Randomizing Human Brain Function Representation for Brain Disease Diagnosis.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Diagnostics

Background:

• Resting-state fMRI (rs-fMRI) quantifies functional connectivity (FC) crucial for understanding brain diseases.
• Traditional FC analysis uses atlas-based region of interest (ROI) parcellation, which has limitations like subjective bias and lack of individual specificity.
• Existing methods struggle with the high dimensionality of fMRI data and integrating parcellation with FC analysis.

Purpose of the Study:

• To propose a novel randomizing strategy for brain function representation to enhance neural disease diagnosis.
• To overcome the limitations of traditional atlas-based ROI parcellation in fMRI analysis.
• To develop a framework that avoids subjective atlas selection and accounts for individual brain specificity.

Main Methods:

• A random sampling strategy is employed to extract brain patches, bypassing the need for atlas-based ROI parcellation.
• A new brain function representation framework is introduced, describing sampled patches based on anchor patches and positional information.
• An adaptive-selection-assisted Transformer network is designed to optimize and integrate patch representations for disease diagnosis.

Main Results:

• The proposed framework demonstrates effectiveness and generality across three independent datasets.
• Experimental results validate the method's ability to improve neural disease diagnosis.
• The approach offers a promising alternative to traditional atlas-based functional connectivity analysis.

Conclusions:

• The novel randomizing strategy effectively generates brain function representations for improved neural disease diagnosis.
• The proposed framework addresses key limitations of atlas-based methods in rs-fMRI analysis.
• This research presents a significant advancement for brain disease diagnosis using neuroimaging data.