Discussion

In this proof-of-concept study, the capability of AIRFIHA for label-free classification of leukocyte subpopulations has been demonstrated on human blood donors. With a well-designed neural network model, high information-content quantitative phase images, and a considerable amount of data collected from human blood donors, our AIRFIHA method has outperformed current reagent-free methods for the classification of granulocytes, monocytes, and B and T lymphocytes. Our preliminary result also shows that the detection accuracy of our method is not severely affected by different donors, thus indicating a potential for use in clinical settings. We have further demonstrated that AIRFIHA can differentiate CD4 and CD8 cells that are normally difficult to distinguish with label-free methods.
Error Analysis. It is important to note that our classification results rely on the accuracy of the separation kits used in this study to select the individual sets of leukocytes. We employed flow cytometry (refer to the details in “Methods”) to measure the percentage population of the specific leukocytes after isolating them using the corresponding kits and the representative results from a donor are presented in Supplementary Figure S7. These negative isolation kits have inherent inaccuracy that can adversely affect the classification results. However, compared with positive selection kits, negative selection kits could better maintain the original cell morphology for our label-free imaging modality, where the morphological attributes form the basis for classification.
Result Evaluation. We compared our result with other reported results using different detection/imaging principles, labeling methods, and experiment instruments, as shown in Table S12 in Supplementary Material. AIRFIHA has a significantly improved accuracy when compared with the methods based on negative isolated leukocyte classification\cite{RN53}. For the classification of monocytes, granulocytes, and lymphocytes, our detection accuracy is slightly lower than the methods using positive fluorescence sorting or complicated purification methods\cite{RN54,RN36,RN55}. It is possible that the negative selection kits have intrinsic lower accuracies in isolating leukocytes when compared with using positive kits, therefore reducing our classification accuracy. If there is a way to sort the leukocytes with higher accuracies without affecting the original morphology states of cells, we expect to further increase the classification accuracy. For the classification of B and T lymphocytes, our result is better than bright and dark field microscopy-based methods for the cross-donor validation experiments\cite{RN36}. To a certain extent, our method benefits from the subtle differences in the refractive index maps of intracellular structure as encoded in the quantitative phase maps. Our classification accuracy is also comparable with 3D QPM based methods that explore expensive and complex instrumentations (note that no human blood test and cross-donor validation have been carried in such methods so far)\cite{RN38}. Notably, both mentioned methods are based on using positive leukocyte extraction methods. As for the classification of CD4 and CD8 cells, our classification accuracy is also compared with that obtained using 3D QPM methods\cite{RN38}.
Further Improvement. With the capability to differentiate very complex leukocyte types, AIRFIHA can provide more comprehensive information for potential disease diagnoses with simplified testing procedures. There are still ways to improve the detection accuracy of our system, such as improving the phase imaging resolution through synthetic aperture phase imaging method\cite{RN56}, deconvolution\cite{RN57}, and using 3D-resolved phase maps, preferably captured through a single image acquisition to avoid taking a large amount of data (such method has been recently made possible; a manuscript is under preparation)\cite{RN71}. The other way to improve accuracy is to expand the dataset and upgrade the neural network model.
Potential Applications. Overall, our results show the potential of AIRFIHA as a fully automated, reagent-free, and high-throughput modality for differential diagnosis of leukocytes at point-of-care and in a clinical laboratory. Additional salient features of this platform include its single-shot measurement, small spatial footprint, and low cost. Of note, owing to its facile and simpler set-up, this platform can be combined with other modalities for blood cell investigation. For example, by combining it with microfluidic devices, AIRFIHA can conduct blood testing and analysis in a fully automated way. Importantly, the need for isolation kits is obviated and the leucocytes separated from blood using a routine centrifugation process can be directly subjected to the AIRFIHA to provide percentage population of leukocyte subtypes. One other example could be its integration with Raman spectroscopy that has been proposed for B lymphocytes acute lymphoblastic leukemia identification and classification\cite{RN4}. While Raman spectroscopy provides biomolecular specificity, spontaneous Raman measurements are not feasible for clinical workflow requiring rapid diagnosis. Importantly, given the potential of the AIRFIHA platform in screening the B cells from other leucocytes, this QPM-based strategy can be used to screen the B lymphocytes where Raman measurements can be performed for B lymphocytes leukemia diagnosis. The combined QPM-Raman system obviates the need of any additional separation method to select B lymphocytes either from the blood or from the leucocyte mixtures for leukemia diagnosis in a label-free manner. Moreover, as AIRFIHA involves a low-cost system that requires minimal sample preparation or chemical consumables, our AIRFIHA has a great potential to be used in point-of-care applications, resource-limited settings, or pandemic situations, e.g., COVID-19 pandemic, in view of a portable and low-cost QPM system recently demonstrated by us\cite{RN58}.