1. Introduction

We present a compact, low-cost deep-ultraviolet (UV) microscope for fast, label-free molecular analysis. This device features simple, inexpensive optics and custom stage control hardware for automated imaging. The microscope can be broadly applied for many imaging applications ranging from high-resolution single-cell molecular imaging [1–6] to automated scanning of tissue samples on slides [7,8]. Here we describe the novel microscope system in detail, characterize its performance, and focus on an application of the device for hematological analysis, which has profound clinical significance.

Specifically, the Complete Blood Count (CBC) is the most commonly performed medical test in the United States with over 45 million tests performed each year [9]. A CBC provides important clinical information including a 5-part differential white blood cell (WBC) count, red blood cell (RBC) count, platelet count, hematocrit (HCT), and hemoglobin (Hb) concentration. Deviations in normal cell counts, HCT, or Hb concentration can be indicative of a number of physiological disorders, including autoimmune diseases, various cancers, infections, and more [10,11], thus making the CBC a powerful and commonly used diagnostic tool. Moreover, for cancer patients undergoing chemotherapy or radiation therapy treatments, it is necessary to frequently monitor WBC counts (specifically neutrophil) and platelet counts via CBC to assess immune health during the course of treatments [12]. Deviations from normal counts can jeopardize cancer treatment course and outcomes, and even impose immediate life-threatening consequences.

Currently, CBCs are performed via hematology analyzers which use various techniques including electrical impedance measurements and flow cytometry to enumerate and characterize blood components [13]. While modern analyzers can perform fast, automated CBCs, they are expensive (∼$\$$100,000), bulky, require multiple biochemical reagents, and must be regularly calibrated. In addition, unusual CBC results typically require additional morphological analysis, which must be performed manually by a trained expert via microscopic analysis of stained blood samples. The complexity and high cost of hematological analyzers limits their availability to large hospitals or commercial laboratories that have the necessary resources and infrastructure; consequently, critical hematological information can take days to process, or requires patients to travel long distances to the few available clinics with a hematological analyzer. Thus, there is a significant clinical need for a point-of-care (POC) CBC analyzer that can be used in low resource settings or at home for fast and inexpensive hematology analysis. Several techniques have been proposed to improve hematology analysis, including fluorescence-based methods, quantitative phase imaging, hyperspectral imaging, and Raman microscopy [14–17]. While these techniques offer the potential for analysis in low-cost or label-free configurations, they have trade-offs between complexity, speed, and cost.

We have advanced deep-UV microscopy [1–4,6,7,18–20] to overcome the aforementioned limitations of CBC and other emerging technologies for hematological analysis. Deep-UV microscopy leverages the absorption properties of biomolecules in the deep-UV region of the spectrum (∼200-300 nm), generating high resolution, label-free, and quantitative molecular images [21]. UV microscopy has been historically overlooked due to phototoxicity concerns, but hardware advancements and proper management of UV exposure have enabled contiguous live cell imaging for several hours [1,2]. In previous work, we demonstrated a multispectral UV microscope that enables label-free imaging of unstained, live blood cells in a smear, along with (1) corresponding pseudo colorization schemes that accurately mimic the gold-standard Giemsa stained appearance of all blood cell types, (2) an automated WBC segmentation and classification framework, and (3) hemoglobin quantification using a single UV wavelength [4,6]. More recently, we have leveraged deep learning for virtual staining, automated segmentation and classification of single-channel, grayscale UV images acquired at 260 nm (corresponding to the peak absorption of nucleic acids), allowing for faster hematology analysis while reducing setup complexity [19]. However, these previous demonstrations of deep-UV microscopy have used bulky and expensive components including a laser-driven plasma light source, scientific grade UV-sensitive cameras, and commercial motorized stages, limiting their applicability in low-resource or POC settings [1–4].

