Abstract:
A key task in the automatic processing and analysis of digital fundus images is to locate the position of the optic disk. We wish to eliminate the optic disk from later segmentation of lesions or exudates, and so reduce the possibility of false classification of lesions. We remark that the optic disk itself may demonstrate features (such as new blood vessels on the optic disk) that are typical indictors of disease (such as diabetic retinopathy).
We present an initial result for optic disk location using an active appearance model (AAM). For more information about AAMs, see, for example, T.F. Cootes, et al., Computer Vision and Image Understanding 61, 38 (1995); T.F. Cootes, et al., IEEE PAMI 23, 681 (2001).
The digital image information regarding the optic disk from the seven subjects was used in constructing an AAM. This was then used in an AAM search (shown in the figure) in order to determine the position of the optic disk in this initial study. The edge of the optic disk predicted by an AAM search is shown in green. We see that an initial guess shown on the left is quickly refined by the iterative search to an accurate final result on the right. (The blue dots are artefacts generated by the plotting program.)
It has been proven that an AAM may be used in order to refine an initial guess for the position of the optic disk. The edges of an optic disk in a digital fundus image are not always well–defined, which may cause problems for, e.g., active snakes, unless one is careful. This should not be the case for an AAM, and so it is imagined that an AAM would be an excellent method of automatic optic disk location. Quantitive assessments of the accuracy and robustness of this approach will now be carried out by determining point–to–point and point–to–curve errors of the predictions of our model with respect to a "gold standard" provided by the sets of "mark–up points" from experts. We will test if either AAM model parameters or features extracted from the optic disk correlate directly with the presence of eye disease (e.g., diabetic retinopathy) by using, for example, ROC curve analysis. We shall also perform neural network classifications based on these results.
Keywords: imaging/image analysis: non-clinical • optic disc • diabetic retinopathy