Lensless On-Chip-Hemocytometry With Secondary Processing Of Cell Images in the Framework of an Unconventional Photometric Model
1.1. Foreign work in the field of lensless on-chip cytometry
The first foreign work in the field of lensless microscopy of blood and lensless cytometry dates back to 2008, when the basic and widely available technologies of lensless on-chip cytometry were annotated and later tested on reference samples. They were first tested using normal erythrocytes, yeast cells and calibration microspheres for metrological calibration , and then applied for a simple morphometric analysis of blood, which isolated red blood cells, white blood cells and platelets . These methods were based on the use of standard types of CMOS image sensors. Such standard photosensitive elements traditionally used in most digital cameras, video and Web cameras, have red – R, green – G and blue – B channels of color image registration. In a standard digital photograph, the recorded signal is an integral characteristic collected from several pixels. In the case of contact work with blood cells per pixel (at one point of the image) only one color is matched, therefore a significant technological achievement was the transition from the usual wide-field lensless monitoring using LUCAS (Lensless Ultrawidefield Cell [monitoring] Array for Shadow [imaging]) to multispectral / color monitoring (Multi – Color LUCAS) .
Since it is difficult to develop a metrological certification method for a fluorescence cytometric method that is implemented on a chip (CMOS image sensor), the efficiency of analysis, counting and classification of cells using this method is low. The reason lies in the fact that almost all cell types are displayed by the time integrated signal in the form of a Gaussian peak of diffuse luminescence, and the task of separating emission and excitation beam on a chip is difficult and almost unbridgeable for the users without special background in physics. Therefore, the problem of metrological certification of the fluorescent cytometric method has not yet been solved . From the point of view of ensuring traceability to the standard (metrological assurance of the uniformity of measurements), the method of differential interference contrast (DIC) on a chip gives a smaller methodological error. To apply this method, it is required to fine-tune the uniaxial crystal and the absorbing analyzer , and the method produces pseudo-color images (due to the properties of DIC, equivalent to the use of Nomarsky or Pluto optics). In the holographic case using non-coherent sources, digital processing of holograms obtained using DIC reconstructs the differential phase-contrast image of the sample in a wide field of view (24 mm2).
Another way to improve metrological characteristics of analytics in lensless holographic blood cytometry is the algorithmic (i. e., implemented using computer programs for image processing) increase in resolution – the method of sub-pixel super-resolution . Until recently, resolvometric properties and the complexity of processing information on holographic on-chip cytometers prevented their introduction into laboratory diagnostic and clinical practice.
From the point of view of diagnostics, the precedents of hematological introduction of lensless holographic on-chip cytometry are of a piece nature, mainly limited to the field of medical parasitology at the level of individual non-clinical self-diagnosis in third world countries and developing countries. Thus, compact devices of super-resolving holographic on-chip lensless microscopy were tested for the automated diagnosis of malaria . The accuracy of the diagnosis of this disease (i. e., the identification of Plasmodium falciparum, and P. vivax using a certain database) is automatically indicated by the use of crowdsourcing platforms and identification of blood cells without verifying the results of the analysis, since the systematization of the material in the case of building a database for the detection of malaria plasmodia was carried out by non-specialists in the order of the "network game" . Obviously, and it goes without saying that with any metrological accuracy of a holographic lensless cytometer as an optical device, this method, when using incorrectly validated morphometry databases, cannot be accepted by physicians for incorporation into diagnostic technology.
Probably, this circumstance gave rise to the rejection of the method among the specialists involved in the development of holographic cytometers. Therefore, already in 2013, they switched from the idea of lensless cytometry to the ideas of democratization of hematological analysis in general, including on camera phones, which had already appeared by that time . The main advantage of the lensless cytometry method was the compactness of the systems due to the lack of an optical path on the scanning line and the absence of lenses. This, in turn, determined the economic efficiency of the method. At the same time, the efficiency or robustness (reliability) of such diagnostic systems is no longer achieved due to the physical simplicity of the lensless contact cytometer, but due to the use of powerful algorithms for in-depth training of neural networks responsible for making diagnostic decisions . But in the process of phase reconstruction and reconstruction of the holographic image, neural networks  are also used, whose machine learning techniques take a lot of time. This, in turn, may become another obstacle to the use of lensless on-chip cytometry technology in hematology.
