Stony Brook researchers pioneer a new method to tackle critical medical imaging challenges in radiotherapy, breast cancer screening, and pediatric care.
Stony Brook, NY, Apr 21, 2025 – Researchers at Stony Brook University are developing a new AI-based method that could transform medical imaging, making CT scans faster, safer, and more accurate, without needing huge amounts of training data or exposing patients to high doses of radiation.
Supported by Stony Brook University’s AI Seed Grant Program, the team is working with a new deep learning technique called Deep Image Prior (DIP). Unlike most AI tools, DIP doesn’t need to learn from large databases of past patient images. Instead, it builds a high-quality image from scratch, using only the data from the scan in front of it. This means it can still work well even when data is limited or when a patient’s medical history isn’t available.
Senior postdoctoral associate Dr. Ziyu Shu, Department of Radiation Oncology, has taken the idea further with a new method called RBP-DIP (short for Residual Back Projection with Deep Image Prior). His approach uses a smart, step-by-step process to improve image clarity, especially in tricky situations, like when a patient moves during a scan, or when doctors need to reduce the amount of radiation.
The team, led by Principal Investigator Dr. Xin Qian, clinical assistant professor, Department of Radiation Oncology, is focusing on three major challenges in current medical imaging systems:
First, during radiation therapy, doctors often use CT scans to track a patient’s internal anatomy and adjust treatment. But those scans take about a minute, long enough for natural movements like breathing to blur the image. With this new method, doctors could get clear images much faster, potentially in just one breath-hold, making treatments safer and more precise.
Second, in breast cancer screening, a technique called digital tomosynthesis takes images from limited angles. It is useful, but the image quality isn’t always good enough to spot small tumors. The researchers believe their AI approach can reconstruct clearer, more detailed images from the same scan without needing extra time or higher radiation.
Third, the team is seeking ways to improve low-dose CT scans for sensitive patient groups like children or people needing frequent screenings. Current methods often require trade-offs between image quality and safety. But with RBP-DIP, CT systems will need fewer X-ray images and deliver better results with less radiation.
Dr. Xian said, “What makes this approach stand out is that it doesn’t rely on a database of previous patients. While it can use a prior image of the same patient if available, it’s not required. That’s especially important in healthcare settings where privacy, variety, and fast decision-making are key.”
“Instead of relying on the past, we’re optimizing for the patient in front of us,” added Dr. Shu. “This makes the system flexible, fast, and safer for real clinical use.”
The team has already seen promising results. Using experimental CT machines, they’ve managed to create clear images using far fewer X-ray angles than usual, without the grainy artifacts or blur that older methods often produce. In some cases, just 51 projections were enough to build a sharp, detailed scan.
This project is part of a growing shift in how AI is being used in medicine. Rather than treating AI as a mysterious black box, researchers are finding ways to make it more transparent, controllable, and tailored to individual patients.
As the work continues, the team at Stony Brook is gathering data to support future clinical studies and larger funding opportunities. Once perfected, their method could help build the next generation of CT scanners — machines that are quicker, more accurate, and better suited to patient needs.
Communications Assistant
Ankita Nagpal