Skip to main content
. 2025 Dec 17;6(2):101038. doi: 10.1016/j.xops.2025.101038

Figure 5.

Figure 5

Qualitative characterization of agreement and segmentation errors by ReticularNet, by comparison with OCT, for 2 representative visits (rows 1 and 2). The top row shows fair agreement between ReticularNet predictions and ground truth, whereas the bottom row demonstrates poorer performance with multiple false-positives, likely due to drusen being misclassified as RPD in ReticularNet predictions. From left to right in each row (1 and 2): (A) original NIR image with overlay of RPD contour from reading center grading (green); (B) NIR image with overlay of ReticularNet predictions versus ground truth (yellow = agreement; cyan = false-negative; magenta = false-positive), with DSC displayed in yellow at the top right; (C) OCT B-scan; and (D) the same OCT B-scan with vertical bars color-coded for agreement (yellow), false-negative (cyan), and false-positive (magenta). DSC = Dice similarity coefficient; NIR = near-infrared reflectance; RPD = reticular pseudodrusen.