Table 1 Discordant cases were reviewed by a senior retina specialist grader (SS).

From: Automated identification of clinical features from sparsely annotated 3-dimensional medical imaging

Post hoc analysis of discordant cases between algorithm and ground truth

 

Discordant after review

Concordant after review

Observations from post hoc review

IHRF FP (N = 5)

1

4

Small IHRFs could be observed but were close to the minimum threshold size to be included

IHRF FN (N = 4)

2

2

IHRF were in close proximity to the RPE band making separation from the band more difficult to discern

SDD FP (N = 10)

2

8

Poor quality of B-scan images makes it more difficult to separate the SDD from the outer retinal bands (EZ, RPE)

SDD FN (N = 7)

1

6

SDDs very small in size

hDC FP (N = 10)

5

5

Drusen of smaller size making assessment of internal reflectivity difficult. Level of hyporeflectivity was borderline

hDC FN (N = 1)

0

1

Feature missed by grader

  1. Upon re-review the senior retina specialist disagreed with the original ground-truth grading in some cases, but in all discordant cases the findings were borderline. Observations with regards to the cause for difficulty in ground-truth assessment are provided.
  2. FP false positive, FN false negative, IHRF intraretinal hyperreflective foci, SDD subretinal drusenoid deposits, hDC hyporeflective drusen core, RPE retinal pigment epithelium.
Back to article page