Direct Detection and Measurement of Nuchal Translucency with Neural Networks from Ultrasound Images

Nuchal Translucency (NT) in ultrasound images are commonly used to detect genetic disorder in fetuses. Due to lack of distinctive local features around NT region, existing NT detection methods first model some other prominent body parts, such as the fetal head. However, explicit detection of other body parts requires additional annotation, development and processing costs. It may also introduce cascading error in cases of unclear head location or non-standard head-NT relations. In this work, we design a convolutional neural network with fully connected layers to detect NT region directly. Furthermore, we apply U-Net with customized architecture and loss function to obtain precise NT segmentation. Finally, NT thickness measurement is calculated using principal component analysis. A dataset containing 770 ultrasound images were used for training and evaluation. Extensive experimental results show that our direct approach automatically detects and measures NT with promising performance.