An overlapping wave segmentation method based on a deep convolutional network

Abstract

The present invention relates to the field of medical technology, and specifically relates to an overlay wave segmentation method based on a deep convolutional network. The features of the histogram formed by the overlapping peaks are extracted through the deep convolutional network, and the optimal segmentation point of the two peaks on the histogram is identified through the features. After the network model is trained and deployed to the terminal for calculating the segmentation line of the histogram, the result can be obtained only by performing a single forward convolution operation, without multiple iterations, derivations, and fitting calculations, and the computational amount is small. At the same time, the calculated segmentation point can make the areas on both sides of the segmentation point most conform to the areas after the overlapping peaks are separated. That is, when using the present invention in flow cytometry technology, higher cell classification accuracy can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to an overlapping wave segmentation method based on a deep convolutional network. Background Art

[0002] A flow cytometer is an important instrument for analyzing human cell distribution data. It calculates the number and proportion of various types of cells in a measured specimen by analyzing the histograms and scatter plots of blood and urine, and indicates whether there may be lesions in the human body, providing a basis for doctors' diagnosis.

[0003] When the flow cytometer classifies two types of cells through a single signal, if the two types of cells overlap in this signal, then when this signal is statistically presented on the histogram, it will be shown as a histogram where two waveforms overlap and add up.

[0004] Currently, the flow cytometer usually uses the lowest point between the two wave peaks as the segmentation point for such overlapping histograms. The statistics to the left of the segmentation point are counted as one type of cell, and the statistics to the right of the segmentation point are counted as another type of cell. This statistical method has a small computational amount and is easy to implement, but there are classification errors, and as the overlapping range of the two waves increases, the error will also gradually increase, ultimately affecting the accuracy of the flow cytometer detection. Summary of the Invention

[0005] The purpose of the present invention is to provide an overlapping wave segmentation method based on a deep convolutional network, aiming to solve the technical problems in the prior art that the flow cytometry statistical method has a small computational amount and is easy to implement, but there are classification errors, and as the overlapping range of the two waves increases, the error will also gradually increase, ultimately affecting the accuracy of the flow cytometer detection.

[0006] To achieve the above purpose, the present invention provides an overlapping wave segmentation method based on a deep convolutional network, including the following steps:

[0007] Using a flow cytometer to collect the individual histograms of each detection signal of various types of cells, and processing to obtain a sample data set;

[0008] Constructing 7 sequentially arranged feature extraction units, and then constructing a feature processing unit to form a deep convolutional network;

[0009] Using the sample data set to complete the training of the model;

[0010] After the model training is completed, it can be put into use to output the segmentation point.

[0011] Among them, the step of using a flow cytometer to collect the individual histograms of each detection signal of various types of cells and processing to obtain a sample data set includes:

[0012] Collect individual histograms of each detection signal of various types of cells using a flow cytometer;

[0013] Manually select two appropriate histograms, select the signal waveforms with overlapping parts, and the superimposed histogram obtained by adding the two histograms should have two peaks, and there should be a trough between the peaks. Here, the histogram with a smaller peak position is called A, the other histogram is called B, and the overlapping histogram formed after superposition is called C;

[0014] Calculate an optimal segmentation point p on C such that the area to the left of point p in C is closest to the area of A, and the area to the right of point p is closest to the area of B;

[0015] Let the peak coordinate of A be a, the peak coordinate of B be b. After A and B are superimposed to form C, the peak coordinate of C in A is a′, and the peak coordinate of B is b′; On C, intercept from a′ to b′ to obtain the intercepted histogram segment C′, and its length is L C′ = b′ - a′, and at this time, the segmentation point p′ on C′ = p - a′;

[0016] Normalize C′ to 0 - 1 to obtain C1′;

[0017] Use linear interpolation to adjust the length of C1′ to 64 to obtain a histogram Hist64 with a length of 64;

[0018] Calculate the segmentation point p on Hist64 Hist64 .

