Mask attention layer modified residual vision transformer with transfer learning based features extraction for Alzheimer disease classification.
Authors
Affiliations (2)
Affiliations (2)
- Department of Computer Science & Engineering, School of Computational Sciences, Faculty of Science & Technology, JSPM University Pune, Maharashtra, 412207, India. Electronic address: [email protected].
- Department of Computer Science & Engineering, School of Computational Sciences, Faculty of Science & Technology, JSPM University Pune, Maharashtra, 412207, India.
Abstract
AD is an advanced neurodegenerative condition that requires precise and prompt examination for clinical involvement. Many deep learning (DL) algorithms for AD classification are computationally complex and unsuitable for real-time clinical applications. Hence, this study proposes a Mask Attention Layer Modified Residual Vision Transformer (MRViT) combined with transfer learning and Hiking Troops Optimization (HkTO) for effectual MRI-based AD classification. MRI images are gathered from the Open Access Series of Imaging Studies (OASIS) and the AD Neuroimaging (ADNI) dataset. Primarily, a Modified Pixel Gaussian filter is applied to eradicate noise and improve image quality. Afterward, deep features are extracted using transfer learning models, namely InceptionResNetV2, DenseEfficientNet-B3, and ConvoGoogleNet, to attain discriminative feature representations. The extracted features are then optimized using the HkTO algorithm to select the most relevant and non-redundant features, thereby dropping feature dimensionality. Also, HkTO is used to optimize the hyperparameters of the proposed MRViT classifier, refining classification accuracy and model robustness. Lastly, the certain optimal features are classified using the proposed MRViT architecture, including a mask attention mechanism to highlight disease-relevant regions while suppressing irrelevant information. Testing on the OASIS and ADNI datasets shows classification accuracies of 99.66 % and 99.89 %, proving the framework's precision and computational efficiency.