Impedance signal processing medical analysis device

Abstract

Impedance Signal Processing Medical Analysis Device: A cellular impedance signal processing medical analysis device comprising a memory (4) arranged to receive pulse datasets, each pulse dataset comprising impedance value data associated with a time marker, these data together representing a curve of cellular impedance values ​​measured during the passage of a cell through a polarized orifice. This device further comprises a classifier (6) comprising a convolutional neural network receiving the pulse datasets as input and being provided with at least one convolutional layer, which convolutional layer has a depth greater than or equal to 3, and at least two fully connected layers, as well as an output layer providing a classification of the cell that generated a pulse dataset. (Short Figure: Fig 10)

Description

Title of the invention: Impedance signal processing medical analysis device

[0001] The invention relates to the field of hematology, and in particular to cell counting and classification devices.

[0002] Since the 1950s, the counting and volumetric analysis of different blood cells in automated hematology analyzers have been performed by impedance measurement, according to a method known as the Coulter Principle. This method consists of passing cells suspended in a conductive liquid through a polarized microorifice and detecting the variations in electrical resistance (or impedance variations) induced by the passage of particles through the orifice. The detection of the different pulses thus generated allows for the counting of the cells.

[0003] Various solutions have been developed to address the problems associated with flow within an orifice (rotations due to edge or hydrodynamic effects, doublet masking, etc.). Since none of these solutions proved satisfactory, hydrodynamic focusing, or hydrofocusing, technology was developed. This solution involves creating a hydrodynamic sheath around the flow of cells to be analyzed, which centers it within the orifice and limits the effects related to edge flow. However, this technique is very complex to implement and particularly expensive.

[0004] The Applicant has recently developed a solution that provides highly satisfactory results while eliminating the need for hydrofocusing. This solution has been protected in patent application FR 1904410. During this development, the Applicant realized that its work could also be used to characterize cells based on their impedance signals, in order to provide information on normality or abnormality, or to characterize the morphology of these cells.

[0005] The state of the art on existing techniques for measuring the deformability of red blood cells reveals two main families.

[0006] The first family concerns precise measurement techniques that are time-consuming and complex to implement. For some of these, analytical models allow for the determination of rheological parameters such as the elastic and viscous moduli of the membrane, among others. One such method is micropipette aspiration, the principle of which is to aspirate a portion of a red blood cell into a pipette by applying a known vacuum. By measuring certain quantities related to the shape of the red blood cell after aspiration, the shear modulus or the viscosity of Shear stress can be deduced (see, for example, the article by EA Evans, "New membrane concept applied to the analysis of fluid shear and micropipette deformed red blood cells," Biophysical Journal, 1973). Other methods, such as the "optical tweezer" method (see, for example, the article by 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, the article by Guck et al., "The optical stretcher: a novel laser tool to micromanipulate cells," Biophysical Journal, 81:767-784, 2001), consist of stretching the globule using optical lasers and observing its shape.

[0007] The second family includes techniques that allow for the faster and more autonomous processing of a large number of cells, thus providing a statistical indication of the deformability of red blood cells in a sample. This type of technique studies the deformation of red blood cells using a shape parameter called the Deformation Index (DI), which is essentially a measure of the elongation of the red blood cell subjected to a well-known mechanical stress. The deformation index combines the mechanical and morphological parameters of the red blood cell; it is therefore simpler to study but does not provide precise rheological information. Among these techniques, we can consider the one presented in the article by Cha et al., "Cell stretching measurement utilizing viscoelasticparticle focusing," Analytical Chemistry 2012, in which red blood cells are stretched sequentially in an extensional flow and observed using a camera.Following the same principle, but by subjecting red blood cells to shear flow, the method described in the article by Dobbe et al., "Analyzing red blood cell-deformability distributions," *Blood Cells, Molecules, and Diseases*, 28:373-384, 2002, demonstrated the impact of certain pathologies on the distributions of the red blood cell deformability index. The article by Mohandas et al., "Analysis of factors regulating erythrocyte deformability," *Journal of Clinical Investigations*, 66:563-573, 1980, describes an ektacytometer in which red blood cells are subjected to shear flow, and for which light diffraction spectra are observed to measure the deformability index. By studying the deformability index curve as a function of the osmolarity of the suspending medium (see, for example, the article by Clark et al.),“Osmotic gradient ektacytometry: comprehensive characterization offered cell volume and surface maintenance”, Blood, 61:899-910, 1983), it was shown that certain rheological and / or morphological parameters can be measured.

