Medical analysis device with impedance signal processing function

Abstract

A medical analysis device with cell impedance signal processing functionality, the medical analysis device comprising a memory (4) arranged to receive pulse data sets, each pulse data set comprising impedance value data each time associated with a time marker, the data together representing a cell impedance value curve measured when a cell passes through a polarization opening. The device further comprises a classifier (6) comprising a convolutional neural network receiving pulse data sets as input, and said convolutional neural network is provided with at least one convolutional layer having a depth greater than or equal to 3, and at least two fully connected layers, and further an output layer presenting a cell classification, from which output layer the pulse data sets are derived.

Description

Technical Field

[0001] This invention relates to the field of hematology, and more particularly to cell counting and sorting equipment. Background Technology

[0002] Since the 1950s, the counting and volume determination of various blood cells in hematology analyzers have been performed by measuring impedance using a method known as the Coulter principle. This method involves passing cells suspended in a conductive liquid through polarized micro-orifices and detecting the change in resistance (or impedance) caused by the passage of particles through the orifices. The detection of the resulting pulses allows for element counting.

[0003] Different solutions have been developed to account for issues related to flow within the orifice (due to edge or hydrodynamic effects, rotation caused by double masking, etc.). Since these solutions are unsatisfactory, hydrodynamic focusing techniques, or hydrodynamic focusing techniques, have been developed. This solution involves executing a hydrodynamic sheath over the cell flow to be analyzed, which centers the cell flow within the orifice and limits the effects associated with flow through the edges. However, this technique is very complex to implement and particularly expensive.

[0004] Recently, the applicants developed a solution that provides very satisfactory results while still overcoming affranchissant hydrodynamic focusing. They have protected this solution in patent application FR 1904410. During this development, the applicants realized that their research could also be used to characterize cells based on their impedance signals, thereby returning information about normal or abnormal cells or characterizing the morphology of these cells.

[0005] Recent advancements in existing techniques for measuring red blood cell deformability have revealed two main families.

[0006] The first family involves precise but time-consuming and complex measurement techniques. For some of these, the analytical models can be traced back to rheological parameters such as elasticity and membrane viscosity modulus. Among these methods, aspiration via a micropipette can be mentioned, which works by applying a known low pressure to aspirate a portion of red blood cells from the pipette. By measuring certain quantities related to the morphology of the red blood cells after aspiration, the shear modulus or shear viscosity can be inferred (see, for example, EAEvans' article "New membrane concept applied to the analysis of fluid shear and micropipette deformed red blood cells," Biophysical Journal, 1973). Other methods, such as those described below, include using optical lasers to stretch blood cells and observe their shape: for example, the “optical tweezer” method (see, for example, Brandao et al., “Optical tweezers for measuring red blood cell elasticity: application to the study of drug response in sickle cell disease”, European Journal of Haematology, 70:207-211, 2003) or the “optical stretcher” method (see, for example, Guck et al., “The optical stretcher: a novel laser tool to micromanipulate cells”, Biophysical Journal, 81:767-784, 2001).

[0007] The second family includes techniques that can process large numbers of cells more quickly and autonomously, thus providing a statistical understanding of the deformability of a sample of red blood cells. These techniques use a shape parameter called the Deformability Index (DI) to study red blood cell deformation; it is essentially a measure of the stretching of red blood cells after being subjected to known mechanical stress. The Deformability Index combines the mechanical and morphological parameters of red blood cells: therefore, it is simpler to study, but does not provide precise rheological information. Among these techniques, the method described in Cha et al.'s article "Measurement of Cell Stretching Using Viscoelastic Particle Focusing" (Analytical Chemistry, 2012) can be retained, where red blood cells are stretched in a chain in a stretching flow and observed using a camera. Based on the same principle, but by applying stress to red blood cells in a shear flow, the method described in Dobbe et al.'s article "Analyzing Red Blood Cell-Deformability Distributions" (Blood Cells, Molecules, and Diseases, 28:373-384, 2002) allows for showing the effect of certain diseases on the distribution of the red blood cell Deformability Index. The article "Analysis of factors regulating erythrocyte deformability" by Mohandas et al. (Journal of Clinical Investigations, 66:563-57, 1980) describes a laser diffraction method in which stress is applied to erythrocytes in a shear flow and the diffraction spectrum against them is observed to measure the deformability index. By studying the curves of the deformability index as a function of the osmotic pressure of the suspension medium (see, for example, Clark et al., "Osmotic Gradient Laser Diffraction: A Comprehensive Characterization of Erythrocyte Volume and Surface Maintenance," Blood, 61:899-910, 1983), it was shown that certain rheological and / or morphological parameters can be measured.

