CytoDiffusion, an AI blood cell analysis system, outperforms human experts in detecting leukemia
CytoDiffusion, an AI-based blood cell analysis system developed by a research team including
Deep generative classification of blood cell morphology | Nature Machine Intelligence
https://www.nature.com/articles/s42256-025-01122-7
SciTechDaily
https://scitechdaily.com/ai-blood-cell-analyzer-outperforms-human-experts-in-detecting-leukemia/
Identifying subtle changes in the size and shape of blood cells, a skill essential for diagnosing blood disorders, requires extensive training, and even experienced clinicians can have difficulty making judgments in difficult cases.
A research team from the University of Cambridge,
'Our bodies contain many different types of blood cells, each with different properties and roles,' said Simon Deltadahl, a researcher at the University of Cambridge's Department of Applied Mathematics and Theoretical Physics. 'A blood sample contains thousands of cells, and it's impossible for a human to analyse every single one. CytoDiffusion automates that process, highlighting any abnormalities for human review.'
◆Train with over 500,000 images
The research team trained the system on images of over 500,000 blood smears collected at Addenbrooke's Hospital in Cambridge, including both common and rare blood cells. By modeling the entire distribution of cell appearances rather than simply classifying them into categories, the AI was able to adapt to differences in medical facilities, microscopes, and staining methods, allowing it to more accurately recognize rare or abnormal cells. Tests demonstrated that the system was able to detect abnormal cells associated with leukemia with much higher sensitivity than existing systems, and also indicated when it was unsure of its diagnosis.
◆ Strengths of being able to grasp uncertainty
'When we tested the accuracy, the system was slightly better than humans, but what really stood out was its ability to grasp uncertain situations. Our model never makes mistakes that we judge with certainty, but humans sometimes do,' Deltadahl said.
◆As a partner supporting clinicians
'As a hematology trainee, I would often be faced with analyzing large numbers of blood smears after a day's work,' said Dr. Steshu Sivapalaratnam of Queen Mary, University of London. 'After working late into the night, I became convinced that AI could do a better job than I could.'
The research team does not intend for CytoDiffusion to replace clinicians, but rather to assist them by quickly identifying abnormal cases and automatically processing routine cases. Future work will involve making the system faster and testing it on a diverse patient population to ensure fairness and accuracy. Furthermore, by providing a public dataset of over 500,000 blood smear images, the team hopes to create an environment where researchers worldwide can build and test new AI models, ultimately contributing to improved patient care.
Related Posts:
