Normative volumetric growth modeling of the whole fetal body, placenta and amniotic fluid for 3-dimensional T2-weighted magnetic resonance imaging

Technical Summary: Normative Volumetric Growth Modeling of the Whole Fetal Body, Placenta, and Amniotic Fluid

Problem Statement
While fetal MRI offers superior soft-tissue contrast and true 3-D information compared to ultrasound, quantitative volumetric assessment of the fetus, placenta, and amniotic fluid remains rare in clinical practice. Current limitations include:

1. Labor-Intensive Manual Segmentation: Existing methods rely on manual segmentation of motion-corrupted 2-D image stacks, which is slow, operator-dependent, and error-prone, particularly for large structures like the amniotic cavity.
2. Lack of 3-D Automation: Deep learning approaches typically segment single structures on raw 2-D stacks, suffering from slice misalignment and limited accuracy due to through-plane motion. No existing method provides a reliable automated solution for simultaneous whole-uterus volumetry in 3-D reconstructed MRI.
3. Absence of Normative Data: There are no combined normative growth charts for fetal, placental, and amniotic fluid volumes derived specifically from healthy control pregnancies across multiple MRI field strengths. Clinicians lack standardized centiles and z-scores to interpret measured volumes or evaluate the inter-relationships of these compartments.

Methodology
The study developed an integrated automated pipeline for whole-uterus volumetry using 3-D T2-weighted (T2W) fetal MRI.

• Data Acquisition and Preprocessing:

  • Cohort: The study utilized 506 scans from 16–41 weeks gestational age (GA) acquired on 0.55T, 1.5T, and 3T MRI systems.
  • Groups: 357 healthy term controls (confirmed delivery ≥37 weeks), 86 preterm/high-risk cases (delivery ≤32 weeks, including PPROM), and 63 additional datasets for training.
  • Reconstruction: Deformable slice-to-volume 3-D reconstruction (DSVR) was applied to motion-corrupted T2W stacks using the SVRTK framework to generate high-resolution (1.2 mm isotropic) 3-D images, correcting for inter-slice misalignment.

• Automated Segmentation:

  • Architecture: A 3-D U-Net model was trained within the MONAI framework.
  • Training Data: 223 manually curated 3-D DSVR reconstructions spanning diverse field strengths, GA ranges, and anatomical variations (including anomalies and preterm cases).
  • Labels: The model segments four labels: whole fetus (head and body separated), placenta, and amniotic fluid. The umbilical cord is excluded from volumetry.
  • Ground Truth: Labels were generated via an iterative process combining existing DL models, manual segmentation in ITK-SNAP, and interleaved refinement.

• Normative Modeling and Analysis:

  • Model Derivation: The trained model was applied to 357 healthy term-control datasets. After visual inspection and minor manual refinements (<25% of cases), quadratic regression was used to generate normative growth curves (5th, 50th, 95th centiles) for volumes and z-scores.
  • Statistical Analysis: Analysis of Covariance (ANCOVA) assessed correlations between volumes, GA, maternal characteristics (height, weight, ethnicity), and birth weight.
  • Validation: The pipeline was tested on 43 independent datasets and evaluated on 95 longitudinal scans from 42 fetuses and 86 preterm pregnancies to assess clinical utility.

Key Contributions

1. First Integrated Pipeline: Introduction of the first automated deep-learning pipeline for simultaneous segmentation and volumetry of the fetus, placenta, and amniotic fluid in motion-corrected 3-D reconstructed T2W fetal MRI across 0.55T–3T field strengths.
2. Normative Growth Models: Generation of the first combined normative volumetric growth charts for all three intrauterine compartments derived from a large cohort (n=357) of confirmed healthy term pregnancies.
3. Clinical Reporting Tool: Development of a reporting tool that computes centiles and z-scores, generates structured HTML reports with segmentation visualizations, and calculates estimated fetal weight based on fetal volume.
4. Public Availability: All automated segmentation components, centile calculators, and reporting scripts are made available via the auto-SVRTK repository.

Results

• Segmentation Accuracy: The automated pipeline produced accurate 3-D labels with high qualitative scores (average >3.6/4). Minor local corrections (<1% volume difference) were required in less than 25% of the control cohort, primarily at the placenta-myometrium interface and around fetal limbs.
• Growth Trajectories:
  • Fetal Volume: Showed a strong positive association with GA, with variability increasing after 30 weeks. Male fetuses had significantly higher volumes than females.
  • Placental Volume: Increased significantly with GA and correlated positively with maternal height, weight, and fetal volume.
  • Amniotic Fluid: Followed a quadratic pattern with a mid-gestation peak and decline toward term, exhibiting high inter-subject variability.
  • Correlations: Volumes and centiles correlated strongly with birth weight. Fetal and placental growth rates were weakly correlated, while fetal growth showed a moderate negative correlation with amniotic fluid slope.
• Field Strength Effects: While general trends overlapped, ANCOVA revealed significant differences in fetal and placental volumes between 1.5T and 3T cohorts, likely driven by GA distribution and protocol factors rather than biological variation.
• Clinical Utility (Preterm vs. Term):
  • Preterm pregnancies (≤32 weeks) demonstrated significantly lower fetal, placental, and amniotic fluid volumes and centiles compared to term controls.
  • The pipeline successfully identified deviations in high-risk cases (e.g., PPROM), where amniotic fluid centiles were often <5th.
  • Longitudinal scans showed individual trajectories closely tracking normative curves for fetal and placental volumes, validating the model's applicability for monitoring growth.

Significance
The authors claim this study provides a robust, standardized framework for quantitative fetal MRI that overcomes the limitations of manual 2-D segmentation. By establishing the first combined normative models for the whole uterus, the work enables the calculation of z-scores and centiles, which are essential for determining whether measured volumes fall within expected ranges. The pipeline offers a practical tool for evaluating growth patterns in both normal and high-risk pregnancies, supporting early identification of abnormal developmental trajectories and serving as a foundation for future organ-specific volumetry and clinical integration.