[0008] While providing richer information, the first family of methods cannot perform high-speed hematological analyses. In fact, they are too lengthy and complex to implement and require the intervention of specialized manipulators. On the other hand, while the second family of methods are easier to implement, industrializing them within the scope of medical analysis laboratories remains complex, although studying their more global response to the deformability index makes it possible to isolate pathological red blood cell subsets.

[0009] All the methods described above require optical acquisition systems or video microscopes. This is clearly more complex and expensive to implement than impedance measurement systems. However, counters operating on the Coulter principle cannot determine the deformation index or a simplified version of it.

[0010] Therefore, there is a need to provide a simple measuring device that can distinguish cell populations based on their morphological characteristics. Summary of the Invention

[0011] This invention improves upon this situation. To this end, the present invention proposes a medical analysis device with cellular impedance signal processing capabilities, comprising a memory arranged to receive pulse datasets, each pulse dataset including impedance value data associated with a time stamp each time, which collectively represent a cellular impedance curve measured as a cell passes through a polarization opening. The device further includes a classifier comprising a convolutional neural network receiving the pulse datasets as input; and the convolutional neural network having at least one convolutional layer having a depth greater than or equal to 3, at least two fully connected layers, and an output layer presenting cell classifications from which the pulse datasets are derived.

[0012] This device is particularly advantageous because it can perform cell classification using simple measuring equipment based on cell impedance measurements in a Coulter principle hematology counter. Furthermore, existing hematology counters can be modified to incorporate the advantages of this invention, thus avoiding the need to purchase new equipment.

[0013] According to various embodiments, the present invention may have one or more of the following features:

[0014] - A convolutional neural network consists of two convolutional layers, one of which is connected to the input layer that receives the pulse dataset.

[0015] - The fully connected layer consists of 4 layers of neurons.

[0016] - The activation function for all layers of the classifier is the sigmoid function.

[0017] The pulse dataset received in the memory was obtained from impedance measurements taken when red blood cells passed through the polarization opening.

[0018] - The output layer returns a value indicating whether the cell associated with a given pulse is normal or abnormal.

[0019] - The output layer returns a triplet that indicates the morphological characteristics of the cell.

[0020] - The output layer returns a value indicating whether the cell associated with a given pulse indicates the presence of a disease (such as malaria or sickle cell disease), and

[0021] - The output layer returns a value that indicates information about how the samples from which the pulse dataset is derived have changed over time.

[0022] This invention also relates to a method for classifying blood samples, the method comprising the following operations:

[0023] a) Receive pulse datasets, each pulse dataset including impedance data associated with each time stamp, which together represent a curve of cell impedance values ​​measured as the cell passes through the polarization opening, where the cell is derived from a blood sample.

[0024] b) For each pulse dataset

[0025] b1. Determine the maximum impedance value of the pulse dataset.

[0026] b2. The upper impedance value is calculated by multiplying the maximum impedance value by an upper coefficient selected in the range [0.7; 0.95], by determining the time marker in the pulse dataset where the relevant impedance value in the pulse dataset equals the upper impedance value, and by calculating the upper duration corresponding to the maximum duration between these time markers; and the lower impedance value is calculated by multiplying the maximum impedance value by a lower coefficient selected in the range [0.1; 0.6], by determining the time marker in the pulse dataset where the relevant impedance value in the pulse dataset equals the lower impedance value, and by calculating the lower duration corresponding to the maximum duration between these time markers.

