Method

Fresh blood sample procurement. The fresh blood samples from six anonymous healthy adult donors were purchased from StemCell Technologies (Vancouver, Canada) and all the experiments were conducted within 24 hours of blood extraction. The purchased blood samples contained ethylenediaminetetraacetic acid (EDTA) as the anti-coagulant.
Leukocyte isolation from fresh blood. Four types of leukocytes, namely monocytes, granulocytes, and B and T lymphocytes, were isolated from fresh blood samples using isolation kits from Stemcell Technologies. From each donor the amount of blood was in the 1-3 ml range, depending on the minimum volume requirement as per manufacturer's instruction for each leukocyte subpopulation. To isolate these four subpopulations, we used EasySep Direct Human Monocyte Isolation Kit, EasySep Direct Human Pan-Granulocyte Isolation Kit, EasySep Direct Human T Cell Isolation Kit, and EasySep Direct Human B Cell Isolation Kit (Stemcell Technologies Inc). These separation kits used immunomagnetic negative selection for isolating each specific leukocyte type from the whole blood sample. Two additional negative separation kits, i.e., EasySep Direct Human CD4+ T Cell Isolation Kit and EasySep Direct Human CD8+ T Cell Isolation Kit, were used for the isolation of CD4 and CD8 cells, respectively. Phosphate-buffered saline free from Ca++ and Mg++ (Gibco, Thermo Fisher Scientific) was used as the recommended medium for the EasySep Isolation kits. The isolation was carried out following the manufacturer’s instructions with multiple cycles of mixing and incubation with the provided RapidSpheres and cocktail from the isolation kits. The final incubation yielded the isolated leukocytes in a 14 ml polystyrene round-bottom tube (Thermo Fischer Scientific), which were centrifuged at 400g for 5 minutes. The cell pellet was resuspended in PBS before the cells were imaged. 
Flow cytometry analysis. Flow cytometry was performed on the isolated leukocytes after the EasySep procedure to confirm the purity of the isolation. The viability of the leukocytes was checked with Acridine Orange and Propidium Iodide (AO/PI) staining (Invitrogen, Thermo Fischer Scientific) using a cell counter. The isolated leukocytes were counted and 50,000 of them were resuspended in cold PBS (Gibco, Thermo Fisher Scientific) at a density of 107/ml. 100 ml of this cell suspension was added to each well in a 96 well plate. 1 µl of the required fluorophore-conjugated antibody was added to each well and incubated in the refrigerator for 20 mins. Anti-CD-14- PerCP was used for monocytes, Anti-CD-66b-FITC was used for granulocytes, Anti-CD-19- APC was used for B lymphocytes, and Anti-CD3- PE was used for T lymphocytes.  The leukocytes were washed thrice with cold PBS and resuspended in 100 µl of cold PBS. The leukocytes were used for the flow cytometry analysis (MACSQuant Analyzer) and the data were analyzed with FlowJo software.
Leukocyte sample preparation for quantitative phase imaging. After the isolation of the leukocytes, we suspended them in PBS solution and diluted five-ten times. DNase solution (1mg/ml) (Stemcell Technologies Inc) was added to the isolated cells to decrease the clumping and adsorption of protein fragments. Typically, 10 µl of the isolated cell suspension was sandwiched between two quartz coverslips and a secure seal spacer. Then, the sample was placed onto the sample-stage of the home-built system for quantitative phase imaging. We repeated this sample preparation procedure for collecting all the required phase images of leukocytes from each donor.
Training of the classification model. Phase maps of the leukocytes were obtained by cropping the phase images retrieved from the measured interferograms. Each phase map, containing one leukocyte, was then resized to 300х300 pixels to be used as the input of the network. In the training process, a 5-fold cross-validation method was used to tune the hyperparameters, including network depth, batch size, learning rate, etc. During the training, to ensure all leukocyte types were trained under the same condition (i.e., each type has the same number of training samples), the datasets of unbalanced leukocyte types were augmented by rotation, position shifting, and flipping. For the monocyte-granulocyte-lymphocyte classifier, B and T lymphocytes were treated as one type, i.e., lymphocytes, and then all granulocytes, monocytes and lymphocytes were used to train and test the classifier. Categorical cross-entropy loss and Adam optimizer (learning rate = 1 * 10-3, β= 0.9, β= 0.999, learning rate decay = 0)\cite{RN59} were applied to optimize the model. In the end, the model with the best average validation accuracy was chosen as the final monocyte-granulocyte-lymphocyte classifier. For the B-T lymphocyte classifier, the dense layer of the obtained monocyte-granulocyte-lymphocyte classifier was first replaced with a new dense layer that has two outputs. All the B and T lymphocytes were used to fine-tune the entire network. Categorical cross-entropy loss and SGD optimizer (learning rate = 1 * 10-3 , learning rate decay = 1 * 10-6 , momentum = 0.9)\cite{RN60} were used. The network model with the best validation result was chosen as the final B-T lymphocyte classifier. By connecting these two network models, the final cascaded network model was obtained, from which the testing was conducted. The CD4-CD8 classifier was fine-tuned from the B-T lymphocyte classifier and trained and tested within the same donor. These frameworks were implemented with Tensorflow backend Keras framework and Python in the Microsoft Windows 10 operating system. The training was performed on a computer workstation, configured with an Intel i9-7900X CPU, 128 GB of RAM, and an Nvidia Titan XP GPU.