Fetal Cardiac Structure Detection Using Multi-task Learning

The prenatal ultrasound examination is a highly effective method for evaluating fetal health, wherein the acquisition of standard fetal ultrasound planes and accurate identification of key structures within them are essential for precise measurement of fetal growth parameters. However, manual acquisition of standard planes is not only time-consuming and laborious, but may also introduce subjectivity and inconsistency due to individual experience differences among sonographers. To mitigate these challenges, we propose a fetal cardiac structure detection model based on multi-task learning (FCSD), which adopts the ConvNeXt network structure to extract underlying shared multi-scale features. Within the detection architecture, the Feature Pyramid Network (FPN) integrates disparate feature layers from ConvNeXt, enabling the comprehensive utilization of semantic information encapsulated in high-level feature maps, thereby enhancing the capability to extract crucial features. Additionally, we design a classification module with stacked multi-layer residual networks and feature fusion, which enables efficient and accurate discrimination of standard planes in fetal ultrasound images. The experimental outcomes show that the proposed FCSD model surpasses other competitive baseline models in terms of detection accuracy and classification precision, which can assist sonographers in prenatal screening.