[0027] b3. Calculate the peak position value, which is equal to the difference between the moment associated with the maximum impedance value and the first moment corresponding to the lower impedance value, divided by the following duration. Optionally, calculate the rotation value, which is equal to the upper duration divided by the following duration.

[0028] c) Determine the statistical distribution of the pulse dataset based on the duration / peak position value pairs or rotation value / peak position value pairs, wherein the statistical data is established for a range of values ​​of the pairs.

[0029] d) The blood sample is classified by comparing the distribution of operation c) with a reference distribution of a healthy blood sample. Attached Figure Description

[0030] Other features and advantages of the invention will become better apparent as one reads the following description, which is extracted from examples given in a non-limiting manner for informational purposes and from the accompanying drawings, wherein the drawings are:

[0031] - Figure 1 A top view of the measuring aperture within the scope of this invention is shown, along with the trajectory that cells can take therein.

[0032] - Figure 2 Showing the target Figure 1 The pulse measurement value obtained from the trajectory,

[0033] - Figure 3 and Figure 4 It shows the use of Figure 1 The characteristic diagram of the red blood cell impedance pulse measured by the arrangement,

[0034] - Figure 5 Feature maps of the region of interest are shown.

[0035] - Figures 6 to 9 It shows Figure 5 The proportion of pulses in some regions of interest,

[0036] - Figure 10 A diagram illustrating one embodiment of the device according to the present invention is shown.

[0037] - Figure 11 It shows in Figure 10 The diagram of the neural network implemented in the implementation scheme.

[0038] - Figure 12 A diagram of a neural network implemented in another embodiment of the invention is shown.

[0039] - Figure 13 Feature maps of the region of interest are shown.

[0040] - Figures 14 to 16 It shows Figure 13 A schematic diagram of pulse population statistics in some regions of interest, and

[0041] - Figure 17 The analysis of two healthy blood samples over an 11-day period is shown, and

[0042] - Figure 18 It shows Figure 17 The statistical change of the measured value over time corresponds to Figure 5 3' of a square. Detailed Implementation

[0043] The accompanying drawings and description below essentially contain elements with definite properties. Therefore, they can not only be used to better understand the invention, but also help to define it where applicable.

[0044] Figure 1 A top view of a measurement orifice within the scope of this invention is shown, along with the trajectory that a cell can take therein. The orifice has walls indicated by dashed lines, and the x and y coordinates are expressed in μm.

[0045] Figure 2 Showing the target Figure 1 The impedance pulses measured for each trajectory are extracted from signals corresponding to changes in system impedance caused by the passage of a cell through a micropore. The x-axis is expressed in μs, and the y-axis in ohms. It can be seen that the closer the cell's incident trajectory is to one of the walls of the pore opening, the more disordered the measurements become, and this is the cause of the errors described in the introduction.

[0046] The applicant's research, disclosed in application FR 1904410, has led them to develop a new numerical value for representing impedance pulses. This value, called WR, is the ratio of two pulse widths. These widths can indicate the presence of a peak in the pulse, or conversely, the presence of a bell-shaped pulse.

[0047] Therefore, the maximum pulse height in the pulse dataset is first determined. This maximum height is used to calculate the upper and lower impedance values.

[0048] The upper impedance value is obtained by multiplying the maximum impedance value (which corresponds to the maximum height) by an upper coefficient. This upper coefficient is used to determine two moments, which typically allows for an accurate estimation of the width of the pulse impedance peak. For this purpose, the upper coefficient is selected in the range [0.7; 0.95], preferably [0.8; 0.9], which ensures that there are at least two moments, and that these moments actually correspond to the peak value of the pulse (to limit the possibility of multiple peaks).

