Medical analysis equipment with impedance signal processing function

Abstract

A medical analysis device with a cell impedance signal processing function comprises a memory (4) configured to receive a pulse data set, each pulse data set comprising impedance value data associated with a time mark each time, the data together representing a curve of the cell impedance value measured when the cell passes through a polarization opening; a calculator (6) configured to process the pulse data set by determining a rotation value, the rotation value indicating whether the cell has undergone rotation during its passage through the polarization opening, the pulse data set having been obtained from the cell; and a classifier (8) configured to retrieve a given pulse data set from the calculator (6) and classify the given pulse data set into a rotation pulse data set group (10) or a non-rotation pulse data set group (12) using the generated rotation value.

Description

Technical Field

[0001] The present invention relates to the field of hematology, and in particular to a cell counting device. Background Art

[0002] Since the 1950s, the counting and volume determination of different blood cells in hematological systems has been performed by impedance measurement according to a method known as the Coulter principle. This method consists in passing cells suspended in a conductive liquid through polarized micro-openings and detecting the change in resistance (or impedance) caused by the passage of the particles through the openings. The detection of the different pulses thus generated makes it possible to count the elements.

[0003] Several authors have described analytical developments in the literature, including Grover et al., "Electrical Size of Particles in Suspension: I. Theory" (Biophysical Journal, 9: 1398-1414, 1969) and Hurley, "Particle Size Measurement Using a Particle Sizer" (Biophysical Journal, 10: 74-79, 1970), which describe how the amplitude of the electrical disturbance can be used to retrieve the cell volume in specific cases according to the following formula: U = (r.fs.i)V / S^2, where U is the voltage change at the electrode terminals, V is the volume of the particle, S is the opening cross-section, r is the conductivity of the medium, i is a constant applied current, and fs is the form factor of the cell, which depends on the orientation and shape of the cell.

[0004] This formula makes it possible to construct volume distributions for cells analyzed under uniform electric fields, whose cell dimensions are much smaller than the opening and have a constant shape factor. For example, in the case of red blood cells (RBCs), the mean corpuscular volume (MCV) and the distribution index (DI) are defined based on the pulse height distribution and represent important hematological parameters. Based on this formula, it is also possible to retrieve the actual volume by multiplying the volume V by a coefficient involving the shape factor fs.

[0005] These calculations are based on the assumption that the shape factor fs is constant and that the electric field is uniform. However, in conventional devices these assumptions are usually not observed. In fact, on the one hand, the electric field is not uniform (stronger near the edges), and on the other hand, when the cells are close to the walls, the hydrodynamic effects are large (especially due to shear stresses) and they can rotate or adopt complex shapes. In these cases, the shape factor changes with respect to the cell channel in the center and must be corrected if one wants to get the actual volume correct.

[0006] In the current device, the shape factor is not corrected for cell trajectory, and the associated impedance signal for cell passage close to the wall has a pulse height, leading to an overestimation of cell volume due to electrostatic and / or hydrodynamic artifacts.

[0007] For this reason, the volume distribution obtained is usually asymmetric, with a longer tail on the right (called a "right-skewed" distribution), and thus the measurements of volume parameters of the cells (MCV and DI) are distorted.

[0008] Several solutions have been proposed to address these issues:

[0009] - Spheroidizing erythrocytes can eliminate the effect on shape factor but involves the development of specific reagents to alter osmotic pressure,

[0010] - Enlarging the opening homogenizes the electric field, but creates large impedance losses and a higher risk of simultaneous passage (when two cells pass through the opening at the same time, making them indistinguishable),

[0011] - Sorting the pulses to remove those corresponding to cell tracks close to the edge, but these pulse width-based approaches are ineffective even when taking pulse height into account (especially when cell size varies), and these approaches fail to allow identification of cell rotation information.

[0012] Since none of these solutions were satisfactory, the hydrodynamic focusing technique or hydrodynamic focusing was developed. This solution consists in a hydrodynamic sheathing of the flow of cells to be analyzed, which makes it possible to center it in the opening and limit the effects associated with the edge channels. Nevertheless, this technique is very complex to implement and particularly expensive. Summary of the invention

[0013] The present invention improves this situation.

