Surface-Based Multi-axis Longitudinal Disentanglement Using Contrastive Learning for Alzheimer’s Disease

Comprehensive Summary

The objective of this study by Zhang and Shi was to generally discuss improvements to the model of Alzheimer’s disease progression made possible through neuroimaging scans that were able to distinguish the progression due to Alzheimer’s disease from that associated with the Aging process. A highly successful surface-based model was created that uses the longitudinal measurement of cortical thickness, containing a space that defines one axis for aging and other axes for diseases. Testing on 1,321 patients whose MRI scans were made available through the ADNI database showed significantly better separation between CN, MCI, and AD patients, better discrimination between stable and progressing MCI patients, better correlation with cognitive scores, and better accuracy in the classification of amyloid status compared to single-axis and standard autoencoder models. An ablation study also showed that both multi-axis formulation and contrastive loss were essential for optimal performance. The researchers emphasize that Alzheimer's progression differs from typical single disease progression and cannot adequately be represented, while a multi-axis formulation has the flexibility and interpretability to be properly visualized.

Outcomes and Implications

The study is important because proper differentiation of aging from disease-specific neurodegeneration in Alzheimer’s disease is essential for early diagnosis, prognosis, and stratification. The proposed model accounts for a variety of disease trajectories and thus has significant implications for the improvement of correlations regarding cognitive decline and biological markers. In experimental work, the proposed model could have better personalized monitoring using MRI data from improved risk prediction of MCI conversion and patient selection for clinical trials. Although there are no indications within the proposed model regarding the direct applications of it to clinical and preclinical applications at the current level of development, the validation of the model at the level of large-scale databases has implications related to the preclinical stage at the short-term level.