Abstract
Purpose: :
Evaluation of pixel feature classification, an automated image segmentation technique, to segment the cup and rim in optic disc stereo color images. The results of the computer algorithm and two glaucoma fellows in training were compared to the results of three glaucoma experts.
Methods: :
102 randomly chosen eyes of 102 patients with glaucoma of varying severity underwent stereo color imaging with the Nidek 3DX camera (Nidek, Fremont, CA). The cup and rim of the optic discs in the stereo pairs were drawn by three glaucoma experts (LMA, YHK and EG) using stereo planimetry in a masked fashion, and each pixel was assigned to either cup, rim or background. Two glaucoma fellows also evaluated cup/rim in the same way. Optimal feature selection was performed on a randomly chosen training set of 51 (out of the 102) stereo pairs to select those features that resulted in an optimal Area under the Receiver Operator Characteristics Curve (AROC) on cp and rim segmentation. The 313 available features included texture, color, edge orientation, depth from stereo, and gradient features at various scales and orientations. When 10 principal features had been obtained, the algorithm the assigned each pixel in the remaining 51 stereo pairs as either cup, rim or background, using a k–nearest neighbor classifier.
Results: :
The cup to disc (area) ratio (CDRa) average was 0.33 (range 0.06 – 0.84). The correlation of CDRa from the automated segmentation with the 3 experts was 0.88 (95% Confidence Interval (CI) 0.80 – 0.93), and from the glaucoma fellows was 0.82 (CI 0.76 – 0.90) and 0.90 (CI 0.93 – 0.95) respectively. The algorithm had a tendency to overestimate CDRa compared to the experts and the glaucoma fellows. The first three principal features explaining were edge at 0 degrees and scale 64 pixels (AROC 0.67), variance in green channel (AROC 0.76), and red radial gradient at 16 pixels (AROC 0.83).
Conclusions: :
A computer algorithm has been used to automatically segment cup and rim in stereo color photographs. Though the pixel feature classification algorithm was not trained on the CDRa, its accuracy on CDRa was comparable to that of glaucoma fellows.
Keywords: imaging/image analysis: clinical • optic disc • depth