An artificial intelligence (AI) system developed by researchers at the University of Oxford’s Radcliffe Department of Medicine could allow many heart MRI patients to avoid contrast dye injections, after passing one of the first rigorous real-world clinical tests of its kind.
Normally, cardiac MRI scans require injection of a contrast 'dye', called gadolinium-based contrast agent. This allows visualisation of scars and other damage inside the heart muscle. This is a unique capability of cardiac MRI scans, and can inform important clinical decisions for a patient.
The study is the first prospective, multicentre, blinded validation of an AI technique called Virtual Native Enhancement (VNE), which creates 'virtual contrast' heart MRI images without injecting dye into the bloodstream. Researchers say the technology could eventually shorten scans, reduce costs and make cardiac MRI available to more patients.
Professor Qiang Zhang, lead author from the University of Oxford, said: 'Many AI tools show promise during technical development, but very few are tested rigorously in real clinical environments. This study shows that AI-powered virtual contrast imaging can work reliably across independent hospitals and clinical teams, which is an important step towards real-world patient care.'
Current MRI injections time consuming and unsuitable for some
Heart MRI scans are commonly used to detect scarring caused by heart attacks and other heart diseases. Doctors normally rely on a method called late gadolinium enhancement (LGE), which requires patients to receive an injection of a rare-earth metal gadolinium-based contrast agent. Although this is effective, the injections can cause discomfort, require nursing and clinician's attendance, and increase scan time - and are less suitable for some patients, including some people with advanced kidney disease, mothers who are pregnant or lactating, and children.
The Oxford-led team developed VNE to produce images similar to traditional contrast-enhanced images using only MR images collected before contrast dye would normally be given. The AI analyses standard heart motion and tissue images, then generates enhanced images showing possible areas of damage.
To test whether the technology worked outside the Oxford laboratory, researchers studied 136 patients scanned at hospitals in the UK and China. Importantly, the Oxford researchers generating the AI images had no access to the original 'ground truth' contrast scans or patient clinical information during image generation and analysis. Independent clinicians then assessed the images blindly, reducing the risk of bias.
When the AI-generated images were judged to be high quality and confident, VNE detected heart attack scars with around 94% accuracy - comparable to conventional contrast-enhanced MRI. Independent readers concluded that nearly 70% of these patients have high-quality and confident VNE images, and could potentially avoid contrast injections altogether, without reducing diagnostic accuracy.
The researchers say the work highlights a broader challenge in medical AI. Many systems perform well in early studies in single laboratories, but fail to translate into routine healthcare, because they struggle with image quality, reproducibility, or differences between hospitals and scanners.
Professor Stefan Piechnik, joint senior author, added: 'A major problem with medical AI is that systems which work well in one hospital often fail when tested somewhere else. We designed this study very carefully to make sure the scans were performed consistently and the AI was tested fairly in real clinical settings.'
'That is why this work matters beyond heart imaging alone. It shows how AI can be developed, tested and introduced into healthcare in a safe and reliable way.'
Professor Vanessa Ferreira, joint senior author of this study, and President of the Society for Cardiovascular Magnetic Resonance (SCMR), said: 'This technology could make heart MRI scans quicker, simpler and easier for many patients by reducing the need for contrast injections. In this study, we found that when the AI-generated images were clear and high quality, they performed as well as standard contrast scans for detecting heart attack scars.
'We still need larger studies and broader testing of different cardiac diseases before this could replace standard contrast-enhanced cardiac MRI scans for everyone. But the technology is now near the stage where it could begin helping doctors decide which patients really need contrast injections, and which do not. Our long-term goal is to make cardiac MRI as straightforward and accessible as an ultrasound heart scan.'
The study is a collaboration with the University of Leeds, and Fuwai Hospital of the Chinese Academy of Medical Sciences in China, and funded by the British Heart Foundation and Kusuma Trust.
Dr Sonya Babu-Narayan, clinical director at the British Heart Foundation and consultant cardiologist, said: 'Cardiovascular MRI with contrast is the clinical gold standard tool to detect heart scars. For two thirds of heart attack patients in this study, AI-generated images were able to accurately mimic the kind obtained after contrast – this implies that for a large number of patients cardiac MRI could be quicker and needle-free in future. In the meantime, this tool may allow doctors to better triage those patients that will need contrast given at cardiac MRI.'
The study is published in the Journal of the American College of Cardiology.