[0008] Although richer in information, the methods of the first family do not allow for high-throughput hematological analysis. Indeed, they are too long and complex to implement and require the intervention of a specialized manipulator. On the other hand, the methods of the second family, although easier to implement, remain complicated to industrialize within the framework of a medical analysis laboratory, although the study of a more global response with regard to the deformation index they offer makes it possible to isolate subpopulations of pathological red blood cells.

[0009] All the methods mentioned above require an optical or video-microscopy acquisition system. This is significantly more complicated and expensive to implement than an impedance measurement system. Furthermore, meters operating under the Coulter Principle are not capable of determining the strain index or a simplified version thereof.

[0010] There is therefore a need to offer a simple measurement device that allows for the discrimination of cell populations according to their morphological characteristics.

[0011] The invention improves the situation. To this end, it proposes a medical analysis device for processing cell impedance signals, comprising a memory arranged to receive sample datasets. Each sample dataset comprises impedance value data associated with a time marker. Together, these data represent a curve of cell impedance values ​​measured during the passage of a cell through a polarized orifice. This device further comprises a classifier including a convolutional neural network that receives the sample datasets as input and is equipped with at least one convolutional layer, which convolutional layer has a depth greater than or equal to 3, and at least two fully connected layers, as well as an output layer providing a classification of the cell that generated a sample dataset.

[0012] This device is particularly advantageous because it allows, with a simple measuring device, the classification of cells based on the measurement of their impedance in a Coulter principle flow cytometer. Furthermore, it is possible to adapt existing flow cytometers to incorporate the advantages of the invention, thus avoiding the need to purchase new devices.

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

[0014] - the convolutional neural network comprises two convolutional layers, of which one is connected to the input layer which receives the data sets from the source.

[0015] - the fully connected layer comprises 4 layers of neurons,

[0016] - the function for activating all layers of the classifier is the function sigmoid,

[0017] - the cells are red blood cells,

[0018] - the output layer returns a value indicating whether a cell associated with a draw Given is normal or abnormal,

[0019] - the output layer returns a triplet indicating a morphological characteristic of a cell, and

[0020] - the output layer returns a value indicating whether a cell associated with a draw given indicates the presence of a disease such as malaria or sickle cell disease.

[0021] The invention also relates to a method for classifying a blood sample comprising the following operations:

[0022] a) receiving sets of draw data, each set of draw data comprising impedance value data associated with a time marker, these data together representing a curve of cellular impedance values ​​measured during the passage of a cell through a polarized orifice, the cells being derived from the blood sample,

[0023] b) for each dataset of the source

[0024] b1. determine a maximum impedance value of the source data set,

[0025] b2. calculate a high impedance value by multiplying the impedance value maximum by a high coefficient chosen from the range [0.7; 0.95], and by determining in the data set the time markers whose associated impedance value in the data set is equal to the high impedance value, and by calculating a high duration corresponding to the maximum duration between these time markers, and a low impedance value by multiplying the maximum impedance value by a low coefficient chosen from the range [0.1; 0.6], and by determining in the data set the time markers whose associated impedance value in the data set is equal to the low impedance value, and by calculating a low duration corresponding to the maximum duration between these time markers,

[0026] b3. calculate a peak position value equal to the division of the difference between the instant associated with the maximum impedance value and the first instant corresponding to the low impedance value, and the low duration, and optionally a rotation value equal to the division of the high duration by the low duration,

[0027] c) determine the statistical distribution of the power datasets as a function of the pairs low duration / peak position value or the pairs rotation value / peak position value, the statistics being established with respect to a set of ranges of values ​​of pairs,

[0028] d) classify the blood sample by comparing the distribution of operation c) to a reference distribution of a healthy blood sample.

