The algorithm and the dermatologist equally assess the thickness of melanoma

When more than 400 international dermatologists had to evaluate just under 1,500 dermatoscopy images of melanomas, they succeeded in correctly evaluating whether or not melanomas were invasive (so-called in situ melanomas) in 76 percent of cases. For the machine learning algorithm, the corresponding percentage was 74 percent.

This is a step toward finding better and more sophisticated tools for doctors to assess the thickness of melanoma, according to one of the researchers the studySam Polisi, lecturer in dermatology and venereology at the Sahlgrenska Academy at the University of Gothenburg, and also a dermatologist.

Although it is relatively easy for dermatologists to detect melanoma, it is much more difficult to assess the thickness. When clinicians in this study were asked to quantify the thickness further in three groups (melanoma in situ and invasive melanoma under 1.0 mm and above), only 56 percent of the assessments were correct.

The doctors’ experience or training had no bearing on how well they performed in their ratings, which Sam Polisi sees as a sign of how difficult it is. It is still important to estimate how quickly patients can be prioritized for the first process, and thus Sam Polesie believes that machine learning algorithms can serve as a valuable complement in future clinical work.

This is an upcoming development and dermatologists are generally positive. The big challenge is getting these algorithms to work in clinical everyday life, so they are not seen as competition but rather as decision support, he says and continues:

The controversy surrounding the introduction of machine learning algorithms into dermatology often revolves around where we should put them in the clinical flow: before or after a dermatologist’s initial evaluation? If the algorithm is put in place first, we could potentially increase sensitivity and have a time-saving effect, but then the clinical flow changes as well. If it is then placed then we would probably increase the specificity of our ratings, but there would be no time gain and even more time consumption as it could lead to conflicting ratings to be dealt with.

A limitation of the study is that the images that doctors interpreted were all melanoma, while dermatologists have to consider many other differential diagnoses in a clinical setting. Another is that the algorithm was compared with the collective evaluation of dermatologists.

Our next step is to evaluate the algorithms in future studies and then compare the performance of the algorithms in comparison to the evaluations of individual dermatologists, says Sam Polesie and continues:

Only then can we begin to comment on how these algorithms fit into healthcare. I hope such tools will be available in Swedish skin care in 5-10 years.

The study was published in the journal of the European Academy of Dermatology and Venereology, JEADV, and was carried out in collaboration with researchers at the Medical University of Vienna in Austria.

Lakartidningen.se 2022-08-17