[0014] To this end, the present invention relates to a medical analysis device with a cell impedance signal processing function, comprising a memory, which is configured to receive pulse data sets, each pulse data set including impedance value data associated with a time mark each time, which data together represent a curve of the cell impedance value measured when the cell passes through a polarization opening; a calculator, which is configured to process the pulse data set by determining a rotation value, which indicates whether the cell has undergone rotation during its passage through the polarization opening, through which the pulse data set has been obtained; and a classifier, which is configured to provide a given pulse data set to the calculator, and use the generated rotation value to classify the given pulse data set into a rotation pulse data set group or a non-rotation pulse data set group.

[0015] The device is advantageous because by considering sets of rotation-free pulse data sets, results similar to those that can be obtained on hydrodynamic focusing systems can be obtained, but at significantly reduced cost and complexity of use.

[0016] Alternatively, the device according to the invention may have one or more of the following features:

[0017] -The calculator was set up to process the pulse data set in the following way:

[0018] * Determine the maximum impedance value of the pulse data set,

[0019] * 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 marks in the pulse data set at which the associated impedance value in the pulse data set is equal to the upper impedance value, and by calculating the upper duration corresponding to the maximum duration between these time marks,

[0020] * 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 marks in the pulse data set at which the associated impedance value in the pulse data set is equal to the lower impedance value, and by calculating the lower duration corresponding to the maximum duration between these time marks,

[0021] * Returns the rotation value determined by the ratio of the upper duration and the lower duration.

[0022] -The upper coefficient is selected in the range [0.8; 0.9],

[0023] -The coefficient is selected in the range [0.3; 0.6],

[0024] - the upper coefficient is equal to 0.875, the lower coefficient is equal to 0.5,

[0025] - the calculator returns a rotation value, the rotation value being equal to the upper duration divided by the lower duration, and wherein the classifier is configured to classify pulse data sets having rotation values ​​strictly less than 0.45 into a rotated pulse data set group, and to classify pulse data sets having rotation values ​​greater than 0.45 into a non-rotated pulse data set group,

[0026] - the calculator is a neural network, comprising, on the one hand, a converter arranged to process the pulse data set to adapt it for processing by the neural network, arranged to make the time stamps and the impedance values ​​dimensionless and to resample the dimensionless data to a selected number of impedance values ​​for the neural network, and, on the other hand, a multilayer perceptron comprising an input layer, at least one hidden layer and an output layer for determining the rotation values, the input layer comprising an input for each value from the resampling performed by the converter.

[0027] -The input layer also includes an input for receiving the width of the pulse data set processed by the converter,

[0028] -The multilayer perceptron includes a hidden layer of 20 neurons, and the activation function of these neurons is the sigmoid function,

[0029] The multilayer perceptron is trained by stochastic gradient descent after back-propagation of the error gradients, the training error is calculated according to a cost function based on the root mean square error, and the output is set to 0 for the input corresponding to the cell orientation of the associated spike data set being greater than the maximum impedance value of 0.3 rad among the impedance values ​​for the spike data set, and is set to 1 otherwise.

[0030] The present invention also relates to a method for processing impedance signals for medical analysis, comprising the following operations: a. receiving pulse data sets, each pulse data set comprising impedance value data associated with a time mark each time, which data together represent a cell impedance value curve measured when the cell passes through a polarization opening, b. processing the pulse data set from operation a by determining a rotation value, which indicates whether the cell has undergone rotation during its passage through the polarization opening, and the pulse data set has been obtained by the cell, c. classifying the pulse data set into a rotation pulse data set group or a non-rotation pulse data set group according to the rotation value obtained in operation b.

