Neurocognitive Latent Space Regularization for Multi-Label Diagnosis from MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Cognitive Neuroscience

Background:

• Deep learning models are crucial for neuroscientific discovery using brain MRI.
• Interpretability of these models is essential for reliable scientific conclusions.
• Current methods lack sufficient alignment with clinically meaningful variables in the latent space.

Purpose of the Study:

• To enhance the interpretability of deep learning models in brain MRI studies.
• To regularize the latent space of a multi-label classifier using pairwise disentanglement.
• To align the latent space representation with neuropsychological test scores.

Main Methods:

• Applied pairwise disentanglement to regularize the latent space of a multi-label classifier.
• Used brain MRI data from controls, mild cognitive impairment (MCI), and HIV-associated cognitive disorder (HAND) cases.
• Disentangled the latent space with respect to the neuropsychological z-score (NPZ).

Main Results:

• The proposed disentanglement method achieved statistically significantly higher balanced accuracy compared to a model without disentanglement.
• The difference in latent space representations along the disentangled direction significantly correlated with the difference in NPZ scores.
• Identified brain regions crucial for classification that align with existing neuroscientific literature.

Conclusions:

• Pairwise disentanglement effectively enhances the interpretability of deep learning models in brain MRI analysis.
• The method provides a more clinically meaningful latent space representation, correlating with cognitive status.
• This approach offers a promising tool for neuroscientific discovery and understanding cognitive impairment.