An AI tool that can analyse abnormalities in the shape and form of blood cells, and with greater accuracy and reliability than human experts, could change the way conditions such as leukaemia are diagnosed.
Researchers have created a system called CytoDiffusion that uses generative AI – the same type of technology behind image generators such as DALL-E – to study the shape and structure of blood cells.
Unlike many AI models, which are trained to simply recognise patterns, CytoDiffusion – developed by researchers at the University of Cambridge, University College London and Queen Mary University of London – could accurately identify a wide range of normal blood cell appearances and spot unusual or rare cells that may indicate disease. The results are reported in the journal Nature Machine Intelligence.
Spotting subtle differences in blood cell size, shape and appearance is a cornerstone of diagnosing many blood disorders. But the task requires years of training, and even then, different doctors can disagree on difficult cases.
“We’ve all got many different types of blood cells that have different properties and different roles within our body,” said Simon Deltadahl from Cambridge’s Department of Applied Mathematics and Theoretical Physics, the study’s first author. “White blood cells specialise in fighting infection, for example. Knowing what an unusual or diseased blood cell looks like under a microscope is an important part of diagnosing many diseases.”
However, a typical blood ‘smear’ contains thousands of cells – far more than any human could analyse. “Humans can’t look at all the cells in a smear – it’s just not possible,” said Deltadahl. “Our model can automate that process, triage the routine cases, and highlight anything unusual for human review.”
“The clinical challenge I faced as a junior haematology doctor was that after a day of work, I would have a lot of blood films to analyse,” said co-senior author Dr Suthesh Sivapalaratnam from Queen Mary University of London. “As I was analysing them in the late hours, I became convinced AI would do a better job than me.”
To develop CytoDiffusion, the researchers trained the system on over half a million images of blood smears collected at Addenbrooke’s Hospital in Cambridge. The dataset – the largest of its kind – included both common blood cell types and rarer examples, as well as elements that can confuse automated systems.
By modelling the full distribution of cell appearances rather than just learning to separate categories, the AI became more robust to differences between hospitals, microscopes and staining methods, and better able to recognise rare or abnormal cells.
In tests, CytoDiffusion could detect abnormal cells linked to leukaemia with far greater sensitivity than existing systems. It also matched or surpassed current state-of-the-art models, even when given far few training examples; and quantify its own uncertainty.
“When we tested its accuracy, the system was slightly better than humans,” said Deltadahl. “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.”
“We evaluated our method against many of the challenges seen in real-world AI, such as never-before-seen images, images captured by different machines and the degree of uncertainty in the labels,” said co-senior author Professor Michael Roberts, also from Cambridge’s Department of Applied Mathematics and Theoretical Physics. “This framework gives a multi-faceted view of model performance which we believe will be beneficial to researchers.”
The team also showed that CytoDiffusion could generate synthetic blood cell images indistinguishable from real ones. In a ‘Turing test’ with ten experienced haematologists, the human experts were no better than chance at telling real from AI-generated images.
“That really surprised me,” said Deltadahl. “These are people who stare at blood cells all day, and even they couldn’t tell.”
As part of the project, the researchers are releasing what they say is the world’s largest publicly available dataset of peripheral blood smear images: more than half a million in total.
“By making this resource open, we hope to empower researchers worldwide to build and test new AI models, democratise access to high-quality medical data, and ultimately contribute to better patient care,” said Deltadahl.
While the results are promising, the researchers say that CytoDiffusion is not a replacement for trained clinicians. Instead, it is designed to support them by rapidly flagging abnormal cases for review and handling more routine ones automatically.
“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,” said co-senior author Professor Parashkev Nachev from UCL. “Our work suggests that generative AI will be central to this mission, transforming not only the fidelity of clinical support systems but their insight into the limits of their own knowledge. This ‘metacognitive’ awareness – knowing what one does not know – is critical to clinical decision-making, and here we show machines may be better at it than we are.”
The researchers say further work is needed to make the system faster and to test it across diverse patient populations to ensure fairness and accuracy.
The research was supported in part by the Trinity Challenge, Wellcome, the British Heart Foundation, Cambridge University Hospitals NHS Trust, Barts Health NHS Trust, the NIHR Cambridge Biomedical Research Centre, NIHR UCLH Biomedical Research Centre, and NHS Blood and Transplant. The research was conducted by the Imaging working group within the BloodCounts! consortium, which aims to use AI to improve blood diagnostics globally.
