Artifacts, Segmentation Errors & Documentation

Artifacts, Segmentation Errors & Documentation Standards

Every OCT platform makes algorithmic mistakes. Automated segmentation — the software that draws layer boundaries — fails in predictable patterns when disease disrupts the layer architecture that segmentation algorithms were trained on. When you cannot distinguish real pathology from algorithm error, you risk misdiagnosis in both directions: missing real disease or treating normal variation.

Common Segmentation Errors

OCT segmentation algorithms identify layer boundaries by detecting transitions between hyper- and hypo-reflective bands. When disease creates unusual reflectivity patterns — or when image quality degrades — algorithms fail in characteristic ways:

The most important rule: always cross-reference the B-scan boundary lines when en face findings look unusual. En face pattern inconsistency (e.g., drusen appearing on the GCL slab) is almost always a segmentation artifact, not pathology.

Shadow & Projection Artifacts

Shadow artifacts arise when a hyper-reflective structure anterior in the retinal stack absorbs or deflects the OCT beam, creating a dark "shadow" on all structures below it. Projection artifacts arise when flow signal from anterior vessels is projected onto deeper OCTA slabs, creating false flow foci.

• Vessel shadows: The large inner retinal vessels cast vertical dark shadows on all deeper layers. On the RPE en face slab, these appear as dark arcs or streaks radiating from the disc — do not misinterpret as GA or atrophy. They follow the vessel course on fundus photography exactly.
• Hard exudate shadows: Calcified or lipid-laden exudates are highly hyper-reflective and cast dense shadows. On en face, a hard exudate produces a bright spot (at its depth level) with a dark shadow column on all slabs below.
• OCTA projection artifacts: Superficial vessel flow signal projects downward onto deeper OCTA slabs. On the outer retinal OCTA slab, this creates linear bright structures that mimic CNV flow. The pattern follows inner retinal vessel topography exactly — a dead giveaway of projection artifact.
• Sub-retinal blood: Fresh sub-retinal hemorrhage is highly absorbing and creates broad shadow zones. En face imaging below the level of hemorrhage will show a large, irregular dark zone — do not interpret as GA. Correlate with fundus exam.

Signal Strength Pitfalls

OCT image quality is quantified as a signal strength (SS) index, typically on a scale of 0–10. Signal strength affects every interpretation decision — but the relationship between signal strength and image reliability is not linear, and there are important nuances:

• SS below 5: Segmentation is unreliable. Layer boundaries may be misidentified. En face images will show increased noise, reduced layer contrast, and higher false segmentation error rate. Rescan before interpreting.
• SS 5–6: Marginal. Segmentation may be acceptable for central foveal analysis but unreliable at the periphery. Note the signal strength explicitly in your documentation.
• SS 7 and above: Generally acceptable. Segmentation reliable for most clinical purposes. OCTA requires SS ≥7 for reliable flow detection.
• Focally reduced signal: Even high overall SS scans can have focal signal reduction — posterior staphyloma, dense vitreous floaters, or media opacity can degrade signal in one sector while the rest of the scan is adequate. Check B-scan appearance across the entire cube, not just the central lines.
• Corneal or lens opacity: Media opacities reduce signal uniformly but variably by wavelength. SS-OCT (1050 nm) penetrates media opacity better than SD-OCT (840 nm) — if a patient's lens or cornea is compromising the signal, note it and consider whether the platform is appropriate.

Manual Correction Techniques

All major OCT platforms provide manual segmentation correction tools. Knowing when and how to use them is as important as recognizing the error.

• When to correct: Correct segmentation when the error affects a clinically relevant measurement or when the en face image is obviously displaced from its intended layer. Minor peripheral segmentation wobble that does not affect the central 3 mm generally does not require correction.
• ILM correction: Available on Zeiss Cirrus, Heidelberg Spectralis, and Topcon Triton. On Spectralis, the "Manual Layer Editing" tool allows point-by-point ILM adjustment. Move the boundary to where the true ILM surface is visible on the B-scan, then re-export the en face image.
• RPE correction: Most critical for GA measurement. Correct the RPE boundary to the inner edge of visible RPE, not to the outer border of the hypo-reflective zone below GA (which may be choroid, not sub-RPE space).
• Slab re-export: After manual correction, always re-export the en face image — the displayed en face does not automatically update in all platforms until explicitly refreshed or re-generated.
• Documentation: Note in your report when manual segmentation correction was performed. "En face image after manual ILM correction at foveal ERM traction site" provides clinical context and protects against misinterpretation by subsequent reviewers.

Documentation Standards

Consistent documentation of OCT findings — including segmentation decisions, artifact identification, and image quality — is the foundation of reliable serial monitoring. Without documentation standards, comparison across visits becomes unreliable even when the imaging is technically excellent.

• Minimum documentation elements: Signal strength index, scan protocol used (cube size, line spacing), segmentation correction status (auto or manual), any artifacts noted, and the slab depth used for en face images if different from software defaults.
• Artifact notation: Explicitly note artifact presence in the report when it affects interpretation: "Vessel shadow artifact across superior temporal quadrant of RPE slab — the dark arc at the arcade represents shadow, not atrophy."
• Serial comparison: Use the software's serial comparison or "follow-up" mode to ensure the same scan position, orientation, and slab parameters are used across visits. Manual re-positioning introduces measurement variability.
• GA measurement: When measuring GA area on en face, specify whether measurement is automated or manual, the boundary criterion used (e.g., RPE signal drop to baseline noise), and whether ORT or subretinal blood was excluded from the measurement zone.
• Report language: Avoid ambiguous terms like "possible mild artifact" — state specifically: "ILM segmentation error at temporal ERM margin corrected manually; en face re-generated and interpreted from corrected segmentation." Precise language protects clinical decision-making and supports medico-legal documentation.