A new artificial intelligence system designed to analyse the shape and structure of blood cells is showing signs it could improve how conditions such as leukaemia are detected.
Researchers behind the system, known as CytoDiffusion, say it can identify abnormal cells more accurately and consistently than human specialists, potentially reducing missed or uncertain diagnoses in busy clinical settings.
CytoDiffusion uses generative AI, similar to image-generation systems such as DALL-E, to examine subtle visual variation in blood cells. Traditional medical AI tools are typically trained to assign images to fixed categories. The researchers argue this approach struggles with rare or ambiguous cases.
“Knowing what an unusual or diseased blood cell looks like under a microscope is an important part of diagnosing many diseases,” said Simon Deltadahl, the study’s first author.
A single blood smear can contain thousands of cells, far more than a clinician can realistically examine in full. “Humans can’t look at all the cells in a smear – it’s just not possible,” Deltadahl said.
He described CytoDiffusion as a triage tool that flags unusual cells for human review.
“After a day of work, I would face a lot of blood films to analyse,” said Dr Suthesh Sivapalaratnam. “I became convinced AI would do a better job than me.”
The system was trained on more than half a million blood smear images from Addenbrooke’s Hospital. In tests, it showed higher sensitivity in identifying leukaemia-related abnormalities and was able to assess its own uncertainty. “When we tested its accuracy, the system was slightly better than humans,” Deltadahl said. Researchers stress the system is designed to support, not replace, doctors.