Source-linked AI summary
An automatic multi-tissue human fetal brain segmentation benchmark using the Fetal Tissue Annotation Dataset
Kelly Payette, Priscille de Dumast, Hamza Kebiri, Ivan Ezhov, Johannes C. Paetzold, Suprosanna Shit, Asim Iqbal, Romesa Khan, Raimund Kottke, Patrice Grehten, Hui Ji, Levente Lanczi, Marianna Nagy, Monika Beresova, Thi Dao Nguyen, Giancarlo Natalucci, Theofanis Karayannis, Bjoern Menze, Meritxell Bach Cuadra, Andras Jakab
TL;DR
Automatic fetal brain segmentation is limited by the scarcity of accessible, curated, annotated data despite its importance for quantitative analysis of normal and pathological development. The paper introduces FeTA, evaluates annotation variability and segmentation approaches, and releases a dataset spanning seven tissue classes. Variability was higher for poorly defined tissues in low-quality images, while the dataset provides training and validation resources for future algorithms.
Problem
Automatic fetal brain analysis is limited by the lack of accessible, curated, annotated data for multi-tissue segmentation.
Method
The paper constructs FeTA from reconstructed fetal MRI volumes with seven manual tissue labels, evaluates annotator variability, and presents automatic segmentation approaches using 2D/3D U-Nets and SeBRe.
Results
External-CSF DSC was 0.13±0.11 and GM DSC was 0.45±0.10 in low-quality images, compared with 0.88±0.02 and 0.71±0.03, respectively, in high-quality volumes.
Takeaways & Limitations
FeTA supplies 40 segmented volumes and 10 unsegmented validation volumes to support development and evaluation of automatic fetal brain segmentation algorithms.
Takeaways & Limitations
The dataset uses single segmentations for most cases, and only nine volumes received multiple annotations, limiting conclusions about inter-annotator variability.
Abstract
from arXiv · showhide
It is critical to quantitatively analyse the developing human fetal brain in order to fully understand neurodevelopment in both normal fetuses and those with congenital disorders. To facilitate this analysis, automatic multi-tissue fetal brain segmentation algorithms are needed, which in turn requires open databases of segmented fetal brains. Here we introduce a publicly available database of 50 manually segmented pathological and non-pathological fetal magnetic resonance brain volume reconstructions across a range of gestational ages (20 to 33 weeks) into 7 different tissue categories (external cerebrospinal fluid, grey matter, white matter, ventricles, cerebellum, deep grey matter, brainstem/spinal cord). In addition, we quantitatively evaluate the accuracy of several automatic multi-tissue segmentation algorithms of the developing human fetal brain. Four research groups participated, submitting a total of 10 algorithms, demonstrating the benefits the database for the development of automatic algorithms.
Background & Summary
The paper introduces FeTA, a publicly available dataset of reconstructed fetal brain MRI volumes with seven manually segmented tissue classes, spanning normal and pathological brains across 20–33 gestational weeks. It also examines annotation variability and releases the data to support automatic multi-tissue segmentation.
- Motivation: The dataset addresses limited accessible, curated, annotated fetal MRI data needed to develop automatic multi-class segmentation methods.Existing approaches were constrained by normal-fetus atlases, low-resolution directional scans, or single-class segmentation.
- Dataset contribution: FeTA provides reconstructed fetal brain volumes and manual segmentations into seven tissue classes for normal and pathological brains aged 20–33 gestational weeks.The classes are external CSF, grey matter, white matter, ventricles, cerebellum, deep grey matter, and brainstem/spinal cord.
- Data collection: 50 pregnant women contributed fetal MRI data, including 18 non-pathological and 32 pathological brains, with an average gestational age of 26.9 ± 3.3 weeks.The pathological group consisted of fetuses with spina bifida before or after fetal spinal lesion repair surgery.
