Variational AutoEncoder for Regression: Application to Brain Aging Analysis

Abstract

While unsupervised variational autoencoders (VAE) have become a powerful tool in neuroimage analysis, their application to supervised learning is under-explored. We aim to close this gap by proposing a unified probabilistic model for learning the latent space of imaging data and performing supervised regression. Based on recent advances in learning disentangled representations, the novel generative process explicitly models the conditional distribution of latent representations with respect to the regression target variable. Performing a variational inference procedure on this model leads to joint regularization between the VAE and a neural-network regressor. In predicting the age of 245 subjects from their structural Magnetic Resonance (MR) images, our model is more accurate than state-of-the-art methods when applied to either region-of-interest (ROI) measurements or raw 3D volume images. More importantly, unlike simple feed-forward neural-networks, disentanglement of age in latent representations allows for intuitive interpretation of the structural developmental patterns of the human brain.

Recent Advances in Variational Autoencoders With Representation Learning for Biomedical Informatics: A Survey

“…2) Improve the interpretability of representation learning [79,87,89,90]. Table 4 shows the literatures for the application of VAEs in medical imaging and image analyses.…”

“…Recent scientific advances have combined the interpretability of supervised settings with the power of VAEs. Zhao et al [89] proposed a VAEs based unified probabilistic model for learning the latent space of imaging data and performing supervised regression. Their results allow for intuitive interpretation of the structural developmental patterns of the human brain.…”

Cognitive Computing-Based CDSS in Medical Practice

“…For the representations of medical images, studies have been performed on layer-wise representing images from local regions till aggregating to the representation of the whole image including both 2D images [85,86] and 3D images [87]. Despite the above supervised methods that depend on the annotated labels of input to learn the representations, there exist unsupervised direct representation learning methods, generally known as an autoencoder, that encode the input into a feature vector in the latent space, and then, another network is used to decode the feature and reconstruct the input 4 Health Data Science [88,89]. By minimizing the difference between the original input and the reconstructed output, the autoencoder automatically learns the representations of the input without supervised labels.…”

Discussion of “Statistical disease mapping for heterogeneous neuroimaging studies”

“…Our modification thus enables the SDM framework to “disentangle” disease region configuration and covariate information. Different approaches to this problem have been considered in the machine learning literature (Zhao et al 2019, among others).…”