AI and the Early Detection of Skin Cancer

Skin cancer is the ideal problem for machine learning, and that is not a compliment to the technology so much as a description of the disease. Diagnosis begins with a visual pattern. The pattern is photographable. There are enormous archives of labelled images. And the cost of missing a melanoma early is measured in years of life, while the cost of an unnecessary referral is measured in an outpatient appointment. Few clinical problems line up so neatly.

The results have followed. Studies comparing convolutional neural networks against panels of dermatologists on curated image sets have repeatedly found the algorithms performing at or above the human average on classification of melanoma versus benign lesions. Regulators in Europe have approved several image-based triage tools, and a number of health systems now run them inside urgent skin cancer pathways, where a photograph taken in a community clinic is assessed before a consultant ever sees the patient.

That last point is the real story. The value is not that a machine can label a mole. It is that the bottleneck in skin cancer care has never been diagnostic skill; it has been access to diagnostic skill. In Britain, urgent suspected-cancer skin referrals have grown far faster than the dermatology workforce. The overwhelming majority of those referrals turn out to be benign. A triage system that can safely and confidently identify the clearly harmless lesions frees consultant time for the ones that matter, and shortens the interval between a worried patient noticing something and a specialist looking at it.

There are serious caveats, and the honest practitioners state them loudly.

The first is training data. Dermatology image archives have historically over-represented lighter skin. An algorithm trained mostly on white European skin performs measurably worse on darker skin, and since melanoma on darker skin more often presents in unusual sites — palms, soles, nail beds — and is diagnosed later, an unequal tool risks widening an inequality that already costs lives. Several research groups are now building deliberately diverse image sets, but the gap is not closed.

The second is what the model is actually looking at. There is a well-documented failure mode in which a network learns that a lesion photographed next to a surgical ruler is more likely to be malignant, because clinicians reach for a ruler when they are worried. The model becomes an excellent detector of clinical anxiety and a poor detector of cancer. Interpretability work exists to catch this, but it requires people who are looking for it.

The third is the smartphone problem. Consumer applications that promise a melanoma risk score from a photograph have been the subject of repeated regulatory concern, and one review found that several widely available apps missed a substantial proportion of malignant lesions. The danger is not the false alarm; it is the false reassurance that keeps someone away from a doctor for eight months.

The fourth is calibration under real conditions. A model validated on clean dermoscopic images taken by trained staff will behave differently on a photograph taken on a ward with poor lighting and a smeared lens. Prospective clinical trials, rather than retrospective image-set comparisons, are the only honest test, and there are still far too few of them.

What good deployment looks like is fairly clear by now. The algorithm sits inside a clinical pathway rather than replacing one. It triages rather than diagnoses. A clinician retains the decision and can see why the system flagged what it flagged. Performance is monitored continuously and broken down by skin type, age and site, so that degradation is caught rather than assumed away. And the patient is told plainly that software was involved.

Used that way, this is one of the most obviously beneficial applications of machine learning in medicine. Melanoma caught at stage one has a five-year survival above ninety-five per cent. Caught at stage four it collapses. Anything that reliably compresses the time between noticing and treating is worth a great deal, and unlike much of what is currently sold as artificial intelligence, this technology is being measured against an outcome that cannot be fudged.

The remaining work is unglamorous: better data from more people, prospective trials, and clear rules about liability when the machine is wrong. None of that is a research problem. It is an organisational one, which is usually the harder kind.