CN114494197A - Cerebrospinal fluid cell identification and classification method for small-complexity sample - Google Patents

Abstract

The invention discloses a complex small sample cerebrospinal fluid cell identification and classification method. The specific method steps are as follows: S1: use a microscope splicing imaging platform to obtain cell images under a glass slide sample, including monocytes, lymphocytes and neutrophils image; S2: Preprocess the resulting image set. In view of the background impurities caused by lens contamination or improper operation in the collected cell samples, as well as the problem of overlapping and adhesion of individual cells, the main purpose is to filter and denoise the image, remove irrelevant factors in the image, and correct Adhesive cells are separated; S3: Transfer training of the model for a small sample set; S4: According to the obtained model, use the BP algorithm to reverse fine-tune the weights and thresholds of the model; S5: Use the trained model to perform training on the test set Recognition of cerebrospinal fluid cell images and further optimization of the algorithm. The invention can effectively identify different types of human cerebrospinal fluid cells.

Description

A complex small sample cerebrospinal fluid cell identification and classification method

technical field

The invention relates to the technical field of medical cell identification and classification, in particular to a complex small sample cerebrospinal fluid cell identification and classification method.

Background technique

Cerebrospinal fluid (CSF) cytology, one of the most important tools for neurologists, includes gross cell counts and cytological classifications, providing important first-hand information on the central nervous system and the range of pathological conditions it covers. CSF samples need to be processed immediately, if possible within 1 hour of collection. In normal CSF cells are mainly T lymphocytes, a small amount of mononuclear macrophages, and occasionally B lymphocytes; the number of cells increases significantly, and neutrophils are more common under microscope in bacterial meningitis, and further search for cells is required. Evidence of endobacteria; cellular background dominated by lymphocytes and monocytes is more common in viral infections and chronic inflammation; can occur in tuberculous meningitis in the background of promiscuous cellular responses; phagocytosis of erythrocytes or macrophages containing fragments of hemoglobin degradation products Phage cells (the latter are called hemosiderin cells) are all indicative of old subarachnoid hemorrhage; when atypical cells are found under the microscope to suspect a tumor, a comprehensive judgment should be combined with clinical and immunocytochemical staining.

In recent years, metagenomic sequencing technology has received extensive attention, and it has certain value in the detection of pathogens of infectious diseases in the central nervous system, but there are still some shortcomings: the specimen is easily contaminated, which affects the detection results and limits the overall sensitivity of pathogen detection; There are false negative and false positive results, and even the test results cannot be analyzed; the test cost is high, which limits its wide application. Therefore, metagenomic sequencing cannot yet replace traditional diagnostic methods.

To date, most clinical laboratories still count and classify cells in CSF by hand, using the method of counting single nuclei (including lymphoid and monocytes) and multiple nuclei directly under the microscope based on their nuclear morphology. A total of 100 cells were counted. This method is cumbersome, time-consuming and labor-intensive. Due to the difference in proficiency and standardization among different operators, it has great subjectivity. The results have low repeatability and large errors. The return time is long, it cannot meet the clinical needs well, and it is not suitable for the development of large-scale clinical work in modern hospitals. Cerebrospinal fluid samples are small in volume compared to blood and urine. In manual counting, the sampling amount is small, and the accuracy of counting cannot be guaranteed. If the automated cell detection of cerebrospinal fluid samples can be realized or partially realized, the above problems can be solved to a certain extent. At present, there is no special automatic analyzer for counting and classifying cerebrospinal fluid cells.

With the development of automated cell detection technology, in recent years, many researchers have tried to use various cell analyzers (such as automatic urine sediment analyzer and blood analyzer) to count and analyze cerebrospinal fluid cells. Now some new models of blood cell analyzers have added the function of body fluid cell analysis, making it possible for laboratories to automatically count and classify cells in pleural fluid, ascites, etc. However, due to the particularity of cerebrospinal fluid, the sample size is small, and the principle and internal design of various instruments limit their application in the detection of cerebrospinal fluid samples. The presence of bacterial impurities and cell adhesion has a huge impact on the identification and classification of cerebrospinal fluid cells.

