A token efficient vision framework using patch residual transformer for Alzheimer's disease diagnosis.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Alzheimer's disease (AD) diagnosis relies on structural magnetic resonance imaging (sMRI), but current AI models like Vision Transformers (ViT) face challenges with data redundancy and feature focus.
• Early detection of AD and prediction of mild cognitive impairment (MCI) progression are critical for effective intervention.

Purpose of the Study:

• To introduce a novel, token-efficient framework, the Patch Residual Transformer (PRT), for enhanced sMRI-based AD diagnosis and MCI conversion prediction.
• To address information redundancy and focus loss issues in ViT models applied to high-dimensional neuroimaging data.

Main Methods:

• The PRT model partitions sMRI data into patches and utilizes a Patch Residual Block (PRB) with Top-K patch selection and linear stitch patch token fusion (LSPTF).
• Top-K identifies and ranks critical patches, while LSPTF focuses the model on essential regions, improving efficiency and performance.
• Validation was performed on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies-3 (OASIS-3) datasets.

Main Results:

• The PRT model significantly outperformed existing slice-, patch-, ROI-, and subject-level methods for both AD diagnosis and MCI conversion prediction.
• The PRT demonstrated strong generalization capabilities across different datasets.
• Consistent and robust salience maps were generated, highlighting the model's interpretability.

Conclusions:

• The PRT framework offers a superior approach for leveraging sMRI data in AD diagnosis and MCI conversion prediction.
• The proposed PRB, incorporating Top-K and LSPTF, effectively mitigates focus loss and enhances model performance in neuroimaging analysis.
• PRT shows promise for improving early detection and intervention strategies for Alzheimer's disease.