Fig. 5: Composite of B-scan images of example cases with disagreement between multiple graders.
From: Automated identification of clinical features from sparsely annotated 3-dimensional medical imaging
Top row: IHRF, middle row: SDD, bottom row, hDC. In a, Aan IHRF is clearly visible (white circle) but is in a region of atrophy. Some graders excluded consideration of the feature as a result. This finding was correctly detected by the algorithm. b A tiny brighter dot (arrow) is observed in the ELM band. This was interpreted by some graders as a possible IHRF. However, the feature is too small and the reflectivity is not as bright as the RPE band. This finding was correctly excluded by the algorithm. c, d The EZ has a slightly “wavy” profile suggestive of possible underlying subretinal drusenoid deposits (within the white circles). In both these cases, the algorithm correctly identified the presence of these subtle SDD. e The drusen (white arrow) is relatively small and its height is borderline for being ≥40 µm, which is the minimum threshold set by the grading protocol in order to be able to assess internal reflectivity. Graders disagreed with regards to whether the lesion met the size criterion. f The internal reflectivity of the drusen is slightly reduced but is clearly brighter than the vitreous overlying the retina. The reflectivity is not sufficiently reduced to be confident that a hDC is present, and hence the disagreement between graders. IHRF intraretinal hyperreflective foci, ELM external limiting membrane, EZ ellipsoid zone, SDD subretinal drusenoid deposits, hDC hyporeflective drusen core.
