Abstract

Purpose: The classification of retinal vessels as arteries (A) and veins (V) is an important step in computer analysis of cardiovascular diseases (e.g. diabetic retinopathy (DR), hypertension) that have different effects on the different vessel types. In this work it is demonstrated that by comparing features of closely running blood vessels A and V along the main retinal arcades (RAs) can be distinguished, as A appear brighter and slightly narrower than V. Methods: DRIVE dataset was used to train and test the algorithm. Vessels running along RAs were detected by ellipse fitting (EF) on sample coordinates of the main vessels. Variations between diameters and pixel intensities for each pair of adjacent vessels were gathered for a supervised training stage. 4023 samples (1394 A-V, 1394 V-A, 430 A-A, 806 V-V occurrences) were selected and probability density functions (pdf) were estimated by kernel density estimation (KDE). Under the assumption of independent and identically distributed (iid) features, maximum likelihood (ML) was computed by implementing an expectation maximization algorithm (EMA). Assuming that there was equal probability (P) for a segment (S) to be A or V, every time
S appeared in a classification window (CW), P was updated by applying the Naïve-Bayes theorem. At the end ML was selected to de ne whether S was A or V.
Results: During testing, results assessment was done by comparing to the manual labelling provided with DRIVE dataset. 1419 S were considered as located along the RAs and thus investigated and achieving a classification accuracy of 83.3%.
Conclusions: Results were consistent with the initial hypothesis. Sometimes pathological conditions (e.g. DR) and bad image illumination invalidated the initial assumptions. Next, vascular network topological information will be added, as we believe that this will reduce error due to BV overlap.