Pattern recognition and melanoma

Acta Eruditorum

Abby Van Voorhees

Dr. Van Voorhees is the physician editor of Dermatology World. She interviews the author of a recent study each month.

Bookmark and Share

In this month’s Acta Eruditorum column, Physician Editor Abby S. Van Voorhees, MD, talks with Jean-Jacques Grob, MD, about his recent Journal of Investigative Dermatology article, “Evidence of a Limited Intra-Individual Diversity of Nevi: Intuitive Perception of Dominant Clusters Is a Crucial Step in the Analysis of Nevi by Dermatologists.”

Dr. Van Voorhees: Let’s begin by talking about what is known about how dermatologists detect melanoma. We teach students the ABCD algorithm; don’t we use it?

Dr. Grob: ABCD was a model invented to summarize and recognize nevi. A lot of nevi are ABC, at least, and sometimes D. Many melanoma are not ABCD, and a lot of seborrheic keratoses are ABC. So although it’s easy to remember, it has limitations. We demonstrated a few years ago that if you tried to teach melanoma recognition by ABCD, by just showing pictures, or by doing both to samples of the general population, the most efficient way to train people is just showing them pictures. When you just give the ABCDs people tend to see melanoma everywhere, and when you mix the two they do worse than with pictures alone.

Why? Recognition is a cognitive process; we don’t know exactly how it works, but we do not use an algorithm. We don’t say something is a cow because it has four legs and is black and white; we recognize a cow because we’ve seen cows. And we each establish a pattern of recognition. A good example is someone’s face. If I tried to describe myself very precisely to you, if we meet in one hour you will not recognize me. But if we meet for only 10 seconds, you will build a pattern of recognition and you will still recognize me in a year. And my face is closer to your face than a melanoma is to a nevus. We are all able to recognize anything provided we have some opportunity to see it beforehand and build up a recognition pattern. [pagebreak]

In fact, recognition is covering different processes. I was speaking about direct recognition — if I see something, I can recognize it if I see it again. If I’ve seen 100 melanomas it’s very likely I will recognize another. There’s also differential identification. That means we are able to detect something which is not expected somewhere. Let’s say I put a tortoise in your office. You’ll immediately see that this is not expected there. This is what we do on the skin of a patient. The analysis of a single lesion out of context is a very difficult task because, in the usual clinical situation when you see a patient, you see all of his skin and all of his pigmented lesions, which provides you with a lot of information; you intuitively tend to build up different clusters of lesions that tend to look the same and there might be one lesion that is different. This is the “ugly duckling.” You see a patient and he has two or three patterns of nevi — and one nevus that is quite different. How it is different in this given individual is different than it would be in other patients. Out of context it might not be suspicious, but it is in this patient. This is something humans can do: see that something is different. A given nevus that is an “ugly duckling” on my skin might not be on yours, depending on the clusters we each have.

Then you have the third process involved in recognition of a potential melanoma being based on chronologic criteria. This is often considered as D, changing diameter, or E, evolution, in the ABCDE algorithm. But you need a reference to know this, either a photo or the memory of the patient. A melanoma is usually growing faster than other pigmented lesions around it.

To summarize: If a lesion is fitting with one of the different patterns of melanoma stored in your brain, and/or if it’s an ugly duckling, and/or the patient says it’s changing fast — you have to consider melanoma.

Another point is that the recognition pattern we each build is not the same. When you meet a young guy and you say he looks like his father, your wife might say he looks like his mother. You’re both right. But you’ve built your pattern of recognition of his face using points that are present in his father; she has built hers using points of recognition that are present in his mother. You don’t use the same patterns of recognition, but they both work — you’re both able to recognize him. When we see a melanoma we can both say it’s melanoma, but not for the same reasons. When I explain my reasons, they spoil your system. Provided your system works, we shouldn’t care how you recognize melanoma and how I recognize it. [pagebreak]

Dr. Van Voorhees: What was the objective of your study?

Dr. Grob: To see whether we were able to see the common patterns of all nevi in a given individual. I was pretty sure that each individual had his own nevus profile, but once you say that you have to demonstrate it. This idea is the basis for the ugly duckling — not like the others — but that presupposes that you have some common pattern for the others. That’s what we called perceived similarity clusters (PSCs).

