Open Access
Add to BibSonomy
Abstract
Label-free imaging technology is a trending way to simplify and improve conventional hematology analysis by bypassing lengthy and laborious staining procedures. However, the existing methods do not well balance system complexity, data acquisition efficiency, and data analysis accuracy, which severely impedes their clinical translation. Here, we propose defocusing phase-contrast imaging under the illumination of 415 nm light to realize label-free hematology analysis. We have verified that the subcellular morphology of blood components can be visualized without complex staining due to the factor that defocusing can convert the second-order derivative distribution of samples’ optical phase into intensity and the illumination of 415 nm light can significantly enhance the contrast. It is demonstrated that the defocusing phase-contrast images for the five leucocyte subtypes can be automatically discriminated by a trained deep-learning program with high accuracy (the mean F1 score: 0.986 and mean average precision: 0.980). Since this technique is based on a regular microscope, it simultaneously realizes low system complexity and high data acquisition efficiency with remarkable quantitative analysis ability. It supplies a label-free, reliable, easy-to-use, fast approach to simplifying and reforming the conventional way of hematology analysis.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Hematology analysis is the most common clinical test which supplies the primary or the only evidence for diseases, such as infection [1–4], anemia [5–7], platelet disorder [8–10], blood cancer [11–13], and blood parasite diseases [14,15]. However, the conventional approach relies on chemical staining to reveal the subcellular structures of blood cells. The staining procedure is laborious and lengthy. It usually requires a trained technician to stain according to the strict staining protocols and assess the staining quality under a microscope, which takes about 45 min [16]. These limitations impede the application of hematology analysis for point-of-care testing and low-resource regions. Besides, the staining styles are changeable due to multiple factors (such as staining time, temperature, pH, and proportion of different stains), which makes the diagnosis decision error-prone [17].
Label-free imaging techniques bypass staining-related problems in hematology analysis by measuring endogenous features of blood cells. Raman and autofluorescence signals carry high specific molecular information that can discriminate blood components but suffer from weak optical signals causing a long data acquisition time [18–23]. In addition, the rather complex light path impedes them from primary clinical applications. Absorbing spectra can also characterize blood cells. Hyperspectral microscopy uses three-dimension spectral data of blood samples in a wide wavelength band, which also requires a complex light path and takes a long time to capture data [24–28]. Blood imaging technique based on deep-ultraviolet (deep-UV) microscopy has a simple imaging light path [16,29,30]. But customized deep-UV optical components are still necessary since the current microscopes are mainly designed for visible light. Quantitative phase imaging techniques characterize blood cells using the inner optical property (refractive index or optical phase distribution) differences [31–34]. They generally rely on the supposition of two beams with a high degree of coherence to retrieve the quantitative phase. The data acquisition time will be limited by the computational reconstruction and the data analysis accuracy will be affected by the susceptible imaging procedures.
The existing label-free imaging techniques can overcome the challenges induced by staining, but do not well balance the system complexity, data acquisition efficiency, and data analysis accuracy. They sacrifice the imaging simplicity or efficiency compared to the regular microscope for obtaining the endogenous features. There are several simple methods to realize label-free phase imaging, including differential phase contrast [35,36], oblique illumination [37,38], programmable illumination [39], Fourier ptychographic microscopy [40,41], and defocusing phase contrast imaging [42,43]. The defocusing phase contrast microscopy is easy to implement. According to the transport of intensity equation (TIE), the axial image intensity depicts the second-order derivative distribution of samples’ inner phase [44,45]. The phase variation of biological samples can be simply visualized by defocusing any bright-field microscope without strict coherent illumination. Even though this has been widely used for quantitative phase retrieval, it requires at least two strictly registered axial images and precise system parameters for calculation [46]. Directly defocusing phase-contrast imaging is a promising way for convenient cell characterizing. However, there are two challenges to applying defocusing phase contrast microscopy in the application of hematology analysis. First, compared to other phase contrast approaches, the contrast of conventional defocusing phase-contrast microscopy of the bright-field microscope is low. It's difficult to detect the subcellular structures of leucocytes. Second, the phase-contrast microscopy cannot reveal the hemoglobin distribution of erythrocytes, which is critical for the diagnosis of anemia and is visible in conventional Giemsa staining blood smear.
In this study, we proposed a label-free hematology analysis technique based on defocusing phase-contrast imaging under transmission illumination of 415 nm light via a regular microscope. The wavelength 415 nm is an absorption peak of hemoglobin so that the hemoglobin distribution of erythrocytes is visible [47]. According to Lambert-Beer law, quantitative hemoglobin measurement is also possible. Furthermore, we found that the contrast of the subcellular structures of leucocytes is much higher under 415 nm illumination rather than in white or RGB single-color illumination. Under the 415 nm illumination, the cytoplasmic granules are even more obvious than in Wright-Giemsa staining imaging. We verified its effectiveness by quantitative leucocyte classification and counts. The results demonstrated that this technique can realize automated, high accurate five-part leucocyte classification (the mean F1 score: 0.986 and mean average precision: 0.980). And the leucocyte counts by the proposed technique of a group of samples (n = 14) show remarkable linear correlations with the manual counting results at a significance level of 0.01 with Pearson coefficients large than 0.900. And the quantitative leucocyte percentages show a high linear correlation (Pearson coefficients large than 0.968) with percentages from commercial hematology analyzer [Sysmex XN-10 (B4)] for the neutrophil, lymphocyte, eosinophil, and monocyte. Further, we demonstrated that the blood defocusing phase-contrast images taken by this technique can be virtually stained into images with conventional staining style by a deep-learning network. This is important for the quick clinical translation of this technique.
In summary, this label-free, easy-to-use hematology analysis technique based on a regular microscope can simultaneously realize low system complexity and high data acquisition efficiency with remarkable quantitative analysis ability. It also has great potential to reform the conventional workflow of hematology analysis in clinics and to be quickly popularized, especially in resource-limited areas.
2. Methods
2.1 Label-free blood cell imaging principle
In this study, we design an imaging modality for high-contrast, label-free blood cell imaging, which is simple and easy to implement. The schematic diagram of the imaging system is shown in Fig. 1. The unstained blood smear sample was captured with a defocusing distance of z under the illumination of 415 nm light. The distance from the upper surface of the illuminator to the lower surface of the object stage is ∼56 mm, which can supply a small illumination numerical aperture (NAill) less than the objective numerical aperture (NAobj). We termed the blood image captured by this proposed method as blood defocusing phase-contrast (BDPC) image. According to the TIE [44,45], the BDPC images satisfy:
Here, k = 2π/λ represents the wave vector, λ represents the illumination wavelength (λ = 415 nm here), I(x) = a(x)2 represents the image intensity in the focal plane, a(x) represents the amplitude distribution of the sample, ▽ represents the gradient operator, p(x) represents the phase distribution of the sample. For leucocytes, the I(x) can be regarded as constant (assuming I(x) ≡1 here) since they are weak absorbing objects under the illumination of 415 nm light. Then the BDPC images for leucocytes under this imaging modality, satisfy:
Fig. 1. The schematic diagram of imaging modality for the proposed label-free hematology analysis technique. The unstained blood smear sample was captured with a defocusing distance of z under the illumination of 415 nm light. The distance from the upper surface of the illuminator to the lower surface of the object stage is ∼56 mm, which can supply a small illumination numerical aperture (NAill) less than the objective numerical aperture (NAobj).
Here, $\Delta$ represents the Laplacian, i.e. the second-order derivative operator. It indicates that the optical phase’s second-order derivative of leucocytes is revealed in the BDPC images directly.
