MRC DiMeN Doctoral Training Partnership: Using artificial intelligence to optimise treatment decisions by analysis of retinal images for patients with blinding diabetic eye disease

Diabetic eye disease accounts for over 28,000 sight impaired people in the UK, with 1280 new patients registered blind each year. The commonest cause of blindness is diabetic macular oedema (DMO), which can be treated with laser or intraocular injections. Spectral-domain optical coherence tomography (SD-OCT) can quickly, and easily detect DMO but current quantitative measures of its severity on SD-OCT are poor at predicting progression and response to treatment. This project will investigate the design, development, testing, and application of deep convolutional neural networks to this problem of DMO quantification utilising several other OCT biomarkers. Based on our recent initial studies, neural architectures within this application setting can achieve close to expert human-level performance on unseen test images.

The proposed project objectives are:
1. Development of a fully automated comprehensive tool to segment and quantify a range of different OCT biomarkers in DMO.
2. Assess the performance of the tool against human grader classification.
3. Assess the ability of the automated quantification to predict outcome after DMO treatment using high quality previously collected datasets.

Experimental Approach:
3D Imaging Data Acquisition and Annotation [PhD,S2,S3]:
High density and resolution SD-OCT images will be acquired using a Heidelberg Spectralis SD-OCT. Annotations of all biomarkers will be created manually using a 3D image annotation tool. There will be 100 3D images annotated in the training dataset, 20 in the validation dataset and 100 in the test set.

3D Model Design and Training [PhD,S1]:
To improve the model’s generalization, we used the following transforms, all performed in 3D: scaled-up image, cropped data, elastic deformations. We experimented with different model depths. For regularization, we will experiment with different values of weight decay for our optimizer. We will train our model separately three times to assess the consistency of our results.

3D Quantitative Analysis of Macular Oedema [PhD,S1,S2,S3]:
The features, in the 3D SD-OCT images, will be automatically segmented, labelled and then quantitatively analysed. Shape representations will include landmarks, implicit representations, parametric representations, medial models, and deformation-based descriptors

Validation of the Developed Automated Quantification Tool [PhD,S2,S3]
The developed automated 3D model clinical utility to predict treatment outcomes will be validated utilising prospectively collected data from 2 previously carried out clinical trials to assess 1-year treatment outcomes in patients with DMO, using general linear modelling and will be compared with both standard OCT thickness, and a human retinal expert prediction of treatment outcome as an initial assessment of this approach in clinical practice.

Additional information:
BO’s lab: boguslawobara.com
DS’s profile: https://www.ncl.ac.uk/bns/about/staff/profile/davidsteel.html#background
MH’s profile: https://www.linkedin.com/in/maged-habib-9a888634/

Benefits of being in the DiMeN DTP:
This project is part of the Discovery Medicine North Doctoral Training Partnership (DiMeN DTP), a diverse community of PhD students across the North of England researching the major health problems facing the world today. Our partner institutions (Universities of Leeds, Liverpool, Newcastle and Sheffield) are internationally recognised as centres of research excellence and can offer you access to state-of the-art facilities to deliver high impact research.
We are very proud of our student-centred ethos and committed to supporting you throughout your PhD. As part of the DTP, we offer bespoke training in key skills sought after in early career researchers, as well as opportunities to broaden your career horizons in a range of non-academic sectors.

Being funded by the MRC means you can access additional funding for research placements, international training opportunities or internships in science policy, science communication and beyond. See how our current DiMeN students have benefited from this funding here: http://www.dimen.org.uk/overview/student-profiles/flexible-supplement-awards

Further information on the programme and how to apply can be found on our website:
https://bit.ly/3lQXR8A