london euretina

This meeting has been awarded 20 CME credits

Security Notice

Please note that Kuoni are our only destination management company. Other venders indicating that are operating for the society should be ignored. We never use western union as a payment portal

The automated detection of longitudinal change in diabetic retinal screening images

Session Details

Session Title: FP-14 Vascular Diseases and Diabetic Retinopathy IV

Session Date/Time: Saturday 13/09/2014 | 16:30-18:00

Paper Time: 17:10

Venue: Boulevard F

First Author: : A.Ahmed UK

Co Author(s): :    R. Staff   A. Fleming   J. Olson   G. Prescott   P. Sharp   S. Philip

Abstract Details

Purpose:

347 million people worldwide had diabetes in 2012 and the number continues to increase. Diabetic retinopathy is the most common vascular complication and, due to risk of vision loss, early detection is critical. The development of features such as microaneurysms and dot haemorrhages among others are signs of retinopathy. While there is considerable research on automatic analysis of retinal images for the signs of retinopathy, there is limited research aimed at studying longitudinal change. This study aims to assess and classify changes over time. Characterising change has the potential to predict future disease progression and consequently inform management.

Setting:

Longitudinal data was acquired from the Scottish Diabetic Retinal Screening service and analysed at the Aberdeen Biomedical Imaging Centre, University of Aberdeen.

Methods:

In this study we used locally written software (publicly available and CE marked) [1] for the extraction of retinopathy features and the Generalized Dual Bootstrap-ICP (GDBICP) fully-automated 2D image registration algorithm [2]. We attempted to register and characterise 214 image pairs (250 images) from 18 patients. The interval between successive images from a particular patient was approximately 12 months. Using the fully automated feature extraction software we extracted the number and the locations of microaneurysms (MA) for each image in its native (unregistered) coordinate. Then using GDBICP we registered follow-up images to the corresponding baseline image. Using the calculated transformation we transformed the follow-up MA coordinates so that they could be compared with the successive image results. For each successive image we calculated the number of new MA’s (MA in the follow-up image not in the previous scan); unchanged MA’s (MA both the follow-up and previous image at the same location); resolved MA’s (MA’s in the previous image not in the follow-up image).

Results:

In this initial study a total of 250 image (214 pairs) registrations were attempted using the GDBICP application; the application successfully registered 201 pairs and failed to register 13 image pairs (6% failure rate). once our software automatically calculated changes in number of MA’s for a patient at approximately 12 month intervals, a regression indicated that the total number of MAs at any given time point remain constant with time due to a combination of new ones appearing and old ones resolving. The rate turnover (new + resolved) increased with time (.06 MA’s per month). In addition the rate of turnover was associated with the total number present at any time point (.59 MA’s per MA present).

Conclusions:

Generalized Dual Bootstrap-ICP (GDBICP) is a fully-automated 2D image registration algorithm. It is designed to register two images taken of the same scene, although perhaps at different times and from different viewpoints. [1]. Using GDBICP’s transformation algorithm, which is unique to each individual transformation, we have successfully automated the registration of a patient’s sequential retinal images so allowing us to automatically assess disease change from one image to the next. Our preliminary conclusions are that this pilot data shows MA’s pathology in the diabetic eye is a dynamic process with MA’s appearing and disappearing across time. But in order to obtain clinically meaningful results on the value of MA turnover we intend to apply this approach to a large data cohort of approximately 50.000 images. This will allow us to analyse patterns to establish whether automatic disease prediction is possible.

Back to previous
EURETINA, Temple House, Temple Road, Blackrock, Co Dublin. | Phone: 00353 1 2100092 | Fax: 00353 1 2091112 | Email: euretina@euretina.org

Privacy policyHotel Terms and Conditions Cancellation policy