Getting Usable Patient Photos for Remote Assessment

The photo arrives at 9:14am. A patient has messaged the practice about a broken upper front tooth, worried, asking whether she needs to come in today. You open the attachment and what you get is a dark, slightly blurred shot of her chin and lower lip, taken at arm’s length, with the flash blowing out a white patch somewhere around the lower left quadrant. The tooth in question is not in frame.

You cannot assess that. Nobody could. And the clinical question itself was not hard: is this an enamel chip that can wait a fortnight, or an exposed dentine fracture that needs seeing before the weekend? That is a thirty-second decision given a usable image. What stopped you was not clinical difficulty. It was a photograph.

This is the pattern across remote assessment in UK general practice, and it is worth saying plainly because it changes where you spend your effort. Most failed remote assessments fail on image quality. They do not fail because teledentistry cannot answer the question, or because AI triage tools are not good enough yet, or because patients are bad at describing symptoms. They fail because the picture does not show the thing.

What the retake data actually looks like

We pulled anonymised figures from a group of six mixed NHS/private practices in the Midlands and the South West running photo-led triage through a mix of Dentally’s patient messaging, standalone WhatsApp Business numbers, and in two cases a Toothfairy-style patient app. Over about fourteen weeks, 1,180 patient-submitted photo submissions for urgent or semi-urgent concerns.

Of those, 487 (41%) required at least one retake request before a clinician would commit to a triage decision. That is not a small inefficiency. Each retake costs a front-desk message out, a wait of between twenty minutes and two days for the patient to respond, and a second clinical review. One practice manager timed it: an average of 11 minutes of staff time per retake cycle, mostly in chasing.

Here is the breakdown of why those 487 were unusable. One primary reason each, coded by the reviewing clinician:

Read the last row again. In 97% of failed submissions, the problem was not that remote assessment was the wrong tool for the question. The problem was framing, focus, light and obstruction. Four fixable things.

Why your AI triage tool makes this worse before it makes it better

Practices buying into AI-assisted triage tend to assume the software absorbs variability in input quality. Some of it does, within limits. What actually happens with poor input is more awkward than a straightforward failure, and it is worth understanding before you sign anything.

Radiograph AI such as Pearl’s Second Opinion or VideaHealth works on a standardised input: a bitewing or periapical taken on your sensor, in your surgery, at roughly consistent geometry and exposure. That is why the published performance figures hold up. Patient-submitted intraoral photography has none of that standardisation. A phone camera at 8cm with mixed daylight and bathroom LED is a fundamentally different imaging problem.

Tools that do accept patient photos (Toothpic’s remote assessment flow, Dental Monitoring’s ScanBox-based system, the photo-upload paths in several NHS 111 dental pathway pilots) handle this in one of two ways. Either they gate on quality, rejecting the image at upload and asking for another, or they produce a low-confidence output that a clinician then has to decide whether to trust. The first is irritating but safe. The second is the dangerous one, because a confidence score of 0.4 on a possible caries flag in a badly lit photo is not information you can act on, and it creates a record that looks like an assessment.

Dental Monitoring’s answer is instructive: they shipped a physical plastic cheek retractor with a phone cradle, the ScanBox, because software alone could not get consistent images from patients. That is a vendor with substantial engineering resources concluding the problem is at the capture end. Note what that tells you about where to aim your own effort.

The capture instruction that fixes most of it

Here is the thing that genuinely surprised the practices in that dataset. They did not buy anything. They changed the text of the message that asks for the photo.

The before version, which is roughly what everyone sends:

Hi [name], thanks for getting in touch. Could you please send
a photo of the area you're concerned about so the dentist can
take a look? Many thanks, [Practice]

The after version, built specifically against the six failure modes in the table above:

Hi [name], to get this looked at today we need 3 photos.
Please follow these exactly:

1. STAND BY A WINDOW in daylight. Turn your flash OFF.
2. Pull your lip right back with a clean finger so you can
   see the gum above the tooth.
3. Photo 1: the tooth close up, filling most of the frame.
   Tap the screen ON the tooth before you take it so the
   camera focuses there.
4. Photo 2: same area, but step back so we can see the
   teeth either side.
5. Photo 3: the same spot in a mirror, whole smile.

Hold your breath while you press the button. If any photo
looks blurry on your screen, delete it and take it again.
Reply with all 3 and tell us which tooth hurts (e.g. "upper
right, 3 back from the front").

[Practice]

Across the four practices that adopted a version of this, retake rate dropped from a baseline of 41% to 16% over the following eight weeks. Not zero. Sixteen percent. The remaining failures were concentrated in posterior teeth, which is predictable, and in patients over about 75, which is also predictable and worth planning for separately.

