Can a Photograph Diagnose Skin Cancer? Where the Evidence Actually Stands
Can a Photo Tell If a Mole Is Skin Cancer? What We Actually Know
Algorithms can classify images of skin lesions with impressive accuracy in test conditions, and several studies have reported performance comparable to specialists on curated image sets. That is a real result, and it is a narrower one than the headlines suggest. The distance between classifying a photograph and diagnosing a patient is where all the difficulty lives, and it is worth understanding before trusting an app with a mole.
Computer programs are now very good at looking at a photo of a mole and saying what it probably is. In tests, some have done as well as skin doctors. That result is real. It is also much smaller than the headlines make it sound. Sorting photos is not the same job as working out what is wrong with a person. That gap is worth understanding before you trust an app with a mole on your own skin.
Can a Photograph Diagnose Skin Cancer? Where the Evidence Actually Stands: the essentials
The short version
High classification accuracy on curated image datasets under test conditions
They get a lot of them right, when the photos are clear and chosen in advance.
Evidence that an app in a pocket performs the same way on real skin
Proof that a phone app does the same job on real skin, in real life.
Training sets have historically under-represented darker skin
The photo collections used to teach these programs have always held too few people with darker skin.
A photograph carries no history, no change over time and no examination
A photo cannot say how long the spot has been there, whether it has changed, or how it feels to touch.
Triage support and workflow, alongside a clinician rather than instead of one
Helping decide who needs to be seen first, working alongside a doctor rather than in place of one.
A lesion that is changing, bleeding or unusual is examined in person
If a spot is changing, bleeding, or just looks odd, a doctor should look at it in person.
What the studies actually show
Since 2017, a series of studies has reported that convolutional neural networks can classify dermatoscopic images of skin lesions at a level comparable to, and in some tests exceeding, that of dermatologists presented with the same images. Those results have been repeated often enough to be taken seriously.
Since 2017, study after study has tested this. The programs were shown close-up photos of moles and other marks on the skin. Skin doctors were shown the same photos. The programs did about as well as the doctors, and in some tests better. That has now been repeated often enough to take seriously.
The important qualifier is in the phrase presented with the same images. The comparison is between an algorithm and a clinician who has been deprived of everything a clinician normally uses: the history, the rest of the skin, the ability to look from another angle, to palpate, to ask when it changed, and to compare it with the person’s other lesions. Under those conditions the comparison is fair to the algorithm and not to the clinician.
The catch is in those last five words: shown the same photos. A skin doctor in real life never works from a photo alone. They ask how long it has been there. They look at the rest of your skin. They move their head to see it from another angle. They press on it with a finger. They compare it with your other moles. In these tests the doctor had none of that. So the test was fair to the program, and unfair to the doctor.
It is also a comparison on curated data. Study images are usually well lit, in focus, taken with a dermatoscope, and drawn from archives where the diagnosis was confirmed. A photograph taken on a phone in a bathroom is a different input, and performance on the first does not transfer automatically to the second.
The photos were also unusually good ones. They were well lit and in focus, taken with the special magnifying lens doctors use, and pulled from records where the true answer was already known. A photo you take on your phone in the bathroom is nothing like that. Doing well on the first kind of photo does not mean doing well on the second.
The problem in the training data
The problem with the photos it learned from
A model learns the distribution it is trained on. Publicly available dermatology image sets have historically contained far fewer images of darker skin, and reviews have repeatedly noted that skin tone is often not even recorded.
A program only ever learns from the photos it is shown. The photo collections used to teach them have always held far fewer pictures of darker skin. Reviews of those collections keep finding that skin tone was often not even written down.
The consequence is predictable and serious. A system that performs well on the skin types it saw in training may perform considerably worse on skin types it did not, and the group most affected is one for which melanoma is already diagnosed later and carries worse outcomes. A tool that works less well precisely where the need is greatest is not a neutral tool.
What happens next is easy to predict, and it is serious. A program that does well on the skin tones it learned from can do much worse on the ones it did not see. And darker skin is already where melanoma, the most dangerous kind of skin cancer, tends to be caught late, when it is harder to treat. A tool that works worst for the people who need it most is not a fair tool.
This is being worked on, and it is not a reason to dismiss the technology. It is a reason to ask, of any specific product, what it was trained on and in whom it has been validated, and to be sceptical of a general claim of accuracy that does not answer.
People are working on this, and it is not a reason to write the whole idea off. It is a reason to ask hard questions about any one product. Which photos was it taught from? Who was it tested on? If a company advertises high accuracy but will not answer those two questions, be careful with the claim.
What a photograph cannot carry
What a photo cannot tell you
Diagnosis in dermatology is substantially about change and context. Is this new? Has it altered in the last three months? Does it look different from the person’s other moles? Has it bled, itched or failed to heal? Is there a family history? None of that is in the pixels.