[0019] Among them, in the step of using a flow cytometer to collect individual histograms of each detection signal of various types of cells and processing to obtain a sample data set:

[0020] Use multiple groups of samples to calculate repeatedly to obtain at least 1000 copies of Hist64 and the corresponding segmentation point p Hist64 of the sample data set.

[0021] Among them, in the step of constructing 7 sequentially arranged feature extraction units, and then constructing a feature processing unit to form a deep convolutional network:

[0022] The input of the first feature extraction unit is data with a length of 64, that is, the processed histogram array, and the inputs of the second to seventh feature units are the output values of the previous feature extraction unit.

[0023] Among them, in the step of constructing 7 sequentially arranged feature extraction units and then constructing a feature processing unit to form a deep convolutional network: The feature processing unit is composed of a first fully connected layer, a ReLU unit, and a second fully connected layer. The first fully connected layer 1 has 512 inputs and 512 outputs, that is, the maximum value of n is 512, and the maximum value of i is 512. This layer is used to combine the local features extracted by the feature extraction units to obtain 512 overall features; the ReLU unit is used to rectify the 512 overall features output by the first fully connected layer and suppress low-validity features with a probability less than 0; the second fully connected layer receives the 512 overall features processed by the ReLU unit and outputs 1 result, with the maximum value of n being 512 and the maximum value of i being 1.

[0024] Among them, after the model training is completed and it can be put into use, the steps for outputting the segmentation point include:

[0025] After the model training is completed and it can be put into use, for the overlapping peak histogram C to be segmented, the highest point coordinate of its A peak is a′, and the highest point coordinate of the peak is b′; on C, intercept from a′ to b′ to obtain the intercepted histogram segment b′, and its length is L C′ = b′ - a′;

[0026] Process C′ to obtain the scaling coefficient K of the histogram Hist64;

[0027] Input Hist64 into the model to obtain the segmentation point p of Hist64 predicted by the model Hist64 ;

[0028] Calculate the optimal segmentation point of the overlapping peak histogram C through p = a′ + p Hist64 ÷K.

[0029] A method for segmenting overlapping waves based on a deep convolutional network according to the present invention extracts the features of the histogram formed by overlapping peaks through the deep convolutional network, and identifies the optimal segmentation point of the two peaks on the histogram through the features. After the network model training is completed and it is deployed to the terminal for calculating the segmentation line of the histogram, only one forward convolution operation is required to obtain the result, without multiple iterations, derivations, and fitting calculations, and the computational amount is small. At the same time, the calculated segmentation point can make the areas on both sides of the segmentation point most conform to the areas after the overlapping peaks are separated. That is, when using the present invention in flow cytometry technology, higher cell classification accuracy can be obtained. Description of the Drawings

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0031] Figure 1 It is the volume histogram of Class A cells provided by the present invention and its peak coordinates.

[0032] Figure 2 It is the volume histogram of Class B cells provided by the present invention and its peak coordinates.

[0033] Figure 3 is Figure 1 and Figure 2 the overlapping histogram formed by superposition.

[0034] Figure 4 is Figure 3 the corresponding histogram with a length of 64 and its corresponding segmentation points.

[0035] Figure 5 It is the overall network structure diagram of the deep convolutional network provided by the present invention.

[0036] Figure 6 It is the structural schematic diagram of the feature processing unit provided by the present invention.

[0037] Figure 7 It is the loss line chart of model training.

[0038] Figure 8 It is the overlapping histogram after the model is put into use.

[0039] Figure 9 is Figure 8 the corresponding histogram segment C1'.

[0040] Figure 10 is Figure 9 the corresponding histogram with a length of 64 and its segmentation points.

[0041] Figure 11 It is the step flowchart of an overlapping wave segmentation method based on a deep convolutional network provided by the present invention. Detailed implementation manners

[0042] The following will describe in detail the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation of the present invention.