The history of lensless microscopy in hematology is a classic example of biomedical engineering, which was not adequately accepted by medical professionals. Because of this, the method yielded a much lower heuristic / diagnostic value than would be the case if the priorities were set correctly. If we consider the product of medical use as the final result of biomedical engineering, it is quite obvious that the emphasis of measurements should be placed in the field of diagnostic qualimetry rather than metrology; in the field of not so much engineering as hematology. Groups of researchers who worked on this topic with a greater emphasis in the biomedical area, by definition, are more focused on medical practice and clinical approbation of the system than on permanent improvement of metrological parameters when creating a recording device.
1.2. Domestic development of lensless on-chip hemocytometers
In Russia, work on lensless on-chip cytometry has been going on since 2009, if we consider the development of devices that are applicable in hematology. But experiments on phase correlometric analysis of the signal of biomedical objects from charge-coupled devices and CMOS image sensors have been conducted since 2007 by groups of authors of this article. The groups obtained the results of experiments using both laser and non-coherent radiation sources. A number of works carried out earlier in the framework of the Russian collaboration (institutes of the Russian Academy of Medical Sciences, departments of higher educational institutions in physiological and anatomical specialties), have led to a number of key results that have not yet been achieved by foreign teams. The basic list of results includes the following works:
1. The applicability of lensless speckle-interferometric, lensless speckle imaging analysis is shown not only for measuring and identifying mature blood cells and obtaining, as a final result, an extended blood leukocyte formula, but also for analyzing proliferative activity and differentiation (for bone marrow cells and on model cultures) . In parallel, within the framework of initiative research, adequate approximations were found for the morphology of different types of cells (Poincaré and Henon, Chirikov and Gumovsky-Mira, Kepler, Volterra, Dufing and Henon-Heiles) . The result of this work was the creation of on-chip cytometers (hemocytometers) in 2009–2011 with a microfluidic chip in the format of chambers for counting blood and cerebrospinal fluid cells, using the grids of which monitoring of the resolution of the reading chips ("on-chip resolvometry") is performed [16, 17].
2. For automated morphometry and cell classification in lensless cytometry, one of the authors of the articles [16–18] (Notchenko A. V.) created a program of deep learning. Its work has been tested both on stained blood samples and on nerve tissue sections. The compatibility of the multi-axis system of robotized sample positioning on Fedorov’s automated table (which determines in part the metrological advantage of our development with respect to foreign analogues) and the cell scanning system in laboratories on a chip (in particular, blood cells) was ensured. In the articles [19, 20] a three-dimensional reconstruction of blood cells is performed (see Fig. 10 in these works and the corresponding Fig. 2 in this article), which is absent in the works of foreign peers. Such visualization can be used not only for cytometry purposes, but also for aggregometry and coagulometry. In other words, reconstruction by slices (i. e., optical slices) could be performed not only in two-dimensional, but also in three-dimensional form, as well as identification of cell types. Fig. 1, 2 explain this approach and the results of its application to blood cells (visualization of the aggregation of red blood cells).
3. On membrane-mimetic model structures with a specific for morphometric determination of hemolytic anemia and adjacent states morphology, the applicability of the algorithms and laser diagnostics of membranopathies (on models of microspherocytes, elliptocytes, stomatocytes, echinocytes, acanthocytes) is demonstrated . Unfortunately, due to financial and technical limitations for validation, which led to incorrect interpretation of some results obtained, this article can be considered as partially erroneous and cannot be used to identify cell types until our next works are published.
4. A scheme for mapping non-optical characteristics of cells / non-optical labels in a lensless on-chip cytometer is proposed. It is based on the use of signal converters of hybridized agents with a non-optical response to an optical signal with positional sensitivity (hybridisable agents are immunomagnetic particles, radioisotope labels, sensitizers used as agents of photodynamic diagnostics and photodynamic therapy, which are excited by radiation sources beyond the visible spectral range; etc.). That is, binding to specific cells takes place and is identified with specific classes of cells in a database updated by machine learning techniques with a teacher [22,23]. In particular, in the recent past, a program was developed for measurements on hemocytoblasts (under experimental conditions) . A similar program is annotated for work in the field of lymphology .
