A comparative study of vision transformer architectures for the detection of Alzheimer's disease using magnetic resonance images.

Authors

Abstract

Alzheimer's disease is a neurodegenerative disease that affects millions of people worldwide. With the increasing global elderly population, early and accurate diagnosis of the disease is becoming increasingly important to slow its progression and improve patient outcomes. This study aims to present a comprehensive vision transformer-based solution proposal for the accurate and rapid detection of Alzheimer's disease using magnetic resonance images. Nine different transformer-based models, namely Vision Transformer, Pooling-based Vision Transformer, Convolutional Vision Transformer, Crossformer, Cross-attention Vision Transformer, Nested Transformers, Multi-Axis Vision Transformer, Separable Vision Transformer, and MobileViT, were used to analyse their effectiveness in diagnosing the disease in detail. Experimental results have shown that transformer-based architectures exhibit high performance in the detection of Alzheimer's disease. In binary classification results, the Multi-Axis Vision Transformer model stood out from other models by achieving the most successful performance in accuracy, sensitivity, specificity, precision, and f1-score metrics. In multi-class classification results, the Crossformer model provided the highest results in terms of sensitivity, specificity, precision, and f1-score, while the Separable Vision Transformer model was the most successful model in the accuracy metric. Furthermore, it was observed that the Mobile Vision Transformer model produced competitive results, especially in sensitivity and specificity metrics. In conclusion, this study provides a promising framework for the use of vision transformers in neuroimaging, providing an innovative solution for the early diagnosis of Alzheimer's disease.

Tags

Topics