[Paper Review] Correcting differences in multi-site neuroimaging data using Generative Adversarial Networks

Harrison Nguyen, Richard W. Morris+3 more|arXiv (Cornell University)|Mar 26, 2018

Advanced Neuroimaging Techniques and ApplicationsMedicine39 references4 citations

TL;DR

This paper proposes a GAN-based method to correct non-linear, site-specific differences in multi-site structural MRI data without prior knowledge of bias sources. By learning a domain-invariant transformation between two scanner sites, the model reduces scanner-induced variance while preserving clinically relevant information, significantly improving classification accuracy for gender and schizophrenia diagnosis compared to linear methods.

ABSTRACT

Magnetic Resonance Imaging (MRI) of the brain has been used to investigate a wide range of neurological disorders, but data acquisition can be expensive, time-consuming, and inconvenient. Multi-site studies present a valuable opportunity to advance research by pooling data in order to increase sensitivity and statistical power. However images derived from MRI are susceptible to both obvious and non-obvious differences between sites which can introduce bias and subject variance, and so reduce statistical power. To rectify these differences, we propose a data driven approach using a deep learning architecture known as generative adversarial networks (GANs). GANs learn to estimate two distributions, and can then be used to transform examples from one distribution into the other distribution. Here we transform T1-weighted brain images collected from two different sites into MR images from the same site. We evaluate whether our model can reduce site-specific differences without loss of information related to gender (male, female) or clinical diagnosis (schizophrenia, bipolar disorder, healthy). When trained appropriately, our model is able to normalise imaging sets to a common scanner set with less information loss compared to current approaches. An important advantage is our method can be treated as a black box that does not require any knowledge of the sources of bias but only needs at least two distinct imaging sets.

Motivation & Objective

• Address the challenge of scanner-induced variability in multi-site neuroimaging studies, which undermines statistical power and reproducibility.
• Develop a post-hoc, data-driven method to normalize MRI data across sites without requiring knowledge of the sources of bias.
• Preserve clinically relevant information (e.g., gender and diagnosis) while reducing site-specific artifacts in T1-weighted brain images.
• Demonstrate that deep generative models like GANs outperform traditional linear or Gaussian process regression in correcting complex, non-linear site differences.

Proposed method

• Train a conditional generative adversarial network (cGAN) to learn a mapping from T1-weighted MRI images acquired at one site to the distribution of images from a target site.
• Use a U-Net-based generator with skip connections to preserve spatial details during image translation, and a PatchGAN discriminator to evaluate local image realism.
• Train the GAN using adversarial loss to minimize the difference between real and generated images, and cycle-consistency loss to ensure structural fidelity.
• Apply the trained GAN to transform mid-sagittal slices from a source site to match the target site’s distribution, enabling domain adaptation.
• Evaluate performance using downstream classification tasks (gender and schizophrenia) to assess preservation of biological signal.
• Compare the GAN approach against linear regression and Gaussian process regression as baselines, using classification accuracy and statistical invariance as metrics.

Experimental results

Research questions

• RQ1 Can a GAN-based method effectively reduce non-linear, site-specific differences in multi-site structural MRI data without prior knowledge of bias sources?
• RQ2 Does the GAN transformation preserve clinically relevant information such as gender and diagnosis while minimizing scanner-induced variance?
• RQ3 How does the GAN-based normalization compare to linear and Gaussian process regression in terms of classification accuracy for diagnostic and demographic labels?
• RQ4 Where in the brain do the most significant image transformations occur, and do these regions correlate with known anatomical differences related to gender or disease?

Key findings

• The GAN-based method significantly improved classification accuracy for gender compared to both linear regression and Gaussian process regression, indicating better preservation of biologically relevant structure.
• The model reduced site-specific differences to the point where transformed images were indistinguishable between scanners in a two-class classification task, suggesting effective domain alignment.
• The GAN outperformed linear and GP regression in distinguishing schizophrenia patients from healthy controls, indicating that non-linear site effects were better corrected by deep generative modeling.
• Most image transformations occurred in regions like the thalamus and brainstem—areas known to differ between males and females—suggesting the method enhances consistency in biologically relevant regions.
• The method successfully reduced bias-field-like variations across sites, even in regions with complex tissue contrast such as gray-white matter boundaries.
• Despite using only mid-sagittal slices, the GAN model captured sufficient spatial context to improve downstream classification, suggesting robustness to limited input dimensionality.