[0049] Therefore, the upper impedance value is less than the maximum impedance value and greater than 70% of the maximum impedance value. The applicant's research shows that this range enables accurate capture of the peak value of the generated pulse. The applicant found that a value of 0.875 is particularly advantageous and produces optimal results: enabling the estimation of the pulse peak value in the most accurate way possible. In practice, the peak value near the maximum height is usually quite narrow.

[0050] The lower impedance value is obtained by multiplying the maximum impedance value by a coefficient. This upper coefficient is used to determine two moments, which typically allows for an accurate estimation of the impedance pulse width. For this purpose, the lower coefficient is selected in the range [0.1; 0.6], and preferably [0.3; 0.6], which ensures that there are two moments, and these moments correspond to the typical pulse width.

[0051] Therefore, the lower impedance values ​​are included between 30% and 60% of the maximum impedance value. The applicant's research shows that this range enables accurate capture of the width of the generated pulse by removing noise. The applicant has found that a value of 0.5 is particularly advantageous and produces optimal results: the slope of the pulse is very steep at less than 50% of the maximum height, and this value allows for the avoidance of any noise measurement risks.

[0052] Once the upper and lower impedance values ​​are determined, the duration between two moments in the pulse dataset that are furthest apart in time and each possesses either an upper or lower impedance value is determined. The duration associated with the upper impedance value is called the upper duration, and the duration associated with the lower impedance value is called the lower duration. Intuitively, it seems that the upper duration essentially corresponds to the width of the impedance peak in the pulse dataset, while the lower duration essentially corresponds to the pulse width. Finally, the numerical value WR is determined by obtaining the ratio of the upper duration to the lower duration.

[0053] The second value, called PP, is associated with the peak value of the pulse. This value is calculated by taking the difference between the moment of maximum pulse value and the first moment corresponding to the lower impedance, and then dividing this difference by the lower duration. The result is a percentage representing the location of the maximum pulse within the latter.

[0054] The applicant has studied the representation of impulses, particularly by constructing typical graphs (PP; WR). In fact, the applicant has found that such graphs allow them to identify properties of interest, especially the impulse characteristics of cells that have become objects of rotation.

[0055] To test their hypothesis, the applicants altered the morphology of red blood cells by adding specific molecules to the electrolyte solution.

[0056] Different concentrations of glutaraldehyde and N-dodecyl-N,N-dimethyl-3-ammonium-1-propanesulfonate (also known as sulfobetaine 3-12, hereinafter referred to as SB3-12) were added to a diluent, and the impedance of erythrocytes from healthy blood added to these solutions was then measured. More precisely, on the one hand, formulations containing glutaraldehyde at concentrations between 0% and 0.5% were prepared, and on the other hand, solutions containing SB3-12 at concentrations between 0 mg / L and 90 mg / L were prepared. These formulations were prepared separately, i.e., each contained only either the glutaraldehyde additive or the SB3-12 additive.

[0057] It is well known that glutaraldehyde has a fixing effect and can harden red blood cells while maintaining their discoid cell shape. In fact, it is well known that the use of SB3-12 will cause cells to become spherical.

[0058] Blood samples from healthy patients (which were verified to be free of abnormalities, hereinafter referred to as "healthy blood") were analyzed in SB3-12 at different concentrations, and another sample was analyzed in glutaraldehyde at all of the aforementioned different concentrations. Each collection was performed twice to preliminarily assess the reproducibility of the proposed development.

[0059] Finally, the applicant calculated a graph (PP; WR) of the pulses generated by each formulation. Figure 3The graphs obtained for formulations binding SB3-12 are shown, while Figure 4 The graphs obtained for formulations that bind glutaraldehyde are shown.

[0060] These charts validated the applicant's intuition that they contained information about the morphological characteristics of red blood cells. Therefore, the applicant has constructed the pulse regions to be studied (squares 1' to 6') in charts (PP; WR) from healthy blood, as shown below. Figure 5 As shown. Select Figure 5 The pulse region is used to highlight the difference between glutaraldehyde acquisition and SB3-12 acquisition (see pulse region). Figure 3 and Figure 4 ).

