Standardizing brain magnetic resonance imaging usin generative adversarial networks: A multisite study approach

Abstract

Background: Magnetic resonance imaging (MRI) intensities vary across sites due to differences in acquisition protocols and hardware. Resolution also differs across centers. This hampers developing multisite deep learning models on MRI data. Objective: To standardize MRI intensities and resolution to enable multisite deep learning. Materials and Methods: T2-weighted brain MRI from 500 subjects across sites were split into training, validation and test sets. A generative adversarial network (GAN) model was developed to convert 64x64 low-resolution inputs to 256x256 standardized outputs. Preprocessing involved skull stripping, interpolation and intensity scaling. The generator used convolutional layers and residual blocks. Discriminator classified real/fake images. VGG perceptual loss was incorporated along with MSE and adversarial losses. Results: The GAN model achieved a structural similarity index of 0.9937 and feature similarity of 0.00122 versus ground truth. Intensity distribution was retained. The proposed pipeline reduced interpolation noise by 94% in extracted features. Conclusion: The proposed GAN pipeline can effectively standardize multisite brain MRI for intensity and resolution. By enabling multi-center data harmonization, this approach facilitates developing deep learning models through federated learning on MRI big data.