Detecting Alzheimer's Disease on Small Dataset: A Knowledge Transfer Perspective

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Computer-aided diagnosis (CAD) for Alzheimer's disease (AD) often requires large datasets, posing challenges for institutions with limited data.
• The heterogeneity between different functional magnetic resonance imaging (fMRI) data sources can hinder the generalizability of CAD models.
• Effective utilization of multi-source data for neurological disease diagnosis remains an area requiring further investigation.

Purpose of the Study:

• To address the challenge of limited training samples in Alzheimer's disease (AD) detection using computer-aided diagnosis (CAD).
• To investigate the impact of data source heterogeneity on the performance of CAD models for AD.
• To develop and validate a knowledge transfer method to improve CAD accuracy with small datasets.

Main Methods:

• Conducted a heterogeneity analysis comparing a small local hospital dataset with a large dataset from the AD neuroimaging initiative.
• Proposed and implemented a novel knowledge transfer method to mitigate data disparity between different sources.
• Evaluated the classification accuracy of the proposed method on datasets with insufficient training samples.

Main Results:

• Identified significant differences in sample distributions across different fMRI data sources.
• The proposed knowledge transfer method effectively diminished data disparity between datasets.
• Achieved an approximate 20% increase in classification accuracy compared to models trained solely on the small dataset.

Conclusions:

• The developed knowledge transfer approach offers a novel and effective solution for CAD in Alzheimer's disease (AD) for hospitals with limited data.
• This method successfully addresses the common challenge of small sample sizes in neurological disease detection.
• The findings highlight the potential of leveraging multi-source data for enhanced neurological disease diagnosis.