Synthesizing pseudo-T2w images to recapture missing data in neonatal neuroimaging with applications in rs-fMRI

Sydney Kaplan, Anders Perrone, Dimitrios Alexopoulos, Jeanette K. Kenley, Deanna M. Barch, Claudia Buss, Jed T. Elison, Alice M. Graham, Jeffrey J. Neil, Thomas G. O'Connor, Jerod Rasmussen, Monica Rosenberg, Cynthia E. Rogers, Aristeidis Sotiras, Damien A. Fair, Christopher D. SmyserNeuroImage 2022 · DOI · OpenReview

TL;DR

This paper introduces two novel methods, a deep learning approach (3DGAN-T2w) and a registration-based method (Kaplan-T2w), to generate synthetic T2-weighted (T2w) MRI images from T1-weighted (T1w) images in neonates. These methods effectively address the challenge of missing or low-quality structural MRI data in infant studies, demonstrating high anatomical accuracy and enabling comparable functional connectivity analyses to those using original T2w images.

What Problem Does This Paper Address?

Acquiring high-quality structural MRI data, specifically T1-weighted (T1w) and T2-weighted (T2w) images, is crucial for detailed analysis of brain development in neonates. However, this population is particularly challenging to scan due to movement artifacts and lower tissue contrast compared to adults. This often leads to missing or unusable structural scans, resulting in subject data loss. While synthetic imaging techniques have shown promise in adult populations, their efficacy and impact on standard neuroimaging pipelines in neonates, especially for resting-state functional MRI (rs-fMRI), remained underexplored. This paper aims to fill this gap by developing and validating methods to synthesize T2w images from T1w images, thereby salvaging valuable neonatal neuroimaging data.

Methods

The authors propose two distinct methods for synthesizing pseudo-T2w images:

  • 3DGAN-T2w (Deep Learning): This method extends existing 2D Generative Adversarial Network (GAN) architectures (specifically CycleGAN) to a 3D volumetric approach. This addresses the banding artifacts often seen in 2D slice-by-slice synthesis. The network learns a non-linear mapping between T1w and T2w images without requiring paired training data. It utilizes two generators and two discriminators, trained with adversarial and cycle-consistency losses to ensure realistic image generation and accurate reconstruction of the original modality.
  • Kaplan-T2w (Registration-Based): This method leverages state-of-the-art non-linear image registration and multi-atlas fusion techniques. It requires a smaller "bank" of high-quality, age-specific paired T1w and T2w reference images. The core idea is to register each reference T1w image to the target individual's T1w image, apply these transformations to the corresponding reference T2w images, and then fuse these transformed T2w images to create a synthetic T2w image that is anatomically aligned with the target T1w. This approach is computationally lighter than deep learning methods and does not require large training datasets.

Datasets

  • eLABE (Early Life Adversity Biological Embedding) Study: This was the primary dataset used for training and evaluating the pseudo-T2w generation methods. It included MRI data from 127 neonates. A subset of 107 neonates was used for training the models, and 20 were reserved for primary analyses.
  • ECHO (Environmental Influences on Child Health Outcomes) Program: This nationwide study provided a replication cohort of 10 infants, used for supplemental analyses to validate the findings.

No new benchmarks were introduced; the study focuses on applying and validating existing methodologies in a novel context.

Key Insights and Ideas

  • Bridging the Gap for Neonatal Neuroimaging: The primary insight is that advanced synthetic imaging techniques, previously explored in adults, are not only feasible but also highly effective for neonatal neuroimaging, a population with unique scanning challenges.
  • 3D GANs for Volumetric Data: Extending 2D GANs to 3D is crucial for avoiding artifacts and achieving clinically relevant volumetric image synthesis in medical imaging.
  • Registration-Based Methods as a Viable Alternative: The Kaplan-T2w method demonstrates that even with a smaller set of reference data, sophisticated registration and fusion techniques can yield high-quality synthetic images, offering a valuable alternative when large training datasets are unavailable.
  • Functional Equivalence: A critical finding is that pseudo-T2w images, when used in standard rs-fMRI preprocessing pipelines, yield functional connectivity results that are virtually indistinguishable from those obtained with original T2w images. This strongly supports their utility for data salvage.
  • Anatomical and Contrast Fidelity: Both methods successfully generate pseudo-T2w images that maintain high anatomical accuracy and comparable contrast properties to real T2w images, essential for downstream processing tasks like segmentation and registration.

