Improving automated retinal image analysis using deep learning v1
Research type
Research Study
Full title
Improving Automated Retinal Image Analysis using Deep Learning
IRAS ID
280170
Contact name
Sam Philip
Contact email
Sponsor organisation
University of Aberdeen and NHS Grampian Research Governance
Duration of Study in the UK
0 years, 11 months, 28 days
Research summary
Description: Diabetic retinopathy (DR) is a major complication affecting people with diabetes. Retinal scans are analysed by medical practitioners to identify those who have features of referable retinopathy and will require assessment by an Ophthalmologist. In previous work we (Ophthalmic imaging group, University of Aberdeen & NHS Grampian) have developed software (DRIA-2, DRIA-3) using traditional computer vision techniques to perform automated analysis of retinal images acquired in the Scottish retinal screening programme.
Since 2014, Deep Learning (DL) has brought significant improvement in machine learning techniques for image analysis. DL is a branch of machine learning (ML) and artificial intelligence (AI) that uses deep neural networks, i.e. sequences of mathematical operations that take data as input and output a prediction. Such models have achieved outstanding results in image classification and are particularly attractive because they learn the useful features of an image by themselves, including those of diabetic retinopathy. However, the performance of a deep learning model is highly dependent on its architecture, i.e. the exact sequence and configuration of its mathematical operations. Finding the right sequence is a time-consuming and iterative exercise that requires expertise. Crucially, this exercise might be required again should the appearance of images in the dataset change significantly, e.g. due to technological evolution.
We therefore propose to use the datasets accumulated as part of the above project (CZH/4/472) to assess the benefits of leveraging deep learning technologies to improve sensitivity, specificity, efficiency and explainability of existing automated screening algorithms. In particular, to circumvent the issues described above around the manual design of deep learning models, we propose to use Neural Architecture Search, a recent branch of Deep Learning aims at automatically finding the optimal neural network architecture for any given task.REC name
East of England - Essex Research Ethics Committee
REC reference
21/EE/0071
Date of REC Opinion
10 Mar 2021
REC opinion
Favourable Opinion