[0029] Other features and advantages of the invention will become more apparent from the following description, taken from illustrative and non-limiting examples shown in the drawings:

[0030] - [Fig. 1] represents a schematic view of the measuring orifice within the frame of the invention, as well as trajectories that a cell can take in it,

[0031] - [Fig.2] represents power measurements for the trajectories of the [Fig.1],

[0032] - [Fig.3], and

[0033] [Fig.4] represent impedance power characterization diagrams of red blood cells measured with the arrangement of [Fig.1],

[0034] - [Fig.5] represents a diagram characterizing areas of interest,

[0035] - [Fig.6],

[0036] [Fig.7],

[0037] [Fig.8] and

[0038] [Fig.9] represent proportions of pumps in some of the areas of interest of the [Fig.5]

[0039] - [Fig. 10] represents a diagram of an embodiment of a device according to the invention,

[0040] - [Fig. 11] represents a diagram of a neural network implemented in the mode of the realization of [Fig. 10],

[0041] - [Fig. 12] represents a diagram of a neural network implemented in another method of embodiment of the invention,

[0042] - [Fig. 13] represents a diagram characterizing areas of interest, and

[0043] -[Fig. 14],

[0044] [Fig. 15] and

[0045] [Fig. 16] represent population statistics representations of pumps in some of the areas of interest of [Fig. 13].

[0046] The drawings and description below contain, essentially, elements of a definite nature. They may therefore not only serve to better understand the present invention, but also contribute to its definition, if necessary.

[0047] Figure 1 shows a schematic view of the measuring orifice within the scope of the invention, as well as the trajectories that a cell can take within it. The orifice has walls shown in dashed lines, with the abscissa and ordinate expressed in pm.

[0048] Figure 2 represents the measured impedance peaks (derived from the signal corresponding to the impedance variation of the system due to the passage of a cell through the micro-orifice) for each of the trajectories of Figure 1. The abscissa is expressed in ps, while the ordinate is expressed in ohms. As can be seen, the closer the cell's trajectory is to one of the orifice walls, the more chaotic the measurement, and the more errors are explained in the introduction.

[0049] The Applicant's work described in application FR 1904410 led it to develop a new quantity to characterize impedance sumps. This quantity is called WR, and it is a ratio between two sump widths. These widths allow us to indicate the presence of a peak in the well, or on the contrary a well in the shape of a bell.

[0050] To do this, the maximum height of the well is first determined in the well dataset. The maximum height is used to calculate a high impedance value and a low impedance value.

[0051] The high impedance value is obtained by multiplying the maximum impedance value (which corresponds to the maximum height) by a high coefficient. This high coefficient is used to determine two instants which, in general, allow for a good approximation of the width of the impedance peak of a well. For this purpose, the high coefficient is chosen from the range [0.7; 0.95], and preferably [0.8; 0.9], which ensures that there are at least two instants, and that these instants correspond well to the peak of the wells (in order to limit the cases where several peaks are present).

[0052] Thus, the high impedance value is lower than the maximum impedance value and greater than 70% of it. The Applicant's work has shown that this range effectively captures the peaks of the produced springs. The Applicant has identified that the value of 0.875 is particularly advantageous and gives the best results: it allows for the most accurate estimation of the spring peaks. Indeed, the peaks around the maximum height are generally quite narrow.

[0053] The low impedance value is obtained by multiplying the maximum impedance value by a low coefficient. This low coefficient is used to determine two instants which, in general, allow for a good approximation of the width of the impedance well. For this purpose, the low coefficient is chosen from the range [0.1; 0.6], and preferably [0.3; 0.6], which ensures that there are two instants, and that these instants correspond to the overall width of the well.

[0054] Thus, the low impedance value is between 30% and 60% of the maximum impedance value. The Applicant's work has shown that this range effectively captures the width of the produced wells while eliminating noise. The Applicant has identified that a value of 0.5 is particularly advantageous and gives the best results: the slopes of wells below 50% of the maximum height are very steep, and this value avoids any risk of noisy measurements.