[0031] In various alternative embodiments, the method may have one or more of the following features:

[0032] - Operation b includes:

[0033] b1. Determine the maximum impedance value of the pulse data set,

[0034] b2. calculating the upper impedance value by multiplying the maximum impedance value by an upper coefficient selected in the range [0.7; 0.95], by determining the time marks in the pulse data set at which the associated impedance value in the pulse data set is equal to the upper impedance value, and by calculating the upper duration corresponding to the maximum duration between these time marks, b3. calculating the lower impedance value by multiplying the maximum impedance value by a lower coefficient selected in the range [0.1; 0.6], by determining the time marks in the pulse data set at which the associated impedance value in the pulse data set is equal to the lower impedance value, and by calculating the lower duration corresponding to the maximum duration between these time marks, and

[0035] b4. Returns the rotation value determined by the ratio of the upper duration and the lower duration.

[0036] -The upper coefficient is selected in the range [0.8; 0.9],

[0037] -The coefficient is selected in the range [0.3; 0.6],

[0038] - the upper coefficient is equal to 0.875, the lower coefficient is equal to 0.5,

[0039] - the rotation value returned by operation b4 is equal to the value of the upper duration divided by the lower duration, and operation c classifies the pulse datasets with rotation values ​​strictly less than 0.45 into the rotated pulse dataset group, and classifies the pulse datasets with rotation values ​​greater than 0.45 into the non-rotated pulse dataset group,

[0040] -The operations are performed by a neural network comprising, on the one hand, a converter arranged to process a pulse data set to adapt it for processing by the neural network, arranged to make the time stamps and the impedance values ​​dimensionless and to resample the dimensionless data to a selected number of impedance values ​​for the neural network, and, on the other hand, a multilayer perceptron for determining the rotation values ​​comprising an input layer, at least one hidden layer and an output layer, the input layer comprising an input for each value of the resampling performed by the converter.

[0041] -The input layer also includes an input for receiving the width of the pulse data set processed by the converter,

[0042] -The multilayer perceptron consists of a hidden layer of 20 neurons, and the activation function of these neurons is the sigmoid function,

[0043] The multilayer perceptron is trained by stochastic gradient descent after back-propagation of the error gradients, the training error is calculated according to a cost function based on the root mean square error, and the output is set to 0 for the input corresponding to the cell orientation of the associated spike data set being greater than 0.3 rad of the maximum impedance value among the impedance values ​​for the spike data set, and to 1 otherwise.

[0044] BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Further characteristics and advantages of the invention will appear more clearly from reading the following description, based on examples given by way of illustration and not limitation, based on the accompanying drawings, in which:

[0046] -[ Figure 1 ] represents the principal views of the measuring openings within the scope of the invention and the trajectories that a cell may take therein,

[0047] -[ Figure 2 ]express Figure 1 Pulse measurement of the trajectory,

[0048] -[ Figure 3 ] represents a diagrammatic representation of a first embodiment of a device according to the invention,

[0049] -[ Figure 4 ] indicates that Figure 3 Example classification of a human blood sample performed by the device in and its comparison with the actual corresponding pulse distribution,

[0050] -[ Figure 5 ] represents a diagrammatic representation of a second embodiment of the device according to the invention, and

[0051] -[ Figure 6 ] is a bar graph showing the actual blood sample results calculated by the rotation values ​​of the second embodiment. DETAILED DESCRIPTION

[0052] The drawings and descriptions hereinafter contain substantially elements of a certain nature and may therefore be used not only to aid in a better understanding of the invention but also, where applicable, to facilitate its definition.

[0053] Figure 1 Principle view showing a measurement opening within the scope of the invention, and the trajectories that a cell can take therein. The opening has walls indicated by dashed lines, the x- and y-coordinates are indicated in μm.

[0054] Figure 2 Indicates for Figure 1 The x-coordinate is expressed in μs, while the y-coordinate is expressed in Ohms. It can be seen that the closer the cell's incident trajectory is to one of the walls of the opening, the more chaotic the measurement becomes, leading to the errors described in the introduction.

[0055] The Applicant has carried out advanced simulation work that has enabled him to discover the fact that the overestimation of the volume is directly related to the rotational movement of the cells in the micro-openings. In other words, the filtration method based on the characterization of the rotation using the impedance signal and implemented on this rotation provides a more reliable measurement of the cell volume parameters than the traditional filtration process based on the analysis of height and width, and the results obtained are comparable to those obtained by the hydrodynamic focusing system.

