1. Introduction

Leukemia is a common malignancy with a high mortality rate, which is predicted to rank the 8th for cancer incidence and the 6th for cancer-related deaths in 2017 [1]. Early diagnosis and personalized therapies are crucial for significant reductions in leukemia mortality, especially for children. As a heterogeneous group of diseases with very different clinical features, courses, treatment programs and prognoses, leukemia is generally classified into four subtypes, namely acute lymphocytic leukemia (ALL), chronic lymphocytic leukemia (CLL), acute myeloid leukemia (AML), and chronic myeloid leukemia (CML) [2]. The establishment of optimal therapeutic regimens greatly depends on the classification of leukemia subtypes, but the accurate classification of leukemia subtypes remains a challenge for clinicians.

Current methods for the classification of leukemia subtypes are expensive, time-consuming and require intensive laboratory studies involving expertise in histology, cytochemistry, immunophenotyping, cytogenetics and molecular biology. For example, flow cytometry, fluorescence in situ hybridization (FISH), polymerase chain reaction (PCR) and karyotype analysis have been recently adopted to reveal immunophenotypic features and karyotypic information for leukemia classification [3–7]. More recently, microarray technology and next-generation sequencing (NGS) that can analyze gene expression profiles (GEP) have served as new tools for molecular studies in leukemia diagnosis [8, 9]. While these methods have greatly improved the classification of leukemia subtypes, the development of novel technologies that are easy to operate and low-cost is of great interest, especially in developing countries due to the insufficient infrastructure and lack of expertise.

Light scattering may serve as a label-free and low-cost technique for the classification of leukemia subtypes. In 1975, Salzman et al. separated label-free human leukocyte cells into lymphocytes, monocytes and neutrophils based on the measurements of forward scattering (FSC) and side scattering (SSC) [10]. It has been recently reported that FSC and backward scattering (BSC) are dominated by cell size, and SSC is more associated with the cellular organelles [11]. The measurements of one dimensional (1D) scattered light provide detailed information for label-free cell analysis [11–16]. Recently, 2D light scattering measurements that capture the scattered light in both the polar and azimuthal angular ranges have been demonstrated [17–25]. Compared with 1D measurements, 2D light scattering contains richer micro- and nano-structural information of biological cells. Our group has recently explored the rich information of the 2D light scattering patterns by employing the machine learning algorithms for automatic label-free cell analysis [26]. In terms of label-free leukemia analysis, we reported a 2D light scattering static cytometer that can obtain scattered light in the angular range from 79 to 101 degrees [27, 28], which has been used to classify the leukemic cells (both acute and chronic myeloid leukemic cells) from normal granulocytes [28]. Light scattering spectroscopy that measures BSC has also been demonstrated for the label-free differentiation of leukemic cells and normal cells [29, 30]. However, the classification of acute leukemic cells from the chronic ones remains unsolved upon label-free light scattering measurements. This may be attributed to the limited angles that were detected in both our label-free static cytometry and the light scattering spectroscopy.

In conventional cytometry, the fluorescence or light scattering signals are measured in the FSC and SSC directions by a photodiode or photomultiplier tube [31, 32]. The development of wide-angle 1D light scattering has improved the capability of light scattering for cell label-free analysis [11, 12]. Compared with 1D light scattering measurements, obtaining of wide-angle 2D light scattering patterns from single cells is challenging. For example, in the microscope-based label-free microfluidic ctyometry, the light scattering angular range is limited by the microfluidic chip and the numerical aperture (NA) of the microscope objective [19]. The 2D light scattering label-free static cytometry developed by our group may help to obtain wide-angle 2D light scattering patterns, considering that a microscope can be used to obtain images from static cells with a high NA microscope objective. With wide-angle 2D light scattering label-free cytometry, leukemia subtyping may be achieved since the wide-angle patterns contain much richer information compared to those 2D SSC patterns within a limited angular range.

In this work, we develop a wide-angle label-free static cytometer for automatic classification of acute and chronic myeloid leukemic cells. The wide-angle 2D light scattering patterns of standard microspheres with different diameters acquired by our cytometer agree well with the simulated ones based on Mie theory. This validates the capability of our cytometer to accurately collect the 2D scattered light from 26 to 154 degrees. Texture features such as contrast (CON), angular second moment (ASM) and entropy (ENT) are extracted from the wide-angle 2D patterns of individual acute myeloid leukemic cells (HL-60 cells) and chronic myeloid leukemic cells (K562 cells) with gray level differential statistics (GLDS) method. By using these three GLDS parameters as inputs for a support vector machine (SVM) algorithm with leave-one-out cross-validation (LOO-CV) [33, 34], precise identification of HL-60 and K562 cells is obtained with a sensitivity of 92% and a specificity of 95%. From a clinical perspective, our results demonstrate that the combination of wide-angle 2D light scattering technique and machine learning algorithms is a valuable and promising approach for automatic, label-free classification of acute and chronic myeloid leukemic cells.

