A generative modeling approach for interpreting population-level variability in brain structure

Abstract

Understanding how neural structure varies across individuals is critical for characterizing the effects of disease, learning, and aging on the brain. However, disentangling the different factors that give rise to individual variability is still an outstanding challenge. In this paper, we introduce a deep generative modeling approach to find different modes of variation across many individuals. To do this, we start by training a variational autoencoder on a collection of auto-fluorescence images from a little over 1,700 mouse brains at 25 micron resolution. To then tap into the learned factors and validate the model's expressiveness, we developed a novel bi-directional technique to interpret the latent space-by making structured perturbations to both, the high-dimensional inputs of the network, as well as the low-dimensional latent variables in its bottleneck. Our results demonstrate that through coupling generative modeling frameworks with structured perturbations, it is possible to probe the latent space to provide insights into the representations of brain structure formed in deep neural networks.Keywords: variational autoencoder · interpretable deep learning · brain architecture and neuroanatomy.

Detection of Abnormal Folding Patterns with Unsupervised Deep Generative Models

“…These three classifiers are used first to ensure that the latent representations are able to capture some relevant information regarding the folding patterns. Then, we can focus our analysis on the features contributing the most to the success of the classification using the generative power of the β − V AE, like in [14]. We travel throughout the latent space modifying only one of these features and observe the generated folding patterns.…”

“…Our β − V AE was composed of fully convolutional encoder and decoder with symmetrical architectures comprising three convolutional blocks and 2 fully connected layers. We did a gridsearch (L=8-100, β=1-20, ranges are based on previous works [14] and reconstruction abilities), where hyperparameters were chosen according to the classification performances of the Deletion classifier applied to 100 controls and 100 synthetic samples using a 5-fold stratified cross-validation. We selected L=100, β=2 and a learning rate of 2e-4.…”

“…Based on these initial results, a first framework was proposed for anomaly detection in computed tomography scans of 3D brains where anomalies consisted in labeled traumatic brain injuries [23]. β-VAE have also been successfully used to model the inter-individual variability in the mouse brain [14]. More recently, very promising self-supervised methods have been applied to anomaly detection problems in medical images [6], but these methods lack the generative aspect provided by GAN and VAE, which is crucial in terms of explainability.…”

“…For this purpose, we propose first a dedicated preprocessing leading to focus the learning on the cortical folding geometry of a specific region of interest (ROI). Then, like in [14], a β − V AE is trained on a set of control data sampling the general population to get a latent representation of the folding pattern distribution in this ROI. We also create several benchmark datasets simulating unusual regional folding patterns to assess the ability of our model to detect them.…”

Disentangled representations: towards interpretation of sex determination from hip bone

“…There is thus a key challenge to learn disentangled representations where variables of interest would be independently and explicitly encoded [Bengio et al, 2013]. There are three main paradigms to learn disentangled representations: unsupervised [Chen et al, 2018, Liu et al, 2020, supervised or semi-supervised ,Zhao et al, 2019, and weakly-supervised [Ruiz et al, 2019]. In the supervised or semi-supervised case, the factors of interest are explicitly labelled in all or in a part of the training set.…”