2.1.

System Overview and Image Acquisition

The imaging system used in these experiments was designed as a POC fluorescence-based system with low-cost components, such as the microcomputer and LED excitation source, as shown in Fig. 2. This system will be referred to as the POC system in this paper. The Raspberry Pi 3 Model B microcomputer ($40, Adafruit) was connected to an epi-fluorescence microscope consisting of a 455-nm LED illumination (Philips), a $4\times$ Nikon Plan Fluor objective ($\mathrm{NA}=0.13$, Nikon Instruments Inc.), a dichroic mirror (475-nm cutoff wavelength, Chroma Technologies), a 500-nm longpass emission filter (Thorlabs), a achromatic doublet (100-mm focal length, Thorlabs), and a USB 2.0 Chameleon color camera (CMLN-13S2C-CS, Point Grey). All images were acquired with an optical magnification of $\sim 5.81\times$, a field of view (FOV) of $1.89\times 1.42\mathrm{mm}$, an exposure time of 150 ms, and a gain of 0 dB. For each image set, six nonoverlapping FOVs were acquired within 1 min intervals.

Fig. 2

The imaging system designed to mimic a POC device. It consists of a Raspberry Pi microcomputer connected to an epi-fluorescence microscope with a $4\times$ Nikon Plan Fluor objective, a 500-nm long pass filter, a 475-nm dichroic mirror and filter set, a 455-nm LED, and a Point Grey Chameleon color camera.

2.2.

Blood and Acridine Orange Preparation

To study the effect of AO staining concentration, we performed a human subject trial in cooperation with the University of Arkansas Pat Walker Health Center. Following informed consent, blood samples were collected from 30 healthy subjects under the approval of the University of Arkansas Institutional Review Board (IRB 15-11-384). The samples were drawn via venipuncture (4-ml whole blood) and were analyzed at least 20 min later on a clinical hematology analyzer (XS 1000i, Sysmex) after gently inverting 10 times. An AO stock solution was prepared by mixing distilled water with AO hemi(zinc chloride) salt (Sigma-Aldrich) for a concentration of $100\mu \mathrm{g}/\mathrm{ml}$; it was maintained at a pH of 7.4, stored at 4°C, and protected from light. This stock solution was diluted into nine concentrations ranging from 10 to $90\mu \mathrm{g}/\mathrm{ml}$. Whole blood was stained at a $1:1$ ratio with the 10 concentrations including the stock concentration producing final concentrations ranging from 5 to $50\mu \mathrm{g}/\mathrm{ml}$. Image acquisition using the POC System was performed within 4 h of the blood draw at room temperature or 24 h if the samples were refrigerated.

2.3.

Automated Differential Algorithm

We have developed an image analysis-based automated differential algorithm, shown in Fig. 3, to calculate the three-part leukocyte differential based on the red to green fluorescence intensity ratio (RG ratio). Following image acquisition, the individual cells were segmented by applying a global threshold to convert the image to a binary image that was applied as a mask. Any object with an area less than 15 pixels and greater than 100 pixels were considered platelets, clusters of cells, and debris, and were therefore excluded. The RG ratio was calculated for each cell by dividing the mean red intensity by the mean green intensity. These intensity values were acquired from the red and green grayscale images separated from the RGB images. Those values were plotted on a histogram displaying the frequency of cells at a certain RG ratio value. The three groups of peaks in the histogram [Fig. 3(d)] indicating the three cell populations were used to define cutoff ranges consisting of the lymphocyte/monocyte (LM) and monocyte/granulocyte (MG) cutoffs. We chose a static approach to define fixed cutoff ranges to avoid issues with variations in the peaks [Fig. 3(c)] among samples. The initial cutoff values (0.42 and 0.59) were based on visual inspection of the groups of peaks from the average of the histogram in Fig. 3(c) and four other histograms. From those ranges, the three-part differential was calculated by automatically counting the number of cells within each cutoff section.

Fig. 3

An overview of the image analysis-based automated differential algorithm. (a) Each cell per image was segmented and measured for mean red and green fluorescence intensity to (b) calculate the red-to-green (RG) ratio. (c) The RG ratios were plotted on a histogram showing the frequency of each ratio value. (d) RG ratio cutoff range (bold lines) was used to calculate the final three-part leukocyte differential results.

2.4.

Investigation of Three Variables in the Overall Staining and Postprocessing Method

Three variables within the overall staining and postprocessing method were analyzed to determine their effect on the differential results (Fig. 4). This included concentration (variable 1) and time gates representing a sample of incubation periods (variable 2) within the staining method and post hoc analysis of the RG ratio cutoff values (variable 3).

