Variational voxelwise rs-fMRI representation learning: Evaluation of sex, age, and neuropsychiatric signatures

Eloy Geenjaar, Tonya White, Vince D. Calhoun

Introduction

Resting-state functional MRI (rs-fMRI) provides a window into the brain's functional dynamics. However, the sheer scale and noise of 4D voxelwise data present significant computational hurdles. Traditionally, researchers rely on linear dimensionality reduction techniques like Principal Component Analysis (PCA) or Independent Component Analysis (ICA) before applying deep learning. This paper challenges that paradigm, introducing a 3D ResNet-based Variational Autoencoder (VAE) designed to learn non-linear representations directly from minimally processed voxelwise rs-fMRI data.

Methodology

The authors developed a 3D ResNet VAE architecture, utilizing weight normalization and Exponential Linear Unit (ELU) activations for training stability. To handle the temporal dimension and noise, they employed Piecewise Aggregate Approximation (PAA), averaging consecutive time windows to reduce data volume without losing essential dynamics.

The model was trained unsupervised on the UK Biobank dataset (12,314 subjects). The resulting latent representations ($z$) were then evaluated on downstream tasks—age regression, sex classification, and schizophrenia diagnosis—using standard classifiers (SVM, kNN, SVR) and a Long Short-Term Memory (LSTM) network to capture temporal dependencies.

Results and Analysis

Age Regression and Sex Classification

The unsupervised VAE successfully captured features predictive of both age and sex. For age regression, a 512-dimensional latent space paired with a Support Vector Regressor (SVR) achieved a Mean Absolute Error (MAE) of 4.014 years and an $R^2$ of 0.5288, significantly outperforming PCA-based baselines.

Age regression performance across different metrics and latent dimensions.

In sex classification, the same representations yielded a near-perfect ROC-AUC of 0.994 when paired with an SVM, demonstrating the richness of the learned latent space.

Schizophrenia Prediction and Transfer Learning

A critical finding was the efficacy of transfer learning. Pre-training the VAE on the large-scale UK Biobank dataset significantly boosted performance on the smaller FBIRN schizophrenia dataset (325 subjects).

Impact of pre-training and fine-tuning on schizophrenia prediction performance (ROC-AUC).

The best results for schizophrenia prediction (ROC-AUC of 0.7452) were achieved by an LSTM network processing the sequence of latent representations. This suggests that the biological markers for schizophrenia are deeply embedded in the brain's non-linear temporal dynamics rather than static snapshots.

Conclusion

This work demonstrates that unsupervised deep learning can extract meaningful biological signatures directly from high-dimensional rs-fMRI data. The success of pre-training on healthy populations to improve clinical diagnostic performance offers a promising path forward for neuroimaging studies constrained by small sample sizes. Future research will focus on optimizing data pipelines and enhancing the temporal interpretability of these non-linear models.

Summary updated

Abstract

This work uses a variational autoencoder (VAE) to perform non-linear representation learning from voxelwise rs-fMRI data. The VAE learns a non-linear dimensionality reduction of the data in the form of a latent vector. These latent vectors retain meaningful information related to a subject’s demographics and clinical diagnosis. The retention of meaningful information in the latent vectors is evaluated using age regression and sex classification tasks on the UK Biobank dataset. The results on these tasks are highly encouraging and a linear regressor trained on the latent vectors to predict age performs almost on par with a supervised neural network. Further, the same latent vectors can almost perfectly linearly separate sex. The model that is pre-trained on UK Biobank is also fine-tuned on a smaller neuropsychiatric dataset for a varying number of epochs. The latent vectors it generates for this dataset are then evaluated by performing a schizophrenia diagnosis classification task. We find that pre-training the model on UK Biobank significantly improves the quality of the latent vectors and that the vectors themselves are fairly discriminative. To understand the structure of the latent vectors with respect to demographic variables or neuropsychiatric disorders we train a variety of supervised models on the latent vectors. The results presented in this work open up more in-depth research into the factors of variation that the VAE models and how they can be improved for voxelwise rs-fMRI data.
BibTeX
@inproceedings{geenjaar2021variational,
  author = {Eloy Geenjaar and Tonya White and Vince D. Calhoun},
  title = {{Variational voxelwise rs-fMRI representation learning: Evaluation of sex, age, and neuropsychiatric signatures}},
  booktitle = {BIBM},
  year = {2021},
  doi = {10.1109/bibm52615.2021.9669775},
}
Figure