Purpose:
Our goal is to analyze and visualize the distribution of blood vessels in fundus images to support the medical diagnosis by providing quantitative measurements.
Methods:
15 color fundus images of healthy subjects and 15 images of diabetic retinopathy (DR) patients are acquired by a CANON CF-60UVi camera (3504x2336 pixels). Two set of segmentation results are used for evaluation: manual segmentation by experts and results of an automatic segmentation method developed by the authors. Both the input and the manual segmentation are available online for research purposes (http://www5.informatik.uni-erlangen.de/research/data/fundus-images/). Using the binary vessel images vessel density and distance maps are generated. Afterwards the vessel thicknesses are calculated. The vessel density image is generated by counting the number of vessel pixels in a large neighborhood (radius is 100 pixels). The distance map image encodes distance to the closest vessel. These images support the localization of regions with decreased blood supply. Histograms are generated to show the distribution of the densities, distances and diameters. The statistical moments of these histograms are used in an AdaBoost classifier to discriminate between healthy and DR subjects. 10-fold cross-validation is used to evaluate the classification.
Results:
The accuracy of the classification using the manual segmentations is 93.3%. The area under the ROC curve (AUC) is 0.953. The accuracy for our segmentation method is 80.0% and AUC is 0.931. The figure shows an input image with manually segmented vessels, corresponding density and distance maps and a graph showing the average distribution of vessel densities in both groups.
Conclusions:
We provide methods to analyze and visualize the distribution of vascular tree of the human eye, and a novel classification method to distinguish between DR and healthy subjects using vessel tree based features only.
Keywords: image processing • imaging/image analysis: non-clinical • diabetic retinopathy