2. Materials and method

2.1 Experimental setup

Figure 1 shows a schematic drawing of the wide-angle 2D light scattering static cytometer. A 100-mW collimated laser beam (Frankfurt Laser Company, 532 nm wavelength, Germany) is coupled into a multimode optical fiber (Thorlabs, 105/125 μm) by a 4 × microscope objective. The optical fiber that guides the excitation laser beam is inserted into a static chip to select and illuminate single microspheres or cells on chip. The static chip is mounted on the stage of a microscope (Olympus, BX 53, Japan), and consists of a bottom glass substrate (1 mm in thickness), a top glass layer (0.13 mm in thickness, refractive index 1.5) and two coverslips (0.15 mm in thickness). These two coverslips are put in between the bottom glass substrate and the top glass layer to form a chamber as shown in Fig. 1, which also serves as a reservoir for the suspensions of microspheres or cells. The dimensions of the chamber are 24 mm × 10 mm with a thickness of 0.15 mm. When the laser light interacts with a scatterer on chip, the scattered light is collected by a CMOS detector (Canon, APS-C, Japan) via a 100 × oil-immersion microscope objective with an NA of 1.25 (Olympus, PlanN, Japan). The 2D light scattering patterns are obtained while the system is working in positive defocusing mode [19, 27], where the microscope objective is defocused about 7 μm further away from the scatterer on chip. Considering the effects of the top glass layer, the immersion oil (refractive index 1.5) and the NA of the 100 × objective, our wide-angle 2D light scattering static cytometer collects the 2D scattered light in the angular ranges of ${26}^o\le \theta \le {154}^o$and ${26}^o\le \phi \le {154}^o$. In this work, all the 2D light scattering patterns are obtained in the same angular ranges despite the slight differences in the defocusing distances. The 2D light scattering patterns are cropped and resized to be 2500 pixels by 2500 pixels according to the same light scattering angular ranges, where the CMOS detector is working in a high resolution mode of 5184 pixels by 3456 pixels.

 

Fig. 1 Schematic of the wide-angle 2D light scattering static cytometer. The device consists of a diode pumped solid state (DPSS) laser as the excitation source, a sandwiched glass chip that holds single suspended microspheres or cells excited by an optical fiber, and a 2D light scattering pattern recording system that composes of a 100 × oil-immersion microscope objective and a CMOS sensor. The acquired experimental 2D patterns are sent to a computer for data analysis. The inset picture shows a cross-section illustration of the scattered light transportation from a scatterer to the objective.

Download Full Size | PDF

2.2 Microsphere and cell samples

Standard polystyrene-divinylbenzene (PS-DVB) microspheres (Bangs Laboratories, PS05N, USA) of 3.87 μm (standard deviation (SD) of 0.25 μm) and 4.19 μm (SD of 0.27 μm) in mean diameters were diluted and suspended in ultrapure water. The number density of the diluted sample was approximately 3000 microspheres/mL for single microsphere analysis. Prior to each test, the microspheres were ultrasonicated for 2 minutes.

Human acute myeloid leukemic cell line HL-60 and chronic myeloid leukemic cell line K562 (Cell Bank of Chinese Academy of Sciences, China) were maintained in Iscove's Modified Dulbecco's Medium (IMDM, Invitrogen, Gibco, USA) containing 10% fetal calf serum, 100 U/mL penicillin, and 10 mg/mL streptomycin, and were kept in a 5% CO₂ incubator at 37°C. After several subcultures, cells were washed two times and re-suspended in phosphate buffer saline (PBS). For wide-angle 2D light scattering experiments, the cells were fixed with Immunology Staining Fix Solution (Beyotime, P0098, China) for 30 minutes at room temperature, and were re-suspended in PBS with a concentration of approximately 1500 cells/mL.