[0056] More specifically, the applicants found in their simulations that cells that adopt trajectories close to the wall and that have complex pulses in response systematically undergo rotations when passing through the opening. Therefore, in order to reliably and accurately calculate MCV and DI, it is useful to remove pulses that are identified as cells that have undergone rotations.

[0057] None of the prior art work has precisely established a correlation between the unavailability of a pulse and the presence of cell rotation, and none of these works discloses a method for identifying whether a cell has undergone rotation in an opening (based on its impedance pulse).

[0058] Applicants have identified two methods for reliably determining and / or quantifying whether a cell undergoes rotation within an opening.

[0059] The first method is based on the processing of the signal, which aims to determine the presence of a peak in the measured impedance pulse. In fact, the cell that passes through the center of the opening and whose impedance pulse can be used to calculate the MCV and reliable DI has a bell-shaped impedance pulse, such as Figure 1 The other trajectories induce pulses with peak values ​​corresponding to the closest 90° rotation of the cell in the opening. The method will be reviewed with reference to the first embodiment.

[0060] The second method is based on a neural network that is trained to recognize Figure 1 The impedance pulse associated with traces 1 and 2. The method will be reviewed with reference to a second embodiment.

[0061] Figure 3 A diagram showing a first embodiment of a device according to the invention is shown. In this embodiment, the device comprises a memory 4, a calculator 6 and a classifier 8.

[0062] The memory 4 may be any type of data storage capable of receiving digital data: hard disk, flash hard disk (SSD), any form of flash memory, random access memory, disk, local or cloud distributed storage, etc. The data calculated by the device may be stored on or above any type of storage similar to the memory 4. After the device has performed its task or saved it, these data may be deleted.

[0063] In the example described here, the memory 4 receives a pulse data set. The pulse data set represents a Figure 2 A set of data for an impedance pulse as shown in . It is therefore a set of pairs (measured impedance value; time stamp) that are related to Figure 2 In practice, the pulse data set is usually a sample of the detection output of the orifice. The pulse data set can also be a continuous curve, in which case the calculator 6 will adjust accordingly.

[0064] Calculator 6 and classifier 8 are elements that directly or indirectly access memory 4. They can be embodied in the form of suitable computational codes executed on one or more processors. The term processor refers to any processor suitable for the following calculations. Such a processor can be embodied in any known manner and in the following forms: a personal computer microprocessor, a dedicated chip of the type FPGA or SoC ("system on chip"), a grid computing or cloud computing resource, a microcontroller, or any other form capable of providing the required computing power of the embodiments described below. One or more of these elements can also be embodied in the form of a dedicated electronic circuit, such as an ASIC. A combination of a processor and an electronic circuit can also be envisioned.

[0065] It is worth noting that the device according to the invention can advantageously be integrated in a hematological analysis device, or be separate. Thus, it can be fully integrated in a hematological analysis device, or, for example, as a network service, with which the hematological analysis device is connected when necessary or desired.

[0066] As described above, the memory 4 receives a pulse data set measured when a cell passes through an opening. The classifier 8 is then configured to call the calculator 6 to determine the rotation value of the pulse data set just measured or another pulse data set stored in the memory 4. Based on the rotation value calculated by the calculator 6, the classifier 8 classifies the pulse data set into a rotation pulse data set group or a non-rotation pulse data set group. Once these pulse data set groups are established, the hematology device is able to reliably determine the MCV and DI using the non-rotation pulse data set groups. Note that the MCV is obtained by keeping the pulse 10% above the maximum pulse value and by averaging the resulting volumes, and the DI is obtained by multiplying the standard deviation of the pulses used to calculate the MCV by 100 and dividing by the MCV. Alternatively, the value of 10% can be reduced to 2%.

