Abstract
Purpose :
To present and evaluate a multi-stage automated framework for the detection and classification of longitudinal retinal changes due to microaenurysms (MAs) for diabetic retinopathy (DR) screening.
Methods :
Fundus image sets from one eye of each of 82 diabetic patients who were screened for DR in 2012 and 2013 were used for training (30 eyes) and testing (52 eyes) the framework. First, the fundus image sets acquired during successive retinal examinations were normalized for illumination variation and registered into a common coordinate system (Adal et.al.,IOVS,2015). Second, candidate spatio-temporal retinal change locations were extracted by a novel multiscale Laplacian of Gaussian (LoG) algorithm. Third, several intensity and shape descriptors were extracted from each candidate region and subsequently used by a support vector machine (SVM) to classify the region as an MA or a non-MA related retinal change.
The fundus mosaics of each eye were independently annotated by two graders for MA related retinal changes between the two screening time-points. Different ways of combining the two graders’ annotations were used to define a ground truth.
Results :
The performance of the proposed framework was evaluated on the image sets of 52 eyes. The system achieves a sensitivity of 90% in finding MA related changes marked by both graders at an average of 5 false change detections per image set (fig 1). Some of these false detections relate to other dark-red lesions that resemble MAs (fig 2).
Conclusions :
The system is able to detect retinal changes, including those that are visually difficult to detect on the color fundus images. The detected MA related changes can be used as a biomarker for objective assessment of DR progression, such as the MA turnover rate, as well as for more efficient human grading by highlighting DR-related changes since the previous exam.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.