Automated Placenta Segmentation with a Convolutional Neural Network Weighted by Acoustic Shadow Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics

Background:

• Routine obstetrical ultrasound primarily assesses placental location, often missing abnormal placental development linked to complications.
• Technical challenges and lack of objective criteria hinder ultrasound diagnosis of placental pathology.
• Automated placental assessment tools are needed to overcome current diagnostic limitations.

Purpose of the Study:

• To develop a fully automated placental segmentation method using a convolutional neural network.
• To incorporate a novel layer for automated acoustic shadow detection to improve artifact recognition in ultrasound images.
• To create a robust algorithm for placental segmentation applicable across diverse imaging scenarios.

Main Methods:

• Developed a convolutional neural network with a specialized layer for acoustic shadow detection.
• Utilized a dataset of 1364 fetal ultrasound images from 247 patients, acquired over 47 months with varied machines and operators.
• Compared automated segmentation results (Dice coefficients) with manual segmentation, both with and without the acoustic shadow detection layer.

Main Results:

• Achieved high mean Dice coefficients (0.92±0.04) for automated segmentation on the full dataset, comparable to manual segmentation.
• Demonstrated significant improvement in segmenting images with acoustic shadows (0.87±0.04) when using the acoustic shadow detection layer, compared to without (0.75±0.05).
• The method requires no user input for tuning, ensuring ease of use.

Conclusions:

• The developed automated placental segmentation method is effective and robust, even with ultrasound artifacts.
• This AI-driven approach can serve as a crucial preprocessing step for large-scale placental image analysis.
• Facilitates further research and development of computer-aided diagnostic tools for obstetrical ultrasound.