Grayscale, preprocessed IR815nm, Green532nm, and FAF images from each eye were automatically classified by using unsupervised K-means cluster analysis and commercially available satellite imaging software (
http://www.pcigeomatics.com/; provided in the public domain by PCI Remote Sensing, Markham, Ontario, Canada) with the following parameters: 16 maximum classes, 16 maximum iterations, and a 0.01 minimum threshold (
Fig. 1B, panel 3). The principles of cluster analysis have been described elsewhere.
25,26,30 Although other clustering algorithms and parameters were explored, we describe herein the best combination as judged by the authors, using simultaneous comparison against the reference fundus photograph (
Supplementary Fig. S2). Class separability was calculated by using transformed divergence (D
T; KCLUS algorithm, PCI Remote Sensing), as previously described.
31,32 Briefly, D
T describes a method of calculating the separability or statistical difference (correct classification) between spectral signatures. It derives from measuring the difference in pixel values between all pairs from a predefined number of groups. D
T measures are presented as a value between zero and two: zero indicates extremely poor classification accuracy or overlap between signatures of two classes, while two indicates complete separation between the two classes (analogous to 100% classification accuracy).
32 Pattern recognition theory was originally developed for satellite image analysis and encompasses clustering analysis and tests for cluster separability.
29 The cluster separability indices include D
T, which is effectively similar to separability indices reported by univariate signal detection theory used in psychology, but D
T is able to report separability in N dimensional space. A larger separability value equates to better classification accuracy, and the statistic is resistant to relatively large deviations from normality. Increasing the number of data channels only increases the value of the statistic if the additional channel contributes further separability.
30