Overview
- Results were published in Nature Machine Intelligence by researchers at the University of Cambridge, UCL and Queen Mary within the BloodCounts! consortium.
- The model was trained on more than 500,000 peripheral blood smear images from Addenbrooke's Hospital, learning full distributions of cell appearances to handle real-world variability.
- In evaluations, it detected leukemia-linked abnormal cells with over 90% sensitivity and 96% specificity, outperforming state-of-the-art systems.
- CytoDiffusion slightly exceeded human accuracy and provided calibrated uncertainty, avoiding confident errors.
- The system can generate realistic synthetic blood-cell images that expert haematologists could not distinguish from real ones, and it is intended as a clinician-assist tool pending speed and fairness validation.