Fig. 4: Examples of discordant cases.

B-scans of example cases where SLIVER-net’s determination disagreed with the expert human graders, with heat map overlay highlighting the most informative regions of the image as determined by the algorithm. Panels a–c show examples of false-negative cases where the feature was detected by the grader on initial review, but not by the algorithm. In a, b, virtually no separation can be seen between retinal pigment epithelial (RPE) band and the drusen, which presumably made it difficult for the algorithm to determine that these were intraretinal hyperreflective foci (IHRF). In fact, on post hoc review, the senior retina specialist sided with the algorithm. In c, the heat map highlights the relevant features, but the algorithm failed to identify these tiny conical or spike-like elevations as subretinal drusenoid deposits (SDD). This was deemed to be a true false negative on post hoc review, yet it should be noted that no clear distinction in reflectivity is observed between the SDD and the underlying RPE. Panels d–f show examples of false positives where the algorithm detected a biomarker but the ground-truth human grader did not on initial review. In d, the heat map highlights a drusen but there are no apparent IHRF. However, there are occasional tiny bright dots in the Henle’s layer which are due to retinal capillaries but may have been confused as IHRF. This is a true false positive. In e, the algorithm detected a drusen with hyporeflective core (hDC), but the drusen was small and <40 µm in height. By definition, graders do not assess the internal reflectivity in lesions this small. The algorithm was able to make this assessment and the internal reflectivity is a bit reduced, but it is a true false positive as it does not match the grading convention. In f, the algorithm also determined hDC to be present, but the internal reflectivity of the drusen, while reduced, is not dark enough to be called hyporeflective. This is also a true false positive.