A Hierarchical Bayesian Classification for Non-Vascular Lesions Detection in Fundus Images

Due to its blood microcirculation, the retina is one of the first organs affected by hypertension and diabetes: retinal damages can lead to serious visual loss, that can be avoided by an early diagnosis. The most distinctive sign of diabetic retinopathy or severe hypertensive retinopathy are dark lesions such as haemorrhages and microaneurysms (HM), and bright lesions such as hard exudates (HE) and cotton wool spots (CWS). Automatic detection of their presence in the retina is thus of paramount importance for assessing the presence of retinopathy, and therefore relieve the burden of image examination by retinal experts. The first step for the automatic detection of retinal lesions has to identify candidate lesions, not losing any of them, and providing with accurate outlines, so to allow the extraction of meaningful features for a possible subsequent classification. To accomplish this, we propose a two stage approach. The first stage identifies the rough location of candidate lesions with one or more seed points, evaluating a measure of the spatial density of pixels selected by a local thresholding. The second stage has the objective of outlining as accurately as possible the lesions surrounding each seed. Due to the high variability of lesions and background appearance, classical region growing approaches often fail and are difficult to calibrate, since in retinal imaging the noisy and highly variable background hides the small homogenous regions representing lesions. To tackle this, we rely on a stochastic modelling of a region of interest around a seed as a Markov random field: This is particularly suited to separate objects with different textures, since it combines feature distribution and spatial connectivity. Responses to a Gabor filters bank spanning different orientations and scales provide the description of the local texture, and the final classification is obtained via a simulated annealing optimization. Along with the classification, we propose a simple post-optimization measure to discard regions where no lesions are present despite the presence of seeds. We show results of the proposed method on a data set of manually segmented 60 images, 6 of which containing retinal lesions.