Inside Automated Wound Edge Detection: How Computer Vision Traces the Border

Manual tracing is the largest source of inter-rater variability in wound measurement. Two clinicians looking at the same ulcer routinely produce area estimates that differ by 20% or more. Automated edge detection collapses that variance.

What the model actually sees

The capture pipeline runs three passes on every frame:

1. Depth and distance check — the phone's sensor confirms the lens is parallel and within the calibrated focal window before the shutter fires.
2. Segmentation — a convolutional model classifies each pixel as wound bed, periwound, or background.
3. Perimeter extraction — the wound boundary is vectorized so length, width, and area are computed on the polygon, not a bounding box.

Why this beats manual tracing

• Reproducibility — the same wound, captured twice, yields the same perimeter within pixel tolerance.
• Speed — a two-second capture replaces a 60–90 second trace.
• Defensibility — the polygon is stored with the image, so measurements can be recomputed and audited.

Where edge cases still need a human

Heavy slough, black eschar, and macerated periwound can confuse any segmentation model. The workflow flags low-confidence captures and asks for confirmation rather than silently guessing — and undermining or tunneling still require manual entry because 2D imaging cannot see beneath the surface.

From perimeter to trajectory

Once the polygon is stored, every subsequent capture feeds a curve. See how those curves become clinical signal in the wound trajectory deep dive and how the resulting measurements power the broader wound healing software stack.