[0043] Please refer to Figures 1 to 11 , the present invention provides an overlapping wave segmentation method based on a deep convolutional network, comprising the following steps:

[0044] S1: Use a flow cytometer to collect individual histograms of various detection signals of various types of cells, and process to obtain a sample data set;

[0045] S2: Construct 7 sequentially arranged feature extraction units, and then construct a feature processing unit to form a deep convolutional network;

[0046] S3: Use the sample data set to complete the training of the model;

[0047] S4: After the model training is completed, it can be put into use to output the segmentation points.

[0048] Among them, the specific steps of step S1 are:

[0049] S11: A flow cytometer measures a large number of individual A cell samples and individual B cell samples. Obtain a large number of A cell volume histograms and B cell volume histograms;

[0050] S12: As Figure 1 and Figure 2 , manually select the A and B histograms with overlapping volumes, and superimpose them to form an overlapping histogram C. Let the peak coordinate of the A cell volume histogram be a, and the peak coordinate of the B cell volume histogram be b. After A and B are superimposed to form C, the peak coordinate of A in C becomes a', and the peak coordinate of B becomes b';

[0051] S13: As Figure 3 , calculate an optimal segmentation point on C such that the area to the left of point p in C (i.e., the shaded area in the figure) is closest to the area of A, and the area to the right of point p is closest to the area of B;

[0052] S14: Intercept from a' to b' on C to obtain the intercepted histogram segment C', and its length is L C′ = b' - a', at this time the segmentation point p' on C' = p - a', it can be seen that the optimal segmentation point p' is not the lowest point between the two peaks;

[0053] S15: Normalize C' to 0-1 to obtain C1', and the formula is: C1'(i) = C'(i) / Max(C'), where i is the coordinate on C', i ∈ [0, L C′ -1], C1'(i) is the value of C1' at position i, and Max(C') is the maximum value in C'.

[0054] S16: Use linear interpolation to adjust the length to 64, obtaining a histogram with a length of 64. The formula is as follows:

[0055]

[0056] where j is the coordinate on Hist64, j ∈ [0, 63], and Hist64(j) represents the value at position j;

[0057] j k is the value of the coordinate j multiplied by the scaling factor, j k = j × k;

[0058] K is the scaling factor,

[0059] j k0 is the result of rounding down the coordinate j k ;

[0060] j k1 is the result of rounding up the coordinate j k ;

[0061] C′(j k0 ) represents the position of C′ at position j k0 ;

[0062] S17: As Figure 4 , calculate the segmentation point p Hist64 on Hist64. At this time, the segmentation point is p Hist64 = p * K;

[0063] S18: Repeat steps S12 to S17 multiple times to obtain a sample data set of 2000 Hist64s and their corresponding segmentation points p Hist64 .

[0064] In step S2, refer to Figures 5 to 7, a deep convolutional network is constructed. The input of feature extraction unit 1 is an array of length 64, which is the processed histogram array. The inputs of feature extraction units 2 - 7 are all the output values of the previous feature extraction unit. Taking feature extraction unit 1 as an example to illustrate its internal working principle and function: input the one-dimensional histogram array data of length 64; since, the input data is processed by a one-dimensional convolutional layer with 8 convolutional kernels, obtaining an output of dimension, that is, this step expands out features; these features have both positive and negative values according to the weights of the convolutional kernels. Generally, positive-valued features are considered better features. Therefore, a ReLU module is connected subsequently to set features less than 0 to 0 and keep features greater than 0 unchanged; it is used to expand the perception range of the features extracted by convolution to output more refined and deeper-level features. The subsequent feature extraction units 2 - 7 have exactly the same working principle as feature extraction unit 1, except that the input data of each unit is different and the number of convolutional kernels is different. Their functions or purposes are all to extract deeper-level histogram features with a wider perception range and higher effectiveness based on the features of the previous layer.

[0065] The subsequent feature extraction units 2 - 7 have exactly the same working principle as feature extraction unit 1, except that the input data of each unit is different and the number of convolutional kernels is different. Their functions or purposes are all to extract deeper-level histogram features with a wider perception range and higher effectiveness based on the features of the previous layer.

