Using Artificial Intelligence, Machine Learning to Aid in Skin Cancer Diagnoses

Skin cancers are the most common cancer, and most individuals diagnosed with skin cancer first present to primary care. There, primary care providers face the difficult task of differentiating common benign skin lesions from rarer skin cancers.

“There has been a lot of hype recently about the potential of artificial intelligence and machine learning (AI/ML) algorithms to help diagnose skin cancers,” said Owain Jones, clinical research fellow in the Department of Public Health & Primary Care at the University of Cambridge. “Primary care is where AI/ML algorithms could make a big difference. More accurate assessment of skin lesions could lead to earlier diagnosis of skin cancer, potentially improving outcomes for patients and boosting survival rates.”

The burden on specialist dermatology services could also be reduced and patients reassured when concerning lesions are shown to have a low likelihood of cancer.

This thinking led to Jones and a team of researchers at the University of Cambridge to conduct a study establishing an overview of the research landscape for AI/ML algorithms to evaluate the currently available evidence for their efficacy and safety.

“We were aware that commercially developed AI/ML algorithms and technologies to help diagnose skin cancer were emerging, but that evidence for the safety and effectiveness of these technologies was not commonly available,” Jones said.

For the study investigators considered carefully which types of study to include in the systematic review.

“We were interested in the application of AI/ML algorithms to the diagnosis of skin cancers in primary care, but initial searches showed that there were no research studies that had developed AI/ML algorithms in primary care settings,” Jones said. “Hence, we decided to include all studies that had developed AI/ML algorithms and that had the potential to be used in primary care settings. This led to the inclusion of a large number of studies in our systematic review, but meant that we had included a very broad overview of all of the available evidence in our review.”

Out of the 272 studies included in the review there were no studies that used data from primary care and only 2 studies used data from populations which could be considered as similar to primary care clinical populations

The data showed that AI/ML algorithms demonstrate promising diagnostic accuracy for skin cancer in research settings. However, Jones noted, there is a lack of evidence on the implementation of AI/ML algorithms and on their accuracy in real life clinical settings.

“Our primary interest in this review was for primary care settings and we identified that there is currently a lack of evidence in settings where the prevalence of skin cancer is low, so widespread adoption into primary care practice cannot currently be recommended,” he said. “Our review also highlighted concerns about the datasets used to develop many AI/ML algorithms, and whether these were sufficiently representative of the general population to ensure that these technologies would not suffer from biases in certain minority populations.”

The apparent lack of patients from Black and ethnic minority backgrounds included in the datasets used to develop the AI/ML algorithms was a surprising outcome, and Jones shared that the researchers were not anticipating this result when they started the review.

“Another surprising outcome was the lack of implementation research and studies in real life clinical settings, particularly in view of the huge amount of research that has been done in this field in the last few years,” he said. “This indicated that AI/ML technologies aimed at the diagnosis of skin cancer are probably at an earlier stage of development than we anticipated when started this piece of research.”

The study authors concluded that AI/ML algorithms have great potential to support clinicians in the accurate detection of skin cancers in primary care settings. However, the review shows that research in this area is at an early stage of development and there are concerns as to whether the diagnostic performance demonstrated in the included studies would be maintained among populations with lower skin cancer prevalence or in settings with non-dermatoscopic or lower quality images.

“These algorithms need to be evaluated carefully to ensure that they are accurate, effective, cost-effective, and safe enough for clinical use, and that increased access to skin lesion assessment will not add to the biopsy burden on specialist care providers or contribute to an overdiagnosis of melanoma,” Jones said.

For future studies, the investigators produced a checklist which highlights important aspects AI/ML developers should take into account in the development of these algorithms.

“We hope that the results from our systematic review in combination with this checklist will help improve the quality of research in this area, and help to develop implementable technologies that will produce clinical benefit for patients and clinicians,” Jones said. “Following on from this research we are currently working on a qualitative study evaluating the views of patients, the public, healthcare professionals, and data scientists on the risks and benefits of using AI/ML technologies to help diagnose skin cancer in primary care settings.”