The illumination with small NAill under 415 nm can effectively enhance the image contrast. We compared the BDPC images (case 1) with images captured under other five imaging modalities, i.e. imaging with defocusing and NAill = NAobj under 415 nm light (case 2), imaging at the in-focus plane with NAill = NAobj under 415 nm light (case 3), imaging with defocusing and NAill < NAobj under white light (case 4), imaging with defocusing and NAill = NAobj under white light (case 5), and imaging at the in-focus plane with NAill = NAobj under white light (case 6). The results are shown in Fig. 2 (a). In this experiment, we substituted the LED with a center wavelength of 415 nm (410 nm ∼ 420 nm) for the light source (white light) of the conventional microscope for consistent imaging optical path between imaging with 415 light and white light. The intensity curves of the white lines shown in leucocyte 1 and leucocyte 2 are presented in Fig. 2 (b) and (c). The contrast (C) of the intensity curves is calculated as:
Here, the ICmax and ICmin represent the maximum value and the minimum value of the intensity curve respectively. The contrasts for each case are presented in Fig. 2(a). From the results, the defocusing can enhance the image contrast (C2 > C3, C5 > C6, both in leucocyte 1 and leucocyte 2) and the small NAill can further enhance the image contrast (C1 > C2, C4 > C5, both in leucocyte1 and leucocyte 2). Here Ci represents the contrast for case i (i = 1, 2, 3, 4, 5, 6) respectively. And the illumination of 415 nm is also significant for enhancing the imaging contrast by comparing the contrasts of the images taken under light 415 nm and white light. For the line on leucocyte 1, the contrast for the BDPC image is 0.5586. The image contrasts are lower than 0.2800 for case 4 to case 6. For the line on leucocyte 2, the contrast for the BDPC image is 0.2118. The image contrasts are lower than 0.0900 for cases 4 to 6. The results indicate that the three imaging conditions of 1) illumination of 415 nm, 2) small NAill (NAill < NAobj), and 3) defocusing imaging contribute a lot to the high contrast of the BDPC images. And the image contrast of the leucocyte is remarkably enhanced under the illumination of 415 nm light compared to that using white light. Morphological information of blood cells is the gold standard for blood cell identification. The high contrast of the leucocyte BDPC images is the basis for accurate leucocyte differential in this study.
Fig. 2. Comparing the BDPC images (case 1) with the images captured using other five imaging modalities, i.e. imaging with defocusing and NAill = NAobj under 415 nm light (case 2), imaging at the in-focus plane with NAill = NAobj (case 3), imaging with defocusing and NAill < NAobj under white light (case 4), imaging with defocusing and NAill = NAobj under white light (case 5), and imaging at the in-focus plane with NAill = NAobj under white light (case 6). (a) Two leucocyte images were captured from the six imaging modalities. (b) The intensity curves of the white line position (shown in leucocyte 1). The contrasts of the intensity curves for the six modalities are 0.5586, 0.3218, 0.2080, 0.2722, 0.1017, and 0.0855 respectively. (c) The intensity curves of the white line position are shown in leucocyte 2. The contrasts of the intensity curves for the six modalities are 0.2118, 0.0913, 0.0289, 0.0828, 0.0690, and 0.0240 respectively. The BDPC images show higher contrast than the other five modalities.
Besides, the use of a light source with 415 nm light contributes to the high contrast imaging of erythrocytes. The reason for this is that the hemoglobin, which is the main component of erythrocyte, shows prominent absorption peart around the wavelength of 415 nm [48]. We compared the BDPC image and the defocusing bright-field image of the erythrocytes from a hypochromic anemia sample. The imaging results are presented in Fig. 3(a). The intensity curves of the two cases for the white line shown in Fig. 3(a) are presented in Fig. 3(b). The erythrocyte region with hemoglobin absorbs more light under the illumination of 415 nm compared to that using a bright-field microscope with defocusing. Then the image intensity for erythrocytes is much darker than the background image intensity in the BDPC image [one example was indicated by the red arrow in Fig. 3(b)]. Further, the contrast of the white line in the BDPC image is about 1.7 times as large as the defocusing bright-field image (CBDPC = 0.2976, CBF = 0.1714 for bright-field). With the illumination of 415 nm light, the proposed method shows high sensitivity to hemoglobin, which is important for visualizing the morphology of erythrocytes. Further, the anemia, including hypochromic anemia, can be easy to detect by the proposed method.
Fig. 3. Comparing the BDPC image for erythrocytes with the defocusing bright-field image (a) the BDPC image (left) and defocusing bright-field image (right). (b) The intensity curves of the white line position [shown in (a)]. The BDPC image show higher contrast (C = 0.2976) than the image captured with bright-field microscope (C = 0.1714).
The use of illumination of 415 nm also results in higher image contrast compared to three other illumination wavelengths, i.e. 465 nm, 520 nm, and 650 nm. The blood images under the illumination of 415 nm light (BDPC image) and the three illumination wavelengths are shown in Fig. 4 (a). All the images were captured with the same NAill (NAill < NAobj) using defocusing imaging. The intensity curves of the white line position shown in Fig. 4(a) are presented in Fig. 4(b). The contrasts for the four curves are 0.1347, 0.0428, 0.0843, and 0.0309 respectively. The BDPC image shows the highest contrast among the four cases. The image contrast enhanced a lot under the illumination of 415 nm light compared to image contrast under the other three visible lights.
Fig. 4. Comparing the BDPC image with images captured under the other three illumination wavelengths of 465 nm, 520 nm, and 650 nm. The imaging modalities are the same except for the illumination wavelength for the four cases. (a) The BDPC image (top left) and the three images under 465 nm (top right), 520 nm (bottom left), and 650 nm (bottom right). (b) The intensity curves of the white line position [shown in (a)]. The Contrasts are 0.1347, 0.0428, 0.0843, and 0.0309 for BDPC, 465 nm, 520 nm, and 650 nm respectively. The contrast for the BDPC image is much large than that under the three other illumination wavelengths.
In short, the proposed imaging modality of defocusing imaging with small NAill under the illumination of 415 nm can realize high contrast blood cell imaging, which can help realize reliable hematology analysis. Especially, the BDPC images of leucocytes show remarkable image contrast compared to conventional bright-field imaging. And the rich morphological characteristics of the erythrocytes can also be visualized using illumination of 415 nm.
2.2 Experimental setup
The experiments in this study were implemented with a regular microscope with a microscope objective lens (UPlanApo, 40X, 0.95 NA, SOPTOP) and a camera (MV-SUA202GM-T, image size of 1920 × 1200, pixel size of 4.8 µm x 4.8 µm, MindVision). An easy-to-implement illumination was applied to realize small NAill by placing a homemade illuminator on the microscope condenser holder (Fig. S1). The illuminator is composed of an Abbe condenser (NA = 1.25), a 3D-printed outer shell, and a LED (EPILEDS, XP-3535) with a center wavelength of 415 nm (410 nm ∼ 420 nm). The LED is fixed on the bottom center of the illuminator. The distance between the upper surface of the illuminator and the lower surface of the blood smear sample is ∼56 mm.
