The real value of healthcare AI lies not in its low-cost approximation of human expertise, but in enabling greater diagnostic, prognostic, and prescriptive power than experts or simple statistical models can achieve.
Parasikev Nachev
In tests, CytoDiffusion was able to detect abnormal cells associated with leukemia with much higher sensitivity than existing systems. It also quantified its own uncertainty as well as or better than current state-of-the-art models, even when given far fewer training examples. “When we tested its accuracy, the system was slightly better than humans,” Deltadahl said. “But what really set it apart was knowing when to be uncertain. Our model was never going to say it was certain and then be wrong, which is something humans do sometimes.”
“We evaluated our method against a number of challenges found in real-world AI, including never-before-seen images, images captured on different machines, and degrees of label uncertainty,” said Associate Senior Author Professor Michael Roberts, also from the University of Cambridge’s Department of Applied Mathematics and Theoretical Physics. “We believe this framework will be useful for researchers as it provides a multidimensional view of model performance.”
The research team also showed that CytoDiffusion can generate synthetic blood cell images that are indistinguishable from real blood cell images. In a “Turing test” conducted by 10 experienced hematologists, it was only by chance that human experts could tell the real thing from an AI-generated image. “I was really surprised,” Deltadahl said. “They look at blood cells all day long, and even they couldn’t tell them apart.”
As part of the project, researchers are releasing the world’s largest publicly available dataset of peripheral blood smear images, totaling more than 500,000 images. “By opening up this resource, we hope to enable researchers around the world to build and test new AI models, democratizing access to high-quality medical data and ultimately contributing to better patient care,” said Deltadahl.
Although the results are promising, researchers say CytoDiffusion cannot replace a trained clinician. Instead, it is designed to support them by quickly flagging and reviewing unusual cases and automatically handling more routine cases. “The real value of healthcare AI lies not in approximating human expertise at a lower cost, but in enabling greater diagnostic, prognostic and prescriptive power than experts or simple statistical models can achieve,” said co-senior author Professor Palaszkev Natchev from UCL. “Our research suggests that generative AI will be central to this mission, transforming not only the fidelity of clinical support systems but also our insight into the limits of our own knowledge. This ‘metacognitive’ awareness, or knowing what we don’t know, is critical to clinical decision-making, and here we show that machines may be better than us in this area.”
Researchers say more work is needed to speed up the system and test it on a diverse patient population to ensure fairness and accuracy.
This research was supported in part by the Trinity Challenge, Wellcome, the British Heart Foundation, Cambridge University Hospitals NHS Foundation Trust, Barts Health NHS Trust, NIHR Cambridge Biomedical Research Center, NIHR UCLH Biomedical Research Center, and NHS Blood and Transplant. This study was conducted by BloodCounts!’s Imaging Working Group. A consortium aimed at using AI to improve blood diagnostics globally. Simon Deltadahl is a Fellow of Lucy Cavendish College, Cambridge.
sauce: © University of Cambridge (CC BY-NC-SA 4.0)
