An ultrasound standard plane detection model of fetal head based on multi-task learning and hybrid knowledge graph

Prenatal ultrasound examination is a powerful tool to prevent birth defects and assess fetal health. Obtaining ultrasound standard planes is a prerequisite for prenatal ultrasound diagnosis. However, ultrasound standard plane detection depends heavily on the sonographer’s sufficient clinical experience and solid knowledge of fetal anatomy. In this study, to lighten the workload of the sonographer and promote the accuracy, efficiency, and interpretability of ultrasound standard plane detection, we propose an ultrasound standard plane detection (USPD) model based on multi-task learning and a hybrid knowledge graph. We first design a multi-task learning strategy to learn the shared features of fetal ultrasound images through convolutional blocks. Then, we optimize the generalization performance by extending the shared features into the task-specific output streams. In addition, USPD integrates clinical prior knowledge graphs to reduce the error rate and missed detection rate. The USPD model can recognize the key anatomical structures of fetal heads and analyze the types of ultrasound planes. Furthermore, unlike most “end-to-end” automatic detection models, the USPD model not only outputs the prediction results but also provides consistent interpretation for professional sonographers, thereby increasing the interpretability of the model without the sonographer’s intervention. We conduct extensive experiments on a fetal head ultrasound image dataset to assess the proposed USPD model via comparison with competitive methods. Experimental results illustrate that the proposed USPD model outperforms the competitive methods with regard to accuracy and performance, and it can meet the clinical requirements in practical application.