- Image processing: SR reconstructions were assembled from multiple scans, whose quality was independently rated from 1 to 3 because fetal motion produces variable image quality.A rating of 1 was unusable for segmentation, 2 was usable despite some blurring, and 3 was excellent.
- Annotation variability: Annotator variability was substantial for poorly defined tissues in low-quality images, with external-CSF DSC 0.13±0.11 and GM DSC 0.45±0.10.High-quality volumes showed external-CSF DSC 0.88±0.02 and GM DSC 0.71±0.03; the authors recommend investigating consensus delineations in future releases.
- Data release: The release contains 40 reconstructed volumes with corresponding manual segmentations, 10 additional unsegmented volumes for validation, annotation guidelines, and segmentation-generation code.The data are organized in BIDS format and are available through the FeTA dataset on Synapse.
Technical Validation
The benchmark evaluates 10 automatic multi-tissue fetal brain segmentation algorithms developed by four research groups using the FeTA dataset. Results compare neural-network, multi-atlas, orientation-specific, and combined approaches while documenting tissue-specific challenges and dataset limitations.
- Methods: The benchmark included four research groups and 10 algorithms, spanning multi-atlas registration, 2D and 3D U-Nets, orientation fusion, and Mask R-CNN-based instance segmentation.The multi-atlas pipeline selected atlas candidates by gestational-age difference and reconstruction quality, followed by registration and label fusion.
- Results: External CSF and grey matter were consistently difficult to segment because tissue boundaries, cortical morphology, and contrast change across pathology and gestation.Deep grey matter was also difficult to define in subthalamic and hypothalamic regions, while other structures showed no consistent cross-algorithm error pattern.
- Results: Multi-atlas segmentation performed well, particularly on poor-quality super-resolution volumes, where atlas shape priors can compensate for unclear structures.The method also scored highly on HD95 and VS metrics.
- Results: Except for poor-quality volumes, the Kispi 2D U-Net outperformed its 3D counterpart, partly because 2D slicing produced more training samples than 3D patching.The 2D setup yielded 256 samples per volume, compared with 64 patches for the 3D setup.
- Results: The combined IBBM method ranked first, and its combined segmentations outperformed individual orientations in all cases.Axial and coronal directions performed well, whereas the sagittal version performed less well.
- Limitations: The dataset provides manually annotated multi-class super-resolution volumes but is limited by the chosen reconstruction method, variable image quality, and repeated annotations for only nine volumes.The authors plan alternative reconstructions and broader pathology, gestational-age, and non-pathological coverage.
Usage Notes
The paper publicly releases the dataset used in its analysis and invites fully automated methods to participate in a fetal brain segmentation challenge.
- The challenge requires fully automated segmentation methods using only fetal brain super-resolution volumes as input.
- The released dataset and challenge aim to encourage medical image analysis and machine learning research on fetal brain segmentation.
- Developed methods can support quantitative analysis of fetal brain growth in pathological and non-pathological cases.
Code Availability
The paper provides dataset-development code through Synapse and presents algorithm metrics, rankings, and subgroup evaluations in Figures 5–7.
- The dataset-development code is available on Synapse under the Fetal Tissue Annotation Challenge FeTA Dataset.
- Figure 5 summarizes DSC, HD95, and VS across all tissue labels and ranks the evaluated algorithms.
- Figure 6 compares algorithm performance across excellent, good, and poor-quality super-resolution volumes, including quality-specific rankings.
- Figure 7 separates algorithm evaluations and rankings for pathological and non-pathological super-resolution volumes.
Supplementary Information Overview of each algorithm by label
The supplementary information reports evaluation metrics separately by tissue label and documents the label terminology and annotation guidance.
- Supplementary analyses display evaluation metrics for each algorithm separated by label number.
- The supplementary label notation identifies CSF as external cerebrospinal fluid, GM as grey matter, and WM as white matter.
- Annotation guidelines for creating the label maps are available in the project database on Synapse.