Omni-supervised domain adversarial training for
white matter hyperintensity segmentation the UK Biobank

ISBI 2022 (Oral Presentation)
Vaanathi Sundaresan1
Nicola Dinsdale1
Mark Jenkinson1,2,3
Ludovica Griffanti1

1Wellcome Centre for Integrative Neuroimaging, University of Oxford
2Australian Institute for Machine Learning (AIML), University of Adelaide
3South Australian Health and Medical Research Institute (SAHMRI)

[Paper]
[GitHub]
[TrueNet Code]
[Age Prediction Code]

Abstract

White matter hyperintensities (WMHs, or lesions) appear as hyperintense, localized regions on T2-weighted and FLAIR brain MR images. The heterogeneity in lesion characteristics due to subject-level (e.g., local intensity/contrast) and population-level (e.g., demographic, scanner-related) variations make their segmentation highly challenging. Here, we propose a framework for adapting a state-of-the-art WMH segmentation method with high accuracy from a small, labeled source data (MICCAI WMH segmentation challenge 2017 training data) to a larger dataset such as the UK Biobank without the need of additional manual training labels, using domain adversarial training with omni-supervised learning. Given the well-known association of WMHs with age, the proposed method uses a multi-tasking model for learning lesion segmentation, domain adaptation and age prediction simultaneously. On a subset of the UK Biobank dataset, the proposed method achieves a lesion-level recall, lesion-level F1-measure and Dice overlap value of 0.95, 0.65 and 0.84 respectively, when compared to values of 0.75, 0.49 and 0.80 obtained from the pretrained state-of-the-art baseline method.

Pipeline

We combine domain adaptation and omni supervised learning to significantly improve the segmentation performance on white matter hyperintensities. Data from the MICCAI challenge dataset was used as the source data, and unlabelled UK Biobank data was the target. The data points used within the omnisupervised framework were selected using Age Prediction Deltas, as age is known to correlate with the presence of WMHs.
A network composed of three parts is trained: 1) WMH segmentation network (TrueNet) 2) Domain Prediction 3) Age Predictor

 [GitHub]

Results

Ablation Study

The approach leads to a significant improvement across metrics, especially lesion level recall, showing significant improvement in detecting small lesions.


Visual Results

Sample segmentation results shown for (a) Baseline pretrained on source, (b) unsup-DANN, (c) ssDANN-OL and (d) ssDANN-OL-age pred.

Yellow, blue and red voxels indicate true positive, false negative and false positive voxels respectively.

Acknowledgements

This research is funded by the Engineering and Physical Sciences Research Council (EPSRC) and Medical Research Council (MRC) [EP/L016052/1] and Wellcome Centre for Integrative Neuroimaging, which has core funding from the Wellcome Trust [203139/Z/16/Z]. The computational aspects of this research were funded from National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC) with additional support from the Wellcome Trust Core Award [203141/Z/16/Z]. The authors acknowledge the MRC Dementias Platform UK (MR/L023784/2), the NIHR Oxford Health BRC and the NIHR Oxford BRC. The UK Biobank data was obtained under app. number 8107. We are grateful to UK Biobank for making the resource data available and are extremely grateful to all UK Biobank study participants. This template was originally made by Phillip Isola and Richard Zhang for a colorful ECCV project; the code can be found here.