Tumor single cell identification method based on residual network, microfluidic device and equipment

Abstract

The invention discloses a tumor single cell identification method based on a residual network, a microfluidic device and microfluidic equipment, and relates to the technical field of microfluidics, the method comprises the following steps: adopting an impedance flow cytometry method to obtain an initial electrical response signal, and then obtaining an electrical impedance signal corresponding to each frequency point through a lock-in amplifier; normalizing the electrical impedance signal of each frequency point, converting the normalized electrical impedance signal into a grey-scale map, and splicing the grey-scale maps of the corresponding frequency points at the same moment to obtain a two-dimensional image feature; constructing a residual network, correspondingly reducing the scale of a residual structure according to the two-dimensional image features, and training the network by using a data set marked with various cell categories; and according to an electrical response signal measured in real time, converting the electrical response signal into a two-dimensional image feature, and inputting the two-dimensional image feature into the trained residual network to obtain the category of the tumor cells. According to the method, the overfitting problem and the network model degradation problem which are easily caused when the impedance peak value is singly used as the input feature to be input into the full-connection neural network for cell classification in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the field of microfluidics, and in particular relates to a tumor single cell recognition method based on a residual network, a microfluidics device and equipment. Background Art

[0002] For the detection technology of tumor cells, existing technologies such as impedance flow cytometry combine microfluidics, electrical impedance technology and flow cytometry. It has the advantage of realizing continuous and non-destructive impedance detection of single cells in flow state under the precise control of microfluidics, and provides a powerful tool for cell type identification and status monitoring. Therefore, how to efficiently analyze the impedance flow data obtained by detection and finally analyze the type of cells has become the focus of research. The electrical impedance signal collected by impedance flow cytometry is a one-dimensional time series. The traditional method uses the peak value of the electrical impedance signal as the input feature to train the network, which will lead to two problems. First, it ignores the characteristics expressed by the process of the impedance signal reaching the peak value, which is not conducive to the subsequent feature extraction. Second, due to the number of characteristic parameters, the network model will also be small, which is easy to cause overfitting. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a tumor single cell identification method based on impedance flow cytometry and residual network. The present invention converts the one-dimensional initial impedance electrical signal into a two-dimensional image feature, retaining more characteristic information of cells passing through the detection electrode, such as cell morphology, size and flow velocity.

[0004] This is achieved through the following solutions:

[0005] The tumor single cell recognition method based on residual network includes the following steps:

[0006] S1: The initial electrical response signal is obtained by impedance flow cytometry, and then the electrical impedance signal corresponding to each frequency point is obtained by a lock-in amplifier;

[0007] S2: normalize the electrical impedance signal at each frequency point and convert it into a grayscale image, and splice the grayscale images of each frequency point corresponding to the same time to obtain a two-dimensional image feature;

[0008] S3: Build a residual network, reduce the size of the residual structure according to the corresponding 2D image features, and train the network using a dataset labeled with multiple cell types;

[0009] S4: According to the electrical response signal measured in real time, after being converted into two-dimensional image features in steps S1 and S2, the features are input into the trained residual network to obtain the category of tumor cells.

[0010] Beneficial effects (1) The present invention obtains impedance signals at each frequency point through a phase-locked amplifier for the original one-dimensional composite electrical impedance signal, and converts it using the Gram angular field algorithm, effectively mapping the target information contained in the one-dimensional signal to features on a two-dimensional image. Compared with the one-dimensional signal, the two-dimensional image can retain more cell information, including cell morphology, size and flow rate, etc., thereby more comprehensively expressing the characteristics of the signal. (2) The present invention combines the ResNet image processing method to extract and classify the features of the converted image, effectively improving the accuracy of machine algorithm model recognition.

[0011] Preferably, the structure of the residual network includes sequentially connected convolution blocks, maximum pooling layers, residual blocks, Flatten layers, Dropout layers, fully connected layers and classification layers, and the input two-dimensional image features are output as cell categories after passing through convolution blocks, maximum pooling layers, residual blocks, Flatten layers, Dropout layers, fully connected layers and classification layers. The residual block includes a 3*3 convolution layer, a batch normalization layer, and a ReLu activation function layer connected in sequence, and L2 regularization is introduced in the convolution layer. The Adam optimizer is used in the network training process, the learning rate is 0.0001, the learning rate scheduler uses the StepLR method, the step size is set to 20, the attenuation factor is set to 0.1, and the loss function is set to cross entropy loss.

[0012] The core residual structure of the original ResNet34 network model is not suitable for the two-dimensional images generated by electrical signals due to the excessive number of layers and large number of parameters, which results in a time-consuming training process and unsatisfactory training results. The present invention appropriately reduces the scale of the residual structure in combination with the size of the generated two-dimensional image, and performs a Dropout operation on the fully connected structure after the Flatten layer. By combining the residual network with the Dropout method, the improved network model not only improves the training speed of the model, but also solves the overfitting problem and network model degradation problem that are prone to occur when the traditional single impedance peak is used as the input feature input into the fully connected neural network for cell classification.