The LED-based, low-cost and compact deep-UV microscope presented here is capable of serving as a POC CBC device, among many other potential applications in biology and medicine. We discuss our approach of using simple, inexpensive optics and hardware to design a compact and low-cost system. We also characterize stage performance with positioning repeatability, step linearity, and axis drift experiments. Furthermore, we demonstrate the system’s capability for hematology analysis by imaging fresh blood smears and performing automated WBC segmentation, classification, and colorization, enabled by our single-color neural networks [19]. Finally, we show how this system can be used to image blood in recently developed custom passive microfluidic devices [20], allowing for rapid and inexpensive hematology analysis from a 1 µL blood sample. Ultimately, we show that this low-cost UV microscopy system can be broadly used for simple and label-free molecular imaging applications.

2. Methods and results

2.1 System design

2.1.1 Optics

Low-cost system components were intentionally selected to facilitate translation of the system to a POC setting. While our previous demonstrations of deep-UV microscopy employed a broadband laser-driven plasma light source for sample illumination (>$\$$10,000), we now implement a single, narrowband 255 nm surface mount LED ($\$$6, Photon Wave Co., Ltd.) as shown in Fig. 1. The use of a single-color LED substantially reduces size and cost of the system, while also improving optical power (up to ∼10 mW at the sample plane compared to ∼0.2 mW in previous setups) and thus throughput/speed. Further, the use of a narrow band LED illumination eliminates the need for UV bandpass filters which are necessary when using a broadband light source, and are expensive and have low efficiency. The chosen wavelength of 255 nm yields images with high intracellular nuclear contrast due to the strong absorption of nucleic acids in the spectral region, and is sufficient to enable facile cell segmentation and classification, as well as AI-enabled virtual Giemsa staining [19].

 

Fig. 1. Compact UV microscope schematic comprising illumination components, stage translation hardware, and detection optics. The system microcontroller, LED driver, and motor drivers are housed directly underneath the system (wiring not shown). Experimental resolution inset shows image of high-resolution USAF test target (Newport) and line plots corresponding with 345 nm (green) and 308 nm (red) line pairs at center and periphery of FOV. System dimensions are approximately 6 × 6 × 10”.

Download Full Size | PDF

Between the LED and sample, we placed a UV fused silica ground glass diffuser (DGUV10-600, Thorlabs) which decreases the spatial coherence of light delivered to the sample and reduces background inhomogeneity that can arise from on-chip LED elements. This serves as a more compact and low-cost solution for illumination compared to a traditional Kohler illumination setup. Placing the LED and diffuser close to the sample (i.e., within 5 mm), eliminates the need for a condenser lens while still allowing for efficient light delivery with an effective illumination NA of 0.63, characterized by the largest angle of light rays illuminating the sample, calculated using the diffuser diameter and distance to sample. The high effective illumination NA, combined with NA of the objective (0.5) and short illuminating wavelength results in a spatial resolution of ∼308 nm (Table 1).

Table 1. Summary of Compact UV Microscope and Stage Capabilities

View Table

After passing through the sample, light is collected by a 40X, 0.5 NA UV microscope objective (LMU-40XUVB, Thorlabs) and focused onto the detector plane by a 75 mm focal length, biconvex UV fused silica tube lens (LB4330, Thorlabs). This tube lens allows for a compact system size without introducing severe optical aberrations. Experimental resolution was measured using a high-resolution USAF test target (Newport) at the center and periphery of the microscope field of view (FOV). The minimum separable line pair width at the FOV center was 308 nm (Fig. 1), which is approximately the diffraction limited resolution for this system. The measured experimental resolution at the FOV periphery was 345 nm, which is slightly worse due to coma and spherical aberrations observed. Corresponding Zemax spot diagrams are included in Supplemental Information (Fig. S1) which are in excellent agreement with the experimental results. Even at the periphery, this low-cost and compact microscope offers superior resolution to our previous, more expensive system, which has an experimental resolution of approximately 388 nm. The improvement is again due to the higher effective NA of the illumination light of the compact system.

