Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Augment to Augment: Diverse Augmentations Enable Competitive Ultra-Low-Field MRI Enhancement

Felix F Zimmermann

TL;DR

The paper tackles enhancing ultra-low-field MRI by translating low-field contrasts to high-field appearances under stringent data constraints (50 paired volumes). It introduces a FiLM-conditioned, multi-contrast 3D U-Net trained with a hybrid reconstruction–adversarial objective and reinforced by diverse, auxiliary high-field tasks and robust data augmentations. The auxiliary contrast-synthesis and restoration tasks, together with a strong augmentation strategy, significantly improve fidelity in limited-data settings. On the ULF-EnC dataset, the approach achieves competitive rankings (3rd by brain-masked SSIM on public validation and 4th on the final test) with code available for reproducibility.

Abstract

Ultra-low-field (ULF) MRI promises broader accessibility but suffers from low signal-to-noise ratio (SNR), reduced spatial resolution, and contrasts that deviate from high-field standards. Image-to-image translation can map ULF images to a high-field appearance, yet efficacy is limited by scarce paired training data. Working within the ULF-EnC challenge constraints (50 paired 3D volumes; no external data), we study how task-adapted data augmentations impact a standard deep model for ULF image enhancement. We show that strong, diverse augmentations, including auxiliary tasks on high-field data, substantially improve fidelity. Our submission ranked third by brain-masked SSIM on the public validation leaderboard and fourth by the official score on the final test leaderboard. Code is available at https://github.com/fzimmermann89/low-field-enhancement.

Augment to Augment: Diverse Augmentations Enable Competitive Ultra-Low-Field MRI Enhancement

TL;DR

The paper tackles enhancing ultra-low-field MRI by translating low-field contrasts to high-field appearances under stringent data constraints (50 paired volumes). It introduces a FiLM-conditioned, multi-contrast 3D U-Net trained with a hybrid reconstruction–adversarial objective and reinforced by diverse, auxiliary high-field tasks and robust data augmentations. The auxiliary contrast-synthesis and restoration tasks, together with a strong augmentation strategy, significantly improve fidelity in limited-data settings. On the ULF-EnC dataset, the approach achieves competitive rankings (3rd by brain-masked SSIM on public validation and 4th on the final test) with code available for reproducibility.

Abstract

Ultra-low-field (ULF) MRI promises broader accessibility but suffers from low signal-to-noise ratio (SNR), reduced spatial resolution, and contrasts that deviate from high-field standards. Image-to-image translation can map ULF images to a high-field appearance, yet efficacy is limited by scarce paired training data. Working within the ULF-EnC challenge constraints (50 paired 3D volumes; no external data), we study how task-adapted data augmentations impact a standard deep model for ULF image enhancement. We show that strong, diverse augmentations, including auxiliary tasks on high-field data, substantially improve fidelity. Our submission ranked third by brain-masked SSIM on the public validation leaderboard and fourth by the official score on the final test leaderboard. Code is available at https://github.com/fzimmermann89/low-field-enhancement.

Paper Structure

This paper contains 15 sections, 1 equation, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Network Architecture
  4. Training Objective
  5. Auxiliary Tasks
  6. Auxiliary Tasks
  7. Data Augmentations
  8. Geometric Augmentations
  9. Intensity augmentations
  10. Input degradations
  11. Results
  12. Ablation Study
  13. Discussion
  14. Acknowledgments.
  15. Disclosure of Interests.

Figures (2)

  • Figure 1: Schematic overview of the proposed method. A single 3D U-Net is trained on three tasks. The primary task (top) is to translate the three low-field contrasts into a target high-field contrast, specified via FiLM conditioning. Two auxiliary tasks augment the training by leveraging only the high-field data: (1) contrast synthesis and (2) image restoration (denoising/deblurring).
  • Figure 2: Exemplary result of the central axial slice of a held-out training sample. Given FLAIR (top), T1-weighted (center), and T2-weighted (bottom) input volumes (left), our trained network predicts 3 T-like image volumes (right), closely matching the ground truth targets (center).