Randomized trial investigates latent representations in brain MRI, indicating importance in Alzheimer’s research.
Key Points
The research aims to explore latent representations in brain MRI using a convolutional autoencoder to understand clinical variability in Alzheimer’s disease.
Developed a convolutional autoencoder with a hierarchical encoder and compact latent space.
Applied dimensionality reduction techniques like PCA, t-SNE, PLS, and UMAP for visualization.
Introduced the Latent–Regional Correlation Profiling (LRCP) framework to identify clinically relevant brain regions.
Minimal architectures captured meaningful progression patterns to Alzheimer's disease.
Validation through SHAP-based regression predicted reconstruction error from gray matter intensities.
Statistical agnostic methods confirmed the results, emphasizing rigorous evaluation in neuroimaging.