FBSPM-Net: Automatic recognition of fetal brain ultrasound images in early pregnancy

Ultrasound imaging is used to evaluate fetal development during early pregnancy. However, sonographers frequently make subjective choices when selecting standard fetal brain planes. To enhance objectivity and consistency, this work presents a multi-task model based on YOLOv5, named FBSPM-Net, for identifying fetal brain planes and detecting anatomical structures. FBSPM-Net incorporates a Ghost module to lightweight, a convolutional block attention module to increase accuracy, the SIoU loss function for rapid training, and a decoupled head to enhance the feature usage rate. FBSPM-Net achieves significant improvements, increasing plane identification accuracy by 12.9% and recall by 13.3% while maintaining the same parameters as the baseline model. Experiments demonstrate that FBSPM-Net can detect anatomical structures and recognize the fetal brain plane, thereby reducing the effort required of sonographers in acquiring standard planes.