Adaptive feature fusion network for fetal head segmentation in ultrasound images

Abstract

The measurement of fetal biometrics from ultrasound images plays a vital role in assessing potential development during pregnancy. However, existing fetal segmentation methods failed to accurately segment and asses the head circumference that gives inaccurate segmentation results. To overcome this limitation, a feature feedback and global feature with adaptive feature fusion network (FGA–Net) model is proposed to enhance fetal head segmentation (FHS). It involves four key components for feature extraction, fusion, and correction, respectively. The adaptive feature fusion module (AFFM) and correction map integrate the local features and global features and refine the features to enhance accurate FHS from the ultrasound images efficiently. Initially, ultrasound images are obtained from the two publicly available datasets and preprocessed using normalization and data augmentation techniques. Finally, preprocessed images are fed to FHS by proposed FGA Net utilizing EfficientNet-B0 as the backbone network for efficient feature extraction. Experimental results of proposed FGA-Net are evaluated using the dice coefficient (DC) of 95.78% and 98.95% for FH-PS-AoP and HC-18 datasets, which shows better results than the existing segmentation approaches like inverted bottleneck patch expanding (IBPE) method.