Minimally interactive placenta segmentation from three-dimensional ultrasound images

Area of Science:

• Medical Imaging
• Obstetrics
• Computational Biology

Background:

• Placental size in early pregnancy is linked to fetal growth.
• Current 3D ultrasound (3DUS) placental volume measurement is time-consuming and user-dependent.

Purpose of the Study:

• To develop a semiautomated segmentation technique for robust placental volume measurement from 3DUS.
• To minimize user input and variability in placental biometry.

Main Methods:

• A multi-atlas label fusion (MALF) algorithm was initialized with a single 2D slice annotation.
• 47 singleton pregnancy 3DUS volumes (11-14 weeks) were analyzed.
• Segmentation accuracy was compared to expert manual segmentations using Dice overlap and surface distance.

Main Results:

• MALF achieved segmentation overlap (mean Dice: 0.831) comparable to methods with extensive manual input.
• High test-retest reliability was demonstrated (correlation coefficient: 0.99, intraclass correlation: 0.99).
• The method showed accurate and reliable placental segmentation with minimal user interaction.

Conclusions:

• Multi-atlas label fusion (MALF) offers an accurate and reliable approach for placental segmentation.
• This technique requires minimal user interaction, addressing limitations of current methods.
• Further development could enable routine clinical use for pregnancy surveillance.