We have tested the algorithm on an image database of 30 digital images 640 x 480 taken with a Sony color video 3CCD camera on a Topcon TRC 50IA retinograph. These images have not been used for the development of the algorithm. Fifteen of these images did not contain exudates, and in 13 of these 15 no exudates were found by our algorithm. In two images, few false positives were found (less than 20 pixels).

We asked an ophthalmologist to mark the exudates in the 15 images and compared the results obtained by the algorithm to his. The comparison was done pixel-wise (with 1 pixel tolerance), and as for exudates, the number cannot be determined; it is the surface and the position rather than the number which can be used for diagnostic purposes.

We obtained a mean sensitivity of 92.8% and a predictive value of 92.4%. In Fig. 7.35, an example for the automatic detection of exudates is shown (see also Figs. 7.36 and 7.37).

(a) The top-hat image (b) Algorithm result

Figure 7.35: The result of exudates detection.

Figure 7.35: The result of exudates detection.

0 0

Post a comment