Deep learning for automatic segmentation of the gestational sac and yolk sac from 3D ultrasound in early pregnancy

Abstract

Abstract The current standard manual measurements from 2D ultrasound in the early weeks of pregnancy are operator-dependent and subject to inter- and intra-rater variability, particularly for the non-spherical gestational sac. This leads to diagnostic uncertainty, prolonged follow-up and patient anxiety. Several studies have proposed models for automatic standard plane selection and segmentation of the gestational sac from 2D ultrasound. However, these methods do not eliminate variability caused by natural differences in gestational sac shape. As 3D ultrasound is becoming increasingly accessible in the clinics, an automated method for 3D volumetric measurement would be beneficial for both patients and clinicians in terms of reproducibility, accuracy and ease of use. In this study, we trained deep learning models using the nnU-Net framework for automatic segmentation and volumetric measurement of the yolk sac and gestational sac in 3D ultrasound volumes of 42 women pregnant after assisted reproductive technology, for an initial feasibility assessment of automated volumetry in this setting. The best model performance was similar to the human intra-rater variability for the gestational sac, with median Dice scores of 92.2% versus 92.8%, while yolk sac segmentation remained more challenging, with median Dice scores of 84.3% versus 92.8%. Estimated volumes and diameters showed higher agreement with expert volumetric measurements than the manually measured 2D diameters did, with an improved intraclass correlation coefficient (ICC) from 0.83 to 0.99 for the gestational sac, and comparable ICCs of 0.64 and 0.68 for the yolk sac. These results support the feasibility of automated 3D volumetry as a promising alternative to the current standard manual measurement method. The proposed method could potentially reduce examination time, facilitate new research on early pregnancy development, and in the future reduce the diagnostic uncertainty that leads to prolonged follow-up in early pregnancy.