Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Domain generalization in fetal brain MRI segmentation \\with multi-reconstruction augmentation

Priscille de Dumast, Meritxell Bach Cuadra

TL;DR

This work proposes to leverage the power of fetal brain MRI super-resolution (SR) reconstruction methods to generate multiple reconstructions of a single subject with different parameters, thus as an efficient tuning-free data augmentation strategy.

Abstract

Quantitative analysis of in utero human brain development is crucial for abnormal characterization. Magnetic resonance image (MRI) segmentation is therefore an asset for quantitative analysis. However, the development of automated segmentation methods is hampered by the scarce availability of fetal brain MRI annotated datasets and the limited variability within these cohorts. In this context, we propose to leverage the power of fetal brain MRI super-resolution (SR) reconstruction methods to generate multiple reconstructions of a single subject with different parameters, thus as an efficient tuning-free data augmentation strategy. Overall, the latter significantly improves the generalization of segmentation methods over SR pipelines.

Domain generalization in fetal brain MRI segmentation \\with multi-reconstruction augmentation

TL;DR

This work proposes to leverage the power of fetal brain MRI super-resolution (SR) reconstruction methods to generate multiple reconstructions of a single subject with different parameters, thus as an efficient tuning-free data augmentation strategy.

Abstract

Quantitative analysis of in utero human brain development is crucial for abnormal characterization. Magnetic resonance image (MRI) segmentation is therefore an asset for quantitative analysis. However, the development of automated segmentation methods is hampered by the scarce availability of fetal brain MRI annotated datasets and the limited variability within these cohorts. In this context, we propose to leverage the power of fetal brain MRI super-resolution (SR) reconstruction methods to generate multiple reconstructions of a single subject with different parameters, thus as an efficient tuning-free data augmentation strategy. Overall, the latter significantly improves the generalization of segmentation methods over SR pipelines.
Paper Structure (15 sections, 4 figures, 2 tables)

This paper contains 15 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials
  3. Clinical MR exams
  4. In-SR domain dataset -- CHUV-set
  5. Out-of-SR domain pure testing set -- FeTA-KCL
  6. Methodology
  7. SR reconstruction-based data augmentation
  8. Evaluation
  9. Model and training strategy
  10. Results
  11. Data augmentation
  12. Domain generalization
  13. Conclusion
  14. Compliance with ethical standards
  15. Acknowledgments

Figures (4)

  • Figure 1: Overall framework.
  • Figure 2: Illustration of the intensity variation depending on the regularization parameters $\lambda$ and $\alpha$ for MIALSRTK and NiftyMIC reconstructions on a 32 weeks of GA subject.
  • Figure 3: Sagittal view of a 27.8 weeks-old (top) and coronal view of the 33.1 weeks-old (bottom) fetal brain tissue segmentation obtained in the different configurations studied. White arrows and circles show representative areas where our multi-reconstruction approach improves the segmentation accuracy.
  • Figure 4: Mean DSC (plain) and ASSD (dashed) performance in the domain generalization task as a function of GA in weeks.