Fetal Ultrasound Image Segmentation for Automatic Head Circumference Biometry Using Deeply Supervised Attention-Gated...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Fetal head segmentation and head circumference (HC) measurement from ultrasound images present challenges due to image variability across pregnancy stages.
• Accurate HC biometry is crucial for fetal growth assessment and monitoring.

Purpose of the Study:

• To develop and evaluate a novel deep learning method for automated fetal head segmentation and HC biometry.
• To improve the accuracy and reliability of HC measurements in fetal ultrasound.

Main Methods:

• Proposed a deeply supervised attention-gated (DAG) V-Net model, integrating attention mechanisms and deep supervision into the V-Net architecture.
• Utilized a multi-scale loss function for deep supervision and data augmentation to expand the training dataset.
• Applied morphological processing, edge detection, and ellipse fitting for HC measurement post-segmentation.

Main Results:

• The DAG V-Net achieved a Dice Similarity Coefficient (DSC) of 97.93% and a Hausdorff Distance (HD) of 1.29 ± 0.79 mm on the HC18 test set (n=355).
• HC measurement accuracy demonstrated a difference (DF) of 0.09 ± 2.45 mm and an absolute difference (ADF) of 1.77 ± 1.69 mm.
• The method ranked fifth in the HC18 Challenge, outperforming conventional U-Net and V-Net models and showing comparable or superior results to state-of-the-art methods.

Conclusions:

• The proposed DAG V-Net offers a robust and accurate solution for fetal ultrasound image segmentation and HC biometry.
• The integration of attention mechanisms and deep supervision significantly enhances segmentation performance.
• DAG V-Net shows potential as a valuable tool for clinical application in prenatal care.