Multi-modal cross-attention network for Alzheimer's disease diagnosis with multi-modality data

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) is the leading cause of dementia, making early diagnosis of AD and mild cognitive impairment (MCI) critical.
• Existing deep learning models often concatenate multi-modal data features, overlooking differences in their representation spaces.
• Neuroimaging (sMRI, FDG-PET) and cerebrospinal fluid (CSF) biomarkers offer complementary diagnostic information.

Purpose of the Study:

• To propose a novel Multi-modal Cross-attention AD diagnosis (MCAD) framework for improved AD and MCI detection.
• To effectively learn interactions between different data modalities for enhanced diagnostic accuracy.
• To leverage cross-modal attention to integrate imaging and non-imaging data.

Main Methods:

• Developed an MCAD framework utilizing cascaded dilated convolutions for imaging data and a CSF encoder for non-imaging data.
• Introduced a multi-modal interaction module employing cross-modal attention to fuse sMRI, FDG-PET, and CSF biomarker information.
• Designed an objective function to minimize inter-modal discrepancies and improve feature fusion.

Main Results:

• The MCAD framework achieved superior performance in AD-related classification tasks on the ADNI dataset compared to existing methods.
• Cross-modal attention significantly improved the integration of multi-modal data for diagnostic tasks.
• Analysis confirmed the contribution of each modality and the effectiveness of the cross-attention mechanism.

Conclusions:

• The proposed MCAD framework demonstrates the efficacy of cross-modal attention in integrating diverse biomarkers for accurate AD diagnosis.
• Combining multi-modal data through cross-attention offers a promising approach for early detection of Alzheimer's disease and MCI.
• This method provides a robust tool for enhancing diagnostic performance in neurodegenerative disease research.