Realistic AI-Generated Ultrasound Images Advance Cardiovascular Diagnostics

Daniek van Aarle stood before rooms full of ultrasound researchers and asked them a simple question: which image is real? She would show them two scans side by side—one pulled from a patient's body, one built entirely by computer. About half the audience guessed wrong. That coin-flip moment of uncertainty was the whole point of her four years of work.

Van Aarle had spent her Ph.D. research learning to generate ultrasound images so faithful to reality that even experts could not reliably tell them apart from the genuine article. On September 29, she defended that work. The goal sounds counterintuitive at first: why simulate an ultrasound when real patient scans already exist? The answer sits in a practical problem that has plagued medical research for years.

When researchers develop new software—say, a program designed to automatically trace the wall of a blood vessel—they need hundreds or thousands of ultrasound images to train and test it. With real patient scans, that work falls to physicians, who must manually outline each vessel wall by hand. The process is tedious. Worse, it is inconsistent. Ask three different doctors to draw the boundary of the same vessel wall, and you get three slightly different lines. That variability introduces noise into the training data, making it harder to build reliable tools.

Van Aarle's virtual ultrasounds solve that problem at its root. Because she built the digital blood vessels herself from scratch, she knew exactly where every boundary lay. She knew the vessel's diameter, how it moved, what the correct answer was supposed to be. Software could then be tested against images where the ground truth was certain. "We're essentially trying to fool the computer a little," she explained. "The simulated ultrasound should resemble a real ultrasound as closely as possible."

Much of her research focused on abdominal aortic aneurysms—dangerous widenings of the body's largest blood vessel. During nearly thirty procedures at Catharina Hospital, physicians inserted a miniature ultrasound probe attached to a thin catheter directly into the artery. Van Aarle captured those clinical images and compared them against virtual ultrasounds she had generated from medical scans of the same patients. The work connected to earlier research by Floor Fasen, who was studying how vessel wall characteristics—thickness, calcifications, blood clots—might improve predictions of aneurysm risk. Simulated images gave Fasen's team a way to test ideas without relying solely on patient data.

Van Aarle conducted her research within e/MTIC, a collaboration between Catharina Hospital, TU/e, Philips, Máxima Medical Center, and Kempenhaeghe. The partnership was designed to move medical innovations from the laboratory into clinical practice faster. For Van Aarle, it meant spending time in the operating room, watching how equipment actually worked in real conditions. She recreated the catheter and the ultrasound probe digitally with as much precision as she could manage. "It was fascinating to see in the hospital how the equipment is actually used in practice," she said.

The virtual ultrasounds are not yet perfect. Van Aarle worked to weave subtle properties of real biological tissue into her models. Highly detailed simulations also demand substantial computing power, and further research will determine how much detail is truly necessary. But she sees her work as a foundation. The simulation toolbox she developed is already being used in follow-up projects within her research group, with more on the way. Researchers can now create datasets that would not otherwise exist—training material for new image-generation methods and new ways of extracting information from ultrasound scans. If experienced ultrasound researchers can no longer tell which image came from a patient and which came from a computer, the virtual images have already arrived at something close to the real thing.