Abstract

The placenta is a critical organ connecting the mother and fetus, playing a vital role in fetal development and maternal health. However, research on placental MRI segmentation, particularly across different gestational stages, remains limited. This study systematically explores the performance of nnUNet in automatic placental MRI segmentation, focusing on three gestational stages: early (0–19 weeks), mid (20–30 weeks), and late (31–40 weeks). Comprehensive experiments, including single-stage, mixed-stage, and cross-stage segmentation tasks, were conducted.

The results demonstrate nnUNet’s remarkable adaptability, achieving validation Dice coefficients exceeding 82% in single-stage training. The mid-stage placenta, due to its transitional characteristics, excelled in cross-stage predictions, with Dice reaching 61.66% for late-stage and 56.40% for early-stage. In mixed-stage training, increasing the proportions of early and late-stage data, especially early-stage data, further improved segmentation performance across all stages, with a combined test Dice of 83.86% for mid-stage segmentation.

This study significantly enhances the efficiency and accuracy of placental MRI segmentation, reducing manual annotation time, improving annotation consistency, and providing a foundation for large-scale placental data analysis. It provides a solid foundation for large-scale placental data analysis, addressing the lack of research in this domain and offering critical support for obstetric medical research and clinical applications.

Document Type

Article

Author's School

McKelvey School of Engineering

Author's Department

Electrical and Systems Engineering

Class Name

Electrical and Systems Engineering Undergraduate Research

Language

English (en)

Date of Submission

12-2-2024

Share

COinS