3D Wasserstein generative adversarial network with dense U-Net based discriminator for preclinical fMRI denoising

TL;DR

The paper tackles the challenge of denoising preclinical fMRI data, which suffer from high noise, variable brain geometry, and limited SNR. It introduces 3D U-WGAN, a GAN framework with a dense 3D U-Net discriminator and a 4D patch-based data configuration to jointly leverage temporal and spatial information, guided by a composite loss that includes MSE, perceptual, and adversarial terms. Across CTNI rsfMRI and task fMRI datasets, the method achieves superior PSNR and SSIM, preserves important structural details, and improves downstream analyses such as phantom activation recovery and seed-based functional connectivity, with favorable computational efficiency. The work demonstrates robust denoising performance against state-of-the-art methods and lays groundwork for extending to additional artifacts in preclinical neuroimaging.

Abstract

Functional magnetic resonance imaging (fMRI) is extensively used in clinical and preclinical settings to study brain function, however, fMRI data is inherently noisy due to physiological processes, hardware, and external noise. Denoising is one of the main preprocessing steps in any fMRI analysis pipeline. This process is challenging in preclinical data in comparison to clinical data due to variations in brain geometry, image resolution, and low signal-to-noise ratios. In this paper, we propose a structure-preserved algorithm based on a 3D Wasserstein generative adversarial network with a 3D dense U-net based discriminator called, 3D U-WGAN. We apply a 4D data configuration to effectively denoise temporal and spatial information in analyzing preclinical fMRI data. GAN-based denoising methods often utilize a discriminator to identify significant differences between denoised and noise-free images, focusing on global or local features. To refine the fMRI denoising model, our method employs a 3D dense U-Net discriminator to learn both global and local distinctions. To tackle potential over-smoothing, we introduce an adversarial loss and enhance perceptual similarity by measuring feature space distances. Experiments illustrate that 3D U-WGAN significantly improves image quality in resting-state and task preclinical fMRI data, enhancing signal-to-noise ratio without introducing excessive structural changes in existing methods. The proposed method outperforms state-of-the-art methods when applied to simulated and real data in a fMRI analysis pipeline.

Figures (11)

• Figure 1: Overall framework of 3D U-WGAN for fMRI denoising. (a) Denoising process: the generator generates denoised fMRI. (b) The dense U-Net based discriminator provides both general structure and local feedback to the generator.
• Figure 2: Encoder-decoder based generator architecture.
• Figure 3: 3D dense U-Net based discriminator architecture.
• Figure 4: 3D input patches from 4D fMRI data by a) concatenating volumetric slices (slice-based), b) concatenating time sequences (time-based).
• Figure 5: Denoised example from CTNI rsfMRI DB (a) Noise-free image, (b) Noisy image (noise level: 9%), (c) BM4D maggioni2012nonlocal, (d) DU-GAN huang2021gan, (e) RIRGAN yu2023rirgan, (f) 3D U-WGAN$_{xyt}$ (ours).