Diagnosis and typing of leukemia using a single peripheral blood cell through deep learning.

Leukemia is highly heterogeneous, meaning that different types of leukemia require different treatments and have different prognoses. Current clinical diagnostic and typing tests are complex and time-consuming. In particular, all of these tests rely on bone marrow aspiration, which is invasive and leads to poor patient compliance, exacerbating treatment delays. Morphological analysis of peripheral blood cells (PBC) is still primarily used to distinguish between benign and malignant hematologic disorders, but it remains a challenge to diagnose and type these diseases solely by direct observation of peripheral blood(PB) smears by human experts. In this study, we apply a segmentation-based enhanced residual network that uses progressive multigranularity training with jigsaw patches. It is trained on a self-built annotated dataset of 21,208 images from 237 patients, including five types of benign white blood cells(WBCs) and eight types of leukemic cells. The network is not only able to discriminate between benign and malignant cells, but also to typify leukemia using a single peripheral blood cell. The network effectively differentiated acute promyelocytic leukemia (APL) from other types of acute myeloid leukemia (non-APL), achieving a precision rate of 89.34%, a recall rate of 97.37%, and an F1 score of 93.18% for APL. In contrast, for non-APL cases, the model achieved a precision rate of 92.86%, but a recall rate of 74.63% and an F1 score of 82.75%. In addition, the model discriminates acute lymphoblastic leukemia(ALL) with the Ph chromosome from those without. This approach could improve patient compliance and enable faster and more accurate typing of leukemias for early diagnosis and treatment to improve survival.