2.3 Determining the optimal defocusing distance according to the image granularity
In this section, the index of image granularity (IG) was put forward to determine the optimal defocusing distance for BDPC imaging. The IG is defined as:
Here, the Iz(x) is the image intensity with defocusing distance of z and the x represents the image pixel elements. C (·) presents the canny detection operator [49]. The two input thresh parameters of the canny operator were set to 5 and 60 respectively in this study. We defined the direction toward the camera sensor as the positive defocusing direction. Then the defocusing direction applied in our study is negative. The IG curve for a BDPC image was shown in Fig. 5(a). 33 axial images (including images at the in-focus plane) with image size of 1920 × 1200 were captured to calculate the IG values. The red marks represent the actual measurements of the IG values at the 33 axial positions. And the black curve represents the fitting curve (by cubic spline interpolation) according to the actual measurements. From the results, the image at the in-focus plane has IG value of 5.25 × 107 [corresponding to the blue dot in Fig. 5(a)], which is approximate to the local minimal value of the fitting IG curve. The axial image with a maximum IG value of 7.60 × 107 is captured with defocusing distance of -1.1 µm. Then the defocusing distance for this region can be chosen as -1.1 µm. We didn’t search the focal plane (z = 0) first and then defocus it with -1.1 µm. Actually, we adjusted the z position manually to find the position where the image is the sharpest. At this position, the IG value is just maximum. The IG value could be used for the auto-focusing criterion in our future work. Local leucocyte images at different axial positions were presented in Fig. 5(b). From the images, the leucocyte image with defocusing distance z of -1.1 µm shows clear and discernible morphologic characteristics, which can validate the effectiveness of the IG curve. Further, according to the experiment results, the leucocyte images with defocusing distances ranging from -0.5 µm to -1.1 µm are acceptable through manual observation. The acceptable defocusing range for this region is around 0.6 µm. Other local leucocyte images from the 33 axial images (Fig. S2) and experimental results for the other four groups of axial images with image size of 1920 × 1200 were presented in Fig. S3 to Fig. S7. Consistent conclusions can be inferred.
Fig. 5. Determining the optimal defocusing distances according to the image granularity. (a) Image granularity curve of the axial images (with image size of 1920 × 1200, image amount of 33). The red marks represent the actual measurements of the IG values at the 33 axial positions. And the black curve represents the fitting curve (by cubic spline interpolation) according to the actual measurements. The images at the focal plane have IG values of 5.25 × 107, which is approximate to the local minimal value of the fitting IG curve. The axial image with a maximum IG value of 7.60 × 107 is captured with defocusing distance of -1.1 µm (red dot). (b) Local leucocyte images at different axial positions.
2.4 Blood sample and data preparation
The unstained peripheral blood smears (n = 35) were from the Tongji Hospital (Tongji Medical College, Huazhong University of Science and Technology) without any patient-related information [Shown in Fig. S8(a)]. The blood was taken from each test tube (about 10 µl) to make blood smears without staining. After drying in the air, the BDPC images (image size of 1920 × 1200) were captured from the single-layer region of each blood smear [Fig. S8(b)]. The experimental procedures were in accordance with Chinese law and were approved by the Scientific Committee of Huazhong University of Science and Technology.
2.5 Registration algorithm
The Wright-Giemsa images were required to supply the leucocyte subtype labels and the virtual-staining labels for the five-part leucocyte differential and virtual staining tasks based on deep learning in this study. After the BDPC images were collected, the blood smear was stained with Wright-Giemsa stain. Then the stained blood smears were scanned by a whole slide imaging system (WSI, SR-RBF-01-V3, ConvergenceTech) to obtain the Giemsa images [Fig. S9(a)]. The WSI system contains a camera and objective lens with the same type specification as those applied in our imaging system. A two-step registration algorithm, including a coarse registration step and a fine registration step, was applied to match BDPC images and the corresponding Wright-Giemsa images [Fig. S9(b)]. The registration algorithms were executed on the platform of MATLAB R2021a. The coarse registration step is based on the template matching algorithm. The scanned Giemsa images were converted into gray style first. Given an input BDPC image (image size of 512 × 512), the image similarities between the input image and image patches (image size of 640 × 640) from the Wright-Giemsa image were calculated. The root-mean-square error (RMSE) was applied as the image similarity criterion. The location where the RMSE achieve a minimum was found and the coarse-registration image (image size of 640 × 640) was cropped from the whole Wright-Giemsa image. The fine registration step was carried out to correct subpixel transformations. The multimodal registration function from MathWorks was applied in this step [50]. The image dataset containing precisely matched BDPC images and Giemsa images was created by this method.
2.6 YOLOv5 model for five-part leucocyte classification
The classification task is based on a target detection network of YOLOv5 [51]. We used BDPC images (image size 512 × 512) and their leucocyte labels from the corresponding Wright-Giemsa images to train and test the network. The leucocyte subtype labels include the position coordinates of the bounding boxes with leucocyte and the leucocyte type). The dataset contains 3711 BDPC images (3319 for training and validating and 370 for testing). In the training phase, the BDPC images were fed randomly to the network. Data augmentations including intensity random gamma transformation, random brightness adjustment, and histogram equalization were performed. This can confront the intensity variations caused by experimental parameters (different exposure times, white balance settings, etc.) and increase the robustness of the network. In the testing phase, the inference was performed with confidence thresh of 0.75 and IoU (Intersection-Over-Union) thresh of 0.35. When BDPC images were input into the trained network, the position coordinate and the leucocyte subtype will be output. After training, the inference time for each BDPC image with image size 512 × 512 is around 9 ms.
2.7 Automated five-part leucocyte differential count
To verify the proposed method’s performance for the automated five-part differential count, blood smear samples from 14 patients were analyzed by the trained network. First, the BDPC images (image size 1920 × 1200) of each blood sample were captured. The average amount of BDPC images for each sample is 101. The exposure time to capture an image is from 5 ms to 12 ms. Second, the automated leucocyte differential count was performed by the trained network. The BDPC images with image size 1920 × 1200 will be cut into image patches (with image size 512 × 512) with an overlap of 25%. The image patches will be input into the network to detect the leucocyte instances (indicated by the bounding boxes). Then the detection results in each image patch will be collected to obtain the results for the original input image. After this, the non-maximum suppression algorithm [52] will be applied to remove redundant boxes. Then the automated five-part leucocyte differential counts were compared with the manual counts which is the gold standard based on the Wright-Giemsa images. And the quantitative leucocyte percentages were compared with the quantitative leucocyte percentages from the automated hematology analyzer [Sysmex XN-10 (B4)] used in the Tongji Hospital. The original data was supplied (Table S1 to S6 in Supplementary 1). The inference time for each BDPC image with an image size of 1920x 1200 is around 162 ms. The BDPC image amounts and the leucocyte amounts for each sample were supplied (Table S7 in Supplementary 1).
2.8 Conditional adversarial network for virtual staining
The virtual staining is performed with a network architecture of pixel2pixel conditional adversarial networks (CAN) [53] with a generator G and a discriminator D. Image pairs containing the BDPC images and their corresponding actual Giemsa images were used to train the network. The input image size is 512 × 512. The dataset contains 3970 image pairs (3875 for training and 95 for testing). The loss function of the network for the work is defined as:
Here:
In the testing phase, the BDPC images were stained with the trained generator directly. The virtual staining for each image with size of 512 × 512 takes ∼4.49 ms. For the virtual staining for large-area blood smear region, the BDPC images with image size of 1920 × 1200 were captured continuously. Then the large-area BDPC images (area of 676.8 µm x 676.8 µm, image size of 5640 × 5640) were obtained by registration and fusion with the Grid stitching plugin [54]. After this, the large-area BDPC image was cut into image patches with image size of 512 × 512 with an overlap of 25% to perform the virtual staining. Finally, the virtually stained image patches were stitched again according to the original positions to obtain the large-area virtual staining images. The stitching work of the virtual staining patches was implemented with Python 3.7.0. The virtual staining time for the large-area BDPC image is ∼544.83 ms.
The networks for leucocyte classification and Visual staining of the BDPC images were implemented using Python version 3.7.0, with torch framework version 1.8.1. This work was performed on a desktop computer with Intel Core i7-8700 CPU @ 3.20 GHz and 16GB of RAM, running with a Windows 10 operating system. The network training and testing were performed using NVIDIA GeForce GTX 1060 GPU.
