Automatic recognition of fetal facial ultrasound standard planes based on improved YOLOv4

Accurate acquisition of standard planes of fetal facial ultrasound is essential for subsequent biometry and disease diagnosis. Foreign scholars have extensively researched algorithms for the automatic acquisition of ultrasound standard planes. We view standard plane identification as a detection task, unlike previous classification studies. This study proposes a lightweight target detection network for identifying fetal facial ultrasound standard planes. Methods: Firstly, the model is based on the YOLOv4 algorithm, and given the storage resource limitations of the ultrasound device, we used a lightweight network (GhostNet) to replace the YOLOv4 backbone feature extraction network (CSPDarkNet53). Results: The experimental results show that the average accuracy of the improved YOLOv4 algorithm is 98.06%. The model size is 42.7 MB, a reduction of 85% compared to the original YOLOv4. It takes only 0.07 seconds to detect an ultrasound image, which can fully meet the real-time clinical requirements. It has high detection speed and accuracy, and the model's size is reduced substantially. The algorithm can assist young ultrasonographers in better acquiring high-quality ultrasound images and, to some extent, can address the limitations of the traditional manual approach to acquiring standard planes.