We propose an unsupervised deep learning method for atlasbased registration to achieve segmentation and spatial alignment of the embryonic brain in a single framework. Our approach consists of two sequential networks with a specifically designed loss function to address the challenges in 3D first trimester ultrasound. The first part learns the affine transformation and the second part learns the voxelwise nonrigid deformation between the target image and the atlas. We trained this network end-to-end and validated it against a ground truth on synthetic datasets designed to resemble the challenges present in 3D first trimester ultrasound. The method was tested on a dataset of human embryonic ultrasound volumes acquired at 9 weeks gestational age, which showed alignment of the brain in some cases and gave insight in open challenges for the proposed method. We conclude that our method is a promising approach towards fully automated spatial alignment and segmentation of embryonic brains in 3D ultrasound.

No abstract

Objectives: Periconceptional health influences embryonic brain development, which can have consequences for the neurodevelopmental health of the child. In order to further understand this, our overall aim is to create a comprehensive spatiotemporal atlas of the embryonic brain. The current approach to assess embryonic brain development requires manual measurements of individual brain structures in ultrasound datasets, which is time-consuming. Therefore, our first step is to put all embryos in the same position (alignment) and extract the embryo from the data (segmentation). Here, we present a method to achieve this fully automatically using artificial intelligence (AI). Methods: We marked the Crown-rump landmarks in 3D ultrasound datasets acquired in the Rotterdam Periconception Cohort (Predict study) at 9 weeks gestational age. One image was manually segmented and put in standard position to serve as a reference. During training, the method learns to align the other images to this reference using the landmarks, and thereby achieves segmentation. 47 scans, kept separate during training, were used to evaluate their alignment with the reference. Note the landmarks are only needed for training, afterwards the method is fully automatic. Results: Of the 47 validation images, 37 (79%) were correctly aligned to the reference image. This means our method can automatically align and segment most new images. Figure 1 gives an impression of the results. Conclusions: From these results we conclude that our method is a promising approach for fully automated preprocessing of ultrasound images, paving the way for creation of a spatiotemporal embryonic brain atlas.

No abstract