AI Public Health NHS Diagnostics Research

Your Phone Might Catch Oral Cancer Before Your Dentist Does

Dr Ali Vatan Ali Vatan
·

Deep learning models using smartphone cameras are detecting oral cancer with 94% accuracy in under five seconds. For the NHS, this could be transformative.

Your Phone Might Catch Oral Cancer Before Your Dentist Does

Researchers have built a deep learning model that screens for oral cancer using ordinary smartphone camera images. The system, built on an EfficientNetB0 convolutional neural network, achieved 94.29% accuracy, 95.45% precision, and an AUC of 0.99. Average inference time: under five seconds. You photograph a suspicious oral lesion with your phone, and the AI tells you whether it’s likely malignant.

Of all the dental AI applications I’ve written about, this one matters most.

Early detection changes everything

Oral cancer is a condition where timing determines survival. When caught early at a localised stage, the five-year survival rate sits around 80-85%. When detected late, after spread to lymph nodes or distant sites, that drops to roughly 20-39% depending on the site and extent.

The tragedy is that oral cancer is often diagnosed late. Many patients don’t present until they have pain, difficulty swallowing, or a visible mass, by which point the disease has frequently progressed. Some haven’t seen a dentist in years. Others have been seen regularly but the lesion was missed, attributed to trauma, or monitored when it should have been referred.

Every dentist performs soft tissue examinations at every check-up. But we’re human, and screening in general practice is limited by the clinician’s experience, the quality of the examination, and the sheer rarity of the condition relative to the thousands of benign mucosal variations we see daily.

Why this application is different

I’ve been critical of dental AI where I think it’s solving problems that didn’t need solving. This isn’t one of those cases.

The technology is accessible. You don’t need a CBCT scanner, a specialist imaging suite, or expensive proprietary hardware. You need a smartphone, something virtually everyone already owns. The AI model is lightweight enough to run on a web-based interface, meaning it could be deployed anywhere: a dental practice, a pharmacy, a community health centre, a mobile screening unit.

For the NHS, this matters enormously. Access to dental care in this country is in crisis. Millions of people can’t get an NHS dentist. Oral health inequalities are widening. The communities most at risk of oral cancer (higher rates of tobacco use, alcohol consumption, and deprivation) are precisely the communities least likely to have regular dental check-ups.

A smartphone screening tool doesn’t replace the dentist. But it could act as a triage system, identifying high-risk lesions in settings where dental professionals aren’t available. Community pharmacists, health visitors, practice nurses: any healthcare worker with a smartphone could potentially screen for oral cancer and flag suspicious lesions for urgent referral.

What the evidence shows

The medRxiv preprint describes a model trained on 1,071 smartphone photographs, classified into normal or malignant categories. The results are striking:

  • Accuracy: 94.29%
  • Recall: 93.33%
  • F1-score: 94.38%
  • AUC: 0.99

Broader meta-analyses of AI in oral cancer detection, using various imaging modalities, have reported pooled sensitivity of 0.87 and specificity of 0.81. The smartphone-specific model performs at the upper end of this range, which is impressive given the simplicity of the imaging hardware.

The caveats are real. This is a preprint, not a peer-reviewed publication. The training dataset is relatively small. Real-world performance in diverse populations will likely differ from controlled research conditions. And no screening tool is useful unless it’s embedded in a clinical pathway that can act on positive results; there’s no point detecting a suspicious lesion if the patient can’t access a specialist for a biopsy.

What this could look like in practice

Imagine a national oral cancer screening programme, something the UK has never had, powered by smartphone AI. Pop-up screening events in communities with the highest rates of oral cancer, staffed by trained healthcare workers with nothing more than a phone and an internet connection.

This is AI extending the reach of the healthcare system into communities it currently fails. Not diagnosing diseases that good dentists already catch, but catching disease in people who never see a dentist at all.

Oral cancer kills roughly 3,000 people a year in the UK. Many of those deaths are preventable with earlier detection. If a smartphone and a five-second AI analysis can shift even a fraction of late-stage diagnoses to early-stage ones, we’re talking about lives saved.

That’s the kind of AI application I can get behind without reservation.

References