A 3D decoupling Alzheimer's disease prediction network based on structural MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Current Alzheimer's disease (AD) prediction methods struggle to fully utilize structural magnetic resonance imaging (sMRI) data.
• Traditional convolutional neural networks face challenges in accurately identifying AD lesion structures within sMRI scans.

Purpose of the Study:

• To develop an advanced three-dimensional (3D) Alzheimer's disease (AD) prediction method using sMRI.
• To overcome the limitations of existing predictive models in harnessing the full potential of 3D sMRI data for AD diagnosis.

Main Methods:

• A 3D decoupling, self-attention network is proposed for AD prediction.
• A multi-scale decoupling block enhances feature extraction by segregating convolutional channels.
• A self-attention block adaptively fuses features from sagittal, coronal, and axial directions to focus on brain lesion areas.
• A joint loss function combining clustering and cross-entropy loss improves sample discrimination.

Main Results:

• The model achieved high accuracy: 0.985 on the ADNI dataset and 0.963 on the AIBL dataset.
• The model accurately differentiates between normal controls (NC) and Alzheimer's disease (AD).
• The model effectively distinguishes between stable mild cognitive impairment (sMCI) and progressive mild cognitive impairment (pMCI).

Conclusions:

• The proposed 3D AD prediction network demonstrates competitive performance against state-of-the-art methods.
• The model addresses challenges in 3D sMRI analysis, improving AD diagnosis and treatment prediction.
• This advancement enhances the utility of predictive methods for Alzheimer's disease management.