The first photographs many parents have of their children are sonograms, used to check on the health of developing fetuses. These sonograms use noninvasive ultrasound technology, but the grainy, gray images they produce can lack diagnostic clarity.
University of Virginia engineering student Soumee Guha wants to change that. She uses artificial intelligence to more clearly peek inside areas of the body that can be medically imaged, including the uterus, and better diagnose issues, such as those that may impact a developing fetus, liver disease, heart function and more.
Guha, a Ph.D. candidate in the University of Virginia’s Department of Electrical and Computer Engineering, builds generative artificial intelligence models to do so. Her research is unique in that it doesn’t require the large volumes of patient data one typically needs to teach diagnostic AI models what to look for. That kind of data isn’t always available.
Why Ultrasound Can Be Ultra-Distorted
To capture ultrasound images of a pregnancy, a doctor or technician glides a specialized wand across the surface of the abdomen.
The device, called a transducer, emits soundwaves through the skin and into the uterus. As those waves bounce off surfaces, they send back data from various depths and angles, resulting in those distorted images that give sonograms their distinctive look, which sometimes inhibits a doctor’s ability to diagnose issues.
Ultrasound isn’t the only form of biomedical imaging that can suffer from complex distortions. The endoscopes, or small cameras, that doctors thread into the body — from the gastrointestinal tract to the respiratory system looking for signs of disease such as tumors, fibrosis and infection — can also send back microscopic images that are hard to read in places.
“Those granular patterns you get in medical imaging are due to light or sound waves being scattered from different points on the image surface. These distortions can obscure details that clinicians are looking for to make a diagnosis,” Guha explained. “The good news is that AI is powerful enough now that it can help us sidestep that problem.”
What Makes Guha’s Research Vital
The specific models Guha builds are called diffusion models and, like all machine learning models, they require huge data sets for training — which consumes time and costs money. Guha eliminates that need by adding mathematical tools that help the algorithm tune into specific ranges for greater accuracy.
Her approach also considers “domain-specific” characteristics of imaging systems and the modes they use to capture images. That’s because each mode of capturing an image is different, leading to variations in the acquired images. The models she developed for her dissertation, such as one for de-speckling images, can be used to improve the quality of sonar, radar and laser images, all of which are degraded by speckling.
“Deep learning models for image enhancement often ignore the underlying physics captured in imaging systems, which leads to results that don’t hold up in clinical or research settings,” Guha said.
“My work bridges that gap by integrating established physical models directly into the algorithmic framework.”
Guha’s doctoral project, titled “Learning Under Multiplicative Noise: Principled Image Enhancement Frameworks for Coherent Imaging,” began in the fall of 2021. She successfully defended it this spring and is preparing to graduate.
Her advisor, American Telephone and Telegraph Company Professor of Engineering and chair of the department Scott Acton, said that Guha is the first researcher to develop diffusion models for de-speckling and the first to address the problem of blurring caused by how an imaging system receives single points of light.
“These aren’t just one-off improvements,” he noted. “Soumee is building solutions as part of a unified framework that will move the field forward.”
So where does the generative aspect come in?
How Artificial Intelligence Can Inform Reality
Many medical settings already apply artificial intelligence to existing medical images to detect problems that may be unobservable or otherwise go overlooked. For example, AI can analyze X-rays, magnetic resonance imaging and computed tomography to see the tiniest fracture or find subtle changes to organs over time that a person could miss. Guha, an experienced researcher in Acton’s image analysis lab, has entered the most cutting-edge segment of the diagnostic AI space.
Generative AI is better known to the public through companies such as ChatGPT, which can produce cogent written text based on the user’s instructions or a photo of something that never existed in reality.
Guha’s approach is a little different in that it uses the AI-generated image as a way to better understand what’s real. Her process deploys mathematical formulas to create diffusion models that make new, clearer images from the original, distorted ones, resulting in more accurate information for medical diagnostics.
The approach works by progressively diffusing, or spreading, “visual noise” within the image field. The AI then dials the noise back, looking for patterns and creating the image it expects to see based on the data and models it trained on previously.
“The result is two different sets of images — one from the imaging equipment and one from AI — that a doctor can compare and contrast, or even combine, to better understand the patient information, including why images might differ from what’s expected,” Guha said.
The work has broad implications, Acton added.
“Soumee’s AI images can even be used to train future AI models, and to judge the accuracy of other tasks such as tumor classification or volume measurements,” he said.
The ‘Sheer Joy of Discovery’
Guha’s journey into image processing began unexpectedly during her undergraduate studies.
“It started with a simple data analysis project,” she recalled. “But soon I was spending countless hours experimenting with image processing techniques, just for the sheer joy of discovery.”
That early passion guided her from a bachelor’s degree in electrical engineering to a master’s in computer science, eventually leading to her Ph.D. research at UVA.
What motivates her now is the deeply interdisciplinary nature of biomedical imaging research — where abstract mathematical theory meets real-world impact.
She is pursuing postdoctoral research opportunities to continue that work.
“There’s something profoundly satisfying about transforming mathematical concepts into algorithms that can directly improve health care outcomes,” she said.