Researchers at the University of Southern California have created a hybrid quantum-classical artificial intelligence system that could help doctors identify tumors more precisely while dramatically reducing the computing power required. The breakthrough, announced this week, combines quantum computing with conventional AI to improve how medical imaging systems analyze cancer scans, potentially making advanced diagnostic tools more practical for hospitals with limited computing resources.

How Does Quantum AI Improve Cancer Imaging?

The challenge that motivated this research is fundamental to modern cancer care: doctors need to identify exactly where a tumor ends and healthy tissue begins. This process, called image segmentation, is critical because it affects surgery planning, radiation targeting, and treatment tracking. Radiologists currently spend significant time manually outlining tumors and nearby organs before treatment, a labor-intensive task that can delay care.

Amir Kalev, lead quantum scientist at USC Viterbi's Information Sciences Institute, suspected that quantum computing could strengthen AI's ability to draw these boundaries accurately. Working with former student Naman Jain, who earned a master's degree in quantum information science in 2025, Kalev developed a quantum module called Quantum Feature Extraction, or QuFeX. Rather than replacing conventional computing, QuFeX was designed to enhance existing AI systems.

"Quantum computing's value may depend partly on efficiency rather than simply outperforming conventional machines in speed," noted Naman Jain.

Naman Jain, Quantum Information Science Researcher at USC

The researchers incorporated QuFeX into an established medical imaging system called U-Net, a neural network architecture widely used for image segmentation. The resulting hybrid system, called Qu-Net, combines conventional artificial intelligence with quantum computing techniques. When tested against leading conventional AI models across several image-segmentation benchmarks involving medical images, Qu-Net achieved roughly a 7% improvement in accuracy.

Why Does Using Fewer Computing Parameters Matter?

The efficiency gains may prove just as important as the accuracy improvements. During testing, Qu-Net used approximately 250,000 parameters, internal values that an AI system learns during training to interpret information and make predictions. By contrast, the conventional U-Net required approximately 1.5 million parameters, or about six times as many. This dramatic reduction in computational complexity could have significant real-world implications.

Smaller models require less computing power to train and deploy, making sophisticated AI tools more accessible to hospitals and clinics where computing resources remain limited. The efficiency gains also mean faster processing during clinical use, which could speed up the time-consuming process of preparing radiation treatment plans. Kalev ultimately wants to shorten personalized radiation-treatment planning from days to a single clinical visit, allowing doctors to scan a patient, develop the plan, and potentially begin treatment much sooner.

Steps to Bring Quantum AI Into Clinical Practice

• Clinical Collaboration: Kalev has started working with physicians at USC's Keck School of Medicine, including Dr. Eric Chang, chair of the Department of Radiation Oncology, to investigate practical applications of the quantum machine learning system.
• Real-World Testing: The research team plans to test their technology using both simulated scans and images from actual patients to validate performance in clinical settings.
• Adaptive Treatment Planning: Successful testing could lead to systems that help physicians adjust radiation plans as tumors change during treatment, since tumors can shrink, grow, or shift position between sessions.

The research, published in the journal Quantum Science and Technology, was motivated by a significant problem facing medical AI: limited training data. Large AI models typically improve when developers train them on enormous datasets, but medical researchers often lack that luxury. Patient privacy concerns, specialized imaging requirements, and limited examples of certain diseases constrain available training information. Medical images can also vary substantially in quality and consistency, making it harder for conventional AI systems to learn robust patterns.

Jain said the team wanted to determine whether quantum technology could help artificial intelligence overcome these data limitations. The hybrid approach appears to have succeeded, producing more accurate boundaries around suspicious tissue during experiments. Those improvements could eventually help radiation oncologists determine which areas require treatment while carefully limiting radiation exposure to surrounding healthy tissue.

However, the research remains at an early stage and has not yet established Qu-Net as a clinical diagnostic tool. The technology would still require extensive testing before researchers could establish whether it delivers the promised benefits in real clinical practice. Beyond cancer care, the underlying technology could eventually assist doctors working with brain disorders, cardiovascular disease, and surgical planning, suggesting the quantum-AI approach may have applications far beyond oncology.