CAM-002 Synthetic Auto-Detection Validation¶
Date: 2026-05-26 Issue: #233 Scope: CAM validation for anterior, posterior, lateral derecha, and lateral izquierda views under lighting, skin-tone background, sticker color, and semisphere marker appearance variants.
Data Safety¶
This validation uses only synthetic OpenCV-rendered images: flat anonymous backgrounds plus generated marker circles at EPPA anatomical priors. No real patient data, clinical photos, names, identifiers, secrets, tokens, database dumps, or clinical payloads are included.
Method¶
The harness in python/cam_validation.py renders six synthetic conditions for each supported clinical view and evaluates the existing HSV green-marker segmentation plus anatomical-prior matching from python/detect_markers.py. Each expected marker is classified as:
detected: marker was segmented and matched to the correct expected anatomical point.confused: marker was segmented but assigned to a different anatomical point.missing: expected marker was not recovered.
The metric by condition is precision = detected / expected_markers. The supported views are anterior, posterior, lateral derecha, and lateral izquierda.
Results By View And Condition¶
| View | Expected markers | Bright green stickers | Green semisphere | Under-lit green | Red sticker control | Manual-capture conditions |
|---|---|---|---|---|---|---|
| anterior | 17 | 17 detected / 0 missing / 1.000 precision | 17 detected / 0 missing / 1.000 precision | 0 detected / 17 missing / 0.000 precision | 0 detected / 17 missing / 0.000 precision | 2 of 6 |
| posterior | 19 | 17 detected / 2 missing / 0.895 precision | 17 detected / 2 missing / 0.895 precision | 0 detected / 19 missing / 0.000 precision | 0 detected / 19 missing / 0.000 precision | 6 of 6 |
| lateral derecha | 14 | 12 detected / 2 missing / 0.857 precision | 12 detected / 2 missing / 0.857 precision | 0 detected / 14 missing / 0.000 precision | 0 detected / 14 missing / 0.000 precision | 6 of 6 |
| lateral izquierda | 14 | 12 detected / 2 missing / 0.857 precision | 12 detected / 2 missing / 0.857 precision | 0 detected / 14 missing / 0.000 precision | 0 detected / 14 missing / 0.000 precision | 6 of 6 |
Criteria For Manual Capture¶
The app should ask for manual capture/marking when any of these conditions is true:
- Any expected marker is missing.
- Any marker is confused with another anatomical point.
- Precision for the capture condition is below 95%.
- Anatomical plausibility fails because of extreme asymmetry, improbable vertical ordering, or duplicate marker positions.
Interpretation¶
Bright green stickers and green semispheres are robust for the anterior synthetic priors. Posterior and lateral priors still lose close-together markers in this synthetic harness, so the app should stay conservative and require manual review/capture for those views until matching and real anonymized fixture validation improve. The under-lit and non-green sticker failures are expected and clinically useful: the segmentation refuses weak/non-green markers instead of hallucinating marker positions.
This report is not a substitute for multicenter validation with approved anonymized fixtures. It establishes a reproducible synthetic baseline and conservative manual-capture criteria until real fixtures are explicitly approved.