3. Results
3.1 Visualizing specific morphology features of unstained blood cells
The proposed technique visualized the unstained blood cells by recording the second-order derivative of the blood cells’ optical phase distribution. In general, the phase changes rapidly at the edge of blood cells’ inner structures. Then the inner structures of the blood cells can be visualized. We compared the BDPC images (gray images in Fig. 6) with their corresponding Wright-Giemsa images (color images in Fig. 6).
Fig. 6. Comparison between the BDPC images (gray images) and the conventional Wright-Giemsa images (color images). Image pairs of the BDPC images and corresponding Wright-Giemsa images for (a) five leucocyte subtypes, (b) atypical leucocytes from leukemia patients (c) atypical platelets with large volume (light pink arrowheads), snakelike shape (light-yellow arrowheads), aggregate platelets (light purple arrowheads), (d) kinds of atypical erythrocytes with morphological abnormity (pointed by white arrowheads), structure abnormity (Howell-Jolly’s body pointed by orange arrowheads) and hemoglobin abnormity content (hypochromic erythrocytes pointed by red arrowheads).
The leucocyte has the most complex morphological features among the blood cells. The results show that the profiles of the leucocytes’ cell body, cell nuclei, and cytoplasm textures are consistent with their corresponding Wright-Giemsa images [Fig. 6(a) and 6(b)]. This technique characterizes blood cells by the inner optical property differences. Then the images of eosinophil and basophil present rough image textures due to coarse cytoplasmic granules with variable refractive index. The BDPC images for neutrophils and monocytes show that this technique is quite sensitive to phase variation. The fine folding textures of the cell nuclei [pointed by the white arrowheads in Fig. 6(a)] are revealed by this technique, which can hardly be seen in the Wright-Giemsa images. We also verified that the atypical leucocytes from leukemia patients’ blood smear samples can be characterized by the proposed technique [Fig. 2 (b)]. Dense blast cells can be seen in the BDPC images, including the atypical leucocytes with strumaes [pointed by the red arrowheads in Fig. 6(b)], atypical leucocytes with pseudopods [pointed by the orange arrowhead in Fig. 6(b)], and plasmocyte [pointed by the blue arrowhead in Fig. 6(b)].
The platelets have brighter image intensity than the background in the BDPC images, which makes them easy to be identified. Atypical platelets, including large platelets [the light pink arrowheads in Fig. 6(c)], snakelike platelet [the light-yellow arrowheads in Fig. 6(c)], and aggregate platelets [the light purple arrowheads in Fig. 6(c)] can be easily discriminated. The erythrocyte can be revealed by this technique under the illumination of 415 nm light. Erythrocytes with morphological abnormity [pointed by white arrowheads in Fig. 6(d)], structure abnormity [Howell-Jolly’s body pointed by orange arrowheads in Fig. 6(d)], and hemoglobin abnormity content [hypochromic erythrocytes pointed by red arrowheads in Fig. 6(d)] can be visualized. These results indicate that the significant morphological features for diagnosis decisions can be supplied by the BDPC images.
3.2 Five-part leucocyte classification by the trained YOLOv5 model
The BDPC images of five leucocyte subtypes and their corresponding type labels from the Wright-Giemsa images were used to train the network of YOLOv5. Some BDPC images of the five leucocyte subtypes from the training dataset were shown [ Fig. 7(a)]. We then used the UMAP (uniform manifold approximation and projection) to visualize the image features (3594 leucocyte images) extracted by the trained network in the two-dimension space. The separated points for the five leucocyte subtypes indicated that this technique has favorable separability for the five leucocyte subtypes [Fig. 7(b)].
Fig. 7. Automated and accurate five-part leucocyte differential with a deep-learning algorithm (YOLOv5). (a) BDPC images of five leucocyte subtypes from the training dataset. (b) The UMAP feature visualization for 3594 leucocytes BDPC images in a two-dimension space. The result shows a favorable separability for the five leucocyte subtypes with this technique. (c) Confusion matrix for the differential results on the test dataset. The confusion matrix presents counts from the predicted and actual labels. Each column represents the actual counts for each subtype and each row represents the predicted counts for each subtype. (d) Quantitative metrics, containing precision, recall, and F1 score, for evaluating the differential performance. (e) PR curve evaluating the classifier performance. The mAP achieves 0.980 for the five leucocytes. NEU: neutrophil, LYM: lymphocyte, EOS: eosinophil, BAS: basophil, MON: monocyte.
The classification results on the test dataset are shown in Fig. 7(c)–(e). The confusion matrix represents the counts from the predicted and actual leucocyte labels [Fig. 7(c)]. With the matrix, three quantitative metrics of the precision, recall, and F1 score were calculated [Fig. 7(d)]. The recall is the ratio of the true positive counts and the actual positive counts, the precision is the ratio of the true positive counts and the predicted positive counts, and the F1 score is the harmonic mean of the precision and the recall [2·precision·recall/(precision + recall)]. The mean precision, mean recall, and the mean F1 scores for the five leucocytes are 0.984, 0.978, and 0.986 respectively. From Fig. 7(d), the classification performance of the monocyte is weaker than that of other leukocyte subtypes (precision of 0.953, recall of 0.911, F1 score of 0.931). From the precision-recall (PR) curve [Fig. 7(e)], the AP (i.e. the integral area under the PR curve) are 0.992, 0.990, 0.995, 0.984 0.937 for the neutrophil, lymphocyte, eosinophil, basophil, and monocyte respectively. The mean AP (mAP) for the five leucocyte subtypes achieves a high value of 0.980. The high mean F1 score (0.986) and mAP (0.980) mean that the specific features of the BDPC images for the five leucocyte subtypes can be extracted by the trained network making accurate five-part leucocyte classification.
3.3 Automated five-part leucocyte differential counts versus manual counts
Leucocyte differential counts are important hematology indexes for disease diagnosis. We analyzed the five leucocyte counts of the peripheral blood samples from 14 patients with the proposed technique. The BDPC images of each unstained blood sample were collected and the automated five-part leucocyte differential counting was performed based on the trained deep-learning network. Then the corresponding blood samples were stained by the Wright-Giemsa stains and the Wright-Giemsa images of each blood sample were collected. The real counts of the five-part leucocytes of each blood sample were obtained from the Wright-Giemsa images by manual counting.
The counting results from the proposed method (BDPC counts) and the manual method based on Wright-Giemsa images (manual counts) for all kinds of leucocyte subtypes including the total leucocyte counts were presented in Fig. 8. Each green dot in each figure presents a measurement for one sample. For basophil, there are 9 among the 14 samples having counts of 0 both from the proposed method and the manual method due to the extremely low percentage (the clinical reference range is between 0.0%∼1.0%) of the leucocyte subtype in the peripheral blood. The red dashed line (y = x) indicates the ideal positions for all the measurements, the blue line is the fitting straight line (from Passing-Bablok linear regression) for the counting results from the two methods. The expressions of each fitting straight line were marked in each figure. The fitting slopes are 0.981, 1.000, 1.000, 1.000, 0.960, and 1.000, respectively. The intercepts are 1.308, 1.000, 0.500, 0.000, 0.177, and 0.000, respectively. The fitting straight lines are all approximate to the ideal position (dashed red lines). By a two-sided significance test (Student's t-test), the counts for the five leucocyte subtypes and the total leucocyte from the proposed method show a significant correlation at significance levels of 0.01. The P values were supplied (Table S8 in Supplement 1). The Pearson coefficients (R) are 1.000, 1.000, 0.990, 0.908, 0.998, and 1.000, respectively, which indicates a strong linear correlation between the two methods.
