Studies show darker-skinned patients suffer greater mortality from skin cancer. A tumor detection model built by a team that includes mathematics and computer science double-degree student Dhruv Dewan could help close the gap.

For a student so interested in technology, Dhruv Dewan finds a surprising level of comfort in being unplugged. When he’s not exploring the inner workings of artificial intelligence (AI) models, he loves hiking through the woods. 

The University of Maryland double-degree senior in mathematics and computer science is an Eagle Scout and avid backpacker—his favorite trek is a weeklong, 50-mile route in Virginia on the High Knoll Trail in the Blue Ridge Mountains. 

Dewan sees a common thread that unites his two passions. Both foster a mindset of exploration and require transferable skills in problem-solving and resourcefulness—whether it’s to solve a mathematical proof or devise a way to keep food safe from bears. 

Now, Dewan brings that skill set to his studies at UMD. For his four-year research project in the Gemstone program in the Honors College, Dewan and his teammates on “Team Artificial Intelligence Diagnosis” (AID) are developing AI models that can more equitably and accurately detect skin cancer through photos. The three-person team consists of Dewan, a computer science major and a bioengineering major working under the supervision of Heng Huang, Brendan Iribe Endowed Professorship in Computer Science. 

Dewan hopes that doctors can use these models to better diagnose skin cancer in people with darker skin, who have statistically lower survival rates than white patients. 

“I've really seen how exponentially helpful having a deep understanding of math and statistics is for understanding how machine learning and AI work,” Dewan said. “These models just multiply matrices by other matrices. At the end of the day, it’s all truly, purely math.” 

Improving AI for Cancer Detection

Many models can detect skin cancer from images, Dewan said, but they are mostly trained on people with lighter skin tones. 

“That introduces a really big issue in how equitable they are,” he said. “Many models are good at diagnosing skin cancer for whiter and lighter-skinned patients. But when we tested them on a diverse dataset of skin types that included various darker skin tones, we found that they performed extremely poorly.”

Thus, one approach his team takes is to incorporate more diverse data into training sets and ensure that the models pay attention to those data points.

Additionally, the Gemstone team integrates self-supervised learning into the training process. Existing methods train models on photos labeled as cancerous or non-cancerous. By contrast, self-supervised learning provides the model with a larger sample of unlabeled data. Their model learns deeper features of the images using that expansive dataset, which it can later use to identify telltale signs of cancer. This could prevent the model from overfitting on skin tone as its primary diagnostic criterion. 

Dewan finds that his background in mathematics benefits his AI and machine learning research.  

“Many of the techniques that we’re using to improve current methods are pure statistics,” he said. “Having an intuition for statistics and math allows me to understand how the model works and diagnose how to improve it.” 

Dewan expects that the final model, which his team will present at the end of the Spring 2026 semester, will be able to generalize skin cancer beyond only the skin tones it has seen during training. 

"We hope this model helps to provide early diagnosis for darker-skinned patients who have a much lower survival rate for skin cancer because they’re diagnosed too late or because doctors can't diagnose them,” he said. “Hopefully, doctors can use this model as a tool. It feels really fulfilling that this could have an impact.”

Wherever Dewan ends up, he wants to work on robust and scalable software while keeping equity in mind. And, he hopes to quench the thirst for exploration that he developed as an Eagle Scout. 

“Whether I am in academia or in industry,” he said, “I hope to carry that research and exploration mindset going forward.” 

Written by Jason P. Dinh