Fully-automated deep learning pipeline for 3D fetal brain ultrasound

Thesis

Three-dimensional ultrasound (3D US) imaging has shown significant potential for in-utero assessment of the development of the fetal brain. However, in spite of the potential benefits of this modality over its two-dimensional (2D) counterpart, its widespread adoption remains largely limited by the difficulty associated with its analysis.

While more established 3D neuroimaging modalities, such as Magnetic Res- onance Imaging (MRI), have circumvented similar challenges thanks to reliable, automated neuroimage analysis pipelines, there is currently no comparable pipeline solution for 3D neurosonography.

With the goal of facilitating medical research and encouraging the adoption of 3D US for clinical assessment, the main objective of my doctoral thesis is to design, develop, and validate a set of fundamental automated modules that comprise a fast, robust, fully automated, general-purpose pipeline for the neuroimage analysis of fetal 3D US scans.

For the first module, I propose the fetal Brain Extraction Network (fBEN), a fully-automated, end-to-end 3D Convolutional Neural Network (CNN) with an encoder-decoder architecture. It predicts an accurate binary brain mask for the automated extraction of the fetal brain from standard clinical 3D US scans.

For the second module I propose the fetal Brain Alignment Network (fBAN), a fully-automated, end-to-end regression network with a cascade architecture that accurately predicts the alignment parameters required to rigidly align standard clinical 3D US scans to a canonical reference space.

Finally, for the third module, I propose the fetal Brain Fingerprinting Net- work (fBFN), a fully-automated, end-to-end network based on a Variational AutoEncoder (VAE) architecture, that encodes the entire structural information of the 3D brain into a relatively small set of parameters in a continuously distributed latent space. It is a general-purpose solution aimed at facilitating the assessment of the 3D US scans by recharacterising the fetal brain into a representation that is easier to analyse.

After exhaustive analysis, each module of this pipeline has proven to achieve state-of-the-art performance that is consistent across a wide gestational range, as well as robust to image quality, while requiring minimal pre-processing. Additionally, this pipeline has been designed to be modular, and easy to modify and expand upon, with the purpose of making it as easy as possible for other researchers to develop new tools and adapt it to their needs. This combination of performance, flexibility, and ease of use may have the potential to help 3D US become the preferred imaging modality for researching and assessing fetal development.

Authors: Moser, FA

• DOI: 10.5287/ora-d5bd8g9gj
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English