Automatic placental and fetal volume estimation by a convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate estimation of placental and fetal volumes is crucial for monitoring pregnancy progression and fetal well-being.
• Manual measurement of these volumes from Magnetic Resonance (MR) scans is time-consuming and labor-intensive.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) deep learning algorithm for efficient and accurate estimation of placental and fetal volumes using MR imaging.
• To compare the performance of the AI algorithm against manual measurements.

Main Methods:

• A deep learning algorithm, DenseVNet, was trained using manually annotated MR images from 193 normal pregnancies (gestational weeks 27 and 37).
• The dataset was divided into training (163 scans), validation (10 scans), and testing (20 scans) sets.
• The algorithm's segmentation accuracy was quantified using the Dice Score Coefficient (DSC) compared to ground truth manual annotations.

Main Results:

• The AI algorithm achieved a mean DSC of 0.925 for placental volume estimation and high DSC values for fetal volume estimation (0.952-0.970).
• Estimated placental volumes were 870 cm³ (DSC 0.887) at week 27 and 950 cm³ (DSC 0.896) at week 37.
• Fetal volume estimations showed mean DSCs of 0.952 at week 27 and 0.970 at week 37.
• Volume estimation time was drastically reduced from 60-90 minutes for manual annotation to under 10 seconds using the AI algorithm.

Conclusions:

• The developed AI deep learning algorithm provides accurate estimations of placental and fetal volumes from MR scans.
• The AI approach significantly enhances efficiency, reducing estimation time by over 99% compared to manual methods.
• The AI's performance is comparable to human accuracy, offering a substantial improvement in clinical workflow for prenatal imaging analysis.