Harmonisation and Domain Adaptation

Different MRI scanners produce images with different characteristics which leads to an increase in non-biological noise when the images are pooled. This is known as the harmonisation problem. We propose to model harmonisation as a domain adaptation problem, and have proposed a number of methods to enable models to be trained with data from a range of scanners.

Home  /  Harmonisation  /  Segmentation  /  Translation  /  Privacy  /  Explainable AI

Papers

We, therefore, propose UniFed, a unified federated harmoni- sation framework, which enables three key processes to be completed: 1) the training of a federated partially labelled harmonisation network, 2) the selection of the most appropriate pretrained model for a new unseen site, and 3) the incorporation of a new site into the harmonised federation.

We propose an Unsupervised Source-Free Domain Adaptation (SFDA) method, SFHarmony, and demonstrate the approach for classification, regression and segmentation tasks.

Adapted my harmonisation framework to work in a distributed setting, through modelling site features as Gaussian distributions.

Exploring the use of omni-supervised learning for white matter hyperintensity segmentation, using age prediction as the selection criteria. Leads to significant increase in sensitivity especially for small lesions.

IDP harmonisation using iterative unlearning framework, applied across tasks and architectures.

Parts of this work were presented at MIUA 2020 and MICCAI 2020 [Early Acceptance].

Comparison of DA methods for white matter hyperintensity segmentation, comparing methods including my proposed method: paper.

Segmentation of cardiac MRI across sites, using the method proposed in my paper

MRI harmonisation using iterative unlearning framework, explored for age prediction.

MRI harmonisation using iterative unlearning framework, explored for segmentation tasks.