Fig. 8. Leucocyte counts from the proposed method (BDPC counts) versus the manual counts based on Wright-Giemsa images for blood samples from 14 different patients. Each green dot represents the count of one sample. The blue lines indicate the fitting straight lines (from Passing-Bablok linear regression) for the counts from the two methods (the expressions of each fitting line were marked). The dashed red lines (y = x) indicate the ideal position of the measurements. The results show a significant correlation at significance levels of 0.01 (indicated by the ** beside the R values) for the six different leucocyte types, i.e. neutrophil, lymphocyte, eosinophil, basophil, monocyte, and the total leucocyte count, by a two-sided significance test (Student's t-test). The Pearson coefficients are 1.000,1.000, 0.990, 0.908, 0.998, and 1.000, respectively. The insets on the right bottom in each figure are the box plots generated from count differences (BDPC counts minus manual counts). The medians (M) of the count difference are 0.5, 1.0, 0.5, 0.0, 0.0, and 0.0, respectively. Outliers are indicated by red cross-shaped marks. The number of outliers is 1, 0, 1, 2, 1, and 1, respectively. The ratio of the difference values within the normal range up to a percentage of 92.96% (78/84), which shows a high consistency between the two methods.
We further evaluated the consistency between the two methods through the boxplots. Counting differences between the proposed method and the manual method (BDPC counts minus manual counts) were calculated to generate the boxplot for each leucocyte type (Subgraph in each figure). The central mark (red line) indicates the median, and the bottom and the top edge of the blue box indicate the upper quartile and the lower quartiles, respectively. The whiskers indicate the most extreme value not considered outliers, and the red cross-shaped marks indicate the outliers. The medians (M) of the count difference are 0.5, 1.0, 0.5, 0.0, 0.0, and 0.0, respectively, which means that no significant systematic biases exist. From the box plots, there are only 6 outliers among the total 84 (14 × 6) measurements. The percentage of the measuring differences within the normal range is 92.96% (78/84), which further indicates a high consistency between our method and the manual method. The number of outliers is 1, 0, 1, 2, 1, and 1, respectively for the 6 cases. In short, reliable five-part leucocyte differential counts can be obtained by comparing the results with the manual method. Related data of the boxplots were supplied (Table S9 in Supplement 1).
3.4 Five-part leucocyte differential percentages from the proposed method versus percentages from the automated hematology analyzer
In this section, we further compared the five-part leucocyte percentages of the 14 samples from the proposed method with the percentages from the automated hematology analyzer (Sysmex XN-10). The leucocyte percentage is the ratio of the specific leucocyte count to the total leucocyte count. The results were shown in Fig. 9.
Fig. 9. Leucocyte percentages from the proposed method (BDPC percentages) versus the percentage from the automated hematology analyzer (Sysmex XN-10) based on flow cytometry for blood samples from 14 different patients. Each green dot represents the measurement of one sample. The blue lines indicate the fitting straight lines (from Passing-Bablok linear regression) for the leucocyte percentages from the two methods (the expressions of each fitting line were marked). The dashed red lines (y = x) indicate the ideal position of the measurements. The results show a significant correlation at significance levels of 0.01 (indicated by the ** beside the R value) for the neutrophil, lymphocyte, eosinophil, and monocyte by a two-sided significance test (Student's t-test). The Pearson coefficients (R) are 0.971, 0.983, 0.984, and 0.968 for these four leucocyte subtypes. The two-sided significance test indicates no significant correlation for basophil mainly due to the sampling error (P = 0.162, R = 0.395). The insets on the right bottom of each figure are the box plots generated from percentage differences (percentages from the proposed method minus percentages from the hematology analyzer). The medians (M) of the difference are 0.33%, 2.1%, 0.67%, -0.60%, and -2.4%, respectively. Outliers are indicated by red cross-shaped marks. The number of outliers is 0, 0, 1, 4, and 1, for neutrophil, lymphocyte, eosinophil, basophil, and monocyte respectively.
The same statistic models (see section 3.3) were applied in this section to compare the quantitative percentages for neutrophil, lymphocyte, eosinophil, basophil, and monocyte from the proposed method (BDPC percentages) and the Sysmex XN-10. By the Student's t-test, the percentages for neutrophil, lymphocyte, eosinophil, and monocyte from the proposed method show a significant correlation at significance levels of 0.01 comparing the measurements from the Sysmex XN-10 (Table S10 in Supplement 1). The Pearson coefficients are 0.971, 0.983, 0.984, and 0.968 for these four leucocyte subtypes, which indicates a strong linear correlation. The slopes of the fitting straight line for neutrophils and lymphocytes are 0.956 and 1.003 respectively and the intercepts of the fitting straight line are 0.377 and 0.021 respectively, which are approximate to the ideal line (y = x). For eosinophil and monocyte, the slopes are 1.311 and 0.645 respectively, which shows slight bias from the ideal line. The intercepts for the eosinophil and the monocyte are 0.006 and 0.005 respectively, which are approximate to zero. In this section, the percentage differences between the two methods (BDPC percentages minus Sysmex XN-10 percentages) are calculated to generate the boxplots (Table S11 in Supplement 1). From the results, the medians (M) of percentage difference are 0.33%, 2.10%, 0.67%, and -2.4% for neutrophil, lymphocyte, eosinophil, and monocyte respectively. There is no more than one outlier for the four leucocyte subtypes (the number of outliers is 0, 0, 1, and 1, respectively). From the above results, the quantitative percentages from the proposed method show good correlation and consistency with the commercial hematology analyzer for neutrophils, lymphocytes, eosinophils, and monocytes.
While for basophil, the Student's t-test shows no significant correlation between the two methods. The corresponding P value is 0.162 which is large than 0.05 and the Pearson correlation R is equal to 0.395. From the fitting results based on Passing-Bablok linear regression, the fitting result is also worse for basophil with a fitting straight line of y = 0. And from the boxplot, four outliers exist. The main reason for this is that the basophil account for an extremely low percentage of the total leucocyte count with a reference range between 0%∼1%. For the hematology analyzer based on flow cytometry, thousands of leucocytes will be detected to generate the final results. For the proposed method, about 100 BDPC images were analyzed which contains less account of leucocytes. This will affect the quantitative basophil percentage mainly due to the sampling error.
3.5 Transferring the BDPC images into a virtual Wright-Giemsa image
To assist the diagnosis and accelerate the clinical translation of the technique, the BDPC images were virtual staining into images with Wright-Giemsa staining appearances which are familiar to pathologists [schematic diagram can be seen in Fig. 10(a)]. We captured a BDPC image of an unstained blood smear region with an area of 676.8 µm x 676.8 µm (Fig. S11, image size of 5640 × 5640). Then the BDPC image was virtually stained by the trained network [Fig. 10(c)]. The virtual staining images and their Wright-Giemsa images within four rectangular regions were shown. The original BDPC images within the regions are presented in Fig. 10 (b). The structural similarity (SSIM) scores were applied to evaluate the virtual staining results here. The SSIM scores are 0.7666 (for purple box), 0.6901 (for green box), 0.6747 (for yellow box), and 0.7605 (for blue box) for the four regions respectively. It is possible that the random artifacts (which can be seen in virtual staining images) occur on the image background after virtual staining resulting in the ordinary SSIM scores here. Despite this, combined with the virtual staining technique, the BDPC images can supply normalized images with Wright-Giemsa staining appearances without a cumbersome staining process, saving time and labor for staining. More virtual staining results were presented in Fig. S12.
