A Public Database for the Evaluation of Fundus Image Segmentation Algorithms

Abstract

 

Purpose:

 

The quantitative evaluation of automatic fundus image segmentation methods needs a database with manually labeled gold standards. Only a few online databases are available with outdated low resolution images and gold standards for vessel segmentation only. Our aim is to establish a database of high resolution fundus images with a gold standard for most kinds of segmentation algorithms. This includes the differentiation of arteries and veins, localization of the macula, and the optic disk.

 

Methods:

 

The images are taken with a CANON CF-60UVi camera with a resolution of 3504x2336 pixels. The labeling is done by experts and is corrected by ophthalmologists. The fundus images and the gold standards will be available at a web page "www5.informatik.uni-erlangen.de/research/data/fundus-images", and will be free to use for research purposes. Currently the database contains 15 images of healthy subjects with gold standards for vessel segmentation. An example is shown in the figure. We intend to expand the database with additional images of healthy eyes, glaucomatous and diabetic retinopathy images. We evaluated our already published vessel segmentation methods using the proposed database, and compared the results to the well-known DRIVE database.

 

Results:

 

The first author's vessel segmentation method achieved an accuracy of 95% with a sensitivity of 74% for the DRIVE database. On the proposed database, these values change to 94% accuracy and 70% sensitivity.The second author's method achieved an accuracy of 95% with a sensitivity of 75% for the DRIVE database, and 95% accuracy with 80% sensitivity using the proposed one.

 

Conclusions:

 

We provide a high resolution fundus image database for the evaluation of segmentation methods, and establish a homepage where any author can compare his results to other authors.  

 

Keywords: grouping and segmentation • image processing • blood supply