Fetal Brain Tissue Annotation and Segmentation Challenge

Kelly Payette, Priscille de Dumast, András Jakab, Meritxell Bach Cuadra, Lana Vasung, Roxane Licandro, Bjoern Menze, Hongwei Li ZurichZenodo 2021 · DOI

Introduction

Understanding human brain development in utero is essential for diagnosing congenital disorders and predicting neurodevelopmental outcomes. However, analyzing fetal MRI is notoriously difficult due to the small size of the fetal brain, constant fetal movement, and rapid changes in anatomy and tissue contrast throughout gestation. Traditionally, clinicians have relied on manual segmentation—a slow, error-prone process that varies significantly between experts.

The Fetal Brain Tissue Annotation and Segmentation (FeTA) 2021 challenge was established to address these hurdles by providing the first international benchmark for automated multi-tissue fetal brain MRI segmentation. By creating a standardized, large-scale public dataset, the challenge aimed to identify the most robust deep learning architectures for this complex task.

The FeTA Dataset

The challenge utilized the FeTA Dataset, comprising 120 high-resolution 3D fetal MRI reconstructions (80 for training, 40 for testing). The dataset is unique in its inclusion of both neurotypical brains and those with pathologies such as ventriculomegaly or Chiari-II malformations, spanning a gestational age range of 20 to 35 weeks.

Distribution of subjects by gestational age and pathology across training and testing sets.

Experts manually segmented seven distinct tissue types: external cerebrospinal fluid (eCSF), grey matter (GM), white matter (WM), ventricles, cerebellum, brainstem, and deep grey matter. To mitigate motion artifacts, the dataset employed super-resolution algorithms to create clear 3D volumes from raw, blurry slices.

Methodology: The U-Net Standard

The challenge saw 21 algorithm submissions from 20 international teams, all of which utilized deep learning. This confirms that classical image processing is no longer competitive for complex medical segmentation tasks.

The U-Net architecture emerged as the undisputed gold standard, with 19 out of 21 teams employing some variation of this encoder-decoder structure. The top-performing team, NVAUTO, achieved the winning edge by implementing an asymmetrical U-Net. Unlike standard symmetric models, NVAUTO used a significantly larger encoder to extract complex features and a smaller, efficient decoder for reconstruction. Additionally, four of the top five teams utilized ensemble learning, combining multiple model predictions to minimize errors.

Key Results and Performance Plateau

Team NVAUTO secured first place with an average Dice Similarity Coefficient (DSC) of 0.786. While these results are promising, the challenge revealed a performance plateau; there was no statistically significant difference between the top 10–12 teams. This suggests that standard U-Net architectures may have reached their maximum potential given the current volume of training data.

Fetal brain MRI slices and corresponding multi-tissue segmentations across gestational ages 21 to 35 weeks.

Challenges in Segmentation

  • Pathology: Models performed significantly worse on pathological brains. The high variability in shape and structure in these cases makes generalization difficult for AI models trained on limited samples.
  • Fuzzy Boundaries: Cortical grey matter and the brainstem proved the hardest to segment due to partial volume effects, where boundaries appear "fuzzy" on MRI.
  • Clinical Utility: Despite these difficulties, most teams estimated total intracranial volume within 1% of the manual ground truth, indicating that these tools are nearing readiness for specific clinical applications.

Limitations and Future Directions

While the FeTA challenge represents a major milestone, several limitations remain. The current dataset is sourced from a single center in Zurich, which may introduce bias. Future iterations of the challenge will focus on multi-center data and a broader range of rare pathologies to ensure global robustness. Furthermore, addressing the "ground truth" problem—where human expert labels contain inherent noise—will be critical for pushing past the current performance ceiling. The results suggest that future breakthroughs may require novel asymmetrical architectures or significantly larger, more diverse datasets.

