Abstract
Purpose: :
to evaluate the hypothesis that the addition of an automated fovea location system, and of the distance of the fovea to each potential lesion, increases the performance of our system for automated detection of diabetic retinopathy (DR).
Methods: :
A previously unstudied set of 15,000 retinal exams (60,000 digital fundus photographs), with sufficient image quality, were randomly selected from a DR screening project (DR prevalence ~0.05). We determined the area under the Receiver Operator Characteristics curve (AUC), of the standard system using the published feature set [Abramoff et al, Diabetes Care 2008] as well as of this system augmented with additional features using the fovea location, in comparison to the reading of a single retinal specialist (referable DR - no referable DR). The added features consisted of the maximum and average lesion probabilities of for bright and red lesions within a circle with radius 50 pixels around the fovea and within a circle with radius 100 pixels around the fovea.
Results: :
Without fovea distance features, AUC was 0.900, and with fovea distance, AUC was 0.902, a non-significant difference (p<0.8931).
Conclusions: :
In this preliminary study, no performance improvement resulted from adding distance of potential lesions to the fovea. The location of diabetic retinopathy (DR) lesions with respect to the fovea, especially for exudates, is an important characteristic for human experts evaluating color retinal images for DR. Central lesions are easier to detect given the decreased pigmentation variation around the fovea compared to more peripheral, and an exudate of a certain size is more alarming when it is closer to the fovea than if it is more peripheral. This result is therefore counterintuitive. Possibly, the number of cases with a small number of exudate-like lesions close to the fovea is too small in this dataset.
Keywords: diabetic retinopathy • imaging/image analysis: clinical • diabetes