J2 f x

xeEx(Ci)

• The color: We have already seen in the section 5.2 that the green channel contains the most important information about blood-containing elements in the retina and this is why it is used for the detection of microaneurysms. However, there is also some information in the red, and sometimes in the blue channel. We have studied a lot of color features; the most efficient are the following two:

1. Color Contrast in the Luv color space: In the Luv color space, the euclidean distance can be seen as the "true" distance, i.e. the perceptible distance. We used, therefore, the euclidean distance between the color on the candidate region and the color on its external gradient:

contre (Ci) = [(^ext(L) - ^int(L))2 + (^ext(u) - Mint(u))2

2. Contrast of the principal components of the red and the blue channel: In order to find color information complementary to the information in the green channel, we use the principal component cprb of the blue and the green channel as a feature:

These two features do not depend strongly on each other. They help identifying some false positives, but their efficiency is limited.

For the classification, a KNN-classifier is used (K-nearest neighbors), for it has been shown to work well even if there are outliers [23,24]. We do not detail this method, for it is a standard method of classification.

As training set, we used a set of 16 images. We asked two ophthalmologists to mark the microaneurysms independently and then to compare and discuss their results. They finally agreed on 201 microaneurysms; this has been taken as a golden standard. Our algorithm was then applied on these images; 924 candidates were found. Among them were 199 true positives. These candidates were used to train the classifier.

0 0

Post a comment