Incomplete Multi-Modal Disentanglement Learning With Application to Alzheimer's Disease Diagnosis.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Multi-modal neuroimaging (MRI, PET) aids Alzheimer's disease (AD) diagnosis.
• Incomplete data is a significant challenge in current computer-aided diagnosis.
• Existing methods for handling missing data reduce sample size or introduce noise.

Purpose of the Study:

• To develop a novel method for AD diagnosis using incomplete multi-modal neuroimaging data.
• To address the limitations of conventional strategies for handling missing data in neuroimaging analysis.
• To improve the accuracy and robustness of computer-aided diagnosis for AD.

Main Methods:

• Proposing Incomplete Multi-modal Disentanglement Learning (IMDL) for AD diagnosis.
• Utilizing modality-wise variational autoencoders and a Transformer for feature fusion.
• Implementing cross-modality contrastive learning and adversarial learning to harmonize representations.
• Developing a local attention rectification module for enhanced localization of atrophic areas.

Main Results:

• IMDL demonstrated superior performance in AD diagnosis on ADNI and AIBL datasets.
• The method effectively handles incomplete multi-modal neuroimaging data without scan synthesis.
• Validation on the HABS-HD dataset showed effectiveness for general dementia diagnosis with different imaging modalities.

Conclusions:

• IMDL offers a robust and effective solution for AD diagnosis with incomplete multi-modal neuroimaging data.
• The proposed method overcomes limitations of traditional approaches by avoiding sample reduction and noise introduction.
• IMDL shows promise for improving diagnostic accuracy and applicability across different neuroimaging datasets and dementia types.