Supervisor

Suitable for

Abstract

Increasing digitisation of the medical domain has led to the creation of large amounts of data, which will continue to rise in the near future. As these datasets become increasingly available, there is an opportunity to combine data from multiple sites to generate unified “big data” stores. This has the advantage of improving the anatomical diversity of training datasets, thereby providing a better overall understanding of disease mechanisms, and increasing statistical power. However, multi-site medical data tends to contain hidden biases that are imperceptible to a human observer, but result in a drop in the predictive power of deep learning models. Heterogeneous data from multiple sites may also contain undesirable non-biological variance, even when attempts are made to reduce it by following standardised acquisition protocols or using identical medical equipment.

The goal of this project is to design deep learning algorithms that are applicable to heterogeneous ultrasound image datasets. Specifically, the student will be tasked with implementing convolutional neural network (CNN) architectures to automatically delineate subcortical structures (e.g. cerebellum, lateral ventricles, choroid plexus) in 3D ultrasound images of the fetal brain (Fig. 1). The volumetric measurements will then be used to construct population growth charts that can serve as a reference for the global population. Domain adaptation techniques will also be explored to solve the “harmonisation problem” by remove scanner-specific (and/or site-specific) biases in the model’s predictions.

This project is suitable for a student interested in deep learning, algorithm development, and image analysis. While programming experience is desirable (ideally Python), no prior knowledge of brain anatomy is necessary.

References: The first two papers are on deep learning-based approaches to the harmonisation problem, and the last two papers demonstrate the feasibility of segmenting subcortical structures in 3D ultrasound images. Please visit our website and the project pages for further information.

Model harmonisation [Project page] • Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal Dinsdale, NK, Jenkinson, M, Namburete, AIL. NeuroImage 2021 • Unlearning scanner bias for MRI harmonisation Dinsdale, NK, Jenkinson, M, Namburete, AIL. MICCAI 2020 Structural segmentation [Project page] • Subcortical Segmentation of the Fetal Brain in 3D Ultrasound using Deep Learning Hesse, LS, Aliasi, M, Moser, M, the INTERGROWTH-21st Consortium, Jenkinson, M, Haak, M, Namburete, AIL. NeuroImage 2021 • Cortical plate segmentation using CNNs in 3D fetal ultrasound Wyburd, MK, Jenkinson, M, Namburete, AIL. MIUA 2020