The automatic identification technology of cerebrospinal fluid cells based on deep learning can help neurologists to quickly establish a more scientific differential diagnosis model, which can reduce the adverse effects of subjective factors on the results of microscopic examinations, help doctors to count and classify cells, and greatly improve the diagnosis. And it can be integrated with the resources of high-level medical institutions, so that the overall diagnosis model tends to be standardized and unified, greatly improving the radiation effect of high-quality medical resources to primary medical institutions, and improving the differential diagnosis level of primary hospitals. Therefore, the construction of an automatic identification system of cerebrospinal fluid cells based on deep learning is of great significance for improving the diagnosis rate of infectious diseases in the central nervous system, solving the problems of regional medical differences, low seniority and misdiagnosis by primary physicians, thus ultimately benefiting the majority of patients.

SUMMARY OF THE INVENTION

1. Technical problems to be solved

The purpose of the present invention is to solve the problems in the prior art due to the particularity of cerebrospinal fluid, the small sample size, and the principles and internal design of various instruments themselves limit their application in the detection of cerebrospinal fluid samples. From the sampling point of view, the sample slides have certain bacterial impurities and cell adhesion, which has a huge impact on the identification and classification of cerebrospinal fluid cells. A complex small sample cerebrospinal fluid cell identification and classification method is proposed.

2. Technical solutions

In order to achieve the above object, the present invention adopts the following technical solutions:

A complex small sample cerebrospinal fluid cell identification and classification method, comprising the following steps:

S1: use the microscope automatic scanning platform to acquire the image of the sample slide to obtain a complete image set of the cerebrospinal fluid cell slide with multiple cells;

S2: Preprocess the obtained image set, filter and denoise the image, remove irrelevant factors in the picture, and separate the cells that adhere to each other, and form the obtained sample image set into training set and test set in batches;

S3: Carry out model transfer training for small sample sets, and use deep learning networks trained in similar fields to perform transfer learning on small sample data sets;

S4: Use the BP algorithm to fine-tune the weights and thresholds of the trained model to further optimize the model;

S5: Input the test set into the model, and the output result is the cerebrospinal fluid cell identification result.

Preferably, in the step S1, the cerebrospinal fluid cell slide is placed on the electric translation stage of the microscope, and a software system is used to locate the diagonal coordinates of the scanning range of the slide, determine the scanning range of the image, and record the scanned image The size of the range, the software system platform is used to stitch the collected images to obtain a complete picture of the cell sample, and this step is repeated for the subsequent slide image collection.

Preferably, the specific steps of preprocessing the image set in S2 are:

Step 1: For the irrelevant impurities in the background of the sample, firstly separate the background of the image, obtain the binary image by the maximum inter-class variance method, and use the morphological opening operation to smooth the contour of the target in the binary image. In the open operation of learning, the impurities in the background that are not the target can be removed, and finally the Canny boundary detection algorithm is used to obtain the contour edge information of the target;

Step 2: Use the concave point detection to segment the adhesion cells. The concave point of the adhesion cell refers to the local maximum curvature point in the concave area formed by the adhesion of two or more quasi-circular objects due to overlapping with each other. For near-circular images, there is no sudden change in curvature unless there are two or more cells;

Step 3: Ellipse fitting. In order to obtain the outline boundary of the sticking target lost due to sticking, the algorithm uses the prior knowledge that the target generally appears to be a circle, and uses the ellipse fitting method based on the least squares method for fitting to complete Adhesive segmentation.

Preferably, the multi-layer ResNet model that has been trained in other fields is used in the S3, the part in front of the fully connected layer of the model is intercepted, the output part sets three output nodes according to the required classification type, and then uses the pre-trained migration method, take the parameters of the current multi-layer ResNet as the initial parameters of the present invention, and then use the image data processed in S2 to train the network. The specific process is:

1) Determine the number of nodes in the first layer of the network according to the dimension of the input data, that is, the number of nodes in the input layer;

2) Input data to the residual network unit. According to the characteristics of the ResNet network, that is, the identity mapping function of the residual network, the output of each module is the current input plus the residual, and the training data is used to train the network layer by layer;

3) Use the trained ResNet network for network transfer learning, and use its well-trained parameters as the initial training parameters of this model, which saves part of the training time and training samples, which is very suitable for training and learning with small samples.

Preferably, the specific steps of optimizing the model in the S4 are:

1) When the training is completed, the model is supervised by adding label data to the top layer of ResNet, that is, using the back propagation algorithm (BP) to fine-tune the relevant parameters of the network;

2) Input the classified labeled data into the top layer of ResNet, fine-tune the weight and threshold of ResNet through BP algorithm, and further reduce the training error and improve the accuracy of the transfer learning recognition model through supervised training.

