Automatic fetal biometry prediction using a novel deep convolutional network architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Fetal Medicine

Background:

• Fetal biometric measurements are crucial for prenatal assessment but face challenges like image quality and low amniotic fluid.
• Accurate measurements of biparietal diameter (BPD), head circumference (HC), abdominal circumference (AC), and femur length (FL) are essential for monitoring fetal growth and development.

Purpose of the Study:

• To develop and evaluate an automated convolutional neural network (CNN) system for fetal biometric parameter segmentation and measurement.
• To address challenges in ultrasound image analysis using an attention-gated multi-feature pyramid Unet (MFP-Unet) architecture.

Main Methods:

• The Attention MFP-Unet model was developed, incorporating attention gates for salient region detection.
• Pre-processing involved Niblack's thresholding for head/abdomen and a novel algorithm for femur extraction.
• The algorithm was trained and validated on a public dataset (HC18) and clinical data from 1334 subjects.

Main Results:

• Attention MFP-Unet achieved high performance in fetal anatomy segmentation, with Dice Similarity Coefficient (DSC) of 0.98 and Hausdorff Distance (HD) of 1.14 mm.
• Accuracy of biometry predictions was evaluated using correlation analysis, good contours, and conformity, yielding excellent results (e.g., 0.95 conformity).
• The model demonstrated superior performance in Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), percentage of good contours, conformity, and Average Perpendicular Distance (APD).

Conclusions:

• The Attention MFP-Unet demonstrates superior performance for automatic fetal biometric parameter measurement compared to existing state-of-the-art methods.
• The proposed AI-driven approach offers a robust solution for overcoming challenges in fetal ultrasound image analysis.