Getting the most out of MedGemma — an accuracy roadmap for OpenFootLab

How we go from "it produces plausible observations" to "it produces observations a podiatrist trusts." Grounded in FL-STUDY-001 (Gemma 3 27B vs MedGemma 27B on 35 real foot photos).

Where we actually are

The study told us three things, honestly:

1. Reliability is solved — both models return valid closed-schema JSON 100% of the time.
2. The models disagree on what they see — near-zero κ on redness, swelling, pressure. The

specialization changes the observation, not just the wording.

1. We cannot yet measure accuracy — there is no ground truth on the box. We know the models

differ; we don't know which is right.

So the first move is not a bigger model. It's building the thing that lets us measure accuracy at all — ground truth. Everything else is unfalsifiable without it.

The levers, in ROI order

0. Ground truth first (the unlock)

Accuracy is undefined until we have labels. The Tier-2 blinded clinician packet is not just a verdict — it is a label factory. Every photo a podiatrist adjudicates becomes a golden-set label. Target: a few hundred clinician-labeled foot photos → the golden set and the regression harness. From then on, "better" means higher agreement with the golden set, tracked per model version. No golden-set movement, no accuracy claim. (Skills: golden-set, eval-curator, local-study.)

1. Capture quality — the biggest cheap win

MedGemma normalizes every image to 896×896, 256 tokens. A blurry, oblique, badly-lit photo is information-destroyed before the model sees it. Standardize capture and accuracy moves more than any model swap:

• consistent white-matte background, even lighting, fixed distance/angle (already in the

baseline protocol),

• a scale/color reference in frame (a sticker) so size and redness are calibratable,
• multiple standard views per session, not one hero shot (SENTINEL 4-view: dorsal, plantar,

medial, lateral + the site close-up). One angle hides half a foot. (Skill: sentinel-capture.)

2. Give the model context, not a lone frame

• Prior image for comparison (the ComparisonEngine / vault-longitudinal) — change detection

is the whole point, and a model reasons better about "vs. yesterday" than about an absolute.

• The patient profile in-context — "diabetic, insensate, post-amputation" vs "marathon

runner, intact sensation" should change how a red patch is weighted.

• Multi-view in one call — feed the standard views together so the model sees the whole foot.

3. Prompt + few-shot (cheap, but gated)

MedGemma is prompt-sensitive. In-context reference exemplars ("this image = moderate redness," "this = an open wound") anchor the closed vocabulary far better than words alone. Every prompt change is pinned, hashed, and must pass the golden set before it ships (canary-then-cook).

4. Keep the division of labor (already right)

The model observes; deterministic code tiers and escalates. Never ask the model to be the whole system — a hallucinated tier is unbounded, a hallucinated observation is caught by the closed schema and the safety layer. This is why the study showed 0 diagnosis/dosing leaks. Hold this line as we add capability.

5. Fine-tune a foot-specific LoRA (where domain accuracy is won)

MedGemma's SigLIP encoder was pre-trained on dermatology (skin) — a head start, but diabetic foot wounds are out-of-distribution (FL-STUDY-001's whole premise). The durable fix is a foot-specific LoRA on MedGemma, trained on the clinician-labeled golden set + the opt-in passport-donation flywheel. Cook the 27B for the ceiling; distill/LoRA a 4B for the edge box. (Skills: medgemma-finetune, harvest, canary-then-cook.)

6. Hybrid: MedSigLIP classifier for the closed fields

For the categorical observations (wound present? redness severity?), a fine-tuned MedSigLIP classifier is more measurable and more reliable than generative text — and gives calibrated probabilities the safety layer can threshold. Best system = SigLIP classifier for the closed tags + MedGemma for the human-readable narrative, cross-checked against each other.

7. Ensemble & self-consistency

Sample N times (or run base + specialized) and majority-vote the categorical fields; flag disagreement for human review instead of guessing. Turns the study's low-κ divergence into a signal ("the models disagree here — look closer") instead of a coin flip.

8. The flywheel (the compounding lever)

Every clinician correction in the passport becomes a new training label, locally. The system gets more accurate the more it's used — the harvest loop, not a one-time cook. (Skill: harvest.)

How we prove "better" (not vibes)

Golden set + regression harness. Accuracy = agreement with clinician labels, per model version, published as a LocalStudy re-run. Deterministic gates, never a model grading a model. The number has to move on labeled data, or the change didn't help.

Edge reality — the NAS is the appliance

• Cook tier: 27B BF16 on the rails rig (2× RTX PRO 6000) — training, golden-set eval, the

ceiling number.

• Serve/edge tier: a quantized 4B foot-LoRA on the NAS / edge box — the thing that

actually runs in someone's home, all local. The study's ~21 s/photo on the big rig sets the ceiling; the edge number is what we report for the appliance. (Skills: edge-sku-deploy, medgemma-ops.)

The honest sequence

1. Standardize capture (multi-view, calibrated).
2. Build the clinician-labeled golden set (ground truth) from the Tier-2 packets.
3. Add context: prior-image comparison, profile, few-shot exemplars.
4. Fine-tune the foot LoRA; add the SigLIP classifier hybrid.
5. Measure every step against the golden set; ship only what moves it.

Bigger model is step 4, not step 1. Accuracy is won on capture, ground truth, and domain fine-tuning — in that order.