Detection of deep-UV light can be challenging with conventional cameras due to coverglass layers, Bayer arrays on detector active areas, and IR filters found in front of the sensors, all of which absorb UV light. Therefore, commercial UV imaging solutions have been mostly limited to expensive sCMOS and CCD cameras. For this work, we use a modified de-bayered, 12MP SONY IMX477 detector (Maxmax, LLC) with the coverglass removed, following previously demonstrated procedures for de-bayering sensors [22,23]. This procedure does not affect the stability or durability of these sensors but does expose the active area, thus requiring care to not damage or introduce dust on the sensor. This sensor has a small pixel size of 1.55um, corresponding to ∼110 nm at the sample, which enables critical sampling even with the short focal length tube lens (and hence lower 13.5X magnification). Given the 12MP sensor, the field of view of the system is approximately 300 × 300µm. Background normalization is performed on the RAW sensor images to remove residual Bayer pattern artifacts and produce full resolution UV images of the sample at up to 15 frames/second. The use of a compact, on-chip sensor which costs around $\$$50 per module further reduces the cost of this UV microscopy setup, as compared to commercial scientific cameras in the $\$$10,000 range.

2.1.2 Hardware

Stage precision in this UV microscope had to be sufficient to scan a target sample region without overlap, or within a known range of overlap, to allow for accurate analysis. Previous demonstrations of UV microscopy used high-precision commercial stages and stage controllers costing upwards of $\$$20,000. To achieve accurate and inexpensive stage control, we use a stepper motor and micrometer configuration adapted from the OpenStage project [24]. The use of high-resolution motors in conjunction with low gear-ratio micrometers enables sub-micron step sizes in all three X, Y, and Z axes for a total cost of around $\$$1,000. The X/Y stage micrometers are coupled to their respective motors via custom machined adapters, which have bores corresponding with micrometer and motor spindle diameters and desired set screw taps. The Z-stage micrometer is coupled to its motor via a flexible driveshaft (Fig. 1). Our compact UV microscope features 0.9° step angle bipolar stepper motors (Vexta) mounted to an X/Y stage (T12XY, Thorlabs) and Z-axis translator (MT1, Thorlabs). This particular configuration allows for step sizes of 0.62µm at full, single steps corresponding with a stage translation speed of 2.5 mm/s. Using a technique called microstepping [25], the overall step size achievable with this system is ∼0.04µm, which outperforms many commercially available translation stages. The scan range in all three axes is 0.5 in. and can be extended by choice of stages.

The optical and stage components of our system are fully controlled by a Raspberry Pi microcontroller (Raspberry Pi 3 Model B+, Raspberry Pi). This controller precisely integrates illumination stage translation, and image capture to control LED output power and limit UV dosage to the sample. The use of the SONY IMX477 detector in a Raspberry PiCamera module allows for desired imaging exposure and frame rate to be fixed with the microcontroller for consistent imaging. The GPIO pins in the microcontroller also allow easy and fast interfacing with the LED driver (FemtoBuck LED Driver, Sparkfun) and motor drivers (A4988, Polulu). These electronics are housed directly underneath the compact UV system as they do not need to be modified after assembly (Fig. 1). Captured images are currently saved to a nearby PC for simultaneous image processing and analysis, which does not require high-end computing. Future iterations of this system can leverage cloud-based image saving and processing to eliminate the need for a nearby PC. Without the computing unit, this system has dimensions of 6 × 6x10”, which is well suited for POC applications. Table 1 summarized the capabilities of the system.