[0055] Once the high impedance value and the low impedance value have been determined, the time interval between the two times in the source dataset that are furthest apart in time and that have the high impedance value or the low impedance value, respectively, is determined. The time interval associated with the high impedance value is called the high time interval, and the time interval associated with the low impedance value is called the low time interval. Intuitively, it appears that the high time interval corresponds approximately to the width of the impedance peak in a source dataset, and that the low time interval corresponds approximately to the width of the well. Finally, the WR value is determined by calculating the ratio between the high duration and the low duration.

[0056] The second quantity is called PP and is associated with the peak of the power. This quantity is calculated by taking the ratio between the difference between the instant when the power is at its maximum and the first instant corresponding to the low impedance, and the low duration. The result is a percentage that indicates the position of the power maximum within this range.

[0057] The Applicant has worked on the representations of the pools, in particular by establishing graphs of the type (PP; WR). Indeed, it has discovered that this type of graph allows it to identify interesting behaviors, in particular for pools characteristic of cells that have been rotated.

[0058] In order to validate its hypotheses, the Applicant altered the morphology of red blood cells by adding specific molecules to the electrolytic solution.

[0059] Different concentrations of glutaraldehyde and N- Dodecyl-N,N-dimethyl-3-ammonio-l-propanesulfonate (also known as sulfobetaine 3-12, hereafter SB3-12) was added to the diluent, and the impedances of healthy red blood cells immersed in these solutions were then measured. Specifically, preparations containing glutaraldehyde at concentrations between 0% and 0.5% were prepared, and solutions containing SB3-12 at concentrations between 0 mg / L and 90 mg / L were prepared. The preparations were made separately, meaning that each contained only one addition of glutaraldehyde or one addition of SB3-12.

[0060] It is known that the use of glutaraldehyde has a fixing effect and allows red blood cells to become rigid while maintaining their discocyte shape. Similarly, it is known that the use of SB3-12 tends to spherify cells.

[0061] A blood sample from a healthy patient (whose absence of abnormalities had been verified, hereinafter referred to as "healthy blood") was analyzed with different concentrations of SB3-12, and another was analyzed with all the different concentrations of glutaraldehyde. Each acquisition was performed twice, for a preliminary evaluation of the repeatability of the proposed developments.

[0062] Finally, the Applicant calculated, for each preparation, the (PP; WR) graph of the generated pumps. Figure 3 shows the graphs obtained for the preparations incorporating SB3-12, while Figure 4 shows the graphs obtained for the preparations incorporating glutaraldehyde.

[0063] These graphs validated the Applicant's intuitions, namely that they contain information on the morphological characteristics of red blood cells. Thus, the Applicant established sampling areas (Box1' to Box6') to be studied from the graph (PP; WR) of healthy blood, as shown in [Fig. 5]. The sampling areas in [Fig. 5] were chosen to highlight the differences between the acquisitions with glutaraldehyde and acquisitions with SB3-12 (see [Fig.3] and [Fig.4]).

[0064] In [Fig.5], the draw zones have been defined as follows, each pair expressing a range PP (min; max) and WR (min; max):

[0065] Boxl': (25; 60) - (58; 76)

[0066] Box2': (60; 83) - (65; 76)

[0067] Box3': (25.85) - (76.86)

[0068] Box4': (5.20) - (5.70)

[0069] Box5': (5.25) - (70.85)

[0070] Box6':(20,85)-(10,58)

[0071] In each zone of [Fig.5], the Applicant calculated the proportion of pumps, and the averages of the quantities PP and WR. For a given sample, there are therefore 18 parameters in total (3 parameters for each of the 6 zones).

[0072] Next, the evolutions of the proportion of pumps in the Box3' zone and the Box5' zone were represented, as a function of the concentration in SB3-12 (in figures 6 and 7 respectively), and as a function of the concentration in glutaraldehyde (in figures 8 and 9, respectively).

[0073] In each of Figures 6 to 9, normality is represented by horizontal lines. The margin of error is represented by dashed lines and is defined as being equal to twice the standard deviation. The solid line between the dashed lines represents the mean evaluated over 22 healthy blood samples defining normality.

[0074] Analysis of Figures 6 to 9 reveals that when the concentration of SB3-12 or glutaraldehyde is increased, the calculated parameters deviate from the normal range. These figures clearly show the influence of SB3-12 and glutaraldehyde on red blood cells, and how they visibly alter the morphological characteristics of these cells in the blood samples.