[0067] The calculator 6 is arranged to process the pulse data set to calculate a value indicating whether the cell has undergone a rotation. To this end, the applicant observed different pulse shapes based on the trajectory and inferred a value that can reliably separate the two cell populations. To this end, it has established a value called WR. The WR value is the ratio between the two pulse widths. These widths obtained at a specific time can indicate the presence of a peak in the pulse or (on the other hand) a bell-shaped pulse.

[0068] To this end, the calculator 6 first determines the maximum pulse height in the pulse data set. The maximum height is used to calculate the upper impedance value and the lower impedance value.

[0069] The upper impedance value is obtained by multiplying the maximum impedance value (which corresponds to the maximum height) by the upper coefficient. This upper coefficient is used to determine two times, which generally makes it possible to accurately estimate 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], to ensure that at least two times are obtained and that these times do correspond to the pulse peak (to limit the scenario where there are multiple peaks).

[0070] The upper impedance value is therefore less than the maximum impedance value and greater than 70% thereof. The applicant's work has demonstrated that this range makes it possible to capture the peak of the generated pulse. The applicant has found that a value of 0.875 is particularly advantageous and gives the best results: it is possible to estimate the pulse peak in the most accurate way. In practice, the peak near the maximum height is usually relatively narrow.

[0071] The lower impedance value is obtained by multiplying the maximum impedance value by the following coefficient. The lower coefficient is used to determine two times, which generally makes it possible to accurately estimate the impedance pulse width. For this purpose, the lower coefficient is selected in the range [0.1; 0.6], preferably [0.3; 0.6], which ensures that two times are obtained and these times correspond to the general pulse width.

[0072] The lower impedance value is therefore between 30% and 60% of the maximum impedance value. The applicant's work has demonstrated that this range allows to accurately capture the width of the generated pulse by removing the noise. The applicant has found that a value of 0.5 is particularly advantageous and gives the best results: the slope of the pulse below 50% of the maximum height is very steep and this value prevents any risk of measurement being affected by noise.

[0073] Once the upper impedance value and the lower impedance value are determined, the calculator 6 determines the duration between two times of the pulse data set: at the maximum time distance from each other and having the upper impedance value or the lower impedance value, respectively. 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 basically corresponds to the width of the impedance peak of the pulse data set, while the lower duration basically corresponds to the pulse width.

[0074] Finally, calculator 6 determines the rotation value by calculating the ratio between the upper duration and the lower duration. Again, intuitively, a low rotation value indicates the presence of a peak (and therefore rotation), while a high rotation value indicates a bell-shaped pulse. Obviously, the ratio can be reversed (and the conclusion is the same). Similarly, the rotation value can be calculated based on this high ratio at the selected power or an exponent of this ratio.

[0075] Our simulations show that a rotation value of 0.45 has strong discrimination and can separate cells with the same WR as other cells (for WR values ​​between 0 and 1, Figure 2 The ratio is between 0 and 100) Figure 2 Therefore, the classifier 8 uses the threshold value to place the pulse dataset into the rotation pulse dataset group or the non-rotation pulse dataset group.

[0076] Obviously, one skilled in the art will know how to adjust the upper coefficient, the lower coefficient, the use of powers for calculating the rotation values, and the thresholds for the classifier.

[0077] Several calibration tests may also be performed to optimize the selection of the upper coefficient, the lower coefficient and the threshold value. In fact, it is known that the set of data sets of non-rotated pulses should have a Gaussian distribution. With relatively small computational costs, it is conceivable that the calculator 6 operates with several pairs of upper and lower coefficients and retains the pair that enables the "most gaussian" (la plusgaussienne) distribution to be obtained at the output.

[0078] Figure 4 An example of classification performed by the device according to the invention is shown: an experimental pulse collection using a blood sample from a healthy subject. Separately from their processing by the device according to the invention, the pulses were manually studied to manually classify them. In the figure, square dots represent bell-shaped pulses and round dots represent pulses with peaks. It is clear that the device distinguishes these two pulse populations with very high accuracy.