Working out what a spot is has a lot to do with change, and with the rest of you. Is it new? Has it changed in the last three months? Does it look different from your other moles? Has it bled, itched, or refused to heal? Has anyone in your family had skin cancer? None of that is in the photo.
The ugly duckling sign, the observation that a lesion which looks unlike the person’s others deserves attention, requires seeing the others. A single photograph cannot apply it. Nor can a photograph feel that a lesion is firm, or notice the second one nearby that the person had not mentioned.
Doctors use a simple idea called the ugly duckling. Most of your moles look like brothers and sisters of each other. The one that looks like the odd one out is the one worth checking. To use that idea, you have to see all the others. A photo of one mole cannot. A photo also cannot feel that a lump is firm, or notice the second spot nearby that you never thought to mention.
There is also the asymmetry of errors. A false reassurance from an app that delays a melanoma diagnosis costs more than a false alarm that leads to an unnecessary appointment, and the two are not equally weighted in a user’s mind. People act on reassurance more readily than on alarm.
The two kinds of mistake are not equal either. If an app wrongly says a spot is fine, and that delays finding a melanoma, the cost can be very high. If it wrongly says a spot looks worrying, the cost is one appointment you did not need. People also treat the two differently. Being told you are fine is much easier to believe, and much easier to act on, than being told to worry.
Where it genuinely helps
The realistic near-term value is as support rather than substitution: helping to prioritise which patients are seen first, supporting a clinician’s read, flagging lesions for review in total-body imaging, and extending reach where specialist access is limited.
For now, the sensible use is help, not replacement. It can help decide who gets seen first. It can act as a second pair of eyes for a doctor. It can flag spots worth a look when someone’s whole body has been photographed to track changes. And it can reach people who live nowhere near a skin specialist.
Used that way, with a clinician in the loop and clear accountability for the decision, it is a promising tool. Used as a substitute for examination, it is a way of converting a clinical question into a photograph and then answering the photograph.
Used that way, with a real person making the decision and answerable for it, it is a promising tool. Used instead of being examined, it does something strange: it turns a question about you into a question about a photo, and then answers the photo.
The practical guidance is unchanged by any of this. If something on your skin is new, changing, bleeding, not healing, or simply unlike your other spots, it should be examined by a clinician. An app may be a useful prompt to make that appointment. It is not a reason not to.
None of this changes what you should do. If something on your skin is new, changing, bleeding, not healing, or simply unlike your other spots, get a clinician to look at it. An app can be a useful push to book that appointment. It is never a good reason to skip it.
Questions patients ask
Are skin cancer apps accurate?
Accuracy varies widely between products and between the conditions of testing and real use. Published performance on curated images does not reliably predict performance on a phone photograph, and independent evaluation of consumer apps has often been less favourable than headline claims.
It varies a lot. Products are not all equally good, and every one of them does better in a test than in real life. Doing well on clean, hand-picked photos does not tell you how a program will handle a photo from your phone. When consumer apps have been tested by people who did not build them, the results have often been worse than the advertising.
Can I use an app instead of seeing someone?
No. An app cannot take a history, examine the rest of your skin, feel a lesion or compare it with your others. Use it as a prompt to seek an opinion, never as a reason to delay one.
No. An app cannot ask you questions, look at the rest of your skin, feel a lump, or compare one spot against your others. Use it as a reason to go and get an opinion. Never use it as a reason to put one off.
Does it work equally well on all skin tones?
Does it work as well on all skin tones?
Not necessarily. Training datasets have historically under-represented darker skin, and performance can be lower where the training data was thin. Ask what a specific product was validated on.
Not always. The photo collections used to teach these programs have held too little darker skin, and a program is usually weaker wherever it saw less. If you are looking at a particular product, ask who it was tested on.
What should make me book an appointment regardless?
What should make me book an appointment anyway?
Anything new, changing, growing, bleeding, itching persistently, failing to heal, or looking different from your other spots. Those features are the reason to be examined, and no image analysis changes that.
Anything new. Anything changing or growing. Anything that bleeds, or itches and will not settle, or will not heal. And anything that looks different from your other spots. Those are the reasons to get looked at, and no photo app changes that.
Will this technology replace dermatologists?
Will this replace dermatologists?
There is no current evidence supporting that, and the more plausible direction is support for clinical work: triage, second reads and monitoring. Anyone stating confidently what will happen is speculating.
There is no evidence for that today. The more likely path is that it helps with the work: sorting out who is seen first, offering a second opinion, and keeping track of changes over time. Anyone who tells you confidently what will happen is guessing.
Talk it through with a physician first.
Every plan starts with an assessment of your skin, your anatomy and what you actually want to change.
This page is general health information, not medical advice. It cannot tell you whether a treatment is right for you. Approved uses, risks and results differ from person to person, and product labelling changes. Ask us at Dermatology and Laser Centre about your own situation, or make an appointment, and seek urgent care for any serious symptom.
