UB computer scientist applies AI to improve medical imaging

From X-rays to ultrasounds, medical imaging provides health care professionals with critical information to treat patients with innumerable conditions.

These technologies are poised for greater advancement in the near future, UB computer scientist Mingchen Gao says, due to the computing power of artificial intelligence (AI).

AI-assisted medical imaging tools offer promise in detecting subtle signs of illness that might be overlooked in traditional examinations, such as minor tissue changes indicative of early-stage cancer, says Gao, assistant professor of computer science and engineering, School of Engineering and Applied Sciences.

“AI models have the potential to go beyond traditional diagnostic methods. They offer a level of detail and efficiency that can significantly aid doctors,” she says. “Life-threatening diseases could be caught early or avoided entirely with these advancements.” 

Last year, Gao received a $578,519 National Science Foundation CAREER award to study AI-assisted medical imaging diagnostics.

Her research centers on developing algorithms that enable machine learning models to analyze medical images. Machine learning is a subset of AI that identifies patterns in datasets to make or refine predictions that they generate without being explicitly programmed to do so.

While there have been advances in medical imaging through deep learning — a subset of machine learning that uses synthetic neural networks to mimic the human brain — AI-assisted medical imaging is not yet widely used in health care settings, says Gao.

This is due, in part, to several common obstacles that researchers have encountered in trying to create practical tools for medical professionals.

These hurdles, all of which Gao’s lab is addressing, include:

• Creating AI tools that adapt to the real world, as opposed to lab-controlled lab experiments.
• Achieving reliable results with limited data, particularly for rare or newly emerging diseases.
• Ensuring AI-assisted medical imaging does not exacerbate health inequities. AI tools can make health care more accessible and equitable, especially in underserved medical areas, she says, adding that in the future AI might be able to recommend treatments or monitor health just through smartphones.
• Overcoming “catastrophic forgetting,” which occurs when deep learning models are sequentially trained on multiple tasks. “Our goal is to learn on a sequence of tasks without losing prior knowledge,” Gao says. 

Gao says her lab has a novel approach to addressing these aforementioned challenges.

“We focus on leveraging the geometric interpretation of deep learning and explicitly bringing that geometric information to design a set of algorithms to tackle these issues,” she says.

Such an approach is advantageous when considering another area of concern: patient privacy. It’s essential to ensure the confidentiality and security of patient data through strict protocols for data handling and model training, she says.

“Another new direction of my lab’s work is on the defense against model extraction, preventing the model to be replicated just through black-box public access,” Gao says.

AI acts as a support tool, enhancing, but not replacing, the expertise of health care professionals who use AI to identify patterns and anomalies that are difficult for the human eye to detect, she says.

“The goal of AI algorithms that we’re developing is to provide assisting information to doctors,” she says. “The doctors are the ones who confirm and make the final decisions.”

Gao says her lab’s work will lead to more accurate clinical decisions while bolstering public confidence in AI-assisted health care.

“The journey is just beginning and the potential impact on global health is immense,” she says.