[0079] Now refer to Figure 5 A second embodiment is described. In this embodiment, the calculator 6 is no longer centered around signal processing, but is a neural network. In practice, the pulses can be compared to the image, and with appropriate training, the applicant believes that the neural network may be particularly effective in classifying the pulses into a group of rotated pulse data sets and a group of non-rotated pulse data sets.

[0080] To this end, the calculator 6 comprises a converter 14 and a multilayer perceptron 16. In practice, the pulses have very different widths and amplitudes, so it is necessary to normalize them by an operation to make them dimensionless, so that the training of the multilayer perceptron 16 and its detection quality are reliable. Therefore, the converter 14 is arranged to receive a pulse data set and perform an operation to make it dimensionless in amplitude and time.

[0081] For dimensionless operations on amplitude, the pulse dataset is normalized by the maximum amplitude value of the dataset. This results in a pulse dataset with values ​​between 0 and 1.

[0082] For dimensionless operation in time, the present application relies on the shape of the pulse. In fact, as mentioned above, the slope of the pulse near the edge is very significant. Therefore, considering the pulse width at 50% of the maximum impedance value, a correct estimate of the width can be obtained. Therefore, the converter 14 scales the pulse data set based on this width. In order to ensure that the resulting pulse data set contains enough data points for the multilayer perceptron 16, the time-dimensionless pulse data set can be resampled if necessary.

[0083] Operations that make amplitude and time dimensionless can be performed sequentially (amplitude first, then time, or time first, then amplitude) or simultaneously.

[0084] Once the pulse data set is formatted for the multilayer perceptron 16, it is processed by the multilayer perceptron 16, which in the example described here returns a rotated value between 0 and 1. In the example described here, the multilayer perceptron 16 includes an input layer, a hidden layer, and an output layer. The input layer includes a number of inputs, which is equal to the number of samples of the pulse data set formatted by the converter 14 plus an input receiving the pulse width. The applicant has confirmed that the input related to the pulse width is not essential (and therefore it is optional), however, the best results can be obtained with it. In the example described here, the number of samples is equal to 20.

[0085] Regarding the multilayer perceptron 16, the applicant has studied various training and operation strategies and has obtained the most favorable results in terms of training time and computational cost with a hidden layer comprising 20 neurons and an activation function (which is the same for each neuron), namely a sigmoid function. The study of the output values ​​of the multilayer perceptron 16 shows that a threshold of 0.5 on the output layer allows effective isolation and a Gaussian distribution of the population, which confirms the quality of the results obtained.

[0086] This is achieved thanks to the training of the multilayer perceptron 16 using 120 pulses generated by varying the trajectory and parameters of the cell (sphericity, internal viscosity, shear modulus of the membrane, etc.). The pulses are then formatted by the converter 14 and associated with an output value equal to 0 or 1, depending on whether the orientation of the cell at the maximum pulse amplitude is greater or less than 0.3 rad (able to maintain the same Figure 2 The thresholds for the cells corresponding to trajectories 1 and 2 in Fig. 1 are used). Then, training is performed by stochastic gradient descent after back-propagation of the error gradients. The ADAM algorithm (in Diederik et al.: International Conference on Representation Learning, Methods for Stochastic Optimization, pp. 1–13, 2015) is used to update the weights and the training error is evaluated using a cost function based on the root mean square error.

[0087] To validate this second embodiment, the multilayer perceptron 16 was tested on a system without hydrodynamic focusing using experimental pulse acquisition using blood samples from healthy subjects. Figure 6The distribution of the predicted values ​​of the neural network for the experimental pulses is shown. Two high-concentration regions are observed ([0; 0.1] and [0.9; 1]) forming an intermediate low-concentration region between 0.1 and 0.9. The high-concentration regions correspond to pulses where the model perfectly succeeds in determining the presence of rotation. The intermediate region, where the network has a harder time processing, can be explained by pulse features not present in the digitized training set, such as pulses passing simultaneously or outside the simulated parameter range, or by errors and / or measurement noise. Nevertheless, in the case of a separation threshold of 0.5, the population of rotation-free pulses has a Gaussian distribution, which validates the results.