The purpose of this work is to demonstrate the heuristic value of the results obtained from the domestic design of lensless on-chip cytometers. The development allows not to use either holographic technologies at the stage of collecting primary data, or to use multi-angle scanning from a planar image sensor for data collection. The article demonstrates the methods / algorithms for processing data from lensless on-chip cytometers. Their use will allow a hematologist not to waste time on a preliminary analysis of a hologram using a complex mathematical apparatus.
2. MATERIALS AND METHODS
2.1. Design of a multi-angle on-chip hemocytometer
The hemocytometer is built on a lensless CMOS image sensor (for fixing sample images) with a radiation source positioned at programmed angles / in different coordinate systems. As a radiation source, we used either spectral LEDs (red for AlGaAs, GaP and AlGaInP; green for InGaN, AlGaP and GaN; blue for ZnSe, InGaN and SiC), or laser diodes, or DPSS (diode-pumping solid-state) lasers. The radiation of a laser or an incoherent source was projected with a given step increment on the image sensor on which the sample was localized (a drop of blood under supported conditions).
The described system uses the CMOS sensor of the iPhone 4S with a pixel size of 1.4 microns. A stepper motor was used to position the radiation sources (LED / DPSS) at different angles. It is controlled via the Arduino hardware and software platform, which accepts the commands of G-CODE, a programming language for devices with numerical control (Electronic Industries Alliance, RS274D standard; also, ISO 7-bit / ISO 6983–1 : 2009 or GOST 20999-83 standards). The software of the device allowing to save data and transfer it to a computer – MKIII version of 645 PRO application for registering high-quality lensless illustration in a lossless compression format (in particular, Lempel-Ziv and Lempel-Ziv-Welch algorithms for TIFF).
2.2. Materials and samples
A conditionally healthy anonymous donor’s blood was used, the analyzes of which, according to data from an independent laboratory, did not show any pathologies that could affect the measurement result. Heparinization and citrate buffer were used only for some calibration samples with deliberately uninformative diagnostic results. In most of the samples, the formation of microstructures was also investigated (especially when the sample was dehydrated). Such samples were usually not subjected to any preliminary sample preparation. The sampling was carried out with a lancet-scarifier. Since the time from sampling before analysis did not exceed, as a rule, several minutes (considering the exposure of the sample already on the chip, sometimes it lasted longer, up to several tens of minutes), blood preservation or other preliminary manipulations were not required. Blood was injected onto a chip (lensless array) in a volume of not more than one drop (from 5 to 130 µl, depending on what the duration of the device operation in the recording mode changed).
2.3. Algorithms for data processing and software
After data was recorded by the MKIII version of 645 PRO application, the results were imported into a format compatible with the analyzing software (with partial changes in the quality and size of the file). To reconstruct the three-dimensional structure of blood cell samples, the usual formulas for photometric calculations were not used (the Bouguer-Lambert-Beer law). For these purposes, a photometric model of Lommel-Seeliger was adopted. This model is used mainly in astrophotometry, geodesic and astrogeological measurements. It consists of an exponential term and a fourth-order polynomial in the phase function. At the same time, isophotes of cells were superimposed on the optical density map / cell reconstruction when using the schlieren-method (shadow method) in the pseudo-three-dimensional representation, i. e., the isolines of their local optical characteristics.
The resulting image with isolines of optical characteristics correlated to cells was studied in real time (without discontinuing the measurement process, in situ; or directly from the screen) using QAVIS, a software developed in the Far Eastern Branch of the Russian Academy of Sciences. QAVIS allows you to analyze images with obtaining correlograms, waveletgrams, correlation and spectral characteristics with the identification of individual cells / their types according to the result of identification of correlation and spectral characteristics – integral frequency characteristics (IFC) and integrated spatial characteristics (ISC). At the same time, the user doctor should not understand the mathematical aspects of diagnostics, since the friendly GUI user interface provides information (with the ability to save to a file) at an intuitively understandable output level.
3.1. Registration results
Figure 3 shows the results of registering a dehydrating blood drop with the size more than 1 / 16 of the chip surface. Registration was carried out using the shadow (schlieren-) method for constructing the relief, the photometric model of Lommel-Seeliger with the settings at 45° and 135°. Figure 4 shows a pseudo-three-dimensional representation of cells from the same sample with an arbitrary scale of heights and isophots overlays. Based on these images, one can see that without resorting to the holographic technique, as well as optical tomography, using lensless recording, LED and an algorithm built on the Lommel-Seeliger model, one can get high-quality graphic / photometric representations of blood cells. These images were obtained without using either phase reconstruction or wavefront analysis methods. The level of interference artifacts around the cells with this approach is much lower than when using standard UCLA methods (A. Ozkan group, see above), used for reconstruction of lensless holograms.