Preferably, in the step S5, the test set data is input into the trained classification model, and after multi-layer ResNet mapping, the number of output layer nodes is the number of recognition states, and the input vector successfully activates the corresponding category nodes in the output layer.

Preferably, in the category nodes in S5, monocytes are node 0, lymphocytes are node 1, and neutrophils are node 2.

3. Beneficial effects

Compared with the prior art, the advantages of the present invention are:

(1) In the present invention, the problem of difficulty in feature extraction due to the existence of background impurities in the collected images can be effectively solved, and the recognition of cerebrospinal fluid cells under complex backgrounds can be well adapted; since the trained model parameters are The initial training parameters of the present invention reduce part of the training time to a certain extent, and the trained model parameters are generally more reliable than randomly selected parameters, so they are suitable for small sample learning and training; For the sample, the present invention also utilizes the near-circularity of its cells, and predicts its center point to segment it, so that the application of the present invention is diversified.

(2) In the present invention, the automatic identification technology of cerebrospinal fluid cells based on deep learning can help neurologists quickly establish a more scientific differential diagnosis model, can reduce the adverse effects of doctors on the results of microscopic examination due to subjective factors, and help doctors in cell Counting and classification can greatly improve the diagnosis rate; and can be integrated with the resources of high-level medical institutions, so that the overall diagnosis model tends to be standardized and unified, greatly improving the radiation effect of high-quality medical resources to primary medical institutions, and improving the differential diagnosis of primary hospitals. Therefore, the construction of an automatic identification system of cerebrospinal fluid cells based on deep learning is of great significance for improving the diagnosis rate of infectious diseases in the central nervous system, solving the problems of regional medical differences, low seniority and misdiagnosis by primary physicians, so as to eventually make the majority of patients benefit.

Description of drawings

Fig. 1 is a technical flowchart of a method for identifying and classifying complex small sample cerebrospinal fluid cells proposed by the present invention;

FIG. 2 is a schematic diagram of a concave point of a complex small sample cerebrospinal fluid cell identification and classification method proposed by the present invention;

Fig. 3 is the ResNet model of transfer learning in the present invention;

FIG. 4 is an example of transfer learning in the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments.

Example 1:

Referring to Figure 1, a complex small sample cerebrospinal fluid cell identification and classification method includes the following steps:

S1: use the microscope automatic scanning platform to acquire the image of the sample slide to obtain a complete image set of the cerebrospinal fluid cell slide with multiple cells;

Place the cerebrospinal fluid cell slide on the electric translation stage of the microscope, use the software system to locate the diagonal coordinate points of the scanning range of the slide, determine the scanning range of the image, and note the size of the scanned image range, use the software system platform Perform the stitching of the collected images to obtain a complete picture of the cell sample, and repeat this step for the subsequent slide image collection;

S2: Preprocess the obtained image set, filter and denoise the image, remove irrelevant factors in the picture, and separate the cells that adhere to each other, and form the obtained sample image set into training set and test set in batches;

The specific steps of preprocessing the image set are as follows:

Step 1: For the irrelevant impurities in the background of the sample, firstly separate the background of the image, obtain the binary image by the maximum inter-class variance method, and use the morphological opening operation to smooth the contour of the target in the binary image. In the open operation of learning, the impurities in the background that are not the target can be removed, and finally the Canny boundary detection algorithm is used to obtain the contour edge information of the target;

Step 2: Use the concave point detection to segment the adhesion cells. The concave point of the adhesion cell refers to the local maximum curvature point in the concave area formed by the adhesion of two or more quasi-circular objects due to overlapping with each other. For near-circular images, there is no sudden change in curvature unless there are two or more cells;

Step 3: Ellipse fitting. In order to obtain the outline boundary of the sticking target lost due to sticking, the algorithm uses the prior knowledge that the target generally appears to be a circle, and uses the ellipse fitting method based on the least squares method for fitting to complete Adhesion segmentation;

S3: Carry out model transfer training for small sample sets, and use deep learning networks trained in similar fields to perform transfer learning on small sample data sets;

