spotlight with Albert Sean Chiou, MD, MBA

AI, a game changer for melanoma?

Dr. Albert Sean Chiou, Clinical Associate Professor of Dermatology, directs Stanford Skin Innovation and Interventional Research Group and clinical research at the Stanford Department of Dermatology. Dr. Chiou and his team are exploring ways to advance the prevention, early detection, and diagnosis of melanoma through artificial intelligence technology. 

For this issue, I spoke with Dr. Chiou about his team’s findings on the application of artificial intelligence (AI) in patients, where the research currently stands, and how AI will impact the future of melanoma patient care.

Please describe the use of Artificial Intelligence in skin cancer diagnosis. How can this technology advance prevention and early detection of melanoma? 

Artificial Intelligence can be used for very different purposes. Dermatologists have been exploring it to support early detection and diagnosis of skin cancer.

Skin cancer detection devices have been around for a number of years. A handful have been FDA approved. These technologies are used by clinicians who have existing expertise in skin cancer (typically dermatologists) to help them make better decisions about which skin lesions could potentially represent melanoma and require a biopsy. We've seen some potential applications that move this further upstream to non-specialists. This includes the recent authorization of a tool for use by primary care doctors to help them decide which lesions might be concerning enough to refer to a dermatologist.

A lot of the recent excitement is around software-based computer vision AI skin cancer detectors. In theory, these have the ability to put AI decision support in the hands of anyone with a smartphone that can take pictures of lesions. Since this is the “newer kid on the block”, we're still trying to figure out full potential benefits as well as harms, like false alarms. But, the overall trend is that we're getting more and more widely accessible tools that can potentially identify melanomas.

To answer your second question, where does AI contribute to early detection and prevention of melanoma--Increased access is probably one of the main anticipated benefits from these technologies. In terms of where and how to use these tools, we are still assessing where the safest and most effective places are to insert them into the melanoma diagnostic pathway, as there are different potential touchpoints involving patients, their primary care doctor, and ultimately specialists. It’s about trying to decide where AI works best and the safest way to use it.

How does skin cancer AI technology work?

One simple way to think about how skin cancer AI algorithms work is that they often involve photos taken of the lesion or a magnified picture of the lesion called a dermoscopic image. The AI algorithm does some complex math in the background and gives you a risk score. It’s up to dermatologists or others who receive this risk score to decide what to do with it.

The power of AI is that it is often trained on tens of thousands, sometimes hundreds of thousands of images of melanomas and nonmelanoma skin lesions. Just imagine a doctor having had that much experience, seeing that many lesions, and how good they can get at differentiating between them with all that information. If the information is very well annotated and the diagnoses are very clear, the technology can potentially be very good at picking out the lesions that are concerning for melanoma.

What motivated your interest in leading the Stanford Skin Innovation and Interventional Research Group as well as the potential use of AI for reducing disease burden in melanoma?

As dermatologists, we unfortunately end up encountering melanomas far more often than we’d like. Fortunately, most of our patients do very well.  We frequently catch them early, and the treatment is often successful. However, we all see cases where that doesn't happen, and knowing how aggressive melanoma can be if it presents at a more advanced stage, all of us are motivated to find ways we can do better at this. For me, I'm really interested in exploring new technologies that potentially make us more effective at catching early-stage melanomas, and that can help patients figure out when to see a healthcare provider. 

My colleagues and I get fired up about the process of trying to figure out if a new skin cancer diagnostic or some new dermatologic treatment is actually going to be clinically helpful. We frequently engage in a lot of collaborations with other doctors, other researchers within our university, other schools, or industry partners. There’s much nuance to thinking through how to rigorously evaluate these technologies and figure out if they're useful.

The Stanford Skin Innovation and Interventional Research Group has been around about 7 years. It grew organically from Stanford dermatologists who wanted to investigate new treatments and diagnostics and work together to build a team that would investigate these strategies to help our patients. Our team often includes clinical experts and research coordinators who go out in the clinic and do the hard work. We also partner with computer and data scientists and translational researchers in the lab.

What kinds of evidence currently exist or are needed to support the use of AI in routine care for the early detection of melanoma?

Currently, there's a rigorous framework in place from our regulatory bodies for AI enabled skin cancer detection devices. This strikes an important balance in promoting innovation, but also in requiring robust clinical evidence so the devices help patients and not produce unnecessary harms. For the newer software-based computer vision technologies, we’re still early in the process of their evaluation. 

Most of the evidence for them frequently comes from experimental settings, often performed using historical photos of skin cancers, which are sometimes not comparable to real world situations. At this point, we're seeing that under these very constrained scenarios that many of these software-based technologies can perform very well in identifying potential melanomas from benign skin spots. However, they sometimes have unpredictable results. One example is how changes in lighting or having a hair or shadow cross over a lesion can throw the algorithms off.

Furthermore, our own group has shown that AI performance can change significantly, depending on the tone of your skin. This, along with similar findings from other groups on the effects of what should be seemingly innocent factors, raise concerns about when these technologies might perform less predictably. Based on this, many of us favor having a doctor in the loop to help interpret the results and guide patients on how to act on them.

What do we know about how AI might work in diverse patient populations?

There are many considerations here. One is that historically within dermatology we often have fewer photos of patients with darker skin tones. These AI devices and software are basically as good as the data you use to train them. As a result, the photos we have for these image classifiers do not represent the broad distribution of race and ethnicity that we see in our patient population. We've shown through some of our work that the clinical image-based algorithms may not work as well or as predictably during encounters with a patient who has different factors than the device’s original training, including if they have darker skin tones. Fairness in AI is an important consideration as we hope each patient who uses it has the best shot at getting a result that’s accurate and safe. The good news is that there are strategies to improve AI by increasing the diversity of photos used, and by technical approaches like “fine-tuning” that can improve the performance of these algorithms.

What final thoughts would you like readers to take away from this interview?

With AI, there's potential to significantly increase access to melanoma early detection tools. If we integrate them wisely into our current healthcare system, they could expand our toolkit for early detection and diagnosis. Still, similar to many new diagnostic technologies, we’re seeing a potential complicated picture where access is probably going to be balanced against occasional unpredictable results and false alarms. The statistical reality is that for every new melanoma these technologies help us catch, they will also be responsible for generating a larger number of false alarms and the worry associated with this. 

Overall, we're in for a very human conversation over the next few years to try to find the best balance between all these factors. Having a clinician involved to help interpret the results is likely going to be very important, and we have evidence that doctors using these types of AI for decision support can perform better than the AI alone or the doctor alone. We will likely be learning more about this in the coming years.

Learn more

• Realities of Artificial Intelligence for Early Melanoma Detection

• Artificial Intelligence and Cancer

ELLEN DINUCCI is a contributing writer of WOM-California Gazette and staff member of the Office of Cancer Health Equity at Stanford Cancer Institute.

Health information and cited sources in this newsletter are for educational purposes only. The material from this newsletter is not intended to be a substitute for professional medical/health advice, diagnosis, or treatment.