[0013] Preferably, the conversion into a grayscale image is calculated using the Gram angular field method. Furthermore, the overall solution is a tumor single cell recognition method based on impedance flow cytometry and GAF-ResNet.

[0014] A microfluidic device is further proposed, which is used to implement impedance flow cytometry in a tumor single cell identification method and output an initial electrical response signal.

[0015] A computer device is provided, the device comprising

[0016] a memory for storing one or more computer instructions,

[0017] The processor receives the electrical impedance signal corresponding to each frequency point, and the one or more computer instructions are called by the processor and executed in combination with the electrical impedance signal to implement the tumor single cell identification method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0019] Figure 1 It is a top view of the microfluidic chip structure;

[0020] Figure 2 It is a longitudinal cross-sectional view of the microfluidic chip structure along the channel length line;

[0021] Figure 3 One-dimensional four-frequency impedance signal waveform obtained by the lock-in amplifier;

[0022] Figure 4 The two-dimensional grayscale image generated by splicing the one-dimensional impedance signals of four tumor cells after Gram angular field conversion;

[0023] Figure 5 The improved residual network model diagram of the present invention;

[0024] Figure 6 Flow chart of the tumor single cell identification method of the present invention. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with the embodiments. The following embodiments are explanations of the present invention and the present invention is not limited to the following embodiments. Any changes or substitutions within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

[0026] Example 1: Tumor single cell recognition method based on residual network, refer to Figure 6 , including the following steps:

[0027] S1: The initial electrical response signal is obtained by impedance flow cytometry, and then the electrical impedance signal corresponding to each frequency point is obtained by a lock-in amplifier;

[0028] S2: After normalizing the electrical impedance signal at each frequency point, calculate the Gram matrix and convert it into a grayscale image. Then, concatenate the grayscale images of each frequency point corresponding to the same time to obtain a two-dimensional image feature.

[0029] S3: Build a residual network, reduce the size of the residual structure according to the corresponding 2D image features, and train the network using a dataset labeled with multiple cell types;

[0030] S4: According to the electrical response signal measured in real time, after being converted into two-dimensional image features in steps S1 and S2, the features are input into the trained residual network to obtain the category of tumor cells.

[0031] Regarding S1, specifically, first, use a signal generator to give the input electrode, such as Figure 1 , the input electrical signal is a composite sinusoidal signal:

[0032]

[0033] Wherein, A is the signal amplitude, which is set according to the experimental conditions and is set to 2V in the present invention, f i They are 100kHz, 300kHz, 700kHz and 1200kHz respectively.

[0034] The output electrode detects the response composite electrical signal, which is then connected to a lock-in amplifier. The lock-in amplifier detects the signal based on the following principle:

[0035]

[0036] For the desired frequency f i The DC signal obtained after low-pass filtering is the desired frequency f i The corresponding impedance electrical signal. Figure 3 , based on the impedance electrical signals at four frequency points (100kHz, 300kHz, 700kHz, 1200kHz) collected by the phase-locked amplifier, the sampling frequency is 3000 points per second. Based on the peak detection, the time period when tumor cells pass between the detection electrodes is detected, and the impedance electrical signals of the corresponding time are extracted.

[0037] There are 41 sampling points for the extracted electrical signals at each frequency point. There are 3000 samples for each of the four types of tumor cells (PC-3, MCF-7, Hepg-2, HeLa), of which 70% are used as training sets and 30% as test sets.

[0038] S2: The impedance electrical signal output by the lock-in amplifier is a one-dimensional impedance electrical signal. The one-dimensional impedance electrical signal is further normalized by using a maximum-minimum calibrator to normalize the original value to [-1,1]:

[0039]

[0040] Where X is a one-dimensional impedance electrical signal sequence, min(X) represents the minimum value of the sequence, and max(X) represents the maximum value of the sequence. Indicates the normalized value.

[0041] Map the normalized data to polar coordinates:

[0042]

[0043] in, is the normalized time series, is the value of its time series position, t i Indicates the time sequence position, θ i is the angle mapped to the polar coordinates, N is the number of time points contained in the time series data, and the radius r i Keeping the time dependence of the sequence, the polar angle θ i The numerical relationship is preserved.

[0044] According to the polar coordinate transformation, GAF is used to reconstruct the normalized time series to generate an image. The GAF algorithm has periodic, trendy or regular data, such as sensor data. It is more suitable for the electrical signal obtained in the present disclosure.