[0061] exist Figure 5 In the diagram, the pulse regions are defined as follows, with each region representing the ranges PP (minimum; maximum) and WR (minimum; maximum):

[0062] Square 1':(25;60)-(58;76)

[0063] Grid 2':(60;83)-(65;76)

[0064] Square 3':(25,85)-(76,86)

[0065] Grid 4':(5,20)-(5,70)

[0066] Square 5':(5,25)-(70,85)

[0067] Square 6':(20,85)-(10,58)

[0068] exist Figure 5 Within each region, the applicant calculated the proportion of pulses, as well as the average values ​​of PP and WR. Therefore, for a given sample, there are a total of 18 parameters (3 parameters for each of the 6 regions).

[0069] Then, the pulse ratios in grids 3' and 5' as a function of the concentration in SB3-12 (at what concentrations) were shown. Figure 6 and Figure 7 (in the middle) and with the concentration in glutaraldehyde (in the middle) Figure 8 and Figure 9 Changes in (the middle).

[0070] exist Figures 6 to 9 In each of these sections, normality is represented by a horizontal line. The margin of error is represented by a dashed line and is defined as twice the standard deviation. The solid line between the dashed lines represents the mean of an assessment of 22 healthy blood samples, which defines normality.

[0071] right Figures 6 to 9 Analysis showed that the calculated parameters went beyond normal levels when the concentrations of SB3-12 or glutaraldehyde increased. These figures clearly illustrate the effects of SB3-12 and glutaraldehyde on erythrocytes, and how SB3-12 and glutaraldehyde significantly altered the morphological characteristics of erythrocytes in pulses.

[0072] By assuming that the concentrations in glutaraldehyde and SB3-12 are correlated with the rigidity and sphericity of erythrocytes, respectively, it appears possible to measure these parameters. Indeed, by combining the pulse ratio parameter and the average value of the PP value in zone 3' on a graph, it is possible to quantitatively distinguish between erythrocytes mixed with the glutaraldehyde formulation and those mixed with the SB3-12 formulation.

[0073] All these factors make it possible to empirically verify that impedance pulses contain information related to the morphological characteristics of red blood cells, but there does not appear to be a simple function that can measure the normality of red blood cells or, where applicable, accurately characterize their morphological abnormalities.

[0074] Therefore, the applicant's idea is to develop a device that uses a first trained neural network and is configured to indicate the normality or abnormality of cells based on cell impedance pulses, and a device that uses a second trained neural network and is configured to classify cells by indicating whether cells have normal morphological characteristics, rigid cell morphological characteristics, or spherical cell morphological characteristics.

[0075] Figure 10 A general diagram of the device is shown. The device includes a memory 4 and a classifier 6.

[0076] Memory 4 can be any type of data storage capable of receiving numerical data: hard disk, SSD, any form of flash memory, random access memory, disk, local or cloud-distributed storage, etc. Data calculated by the device can be stored on any type of storage similar to memory 4, or stored on memory 4 itself. This data can be deleted after the device performs its task or saves it.

[0077] In the example described herein, memory 4 receives a pulse dataset. The pulse dataset represents data that can be used to represent... Figure 2 This shows all the data for the impedance pulses. Therefore, it is a set of data pairs (measured impedance value; timestamp), which together define... Figure 2 The curve in the diagram. In practice, the pulse dataset is typically a sample of the orifice's output detection. The pulse dataset can also be a continuous curve, in which case the calculator 6 will make corresponding adjustments.

[0078] Classifier 6 is an element that directly or indirectly accesses memory 4. It can be implemented in the form of suitable computer code that executes on one or more processors. The term "processor" refers to any processor suitable for the computations described below. Such a processor can be implemented in any known manner as a microprocessor for a personal computer, a field-programmable gate array (FPGA) or system-on-a-chip (SoC) type application-specific chip, grid or cloud computing resources, a microcontroller, or any other form capable of providing the computing power required to implement what is described below. One or more of these elements can also be implemented in the form of application-specific electronic circuitry (e.g., application-specific integrated circuits (ASICs)). Combinations of processors and electronic circuitry are also contemplated.