Summary updated

Abstract

Congenital disorders are one of the leading causes of infant mortality worldwide [1]. Recently, in-utero MRI of the fetal brain has started to emerge as a valuable tool for investigating the neurological development of fetuses with congenital disorders in order to aid in prenatal planning. Moreover, fetal MRI is a powerful tool to portray the complex neurodevelopmental events during human gestation, which remain to be completely characterized. Automated segmentation and quantification of the highly complex and rapidly changing brain morphology prior to birth in MRI data would improve the diagnostic process, as manual segmentation is both time consuming and subject to human error and inter-rater variability. It is clinically relevant to analyse information such as the shape or volume of the developing brain structures, as many congenital disorders cause subtle changes to these tissue compartments [2]–[5]. Existing growth data is mainly based on normally developing brains [6]–[8], and growth data for many pathologies and congenital disorders is lacking. The automatic segmentation of the developing human brain would be a first step in performing such an analysis. From a technical standpoint, there are many challenges that an automatic segmentation method of the fetal brain would need to overcome. The physiology and structures of the human brain are constantly growing and developing throughout gestation. In addition, the quality of the images can be poor due to fetal and maternal movement and imaging artefacts [9]. The boundary between tissues is often unclear due to partial volume effects. Furthermore, structures in a pathological fetal brain can have a different morphology than those in a nonpathological brain. This can make it challenging for an automatic method to recognize what the structures are. The field of fetal MRI has so far been understudied due to challenges in imaging and due to the lack of public, curated, and annotated ground truth data. The FeTA challenge is an important step in the development of reliable, valid, and reproducible methods of analyzing high resolution reconstructed MR images of the developing fetal brain from gestational week 21-33. Such new algorithms will have the potential to better understand the underlying causes of congenital disorders and ultimately to guide the development of antenatal/postnatal guidelines and clinical risk stratification tools for early interventions, treatments, and care management decisions. <strong>References</strong> [1] “WHO | Causes of child mortality,” WHO. http://www.who.int/gho/child_health/mortality/causes/en/ (accessed Jun. 07, 2020). [2] G. Egaña-Ugrinovic, M. Sanz-Cortes, F. Figueras, N. Bargalló, and E. Gratacós, “Differences in cortical development assessed by fetal MRI in late-onset intrauterine growth restriction,” Am. J. Obstet. Gynecol., vol. 209, no. 2, pp. 126.e1-126.e8, Aug. 2013, doi: 10.1016/j.ajog.2013.04.008. [3] A. Zugazaga Cortazar, C. Martín Martinez, C. Duran Feliubadalo, M. R. Bella Cueto, and L. Serra, “Magnetic resonance imaging in the prenatal diagnosis of neural tube defects,” Insights Imaging, vol. 4, no. 2, pp. 225–237, Mar. 2013, doi: 10.1007/s13244-013-0223-2. [4] C. Clouchoux et al., “Delayed Cortical Development in Fetuses with Complex Congenital Heart Disease,” Cereb. Cortex, vol. 23, no. 12, pp. 2932–2943, Dec. 2013, doi: 10.1093/cercor/bhs281. [5] C. K. Rollins et al., “Regional Brain Growth Trajectories in Fetuses with Congenital Heart Disease,” Ann. Neurol., no. n/a, doi: https://doi.org/10.1002/ana.25940. [6] D. Prayer et al., “MRI of normal fetal brain development,” Eur. J. Radiol., vol. 57, no. 2, pp. 199–216, Feb. 2006, doi: 10.1016/j.ejrad.2005.11.020. [7] D. A. Jarvis, C. R. Finney, and P. D. Griffiths, “Normative volume measurements of the fetal intra-cranial compartments using 3D volume in utero MR imaging,” Eur. Radiol., vol. 29, no. 7, pp. 3488–3495, Jul. 2019, doi: 10.1007/s00330-018-5938-5. [8] V. Kyriakopoulou et al., “Normative biometry of the fetal brain using magnetic resonance imaging,” Brain Struct. Funct., vol. 222, no. 5, pp. 2295–2307, 2017, doi: 10.1007/s00429-016-1342-6. [9] P de Dumast, P Deman, M Khawam, T Yu, S Tourbier, P Hagmann, P Maeder, JP Thiran, R Meuli, V Dunet, M Koob, M Bach Cuadra, “Translating fetal brain magnetic resonance image super-resolution into the clinical environment [abstract],” Marseille, Feb. 2020, vol. 05.
BibTeX
@inproceedings{payette2021fetal,
  author = {Kelly Payette and Priscille de Dumast and András Jakab and Meritxell Bach Cuadra and Lana Vasung and Roxane Licandro and Bjoern Menze and Hongwei Li Zurich},
  title = {{Fetal Brain Tissue Annotation and Segmentation Challenge}},
  booktitle = {Zenodo},
  year = {2021},
  doi = {10.5281/zenodo.4573144},
}
Figure