Detecting melanoma isn’t a one-size-fits-all endeavor. Different skin tones, ethnicity, levels of sun exposure, and the patient’s medical history all play a fundamental role in early cancer detection, yet medical technology has often struggled to reflect this diversity.
Stanford Cancer Institute member Roxana Daneshjou, MD, PhD, assistant professor of biomedical data science and of dermatology, is working to close this gap by developing a more inclusive framework for artificial intelligence (AI) that recognizes the unique needs of every patient rather than relying on a clinical average. This push for personalized, equitable care is the primary focus of two of Daneshjou’s research avenues.
Data is king: Building a diverse dataset
The first pillar of Daneshjou’s research focuses on diversifying the datasets used to train algorithms so they perform accurately across all skin tones. Many existing algorithms were built primarily using images of white skin, creating a dangerous gap in diagnostic accuracy for patients of color, as the hallmark signs of skin cancer can present differently.
When we talk about AI models, data is king."
In order to build better, more fair, accurate AI algorithms for this space of skin cancer and melanoma, we need really good data. When we talk about AI models, data is king," Daneshjou explained. "And you can't really build these models without good data, which is why we've been so focused on dataset development and using that for model evaluation."
To improve representation, Daneshjou developed a dataset of skin cancer across diverse skin tones to help evaluate model equity. Additionally, she has developed multimodal image datasets that include both standard clinical images and dermoscopic images, which provide magnified views that yield deeper insights into a lesion’s structure.
“Because a human dermatologist uses data from both clinical and dermoscopic images to make decisions in a clinical setting, we wanted to have that paired data to be able to build algorithms that more closely mimic the data for decision-making,” she said.
The goal is to move beyond unimodal AI, which handles only one type of data.
The team is now transitioning this technology into real-world evaluation, partnering with Memorial Sloan Kettering Cancer Center to launch large-scale prospective trials. These trials aim to observe how the AI model’s augmentation actually influences a clinician’s behavior.
"We’ve helped with dataset development because we want to have ways to build and evaluate algorithms," Daneshjou said. "And now we're also moving towards doing a lot of evaluation work, like real-world evaluation work to understand which things are ready for prime time and just even how doctors get influenced by AI decision-making."
SkinSentry: democratizing access to dermatological care
The second major pillar of Daneshjou’s research involves a partnership with the Canary Center at the Stanford School of Medicine to address an accessibility issue. Notably, an estimated 3 billion people currently lack access to professional dermatological care.
High-fidelity whole-body imaging systems are often prohibitively expensive, costing upwards of $300,000 on average. This expense makes them nearly impossible to implement in rural settings or under-resourced clinics.
Enter SkinSentry, a project designed to provide high-fidelity imaging at a fraction of the cost.
The project was born of the idea that if imaging is affordable, it can unlock many different diagnostic tasks in dermatology for those who currently lack access, per Daneshjou.
"The unmet need is that basically it's very hard to get access to dermatology, especially in rural communities," she said. "And so the idea would be you could have something like this in your primary care doctor's office or even your local pharmacy."
While not replacing a doctor's diagnosis, SkinSentry functions as a sophisticated triage tool. By identifying high-risk factors such as an increased atypical mole count or significant sun damage, the device can flag individuals who need to see a specialist. This prevents overdiagnosis while ensuring that those at the highest risk do not fall through the cracks.
Currently, Daneshjou is testing a prototype of the device, with plans to refine the software in future clinical trials.
“We're early days but exciting early days,” she remarked. “The improvement never stops.”