[0075] Assuming that glutaraldehyde concentration and SB3-12 concentration are respectively correlated with red blood cell stiffness and sphericity, it appears possible to measure these parameters. In fact, by combining the proportion parameters of the pumps and the averages of the PP quantity for zone 3' on a graph, it becomes possible to quantitatively distinguish between red blood cells that have been mixed with a glutaraldehyde preparation and those that have been mixed with an SB3-12 preparation.

[0076] All these elements have made it possible to empirically validate the fact that impedance sources contain information concerning the morphological characteristics of red blood cells, but there does not appear to be a simple function that allows us to measure the normality of red blood cells, nor to precisely characterize their morphological abnormality if necessary.

[0077] The Applicant therefore had the idea of ​​developing a device using a first a neural network trained and configured to indicate the normality or abnormality of a cell based on its impedance peak, and a device using a second neural network trained and configured to classify a cell by indicating whether it exhibits normal morphological characteristics, rigidified cell morphological characteristics, or spherized cell morphological characteristics.

[0078] Figure 10 represents a generic diagram of this device. The device includes a memory 4, and a classifier 6.

[0079] Memory 4 can be any type of data storage suitable for receiving digital data: hard disk drive, solid-state drive (SSD), flash memory of any kind, RAM, magnetic disk, locally distributed or cloud-based storage, etc. The data processed by the device can be stored on any type of memory similar to, or on, Memory 4 itself. This data can be erased after the device has completed its tasks or retained.

[0080] In the example described here, memory 4 receives pump data sets. A pump data set represents all the data that characterize an impedance pump shown in [Fig. 2]. It is therefore a set of pairs (measured impedance value; time marker), which together define a curve like those in [Fig. 2]. In practice, the pump data set will generally be a sampling of the orifice detection output. The pump data set can also be a continuous curve, in which case the computer 6 will be adapted accordingly.

[0081] The classifier 6 is an element that accesses memory 4 directly or indirectly. It can be implemented in the form of suitable computer code executed on one or more processors. The term "processors" is understood to mean any processor suitable for the calculations described below. Such a processor can be implemented in any known manner, such as a personal computer microprocessor, a dedicated FPGA or SoC (System on Chip) type chip, a computing resource on a grid or in the cloud, a microcontroller, or any other form suitable for providing the computing power necessary for the implementation described below. One or more of these elements can also be implemented in the form of specialized electronic circuits such as an ASIC. A combination of processor and electronic circuits can also be considered.

[0082] It should be noted that the device according to the invention can advantageously be integrated into a hematological analysis device, or be remote. It can therefore be fully integrated into the hematological analysis device or, for example, be a web service to which the hematological analysis device connects when ne- necessary or desired.

[0083] As suggested above, classifier 6 is a neural network. Indeed, the streams can be considered as images, and, with appropriate training, the Applicant considered that a neural network could be particularly effective in classifying the streams into sets of stream datasets with or without rotation.

[0084] In particular, the Applicant identified that a convolutional neural network was the most suitable. Thus, the architecture of the first neural network is shown in [Fig. 11], while that of the second neural network is shown in [Fig. 12].

[0085] In both cases, the neural network is a convolutional neural network comprising two convolutional layers. Thus, a dataset of 100 (composed of 50 variables) is processed by a first convolutional layer 110 which extracts 6 features, then a second convolutional layer 120 extracts 3 features from layer 110.

[0086] The filters (or convolution kernels) of the first convolution layer 110 have a size of 8, those of the second convolution layer 120 have a size of 3.

[0087] The convolution layer 120 is connected to a fully connected layer 130 of the neural network which comprises a chain of 4 layers of neurons comprising respectively 80, 40, 20 and finally 10 neurons.

[0088] In the case of the first neural network, the fully connected layer 130 returns a value 140 in the output layer. In the example described here, the value 140 is 1 if the cell is normal and 0 if it is abnormal.

[0089] For all the neurons composing the different layers of the model, the activation function retained is the sigmoid function.

