Computer Assisted Fundus Color Image Analysis for Diagnosis of Diabetic Retinopathy

Abstract

Purpose: : The aim of this study is to describe the development and evaluation of a computer assisted fully automatic digital fundus color image analysis for diagnosis of diabetic retinopathy (DR). The main objective was to allow an automatic classification of retinal images into two groups ("normal" or "abnormal") and to provide a DR severity level on a four level-scale.

Methods: : Detection algorithms of the main retinal elements (blood vessels, optic disc and fovea) and of the abnormalities considered as highly representative of different DR stages (microaneurysms, haemorrhages and exudates) were designed and integrated in the computer assisted diagnosis software. This software was developed from a 201 color image learning database, and was evaluated on a 761 image database, acquired on diabetic patients with a Topcon TRC-NW6 non-mydriatic retinograph. These images were taken and annotated manually by a specialist in 2 different ophthalmologic departments. The learning and evaluation data bases contained respectively 61% and 58% of abnormal images. The automatic classification was based on the microaneurysms and haemorrhages detection. The images including at least one microaneurysm or one haemorrhage were considered as abnormal.

Results: : From the 761 images composing the evaluation base, 11% of false positives and 9.5 % of false negatives were detected by the algorithm. Therefore, the DR detection sensitivity is 83.8%, and the specificity is 73.4%.

Conclusions: : The DR prevalence during the screening was approximately 20%. These algorithms could avoid the manual analysis of about 70% of the images and therefore could spare precious medical time. A multicentric evaluation of these algorithms on a larger scale will be carried out, in the framework of OPHDIAT telemedical network, developed for DR screening.

Keywords: diabetic retinopathy • imaging/image analysis: clinical • clinical (human) or epidemiologic studies: health care delivery/economics/manpower