3. Results and discussion

3.1 Validation of the wide-angle 2D light scattering static cytometer

Figures 2(a1)-2(a5) show the representative experimental 2D light scattering patterns of 3.87 μm (mean diameter) microspheres with 20 fringes, 21 fringes, 22 fringes, 23 fringes and 24 fringes, respectively. The simulated 2D patterns are obtained by using our Mie theory based algorithm for single microspheres (refractive index, 1.591) immersed in water (refractive index, 1.334) with different diameters of 3.72 µm, 3.82 µm, 3.87 µm, 3.94 µm and 4.07 µm, as shown in Figs. 2(b1)-2(b5), respectively. Similarly, Figs. 2(c1)-2(c4) show the experimental 2D patterns of 4.19 μm (mean diameter) microspheres with 22 fringes, 23 fringes, 24 fringes and 25 fringes. The simulated 2D patterns from single microspheres with different diameters of 3.93 µm, 4.06 µm, 4.19 µm and 4.30 µm are shown in Figs. 2(d1)-2(d4). The diameters of the microsphere models we used for Figs. 2(b) and 2(d) agree with the standard microsphere size ranges of 3.87 ± 0.25 μm and 4.19 ± 0.27 μm, respectively. In this report, the simulated 2D patterns are obtained with an excitation wavelength of 532 nm, and the scattered light is collected in the angular range from 26 to 154 degrees. This angular range is the same as the experimental 2D patterns according to the optical layout of the wide-angle 2D light scattering static cytometer. The comparisons of the experimental and simulated 2D patterns in Fig. 2 reveal a high degree of agreement in the fringe patterns.

 

Fig. 2 Comparisons of wide-angle 2D light scattering patterns between experimental and simulated results of 3.87 μm and 4.19 μm microspheres. (a) and (b) are experimental and simulated 2D patterns from 3.87 μm (mean diameter) microspheres, respectively. (c) and (d) are experimental and simulated 2D patterns from 4.19 μm (mean diameter) microspheres, respectively. The 2D light scattering patterns from the same kind of microspheres (here 3.87 μm or 4.19 μm) are with various fringe distributions in wide-angle light scattering.

Download Full Size | PDF

The microspheres have a SD for their diameters, which will contribute to different fringe numbers in the wide-angle 2D patterns as shown in Fig. 2. In order to compare the experimental patterns with the simulated ones in details, a cross section scanning of the 2D pattern along angle θ at $\phi ={90}^0$ is performed. Figure 3 shows the 1D light scattering intensity comparisons between the experimental and simulated 2D patterns. Figure 3(a) shows the representative comparison between experimental and simulated 2D patterns of Figs. 2(a3) and 2(b3) for the microsphere with diameter of 3.87 μm. For the 4.19 μm microspheres, similar comparison is shown in Fig. 3(b). From Figs. 3(a) and 3(b), it is noticed that in terms of the fringe positions and their relative intensity distributions, the experimental results agree well with the Mie theory simulations for both the 3.87 μm and 4.19 μm microspheres.

 

Fig. 3 Comparisons of light scattering intensity distributions between experimental and simulated results of 3.87 μm and 4.19 μm microspheres. (a) shows a representative comparison of the wide-angle light scattering intensity distributions between experimental and simulated results for 3.87 μm microspheres, and (b) shows the comparison for 4.19 μm microspheres.

Download Full Size | PDF

The above comparisons validate the good performance of our wide-angle 2D light scattering static cytometer for accurately collecting the scattered light in both the polar and azimuthal angular range from 26 to 154 degrees. Furthermore, the microspheres with sub-micrometer size differences contribute to different fringe numbers in their wide-angle 2D light scattering patterns, as shown in Fig. 2. Compared with our previous work in the limited side scattering angular range from 79 to 101 degrees [27, 28], wide-angle 2D light scattering measurements may provide a better resolution and allow us to obtain more vital information of biological cells.

3.2 Classification of acute and chronic myeloid leukemic cells by wide-angle label-free static cytometer

The validated wide-angle static cytometer is applied for label-free classification of acute myeloid leukemic cells (HL-60 cells) and chronic myeloid leukemic cells (K562 cells). For each subtype of leukemic cells, a total of 100 wide-angle 2D light scattering patterns from different single cells are randomly collected by using our label-free static cytometer. Figures 4(a) and 4(b) show the representative 2D patterns from different single HL-60 cells and K562 cells, respectively. Compared with the fringe patterns of standard microspheres, 2D light scattering patterns of leukemic cells are dominated by speckles, which may be attributed to the organelles of the cells.

 

Fig. 4 Experimental wide-angle 2D light scattering patterns from different single myeloid leukemic cells. (a), representative 2D patterns of HL-60 cells. (b), representative2D patterns of K562 cells.

