Automatic detection of reticular drusen using multimodal retinal image analysis

M.J.J.P. van Grinsven, G.H.S. Buitendijk, C. Brussee, B. van Ginneken, T. Theelen, C.C.W. Klaver and C.I. Sánchez

in: Association for Research in Vision and Ophthalmology, 2014


Purpose: Reticular drusen (RD) have been shown to be associated with a high risk of progression to neovascular age-related macular degeneration. RD identification is challenging due to their subtle characteristics on fundus images, especially using a single imaging method. We report a machine learning system to automatically identify RD using multiple retinal imaging modalities. Methods: Color fundus photographs (CFP), fundus autofluorescent images (FAF) and near-infrared reflectance images (NIR) of 175 eyes of 158 patients with presence of either reticular drusen, soft distinct/indistinct drusen, or no signs of drusen were selected from the Rotterdam Study, a population-based cohort. A machine learning system was developed to automatically identify eyes with presence of reticular drusen. First, semi-automatic multimodal affine image registration was performed. After this, features based on Gaussian moments were calculated on the red, green and blue color channels of the CFP as well as on the FAF and NIR images and combined using a random forest classifier to make a classification. Evaluation was performed by comparing the system output with annotations made by an experienced human grader. Results: The human grader identified 44 eyes with reticular drusen, 78 eyes with soft distinct/indistinct drusen and 53 eyes without drusen in the dataset. The system was evaluated using Receiver Operating Characteristics (ROC) curve analysis in a leave-one-out cross-validation scheme. The proposed system was able to identify images with RD with an area under the ROC curve (AUC) of 0.849 and highest accuracy of 0.857. An AUC of 0.887 was obtained for the differentiation between images with RD and images without drusen; whereas an AUC of 0.834 was achieved if the task was to distinguish images with RD from images with soft distinct/indistinct drusen. Conclusions: A machine learning system, using information of different retinal imaging modalities, was developed for the identification of patients with reticular drusen. The proposed system achieved good performance and allows for a fast and accurate reticular drusen detection using several imaging modalities in an automated way.