Deep learning-based method for recognition and segmentation of nuchal translucency in early pregnancy

Abstract Nuchal translucency (NT) refers to the accumulation of subcutaneous fluid in the nape region of a fetus. The thickness of NT is positively correlated with chromosomal abnormalities and structural defects in the fetus. NT measurement is a critical screening method during early pregnancy but is often hindered by challenges such as speckle noise, weak boundaries, and other imaging artifacts in ultrasound images. Manual annotation and measurement of NT regions are time-consuming and difficult for clinical experts, especially given that the NT region constitutes only a small portion of the entire ultrasound image. Accurate extraction of the NT region, particularly in the presence of noise and weak edges, remains a significant challenge. In this article, building upon the classic and effective U-Net architecture, an NT region segmentation approach adapted for this task is proposed, integrating Multi-Scale Feature Fusion (MSFF) and attention mechanisms to enhance feature selectivity. The proposed method enables robust and automated segmentation of NT regions in ultrasound images, providing a reliable foundation for subsequent NT thickness measurement. The proposed method achieves competitive overall performance on the NT dataset compared with the evaluated methods, attaining an Intersection over Union (IoU) of 75.76% and a Dice similarity coefficient (DSC) of 85.51%. Compared to existing methods, our approach demonstrates superior segmentation accuracy, enabling precise delineation of NT regions. This advancement has significant potential for early prenatal screening of chromosomal abnormalities and related conditions, contributing to enhanced clinical decision-making in early pregnancy management.