[0090] For this first neural network, training was performed using data from acquisitions of healthy blood defining normality, acquisitions with SB3-12 concentrations between 50 mg / L and 90 mg / L, and acquisitions with glutaraldehyde concentrations between 0.3% and 0.5%. The neural network was thus trained to detect highly affected cells. In each case, the training samples were labeled with the value 1 if the associated cell was normal and with the value 0 if the associated cell was abnormal.

[0091] The relevance of the training was verified by removing certain data from the training and introducing them into the trained neural network. The results were excellent, and, with a threshold of 0.5 on the output layer (i.e., the value 1 is returned if the output layer returns a value greater than 0.5 and 0 otherwise), the false positive rate was 4.3% and the false negative rate negatives of 3.1% on the validation power set.

[0092] In the case of the second neural network, the fully connected layer 130 returns a triplet 150 in the output layer. In the example described here, the three components of the triplet take the value 0 or 1.

[0093] For this second neural network, the training is carried out in a similar manner to the first neural network, except that the training sets are labeled with triplets indicating whether a given set is normal ([1; 0; 0]), with morphological characteristics of spherization ([0; 1; 0]) or with morphological characteristics of rigidification ([0; 0; 1]).

[0094] The relevance of the training was verified by removing certain data from the training and introducing them into the trained neural network. The results were excellent, with each element on the output layer being reduced to its maximum component; that is, [0.92; 0.02; 0.06] returns the triplet [1; 0; 0], [0.01; 0.99; 0] returns the triplet [0; 1; 0], and [0.05; 0.25; 0.7] returns the triplet [0; 0; 1]. Under these conditions, on the validation set, the false positive rate was 4% for cells classified as normal, 7.5% for cells classified with spheric morphology characteristics, and 8.2% for cells classified with rigidification morphology characteristics.

[0095] These results are excellent and demonstrate the relevance of the device according to the invention, which makes it possible to obtain results of extreme precision with simple impedance measurements without hydrofocusing.

[0096] The Applicant's work has enabled it to establish that a single convolution layer could suffice, as well as a fully connected layer that would contain only 2 layers or less.

[0097] Alternatively, the Applicant considers it possible to use a multi-layer perceptron (MLP) instead of the convolutional neural network described above. Indeed, even though this type of neural network provides less accurate results for an equivalent number of parameters (typically, 5 to 8% more false positives), it nevertheless constitutes a plausible alternative.

[0098] These results paved the way for the detection of pathologies that result in a modification of the morphological characteristics of red blood cells or other cells, such as malaria or sickle cell anemia. In each case, it is sufficient to test healthy and diseased blood samples to label the corresponding blood groups, and to train the neural network shown in Figures 11 or 12 with this data.

[0099] For example, by analyzing a culture sample in which all red blood cells are parasitized by malaria, a series of signatures labeled with this pathology is obtained and makes it possible to generate a classifier specific to this infection.

[0100] The Applicant has also identified that coupling the low duration with the PP quantity or the WR quantity with the PP quantity makes it possible to perform a simple distinction between normal and abnormal cells on a statistical basis. This makes it possible, for example, to implement the invention without using a neural network described below.

[0101] Thus, based on a set of measurements on healthy blood, the Applicant established [Fig. 13]. In this figure, the Applicant grouped all the values ​​taken from the wells, which allowed it to define the well zones Box1 to Box8.

[0102] In [Fig. 13], the draw zones have been defined as follows, each pair expressing a range PP (min; max) and WR (min; max):

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] Boxl: (70; 80)-(15.5; 18) Box2: (69; 79)-(18.5; 21) Box3: (50.58) - (19.24) Box4: (36.42) - (20.27) Box5: (26.32) - (22.32) Box6: (8.22) - (26.32) Box7: (8;22) - (32;38) Box8: (8;22) - (38;44)

[0111] Next, the Applicant repeated the same operation with the draw datasets shown in [Fig. 3] and [Fig. 4] and established two normality criteria by comparing the population statistics for each draw zone, for healthy blood on the one hand and for the blood samples in Figures 3 and 4 on the other. Indeed, the comparison of population statistics reveals significant differences for each draw zone.

