Latent space projections and atlases: A cautionary tale in deep neuroimaging using autoencoders

This study introduces a deep learning pipeline for the unsupervised analysis of 3D brain MRI using a simple convolutional autoencoder architecture. Trained on segmented gray matter images from the ADNI dataset, the model learns compact latent representations that preserve neuroanatomical structure and reflect clinical variability across cognitive states. We apply dimensionality reduction techniques (PCA, tSNE, PLS, UMAP) to visualize and interpret the latent space, correlating it with anatomical regions defined by the AAL atlas. As a novel contribution, we propose the Latent Regional Correlation Profiling (LRCP) framework, which combines statistical association and supervised discriminability to identify brain regions that encode clinically relevant latent features. Our results show that even minimal architectures can capture meaningful patterns associated with progression to Alzheimer Disease. Furthermore, we validate the interpretability of latent features using SHAP-based regression and statistical agnostic methods, highlighting the importance of rigorous evaluation in neuroimaging. This work demonstrates the potential of autoencoders as exploratory tools for biomarker discovery and hypothesis generation in clinical neuroscience.