2.2 System characterization

2.2.1 Stage performance testing

To assess stage performance, a series of tests were conducted using a UV fused silica resolution target (UV FS USAF 1951, Edmund Optics) placed in a custom machined slide holder. First, a lateral positioning repeatability test was performed to assess the system’s ability to return to a starting point. This was achieved by capturing an image of a test pattern at an initial position, translating the stage 400µm in both X and Y axes, capturing another image at this secondary position, and then returning to the same initial position. This process was repeated 10 times to assess positioning repeatability. Results are illustrated in Fig. 2(A)-(B). The maximum measured offset in either axis was 2µm, much smaller than the system’s FOV of approximately 300 × 300 µm. Second, we performed a lateral step linearity test to measure the experimental step size as compared to the theoretical incremental motion of 0.62µm. As shown in Fig. 2(C)-(D), the step linearity is sufficiently close to the target position, with error less than 0.2µm over 20 steps. Next, a lateral stage drift test was conducted by capturing images of a stationary sample at 15-minute increments for 1 hour. Results (Fig. 2(E)) show a dual-axis drift of less than 0.4µm, indicating marginal stage drift over a time period much longer than the duration of samples imaged with this setup. Measurable drift is only observed after 15 mins. Lastly, a Z-axis positioning repeatability test was conducted to ensure that the setup could automatically return to focus position after, for example, moving the objective to change samples. During this test, a USAF test-target image was captured and the Z-stage was translated 100µm out of focus, then the stage was translated back to focus and repeated 10 times. Stage accuracy and focus were evaluated via a normalized image entropy calculation for each image, demonstrating maximum difference in normalized entropy of approximately 2.5% which did not appear to significantly affect image quality (i.e., all images in the in-focus plane were in sharp focus). These results show that the X/Y/Z axis stage solution employed in this device can accurately and consistently translate in all axes, with sufficient precision for most sample scanning and image capture applications.

 

Fig. 2. Characterization of stage performance. (A-B) X/Y axis positioning repeatability demonstrated by the position offset at a primary location and secondary location translated 400 µm in both axes. (C-D) X/Y axis step linearity demonstrated by translating 20 sequential single steps and comparing actual stage position at each step to the respective call position. (E) X/Y axis drift characterized by image capture at 15-minute intervals for 1 hour and measured stage offset at each time point. (F) Z axis positioning repeatability characterized by capturing a USAF test target image at the focal point and then translating the Z-stage 100 µm to defocus and capture a corresponding image. This was repeated 10 times. Image focus was quantified by normalized image entropy at each position by the maximum entropy of the experiment.

Download Full Size | PDF

2.3 Imaging and automated analysis of blood smears

To assess the compact system’s imaging performance for hematology applications, blood smear samples were imaged with an automated scanning and image capture protocol. Blood smearing, a technique performed by spreading a small volume of blood on a slide, produces a monolayer of blood cells near the tip of the smear that can be visually inspected to determine relative blood cell counts and cell morphology [26]. Typically, this procedure takes over 45 mins, due to the fixing and staining process, but the same information can be acquired by imaging the live cells immediately after smearing using label-free deep-UV microscopy. For each scan, 72 images were captured with approximately 20% overlap in an 4mm² region, previously demonstrated as large enough to obtain statistically significant counts of neutrophils, a WBC subtype [18]. The image presented in Fig. 3(A) shows a representative region (24 stitched images of 72) from one scan of a fresh blood smear used to quantify blood cells, demonstrating label-free contrast and clear distinctions between different cell types with 50 ms exposure. Features of interest such as the nuclear morphology and the granularity of the cytoplasm are clearly visible, thereby enabling automated segmentation, classification, and colorization. For example, the green and blue insets in Fig. 3(A). highlight neutrophils, characterized by their distinct nuclear lobes and granular cytoplasm. Similarly, the orange inset highlights a lymphocyte, characterized by its large, round nucleus and high nucleus-to-cytoplasm ratio. It is also worth highlighting that the image quality achieved with this compact system is superior then our previous, more expensive system [13]. This is a result of the low spatial coherence of the illumination light (LED with the diffuser) which completely eliminates coherent artifacts, such as ripples around the edges of cells that were previously observed [13].