[0045] As a preferred solution, the Gram matrix generated in the present disclosure is based on the Gram angle difference field (GADF), as follows:

[0046] G i,j = sin(θ i -θ j )

[0047] Among them, G i,j Represents the element value of the i-th row and j-th column of the matrix.

[0048] According to the generated Gram matrix, the element value is converted to [0,255] through linear mapping:

[0049]

[0050] in, It is the image grayscale matrix, which can be used to generate grayscale image.

[0051] By splicing four grayscale matrices of different frequencies, a grayscale image can be generated as the input feature of the network model, such as Figure 4 As shown, the pixel size of the image is 82*82.

[0052] S3: Build an improved residual network model, such as Figure 5As shown in the figure. The training process uses four types of tumor cells (PC-3, MCF-7, Hepg-2, HeLa). Each type of cell has 3000 samples, of which 70% are used as training sets and 30% are used as test sets. The labeling process is pre-labeled, that is, the four types of cells are pre-separated, and the sample solutions of the four types of tumor cells are measured and tested separately.

[0053] The improved residual network model first passes through a 3*3 convolution block and a 2*2 maximum pooling layer; then passes through a residual block, which is composed of a 3*3 convolution block, a batch normalization layer, a ReLu nonlinear layer, a 3*3 convolution block, and a batch normalization layer connected in sequence. Then it passes through a flattening layer to expand the data into one dimension, and then passes through a fully connected layer to finally output the result.

[0054] Among them, the convolution block samples L2 regularization, and the fully connected layer undergoes a Dropout operation with a probability of 0.5 to prevent overfitting. During the training process, the Adam optimizer is used, the learning rate is 0.0001, the learning rate scheduler uses the StepLR method, the step size (step_size) is set to 20, the decay factor (gamma) is set to 0.1, the loss function is set to cross entropy loss, the training cycle is set to 200, and the batch size (batch size) is set to 32.

[0055] The above improved network model has advantages in the tumor single cell recognition method. Taking the traditional ResNet-18 as an example, the network has 8 residual blocks, and the convolution kernel of each residual block uses a 3*3 convolution kernel, and the number of convolution kernels is more than 64. For the 86*86 grayscale image used in the present invention, the use of a conventional network will have degradation problems, resulting in large amount of calculation, slow calculation speed but no improvement in effect. Therefore, the residual structure convolution kernel used in the present invention is 3*3, the number is 8, and at the same time, the fully connected layer part undergoes a Dropout operation with a probability of 0.5, that is, the neuron has a probability of 0.5 being abandoned, thereby avoiding the overfitting problem.

[0056] The residual structure used in the present invention is improved according to the input features, and the scale of the residual structure is reduced, making the model more efficient. The following table shows the superiority of the method disclosed in this embodiment compared with the traditional method. For example, the impedance peak value + FCNN is used to directly identify the electrical signal, and the accuracy is low. Although the impedance sequence value + LSTM is used to identify the electrical signal, the accuracy has reached more than 91%, but it has not met expectations. The method of first converting to a GAF grayscale image and then performing machine learning can effectively increase the accuracy, and the accuracy of the improved residual network has increased by nearly 2.5%.

[0057] method Recognition accuracy Impedance peak + FCNN 86.53％ Impedance sequence value + LSTM 91.71％ GAF grayscale image + traditional ResNet 94.67％ GAF grayscale image + improved ResNet 96.95％

[0058] S4: After network training is completed, the electrical signal threshold for cell detection is set according to the electrical response signal measured in real time. When the electrical signal is lower than the threshold, it means that a cell has passed through. At this time, the time corresponding to the lowest point of the electrical signal is obtained through peak detection. Based on this time, the signal sequence containing the cell passing through is expanded to both ends of the time axis. The signal sequence is generated into a grayscale feature map through steps S1 and S2, and then the identified cell category is obtained through step S3.

[0059] Embodiment 2:

[0060] A microfluidic device is disclosed. The microfluidic chip is used to implement the impedance flow cytometry in the tumor single cell identification method based on residual network described in Example 1, and output an initial electrical response signal. It includes a PDMS channel manufactured using micro-nano processing technology, and a gold coplanar electrode is arranged on the channel. The width of the PDMS channel is the same as the width of the coplanar electrode, and the width is determined according to the size of the cell flowing through. The gold coplanar electrode means that all the gold electrodes are on the same surface of the substrate and are at the same height or level as each other.

[0061] Preferably, the width of the channel and the width of the coplanar electrode are 20 μm to ensure that the cells pass smoothly and accurately. Figure 1 and Figure 2 As shown. The choice of width is related to the size of the cells. The diameter of the selected test cells is 12-18μm. Therefore, in order to allow the cells to pass smoothly without clogging and improve the measurement accuracy, the size is as small as possible while ensuring the smooth passage of the cells. Taking comprehensive consideration, the present invention selects a sensing channel width of 20μm. The size of the coplanar electrode has less influence than the channel width. The main consideration for the electrode size selected by the present invention is to ensure the transmission of electrical signals. Therefore, the electrode width cannot be too small. Therefore, an electrode width of 20μm is selected, and the test signal can be effectively transmitted.

