Introduction
The sign is easy to miss by eye
Visual identification of digital clubbing is subjective. Myers and Farquhar reported interobserver kappa from about 0.36 to 0.90, with some values near 0.4, and a bedside profile difference of a few degrees [1]. That range is why a careful look at the nail fold can change the call. It is literature, not a LungLens result. It does not prove that a few degrees delay pulmonary diagnosis, and it does not prove that LungLens would have caught a missed sign.
Problem on the project page · 3-minute tour · Myers 2001
Method
Capture. Quality. Landmarks. Geometry. Confidence. Review.
I could have asked a model to output “clubbing.” I did not. AI, if used, only helps locate points. Ordinary geometry measures the unsigned profile angle and depth ratio. uncertainty.py returns Unable to assess if min(landmark, segmentation) confidence is below 0.55. Default stub adapters use fixed confidence 0.88 and 0.86, so a normal stub run does not hit that gate. The live stub safety net is researchOnly plus clinic_safe_observation. A person still interprets. Landmark finding is still experimental, and the default path can return DEFAULT_POINTS.
Engineering writeup · Guided demonstration · AI vs geometry
Results
Software tests, and a system that can refuse a label
A public snapshot on 2026-10-19 shows 17 tests passing (commit 3354f22) across geometry, quality, fusion, and screening labels. Those suites check fixtures and policy, not diagnostic accuracy. The 0.55 abstention branch exists in uncertainty.py. Default stub confidence stays above that cutoff, so a normal stub run still produces a number. Clinic consumers do not receive a morphology label because of clinic_safe_observation. The cutoff is not a calibrated clinical rule, and the October snapshot is not an abstention-accuracy study.
Evidence ledger
Limitations
What is structurally tested, and what is not
The software logic is structurally tested. Clinical accuracy is not. Fairness across pigmentation, lighting, and cameras remains unproven. Landmark finding is still a stub by default. Those studies are the next phase. I measure finger shape for research review. I do not diagnose lung disease.
Responsible design · Research roadmap