3.2. Detection of erythrocyte aggregation
During storage of preserved blood, or as a result of polyglobulia and various pathological and compensatory processes, with a change in ESR, the aggregation of erythrocytes occurs with a pronounced formation of stacks of erythrocytes (in the so-called "coin columns", or in other forms). This process (known as "rouleaux" in English language literature) can be observed on a cytometric lensless device, and can be detected by changing blood characteristics using the on-chip cytometer described above. Thus, in the early stages of dehydration of a blood drop (the first ten minutes; depending on the volume / thickness of the layer), it is possible to see the formation of these structures in situ on the chip in real time. The result of the corresponding observations is shown in Fig. 5. At the same time, it is possible with the kinetic quality (in dynamics and with positional sensitivity) to determine the aggregometric properties of the sample using the waveletgram of the corresponding cytometric patterns / files of registrograms from a lensless chip. The doctor does not have to own the data analysis technique, since the statistical results of the analysis of samples of blood cells passing through the analyzer can be understood from the simple morphological positions, with the exception that the subject of analysis and interpretation of the morphological picture is not micrographs, but waveletgrams. Fig. 6 shows the result of such an automatic distinction between the levels and stages of aggregation for different morphological types of pictures determined by the clinician: a is the norm, b is partial aggregation (which can be characterized as a "chain"), c is a complete aggregation ("stack").
3.3. Correlographic determination
To more accurately determine the classes of cells without using for this purpose standard morphometric databases, it is rational to have strict quantitative criteria for the differences of blood cells according to the analysis of their physical morphological properties. We have shown that the blood cells have a fairly small range of morphophysical parameters of normal forms, relatively to the deviant ones, therefore, correlation-spectral or integral frequency (and integral spatial – to a much lesser extent) analysis methods are ineffective for processing of images from a lensless chip. The reason lies in the fact that a single cell occupies a small number of pixels, which leads to a poor statistical reliability of the analysis. However, the correlographic approach as such worked much better. Therefore, it turned out to be more preferable as a method for analyzing not only normal, but also deviant forms of blood cells, including membranopathies. Fig. 7a below shows a normal erythrocyte (with contrast inversion), and at the top is its correlogram, with clearly defined details of contrast. In the case of analyzing an amorphous, from the point of view of the analyzer chip, agglomerate with sufficient optical density (Fig. 7b – bottom), the correlogram looks like a partially ordered structure with a large fraction of the diffuse "cloud" (Fig. 7b – top). When red blood cells are detected with a changed proportion and a clear deviation from the spherical shape, the correlogram changes in accordance with the polarity of orientation of this cell in the image (Fig. 7c). A similar effect is also observed when several cells that are not normalized in size (for example, two microspherocytes) stick together. When a "stack" begins to form in the region of interest (ROI) of the analyzer (but not yet a column that can contain more than a dozen red blood cells), the picture of the correlogram changes qualitatively, showing the spots created by the emergence of new structures in the ROI (Fig. 7d). If the erythrocyte shadow is fixed in the ROI, then its correlogram has significantly less pronounced spots-poles while maintaining the central zone of the same structure as that of the normal erythrocyte (Fig. 7e). When two or more erythrocytes are superimposed on the frame (without additional adhesion / aggregation), it is possible to detect this by strengthening specific zones of spots (see Fig. 7f), which are also evident for a single normal erythrocyte (Fig. 7a). Moreover, the result of the projection will also be influenced by the mutual orientation of red blood cells, as well as their stratification on the frame (underlying and overlying cell).
Thus, it is shown that methods of lensless on-chip hemocytometry, implemented in the framework of the simplified Russian model of the device and the alternative Lommel-Seeliger photometric model, allow not only to visualize blood cells, specifically in the quasi-three-dimensional version, but also morphologically classify them using wavelet and correlographic techniques. At present, they are being tested on the cell samples with various membranopathies.
The authors express gratitude to the Russian Foundation for Basic Research (project 16-32-00914) for the financial support.