[0066] The working principle of the fully connected layer is to perform weighted summation through linear weighted summation, and finally output a single value by weighted summing multiple input feature values. When there are multiple sets of weighted summations in the fully connected layer, the fully connected layer will output multiple values.

[0067] The formula for linear summation is:

[0068] a i =w i1 ×x1 + w i2 ×x2 +... + w in ×x n +b

[0069] In the formula, i represents that this is the i-th linear summation of this fully connected layer, and a i is the output value of the i-th linear summation;

[0070] w i1 is the weight of the first feature of the i-th linear summation, w i2 is the weight of the second feature of the i-th linear summation... w in is the weight of the n-th feature of the i-th linear summation;

[0071] x1 is the first input feature of the fully connected layer, x2 is the second input feature... x nThe nth input feature, where b is the bias value.

[0072] The convolution stride of each one-dimensional convolution layer of the 7 feature extraction units is 1, and the convolution kernel size is 3. Before convolution, the input is extended by 1 bit to the left and right to ensure that the width of the input and output of the one-dimensional convolution layer will not decrease due to convolution.

[0073] Parameter table of each feature extraction unit module:

[0074] Feature extraction unit <![CDATA[n i > <![CDATA[w i > <![CDATA[n′ i > <![CDATA[w′ i > F1 1 64 8 64 F2 8 64 16 32 F3 16 32 32 16 F4 32 16 64 8 F5 64 8 128 4 F6 128 4 256 2 F7 256 2 512 1

[0075] Note: In the above table, the input data length and output data length of the first feature extraction unit are the same, indicating that its max pooling layer does not participate in data processing.

[0076] After being processed by 7 feature extraction units, the module will output a tensor of 512×1, that is, a 512-dimensional feature vector.

[0077] Such as Figure 6 , the feature processing unit receives the feature data extracted by the feature extraction unit, and outputs a tensor of 1 after fully connected processing. This tensor is the segmentation point coordinate output by the network. The feature processing unit consists of a first fully connected layer, a ReLU unit, and a second fully connected layer.

[0078] Parameter table of the feature processing unit module:

[0079] Feature extraction unit Input Input Fully connected layer 1 512 512 ReLU 512 512 Fully connected layer 2 512 1

[0080] In step S3, the 2000 datasets are divided into 1500 training sets and 500 test sets.

[0081] The batch size for this training is set to 128;

[0082] The loss function uses the MSELoss loss function in PyTorch;

[0083] The optimization algorithm uses Adam, which can find the optimal solution more quickly;

[0084] The initial learning rate is set to 0.0001, and after every 5 rounds of training, the learning rate is halved;

[0085] After training 10 batches of training set samples during the training process, the validation set is input into the model to calculate the validation set loss function once and output it to TensorBoard for display;

[0086] The model training will end after 40 rounds of training.

[0087] In step S4, the following figure is used as an example to calculate the optimal segmentation point in this implementation, such as Figure 8, in the figure, the peak coordinate of peak A is a' = 30, and the peak coordinate of peak B is b' = 110. On C, intercept from a' to b', and obtain the intercepted histogram segment b', whose length is L C′ = b' - a', and calculate to obtain the length L C′ = 80, p' = p - a' = 15

[0088] Such as Figure 9 , normalize C' to 0 - 1 to obtain C1';

[0089] Such as Figure 10 , use the linear interpolation method to adjust the length of C1' to 64 to obtain the histogram Hist64 with a length of 64. In this example, the scaling factor:

[0090]

[0091] Input Hist64 into the model to obtain the output value p Hist64 = 12.183237

[0092] Convert p according to the scaling factor K and a' Hist64 The coordinate on C:

[0093] p = a' + p Hist64 ÷K = 45.229.04626

[0094] After calculation, the actual optimal segmentation point coordinate p of C is 45. The error of this calculation result is about 0.2, and the lowest point coordinate between the two peaks on C is 50, with an error of 5. Therefore, the accuracy of segmenting the histogram by the present invention is obviously higher

