Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating analytical variability in fMRI results with style transfer

Elodie Germani, Camille Maumet, Elisa Fromont

TL;DR

This study tackles the reproducibility challenge in task-based fMRI by treating data-processing pipelines as a style domain and applying unpaired style-transfer methods to convert statistic maps between pipelines. It re-implements three GAN-based frameworks (Pix2Pix, CycleGAN, StarGAN) for 3D fMRI statistics and introduces a diffusion-model approach with latent-space conditioning via an auxiliary classifier to enable unsupervised multi-domain transfers. Results show GAN-based methods, especially Pix2Pix and StarGAN, achieve high fidelity transfers to target pipelines, while diffusion models offer greater diversity and can improve certain transfers when using latent conditioning and multiple targets. The work demonstrates that pipeline-induced variability can be mitigated through learned style transfer, enabling data augmentation and more robust mega-analyses on publicly shared neuroimaging data.

Abstract

We propose a novel approach to improve the reproducibility of neuroimaging results by converting statistic maps across different functional MRI pipelines. We make the assumption that pipelines used to compute fMRI statistic maps can be considered as a style component and we propose to use different generative models, among which, Generative Adversarial Networks (GAN) and Diffusion Models (DM) to convert statistic maps across different pipelines. We explore the performance of multiple GAN frameworks, and design a new DM framework for unsupervised multi-domain styletransfer. We constrain the generation of 3D fMRI statistic maps using the latent space of an auxiliary classifier that distinguishes statistic maps from different pipelines and extend traditional sampling techniques used in DM to improve the transition performance. Our experiments demonstrate that our proposed methods aresuccessful: pipelines can indeed be transferred as a style component, providing animportant source of data augmentation for future medical studies.

Mitigating analytical variability in fMRI results with style transfer

TL;DR

This study tackles the reproducibility challenge in task-based fMRI by treating data-processing pipelines as a style domain and applying unpaired style-transfer methods to convert statistic maps between pipelines. It re-implements three GAN-based frameworks (Pix2Pix, CycleGAN, StarGAN) for 3D fMRI statistics and introduces a diffusion-model approach with latent-space conditioning via an auxiliary classifier to enable unsupervised multi-domain transfers. Results show GAN-based methods, especially Pix2Pix and StarGAN, achieve high fidelity transfers to target pipelines, while diffusion models offer greater diversity and can improve certain transfers when using latent conditioning and multiple targets. The work demonstrates that pipeline-induced variability can be mitigated through learned style transfer, enabling data augmentation and more robust mega-analyses on publicly shared neuroimaging data.

Abstract

We propose a novel approach to improve the reproducibility of neuroimaging results by converting statistic maps across different functional MRI pipelines. We make the assumption that pipelines used to compute fMRI statistic maps can be considered as a style component and we propose to use different generative models, among which, Generative Adversarial Networks (GAN) and Diffusion Models (DM) to convert statistic maps across different pipelines. We explore the performance of multiple GAN frameworks, and design a new DM framework for unsupervised multi-domain styletransfer. We constrain the generation of 3D fMRI statistic maps using the latent space of an auxiliary classifier that distinguishes statistic maps from different pipelines and extend traditional sampling techniques used in DM to improve the transition performance. Our experiments demonstrate that our proposed methods aresuccessful: pipelines can indeed be transferred as a style component, providing animportant source of data augmentation for future medical studies.
Paper Structure (5 sections, 4 figures, 3 tables)

This paper contains 5 sections, 4 figures, 3 tables.

Table of Contents

  1. Architecture and training
  2. Source content preservation
  3. Classifier conditioning
  4. Multi-target images
  5. Architecture and training

Figures (4)

  • Figure 1: Generated images for two transfers and different competitors: Pix2Pix isola_image--image_2018, CycleGAN zhu_unpaired_2020 and starGAN choi_stargan_2018. Correlation with ground truth are indicated above each image (in percent).
  • Figure 2: Generated images for two transfer and different competitors: conditioning with one-hot encoding ho_classifier-free_2022, with a classifier-conditioning N=1 and N=10 target images randomly selected. Correlations with ground truth are indicated above generated and source images (in percent).
  • Figure :
  • Figure :