[0079] It should be noted that the device according to the invention can be advantageously integrated into a hematological analysis device or separated from it. Therefore, it can be fully integrated into a hematological analysis device, or, for example, as a network service to which the hematological analysis device connects when necessary or required.

[0080] As mentioned above, classifier 6 is a neural network. In fact, the pulses can be assimilated into images, and the applicant believes that, with proper training, the neural network can be particularly effective in classifying pulses into rotating pulse datasets and non-rotating pulse datasets.

[0081] More specifically, the applicant has determined that a convolutional neural network is the most suitable. Therefore, the architecture of the first neural network is as follows: Figure 11 As shown in the diagram, the architecture of the second neural network is as follows: Figure 12 As shown in the image.

[0082] In both cases, the neural network is a convolutional neural network containing two convolutional layers. Thus, the pulse dataset 100 (consisting of 50 variables) is processed by the first convolutional layer 110, which extracts 6 features, and then the second convolutional layer 120 extracts 3 features from layer 110.

[0083] The filter (or kernel) size of the first convolutional layer 110 is 8, and the filter (or kernel) size of the second convolutional layer 120 is 3.

[0084] The convolutional layer 120 is connected to the fully connected layer 130 of the neural network, which includes a series of four layers of neurons, namely 80, 40, 20 and the last 10 neurons.

[0085] In the case of the first neural network, the fully connected layer 130 returns value 140 in the output layer. In the example described in this paper, value 140 is 1 if the cell is normal, and 0 if the cell is abnormal.

[0086] For all neurons that make up each layer of the model, the retained activation function is the sigmoid function.

[0087] For this first neural network, training was performed using data from: collections of healthy blood defining normality, collections with SB3-12 concentrations between 50 mg / L and 90 mg / L, and collections with glutaraldehyde concentrations between 0.3% and 0.5%. The neural network was thus trained to detect highly affected cells. Each time, a training pulse was labeled with a value of 1 if the associated cell was normal, and with a value of 0 if the associated cell was abnormal.

[0088] The relevance of the training was validated by taking some data from the training and feeding it into the trained neural network. The results were excellent, with a threshold of 0.5 on the output layer (i.e., returning a value of 1 if the output layer returns a value greater than 0.5, and 0 otherwise), the false positive rate for the validation pulse set was 4.3% and the false negative rate was 3.1%.

[0089] In the case of the second neural network, the fully connected layer 130 returns a triple 150 in the output layer. In the example described herein, the values ​​of the three components of the triple are either 0 or 1.

[0090] The second neural network is trained in a manner similar to that of the first neural network, except that training pulses are labeled with triples that indicate whether a given pulse is normal ([1; 0; 0]), where spherical morphological features are ([0; 1; 0]) or rigid morphological features are ([0; 0; 1]).

[0091] The relevance of the training was validated by taking some data from the training and feeding it into the trained neural network. The results were excellent; each element on the output layer was simplified to its maximum component, i.e., [0.92; 0.02; 0.06] returned the triple [1; 0; 0], [0.01; 0.99; 0] returned the triple [0; 1; 0], and [0.05; 0.25; 0.7] returned the triple [0; 0; 1]. Under these conditions, on the validation pulse set, the false positive rate for cells classified as normal was 4%, for cells classified as spherical morphology it was 7.5%, and for cells classified as rigid morphology it was 8.2%.

[0092] These results are excellent and demonstrate the relevance of the device according to the invention, which allows for extremely accurate results to be obtained through simple impedance measurements without the need for hydrodynamic focusing.

[0093] The applicants’ research enabled them to determine that a single convolutional layer and a fully connected layer containing only two or fewer layers may be sufficient.

