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.
Introduction
In neuroimaging, large-scale datasets are essential for achieving the statistical power necessary to understand complex neurological disorders. To this end, researchers frequently pool data from multiple hospitals and research centers. However, this practice introduces significant "scanner effects." Every MRI machine possesses unique characteristics—ranging from manufacturer variations (e.g., GE vs. Phillips) to magnetic field inhomogeneities—that create non-obvious artifacts in the resulting images. These artifacts act as confounding variables, potentially leading machine learning models to identify the scanner rather than the underlying pathology. Traditional correction methods, such as physical phantom calibration or linear regression, often fail to capture the complex, non-linear nature of these biases.
Methods
The proposed solution utilizes a Generative Adversarial Network (GAN) architecture, specifically drawing from the CycleGAN or DiscoGAN frameworks. This approach treats scanner harmonization as an image-to-image translation task. Just as a GAN can be trained to translate photos from "summer" to "winter" or "horse" to "zebra," it can learn to translate the specific "style" or artifact distribution of one scanner to match another.
The framework relies on adversarial training where a Generator network attempts to transform images from Site A to look like Site B, while a Discriminator network attempts to distinguish between real and transformed images. To preserve the anatomical integrity of the brain during translation, a cycle-consistency constraint is employed: translating an image from A to B and back to A must yield the original image. The authors utilized the Least Squares GAN (LSGAN) objective function to ensure stable training and high-quality image generation.
Datasets and Experimental Setup
The model was validated using three distinct datasets. First, a proof-of-concept was conducted using the MNIST handwritten digit dataset, where half the images were artificially corrupted to test the GAN's cleaning capabilities. The primary study focused on structural T1-weighted MRI scans from 313 subjects across two sites using different 3T scanners (Phillips Achieva and GE Discovery MR750).
As shown in the demographic breakdown, the cohort included healthy controls and patients with schizophrenia. Additionally, a "reconstruction set" of 11 subjects scanned at both sites provided a ground truth for measuring the accuracy of the cross-scanner transformation.
Results
The GAN's effectiveness was evaluated based on its ability to neutralize scanner bias while preserving clinical signals. Before correction, a Support Vector Machine (SVM) could identify the source scanner with 99.3% accuracy. Following GAN harmonization, this accuracy dropped significantly, indicating the successful removal of site-specific fingerprints.
Crucially, removing this noise enhanced the underlying biological signal. Gender classification accuracy improved by 15.8%, and schizophrenia diagnosis accuracy increased by 3.7% compared to uncorrected data. In the reconstruction task, the GAN reduced the Mean Square Error (MSE) between scanners by 27.02%, significantly outperforming traditional linear and Gaussian Process regression methods, which achieved only ~6% reduction.
Discussion and Limitations
The "black box" nature of this GAN approach is a significant advantage, as it requires no prior knowledge of the specific sources of scanner bias. By learning these biases directly from the data, it offers a powerful tool for post-hoc data pooling. However, the current study is limited to 2D sagittal slices. Extending the architecture to 3D convolutional networks is a necessary next step, though it presents challenges regarding GPU memory and computational overhead. Future work will also explore integrating covariates like age and sex directly into the GAN architecture to further refine the translation process.
Summary updated