Utility of color information for segmentation of digital retinal images: Neural network-based approach

Purpose: The goal of this study was to determine the utility of red, green and blue (RGB) color information in segmenting fundus images for two general categories of retinal tissue: anatomically normal and pathological. The pathologies investigated were microaneurysms and dot blot hemorrhages. Background: Classification, whether performed using parametric or nonparametric techniques, requires that features be selected which allow the desired classes to be partitioned. Features from the data which are useful in separating the classes will improve classification, while those which do not help separate the data can increase computation time and actually lead to lower classification accuracy. One must determine which features are useful for the classification task. Methodology: Digitized 35-mm transparencies were segmented using the neural network-based Digital Fundus Image Diagnostic System (DFIDS) and manually labeled by an ophthalmic technician. Histograms of the sample points in RGB space were used to approximate the class distributions. The shape and proximity of these distributions were analyzed to give insight into possible distribution models, quality of class definitions and labeling, and class separation. Results: The class distributions for the RGB channels appear to be normal in shape. Patient images were found to vary significantly both in appearance and RGB distributions. A large overlap in distributions exists between the RGB feature distributions for the classes of vessel, background, and hemorrhages. The vessel and background classes land similarly the hemorrhage and background) have better separation. Conclusions: There appear to be classes of images which generally have different distributions from one another. For example, highly pigmented images yield different RGB distributions than those of less pigmented images. The presence of significant overlap between the vessel and hemorrhage classes suggests that classification based solely on the information present in the RGB channels will produce poor results. The somewhat better separation presented between the background and vessel/hemorrhage classes may allow for better classification between the background and non-background classes.