Automated Image Quality Assessment Is an Essential Step for Automating Diabetic Retinopathy Detection We have already develped autom We have already develped automated grading software for detecting diabetic retinopathy using microaneurysm detection (Hipwell et al. [italic]Diabetic Medicine[/italic] 2000; [bold]17[/bold]:588-594). However upto 10% of images may be of inadequate quality and it is essential to seperate these images prior to automated analysis or else patients with retinopathy may be missed generating false negatives. We developed and evaluated the the effect of a a quality assessment software that checks both field definition and image clarity of retinal photographs. Anonymised 45[deg] Disc/Macula fundus photographs were obtained from the Grampian Diabetes Retinal Screening Programme. A training set of 1067 images was used to develop the automated grading algorithms and tested on [bold]14,406[/bold] images from [bold]6,722[/bold] patients. All images were graded for image quality and retinopathy by the gold standard. The mean age of patients screened was 63 years ([plusmn]15 years) and 3725 (55%) were male. The gold standard classified 8.2% of the patients as technical failures and 62.5% as having no retinopathy. The quality preprocessor was able to identify [bold]99.8%[/bold] of the patients with poor quality images (technical failures). It also flagged up 17% of patients with mild retinopathy and 3% of patients with observable or referable retinopathy that were classified as having no retinopathy. The combined automated quality assessment and microaneurysm detection software had a sensitivity of 90.5% for any retinopathy and [bold]97.9%[/bold] for referable retinopathy. Hence image quality assessment is an essential step prior to implementing automated microaneurysm detection as a screening tool for diabetic retinopathy.
