A prior-sampling conditional variational autoencoder for neuroimaging normative modelling: Benchmarking deep learning...

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Normative modeling quantifies individual brain deviations using covariates.
• Deep learning advances multivariate analysis in neuroimaging.
• Existing conditional variational autoencoders (cVAEs) struggle with reliable probabilistic predictions.

Purpose of the Study:

• To develop an enhanced cVAE framework for improved normative modeling in neuroimaging.
• To leverage deep learning for high-dimensional data analysis in brain imaging.
• To accurately capture individual deviations related to hypertension severity.

Main Methods:

• Proposed an enhanced cVAE framework with prior-sampling inference.
• Utilized 195 imaging-derived phenotypes (IDPs) from UK Biobank participants (hypertensive and normotensive).
• Benchmarked against GAMLSS, MFPR, HBR, and standard cVAE methods.

Main Results:

• The enhanced cVAE framework demonstrated performance comparable to established models.
• The model accurately captured individual deviations associated with hypertension severity.
• The proposed inference strategy showed superior covariate sensitivity compared to existing cVAE methods.

Conclusions:

• Deep learning-based normative modeling shows promise for complex neuroimaging datasets.
• The enhanced cVAE framework offers a robust tool for personalized brain health assessment.
• This approach facilitates early detection of neurological disorders related to conditions like hypertension.