Abstract
Structural connectomes provide critical insights into brain connectivity but are often confounded by high dimensionality and technical "site effects" in multi-center studies. We present a Joint Variational Autoencoder (Joint-VAE) framework that utilizes a hybrid latent space to disentangle continuous biological signals from discrete categorical variations. By employing staged capacity control and Gumbel-Softmax distributions, the model achieves unsupervised discovery of imaging sites and superior clustering performance over traditional dimensionality reduction techniques.
Introduction
Structural connectomes map the intricate white matter architecture of the human brain. While essential for studying neurodegeneration, these "wiring diagrams" are high-dimensional and susceptible to "site effects"—technical variations introduced by different scanner protocols or hospital settings. Traditional methods like PCA assume continuous variation, often failing to isolate discrete technical noise from biological signals. This work addresses the challenge of capturing mixed sources of variability in an unsupervised manner.
Methods: The Joint-VAE Framework
The proposed model utilizes a Joint-VAE with staged capacity control to learn a hybrid latent space.
Hybrid Latent Space
Unlike standard VAEs that use only continuous vectors, the Joint-VAE splits the latent space into two distinct components:
- Continuous Vector: Captures smooth variations such as connectivity strength or age-related changes.
- Discrete Vector: Captures categorical assignments, such as imaging sites or diagnostic groups, using the Gumbel-Softmax distribution to enable backpropagation through discrete choices.
Staged Capacity Control
To prevent the continuous space from dominating the model, the authors introduced an annealing strategy. By gradually increasing the "capacity" of the continuous space, the model is forced to utilize discrete codes early in training, ensuring effective disentanglement of categorical factors.
Experimental Setup and Datasets
The model was evaluated on a large-scale dataset of 5,761 structural connectomes from six major Alzheimer’s disease studies (including BLSA, NACC, and ROS/MAP/MARS). The cohort featured 10 unique acquisition protocols and a wide age range (22–102 years), including cognitively normal, MCI, and AD subjects. Connectomes were parcellated into 121 regions using the BrainCOLOR atlas.
Results and Discussion
The Joint-VAE demonstrated a remarkable ability to "discover" imaging sites without manual labels, achieving an Adjusted Rand Index (ARI) of 0.65. This significantly outperformed traditional PCA + k-means and standard VAE baselines ($p < 0.05$).
Qualitative analysis via t-SNE showed that the Joint-VAE preserved site-specific clusters as coherent groups, whereas traditional methods tended to fracture them. This stability suggests that providing a dedicated "discrete bucket" allows the model to effectively isolate technical noise from the continuous biological variation of interest.
Conclusion and Future Work
The Joint-VAE framework offers a robust solution for disentangling complex neuroimaging data. Future research will focus on validating the biological integrity of the continuous latent space after site-effect removal and extending the framework to other modalities like functional MRI. Incorporating semi-supervised objectives could further enhance the separation of technical and biological factors.
Summary updated