 

Fig. 3. Blood smear and corresponding virtual colorized image. (A) Sample region of a stitched image of a blood smear demonstrating endogenous WBC nuclear contrast in neutrophils (green and blue squares) and lymphocyte (orange square). (B) Corresponding virtual Giemsa stained image generated with a single-channel colorization network developed for a previous UV microscope with a laser-driven plasma light source. Scale bar: 30 µm (stitched images), 10 µm (insets).

Download Full Size | PDF

These images were then virtually stained using a previously validated, pretrained deep neural network developed for our previous UV microscopy setup using a broadband laser-driven plasma light source [19]. First, images from the compact system were interpolated to match the effective pixel size of the images from the previous system. Additionally, a wavelet based denoising method [27] was used to improve the quality. The virtually stained image in Fig. 3(B) shows that our previously demonstrated network is able to recapitulate the gold-standard of Giemsa stained image appearance with this new compact system without any retraining, despite differences in illumination type, magnification, pixel size, and image quality. To validate these results, the blood smears were fixed in methanol and stained using a Giemsa-May Grunwald solution, commonly used to stain WBC nuclei and granules for hematology analysis [10]. The stained image (Fig. S2>) shows excellent agreement with the virtually stained image (Fig. 3(B)). Notably, cells with virtually stained nuclei are validated as nucleated cells in the stained sample. Furthermore, nuclear shape and granules in the virtually stained image correspond with those in the stained image demonstrating our ability to properly identify nuclear and other subcellular structures in WBCs that properly mimic standard stained images.

To further validate the performance of our compact system, we performed automated segmentation and classification of white blood cells from two representative blood smears and compared our results with manual counts from corresponding stained smears. The segmentation and classification pipeline was previously described in detail [19], and is implemented here without modifications. In short, a convolutional neural network is applied to segment the cell and cell nuclei, then the cropped and segmented images (green and purple images in Fig. 4(A)) are passed to the previously trained five-part WBC classifier [19]. Figure 4(A) shows grayscale images, corresponding cropped images generated after cellular and nuclear segmentation, and virtual Giemsa colorized images. The five-part classification yields 96.8% accuracy, correctly recognizing and classifying 121 of 125 WBCs visible in the monolayer regions of the two blood smears. The confusion matrix is shown in Fig. 4(B) and shows very high sensitivity and specificity for the neutrophils and lymphocytes, the most abundant WBC subtypes. The relatively low specificity and sensitivity for the other cell types (basophils and monocytes) is due to the very small number of these cell types present in the two smears. This performance is expected to improve with a larger data set, as we have previously shown [19]. Nevertheless, the overall high classification accuracy achieved with these preliminary data, using a compact, low-cost system rather than the more expensive system used to train the network, highlights the suitability of the low-cost, compact system for POC hematological analysis and the excellent generalization capabilities of our deep neural network. Further improvements in classification accuracy can also be expected if the deep neural network is retrained with more data from the compact system itself.

 

Fig. 4. Automated segmentation and classification of blood smear images. (A) Grayscale images (top row), segmented images (middle row), and the virtual Giemsa colorized images (bottom row). Scale bar: 20 µm. (B) Confusion matrix for five-part WBC classification.

