Automatic detection of microaneurysms in color fundus images for screening of diabetic retinopathy.

Abstract

Abstract: : Purpose: To present an automated fundus photographic image–analysis algorithm for the detection of microaneurysms and the assessment of its performance. Methods:The automatic algorithm works on the green channel of the color image. After a prefiltering step, dark spots are extracted by means of the diameter closing and an automatic noise dependent threshold. Features, calculated for these dark spots, allow their classification into real microaneurysms and false positives. 57 fundus images were used to assess the quality of the automatic detection of microaneurysms by this algorithm. They were taken with a SONY color video 3CCD camera on a Topcon TRC 50 IA retinograph; resolution was 640 x 480. The same set of images was then examined by two independent human graders which pointed manually all the microaneurisms that they could detect. These manual marks were then recorded by the computer and this count formed the reference base. The results of microaneurisms counting, obtained by the automatic algorithm was then compared to this reference base. Results: The 57 images contained 814 microaneurysms as pointed manually (reference base). 16 of the 57 images contained less than 4 microaneurysms. The comparison between the microaneurism counting provided by the algorithm and the reference base gave a mean sensitivity of 88.1% with 2.3 false positives per image. Conclusions: In the framework of computer assisted mass–screening of diabetic retinopathy, an algorithm for automatic detection of microaneurysms in color fundus images has been presented. The results have been compared to the results obtained by two specialists with satisfying results.

Keywords: diabetic retinopathy • detection • imaging/image analysis: clinical