[0062] Embodiment 3:

[0063] A computer device comprises a memory for storing one or more computer instructions; a processor for receiving electrical impedance signals corresponding to each frequency point, wherein the one or more computer instructions are called by the processor and executed in combination with the electrical impedance signals to implement the tumor single cell identification method based on the residual network as described in Example 1.

[0064] Exemplarily, a computer program may be divided into one or more modules / units, one or more modules / units are stored in a memory and executed by a processor, and the input interface and output interface complete the I / O interface transmission of data to complete the present invention. One or more modules / units may be a series of computer program instruction segments that can perform specific functions, and the instruction segments are used to describe the execution process of the computer program in a computer device.

[0065] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may include, but is not limited to, a memory and a processor. Those skilled in the art may understand that this embodiment is only an example of a computer device and does not constitute a limitation on the computer device. The computer device may include more or fewer components, or a combination of certain components, or different components. For example, the computer device may also include an input device, a network access device, a bus, etc.

[0066] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0067] The memory can be an internal storage unit of a computer device, such as a hard disk or memory of a computer device. The memory can also be an external storage device of a computer device, such as a plug-in hard disk equipped on a computer device, a SmartMedia Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Furthermore, the memory can also include both an internal storage unit of a computer device and an external storage device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data in an output device. The aforementioned storage media include various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory ROM, a random access memory RAM, a disk or an optical disk.

Claims

1. A tumor single cell recognition method based on residual network, characterized in that: The following steps are involved: S1: The initial electrical response signal is obtained by impedance flow cytometry, and then the electrical impedance signal corresponding to each frequency point is obtained by a lock-in amplifier; S2: normalize the electrical impedance signal at each frequency point and convert it into a grayscale image, and splice the grayscale images of each frequency point corresponding to the same time to obtain a two-dimensional image feature; S3: Build a residual network, reduce the size of the residual structure according to the corresponding 2D image features, and train the network using a dataset labeled with multiple cell types; S4: According to the electrical response signal measured in real time, after being converted into two-dimensional image features in steps S1 and S2, the features are input into the trained residual network to obtain the category of tumor cells.

2. The method for single-cell tumor recognition based on residual network according to claim 1, characterized in that: The impedance flow cytometry method comprises: Sinusoidal superposition electrical signals of several frequencies including 100kHz, 300kHz, 700kHz, and 1200kHz are input to the input electrode, and the initial electrical response signal of the corresponding frequency is output from the electrode.

3. The method for single-cell tumor recognition based on residual network according to claim 1, characterized in that: The structure of the residual network includes a convolution block, a maximum pooling layer, a residual block, a Flatten layer, a Dropout layer, a fully connected layer and a classification layer connected in sequence. The input two-dimensional image features pass through the convolution block, the maximum pooling layer, the residual block, the Flatten layer, the Dropout layer, the fully connected layer and the classification layer to output the cell category.

4. The method for single-cell tumor recognition based on residual network according to claim 3, characterized in that: The residual block includes a 3*3 convolutional layer, a batch normalization layer, a ReLu nonlinear layer, a 3*3 convolutional block and a batch normalization layer connected in sequence.

5. The method for single tumor cell recognition based on residual network according to claim 1 or 4, characterized in that: In step S3, the Adam optimizer is used in the network training process, the learning rate is 0.0001, the learning rate scheduler adopts the StepLR method, the step size is set to 20, the decay factor is set to 0.1, and the loss function is set to the cross entropy loss.

6. The method for single-cell tumor recognition based on residual network according to claim 1, characterized in that: The conversion into a grayscale image is calculated using the Gram angular field method.

7. A microfluidic device, characterized in that: The microfluidic device is used to implement the impedance flow cytometry in the tumor single cell identification method based on residual network described in any one of claims 1 to 6, and output an initial electrical response signal.

8. A microfluidic device according to claim 7, characterized in that: It includes a PDMS channel, on which a gold coplanar electrode is arranged, and the width of the PDMS channel is the same as that of the coplanar electrode, and the width is determined according to the size of the cells flowing through.

9. A microfluidic device according to claim 8, characterized in that: The width of the channel and the width of the coplanar electrode are 20 μm.

10. A computer device, characterized in that: include a memory for storing one or more computer instructions, A processor receives an electrical impedance signal corresponding to each frequency point, wherein the one or more computer instructions are called by the processor and executed in combination with the electrical impedance signal to implement the tumor single cell recognition method based on the residual network as described in any one of claims 1 to 6.

Images