DEVELOPING AN AUTOMATIC MODEL FOR NUCHAL TRANSLUCENCY THICKNESS SEGMENTATION IN FETAL ULTRASOUND IMAGES

Abstract

The determination of the nuchal translucency (NT) thickness on first-trimester ultrasound images is an important step that should be performed in all pregnant women to detect early signs of Down syndrome, Turner syndrome, fetal anomalies, and other genetic disorders. Currently, NT measurement is manually performed by experienced and certified ultrasound physicians, so the results heavily depend on the physician's expertise. Therefore, there is a need for a method to automatically identify the NT region to assist physicians in quickly and accurately measuring NT thickness. This paper focuses on using deep learning models to segment nuchal translucency regions in ultrasound images. Several popular deep learning segmentation models such as FPN, UNet, UNet++, DeepLabV3, and DeepLabV3+ were selected for implementation and testing. The experiments show that the UNet model with EfficientNetB6 encoder achieved the best results with an accuracy of 99.51%, IoU score of 60.95%, and Dice score of 77.14%. The paper also discusses the challenges and future directions of this field.