Fig. 10. Virtual staining of the BDPC images with a conditional adversarial network. (a) Schematic diagram of the virtual staining with a conditional adversarial network that contains a generator G and a discriminator D. A BDPC image from an unstained blood smear was stained by the G to mimic the Wright-Giemsa image and then the D tries to distinguish between the virtual staining image and the real Wright-Giemsa image. (b) For BDPC images (image size of 256 × 256) within four rectangular regions on the large-area BDPC images with an actual area of 676.8 µm x 676.8 µm (Fig. S11). (c) Virtual staining image of the large-area BDPC image. The four virtual staining images corresponding to (b) within the four rectangle regions and their corresponding Wright-Giemsa images were shown.
4. Discussion and conclusion
This study has demonstrated that a label-free hematology analysis technique can be simply achieved with a regular microscope based on defocusing phase-contrast imaging, with specific transmission illumination of 415 nm light. We adopted an easy and low-cost way to modify the illumination of a regular microscope by directly substituting the microscope condenser with a homemade illuminator of 415 nm light. The homemade illuminator costs only ${\$}$12 (${\$}$10 for an Abbe condenser, ${\$}$1 for a 3D-printed shell, and ${\$}$1 for a LED). More importantly, this will not affect the normal use of the regular microscope for other biological applications. In this study, we proved the high performance of the proposed label-free hematology analysis technique on a regular microscope with a high numerical aperture (NA) objective lens (40X, 0.95NA). Next, we will try to realize hematology analysis using the same imaging modal with a miniature microscope (which has a lower spatial resolution of ∼1 µm). Five-part leucocyte differential count is possible with the miniature microscope for two reasons. First, the miniature microscope has the ability to image the cell nuclei (with an image size large than 1 µm) using defocusing imaging. Second, the cytoplasmic granules of basophils and eosinophils will have special image textures due to defocusing [Fig. 7(a)]. This will also result in different image intensities if a miniature microscope is used.
The combination of defocusing phase contrast and illumination of 415 nm wavelength is the key to this technique. The leucocyte has complex morphological characteristics among all blood cell types due to the variable nuclei structures and different cytoplasmic granules properties. Exogenous stains, such as Wright-Giemsa stains, hematoxylin-eosin (HE) stains, and Acridine Orange (AO) fluorescence stains, have been used for discerning leucocyte subtypes in conventional ways. In this technique, the subcellular morphology of the leucocyte is visualized by defocusing which can capture the second-order derivative of the optical phase. Then the profile of the leucocyte’s subcellular structures can be recorded. Besides, adopting illumination light of 415 nm also contributes to the high contrast of leucocyte images. We compared illumination of white, 415 nm (purple), 460 nm (blue), 530 nm (green), and 630 nm (red), and we found that the contrast of leucocytes is highest at 415 nm. However, 415 nm is the absorption peak of erythrocytes, and it seems not special for leucocytes. We guess it's because the short wavelength of purple light helps to achieve a higher optical resolution than other visible lights. The equivalent optical resolution reaches 266 nm with this system (optical resolution δ = 0.61λ/NA, where λ = 415 nm and an air lens with numerical aperture NA of 0.95), which achieves an optical resolution of a microscope system with an oil lens.
Illumination of 415 nm light is the key for label-free erythrocyte imaging as well. Considering that the hemoglobin (both the oxy- and deoxyhemoglobin) has a prominent absorption peak at the wavelength of 415 nm (corresponding to the Soret peak), this technique shows high sensitivity to trace the amount of hemoglobin. This will be helpful for the imaging of hypochromic erythrocytes [example of this kind of cells can be seen in Fig. 6(d)], which are significant clues for assessing anemia. Besides, according to Lambert's law, the absorbance of hemoglobin is proportional to the hemoglobin concentration. Theoretically, the quantitative hemoglobin content of the erythrocytes can be calculated from the images captured with this technique [30]. But for the factor that the molar extinction coefficient E of hemoglobin varies a lot with oxygenation state around the wavelength of 415 nm light (at the wavelength of 414 nm, the EHbO2 is 524280 cm-1/M, EHb = 342596 cm-1/M [48]), the hemoglobin absorbance will be affected by the actual ratio of the oxyhemoglobin and the deoxyhemoglobin. It will need to consider the actual oxygenation state in the calculation or make a further calibration of the quantitative hemoglobin according to the reference value from a commercial hematology analyzer. Details will be studied in our future work. Anyhow, the 415 nm absorption peak is much higher than other absorption peaks of erythrocytes, resulting in strong intensity signals. The strong intensity signal is valuable for the measurement of hemoglobin because it can utilize the dynamic range of the camera better.
In hematology, there are usually two situations, blood cells on the blood smear, or blood cells suspended in an aqueous environment. We found that the structures of cells are quite different in these two situations. When the blood cells are suspended in an aqueous environment, they are round like balls, instead of flat on the blood smear. It is much more difficult to recognize the subcellular structure when the cells are round like balls. It means that the morphological analysis of blood cells in an aqueous environment is another story. Not because of the BDPC method, but because of the cells themselves. However, the preparation of a blood smear is a problem in the point-of-care application, since it needs a technician as well as staining. There is a conflict that the microfluidic chip for blood cells can simplify the preparation of samples, but the cells suspended in an aqueous environment are hard to recognize, while the cells on a blood smear are easy to recognize, but the preparation requiring technicians is a problem in point-of-care application. One approach is to develop a low-cost easy-to-use tool for the preparation of the blood smear. Another approach is to flatten the cells in the aqueous environment. We found that if we put a cover glass on the blood, instead of smearing it, the morphology of blood cells is quite similar to a blood smear. Although the cells are in an aqueous environment, as shown in Fig. S13.
In this study, we have verified that the deep features of the BDPC images can be well extracted by the deep-learning algorithm to make an accurate classification, which is promising to supply more subjective diagnostic decisions than the manual examination. Images with conventional Giemsa staining appearance can also be obtained by an image translation algorithm based on deep learning. Then the normalization of the staining style for blood samples can be achieved, which can guarantee the consistency of the diagnosis analysis. The performance of the classification and virtual staining algorithm is based on the imaging ability of BDPC. BDPC can reveal the profile of the cell and the cell nuclei since it converts the second-order derivative of the optical phase to intensity. And it can get an even higher contrast of the cytoplasmic granules, which is the critical clue for classification, than conventional Giemsa blood imaging. The actual colors of structures in Giemsa blood imaging are relevant to the cell type, and the structural features are relevant to the cell type as well. Then the deep learning network can learn the relationship between the structural features and the colors. For example, lymphocytes and their nuclei are small and round, with bluer cytoplasm than neutrophils, as shown in Fig. 10.
In this study, the BDPC images were captured manually. Even though this can be easily done by any person with little experience, an automated scanning program is necessary if there are requirements for high throughput applications. A proper autofocus algorithm is the key to realizing it. We found that the IG may be supplied as a suitable autofocus criterion by analyzing the IG curves for different blood smear regions. The autofocus algorithm for capturing the BDPC images is similar to that of conventional bright-field imaging. The target images, i.e. the BDPC images, can be captured by searching the peak of the IG curve directly. According to our experiment results, the acceptable range of the z-plane for defocusing phase contrast is about 0.6 µm from the peak of IG value at NA of 0.95 (as shown in Fig. 5, Fig. S2 to Fig. S7), which is approximately equal to the depth of focus at NA of 0.95 (λn/NA^2 + n*e/(M*NA) ≈ 0.71 µm, where n is the refractive index of air, e is the pixel size of the camera, M is the magnification of the system).
