Clinical photography classification

A computer vision model on a wound, skin, or lesion photograph exists to support acuity triage. It helps a clinic decide which images need a human look now, which can wait, and which do not justify a specialist referral on the photo alone. The matching output is a quality flag that names the image ID and the visual artifact that triggered it. If the image is ordinary, the field stays empty. The model does not invent a diagnosis. A clinician still triages.

Staff will try to read a disease name out of any non-empty cell. If the cell holds motion blur on image 4821, they can check it. If it holds a lesion class, they will copy it into the note.

Clinics already sit this kind of vision work next to the systems that store encounters and photos. Vendors in this class include Skin Analytics, Epic, Microsoft, and Oracle Health. Treat them as the stack that holds the image, runs or hosts the model, and shows the flag where a nurse or dermatologist already works. A product name is not a clinical claim.

The adjacent language job is different. Symptom-to-differential generation starts from a history a person took. This page starts from pixels. Keep those jobs apart so a poorly lit heel photograph does not become a named disease in the chart.

The first operational step is load, not interpret. The photo enters the queue with a stable image ID, the encounter or wound episode it belongs to, and enough context to know which body site was intended. Until that ID exists, nothing downstream should fire. A model that classifies an unindexed file cannot produce a flag a clinician can cite in a huddle.

Loading is also where you refuse to complete the row with fiction. A flash-blown close-up, a gloved finger over the margin, a photo of the dressing instead of the wound, or a frame taken from across the room is still a photograph. It is not yet a classifiable clinical image. The model may still run. What it must not do is fill the quality field with a guessed lesion type so the worklist looks finished.

If your workflow already pushes imaging into a worklist the way radiology AI triage and flagging does for scans, reuse that pattern. The photo is an item with an ID, a timestamp, and a place the flag will land. Do not fork a second inbox that only the model can see.

When the model has something to say, the flag cites two things: which image, and what on that image. Image 4821, motion blur across the wound bed. Image 1194, specular highlight obscuring the inferior margin. Image 3308, cropped so the periwound skin is missing. Those are quality statements a clinician can verify by opening the file. If they cannot point at the artifact in a few seconds, the flag is wrong, even if the tile is red.

When the photograph is ordinary (in focus, site visible, no dressing in the way, lighting usable), the quality field stays empty. Empty is a result. It means the pixels did not present an artifact the model is allowed to name. It does not mean no disease, refer, or discharge. Ordinary empty is how you avoid a dashboard of red tiles that teach staff to ignore the channel.

Do not let the pipeline invent a lesion to justify a non-empty flag. Possible melanoma is not an artifact. If the model cannot point at blur, occlusion, crop, exposure, a color cast that hides the bed, or another named artifact, it writes nothing.

The same discipline shows up in pathology image analysis: the useful machine output points at a region or a quality problem on a specific slide, and a pathologist still reads. Cite the frame. Do not write the disease.

A wound clinic lead opens the afternoon list. One item is a lower-leg photo from a community nurse, stored as image ID 4821. The model returns a flag: 4821, motion blur across the wound bed, inferior edge not resolved. The acuity question is whether this photo can support a decision about referral, dressing change, or watchful waiting. It cannot, because the artifact is in the way. The lead does not treat the flag as venous ulcer, arterial ulcer, or infected. They request a retake, or they examine the patient. The empty rows next to it stay empty. Those photos are ordinary enough to look at. Looking is still a person.

That is the whole example. There is no conversion story and no improvement percentage. The shape of the output is the lesson: ID plus artifact, or silence, then a clinician.

Three failure modes show up in the same UI if you optimize for a busy-looking worklist.

A red tile with no artifact is noise wearing a severity color. If the row is flagged but the cite is missing (no image ID, or an ID with no visual claim), staff cannot check the work. They either open every photo anyway, which wastes the queue, or they trust the color, which is how a quality channel becomes a fake diagnosis channel. Suppress the tile until the flag can name the image and the artifact. Color without a cite is a defect in the pipeline, not a clinical finding.

Treating the flag as a diagnosis is the second failure. Blur is not cellulitis. Occluded margin is not melanoma. If the note or the referral letter copies the flag into an impression field, you have moved a quality message into a disease name. Keep the flag in a quality slot. The impression stays a clinician's sentence, or it stays blank until they write one.

Inventing a lesion is the third. A model that cannot see the inferior edge sometimes still emits a lesion class because every training photo had a label. That output is fiction. Drop it. The allowed non-empty result is the artifact. A differential starts from symptoms and exam, not from a guessed class on a bad photo.

After the flag or the empty field is on the row, a clinician triages. They open the photo, accept the artifact claim or reject it, decide whether the image is usable, and only then choose acuity: same-day review, routine follow-up, retake, or referral. The model does not close that loop.

Referral is a later step. When the photo is usable and the clinician decides the patient needs a specialist, smart referral matching is the routing problem: who, how soon, with what packet. Classification of the photograph does not pick the consultant. It only tells you whether the packet's image is fit to send. Sending a blurred frame with a confident lesion label is worse than sending nothing, because the receiving clinic will triage the label.

If the clinician disagrees with a cite, log it against the image ID. That disagreement is operational data for the next model review, not a prompt to thicken the flag with a disease name.

Load the photo. Flag with a cite, or leave ordinary empty. Do not invent a lesion. A person still triages.