A template matching technique for artifacts detection in retinal images

The continuous development of automatic retinal diseases diagnosis systems based on image processing has shown their potential for clinical practice. However, the accuracy of these systems is often compromised, mainly due to the intrinsic difficulty in detecting the abnormal structures and also due to deficiencies in the image acquisition which affects image quality. Light flares are one of such deficiencies that usually don't compromise the overall image quality, but can be misclassified by an automatic diagnosis system. In this article a method is proposed for detecting light artifacts (flares) on retinal images. The output from the light artifact detection is a binary image mask that is useful to reject those pixels from being further processed. The proposed method uses a template matching algorithm to detect artifacts similar to the predefined template artifact images. Two main types of light artifacts were identified: light flares and the central artifact. To reduce over-segmentation the light artifact candidates are characterized by their shape and color and are classified by a decision tree. The method was developed using a dataset of 61 images from which 20 were used for the classifier training and the remaining 41 for independent testing. With the test dataset the method obtained an average sensitivity/false detection per image pairs of 0.97/0.12 for the central artifact and 0.73/0.36 for the light flares, what were considered good results regarding the heterogeneity of the dataset which contain low and high quality images.