Break down why each line earns its place, because the temptation will be to trim it for friendliness and that is how you lose the effect.

“Turn your flash off” and “stand by a window” together address 33% of failures (darkness plus blowout). Phone flash at intraoral distance is a point source 6cm from a wet, highly reflective surface. It produces a specular hotspot almost exactly where you want to see detail. North-facing window light at 11am is better than any flash you have on a phone.

“Tap the screen on the tooth” is the single highest-yield sentence. Phone autofocus defaults to the largest high-contrast region in frame, which at intraoral distance is usually the lip margin or the edge of the frame, not the tooth. Patients do not know tap-to-focus exists for this purpose. Telling them converts a meaningful chunk of that 24% blur category.

“Pull your lip right back with a clean finger” addresses obstruction and also forces the patient into a position where the tooth is actually presented to the lens rather than at a glancing angle.

Three photos rather than one is doing quality assurance by redundancy. If the close-up fails, the mid-range shot frequently still answers the question. If both fail, the mirror shot at least confirms which tooth they mean, which is what the “wrong area in frame” category is really about: you can usually tell a patient has photographed the wrong quadrant, but not which one they intended.

A worked example

Patient messages Monday 8:50am: “tooth broke last night, upper front, can you fit me in.” Sends one photo, taken in a car, flash on, showing most of a face.

Under the old flow: reception asks for a better photo at 9:20, patient replies 2:40pm with a second photo, slightly better but still dark, clinician reviews 5:10pm, asks for a third, patient sends Tuesday morning, decision made Tuesday 11am. Roughly 26 hours, three clinical reviews, four front-desk touches, one anxious patient.

Under the structured request, sent as an auto-reply triggered by any inbound message containing “broke”, “chipped”, “knocked” or “pain”: patient receives the three-photo instruction at 8:51am, sends all three by 9:05. Clinician reviews at 9:30 and sees a mesio-incisal enamel fracture on the UR1 with no visible dentine exposure and intact adjacent teeth. Books into Thursday’s routine composite slot rather than displacing the day’s emergency appointment. One review, one touch, 40 minutes.

The second version did not use better software. It asked better.

Setting the quality bar you will actually enforce

Decide what “usable” means for your practice, write it down, and make it the same standard the front desk applies before escalating to a clinician. Vague standards get applied inconsistently and your retake rate stays high.

A workable minimum for anterior assessment:

• The tooth of concern and at least one neighbour fully in frame
• Incisal edge and gingival margin both visible
• No specular hotspot over the area of concern
• Sharp enough that you can see the enamel surface texture, not just tooth outline
• Taken within the last 24 hours (patients send old photos surprisingly often)

For posterior teeth, be honest: a patient-submitted photo of a lower second molar is rarely going to be diagnostic, and treating remote photography as capable there will burn your credibility with the clinical team. Use photos posteriorly to confirm gross findings (visible cavitation, a fractured cusp, obvious swelling) and route everything else to an appointment. Our pillar on triage and remote assessment goes into how to set those routing rules by tooth position and presenting complaint, which is where the real efficiency sits.

One more practical note on devices. Do not build your instructions around a specific phone model or camera app, because you will have patients on a four-year-old Android and patients on a current iPhone and the instruction has to work for both. Tap-to-focus, flash-off and window-light all work everywhere. Portrait mode, macro mode and night mode do not, and night mode in particular will produce a composite image with motion artefacts that look like surface cracks. If your instruction tells people to use a mode, you have introduced a new failure category.

Where to put your money instead

If you have budget earmarked for improving remote assessment, the ordering that comes out of this data is uncomfortable for vendors and good for your P&L.

First, rewrite your photo request message. Cost: an hour. Expected effect: retake rate down by roughly half.

Second, automate the sending of it so it fires on keyword without waiting for someone at the desk to be free. Most practice management systems with a messaging module can do this, and Dentally, SOE Exact and Carestream all have some version of templated auto-reply available. Cost: a morning of configuration.

Third, give your front desk a one-page pass/fail checklist so unusable photos get bounced in minutes rather than sitting in a clinician’s queue until lunchtime. Cost: printing.

Only then look at a tool that does automated quality gating at upload or AI-assisted photo triage. Those tools are genuinely useful, and they become substantially more useful sitting on top of a population of patients who have already been told how to hold a phone. Buy them fourth, not first, and when you evaluate them ask the vendor directly what proportion of real patient submissions their quality gate rejects. If they do not know, they are not measuring the thing that determines whether their product works in your practice.

The 41% figure is the one to take into your next supplier meeting. Ask what their tool does with those submissions, and listen for whether the answer is “rejects them clearly” or something softer.