[0095] What is disclosed above is only a preferred embodiment of the present invention. Of course, the scope of rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention

Claims

1. An overlapping wave segmentation method based on a deep convolutional network, characterized in that, It includes the following steps: Collect the individual histograms of various detection signals of various cells by using a flow cytometer, and process to obtain a sample data set; Construct 7 sequentially arranged feature extraction units, and then construct a feature processing unit to form a deep convolutional network; Use the sample data set to complete the training of the model; After the model training is completed, it can be put into use to output the segmentation point; The step of collecting the individual histograms of various detection signals of various cells by using a flow cytometer and processing to obtain a sample data set includes: Collect the individual histograms of various detection signals of various cells by using a flow cytometer; Manually select two appropriate histograms, select the signal waveforms with overlapping parts, and the superimposed histogram obtained by adding the two histograms should have two peaks, and there should be a trough between the peaks. Here, the histogram with a smaller peak position is called A, the other histogram is called B, and the overlapping histogram formed after superposition is called C; Calculate an optimal segmentation point p on C such that the area to the left of point p in C is closest to the area of A, and the area to the right of point p is closest to the area of B; Let the peak coordinate of A be a, and the peak coordinate of B be b. After A and B are superimposed to form C, the peak coordinate of C in A is a′, and the peak coordinate of B is b′; on C, intercept from a′ to b′ to obtain the intercepted histogram segment C′, whose length is L C′ = b′ - a′. At this time, the segmentation point p′ on C′ = p - a′; Normalize C′ to 0-1 to obtain C1′; Use linear interpolation to adjust the length of C1′ to 64 to obtain a histogram Hist64 with a length of 64; Calculate the splitting point p on Hist64 Hist64 ; The step that after the model training is completed, it can be put into use to output the segmentation point includes: After the model training is completed, it can be put into use. For the overlapping peak histogram C to be segmented, the highest point coordinate of its A peak is a′, and the highest point coordinate of the peak is b′; on C, intercept from a′ to b′ to obtain the intercepted histogram segment b′, and its length is L C′ = b′ - a′; Process C′ to obtain the scaling coefficient K of the histogram Hist64; Input Hist64 into the model to obtain the segmentation point p of Hist64 predicted by the model Hist64 ; Calculate the optimal segmentation point of the overlapping peak histogram C by p = a' + p Hist64 ÷K 2. The overlapping wave segmentation method based on a deep convolutional network according to claim 1, characterized in that In the step of collecting the individual histograms of various detection signals of various cells by using a flow cytometer and processing to obtain a sample data set: By repeatedly calculating using multiple groups of samples, at least 1000 copies of Hist64 and the corresponding segmentation point p are obtained Hist64 of the sample data set.

3. The overlapping wave segmentation method based on a deep convolutional network according to claim 2, characterized in that, In the step of constructing 7 sequentially arranged feature extraction units, and then constructing a feature processing unit to form a deep convolutional network: The input of the first feature extraction unit is data with a length of 64, that is, the processed histogram array, and the inputs of the second to seventh feature units are the output values of the previous feature extraction unit.

4. The overlapping wave segmentation method based on a deep convolutional network according to claim 3, characterized in that In the step of constructing 7 sequentially arranged feature extraction units, and then constructing a feature processing unit to form a deep convolutional network: The feature processing unit is composed of a first fully connected layer, a ReLU unit, and a second fully connected layer. The first fully connected layer 1 has 512 inputs and 512 outputs, that is, the maximum value of n is 512, the maximum value of i is 512. This layer is used to combine the local features extracted by the feature extraction unit to obtain 512 overall features; The ReLU unit is used to rectify the 512 overall features output by the first fully connected layer and suppress the low-validity features with a probability less than 0; The second fully connected layer receives the 512 overall features processed by the ReLU unit and outputs 1 result, with the maximum value of n being 512 and the maximum value of i being 1.

Images