Abstract
Purpose: :
Telemedicine and computer assisted analysis can aid in the early detection of diabetic retinopathy and other diseases in large patient populations. We have developed an automated system to diagnose macular edema (ME) in retinal fundus images. The algorithm was developed for a telemedicine network of primary care clinics based in Memphis, TN, North Carolina, and Mississippi. We present a comparative study between the automated diagnosis system and the diagnosis of two retina specialists.
Methods: :
Two publicly available datasets are employed: DMED and Messidor. The former is a set of 169 fundus images collected from our telemedicine network which were used to train and develop the automatic diagnosis algorithms. The latter set consists of 1200 images collected and diagnosed by outside experts; these are employed to determine the automatic and the retina specialists’ diagnosis. The ME diagnosis algorithms are based on a high-speed computer-based "hard exudates" detection. The algorithms detect the field of view, normalize the image, then perform edge analysis and evaluation of the global probability of exudation. The final diagnosis is generated using a pure pattern recognition approach: a novel set of features is computed and employed to train a classifier. When a new image is presented to the system, the classifier produces a diagnosis based on the images shown during the training phase.
Results: :
Two retina specialists determined the presence or absence of exudates in order to diagnose ME on a random sample of 300 images of the Messidor dataset, with 104 images exhibiting ME and 196 exhibiting no ME. The automatic system operated on the entire dataset. The reference standard is provided by the Messidor dataset, which has been employed by various research groups. We compared the performance of the automatic system by creating the ROC Curve and overlaid the Specificity/Sensitivity of the two retina specialists. Also we used Kappa value/AC1-statistics as concordance metrics, obtaining: Expert1 0.63/0.73, Expert2 0.62/0.64, Automatic Diagnosis 0.64/0.85.
Conclusions: :
Our approach is a promising method to automatically screen for ME. With this study we demonstrate that the system's performances is comparable to retina specialists. Similar studies on diabetic retinopathy will be conducted in the future to further evaluate the automated system.
Keywords: image processing • edema • imaging/image analysis: clinical