Dr. Van Voorhees: Were perceived similarity clusters an effective tool to look at an individual’s nevi? How did dermatology experts compare with each other? Were PSCs similar amongst them? Was there a lot of group concordance? How did the novice group do? Was there much difference compared with dermatologists?

Dr. Grob: We let a number of people cluster all the nevi of each given patient, just on their perception of similarities, any way they wanted — two groups, three, 10. Each may cluster them differently. What we showed is that when you take the best experts and the people with the least knowledge possible — people who don’t even know what a nevus is — in fact those two groups clustered the nevi in a very similar way. Of course there is much more concordance between experts and other experts than between experts and naives, but even between the two groups there’s a large concordance. Pattern recognition is a characteristic of the human brain, not just of dermatologists. We showed that however many nevi there were, any human being can cluster them into a few groups (usually two or three) that reasonably cover the reality.

We didn’t test the expertise of experts; we just tried to understand how they were functioning. We found that they tended to do the same clustering. Among eight experts who saw the same pictures of patients only one clustered them in a different way. And that one person is still an expert — his brain just works differently. Just like in tennis you may find that most of the best are playing one way, but there may be another path to success. [pagebreak]

Dr. Van Voorhees: How did the novice group do? Was there much difference compared with dermatologists?

Dr. Grob: Not so bad. Not for the diagnosis of melanoma, of course; they weren’t tested for that. But for the clustering of nevi into groups they were not that different than the dermatologists. Even the naives could have picked out an ugly duckling; we’re doing more studies on that now.

Dr. Van Voorhees: Did dermoscopy assist dermatologists in grouping nevi?

Dr. Grob: Dermoscopy is taught in exactly the way it should not be taught, by description: spots, lines, networks. Words to describe images. We should teach dermoscopy by showing pictures and let each of our brains build up its own cognitive process at the dermoscopic scale. The way we teach it now is a less natural recognition process. We use visual recognition every day; dermoscopic scale is less intuitive. This may explain why the concordance in dermoscopy is lower than pure visual recognition. We’re not intuitive in dermoscopy; we’re using artifical algorithms that we have learned.

Dr. Van Voorhees: How can this enhanced understanding of the consensus clusters help us going forward? Can this information be used in computer models? Do you see technological advances that can come from this work?

Dr. Grob: We are working on technical models that could do this. This clustering idea isn’t used in any computer-aided system so far, but it could be. They don’t take into account the environment of the lesion; that’s why dermatologists are better at diagnosing. If you don’t give the computer the information you have it won’t do better than you. [pagebreak]

Dr. Van Voorhees: Do the results of this study tell us more about how we teach dermatology? Do we need to more concretely teach pattern recognition?

Dr. Grob: Yes, we should start young doctors in dermatology by telling them they need to train their brains by looking at a lot of pictures so they build up their own recognition process for many important situations. Show an eczema, tell the student is is an eczema, but do not explain why. Same thing for psoriasis and so on. His brain will make that connection straight from the images, and explanation on color, distribution, and size of lesions may disturb the building of his intuitive recognition process rather than help it. Written description used to be useful, when students did not have enough opportunities to see images, but it is so easy now to see hundreds of images to train one’s own recognition process.

What I’ve told you is broader than just what we found in our paper. Melanoma recognition is an important issue, but recognition of situations applies across both dermatology and other medical specialties, like radiology. This is why you have an instant diagnosis when you see a patient, though you may not be able to explain why  — and when you try to explain you start saying stupid things, because you do not know how your brain does the job. Every time you try to explain visual information by words you lose or distort information. I used to explain why something was psoriasis when teaching — now I just show students 100 cases of psoriasis and 100 eczemas and they become able to recognize each one. When you try to explain you spoil the information — you’d say psoriasis is well-limited and eczema is not. That’s more or less true, but there are exceptions. Once you’ve seen them both, you don’t need to go back to criteria that often fail.

Jean-Jacques Grob, MD, is professor of dermatology at Hopital Ste Marguerite in Marseilles, France. His article was published in the Journal of Investigative Dermatology. J Invest Dermatol 133: 2355-2361; advance online publication, May 16, 2013; doi:10.1038/jid.2013.183.