The proposed technique is potential to be popularized since the regular microscope is one of the most common imaging tools available to pathologists. It can also greatly simplify the conventional way of hematology analysis by bypassing the lengthy and laborious staining procedures. With simple sample preparation within a minute, the hematology analysis can be performed. Furthermore, quantitative leucocyte differential counts and virtual Giemsa images can be automatically obtained, which can be supplied as diagnostic evidence for telemedicine. We will explore this technique’s potential for early diagnosing of various diseases, including leukemia, anemia, and thrombocytopenic purpura, in our following work. This easy-to-use, reliable, label-free technique can also facilitate the development of point-of-care devices for hematology analysis and improve the health level, especially in resource-limited areas.
Acknowledgments
The authors thank Convergence Technology Co., Ltd. (Wuhan, China) supplied the whole slide imaging system to scan the stained blood smears. Tongji Hospital (Tongji Medical College HUST, Wuhan, China) supplied the unstained blood smears.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data/Codes underlying the results presented in this paper are available in Ref. [55].
Supplemental document
See Supplement 1 for supporting content.
References
1. M. H. F. Klinger and W. Jelkmann, “Review: Role of blood platelets in infection and inflammation,” J. Interferon Cytokine Res. 22(9), 913–922 (2002). [CrossRef]
2. L. A. Al-Gwaiz and H. H. Babay, “The diagnostic value of absolute neutrophil count, band count and morphologic changes of neutrophils in predicting bacterial infections,” Med. Princ. Pract. 16(5), 344–347 (2007). [CrossRef]
3. T. Honda, T. Uehara, G. Matsumoto, S. Arai, and M. Sugano, “Neutrophil left shift and white blood cell count as markers of bacterial infection,” Clin. Chim. Acta 457, 46–53 (2016). [CrossRef]
4. T. O’Connor, J.-B. Shen, B. T. Liang, and B. Javidi, “Digital holographic deep learning of red blood cells for field-portable, rapid COVID-19 screening,” Opt. Lett. 46(10), 2344 (2021). [CrossRef]
5. K. de Haan, H. Ceylan Koydemir, Y. Rivenson, D. Tseng, E. Van Dyne, L. Bakic, D. Karinca, K. Liang, M. Ilango, E. Gumustekin, and A. Ozcan, “Automated screening of sickle cells using a smartphone-based microscope and deep learning,” Npj Digit. Med. 3(1), 76 (2020). [CrossRef]
6. C. I. Wilson, P. L. Hopkins, B. Cabello-Inchausti, S. J. Melnick, and M. J. Robinson, “The peripheral blood smear in patients with sickle cell trait: A morphologic observation,” Lab. Med. 31(8), 445–447 (2000). [CrossRef]
7. G. Dhaliwal, P. A. Cornett, and L. M. Tierney-Jr, “Hemolytic anemia,” Am. Fam. Physician 69(11), 2599–2606 (2004).
8. A. Greinacher, A. Pecci, S. Kunishima, K. Althaus, P. Nurden, C. L. Balduini, and T. Bakchoul, “Diagnosis of inherited platelet disorders on a blood smear: A tool to facilitate worldwide diagnosis of platelet disorders,” J. Thromb. Haemost. 15(7), 1511–1521 (2017). [CrossRef]
9. D. S. Velaskar and A. P. Chitre, “A new aspect of platelet aggregation and a test to measure it,” Am. J. Clin. Pathol. 77(3), 267–274 (1982). [CrossRef]
10. J. B. Brière, “Essential thrombocythemia,” Orphanet J. Rare Dis. 2(1), 3 (2007). [CrossRef]
11. G. S. Nowakowski, J. D. Hoyer, T. D. Shanafelt, C. S. Zent, T. G. Call, N. D. Bone, B. LaPlant, G. W. Dewald, R. C. Tschumper, D. F. Jelinek, T. E. Witzig, and N. E. Kay, “Percentage of smudge cells on routine blood smear predicts survival in chronic lymphocytic leukemia,” J. Clin. Oncol. 27(11), 1844–1849 (2009). [CrossRef]
12. S. S. Patel, G. A. Aasen, M. M. Dolan, M. A. Linden, R. W. McKenna, V. K. Rudrapatna, B. J. Trottier, and S. M. Drawz, “Early diagnosis of intravascular large b-cell lymphoma: Clues from routine blood smear morphologic findings,” Lab. Med. 45(3), 248–252 (2014). [CrossRef]
13. A. N. Khvastunova, S. A. Kuznetsova, L. S. Al-Radi, A. V. Vylegzhanina, A. O. Zakirova, O. S. Fedyanina, A. V. Filatov, I. A. Vorobjev, and F. Ataullakhanov, “Anti-CD antibody microarray for human leukocyte morphology examination allows analyzing rare cell populations and suggesting preliminary diagnosis in leukemia,” Sci. Rep. 5(1), 12573 (2015). [CrossRef]
14. M. T. Rinehart, H. S. Park, K. A. Walzer, J.-T. A. Chi, and A. Wax, “Hemoglobin consumption by P. falciparum in individual erythrocytes imaged via quantitative phase spectroscopy,” Sci. Rep. 6(1), 24461 (2016). [CrossRef]
15. M. V. D’Ambrosio, M. Bakalar, S. Bennuru, C. Reber, A. Skandarajah, L. Nilsson, N. Switz, J. Kamgno, S. Pion, M. Boussinesq, T. B. Nutman, and D. A. Fletcher, “Point-of-care quantification of blood-borne filarial parasites with a mobile phone microscope,” Sci. Transl. Med. 7(286), 286re4 (2015). [CrossRef]
16. A. Ojaghi, G. Carrazana, C. Caruso, A. Abbas, D. R. Myers, W. A. Lam, and F. E. Robles, “Label-free hematology analysis using deep-ultraviolet microscopy,” Proc. Natl. Acad. Sci. 117(26), 14779–14789 (2020). [CrossRef]
17. K. Dunning and A. Safo, “The ultimate Wright-Giemsa stain: 60 years in the making,” Biotech. Histochem. 86(2), 69–75 (2011). [CrossRef]
18. A. Ramoji, U. Neugebauer, T. Bocklitz, M. Foerster, M. Kiehntopf, M. Bauer, and J. Popp, “Toward a spectroscopic hemogram: Raman spectroscopic differentiation of the two most abundant leukocytes from peripheral blood,” Anal. Chem. 84(12), 5335–5342 (2012). [CrossRef]
19. C. G. Atkins, K. Buckley, M. W. Blades, and R. F. B. Turner, “Raman spectroscopy of blood and blood components,” Appl. Spectrosc. 71(5), 767–793 (2017). [CrossRef]
20. A. J. Walsh, K. P. Mueller, K. Tweed, I. Jones, C. M. Walsh, N. J. Piscopo, N. M. Niemi, D. J. Pagliarini, K. Saha, and M. C. Skala, “Classification of T-cell activation via autofluorescence lifetime imaging,” Nat. Biomed. Eng. 5(1), 77–88 (2021). [CrossRef]
21. B. P. Yakimov, M. A. Gogoleva, A. N. Semenov, S. A. Rodionov, M. V. Novoselova, A. V. Gayer, A. V. Kovalev, A. I. Bernakevich, V. V. Fadeev, A. G. Armaganov, V. P. Drachev, D. A. Gorin, M. E. Darvin, V. I. Shcheslavskiy, G. S. Budylin, A. V. Priezzhev, and E. A. Shirshin, “Label-free characterization of white blood cells using fluorescence lifetime imaging and flow-cytometry: Molecular heterogeneity and erythrophagocytosis [Invited],” Biomed. Opt. Express 10(8), 4220 (2019). [CrossRef]
