Fig 1.
Example of vertical cup/disc ratio.
Here, the boundaries of the cup and disc were determined using the ORIGA-GT software (modified from [16]). This software generates boundaries by fitting two ellipses using human expert landmark identification and least squares fitting. The cup boundary is given in blue; the disc boundary is in red. In the text, this is referred to as semi-automated segmentation.
Fig 2.
Orientation of the landmarks in the right and the left eye.
The centre of the cup is used for the calculations.
Fig 3.
The profile of 24 cup/disc ratios (pCDR) in two eyes.
One healthy fundus (A) and one glaucomatous fundus image (B) are showed here. The cup and disc were semi-automatically segmented, which is shown by the best-fitting ellipses (C and D). The profile of 24 CDR values were plotted in circular (E and F) and Cartesian systems (G and H).
Fig 4.
The cup/disc ratio profiles (pCDR) of all individual eyes from ORIGA.
Individual healthy (A) and glaucomatous (B) optic nerve images from the ORIGA dataset (n = 650) in circular (C) and Cartesian (D) formats. These profiles come from semi-automated segmentation. The population mean pCDR for healthy (cyan) and glaucomatous (yellow) groups are shown together with the individual pCDR profiles of the two eyes from Fig 3 (black).
Table 1.
Fitted spatial statistical model and association with disease group in the ORIGA dataset.
Fig 5.
Internal validation of the spatial algorithm using automatically segmented images from ORIGA.
A) The grader's semi-automatic segmentation (blue) and the fully automatic segmentation (green). B) The individual automatically segmented profiles with means (thick blue line for healthy, thick red line for glaucomatous). We used the automatically segmented discs and cups to detect glaucoma. C) The AUROC is 99.6%. D) The probability of glaucoma and the decision threshold for 96.6% sensitivity and 99.0% specificity. The size of the testing dataset is n = 163. E) The risk of glaucoma (log(p/(1 − p))) vs Rim-to-Disc at the narrowest rim. F) The Rim-to-Disc at the narrowest rim vs disc size.
Table 2.
Comparison of the spatial algorithm with machine learning (SVM) for the classification of glaucoma.
Fig 6.
External validation of the spatial detection algorithm using the automatically segmented images from RIM-ONE.
Here, all ORIGA-light images were used to train the segmentation and the glaucoma detection. The RIM-ONE images were then automatically segmented and glaucoma detection was tested. A) The grader's semi-automatic segmentation (blue) and the fully automatic segmentation (green). B) The individual automatically segmented profiles of 39 glaucomatous, 85 healthy and 35 suspected optic discs. C) The AUROC in external validation was 91.0% for discrimination between glaucomatous and healthy. The threshold probability of 0.90 (see the circle) yields 89.7% sensitivity and 74.1% specificity. D) The posterior probability of glaucoma in the three RIM-ONE groups with the 0.90 threshold (dashed line). The algorithm identified as glaucomatous: 35 out of 39 glaucomatous (90%), 22 out of 85 healthy (26%), and 13 out of 35 suspected (37%) eyes. E) The risk of glaucoma (log(p/(1-p)) vs Rim-to-Disc ratio at the narrowest rim. F) The risk of glaucoma vs disc size.