Recent advances in the analysis of non vascular features in retinal images

Abstract

Abstract Purpose: To problem of detecting non vascular dark lesions (haemorrhages and microaneurysms) in retinal images for diagnostic purposes is addressed. Methods: The green channel of images is extracted and then normalized by compensating the luminosity and contrast drifts. The detection of clusters of dark pixels is based first on a local thresholding to identify pixels darker than the surrounding, followed by an evaluation of their spatial density, which tells how much the neighborhood of the pixel is dense of similar (dark) pixels. A further threshold is then applied to this density to identify candidate clusters of dark pixels. Results: In a pool of 60 50‐deg fundus images acquired on slide, and then digitized with a 1378 dpi resolution, true colour, 6 images presented microaneurysms or haemorrhages. These images were chosen to evaluate the performance of the proposed method. On average, the algorithm was able to detect at least part of a lesion for 94% of the lesions present in an image. When an image presented only one lesion, so that a hit‐or‐miss situation may compromise the clinical evaluation, the algorithm was able to correctly identify the true candidate. Conclusions: A novel, high sensitivity algorithm for identifying dark lesions in retinal images is presented. It exploits both the local gray‐level information for a first segmentation and the spatial density of segmented pixels to obtain a measure of a region homogeneity. The effectiveness of the algorithm has been successfully tested on six retinal images presenting dark lesions of various shapes and dimensions. This suggest that it can be well suited as first stage of a dark lesion detection system, which will be followed by a classification stage, able to reject possible false candidates.