Use the multi-layer ResNet model that has been trained in other fields, intercept the part in front of the fully connected layer of the model, set three output nodes in the output part according to the required classification type, and then use the pre-training migration method to transfer the current multi-layer ResNet The parameters of the present invention are used as the initial parameters of the present invention, and the image data processed in S2 is used to train the network. The specific process is:

1) Determine the number of nodes in the first layer of the network according to the dimension of the input data, that is, the number of nodes in the input layer;

2) Input data to the residual network unit. According to the characteristics of the ResNet network, that is, the identity mapping function of the residual network, the output of each module is the current input plus the residual, and the training data is used to train the network layer by layer;

3) Use the trained ResNet network for network transfer learning, and use its well-trained parameters as the initial training parameters of the model, which saves part of the training time and training samples, which is very suitable for training and learning with small samples;

S4: Use the BP algorithm to fine-tune the weights and thresholds of the trained model to further optimize the model;

The specific steps to optimize the model are:

1) When the training is completed, the model is supervised by adding label data to the top layer of ResNet, that is, using the back propagation algorithm (BP) to fine-tune the relevant parameters of the network;

2) Input the classified labeled data into the top layer of ResNet, fine-tune the weight and threshold of ResNet through BP algorithm, and further reduce the training error and improve the accuracy of the transfer learning recognition model through supervised training.

S5: Input the test set into the model, and the output result is the cerebrospinal fluid cell identification result.

Input the test set data into the trained classification model, after multi-layer ResNet mapping, the number of output layer nodes is the number of recognition states, the input vector successfully activates the corresponding category node in the output layer, and the monocyte in the category node is node 0 , lymphocytes are node 1, and neutrophils are node 2.

In the present invention, the problem of difficulty in feature extraction caused by the existence of background impurities in the collected images can be effectively solved, and the recognition of cerebrospinal fluid cells under complex backgrounds can be well adapted; since the trained model parameters are used in the present invention The initial training parameters reduce part of the training time to a certain extent, and the trained model parameters are generally more reliable than randomly selected parameters, so they are suitable for small sample learning and training; for samples with cell adhesion, this The invention also makes use of the near-circularity of its cells to predict the center point for segmentation, so that the application of the invention is diversified.

In the present invention, the automatic identification technology of cerebrospinal fluid cells based on deep learning can help neurologists to quickly establish a more scientific differential diagnosis model, can reduce the adverse effects of doctors on microscopic examination results caused by subjective factors, and help doctors to count and classify cells , greatly improving the diagnosis rate; and can be integrated with the resources of high-level medical institutions, so that the overall diagnosis model tends to be standardized and unified, greatly improving the radiation effect of high-quality medical resources to primary medical institutions, and improving the level of differential diagnosis in primary hospitals. Therefore, The construction of an automatic identification system for cerebrospinal fluid cells based on deep learning is of great significance for improving the diagnosis rate of infectious diseases in the central nervous system, solving the problems of regional medical differences, low seniority and misdiagnosis by primary physicians, thus ultimately benefiting the majority of patients.

Example 2:

Referring to Figures 1-4, a complex small sample cerebrospinal fluid cell identification and classification method includes the following steps:

S1: use the microscope automatic scanning platform to acquire the image of the sample slide to obtain a complete image set of the cerebrospinal fluid cell slide with multiple cells;

Place the cerebrospinal fluid cell slide on the electric translation stage of the microscope, use the software system to locate the diagonal coordinate points of the scanning range of the slide, determine the scanning range of the image, and note the size of the scanned image range, use the software system platform Perform the stitching of the collected images to obtain a complete picture of the cell sample, and repeat this step for the subsequent slide image collection;

S2: Preprocess the obtained image set, filter and denoise the image, remove irrelevant factors in the picture, and separate the cells that adhere to each other, and form the obtained sample image set into training set and test set in batches;

The specific steps of preprocessing the image set are as follows:

Step 1: For the irrelevant impurities in the background of the sample, firstly separate the background of the image, obtain the binary image by the maximum inter-class variance method, and use the morphological opening operation to smooth the contour of the target in the binary image. In the open operation of learning, the impurities in the background that are not the target can be removed, and finally the Canny boundary detection algorithm is used to obtain the contour edge information of the target;

Step 2: Use the concave point detection to segment the adhesion cells. The concave point of the adhesion cell refers to the local maximum curvature point in the concave area formed by the adhesion of two or more quasi-circular objects due to overlapping with each other. For near-circular images, there is no sudden change in curvature unless there are two or more cells;

