OCT segmentation algorithms disagree in neovascular AMD and should not be used interchangeably
A study in Eye compared two automated OCT segmentation programs in 50 patients with neovascular AMD and found clinically meaningful differences, including 43.54 μm in central total retinal thickness.
Retinal layer thickness measured by optical coherence tomography (OCT) is used to monitor neovascular age-related macular degeneration (AMD) in both clinical care and research. But the numbers depend on the algorithm doing the automated segmentation. A study published on November 16, 2026, in Eye compared two of these algorithms.
What was done
Researchers from the University of Turin and other Italian centers, working with the US company Voxeleron, conducted a retrospective cohort study of 50 patients with neovascular AMD. Scans were acquired on a Heidelberg Spectralis spectral-domain OCT.
Retinal thickness was averaged across ETDRS grid subfields centered on the fovea using two methods:
- the Spectralis built-in software (version 1.10.4.0);
- the Orion automated segmentation software (Voxeleron, version 3.0.0).
The main outcome was the difference between the two algorithms for inner retina, outer retina, neurosensory retina, and total retinal thickness.
What they found
- Inner retina: mean difference of 8.97 μm in the central subfield (R² = 0.32), with larger variability in the inner (22.09 to 28.73 μm) and outer rings (16.02 to 29.50 μm).
- Outer retina: mean difference of 21.51 μm in the central subfield (R² = 0.36), with modest variation across subfields.
- Neurosensory retina: a small central mean difference (0.13 μm; R² = 0.66) that grew in the peripheral subfields, with an overall mean difference of 45.30 μm (R² = 0.82).
- Total retinal thickness: a 43.54 μm difference in the central subfield (R² = 0.61) and a mean of 18.16 μm across the rest of the grid (R² = 0.60).
The authors conclude that although correlations between the algorithms were moderate to strong for several metrics, mean differences were clinically meaningful in some compartments and the two methods should not be used interchangeably.
Limitations
- Retrospective design, small sample, and a single OCT device.
- The study measures agreement between algorithms, not which one is more accurate against manual reference segmentation or clinical outcomes.
- Two authors are employees of Voxeleron, the maker of Orion, and hold company patents; other authors disclosed ties to pharmaceutical and device companies.
Why it matters in practice
In neovascular AMD, central thickness and fluid guide retreatment decisions. The findings call for caution when comparing scans analyzed with different software, for example after a platform change, between practices, or across sites in a clinical trial. To follow a patient over time, it is best to keep the same device and the same algorithm, and to check the segmentation when the numbers do not match the image.
Read the original paper in Eye.