Encoder-Convolutional Neural Network (CNN) architecture for AD detection using Magnetic Resonance Imaging (MRI). Data augmentation techniques, including rotation, zooming, brightness adjustment and flipping, are applied to increase the number of training images. Images are processed using a Modified Gaussian Filtering technique (modGFT) that implements a spatially adaptive variance to suppress noise while preserving critical anatomical edges and structural details. Relevant features, including shape features, Improved Median Robust Extended Local Binary Pattern (ImpMRELBP) and Pyramid Histogram of Oriented Gradients (PHOG) features, are extracted. The ImpMRELBP descriptor highlights boundaries and captures meaningful texture variations. These features are input into a hybrid modTransEncd-CNN detection model. The model incorporates batch normalization and an enhanced attention module, improving training speed, stability and generalization. The results from both models are merged using soft voting for AD classification. Experimental results demonstrate that the proposed modTransEncd-CNN method outperforms traditional models, achieving 96% accuracy.