Latent Space Projections and Atlases, a Cautionary Tale in Deep Neuroimaging using Autoencoders.
J M Gorriz1, F Segovia1, C Jimenez-Mesa1
1Data Science and Computational Intelligence Institute, Granada 18071, Spain.
International Journal of Neural Systems
|June 30, 2026
Summary
This study presents a deep learning framework to analyze 3D brain MRI data, uncovering patterns related to Alzheimer's disease progression. The model effectively identifies brain regions encoding clinically relevant information for biomarker discovery.
Area of Science:
- Neuroimaging and Artificial Intelligence
- Computational Neuroscience
- Medical Image Analysis
Background:
- Alzheimer's disease (AD) poses a significant challenge, necessitating advanced methods for early detection and understanding.
- Exploring latent representations in 3D brain MRI can reveal subtle neuroanatomical patterns associated with cognitive decline.
Purpose of the Study:
- To develop and validate a deep learning framework for inferential exploration of latent representations in 3D brain MRI.
- To identify brain regions encoding clinically relevant information for Alzheimer's disease progression using a novel Latent-Regional Correlation Profiling (LRCP) framework.
Main Methods:
- A convolutional autoencoder with a hierarchical encoder was trained on segmented gray matter images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
- Dimensionality reduction techniques (PCA, t-SNE, PLS, UMAP) and the proposed LRCP framework were used for visualization and interpretation of the latent space.
- SHAP-based regression was applied for interpretability, assessing the relationship between reconstruction error and regional gray matter intensities.
Main Results:
- The deep learning model learned latent representations preserving neuroanatomical structure and reflecting clinical variability across cognitive status.
- The LRCP framework successfully identified brain regions encoding clinically relevant latent information associated with Alzheimer's disease progression.
- Interpretability analysis highlighted anatomically meaningful regions involved in the model's reconstruction strategies, validated by statistical agnostic methods.
Conclusions:
- Even minimal deep learning architectures can capture meaningful patterns related to Alzheimer's disease progression in 3D brain MRI.
- The proposed framework demonstrates the potential of autoencoders as exploratory tools for biomarker discovery and hypothesis generation in clinical neuroscience.
- Rigorous evaluation is crucial for validating findings in neuroimaging studies.