Fig. 4

The analysis method consisting of three variables: (1) concentration, (2) time gates, and (3) RG ratio cutoffs. During data collection, images were acquired at 10 different concentrations for each time gate. In post hoc analysis, a fixed set of time gate and concentration data was used to determine optimal RG ratio cutoff values when the range was varied.

To evaluate the staining method, 10 AO concentrations varying from 5 to $50\mu \mathrm{g}/\mathrm{ml}$ (final) and incubation periods separated into three time gates (3 to 4, 5 to 6, and 7 to 8 min) were used. The 10 AO concentrations were selected based on other existing AO staining methods.^17,33,37 The time gates were determined based on the need to acquire a large number of cells while minimizing the time-dependent redshift in AO emission. The optimal AO concentration and time gate were analyzed simultaneously with percent deviation from the clinical hematology analyzer as a measure of accuracy. The images were acquired using the three time gates for each of the 10 concentrations, resulting in 30 different image sets per specimen.

For post hoc analysis of the postprocessing method, the RG ratio cutoffs between the LM and MG were varied. Image data acquired using time gate 5 to 6 min and $10\mu \mathrm{g}/\mathrm{ml}$ from all samples were evaluated with the different cutoff values. Initially, the cutoff range was manually selected via visual inspection of histograms with very distinct peaks separating the three populations. However, since this process could introduce observer bias, a range of cutoff values near these were evaluated. Five values from each cutoff (LM $0.42\pm 0.2$, MG $0.59\pm 0.2$, in increments of 0.1) were applied to those image sets comprising 25 different results per specimen. The results were analyzed for percent deviation from the clinical hematology analyzer as a measure of accuracy.

2.5.

Analysis of the Combined Optimal Staining and Postprocessing Methods

All 30 samples measured in triplicates (for repeatability of the system) were used to test the accuracy of the combined optimal staining and postprocessing methods based on the results from the previous section. Three time gates from 3 to 4, 5 to 6, and 7 to 8 min were used to demonstrate the redshift that occurs over time and how it affects the test results. Two metrics were used in the analysis: Bland–Altman and correlation. For the Bland–Altman analysis, the mean difference between the results produced by the POC system and the clinical hematology analyzer was used to calculate the average bias between the systems. To determine the correlation between the systems, all values were plotted and a linear regression was drawn to produce an $R$-squared value with a value closer to one indicative of a closer relation to the clinical value. All three time gates were compared using this same analysis approach.

3.1.

Acridine Orange Concentration on Cell Count Accuracy

AO concentration effects on overall leukocyte differential accuracy are summarized in Fig. 5. The green, blue, and red bars represent the deviation from the clinical standard measurement for each cell population (lymphocytes, monocytes, and granulocytes, respectively). The time gates were included since it is a dependent variable in the AO staining method. The monocyte population demonstrated the greatest inaccuracy at concentrations above $10\mu \mathrm{g}/\mathrm{ml}$ where the deviations exceeded the CLIA-defined limits of $\pm 15\%$, and many exceeded 100%. Similarly, lymphocytes had the greatest inaccuracy at concentrations above $15\mu \mathrm{g}/\mathrm{ml}$; however, the deviations were negative, indicating that the POC system was underestimating the prevalence of this cell population. Overall, a concentration of $10\mu \mathrm{g}/\mathrm{ml}$ and time gate 5 to 6 min, denoted by asterisks in Fig. 5, demonstrated the lowest deviations ($\text{lymphocytes}=-2.47\%$, $\text{monocytes}=9.87\%$, and $\text{granulocytes}=-0.192\%$) from the clinical hematology analyzer.

Fig. 5

A comparison of different concentrations (columns) and time gates (row) using percent deviation as a measure of accuracy. The green, blue, and red plots represent the lymphocytes, monocytes, and granulocytes, respectively. A concentration of $10\mu \mathrm{g}/\mathrm{ml}$ and time gate between 5 and 6 min demonstrated the lowest deviations ($\text{lymphocytes}=-2.47\%$, $\text{monocytes}=9.87\%$, and $\text{granulocytes}=-0.192\%$), denoted by an asterisk.

In Fig. 6, a histogram plot demonstrates the change in the leukocyte frequency peaks when the concentration was increased and the time gate was maintained at 5 to 6 min. In lower concentrations, the peaks were in locations that match the fluorescence characteristics of each cell population, consistent with previous results at $10\mu \mathrm{g}/\mathrm{ml}$. With increased concentrations, the largest peak shifted to the right (redshifts) and expanded where there was less distinction between the groups.

Fig. 6

Histograms demonstrating the change in the peaks when concentrations were increased and the time gate was maintained at 5 to 6 min. The dotted line represents a concentration of $10\mu \mathrm{g}/\mathrm{ml}$, where there are peaks in expected locations ($n=754$ cells). The dashed line represents the highest concentration of $50\mu \mathrm{g}/\mathrm{ml}$, where the peaks are less distinct and the first peak with a greater bandwidth ($n=1457$ cells).