Download Full Size | PDF

Before we extract texture properties of the wide-angle 2D light scattering patterns, a code was developed to remove noise from the 2D pattern and to enhance low-contrast details. Then segmentation of the processed 2D pattern into two layers (target and background) was performed by calculating Euclidean distance [35] between each pixel of a pattern and an RGB vector ([0, 255, 0]). Let $p_{\max }$ denote the maximum pixel intensity of a pattern. If the calculated Euclidean distance is less than or equal to a threshold value T ($T=0.81\times p_{\max }$), the pixel will be assigned to the target layer, otherwise it will be treated as a background pixel with an RGB vector of [0, 0, 0]. The target layer was then reformatted to a gray-level 2D light scattering pattern.

The GLDS method calculates occurrence probability of gray-level differences between neighboring pixels, and is well suited to extract texture features of the 2D light scattering patterns. Here the GLDS method was performed on the processed gray-level 2D patterns and three texture features of CON, ASM, and ENT were obtained. The scatter plots in 3D GLDS parameter space for 100 HL-60 cells and 100 K562 cells are shown in Fig. 5 with CON, ASM, and ENT plotted on the x-axis, y-axis and z-axis, respectively. In Fig. 5, red spherical symbols represent HL-60 cells and black spheres are for the K562 cells. The mean and SD values of these three GLDS parameters for HL-60 cells and K562 cells demonstrate that the two subtypes of leukemic cells can be distinguished in the 3D GLDS space, however they have a very high degree of overlap.

 

Fig. 5 Scatter plots in the 3D GLDS parameter space of 100 HL-60 cells and 100 K562 cells. The CON, ASM, and ENT texture features were extracted from the 2D patterns of HL-60 cells and K562 cells. Blue and green spherical signs indicate the mean values for these two groups of leukemic cells, and their SD values are illustrated by the blue and green ellipsoids, respectively.

Download Full Size | PDF

In order to automatically classify HL-60 cells and K562 cells at single-cell level, the three parameters of CON, ASM and ENT were implemented in a SVM supervised machine learning algorithm. The LOO-CV method was considered as a reliable criteria for parameter selection and was used to provide an unbiased estimation for the classification accuracy. That is, the LOO-CV chose a single 2D pattern from the total of 200 original patterns as a test sample, and left the other 199 patterns as a training set. This process was repeated such that each of the 200 2D patterns was tested. In our SVM based automatic cell classification, the HL-60 cells are set with positive labels and the K562 cells with negative labels. Performance measurements such as sensitivity (Sen = TP/(TP + FN)), specificity (Spe = TN/(TN + FP)), accuracy (Acc = (TP + TN)/(TP + TN + FP + FN)) and the area under curve (AUC) are calculated as shown in Table 1. Here the TP, TN, FP and FN stand for true-positive, true-negative, false-positive and false-negative, respectively. Our SVM algorithm automatically classifies HL-60 cells from K562 cells with a sensitivity of 92% and a specificity of 95%. Table 1 shows that an overall classification accuracy of 93.5% is achieved with an AUC of 0.981. Note that the maximum value of AUC is 1, and a higher value of AUC indicates a better performance of the SVM classifier. These results demonstrate that our wide-angle label-free static cytometer together with machine learning algorithms can be used to automatically classify HL-60 and K562 cells at single-cell level very effectively.

Table 1. Classification of HL-60 cells and K562 cells by SVM with LOO-CV

View Table

4. Summary and conclusions

In this report we developed a wide-angle 2D light scattering static cytometer for label-free automatic classification of acute and chronic myeloid leukemic cells. Our label-free static cytometer collects the scattered light in a very wide angular range from 26 to 154 degrees. This was validated by comparing the experimentally observed 2D light scattering patterns from standard microspheres with the simulated ones based on Mie theory. Our results demonstrated that both 3.87 and 4.19 µm microspheres with a given mean diameter contribute to various 2D wide-angle light scattering patterns. Good agreements of the light scattering intensity comparisons between the wide-angle 2D experimental and simulated patterns were obtained, which confirmed the capability of our wide-angle label-free static cytometer for particle analysis.

Cell subtyping is one of the most critical steps in the diagnosis and treatment of leukemia. Here we demonstrated a label-free and low-cost method for the automatic classification of acute and chronic myeloid leukemic cells. An image analysis algorithm with GLDS method was developed to extract three feature parameters (CON, ASM and ENT) from the experimental wide-angle 2D light scattering patterns of both the acute and chronic myeloid leukemic cells. By combining the wide-angle 2D light scattering cytometry with SVM supervised machine learning, we demonstrated that the HL-60 cells and K562 cells can be automatically classified with an accuracy of 93.5% in a label-free manner. Our wide-angle 2D light scattering static cytometry may find future applications for label-free, automatic screening of leukemia and other diseases at single-cell level.