A new generative AI tool called CytoDiffusion analyzes blood cells with greater accuracy than human experts, potentially improving diagnoses of diseases like leukemia. Developed by researchers from UK universities, the system detects subtle abnormalities and quantifies its own uncertainty. It was trained on over half a million images and excels at flagging rare cases for review.
Researchers from the University of Cambridge, University College London, and Queen Mary University of London have developed CytoDiffusion, a generative AI system that examines blood cell shapes and structures under a microscope. Published in Nature Machine Intelligence, the tool surpasses human specialists in identifying abnormal cells linked to blood disorders such as leukemia, offering higher sensitivity and consistency.
Unlike traditional AI that sorts images into fixed categories, CytoDiffusion models the full spectrum of normal blood cell appearances, making it robust against variations in microscopes or staining methods. It was trained on more than 500,000 blood smear images from Addenbrooke's Hospital in Cambridge, the largest dataset of its kind.
"We've all got many different types of blood cells that have different properties and different roles within our body," said Simon Deltadahl, the study's first author from Cambridge's Department of Applied Mathematics and Theoretical Physics. "White blood cells specialize in fighting infection, for example. But knowing what an unusual or diseased blood cell looks like under a microscope is an important part of diagnosing many diseases."
The system handles the vast scale of blood analysis, where smears contain thousands of cells too numerous for manual review. "Humans can't look at all the cells in a smear -- it's just not possible," Deltadahl noted. "Our model can automate that process, triage the routine cases, and highlight anything unusual for human review."
In tests, CytoDiffusion slightly outperformed humans in accuracy and stood out by reliably assessing its uncertainty. "When we tested its accuracy, the system was slightly better than humans," Deltadahl said. "But where it really stood out was in knowing when it was uncertain. Our model would never say it was certain and then be wrong, but that is something that humans sometimes do."
The AI also generates realistic synthetic blood cell images that fooled hematologists in a Turing test, with experts unable to distinguish them from real ones better than chance.
To advance global research, the team is releasing the dataset publicly. Co-senior author Professor Parashkev Nachev from UCL emphasized its supportive role: "The true value of healthcare AI lies not in approximating human expertise at lower cost, but in enabling greater diagnostic, prognostic, and prescriptive power than either experts or simple statistical models can achieve."
The researchers stress that CytoDiffusion aids clinicians rather than replacing them, with further work needed on speed and diverse populations.
