Automatic detection of eye diseases using automated color fundus image analysis

M.J.J.P. van Grinsven, F.G. Venhuizen, B. van Ginneken, C.B. Hoyng, T. Theelen and C.I. Sánchez

in: Association for Research in Vision and Ophthalmology, 2015


Purpose: Diabetic Retinopathy (DRP) and Age-related Macular Degeneration are the most common visual threatening eye diseases in industrialized countries. Detection of these diseases at an early stage can help to identify patients that would benefit from treatment to slow or prevent their progression into more severe stages with visual loss. We report an automatic software solution for the detection of early stages of DRP and AMD using color fundus (CF) images. Methods: CF images from several public datasets (DiaretB0, DiaretB1, Stare, Messidor, DR1/DR2) and two private data sets were pooled together and used in this study. Only macular centered images with sufficient quality for manual assessment were included. Images in advanced stages of AMD or DRP were excluded. The remaining study set consisted of 2128 CF images. All images were labeled by an expert into one of three classes: early stage DRP, early stage AMD or control. Two existing machine learning systems, one for early stage AMD detection and one for early stage DRP detection, were combined to detect the early stages of DRP and AMD simultaneously. Final classification features of both systems were concatenated and a random forest classifier was trained to make a classification between the target cases (early stage DRP or AMD) and the control cases. Evaluation was performed by comparing the output of the system with the human expert's labels. Results: The human expert labeled 596 and 196 cases as early stage DRP or AMD, respectively. The remaining 1336 CF images were labeled as controls. The automatic system was evaluated using Receiver Operating Characteristics (ROC) curve analysis in a 10-fold cross-validation scheme. The combined system for detection of the early stages of DRP and AMD was able to separate the target cases from the control cases with an area under the ROC curve (AUC) of 0.948. The system achieved a sensitivity of 0.880 and a specificity of 0.886. Conclusions: A machine learning system was developed for the identification of early stages of AMD and DRP. The proposed system achieved good performance and allows for a fast and accurate identification of patients that may benefit from treatment, opening the way to a cost-effective mass screening procedure of patients at risk of AMD and DRP.