Abstract
Magnetic Resonance Imaging (MRI) of the brain is a critical tool for investigating neurological disorders, yet data acquisition remains expensive and time-consuming. Multi-site studies offer a solution by pooling data to increase statistical power, but images are often confounded by site-specific "fingerprints"—differences in scanner hardware and protocols that introduce bias. We propose a data-driven approach using Generative Adversarial Networks (GANs) to harmonize T1-weighted brain images across different sites. By treating site differences as a style-transfer problem, our model reduces site-specific variance while preserving clinical information related to gender and diagnosis. Our results demonstrate that this "black box" method can normalize imaging sets more effectively than current linear approaches, improving the accuracy of downstream clinical classifications.
Introduction
In neuroimaging research, the pursuit of large datasets often leads to multi-center collaborations. However, every MRI scanner possesses unique characteristics—magnetic field stability, manufacturer-specific gradients (e.g., GE vs. Philips), and acquisition protocols—that create subtle but pervasive variations in the resulting images. These "site effects" act as confounding variables; for instance, if one site primarily scans healthy controls and another scans patients, a machine learning model might inadvertently learn to identify the scanner rather than the pathology.
Traditional normalization methods often rely on physical phantoms or simple linear corrections. However, these methods frequently fail to capture the complex, non-linear nature of how hardware affects biological tissue imaging. This paper explores the use of deep learning to treat scanner-specific artifacts as a "style" that can be translated or removed.
Methods: Cycle-Consistent GANs
The core of our approach is the Cycle-Consistent GAN (CycleGAN) architecture. Unlike standard GANs, CycleGAN is designed for image-to-image translation between two domains without requiring paired examples.
In this framework, the system uses two generators and two discriminators. Generator $G$ learns to make images from Site A look like they originated from Site B, while Generator $F$ performs the reverse. To ensure the model does not "hallucinate" or delete real biological features, we employ a cycle loss function: if an image is transformed from A to B and then back to A, the resulting image must be identical to the original. This ensures that only the "style" (scanner noise) is modified, while the "content" (brain anatomy) remains intact.
Datasets and Cohort
The study utilized T1-weighted structural MRIs from 313 participants across two distinct sites. Site A used a Philips Achieva 3T scanner, while Site B utilized a GE Discovery MR750 3T scanner. The cohort included healthy controls and patients diagnosed with schizophrenia.
For training, the middle five 2D sagittal slices were extracted from preprocessed gray matter density maps. Additionally, a small "ground truth" subset of 11 subjects scanned at both sites was used for direct validation of the transformation accuracy.
Key Results
Our evaluation focused on whether the GAN could remove site bias without losing clinical signal.
- Scanner Indistinguishability: Before correction, a Support Vector Machine (SVM) could identify the source scanner with 99.3% accuracy. After GAN correction, this accuracy dropped significantly, indicating that the site-specific "fingerprint" was successfully blurred.
- Clinical Classification: Applying GAN correction improved the accuracy of schizophrenia vs. healthy control classification by 3.7% over the baseline. It also outperformed traditional Gaussian Process regression by 1.2%.
- Gender Classification: On a balanced dataset, GAN correction improved gender classification accuracy by 15.8% compared to uncorrected data, suggesting that removing scanner noise allows biological signals to emerge more clearly.
- Reconstruction Accuracy: For the 11 subjects scanned at both sites, the GAN reduced the Mean Square Error (MSE) between the two different scanner images by 27.02%.
Discussion and Limitations
The primary insight of this work is that hardware artifacts can be treated as a "style" that can be stripped away. By using an unsupervised, non-linear approach, we successfully untangled site effects that linear regression often exacerbated.
However, limitations remain. The current model operates on 2D sagittal slices, which loses volumetric context. Future work will involve moving to 3D convolutional kernels, which, while computationally intensive, would provide a more holistic normalization. Furthermore, while the "black box" nature of the GAN is an advantage for ease of use, further research is needed to interpret exactly which image features (e.g., contrast, edge sharpness) the model is altering during the harmonization process.
Summary updated