Cross-center generalizability of fetal brain MRI segmentation algorithms

Establish robust domain generalizability of automatic multi-class fetal brain MRI segmentation algorithms across different imaging centers that use varying scanners, acquisition parameters, and super-resolution reconstruction methods, to enable reliable real-world clinical applicability.

Background

FeTA 2021 evaluated fetal brain tissue segmentation algorithms in a single-center setting. The 2022 FeTA challenge expanded to a multi-center design to explicitly test generalizability across institutions, scanners, and reconstruction pipelines. Domain shifts between centers are known to affect deep learning performance, and overcoming these shifts is essential for clinical deployment.

The paper emphasizes that differences in acquisition parameters, field strength, coil type, reconstruction methods, and operator expertise cause significant variability in image appearance, which undermines segmentation models trained on limited domains.

References

However, FeTA 2021 was a single center study, and the generalizability of algorithms across different imaging centers remains unsolved, limiting real-world clinical applicability.

— Multi-Center Fetal Brain Tissue Annotation (FeTA) Challenge 2022 Results  (2402.09463 - Payette et al., 2024) in Abstract

This study evaluated one tumor type using a single-institution cohort acquired exclusively on GE scanners; robustness across institutions, vendors, and imaging protocols remains unverified.

— Parameter-Efficient pretrained-CT-to-MRI Transfer for Rectal Cancer Segmentation: Performance-Calibration Trade-offs  (2608.27178 - Rangnekar et al., 27 Aug 2026) in Discussion and conclusions, Section 4