Instead of matching image regions directly to the parameterized models in
equations 1 and
2, which would require matching to a large number of lesion models with different

, we created a
compact filter set derived from a large number of instances of the models, as follows: instances of lesion models were generated from
equations 1 and
2, setting β = 2 and δ = 1, based on preliminary studies, and by varying sizes α
1 and α
2 and rotation θ. Once generated,
lesion model instances were grouped according to their sizes α
1 and α
2, and within each group, a single circular analysis image region was used. The radii for the three scales were 7, 13, and 26 pixels (∼35, 65, and 130 μm, respectively), three times as large as the typical lesion size in each scale group, as this was the optimal region size for microaneurysm detection in our previous work.
29,30 Thus, we obtained 4080 lesion model instances (1360 for each scale), up to a size of 2096 pixels, 1048 in each transformed image (after convolution with the Haar filters). Principal component analysis (PCA) was then applied to the 1360 model instances of each scale. See
Figure 2B for an example of the resulting first few principal components for the 13-pixel scale lesion models. The number (
n) of principal components retained for the compact filter set was chosen so that 99% of the variance in the model instances was retained, resulting in
n = 13, 26, and 46 filters, respectively, for the three different scales.