448 Deformable Medial Modeling to Generate Novel Shape Features of the Placenta Using Automated versus Manual Segmentations

Gabriel Arenas, Alison M. Pouch, İpek Oğuz, Baris Oguz, Brett Byram, Xing Yao, Nadav SchwartzJournal of Clinical and Translational Science 2024 · DOI

TL;DR

Researchers have developed an automated pipeline using Convolutional Neural Networks (CNNs) and "deformable medial modeling" to map the complex 3D geometry of the placenta in early pregnancy. This approach achieves high agreement with manual expert tracing, opening the door for using placental shape—not just size—as a potential biomarker for pregnancy health.

What Problem Does This Paper Address?

The placenta is the "black box" of pregnancy; its function is critical for fetal development, yet we have limited ways to non-invasively assess its health in real-time. While 3D ultrasound allows us to see the placenta, traditional clinical assessments often rely on simple measurements like volume or thickness at a single point.

However, the placenta's overall shape and how it attaches to the uterus (the maternal surface) versus how it connects to the fetus (the chorionic surface) may hold vital clues about its efficiency. The challenge is that manually "segmenting" (tracing) the placenta in 3D ultrasound is incredibly time-consuming and prone to human error. This study addresses the need for an automated, high-fidelity way to extract complex shape features that could eventually predict pregnancy complications.

Methods

The researchers compared two different ways of processing 3D ultrasound data from 20 singleton pregnancies (at 11–14 weeks gestation):

  1. Manual Segmentation: Experts hand-traced the placental boundaries.
  2. Automated Segmentation: A trained Convolutional Neural Network (CNN) pipeline automatically identified the placental boundaries.

Once the boundaries were established, the team applied deformable medial modeling. Think of this as finding the "skeleton" or the central manifold of the placenta. By modeling this middle surface, they could calculate sophisticated morphometric features:

  • Surface Areas: Specifically distinguishing between the maternal and chorionic sides.
  • Thickness and Diameter: Measured across the entire generated medial surface.
  • Non-planarity: A measure of how much the placenta curves or warps, calculated as the greatest height difference across the medial surface.

Datasets

The study utilized a dataset of 3D ultrasound volumes from 20 singleton pregnancies. These were captured during the first trimester (11-14 weeks), a critical window for placental development. While the sample size is small, it serves as a proof-of-concept for comparing the automated CNN approach against the "gold standard" of manual segmentation.

Key Insights and Ideas

  • Shape Matters More Than Size: The core philosophy here is that volume alone is a "noisy" metric. By looking at surface area differences and "non-planarity," researchers can capture the "topography" of the placenta, which may reflect how well it is integrated with maternal blood vessels.
  • The "Medial" Advantage: Using a medial model (a central surface) rather than just a 3D blob allows for more consistent measurements of thickness and circumference across an irregularly shaped organ.
  • High Fidelity Automation: The study demonstrates that deep learning isn't just "close enough"—it is highly correlated with expert manual work. This suggests that automated tools can finally handle the heavy lifting of 3D image analysis in a clinical timeframe.

Key Results

The automated system performed remarkably well when compared to manual experts:

  • Volume Agreement: The mean percent difference in volume was only 3.28%.
  • Dice Overlap Score: They achieved a score of 0.85 ± 0.07 (where 1.0 is a perfect match), indicating high spatial overlap.
  • Strong Correlations: There were strong, statistically significant (p < 0.01) linear correlations for almost all metrics, including thickness, circumference, and surface area.
  • Consistency: For all morphometric parameters, the mean percent difference between the human and the AI was less than or equal to 10%.

While there were some minor statistical differences in maternal surface area and mean diameter, the overall "shape profile" generated by the AI was highly representative of the manual version.

Limitations and Future Work

The primary limitation is the sample size; 20 pregnancies provide a strong technical validation but are not enough to establish clinical norms.

The authors acknowledge that the next big step is clinical utility. Now that we can measure these complex shapes automatically, we need to find out if they actually matter. Future research will likely focus on:

  1. Large-scale longitudinal studies: Do specific "non-planarity" scores in the first trimester predict preeclampsia or fetal growth restriction later in pregnancy?
  2. Refining the Maternal Interface: Investigating why the maternal surface area showed more variance between manual and automated methods.
  3. Real-time Integration: Moving the pipeline from a research setting into the ultrasound suite for immediate feedback during a scan.

Summary updated

Abstract

OBJECTIVES/GOALS: In this study, we implemented deformable medial modeling as a morphometric approach in first trimester placentas to characterize morphometric differences between fully automated and manual segmentations. METHODS/STUDY POPULATION: Twenty placentas from singleton pregnancies between 11-14 weeks’ gestation were manually and automatically segmented from 3D ultrasound volumes. Automated segmentations were produced by a trained convolutional neural network pipeline. Dice overlap scores and volumes were computed between manual and automated segmentations. Deformable medial modeling was applied to both manual and automated segmentations to produce the following metrics: maternal and chorionic surface areas (SA), thickness, circumference, and diameter along the generated medial surface. Placental non-planarity was also determined as the greatest medial surface height difference. A paired t-test and simple linear regression was performed between manual and automated segmentations for each shape metric. RESULTS/ANTICIPATED RESULTS: Mean placental volume measurements between manual and automated segmentations were similar, with a percent difference of 3.28% and a mean Dice overlap score of 0.85 ± 0.07. There were strong, statistically significant (p &lt;0.01) linear correlations with chorionic and maternal SA, SA difference, thickness, circumference, medial surface diameter, and medial surface height difference. No significant differences were noted with chorionic SA, thickness, circumference, maximum medial surface diameter, or medial surface height difference. However, statistically significant differences (p &lt;0.01-0.03) were noted in maternal SA, SA difference, and mean medial surface diameter. Despite these differences, mean percent difference for all morphometric parameters was less than or equal to 10%. DISCUSSION/SIGNIFICANCE: A deformable medial model evaluate unique global and regional shape placental features with highly correlated values between manual versus automated placental segmentations. However, clinical studies are needed to determine if minor differences would impact the clinical utility of these features as potential indicators of placental function.
BibTeX
@article{arenas2024deformable,
  author = {Gabriel Arenas and Alison M. Pouch and İpek Oğuz and Baris Oguz and Brett Byram and Xing Yao and Nadav Schwartz},
  title = {{448 Deformable Medial Modeling to Generate Novel Shape Features of the Placenta Using Automated versus Manual Segmentations}},
  journal = {Journal of Clinical and Translational Science},
  year = {2024},
  doi = {10.1017/cts.2024.383},
}
Figure