Introduction

Leukocytes play an important role in maintaining the normal function of human immune systems and differ in structure and function\cite{RN64,RN65}. For instance, B and T lymphocytes can produce antibodies to defend the body against foreign substances, such as bacteria and viruses. Abnormal leukocyte differential counts are indications of malfunctions of the immune system or infectious diseases\cite{RN1}. For instance, a sharp increase in neutrophil-to-lymphocyte ratio serves as an independent risk factor for SARS-CoV-2 infection\cite{RN62,RN63}. To differentiate basic leukocyte types, volume and granularity parameters are often measured through electrical impedance and light scattering-based cytometry techniques\cite{RN5,RN66}. For more complex leukocyte types with similar morphologies (e.g., B and T lymphocytes), fluorescent molecules bound with antibodies that specifically target the proteins expressed on the surface are typically used to activate fluorescence emission which can be captured by detectors for population counting. Although antibody labeling based flow cytometry methods are widely used in the clinical laboratories, there remain a few drawbacks. Firstly, only the labeled cellular structures are used for differentiating cell types, but not all the cellular structures. Secondly, an extensive list of expensive reagents is required for differentiating many cell types. Lastly, the labeled cells are not suitable for further testing as their original states have changed.
Label-free imaging methods can potentially solve the aforementioned issues\cite{RN7,RN8,RN9,RN3,RN10}. For instance, a hemogram based on Raman imaging has been proposed to discern leukocytes\cite{RN11}. While this innovative approach leverages the unique biochemical attributes for the classification, it is limited by the weak spontaneous Raman signal, thus not suitable for high-throughput applications in a clinical setting. Quantitative phase microscopy (QPM) is a rapidly emerging imaging modality that is suitable for high-speed imaging of unlabeled specimens. In QPM, the exact optical path-length delay associated with the density and thickness at each point in the specimen is mapped, which has enabled label-free imaging of transparent structures (e.g., live cells) with a high imaging contrast\cite{RN12,RN13,RN14}. In recent years, QPM has been used for single-cell analysis by extracting quantitative biomarkers, e.g., measuring cell dry mass to quantify cell growth\cite{RN15,RN16}, studying red blood cell rheology\cite{RN17,RN18}, characterizing cell viability\cite{RN19}, analyzing large cell population\cite{RN20,RN21}, and screening cancer\cite{RN22}, etc. However, most studies have primarily relied on interpreting the QPM results in terms of a few principal morphological characteristics. Recently, several laboratories have sought to shift the paradigm by utilizing machine learning (ML) and artificial intelligence (AI) for analyzing and interpreting QPM data\cite{RN24,RN25,RN74}. As for the applications on hematology, QPM with ML/AL has been used to screen cancer cells in blood\cite{RN73,RN76,RN77}, diagnose red blood cell related disorders\cite{RN75,RN78}, detect activation states of leukocytes\cite{RN23}, and classify various leukocytes\cite{RN70,RN38,RN54,RN53,RN79}. The full field and fast imaging attributes of QPM enable availability of volumes of high-dimension imaging and therefore make QPM a unique modality for the application of ML/AI to those tasks involving cell classification and imaging.
With recent developments in ML/AI, e.g., visual geometry group (VGG)\cite{RN26}, inception\cite{RN27}, and residual neural network (ResNet)\cite{RN28,RN29}, abundant training data is available to train a model to extract important image features to classify targeted objects\cite{RN30,RN31}. Compared with previous manual feature extraction analysis methods, the new approaches in ML/AI may offer features with statistically significant higher sensitivity and specificity. Among the recent ML/AI methods, ResNet tackles the gradient vanishing problem by creating shortcut paths to jump over layers. Conversion among different types of biomedical images and the segmentation of certain cell structures have been achieved by using ResNet building blocks\cite{RN32,RN33,RN34}. With such exciting developments, ML/AI have also been applied to label-free imaging systems to tackle complicated cell analysis problems. For instance, machine learning for the differentiation of lymphocytes has been achieved on a bright-field and dark-field microscopy platform\cite{RN36} and a QPM platform using fixed pathology slides\cite{RN70}. To further improve the detection accuracy and specificity of leukocyte subtypes, the method using 3D QPM technique has been proposed and demonstrated\cite{RN24,RN37}.
In this work, we propose a rapid, low-cost, AI-enabled reagent-free imaging hematology analyzer (AIRFIHA) that can classify hierarchical leukocyte subtypes in human blood samples. Note that preliminary results of this study were presented at a conference in April 2019\cite{RN69}. AIRFIHA is based on leveraging the morphological attributes of phase images from a custom-built QPM system and a cascaded-ResNet for leukocyte classification. From this proof-of-principle study on six human donors, we have achieved a classification accuracy of 90.5% on average for monocytes, granulocytes, and B and T lymphocytes. The robustness and applicability of our proposed method have been confirmed by conducting cross-donor validation experiments. We further investigated the potential of AIRFIHA in discerning human CD4 and CD8 T cells. AIRFIHA demonstrated a much higher accuracy when compared with methods based on negative isolated leukocyte classification and a comparable or better accuracy when compared with methods based on positive isolated leukocyte classification. This study shows a promising perspective when applying AIRFIHA for automated clinical blood testing applications, which is especially useful in resource-limited settings and during pandemic situations.  

Results

AIRFIHA system

In this work, the classification of human leukocyte types is achieved using a QPM system and a neural network, as conceptually illustrated in Figure 1. The exact configuration of the QPM system is based on a diffraction phase microscope (DPM)\cite{RN39,RN2,RN40}, which can provide highly stable and accurate phase imaging of cells. The imaging resolution of the QPM system is 590 nm, while the field of view is around 61 μm x 49 μm. Compared with optical diffraction tomography\cite{RN38,RN37}, QPM does not necessitate a complex imaging system and expensive computation requiring a large amount of data, and the system is relatively cost-effective with a smaller footprint. To obtain the dataset for neural network training and classification, the leukocyte samples were isolated from the fresh blood samples of six healthy donors within 24 hours of blood extraction. The blood sample used for the leukocyte separation for each donor was in 1-3 ml range, depending on the minimum volume requirement as per manufacturer's instruction for the leukocyte subpopulations. To minimize the influence of labeling on cell activity, the leukocytes were negatively isolated by using antibody-labeled magnetic particles as illustrated in Figure 1 (a) (refer to detailed sorting procedure in “Methods”). Then, the isolated sample was diluted in PBS (phosphate buffer saline) and mounted between two glass coverslips before placing it onto a home-built QPM system as illustrated in Figure 1(b) (refer to the detailed sample preparation procedure in “Methods). Phase images of each leukocyte type were retrieved from the measured interferograms (refer to the detailed description of the QPM system and the phase retrieval method in Supplementary). After thousands of phase images of labeled leukocytes of different types were measured, all the leukocytes in each phase image were segmented to construct the training and testing dataset\cite{RN41}. A neural network was constructed, trained, and validated for classifying the leukocytes using the phase image dataset (Figure 1c). A detailed description of the neural network is provided in the following section. Finally, the AIRFIHA system was used to identify leukocyte types of new samples (Figure 1d).