[0094] Alternatively, the applicant argues that a multilayer perceptron (MLP) can be used instead of the convolutional neural network described above. Indeed, although such neural networks are less precise with the same number of parameters (typically with an additional 5% to 8% false positives), they constitute a reliable alternative.

[0095] These results open the door to detecting diseases that alter the morphological characteristics of red blood cells or other cells, such as malaria or sickle cell disease. Each time, healthy and diseased blood are tested to label the corresponding pulses, and this data is used to train... Figure 11 or Figure 12 A neural network is sufficient.

[0096] For example, by analyzing culture specimens in which all red blood cells are parasitized by malaria, a series of features labeled with the disease can be obtained, and a classifier specific to this infection can be generated.

[0097] Furthermore, the applicant has determined that pairing the duration with the numerical value PP, or pairing the numerical value WR with the numerical value PP, allows for a simple statistical distinction between normal and abnormal cells. This enables the invention to be implemented, for example, without using the neural networks described below.

[0098] Therefore, based on a set of measurements of healthy blood, the applicant established... Figure 13 In this diagram, the applicant has grouped together all the values ​​extracted from the pulses, which allows them to define the pulse region squares 1 through 8.

[0099] exist Figure 13 In the diagram, the pulse region is defined as follows, with each pair representing the range PP (minimum; maximum) and WR (minimum; maximum):

[0100] Grid 1: (70; 80) - (15, 5; 18)

[0101] Square 2: (69; 79) - (18, 5; 21)

[0102] Square 3: (50, 58) - (19, 24)

[0103] Square 4: (36,42)-(20,27)

[0104] Square 5: (26,32)-(22,32)

[0105] Square 6: (8,22)-(26,32)

[0106] Grid 7: (8; 22)-(32; 38)

[0107] Grid 8: (8; 22) - (38; 44)

[0108] Then, the applicant... Figure 3 and Figure 4 The pulse dataset shown repeats the same operation, and establishes two normality criteria by comparing population statistics for each pulse region, one for healthy blood and the other for... Figure 3 and Figure 4 The blood. In fact, comparing population statistics can identify significant differences in each pulse region.

[0109] Therefore, in Figure 14 In the paper, the applicant shows the population statistical distribution of each pulse region of healthy blood (represented by solid lines, with each region defined by a solid color), and on the other hand, the values ​​of healthy blood are represented by dashed lines.

[0110] exist Figure 15 ( Figure 16 In each of the above, the applicant reproduced the population statistical distribution of each pulse region of healthy blood (represented by solid lines, with a region defined by solid color, and...). Figure 14 (Same), on the other hand, the pulses obtained on blood containing glutaraldehyde (SB3-12 each) are represented by dashed lines.

[0111] Figures 14 to 16 It is clearly shown that population statistics are compiled for pulses extracted from blood samples and compared with... Figure 14 The defined normal envelope (and in) Figure 15 and Figure 16 By comparing the samples with those reproduced in the middle, it is possible to determine whether the samples are healthy or have a tendency to become spherical or rigid.

[0112] Therefore, by analyzing a sufficient portion of the sample (e.g., approximately 10,000 cells, the amount of blood analyzed depending on the counting conditions), information of the type "healthy blood sample" or "abnormal blood sample" can be quickly returned without using a neural network.

[0113] Alternatively, instead of the chart (lower duration; PP), the chart (WR; PP) can be used to establish pulse zones that can define normality criteria.

[0114] Figure 17 The analysis of two healthy blood samples over an 11-day period is shown (each analysis was repeated). The figure illustrates how the distribution of the pulse dataset on the WR / PP chart changes with the age of the samples.

[0115] For each analysis, the statistical distribution of the pulse dataset was calculated based on the WPP and WR / PP ratios, and this statistical data is relative to... Figure 5 and Figure 13 The set of metric ranges shown is established.

