Deep-learning assisted label-free hematology analysis through defocusing a regular microscope

Abstract Hematology analysis is the most fundamental clinical test in differential diagnosis. The conventional way uses a laboratory hematology analyzer for complete blood count and requires highly trained professionals for morphology examination. The whole procedure is time- and labor-intensive, which severely affects the disease diagnosis efficiency. Meanwhile, the requirement for laboratory instruments and professionals impedes its popularization. Here, we proposed a deep-learning assisted, label-free hematology analysis technique through a regular microscope to solve these limitations. We have demonstrated that the subcellular morphology of unstained leukocytes, erythrocytes, and platelets can be revealed by simply defocusing a regular microscope. We have shown that this technique can automatically differentiate five-part leukocytes with high precision (mean average precision achieves 98.0%) and convert the label-free blood images into virtual Giemsa images with the assistance of deep-learning algorithms. The Pearson coefficients for the leukocyte counts from this technique and the manual counting method are large than 0.9083. The proposed technique can not only enhance the clinical hematology analysis efficiency but also has great potential to be quickly popularized since the regular microscope is a broadly used imaging tool.