[0088] Although the applicant has found that the above-described multilayer perceptron is particularly effective, the parameters of the model (number of hidden layers, number of inputs in the input layer, number of neurons in the hidden layer, activation functions, cost functions in training, other hyperparameters of the model) may vary significantly without departing from the spirit of the invention. Further neural network architectures (CNN or other) are also contemplated.

[0089] In the above, ranges and values ​​have been described with respect to a specific application of red blood cells on a machine where measurement parameters may vary (voltage at the electrodes, opening size, flow rate, etc.) Further applications may require adjustments to these ranges and values.

Claims

1. A medical analysis device having a cell impedance signal processing function, the medical analysis device comprising a memory (4), the memory being configured to receive a pulse data set, each pulse data set comprising impedance value data associated with a time stamp each time, the data data collectively representing a curve of the cell impedance value measured when the cell passes through a polarization opening; a calculator (6) configured to process the pulse data set by determining a rotation value indicating whether the cell, from which the pulse data set has been acquired, has undergone a rotation during its passage through the polarization opening; and a classifier (8) configured to provide a given pulse data set to the calculator (6) and classify the given pulse data set into a rotated pulse data set group (10) or a non-rotated pulse data set group (12) using the generated rotation value.

2. The device according to claim 1, wherein The calculator (6) is arranged to process the pulse data set in the following manner: - determining a maximum impedance value of said pulse data set, - by multiplying the maximum impedance value by 0.95] to calculate an upper impedance value, determining in the pulse data set each time mark at which the associated impedance value in the pulse data set is equal to the upper impedance value, and calculating an upper duration corresponding to the maximum duration in the time mark at which the associated impedance value in the pulse data set is equal to the upper impedance value, - calculating a lower impedance value by multiplying the maximum impedance value by a lower coefficient selected in the range [0.1; 0.6], determining in the pulse data set each time mark at which the associated impedance value in the pulse data set is equal to the lower impedance value, and calculating a lower duration corresponding to the maximum duration in the time marks at which the associated impedance value in the pulse data set is equal to the lower impedance value, - Returns a rotation value determined by the calculated ratio of the upper duration and the lower duration.

3. The device according to claim 2, wherein: The upper coefficient is selected in the range [0.8; 0.9].

4. The device according to claim 2, wherein: The lower coefficient is selected in the range [0.3; 0.6].

5. The device according to claim 2, wherein: The upper coefficient is equal to 0.875, and the lower coefficient is equal to 0.

5.

6. The device according to any one of claims 2 to 5, wherein: The calculator (6) returns a rotation value, which is equal to the upper duration divided by the lower duration, and wherein the classifier (8) is configured to classify the pulse data sets whose rotation values ​​are strictly less than 0.45 into the rotated pulse data set group, and classify the pulse data sets whose rotation values ​​are greater than 0.45 into the non-rotated pulse data set group.

7. The device according to claim 1, wherein: The calculator (6) is a neural network, comprising, on the one hand, a converter (14) arranged to process a pulse data set to adapt it for processing by a neural network, arranged to render the time stamps and impedance values ​​dimensionless and to resample the dimensionless data to a selected number of impedance values ​​for the neural network, and, on the other hand, a multilayer perceptron (16) comprising an input layer, at least one hidden layer and an output layer for determining the rotation value, the input layer comprising an input from each value of the resampling performed by the converter (14).

8. The device according to claim 7, wherein: The input layer also includes an input for receiving the width of the pulse data set processed by the converter (14).

9. The device according to claim 7, wherein: The multilayer perceptron (16) comprises a hidden layer including 20 neurons, and the activation function of these neurons is a sigmoid function.

10. The apparatus according to any one of claims 7 to 9, wherein: The multilayer perceptron (16) is trained by stochastic gradient descent after back propagation of error gradients, the training error is calculated according to a cost function based on a root mean square error, and the output is set to 0 for an input corresponding to a cell orientation of an associated pulse data set greater than 0.3 rad of a maximum impedance value among the impedance values ​​of the pulse data set, and is set to 1 otherwise.

Images