EP01.18: Automatic fetal nuchal translucency measurement using deep learning‐based three‐dimensional segmentation network in non‐supine positioned fetus

The increased Nuchal translucency (NT) is related to chromosomal defects and early cardiac failure in first trimester of pregnancy. Detecting the midsagittal plane (MSP) is an essential part to measure NT. However, there are many difficulties in measuring NT, such as inappropriate fetal pose, high fetal mobility and maternal abdominal wall thickness. Though there has been an automated system proposed, its clinical use has been limited to the supine fetal position with an angle of deviations under 20 degrees. In light of this, we present a novel solution for automated NT measurement in non-supine positioned fetus from three-dimensional (3D) ultrasound images. The provided images were obtained by an expert with W10 (Samsung Medison, Seoul, Republic of Korea) ultrasound machine using a 2–6 MHz transabdominal transducer. We propose a method that uses a deep learning-based algorithm. The method initially localises the fetal head and segment key landmark structures and automatically navigate the MSP based on these landmarks. The proposed learning algorithm guides a more accurate segmentation method in a novel way by reinforcing accuracy while compensating for a geometrical feature within the training stage. A total of 200 3D volume dataset of fetal head and upper thorax was obtained from 78 patients during first trimester NT screening at our institution of which 150 were used for training and 50 testing for 3D auto-labelling system, and validation of automated NT detection system, respectively. A success rate of 95% for 3D segmentation and an accuracy of 92.3% for MSP detection was achieved with an inference time of 0.5 seconds. The mean difference in NT measurements was -0.05mm (p 0.18) and the intraclass correlation coefficient was 0.923 between the two-dimensional NT (2D-NT) and the automated 3D volume NT system (5D NT-AI). A novel deep learning algorithm is reproducible and comparable with conventional 2D techniques for NT measurement.