3.2.

RG Ratio Cutoff Effects on Cell Count Accuracy

The RG ratio cutoff values used in the differential algorithm to count the cell populations were evaluated after observing the effects from variables 1 (concentration) and 2 (time gates). In Fig. 7, similar to the previous section, green, blue, and red bars indicate the three cell populations. For granulocytes and lymphocytes, all cutoff values were between $-3.7\%$ to 2.05% and $-11.01\%$ to 4.14%, respectively. Monocytes, as anticipated, demonstrated the greatest variability with half of the tested RG cutoff values producing deviations greater than $\pm 15\%$ compared to the clinical results. Additionally, there were observable trends in deviation values for the lymphocytes and granulocytes, relating to the cutoff value (LM or MG) being studied. For lymphocytes, the LM cutoff directly affected the count by accepting more or less cells between the first two peaks on the histogram [Fig. 3(d)]; therefore, there was greater plot-to-plot variation in accuracy. This result was similar for the granulocyte count except that the trend varied within each plot. The cutoff values of 0.42 (LM) and 0.58 (MG), denoted by asterisks in Fig. 7, produced the lowest deviations from the clinical results ($\text{lymphocytes}=0.706\%$, $\text{monocytes}=-1.534\%$, and $\text{granulocytes}=-0.645\%$). This demonstrates the importance of controlling the staining process since slight variations in the RG ratio cutoff values affected the cell count accuracy.

Fig. 7

A comparison of different cutoff values between LM (0.40 and 0.44) and MG (0.57 and 0.61) using percent deviation as a measure of accuracy using data from the optimal staining method. The cutoff values of 0.42 (LM) and 0.58 (MG) produced the lowest deviations from the clinical hematology analyzer ($\text{lymphocytes}=0.706\%$, $\text{monocytes}=-1.534\%$, and $\text{granulocytes}=-0.645\%$), denoted by asterisks.

3.3.

Time Gate Effects on Cell Count Accuracy

From the previous results in Sec. 3.2, the optimal staining and postprocessing method included a final concentration of $10\mu \mathrm{g}/\mathrm{ml}$ and the RG ratio cutoff values of 0.42 (LM) and 0.58 (MG). Using these constraints, Bland–Altman analysis and linear correlation were used to compare our POC system to the clinical hematology analyzer for the three time gates at 3 to 4, 5 to 6, and 7 to 8 min. In Fig. 8 (left), the Bland–Altman plots demonstrate the agreement between the two systems with the solid lines representing the mean difference (bias) and the dashed lines representing the 95% confidence intervals of the differences (limits of agreement) calculated by the mean $\pm 1.96$ times the standard deviation.^6,21,26 A bias at zero represents perfect agreement between the systems, and the closer the limits of agreement are to this bias, the less likely our POC system will produce different results from the clinical hematology analyzer results that could affect the diagnosis. For example, data points above the mean represent an overestimate of that population, which could lead to a potential false-positive measurement. The linear correlation between the two systems is shown in Fig. 8 (right). Deviation from the dotted line indicates poor agreement with the clinical hematology analyzer. Using these metrics, the three time gate plots demonstrated the trend seen previously in Fig. 5 where the deviation between the systems increased between time gates 5 to 6 and 7 to 8 min. As the time gates increased, the lymphocyte count was underestimated while the granulocyte count was overestimated, demonstrated in the Bland–Altman plots (Fig. 8, left). The time gate of 7 to 8 min had the most deviation, resulting in a bias of 11.30 with a 95% limits of agreement of $-8.29$ and 30.89 for the granulocyte population and an $R^2$ of 0.758. Time gates 3 to 4 and 5 to 6 min produced similar results, yet time gate 5 to 6 min yielded results that were closest to the actual clinical results ($\text{bias}=0.222$ and limits of agreement of $-6.02$ and 6.46 for the monocytes and $R^2=0.981$).

Fig. 8

Bland–Altman analysis and correlation plots for the time gates with an AO concentration of $10\mu \mathrm{g}/\mathrm{ml}$ and cutoff values of 0.42 (LM) and 0.58 (MG). Time gate 3 to 4 and 5 to 6 min had similar linear correlation with $R^2$ values above 0.960. The Bland–Altman analysis for time gate 5 to 6 min resulted in the monocytes having a bias of 0.222 and limits of agreement of $-6.02$ and 6.46, and lymphocytes and granulocytes producing similar results. Time gate 7 to 8 min had to greatest inaccuracy with an $R^2$ value of 0.750 and the granulocytes with the greatest bias of 11.30 and limits of agreement of $-8.29$ and 30.89.

