Pubic Symphysis-Fetal Head Segmentation in Ultrasound Images with Deep Neural Networks

Pubic symphysis-fetal head (PS-FH) segmentation is a crucial task for measuring the angle of progression (AoP) of the fetal head during labor, which is an objective and reliable parameter that reflects the fetal descent and the likelihood of vaginal delivery. However, manual segmentation of PS-FH from transperineal ultrasound (TPU) images is time-consuming, subjective, and prone to inter-observer variability. Therefore, automatic PS-FH segmentation methods based on deep neural networks have been proposed to improve the efficiency and accuracy of AoP measurement. In this paper, we survey different loss functions, U-Net backbones, and attention mechanisms for PS-FH segmentation, and compare their performance on a public dataset. We have found that the combination of U-Net, EfficientNet, IoU loss, and scSE attention achieves the best performance, with an average accuracy of 0.96, an average Dice score of 0.96, and an average IoU of 0.93. This demonstrates the effectiveness of our method, which can extract more effective features from the ultrasound images, while optimizing the overlap ratio between the predicted and ground truth masks, and enhancing the feature maps by recalibrating the channel-wise and spatial-wise importance.