[0116] exist Figure 18 For the sake of simplicity, only results related to square 3' are provided. The accompanying figure shows the results related to... Figure 5 The statistical variation of the pulses associated with the 3' square over time, based on the age of the sample, with the time starting from the date the sample was collected from the patient.

[0117] The attached figure shows that the blood sample remained "normal" (unchanged) for 7 days before becoming "abnormal," meaning that the statistical markers began to deviate significantly.

[0118] Therefore, this invention can be used to monitor changes in blood samples over time. The accompanying figure is consistent with scientific publications in the field: as samples age, the biomechanical characteristics of the red blood cells present in the sample decrease. In particular, it is well known that the elasticity of red blood cells decreases over time. Changes in samples over time can be affected by storage duration, storage conditions, etc. In any case, Figure 17 and Figure 18 The arrangement according to the invention allows for the generation of information relating to changes in a sample over time, which indicates, for example, the validity of the sample.

[0119] While the above primarily refers to research on red blood cells, the present invention is also applicable to any other type of cell whose morphological characteristics can be altered, such as platelets.

Claims

1. Medical analysis device with cell impedance signal processing functionality, said medical analysis device comprising a memory (4) arranged to receive pulse data sets, each pulse data set comprising impedance value data each time associated with a time marker, these data together representing a cell impedance value curve measured when a cell passes through a polarization aperture, characterized in that, The device comprises a classifier (6) comprising a convolutional neural network receiving as input a set of pulse data and provided with at least one convolutional layer (110, 120) having a depth greater than or equal to 3, and at least two fully connected layers (130), and an output layer (140, 150) rendering a cell classification, the set of pulse data being derived from the output layer (140, 150), wherein the output layer returns a triplet indicative of morphological features of the cell.

2. The apparatus of claim 1, wherein, The convolutional neural network comprises two convolutional layers, one of which is connected to an input layer receiving the set of pulse data.

3. The apparatus of claim 1, wherein, The fully connected layers comprise 4 layers of neurons.

4. The apparatus of claim 1, wherein, The activation function of all layers of the classifier (6) is a sigmoid function.

5. The apparatus of claim 1, wherein, The set of pulse data received in the memory (4) is obtained from impedance measurements taken when red blood cells pass through a polarization opening.

6. The apparatus of any one of claims 1-5, wherein, The output layer returns values indicative of information related to changes occurring over time in the sample from which the set of pulse data is derived.

7. Method of classification of a blood sample, the method comprising the operations of: a) receiving a set of pulse data, each set of pulse data comprising impedance value data each associated with a time marker, these data together representing a curve of impedance values of a cell measured when the cell passes through a polarization opening, wherein the cell is from the blood sample, b) for each set of pulse data bl. determining a maximum impedance value of the set of pulse data, b2. calculating an upper impedance value by multiplying the maximum impedance value by an upper coefficient selected in the range [0.7; 0.95], and by determining in the set of pulse data the time markers at which the relevant impedance values in the set of pulse data are equal to the upper impedance value, and by calculating an upper duration corresponding to the maximum duration between these time markers, and calculating a lower impedance value by multiplying the maximum impedance value by a lower coefficient selected in the range [0.1 ; 0.6], by determining in the set of pulse data the time markers at which the relevant impedance values in the set of pulse data are equal to the lower impedance value, and by calculating a lower duration corresponding to the maximum duration between these time markers, b3. calculating a peak position value equal to the difference between the time instant associated with the maximum impedance value and the first time instant corresponding to the lower impedance value divided by the lower duration, and optionally calculating a rotation value equal to the upper duration divided by the lower duration, c) determining a statistical distribution of the set of pulse data as a function of the lower duration / peak position value pair or the rotation value / peak position value pair, wherein the statistical distribution is established for a set of ranges of values of the pair (bin 1, bin 2, bin 3, bin 4, bin 5, bin 6, bin 7, bin 8, bin 1', bin 2', bin 3', bin 4', bin 5', bin 6'), d) classifying the blood sample by comparing the distribution of operation c) with a reference distribution of healthy blood samples.

Images