Purchase this article with an account.
Li Tang, Meindert Niemeijer, Michael Abramoff; Large Retinal Hemorrhage Detector for Automated Diabetic Retinopathy Screening System. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5527.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To automatically detect large retinal hemorrhages in fundus photographs, which account for one of the main causes of false negatives (FNs) in existing diabetic retinopathy (DR) screening systems. Elimination of those FNs is highly desirable in increasing the reliability of screening systems which can be translated into practice safely.
1200 images from 1200 patients in the publicly available Messidor (http://messidor.crihan.fr) were used, which contains 11 expert labeled images with large hemorrhages. An additional 20 images with large hemorrhages were extracted from the false negatives from 34848 patient exams in the Eyecheck fundus image dataset. Our large hemorrhage detection method has the following steps: the RGB image is transformed into PCA color space adapted to the color distribution of the fundus image. Point clouds clustered in the RGB color space were stretched in the PCA color space to improve color discriminability. Two principal image planes were combined to enhance hemorrhage contrast. Difference of Gaussian (DOG) pyramid kernel transform was applied with combinations of scale pairs. 6 scale pairs were obtained from 4 large scales, and normalized to have unit energy. Summation was binarized to label hemorrhage-ness. Fovea was excluded automatically. Positive connected regions larger than a threshold area were identified as large hemorrhages.
The dataset contained 1220 images, 31 of which contained one or more large hemorrhages. At a preset area threshold of 500 pixels, the detector identified 12 of 20 large hemorrhage images from the added 20, and 6 of 11 large hemorrhages, as well as 74 false positives. The large hemorrhage detector achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.77.
The large hemorrhage detector was capable of finding more than half of large hemorrhage, containing images that were previously missed by an automated system. Integration of the hemorrhage detector boosts the system performance by identifying challenging cases of large hemorrhages.
This PDF is available to Subscribers Only