Non-Vascular Lesion Detection

E. Grisan and A. Ruggeri. A Markov Random Field Approach to Outline Lesions in Fundus Images. ECIFMBE 2008, IFMBE Proceedings 22, pp. 472–5, Springer-Verlag, Berlin Heidelberg 2008.

Due to its blood microcirculation, the retina is one of the first organ 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 haemorrhages and microaneurysms (HM), 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 he burden of images examination by retinal experts. In this work we propose a simple and effectivemethod to detect and identify these lesions in retinal images, by a two stage classifier. By considering a pixel-wise classification at the first stage and an object-wise classification at the second, it impose a hierarchy (or a multi-level) geometry to the classifier. By using a Bayesian MAP classifier for the first stage and a simple linear discriminant for the second, it proved to achieve a sensitivity of 0.83, 0.71, 0.73 for the classes HE, HM, CWS respectively, with a specificity of 0.94, 0.99, 0.91. The area of the identified lesions shew a correlation with the clinical severity grading of 0.87, 0.83, 0.78.

Pixel Bayesian Classification

Slices of the probability density functions of the three classes of interest

Connected Regions Extraction

Evaluation of Region-Wide Features

Features-Based Region Classification


return to top