Purchase this article with an account.
Alina V Dumitrescu, Michael David Abramoff, James C Folk, Meindert Niemeijer; Automated Retinal Hermorrhage Detection System. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4829.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Hemorrhage detection is a challenging task in automated diabetic retinopathy screening due to the variability in lesion appearance. In this abstract a new method for automatically detecting the presence of retinal hemorrhages in digital color fundus photographs is evaluated.
The evaluated method is based on an unsupervised learning technique that learns a set of filters that can represent small patches of fundus images with and without lesions. By max-pooling these filter outputs over small 100x100 pixel windows, taken from a fundus image, a set of features can be extracted. Then a classifier can be trained to distinguish windows containing hemorrhages from windows containing normal retina. By moving the window over the fundus image and classifying them, a set of likelihoods for the presence of hemorrhages in the fundus image is obtained. These likelihoods can then be combined using an averaging operation to obtain a likelihood for the complete fundus image. For training the filterbank 200000 randomly sampled 10x10 pixel patches were extracted out of a dataset with 9000 digital fundus images. From these, 400 representative filters were learned. For testing we used the publicly available Messidor dataset. Of the 1200 images in the dataset, 138 were determined to contain hemorrhages by an ophthalmologist. The ophthalmologist marked the location of the hemorrhages in the images. We then extracted 30 100x100 windows from each of the 1200 images, the windows were only extracted from locations where the window fit completely inside the FOV. Each window was labeled as 0 if no hemorrhages were present in the window and as 1 when the number of hemorrhage pixels was higher than 100. The learned filters were applied, and each window yielded 3600 features. On the resulting dataset 3 fold cross validation was then performed. For each image we averaged the likelihoods of the windows from that image above the 40th percentile in likelihood, to obtain the likelihood for the image itself. To assess the performance of the algorithm at the image level, an ROC analysis was performed.
The obtained area under the ROC curve for the presence of hemorrhages in fundus images was 0.832.
The evaluated method represents a promising approach to the detection of the presence of hemorrhages in digital fundus photographs.
This PDF is available to Subscribers Only