A large publicly available single-cell peripheral blood dataset (MLL23)

Distinguishing cell types in peripheral blood smears is critical for differential diagnosis of blood diseases, such as leukemia subtypes. Machine learning can assist physicians in automating cell classification. Still, the generalizability of deep neural networks remains challenging, particularly concerning domain shifts emerging from variations in, e.g., patient cohorts, staining protocols, scanning procedures, and image resolution. We introduce a large, publicly available, fully annotated peripheral blood dataset comprising over 40,000 single-cell images classified into 18 classes by cytomorphology experts.

In the group of lymphoid cells, there are mature ‘typical lymphocytes’ (number of single-cell images = 5,532) and atypical lymphocytes like plasma cells (1,658), large granular lymphocytes (1,849), reactive lymphocytes (33), hairy cells (3,265) and other neoplastic lymphocytes (180), as well as smudge cells (988). In comparison, the group of myeloid cells is divided into mature cells like band neutrophil granulocytes (687), segmented neutrophil granulocytes (7,170), eosinophil granulocytes (2,448), basophil granulocytes (616), monocytes (2510), and immature cells like myeloblasts (8,606), metamyelocytes (483), promyelocytes (745), myelocytes (747), and atypical promyelocytes (2,033). Lastly, normoblasts (2071) are also present in the dataset. The cell types occur with specific frequencies in the peripheral blood in healthy and pathological patients.