Key Results

  • Anatomical Accuracy:
    • Mean Absolute Error (MAE) for Kaplan-T2w was 6.9 ± 0.9%, and for 3DGAN-T2w was 5.6 ± 1.1%.
  • Mean Structural Similarity Index (MSSIM) was 0.72±0.04 for Kaplan-T2w and 0.79±0.04 for 3DGAN-T2w.
  • Cortical ribbon DICE coefficients were 0.76±0.03 for Kaplan-T2w and 0.82±0.03 for 3DGAN-T2w.
  • Contrast Properties: No significant differences in Contrast-to-Noise Ratio (CNR) were found between T2w and pseudo-T2w images (p > 0.15 for all comparisons).
  • Registration Accuracy: Mutual Information (MI) values for BOLD-to-anatomical and anatomical-to-atlas registrations showed substantial overlap between T2w and pseudo-T2w images, indicating comparable registration quality.
  • fMRI Connectivity:
    • Brain-wide functional dense connectomes (dconns) showed high correlation (r = 0.98) between T2w and pseudo-T2w processed data.
  • Average connectivity matrices and variance across subjects were highly consistent between T2w and pseudo-T2w processed data.
  • Statistical differences in functional connectivity (FC) estimates were minimal (0.06% for Kaplan-T2w and 0.03% for 3DGAN-T2w).

Limitations and Future Work

  • Limitations:
    • The study was limited to healthy, term-born neonates. Applicability to premature, injured, or atypically developing brains needs further investigation.
  • The parcellation schemes used for fMRI analysis were derived from adult data, which may not perfectly capture developing neonatal brain networks.
  • Both methods relied on manually drawn brain masks, which is time-consuming.
  • Future Work:
    • Extend the methods to premature infants and other clinical populations.
  • Develop neonatal-specific brain atlases for more accurate fMRI parcellation.
  • Adapt the methods to work with unmasked data, improving efficiency. For 3DGAN-T2w, this involves enhancing memory efficiency for larger inputs. For Kaplan-T2w, it requires developing unmasked image registration for automatic brain-skull boundary delineation.
  • Investigate the impact of different age ranges and potential contrast variations on model performance.

Summary updated

Abstract

T1- and T2-weighted (T1w and T2w) images are essential for tissue classification and anatomical localization in Magnetic Resonance Imaging (MRI) analyses. However, these anatomical data can be challenging to acquire in non-sedated neonatal cohorts, which are prone to high amplitude movement and display lower tissue contrast than adults. As a result, one of these modalities may be missing or of such poor quality that they cannot be used for accurate image processing, resulting in subject loss. While recent literature attempts to overcome these issues in adult populations using synthetic imaging approaches, evaluation of the efficacy of these methods in pediatric populations and the impact of these techniques in conventional MR analyses has not been performed. In this work, we present two novel methods to generate pseudo-T2w images: the first is based in deep learning and expands upon previous models to 3D imaging without the requirement of paired data, the second is based in nonlinear multi-atlas registration providing a computationally lightweight alternative. We demonstrate the anatomical accuracy of pseudo-T2w images and their efficacy in existing MR processing pipelines in two independent neonatal cohorts. Critically, we show that implementing these pseudo-T2w methods in resting-state functional MRI analyses produces virtually identical functional connectivity results when compared to those resulting from T2w images, confirming their utility in infant MRI studies for salvaging otherwise lost subject data.
BibTeX
@article{kaplan2022synthesizing,
  author = {Sydney Kaplan and Anders Perrone and Dimitrios Alexopoulos and Jeanette K. Kenley and Deanna M. Barch and Claudia Buss and Jed T. Elison and Alice M. Graham and Jeffrey J. Neil and Thomas G. O'Connor and Jerod Rasmussen and Monica Rosenberg and Cynthia E. Rogers and Aristeidis Sotiras and Damien A. Fair and Christopher D. Smyser},
  title = {{Synthesizing pseudo-T2w images to recapture missing data in neonatal neuroimaging with applications in rs-fMRI}},
  journal = {NeuroImage},
  year = {2022},
  doi = {10.1016/j.neuroimage.2022.119091},
}
Figure