The prediction from the ResNets is based on the feature vectors which are placed at the end of convolutional layers. The model should produce similar feature vectors for the same input types and very different feature vectors for different input types. To verify the efficacy of our trained ResNets, the t-distributed stochastic neighbor embedding (t-SNE) method\cite{RN48} was used, which has decreased the feature dimension from 256 to 3 for all the cell types. The features are plotted in the same coordinate space as shown in Figure 4d, e. ResNet extracted features could help to distinguish B and T lymphocytes. Compared with principal component analysis (PCA) method, those features have made the boundaries between B and T lymphocytes clearer (refer to Figure S3 in Supplementary Material). As summarized in Table S7 in Supplementary Material, the ResNet method shows increased classification accuracy by over 15% for B and T lymphocytes compared with the PCA method, as evaluated with a trained support vector machine (SVM) classifier. To understand what differences in the cell morphological features contributed to the discrimination in the neural network, we visualized the outputs of each convolutional layer after activation (refer to Figure S4 in Supplementary Material). From the last layers of analyzed leukocytes, we observed that the output features are mainly focused on the nucleus, cytoplasm, and membrane characteristics, elucidating that these cellular features are important in classification. These observations are consistent with the knowledge that subtypes of leukocytes differ in their nuclear, cytoplasmic, and membrane attributes\cite{RN64,RN65}. To explore the cause of classification errors, some mistakenly classified cells are listed out. Apart from the morphological similarities between different cell types, the error could be also caused by the mislabeling in the ground truth dataset (refer to Figure S5 in Supplementary Material). To determine the optimum numbers of cells for achieving a stable and high detection accuracy, we have analyzed the detection accuracy vs. total number of cells used for training (refer to Figure S6 in Supplementary Material). It is found that the detection accuracy becomes stable when the training dataset is > 800 cells (i.e., > 200 cells per type).  In this study, we have collected a total of ~ 2,700 cells with > 600 cells in each type for training, which was appropriate for achieving a reliable classification result.
CD4 and CD8 cells are subtypes of T lymphocytes and have very similar morphological features\cite{RN38}. Routine monitoring of CD4/CD8 cell ratio with point-of-care systems helps monitor immunodeficiency related diseases, e.g. acquired immunodeficiency syndrome (AIDS)\cite{RN49,RN50}. Our proposed AI-powered platform has the potential to offer a unique approach in which the T cells can be virtually isolated and subtyped while also preserving them for subsequent immunophenotypic analysis.  Moreover, such a platform can be expanded to visualize the immunological responses due to its label-free attributes. We had previously demonstrated the use of QPM in identifying the activation state of CD8 cells in a contrast-free manner\cite{RN23}. Building up on our previous study, we conjectured that our QPM can be used for differentiating CD4 and CD8 cells in a label-free manner. To test our hypothesis, we employed our AIRFIHA system on CD4 and CD8 cells from the same blood donor for both training and testing. The classification result is summarized in Figure 4f-h. F1-scores of 80.4% and 77.5% for CD4 and CD8 cells are achieved, respectively (detailed values for recall, precision, and F1-scores are provided in Table S6).  Compared with the F1-scores of 85.7% and 88.8% for CD4 and CD8 cells obtained by using 3D refractive maps\cite{RN38}, our preliminary results have a bit lower accuracy. The AUPRC values for CD4 and CD8 cells are 0.78 and 0.84, respectively. Using the t-SNE method, features are extracted from the CD4-CD8 classifier and plotted (Figure 4h) for visualizing the differentiation capability. Our preliminary results show that our method has a basic differentiation capability for these two subtypes of T lymphocytes. The accuracy can be increased by using high volume of data and further tuning of our neural network.