Download Full Size | PDF

2.4 Absolute neutrophil counts using custom passive-microfluidic devices

As a final demonstration of the capability of the low-cost, compact UV system, we apply a recently developed custom passive microfluidic device to produce a monolayer of blood cells from a 1 µL blood volume [20]. The device mimics a blood smear without requiring a technician to perform sample preparation, while also controlling for the total sampled volume and enabling quantitative absolute cell counts [20]. This type of microfluidic device has been used with previous iterations of our UV microscope to enumerate WBC subtypes for disease diagnosis. For example, an abnormally low Absolute Neutrophil Count (ANC), can indicate neutropenia, a disease characterized by a neutrophil count <1500 neutrophils per µL of whole blood [28]. Here, we collected a fingerprick blood sample from five healthy individuals and measured the ANC using both a commercial CBC analyzer, the clinical gold standard, and an automated scan of the microfluidic device with the developed compact UV microscope. The dimensions of the monolayer region within each channel are approximately 4 × 1 mm, identical to the target region previously imaged on blood smears. Figure 5(A) shows a representative imaged region of whole blood in a microfluidic channel, with distinct nuclear contrast of neutrophils (green arrows) and lymphocytes (blue arrows). The ANC (absolute neutrophil count per µL) is experimentally determined using the number of neutrophils observed in the imaged region (visually assessed) and an estimate of total imaged blood volume via channel height and area measurement. A comparison of the ANC measurements from a commercial CBC and the compact, low-cost UV system reveals a strong linear correlation, with an R² coefficient of ∼0.97 (Fig. 5(B)). While the sample size here is low (n = 5), we have previously demonstrated excellent correlation with larger data sets between ANCs from a commercial analyzer and our microfluidic devices using previous iterations of our UV microscope [20]. As this compact UV system performs similarly to our benchtop microscopes, we expect to see a similar correlation with larger sample sizes. This shows that the low-cost, compact UV microscope system has excellent agreement with a commercial analyzer for identification of blood cell subtypes (specifically neutrophils), but with a much smaller footprint and at a fraction of the cost.

 

Fig. 5. Absolute Neutrophil Counting using custom passive microfluidic devices. (A) Sample image of a fingerprick blood sample in a custom microfluidic device. Arrows indicate neutrophils (green), and lymphocytes (blue) present in the FOV. Scale bar: 25 µm. (B) Comparison of the Absolute Neutrophil Count (neutrophils per µL of blood) of 5 donors using a commercial analyzer and the compact UV microscope.

Download Full Size | PDF

3. Discussion

In this work, we have developed a compact, low-cost UV microscope that can be broadly applied for label-free, molecular imaging applications. The total cost of this system is less than $\$$5,000, roughly a tenth of the cost of previous iterations of UV microscopy. We evaluated the performance of the low-cost 3D translation stages and showed that the positioning accuracy of our inexpensive solution parallels commercial setups and exceeds the requirements for our applications. We demonstrated imaging capability for hematological applications by imaging targeted regions of blood smears (monolayer regions) and showed that neural networks trained on data from previous iterations of UV microscopy systems could be used for automatic segmentation, classification, and colorization of blood cells. We then imaged blood samples using a custom passive microfluidic device and showed that an ANC analysis could be performed with a 1 µL of blood in approximately three minutes, eliminating the need for manual blood smears. We highlight that this compact, low-cost imaging system has improved imaging quality (no coherent artifacts and better resolution) compared to our previous bulkier, more expensive system that we have used to demonstrate the ability of deep UV microscopy as a promising tool for hematology analysis. Thus, the system’s capabilities and embodiment – specifically, compactness (6 × 6x10”), low cost, speed, image performance, and compatibility with fingerprick-based blood collection – are ideal for POC CBC analysis. The imaging time can be further reduced with lower-resolution stepper motors resulting in faster stage translation. Future iterations of this device can use lower magnification objective lenses to increase the FOV, thus reducing the number of images needed to capture the 4 × 1mm² area, and thus total imaging time. The total processing time is under two minutes and can be performed simultaneously with image acquisition. Processing can be further optimized using in-line field programmable gate arrays or cloud-based processing networks, eliminating the need for a complementary PC.

In this work we focused on hematology analysis, but the proposed UV microscope can be applied to any molecular imaging application that leverages nuclear contrast. This can include analysis of unstained histopathology samples [7] and bone marrow aspirates [4], both of which have been previously demonstrated with a benchtop UV microscopy system. Moreover, multispectral UV microscopy can provide label-free contrast from other relevant biomolecules [21], including proteins, collagen, elastin, cytochrome, melanin, and lipids. This compact, low- cost system can be readily applied to a wide range of applications in biology and medicine.