Work in this area aims to use machine learning tools to understand more about cardiovascular changes and development during pregnancy and beyond, both in mothers and children.
An example of the work coming under this theme includes a study using machine learning to understand how the brain develops in utero. Scientists and clinicians can analyse images from 3D ultrasounds during pregnancy to quantify subcortical volume in the brain of a developing foetus, but this usually needs someone to look at the images and decide where the different segments of different parts of the subcortex begin and end.
This takes some time, and it is difficult – ultrasound images are often blurry and unclear, and segmenting different part of a small, developing brain ‘by eye’ on these images needs a skilled operator.
Oxford University researchers, together with colleagues in the Netherlands, trained a neural network (a computing system inspired by the biological neural networks in real brains) to automatically segment different parts of subcortical structure, and the cerebellum, in the developing foetal brain.
The neural network needed only a few sets of ultrasound images with the various parts of the brain labelled to ‘learn’ how to segment ultrasound images that it had not been trained on. The network worked best when the images were first aligned (by a human operator), but it still worked on non-aligned images when they had only a few basic structures labelled.
The research team used their trained neural network on a group of over 200 ultrasound images, and discovered a new pattern of ultrasound-specific growth trajectories for different parts of the developing subcortical foetal brain.
This study demonstrates some of the promise of machine learning – neural networks can learn to do complex segmentation of ultrasound data that usually required a trained human operator, and shows that the results from a neural network can yield new information.
Read the full study:
Subcortical segmentation of the fetal brain in 3D ultrasound using deep learning. Neuroimage. Hesse LS, et al. 2022;254:119117. doi:10.1016/j.neuroimage.2022.119117