Transformer models and deep learning for medical image analysis in the early detection of neurological diseases

Abstract

Early detection of Alzheimer's Disease (AD) remains a major challenge in neuroimaging, as structural changes in the brain may be subtle and complex during the early stages of the disease. Although Convolutional Neural Networks (CNNs) have been widely applied to medical image analysis, their ability to model long-range spatial dependencies in complex brain images is inherently limited by their predominantly local receptive fields. This study investigates the application of the Vision Transformer (ViT) architecture to Structural Magnetic Resonance Imaging (sMRI) for AD classification. A transformer-based framework was developed to analyze sMRI images. By leveraging the self-attention mechanism, the proposed approach enables the modeling of relationships between spatially distant image regions and the extraction of spatial and textural features. Particular attention was given to identifying disease-associated structural alterations, including changes in the hippocampus and cerebral cortical regions. The performance of the proposed approach was compared with that of conventional CNN-based methods. The results indicated that the ViT-based deep learning approach achieved better performance than the comparison methods in terms of the evaluation metrics considered, including classification accuracy and sensitivity. Furthermore, performance evaluation under noisy conditions demonstrated the model's ability to maintain its performance across sMRI images of varying quality. These findings suggest that transformer-based architectures, through their self-attention mechanisms, offer a promising approach to extracting structural features from sMRI images and classifying individuals with AD. Further validation on independent datasets is required to establish their generalizability and clinical applicability.