4.1.

Implications of AO Staining Protocol

We have demonstrated differences in accuracy when AO concentration and incubation periods (represented by three time gates) are varied. This suggests that a strict staining method is required to produce accurate and consistent results. In this case, using $10\mu \mathrm{g}/\mathrm{ml}$ AO and a time gate of 5 to 6 min with fixed cutoffs at 0.42 (LM) and 0.58 (MG) resulted in average deviations of 0.706%, $-1.534\%$, and $-0.645\%$ for lymphocytes, monocytes, and granulocytes, respectively. When the AO concentration and incubation periods were increased a shift in red emission occurred, which correlated to previous publications.³³ This can be attributed to the accumulation of AO, a cationic dye, in acidic vesicles where the dye becomes “acid-trapped” when it is protonated in a lower pH environment.³⁶ The most important inference from this comparison of multiple staining methods is that a drastic redshift occurs over time that reduces the differential count accuracy and, thus, reduces the effectiveness of the dye in distinguishing the cell populations.

Another potential source of this redshift is due to leukocyte proliferation, in which an increase in DNA/RNA synthesis can be observed in applications using AO to quantify changes in cell cycle phase.^34,39 For normal leukocytes in peripheral blood, cells are stable in the $G_0$ phase and require activation in response to an antigen/mutagen to proceed in the cell cycle.⁴⁰ During the next phases of the cell cycle (${\mathrm{G}}_1$ to S phase), there is an increase in DNA and RNA syntheses that occurs over the course of many hours (6+ h).^41–43 The increase in RNA has been shown to increase the red fluorescence emission intensity of AO-stained leukocytes, whereas the increase in DNA content contributed less to the increase in green fluorescence intensity.⁴³ In patients with hematologic malignancies, it may be possible to find some leukocytes in the peripheral blood that exist in a proliferative phase; this may affect the established RG ratio. This could lead to an inaccurate classification, but not necessarily an inaccurate total leukocyte count. Within the environment and time period of the analysis presented in this article, the cells should be in a quiescent state; therefore, the cell cycle should have minimal effect on the red and green fluorescence emissions.

4.2.

Implications of Variations in Postprocessing RG Ratio Cutoff Values

We used a static approach to separate the cell populations based on the frequency of cells at each RG ratio, where fixed cutoff values were defined by the minimum troughs between the peaks in the histogram and were used for every sample tested; see Fig. 3(d). Our results suggest that, using this approach, slight changes in the cutoffs can change the cell count, which could negatively affect the differential results since we have shown that the peaks can change over time due to the redshift in fluorescence intensities with increased AO concentration (Fig. 6) and incubation periods. An alternative method could potentially involve the use of a dynamic approach where the cutoff values are automatically adjusted based on the shifting peaks that can occur with different staining methods by tracking the troughs or maximum peaks of each cell population for each sample tested. Different factors could affect the definition of the histogram curve, such as cell count and bandwidth of the peaks; therefore, further exploration of that method is necessary.

4.3.

Effect of Time Gating Using an Optimal Differential Method

Similar to the effects of AO concentration, different incubation periods, demonstrated by three time gates between 3 and 8 min, can produce different results, as shown in Fig. 8. Our results were comparable to the redshift in AO emission over time and accumulation of AO previously reported in the literature.^33,36 Not only was there a redshift within increased incubation periods but the shift did not appear to be uniform (Fig. 6); therefore, controlling accuracy became more difficult when the staining and postprocessing method was not consistently maintained. Also, the Bland–Altman and correlation plots demonstrated that there was a drastic change in test results at the 7 to 8 min time gate; this could be due to a sudden uptake of AO between 6 and 7 min that caused a greater redshift. This suggests that there is a limit to the incubation period where the AO emissions redshift causes all of the cells to become indistinguishable. Additionally, the difference between 3 to 4 and 5 to 6 min time gates were minimal, yet time gate 5 to 6 min had more test results consistent with the clinical results. Thus, there could still be a difference in results between lower time gates, further demonstrating the need for an optimal staining and postprocessing method.

As current state-of-the-art POC systems rapidly evolve, this information could aid future designs in improving the accuracy and precision of these differential tests. By understanding and quantifying the variations in the red emission, the colorimetric features of AO that are used to distinguish subpopulations of cells for these devices can be utilized more effectively. In general, the method described here is specific to our image-based POC system and AO colorimetric classification; however, this could be extrapolated to other red to green colorimetric POC tests, for instance, image-based methods implementing color cameras and nonimage-based methods such as POC flow cytometry methods implementing microfluidics and fluorescence signal detection using photomultiplier tubes.^7,13 Additionally, future work could incorporate a five-part differential test to explore whether it can be accurately quantified using AO.