On average, doctors who use systems that apply artificial intelligence (AI) to decide whether images of skin abnormalities show signs of cancer are able to make better diagnoses. According to a study by Viennese researchers, doctors with limited experience benefit most from working with AI. However, the “Nature Medicine” journal reports that AI can sometimes also mislead physicians.

Nowadays, systems that more or less independently learn to classify image content on the basis of a large number of images are used in many fields. The application of this type of AI in the medical field has also been under discussion for several years. One of the fields in which this is envisaged is diagnostics based on imaging procedures.

An international team led by the University Clinic for Dermatology and the Institute for Medical Information Management at the Medical University (MedUni) of Vienna in Austria, recently investigated how well this approach works in cases of suspected skin cancer.

To this end, the researchers submitted images showing either benign or malignant skin lesions to 302 specialists from 41 countries. The participants then evaluated what they saw, both with and without the support of AI. The AI generated probabilities of all possible diagnoses that could theoretically be made on the basis of the image. In addition, the system selected images with confirmed diagnoses that were similar to the test images.

In total, this resulted in more than 13,000 decisions. It emerged that the proportion of correct diagnoses increased from 63.6 to 77 percent when the AI showed the physicians the different probabilities for all diagnoses.

AI is of little use to experienced physicians

“Interestingly, less experienced doctors benefited more from the AI support than their more experienced colleagues. Less experienced physicians were also more likely to trust the AI than more experienced ones. The latter only considered the AI’s suggestions to revise their original diagnosis in cases where they themselves were uncertain,” the study authors report.

However, in another part of their study, the researchers showed that the use of AI could also produce undesirable results. Here, they deliberately caused the AI to make incorrect diagnosis recommendations. It turned out that none of the groups of participants were immune to being led astray. This was true even for renowned experts in the field.

The resident physician of the future?

The researchers expect their findings on skin cancer diagnosis to be transferable to other medical fields. In view of the results of their research, they argue that such AI systems should always be tested in cooperation with the users.

For the authors of the study, humans and AI will in future “complement each other and together improve patient care. Not competition, but cooperation between humans and AI should be the focus of attention,” they conclude.