Purchase this article with an account.
Xiwei Zhang, Guillaume Thibault, Etienne Decencière, Guy Cazuguel, Ronan Danno, Bruno Laÿ, Ali Erginay, Zeynep Guvenli-Victor, Pascale Massin, Agnès Chabouis; Automatic Detection Of Exudates In Color Retinal Images. Invest. Ophthalmol. Vis. Sci. 2012;53(14):2083.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Screening for Diabetic Retinopathy through a telemedicine network can improve patient follow-up. In the case of the OPHDIAT network, 70% of the images are considered normal by the readers. To reduce the burden on the specialists, the aim of the TeleOphta project is to develop an automatic system to detect normal images going through the network, by combining image processing and data mining methods. The purpose of the study is to describe the detection of exudates, one of the important elements of the TeleOphta software environment.
Firstly, 47 pathological images containing exudates were randomly extracted from the OPHDIAT database, after excluding bad quality images. The exudates were accurately outlined by an ophthalmologist. In addition, 35 healthy images containing structures or artifacts similar to exudates were added to the database. After pre-processing, a morphological ultimate opening function, combined with a local variance threshold extracts the main candidates, while the small candidates are detected using a strain opening function. An original module was also developed to detect optical reflections along the main vessels. Finally a classifier took the final decision: about 30 characteristics were extracted from each candidate, including local, geometrical and textural properties.
The developed algorithm detected exudates in 45 out of 47 pathological images (the two false negatives correspond to images containing a few small exudates and reflections) and 10 false positives out of 35 healthy images. For comparison, the algorithm proposed by Giancardo et al. has been applied to the same database and it detects presence of exudates in all images for both groups.Secondly, the algorithm has been tested on the public DiaRetDB1 database: out of 89 images, there was a consensus reached by four experts on 84 images for the presence of exudates. Our method, without any specific optimization but a modification of a threshold due to a difference of image quality, detects exudates on 22% of healthy images, without false negatives.
The method is efficient for detecting exudates, but tends to consider some artifacts as exudates. The information brought by this detection improves the performance of the data mining system, which would take the final decision.
This PDF is available to Subscribers Only