Before joining Binghamton University as a postdoctoral associate, Josh Chen worked as a research scientist at various medical centers throughout New York City. His job often involved using medical artificial intelligence technology. But now, through Binghamton's Excellence in Entrepreneurship and Discovery (EXCEED) program, he has gotten to work at the frontlines of cancer detection alongside Nancy Guo, empire innovation professor, as a developer of that very technology.
"I was trained as a computer scientist, so entrepreneurship and commercialization are like my blind spots," Chen said. "Thanks to the EXCEED training program, I got to learn a lot more about how the technology, idea, and concept materializes into real-world production and valuable assets for other people."
EXCEED is housed at Binghamton under the U.S. National Science Foundation's Accelerating Research Translation (ART) grant. This year, three professors won EXCEED Seed Translational Research Project grants to help accelerate the commercialization of their research - including Professor Yu Chen's work in detecting deepfakes and altered media, Associate Professor Pritam Das' research on power conversion and optimization, as well as Guo's tool that integrates AI into biometric data and cancer image analysis.
Chen, who is part of Guo's research group, additionally won an EXCEED Innovation Fellowship that funded his own work in the area of data privacy protection, for both patients and providers.
APPLICATIONS OPEN
Faculty researchers who want to have an EXCEED's Undergraduate Innovation Intern in the fall 2026 semester may now apply online.
"My team and I are trying to develop novel federated learning methods that can protect both data and model privacy. It's like a two-way protection," Chen said. "It protects the user, the medical center, for data privacy, and also protects the provider - us, the model developer - from model privacy leakage. This way, the model IP can remain secured by the developer's side."
Federated learning, Chen said, is a way to train AI models through multiple remote sites, without the data ever leaving its owners. In other words, it's the model that is shared, not the data itself. This method is important for projects like Guo's, which require copious amounts of medical data for precision oncology. But in terms of complete privacy protection, there are still limitations.
Chen is working on a novel dual-protection mechanism that will not only protect the training data, but also the models. His method was inspired by an established technique called "deep mutual learning." In this setting, two types of models - public and private - can learn from each other, with the public or "proxy" model being the one that gets circulated, while the private model remains safe with its developers.
"The public model can learn from the private model, because it's supposed to be stronger and more advanced. At the same time, the private model can also learn from the public model, because the public model has access to all this new training data from external medical centers. It can gain knowledge from that very valuable extra training data, so they can learn from each other," Chen said. "That's why it's called deep mutual learning."
The goal of these efforts, Chen said, is to come up with a model that has been fine-tuned over multiple medical centers, which can provide highly accurate image analysis down to the cellular level. This, in turn, has the potential to improve cancer detection, diagnosis, and treatment down the road.
Team effort
While Chen takes charge of data privacy, other students in Guo's group play their own parts in the bigger project. Binghamton undergraduate Stephen Barnum, who worked with Guo previously under an EXCEED Undergraduate Innovation Internship, focused on GPU parallel processing, which enabled the group to speed up image analysis that discerned normal cells from cancerous ones.
"Rather than waiting a whole week for one slide, we would only need to wait a day," Barnum said.
Augmenting their efforts is graduate research assistant Mason Dziadulewicz Tipton, whose work bridges radiomics - for example, tissue-level samples versus the much tinier scale of mRNA - with image analysis. Some of the features Guo's group looks at include geometric or textural patterns. This visual analysis can fill in the gaps in more molecular-level data, which could be missing simply due to the nature of data collection.
"It serves as a complement and as validation. Even though what I'm working on in the molecular side is fascinating on its own, I feel there are a lot of mysteries due to certain biological mechanisms," Tipton said. "You can have a wide variety of reasons for your data not falling into the perfect neat distributions you can see in a textbook. But it really adds another dimension to it, and takes learning from the classroom to the real world - where that data is not necessarily playing nice."
Tipton has found their work in Guo's lab, in integrating these different modes of computer science with oncology, to be incredibly rewarding - particularly in a world where it can be easy for technology to inflict harm as it advances.
"AI has the potential, in this technocratic libertarian landscape, to change our world as we know it. It already has," Tipton said. "I think that this work pushes it in the right direction and uses AI for good."
"Exceeding" expectations
For novel research, sometimes the ending is publication: a proposed method, then simulated study, then the results broadcasted in a journal.
With support from EXCEED, however, Chen said he could securely navigate the trials and errors of innovation, and see his method through to the next step toward commercialization. He was able to develop a real-world prototype - which could be distributed in medical centers throughout all of New York state, from Albany to the Big Apple.
"Because of this funding, I was able to go the extra mile from a simulation study of this novel method to real-world prototypes that could potentially be customized into a product," Chen said. "This specific federated learning technology will play a very key role in the deployment of our main product in the future, which is this AI assistant for precision oncology. I would say it definitely exceeded the expectation."
Barnum is working at Raymond Corp. through a co-op, using LiDAR technology to enhance warehouse safety and detect potential obstacles. Though Barnum's path has taken him closer to industry, he said his experience with EXCEED was a valuable pointer to what he wanted.
"Even though I realized I didn't want to do research as a career, I think it was a great experience to test out while I was still an undergrad, when I'm still flexible in what I want to do with my career," he said.
EXCEED has opportunities for many members of a lab, from students to principal investigators. Kathryn Cherny, senior program manager, said that in each year since the NSF first awarded this grant to Binghamton, the number of EXCEED applicants and potential projects has grown.
"These programs are basically pilots to show that there is really viable, impactful translational research going on here, and that we should invest in those things at Binghamton," she said.
Chen said it was a one-of-a-kind experience and arguably a necessary one.
"The core of engineering is making things work. You cannot just stay within theory. You cannot just figure things out in your mind or on paper - you have to make it work in the real world," he said. "Taking the training of EXCEED and the follow-up grant application gives you the chance to really see your research from the point of view of a developer or startup expert. This kind of customer-oriented mindset, not just to develop a tool you think could be useful, but one that you know is useful because real-world users have told you so - that is very important, especially at an early stage."
