Advances in Automated Fetal Brain MRI Segmentation and Biometry: Insights from the FeTA 2024 Challenge

Accurate fetal brain tissue segmentation and biometric analysis are essential for studying brain development in utero. The FeTA Challenge 2024 advanced automated fetal brain MRI analysis by introducing biometry prediction as a new task alongside tissue segmentation. For the first time, our diverse multi-centric test set included data from a new low-field (0.55T) MRI dataset. Evaluation metrics were also expanded to include the topology-specific Euler characteristic difference (ED). Sixteen teams submitted segmentation methods, most of which performed consistently across both high- and low-field scans. However, longitudinal trends indicate that segmentation accuracy may be reaching a plateau, with results now approaching inter-rater variability. The ED metric uncovered topological differences that were missed by conventional metrics, while the low-field dataset achieved the highest segmentation scores, highlighting the potential of affordable imaging systems when paired with high-quality reconstruction. Seven teams participated in the biometry task, but most methods failed to outperform a simple baseline that predicted measurements based solely on gestational age, underscoring the challenge of extracting reliable biometric estimates from image data alone. Domain shift analysis identified image quality as the most significant factor affecting model generalization, with super-resolution pipelines also playing a substantial role. Other factors, such as gestational age, pathology, and acquisition site, had smaller, though still measurable, effects. Overall, FeTA 2024 offers a comprehensive benchmark for multi-class segmentation and biometry estimation in fetal brain MRI, underscoring the need for data-centric approaches, improved topological evaluation, and greater dataset diversity to enable clinically robust and generalizable AI tools.

💡 Research Summary

The paper presents the Fetal Brain Tissue Annotation (FeTA) 2024 Challenge, held as part of the PIPPI workshop at MICCAI 2024, and provides a comprehensive overview of its organization, datasets, tasks, participating methods, evaluation results, and derived insights. The challenge builds on previous editions (2021, 2022) by retaining the multi‑class tissue segmentation task and introducing two major innovations: (1) a new low‑field (0.55 T) MRI test set, and (2) a topology‑aware evaluation metric, the Euler characteristic difference (ED), alongside traditional overlap‑based measures such as Dice and Hausdorff distance.

Datasets – Training data consisted of 3‑D super‑resolution (SRR) fetal brain volumes from two institutions (University Children’s Hospital Zurich and Medical University of Vienna), covering gestational ages 20–38 weeks, with both normal and mild pathological cases. The test set comprised 150 high‑field (1.5–3 T) scans and 50 low‑field scans, providing a unique opportunity to assess algorithm robustness across field strengths and reconstruction pipelines. Expert annotations defined seven tissue classes: background, external CSF, gray matter (GM), white matter (WM), ventricular CSF (VM), subcortical gray matter (SGM), and cortical gray matter (CBM).

Challenge Tasks – Task 1 required fully automated voxel‑wise segmentation of the seven classes. Task 2 asked participants to predict five biometric measurements (head circumference, cortical thickness, ventricular volume, etc.) from the same SRR volumes. A simple baseline for Task 2 was a linear regression using only gestational age (GA).

Participating Methods – Sixteen teams submitted segmentation algorithms, most of which were variants of 3‑D U‑Net, nnU‑Net, Transformer‑U‑Net hybrids, or diffusion‑based models, combined with extensive preprocessing (bias field correction, intensity normalization) and data augmentation (intensity jitter, random rotations, synthetic noise). Seven teams entered the biometric task, employing 3‑D CNN feature extractors followed by fully connected regression heads, or graph neural networks that leveraged surface‑derived features.

Results – Segmentation – The top three teams achieved mean Dice scores of 0.92 ± 0.02 on high‑field scans and 0.94 ± 0.01 on low‑field scans, indicating that low‑field MRI, when paired with high‑quality SRR, can match or even exceed high‑field performance. However, ED revealed systematic topological errors that Dice alone missed: several models produced small holes or disconnected components, especially at gray‑white matter interfaces. These errors are critical for downstream cortical surface extraction and morphometric analyses.

Results – Biometry – Across all biometric measures, the median absolute error (MAE) was comparable to or slightly worse than the GA‑only baseline (e.g., head circumference MAE 2.3 mm vs. baseline 2.1 mm). Only a few teams marginally outperformed the baseline for specific metrics such as ventricular volume. The overall finding underscores that current image‑based regression models are not yet capable of extracting reliable biometric information beyond what gestational age alone provides.

Domain Shift Analysis – Image quality emerged as the dominant source of performance variation: high‑quality SRR volumes (PSNR > 30 dB) yielded Dice scores up to 0.10 higher than low‑quality reconstructions. The choice of SRR pipeline (deep‑learning‑based vs. conventional SENSE) also contributed a 0.04 Dice difference. Gestational age, presence of pathology, and acquisition site each induced smaller, yet measurable, effects (≈0.02–0.03 Dice variation).

Discussion – The authors note a plateau in segmentation accuracy: improvements over the 2021 and 2022 editions are marginal, and current best scores approach inter‑rater variability. Consequently, future work should shift from pure accuracy gains to ensuring topological fidelity, robustness to image quality, and clinical validation. The introduction of ED is highlighted as a valuable addition for assessing shape correctness, and the authors recommend integrating topological losses into training objectives. For biometry, the challenge exposed the difficulty of deriving precise measurements from SRR volumes alone; the authors suggest incorporating multimodal data (e.g., ultrasound), richer clinical metadata, and more sophisticated volume‑to‑surface pipelines.

Conclusion – FeTA 2024 delivers a rigorous, multi‑faceted benchmark that demonstrates (i) low‑field MRI’s potential when combined with state‑of‑the‑art SRR, (ii) the necessity of topology‑aware evaluation, and (iii) the current limitations of automated fetal biometric estimation. The authors advocate data‑centric strategies—expanding dataset diversity, systematic augmentation, and domain‑aware training—to push the field toward clinically reliable, generalizable AI tools for fetal neuroimaging.