Abstract
In the last years the research community has
developed many techniques to detect and diagnose diabetic
retinopathy with retinal fundus images. This is a necessary step
for the implementation of a large scale screening effort in rural
areas where ophthalmologists are not available. In the United
States of America, the incidence of diabetes is worryingly increasing
among the young population. Retina fundus images of
patients younger than 20 years old present a high amount of reflection
due to the Nerve Fibre Layer (NFL), the younger the
patient the more these reflections are visible. To our knowledge
we are not aware of algorithms able to explicitly deal with this
type of reflection artefact.
This paper presents a technique to detect bright lesions also
in patients with a high degree of reflective NFL. First, the candidate
bright lesions are detected using image equalization and
relatively simple histogram analysis. Then, a classifier is trained
using texture descriptor (Multi-scale Local Binary Patterns) and
other features in order to remove the false positives in the lesion
detection. Finally, the area of the lesions is used to diagnose diabetic
retinopathy.
Our database consists of 33 images from a telemedicine network
currently developed. When determining moderate to high
diabetic retinopathy using the bright lesions detected the algorithm
achieves a sensitivity of 100% at a specificity of 100% using
hold-one-out testing.