A Review of Latent Representation Models in Neuroimaging

TL;DR

The paper surveys how latent representation models, including variational autoencoders, latent diffusion models, and generative adversarial networks, are applied to neuroimaging to manage high-dimensional data, reveal disease- and age-related patterns, and illuminate brain function. It outlines theoretical foundations (manifold hypothesis and Bayesian inference) and surveys state-of-the-art applications across harmonization, image reconstruction from fMRI, brain aging, disease classification, functional networks, and multimodal integration. Key contributions include demonstrations of high-quality image synthesis, cross-site data harmonization, and biologically meaningful latent spaces, as well as discussions of limitations such as latent space interpretability and data heterogeneity. The work underscores the potential of latent generative models to advance clinical diagnostics, personalized medicine, and our understanding of brain mechanisms, while calling for further methodological developments and rigorous empirical validation.

Abstract

Neuroimaging data, particularly from techniques like MRI or PET, offer rich but complex information about brain structure and activity. To manage this complexity, latent representation models - such as Autoencoders, Generative Adversarial Networks (GANs), and Latent Diffusion Models (LDMs) - are increasingly applied. These models are designed to reduce high-dimensional neuroimaging data to lower-dimensional latent spaces, where key patterns and variations related to brain function can be identified. By modeling these latent spaces, researchers hope to gain insights into the biology and function of the brain, including how its structure changes with age or disease, or how it encodes sensory information, predicts and adapts to new inputs. This review discusses how these models are used for clinical applications, like disease diagnosis and progression monitoring, but also for exploring fundamental brain mechanisms such as active inference and predictive coding. These approaches provide a powerful tool for both understanding and simulating the brain's complex computational tasks, potentially advancing our knowledge of cognition, perception, and neural disorders.

Figures (4)

• Figure 1: Simplified scheme of the latent representation. High-dimensional neuroimaging scans are processed through a model capable of computing the probability distribution of a lower-dimensional manifold that captures relevant patterns in our dataset, whether the model is implicit or explicit.
• Figure 2: Simplified scheme of the VAE. This explicit model captures the priori distribution that encodes high-dimensional data into the latent manifold, allowing to sample and access the latent variables of the model.
• Figure 3: Simplified scheme of the LDM. This latent generative model takes a latent space and generate a forward diffusion, introducing noise into this representations. Then, an inverse process learns to denoise the latent space to generate a reconstructed latent space using a denoising U-Net.
• Figure 4: Simplified scheme of the GAN model. A latent space is fed to a generator that produces fake reconstructed images. Then, a discriminator learns to distinguish between real and fake images, updating the weights of the model according to the loss function