Inside Automated Wound Edge Detection: How Computer Vision Traces the Border
aiengineering
A look under the hood at the computer vision pipeline that finds a wound's border, handles ambiguous edges, and returns defensible measurements.
Auto-tracing a wound edge looks like magic until you try to do it on macerated periwound skin at 6pm in a dim exam room. Here's what actually happens between the shutter click and a signed measurement.
The capture stage
Before any segmentation, the image has to be worth segmenting. On-screen guidance enforces distance, angle, and focus. If the shot is off, the app asks for a retake — bad inputs make bad borders.
Segmentation
A deep segmentation model, trained on tens of thousands of annotated wound images across skin tones, wound types, and anatomic sites, proposes a pixel-level mask of the wound bed. The model is calibrated to over-segment slightly at ambiguous edges rather than clip, because clipping under-reports area — and under-reported area under-supports reimbursement.
Calibration
A depth signal (from the 3D reconstruction or a fiducial) converts pixels to millimeters. Without calibration, every measurement is a guess in the wrong units.
Post-processing
The raw mask is smoothed, holes are filled, and the longest chord defines length; the perpendicular longest chord defines width. Surface area is computed from the mask directly, not from length × width, so irregular wounds aren't over-estimated.
The human in the loop
Edges get ambiguous — undermining openings, satellite lesions, tunneling. The tool exposes control points the clinician can nudge, and the signed border is what enters the chart. The model assists; the clinician decides.
Why it matters over time
Consistent segmentation across visits is what turns a photo album into a healing trajectory. Measurements land in the healing dashboard automatically, and the underlying imaging pipeline is the same one described in AI-powered wound imaging.