Screening programs for the detection of diabetic retinopathy are being introduced in the UK and elsewhere. The use of automated grading software for this purpose has been proposed and shown to perform with sufficient accuracy. Separately, the use of watermarking for authentication is gaining traction in medical imaging. Much work has been conducted on reducing the perceptibility of watermarking to the human visual system, though little has investigated the perceptibility to machine vision applications. This paper describes a study on the effect of watermarking of retinal images on the automated detection of image clarity and disease. Various watermarking algorithms at a range of embedding capacities are compared. Results are presented for differences in clarity and number of microaneurysms (MA) detected for watermarked images compared to the originals. Sensitivity and specificity of disease and inadequate clarity detection are presented, with a significant reduction in specificity found for generalized least significant bit embedding in both MA and clarity detection.
