乌鸦传媒

Dermatologists were more accurate than primary care physicians in diagnosing skin cancer, according to a systematic review and meta-analysis that appears in .

According to the study team, the findings could help inform benchmarks for clinical trials, provider training, and evaluating novel diagnostic tools including artificial intelligence (AI).

A total of 100 studies were ultimately included in the analysis. With experienced dermatologists using clinical examination and clinical images, the respective sensitivity and specificity for diagnosing keratinocytic carcinomas were 79.0% and 89.1%. For dermatologists using dermoscopy and dermoscopic images, sensitivity and specificity were 83.7% and 87.4%, respectively. For primary care providers, respective sensitivity and specificity were 81.4% and 80.1%.

Further, experienced dermatologists had a 2.5-fold higher likelihood of accurately diagnosing keratinocytic carcinomas using in-person dermoscopy and dermoscopic images compared with in-person clinical examination and images.

With experienced dermatologists using clinical exam and images, respective sensitivity and specificity for diagnosing melanoma were 76.9% and 89.1%; for inexperienced dermatologists, 78.3% and 66.2%; and for PCPs, 37.5% and 84.6%. Using in-person dermoscopy and dermoscopic images, sensitivity and specificity were 85.7% and 81.3% for experienced dermatologists; 78.0% and 69.5% for inexperienced dermatologists, and 49.5% and 91.3% for PCPs.

Maria Wei, MD, PhD, is a dermatologist and melanoma specialist with the University of California, San Francisco (UCSF). She also directs the UCSF Pigmented Lesions/Melanoma Surveillance Clinic. Wei recently discussed the investigation, its findings, and its potential implications with the Reading Room. The exchange has been edited for length and clarity.

What was the need for or key goal of this analysis?

Wei: Right now, we are in the era of precision medicine. And the whole basis of being precise in medicine is the diagnosis. You have to get the diagnosis correct. Historically there hasn't been much of an emphasis on diagnostic skills. It's been mostly therapeutics.

But in the last few years, we're understanding that we need to put some research and understanding into diagnostic accuracy. And so, there have been a lot of papers.

But altogether, what do all these papers tell us?

As one example, there are a lot of new and novel diagnostic techniques, including but not limited to AI. But each of these tests has its own little data set, and that didn't seem to me like a good way to go about it. I think what I was looking for was a benchmark to compare novel diagnostic methods.

Another important part of the context or background for this is the United States Preventive Task Force. When they surveyed the literature on skin cancer screenings, they said, "there is not enough evidence to say with certainty that skin screening for skin cancer is useful." For dermatologists, that's our bread and butter. So our reaction was "how could that be?"

I took a look at the studies the task force looked at. All the studies involved primary care physicians screening general adults.

So, in this paper, I wanted to ask the question: How good are dermatologists and primary care physicians at diagnosing skin cancer? This paper was designed in part to look closely at that issue.

Ultimately then the paper served three purposes: to assess how well we diagnose before we launch these huge clinical trials, to set up benchmarks for novel diagnostic modalities, and to explore the need to standardize training related to screening both for dermatologists and primary care physicians.

With so many ambitious questions at hand, how would you summarize what the research team discovered?

Wei: Dermatologists had about 13-fold higher odds of accurate diagnosis of melanoma compared to primary care physicians.

We also found that if you add the dermatoscope to the unaided visual exam, you're gaining about a six-fold increase in accuracy of diagnosing a melanoma.

Our results actually suggest that the optimal way to do a skin exam is to combine a visual unaided exam with the dermatoscope. That's because, in using your unaided eye, you're more specific in your diagnosis. On the other hand, with the dermatoscope, you're more sensitive. So, putting them together, you get the highest sensitivity and the specificity.

Were you surprised by any of the study findings?

Wei: It was very surprising to me that the accuracy of primary care physicians in diagnosing melanoma was under 50% in sensitivity, using either unaided visual or magnified techniques. I fully expected dermatologists who are trained in skin exams to surpass primary care physicians, but I did not expect [primary care physicians] to be less than 50%.

That has a lot of implications for large-scale studies that use primary care physicians who are diagnosing melanoma, rather than dermatologists. I think that these large studies need to be reconsidered.

What are the clinical take-home messages from this investigation for dermatologists?

Wei: For dermatologists, I think it's important to be trained in dermoscopy, because you gain that substantial increase in accuracy for diagnosing melanoma. It's only recently that the training of dermoscopy has become universal in dermatology residencies.

Secondly, I think we need to be aware of the need to either train primary physicians better in skin exams, or empower them with tools such as AI. I think AI can play a real role in boosting the diagnostic accuracy among primary care physicians.

Thirdly, this will help standardize and create benchmarks for clinical trials, training, and new and novel techniques like AI.

What kind of role do you see AI playing in diagnostics in the foreseeable future?

Wei: I think there's a role for AI in working with physicians. As we can see, we're not 100% accurate and neither is AI. But studies have shown that AI and clinicians together, specifically dermatologists, are better than either one alone.

So, I say we're better together. And I think there's a role for AI because of that. I think that AI tools definitely need supervision, but these can still be effective in a complementary role. That's the key.