Amyloid-β Deposition Prediction With Large Language Model Driven and Task-Oriented Learning of Brain Functional...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Amyloid-beta (Aβ) positron emission tomography (PET) is a gold standard for Alzheimer's disease (AD) diagnosis but is costly and involves high radioactivity.
• Functional connectivity networks (FCNs) derived from functional MRI (fMRI) show promise for assessing Aβ deposition non-invasively.
• Current FCN-based methods lack practical effectiveness for widespread Aβ assessment.

Purpose of the Study:

• To introduce a novel deep learning framework for assessing brain amyloid-beta (Aβ) deposition using functional MRI (fMRI).
• To develop a cost-effective and less radioactive alternative to PET imaging for early Alzheimer's disease (AD) screening.
• To identify key functional brain sub-networks predictive of Aβ deposition.

Main Methods:

• A deep learning framework incorporating a Large Language Model Nodal Embedding Encoder for fMRI feature extraction.
• A task-oriented Hierarchical-order FCN Learning module to model complex brain region correlations.
• Task-feature consistency losses to ensure accurate Aβ prediction and downstream classification effectiveness.

Main Results:

• The proposed deep learning framework significantly outperformed existing state-of-the-art FCN-based methods.
• The study successfully identified crucial functional sub-networks critical for predicting Aβ deposition.
• The method demonstrated superior accuracy in assessing Aβ protein deposition compared to other FCN approaches.

Conclusions:

• The novel deep learning framework offers a promising, cost-effective, and low-radioactivity approach for assessing brain amyloid-beta deposition using fMRI.
• This method can aid in large-scale early Alzheimer's disease screening and prevention strategies.
• The findings provide valuable insights into the relationship between functional brain connectivity and amyloid pathology in AD.