22. F. A. B. da Silva, A. P. Racanelli, I. Lorand-Metze, and K. Metze, “Fluorescence lifetime imaging is able to recognize different hematopoietic precursors in unstained routine bone marrow films,” Cytometry A 99(6), 641–646 (2021). [CrossRef]
23. R. R. Jones, D. C. Hooper, L. Zhang, D. Wolverson, and V. K. Valev, “Raman techniques: fundamentals and frontiers,” Nanoscale Res. Lett. 14(1), 231 (2019). [CrossRef]
24. Q. Wang, J. Wang, M. Zhou, Q. Li, and Y. Wang, “Spectral-spatial feature-based neural network method for acute lymphoblastic leukemia cell identification via microscopic hyperspectral imaging technology,” Biomed. Opt. Express 8(6), 3017–3028 (2017). [CrossRef]
25. G. S. Verebes, M. Melchiorre, A. Garcia-Leis, C. Ferreri, C. Marzetti, and A. Torreggiani, “Hyperspectral enhanced dark field microscopy for imaging blood cells,” J. Biophotonics 6(11-12), 960–967 (2013). [CrossRef]
26. X. Liu, M. Zhou, S. Qiu, L. Sun, H. Liu, Q. Li, and Y. Wang, “Adaptive and automatic red blood cell counting method based on microscopic hyperspectral imaging technology,” J. Opt. 19(12), 124014 (2017). [CrossRef]
27. P. Zamora-Perez, D. Tsoutsi, R. Xu, and P. Rivera Gil, “Hyperspectral-enhanced dark field microscopy for single and collective nanoparticle characterization in biological environments,” Materials 11(2), 243 (2018). [CrossRef]
28. Y. Duan, J. Wang, M. Hu, M. Zhou, Q. Li, L. Sun, S. Qiu, and Y. Wang, “Leukocyte classification based on spatial and spectral features of microscopic hyperspectral images,” Opt. Laser Technol. 112, 530–538 (2019). [CrossRef]
29. A. Ojaghi, P. Casteleiro Costa, C. Caruso, W. A. Lam, and F. E. Robles, “Label-free automated neutropenia detection and grading using deep-ultraviolet microscopy,” Biomed. Opt. Express 12(10), 6115 (2021). [CrossRef]
30. N. Kaza, A. Ojaghi, and F. E. Robles, “Hemoglobin quantification in red blood cells via dry mass mapping based on UV absorption,” J. Biomed. Opt. 26(08), 086501 (2021). [CrossRef]
31. M. Mugnano, P. Memmolo, L. Miccio, F. Merola, V. Bianco, A. Bramanti, A. Gambale, R. Russo, I. Andolfo, A. Iolascon, and P. Ferraro, “Label-free optical marker for red-blood-cell phenotyping of inherited anemias,” Anal. Chem. 90(12), 7495–7501 (2018). [CrossRef]
32. D. Ryu, J. Kim, D. Lim, H.-S. Min, I. Y. Yoo, D. Cho, and Y. Park, “Label-free white blood cell classification using refractive index tomography and deep learning,” BME Front. 2021, 1–9 (2021). [CrossRef]
33. B. Javidi, A. Markman, S. Rawat, T. O’Connor, A. Anand, and B. Andemariam, “Sickle cell disease diagnosis based on spatio-temporal cell dynamics analysis using 3D printed shearing digital holographic microscopy,” Opt. Express 26(10), 13614–13627 (2018). [CrossRef]
34. M. Ugele, M. Weniger, M. Stanzel, M. Bassler, S. W. Krause, O. Friedrich, O. Hayden, and L. Richter, “Label-Free High-Throughput Leukemia Detection by Holographic Microscopy,” Adv. Sci. 5(12), 1800761 (2018). [CrossRef]
35. D. K. Hamilton and C. J. R. Sheppard, “Differential phase contrast in scanning optical microscopy,” J. Microsc. 133(1), 27–39 (1984). [CrossRef]
36. H. Lu, J. Chung, X. Ou, and C. Yang, “Quantitative phase imaging and complex field reconstruction by pupil modulation differential phase contrast,” Opt. Express 24(22), 25345 (2016). [CrossRef]
37. G. N. McKay, N. Mohan, and N. J. Durr, “Imaging human blood cells in vivo with oblique back-illumination capillaroscopy,” Biomed. Opt. Express 11(5), 2373 (2020). [CrossRef]
38. J. D. Giese, T. N. Ford, and J. Mertz, “Fast volumetric phase-gradient imaging in thick samples,” Opt. Express 22(1), 1152–1162 (2014). [CrossRef]
39. J. Li, Q. Chen, J. Zhang, Y. Zhang, L. Lu, and C. Zuo, “Efficient quantitative phase microscopy using programmable annular LED illumination,” Biomed. Opt. Express 8(10), 4687 (2017). [CrossRef]
40. P. C. Konda, L. Loetgering, K. C. Zhou, S. Xu, A. R. Harvey, and R. Horstmeyer, “Fourier ptychography: current applications and future promises,” Opt. Express 28(7), 9603 (2020). [CrossRef]
41. G. Zheng, R. Horstmeyer, and C. Yang, “Wide-field, high-resolution Fourier ptychographic microscopy,” Nat. Photonics 7(9), 739–745 (2013). [CrossRef]
42. P. M. S. Roma, L. Siman, B. Hissa, U. Agero, E. M. Braga, and O. N. Mesquita, “Profiling of individual human red blood cells under osmotic stress using defocusing microscopy,” J. Biomed. Opt. 21(9), 090505 (2016). [CrossRef]
43. P. M. S. Roma, L. Siman, F. T. Amaral, U. Agero, and O. N. Mesquita, “Total three-dimensional imaging of phase objects using defocusing microscopy: Application to red blood cells,” Appl. Phys. Lett. 104(25), 251107 (2014). [CrossRef]
44. N. Streibl, “Phase imaging by the transport equation of intensity,” Opt. Commun. 49(1), 6–10 (1984). [CrossRef]
45. C. Zuo, J. Li, J. Sun, Y. Fan, J. Zhang, L. Lu, R. Zhang, B. Wang, L. Huang, and Q. Chen, “Transport of intensity equation: a tutorial,” Opt. Lasers Eng. 135, 106187 (2020). [CrossRef]
46. M. Beleggia, M. A. Schofield, V. V. Volkov, and Y. Zhu, “On the transport of intensity technique for phase retrieval,” Ultramicroscopy 102(1), 37–49 (2004). [CrossRef]
47. S. Z. Lin, R. Chen, and S. F. Xu, “Absorbing spectrum of blood components in man,” Chin. J. Phys. Med. 20(1), 26–28 (1998).
48. S. Prahl, “Optical absorption of hemoglobin,” http://omlc.ogi.edu/spectra/hemoglobin.
49. Z. Xu, X. Baojie, and W. Guoxin, “Canny edge detection based on Open CV,” in 2017 13th IEEE International Conference on Electronic Measurement Instruments (ICEMI) (2017), pp. 53–56.
50. X. Lu, S. Zhang, H. Su, and Y. Chen, “Mutual information-based multimodal image registration using a novel joint histogram estimation,” Comput. Med. Imaging Graph. 32(3), 202–209 (2008). [CrossRef]
51. Ultralytics, “YOLOv5 in PyTorch,” GitHub (2020), http://github.com/ultralytics/yolov5.
52. A. Neubeck and L. Van Gool, “Efficient Non-Maximum Suppression,” in 18th International Conference on Pattern Recognition (ICPR’06) (2006), pp. 850–855.
53. P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” arXiv: 1611.07004 (2018).
54. S. Preibisch, S. Saalfeld, and P. Tomancak, “Globally optimal stitching of tiled 3D microscopic image acquisitions,” Bioinformatics 25(11), 1463–1465 (2009). [CrossRef]
55. D. Chen, “Analysis algorithms for BDPC images,” GitHub (2022), https://github.com/duang-maker.
