Accurate morphological classification of blood leukocytes remains a cornerstone of hematological diagnostics but still relies on manual expertise, leading to variability across laboratories. The CytologIA data challenge was organized to benchmark artificial intelligence (AI) models for automated classification of normal and pathological leukocytes from peripheral blood smears. Twenty hematology laboratories from France, Belgium, and Switzerland contributed to a multicentric, expert-annotated database of 69,168 images encompassing 23 leukocyte classes. A total of 245 teams from academia, hospitals, and industry participated in the challenge. The top-performing model-combining a YOLOX-based detection module with an ensemble of transformer and convolutional classifiers-achieved a balanced accuracy of 0.94 on the hidden test set, markedly outperforming the baseline CNN (0.82). While abundant cell types such as neutrophils were identified with near-perfect accuracy (>0.97), rarer and morphologically overlapping categories remained challenging. All data and models were released openly on data.gouv.fr and GitHub to ensure full reproducibility. CytologIA represents the first large-scale, open, and collaborative AI benchmark in hematological morphology, establishing a new reference for the development of robust, transferable diagnostic algorithms across institutions.
Sourdeau et al. (Fri,) studied this question.