[0112] Thus, on [Fig. 14], the Applicant has represented on the one hand the statistical distributions of the populations for each draw zone for healthy blood (in solid lines, defining a solid colour zone), and on the other hand the value of healthy blood in dashed lines.

[0113] On [Fig. 15] (respectively [Fig. 16]), the Applicant has reproduced on the one hand the statistical distributions of the populations for each draw zone for healthy blood (in solid lines, defining a solid colour area, identical to those of [Fig. 14]), and on the other hand in dashed lines these same distributions but for draws obtained on blood containing glutaraldehyde (respectively SB3-12).

[0114] Figures 14 to 16 clearly show that it is sufficient to perform population statistics for the draws taken from a blood sample and to compare them with the envelope of normality defined by [Fig. 14] (and reproduced in Figures 15 and 16) to determine whether the sample is healthy or shows propensities for spherization or rigidification.

[0115] Thus, by analyzing a portion of a sufficient sample (for example approximately (10,000 cells, the volume of blood analyzed being dependent on the counting conditions), it is possible to quickly return information such as "healthy blood sample" or "abnormal blood sample", without using a neural network.

[0116] Alternatively, instead of the graph (low duration; PP), the graph (WR; PP) could be used to establish the draw zones allowing the normality criteria to be defined.

[0117] Although the foregoing refers mainly to the study of red blood cells, the invention could be applied to any other type of cell whose morphological characteristics are likely to change, such as platelets for example.

Claims

1.

2.

3.

4.

5.

6. Claims Medical analysis device for processing cellular impedance signals, comprising a memory (4) arranged to receive pulse data sets, each pulse data set comprising impedance value data associated each time with a time marker, these data together representing a curve of cellular impedance values ​​measured during the passage of a cell in a polarized orifice, characterized in that it comprises a classifier (6) comprising a convolution neural network receiving the pulse data sets as input (100) and is provided with at least one convolution layer (110, 120), which convolution layer has a depth greater than or equal to 3, and at least two fully connected layers (130), as well as an output layer restoring a cell classification at the origin of a pulse data set (140, 150) which comprises a triplet indicating a morphological characteristic of a cell. Device according to claim 1, wherein the convolutional neural network (6) comprises two convolution layers (110, 120), one of which (110) is connected to the input layer which receives the pulse data sets. The device of claim 1 or 2, wherein the fully connected layer comprises 4 layers of neurons. Device according to one of the preceding claims, in which the activation function of all the layers of the classifier (6) is the sigmoid function. Device according to one of the preceding claims, in which the pulse data sets received in the memory (4) are obtained from impedance measurements measured during the passage of red blood cells in a polarized orifice. A method of classifying a blood sample comprising the following operations: a. receiving pulse data sets, each pulse data set comprising impedance value data associated each time with a time marker, these data together representing a curve of cellular impedance values ​​measured during the passage of a cell in a polarized orifice, the cells being derived from the sample bloody, b. for each data set of draw 1. determine a maximum impedance value of the pulse data set, 2. calculating a high impedance value by multiplying the maximum impedance value by a high coefficient chosen from the range [0.7; 0.95], and determining in the pulse data set the time markers whose associated impedance value in the pulse data set is equal to the high impedance value, and calculating a high duration corresponding to the maximum duration between these time markers, and a low impedance value by multiplying the maximum impedance value by a low coefficient chosen from the range [0.1; 0.6], and determining in the pulse data set the time markers whose associated impedance value in the pulse data set is equal to the low impedance value, and calculating a low duration corresponding to the maximum duration between these time markers, 3. calculate a peak position (PP) value equal to the division of the difference between the time associated with the maximum impedance value and the first time corresponding to the low impedance value, and the low duration, and optionally a rotation value (WR) equal to the division of the high duration by the low duration, c. determine the statistical distribution of the pulse data sets according to the low duration / peak position value (PP) pairs or the rotation value / peak position value (WR) pairs, the statistics being established in relation to a set of ranges of torque values ​​(Box 1, Box 2, Box 3, Box 4, Box 5, Box 6, Box 7, Box 8, Box 1', Box 2', Box 3', Box 4', Box 5', Box 6'), d. classify the blood sample by comparing the distribution of operation c) to a reference distribution of a healthy blood sample. 16

Images