Atlas-Based Segmentation of the Human Embryo Using Deep Learning with Minimal Supervision

We propose an atlas-based segmentation framework to achieve segmentation and spatial alignment of three-dimensional ultrasound images of the human embryo acquired during the first trimester of pregnancy. Our framework is based on deep learning with minimal supervision. The framework consists of two networks, one dedicated to learning an affine transformation and one dedicated to learning a nonrigid deformation. The affine registration network is trained in two stages, were the first stage is minimally supervised and the resulting transformation is refined in the second unsupervised stage. The nonrigid registration network is trained completely unsupervised and in one stage. The framework is trained, validated and tested on a dataset of ultrasound images of the human embryo acquired in the 9th week of pregnancy. We visually assessed the affine alignment, which was accurate in of the images in the test set. For 14 out of 31 test images a manual segmentation was available and we achieved an average Dice similarity coefficient of 0.78. Therefore we conclude that our framework is a promising approach for segmentation and spatial alignment of the human embryo in three-dimensional ultrasound images with minimal supervision.