Automatic Segmentation of Placenta from MR images Using a Novel BiGC U-Net

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate placenta segmentation in Magnetic Resonance (MR) images is crucial for quantitative analysis in diagnosing placenta accreta spectrum (PAS).
• Challenges include low contrast, image noise, blurred boundaries, and manual segmentation variability, hindering diagnostic accuracy.

Purpose of the Study:

• To develop an automated deep learning (DL) model for accurate placental MR image segmentation.
• To enhance diagnostic capabilities for placenta accreta spectrum (PAS) through improved segmentation.

Main Methods:

• Proposed an enhanced U-Net architecture, BiGC U-Net, incorporating a bidirectional gated convolutional module (BiGC) and hierarchical regularization mechanism (HRM).
• Implemented an innovative data augmentation strategy to synthesize new training images.
• Evaluated performance on public, Sheffield Teaching Hospitals (STH), and combined placental MR datasets, comparing against U-Net, Attention U-Net, ResNet, UNet++, TransUNet, nnUNet, and SSM-Mamba.

Main Results:

• BiGC U-Net achieved superior performance on the combined dataset, demonstrating a Dice similarity coefficient of 90.74 ± 0.44.
• The model also exhibited excellent results in 95th percentile Hausdorff distance (3.84 mm ± 0.53 mm) and relative volume difference (9.06 ± 0.41).
• Outperformed existing DL models in key segmentation metrics.

Conclusions:

• The BiGC U-Net model provides an effective and robust solution for automatic placenta segmentation in MR images.
• This automated segmentation can significantly aid in improving diagnostic rates for placenta accreta spectrum (PAS).