As for real clinical applications, the blood test samples normally come from new individuals whose blood samples will not be known by our model. There could be variances in the morphological features of leukocytes of each type between different donors, depending on their age, health status, etc\cite{RN51,RN52}. To verify whether such variances exist among our donors, we plotted the area and dry mass distributions for each donor (Figure 5a,b), from which it was found out that there were indeed distribution differences between donors for certain leukocyte types (note that for several donors the distributions for certain types of leukocytes were missing). Since we had already acquired QPM images as a part of another work involving B cell leukemia, we decided not to measure the B cells from all the donors. This partly helped in ensuring that the extraction and subsequent QPM measurement from all the other leukocytes were completed within 3-4 hours of receiving the samples. In any case, we have more than 3 different donors for every leukocyte sample and we believe that it is sufficient for this proof-of-the-concept investigation. The effect of such differences on the generalization of our model to the new donor was explored. For this purpose, the leukocyte samples from five donors were used for training (>200 cells per type) and the leukocyte samples from the remaining donor were used for testing. This experiment was repeated by rotating the testing donor in the same group. The classification result is plotted in Figure 5c (detailed numerical values are provided in Table S8-S10). The result obtained earlier using all six donors (refer to Figure 4a) is set as the reference for comparison. As for monocyte and granulocyte, all variances of the 10 tests are less than 0.1. For B lymphocytes only 1 out of 3 tests has variances larger than 0.1, while for T lymphocytes only 1 out of 5 tests has a variance larger than 0.1. The over-all average F1-scores for monocytes, granulocytes, and B and T lymphocytes are 91.6%, 94.3%, 83.7%, and 81.4%, respectively. Notable, the average F1-scores for the B-T lymphocyte classifier of the cascaded-ResNet are 84.1% and 85.9%. This has demonstrated a better accuracy when compared with an earlier work on the classification of B and T lymphocytes, where average F1-scores of 79.4% and 75.7% were obtained, respectively, using bright-field and dark-field imaging cytometry systems\cite{RN36}. To further understand the feasibility of our method, we performed an intra-donor analysis to reveal the variations among leukocytes within the same donor (refer to the F1-scores summarized in Table S11 in Supplementary Material). The results show that there are variations within the same donor, albeit small on average.The cross-donor and intra-donor validation results have shown that our method has a high potential for clinical applications.