1) pit detection:

First, an improved curvature scale space algorithm (Curvature Scale Space, CSS) is used to detect the corner points of the target contour. This improved CSS algorithm preserves all ground-truth corners at a relatively low scale, and then compares the curvature of all candidate corners with an adaptive local threshold to remove redundant corners. Usually, the adaptive local threshold of a candidate corner is determined according to the curvature of its neighborhood area, and candidate corners whose absolute curvature is lower than their local threshold will be eliminated. Among the candidates for corner points, although some points are detected as local maxima in the curvature value, they are in the region of support (ROS), and the difference between adjacent points is very small. When supporting regions, also select the appropriate region.

The setting method of the adaptive local threshold is:

是邻域区域的曲率均值，p代表候选角点的位置，R₁与R₂为支持区域的尺寸大小，C为系数；in,

is the mean curvature of the neighborhood area, p represents the position of the candidate corner point, R ₁ and R ₂ are the size of the support area, and C is the coefficient;

2) Contour segment grouping:

Use the concave points obtained in 1) to segment the contour of the adhesion region into multiple contour segments. Since each contour segment does not correspond to a separate target, there may be situations where multiple contour segments belong to the same target. Therefore, it is necessary to group the contour segments belonging to the same target into a group. For a contour segment, for another contour segment s _j within a certain neighborhood range, if s _i and s _j satisfy the condition of being divided into one group, they are divided into the same group. This grouping method includes the following Three conditional constraints:

Condition 1: If the average distance deviation (Average DistanceDeviation, ADD) of the fitted ellipses after being divided into a group is smaller than the average distance deviation produced by the ellipses fitted individually by any contour segment before the combination, then these contour segments are divided into two groups. for the same group.

Condition 2: If the distance between the center of gravity of the fitted ellipse after being divided into the same group and the center of gravity of the ellipse fitted separately by each contour segment is relatively close, it can be divided into one group.

Condition 3: If the centers of gravity between the ellipses fitted by any two contour segments s _i and s _j respectively are very close, they can be grouped into one group;

Step 3: Ellipse fitting. In order to obtain the outline boundary of the sticking target lost due to sticking, the algorithm uses the prior knowledge that the target generally appears to be a circle, and uses the ellipse fitting method based on the least squares method for fitting to complete Adhesion segmentation;

S3: Carry out model transfer training for small sample sets, and use deep learning networks trained in similar fields to perform transfer learning on small sample data sets;

The multi-layer ResNet model that has been trained in other fields is used, and the front part of the fully connected layer of the model is intercepted. The output part sets three output nodes according to the required classification type, and then uses the pre-training migration method, that is, the current multi-layer The parameters of the layer ResNet model are used as the initial parameters of the present invention, and the image data processed in step 2 of claim 3 is used to train the network. An example of transfer learning is shown in Figure 4. The specific implementation is as follows:

ResNet is composed of multiple convolution modules connected in series. Each convolution module includes one layer of convolution and one layer of pooling. Figure 3 shows a ResNet module. During training, the unit target mapping (that is, the optimal solution to be approached) is assumed to be F(x)+x, and the output is: y+x, then the training goal becomes to make y approach F (x). That is, the same body part x before and after the mapping is removed, so as to highlight the small changes (residuals).

Mathematically expressed as:

y=F(x,{W _i })+W _s x (2)

x is the input of the residual unit, y is the output of the residual unit, F(x) is the target map, and {W _i } is the convolutional layer in the residual unit. W _s is a convolution with a 1×1 convolution kernel size, which is used to reduce or increase the dimension of x so that it is the same size as the output y (because it needs to be summed).

The specific process is:

1) Determine the number of nodes in the first layer of the network according to the dimension of the input data, that is, the number of nodes in the input layer;

2) Input data to the residual network unit. According to the characteristics of the ResNet network, that is, the identity mapping function of the residual network, the output of each module is the current input plus the residual, and the training data is used to train the network layer by layer. .

3) The collection of training data in the present invention does not reach the hundreds of thousands of training set samples required by deep learning, but due to the use of the trained ResNet network for network migration, the well-trained parameters are used as this model. The initial parameters of training, which saves part of the training time and training samples, are very suitable for training and learning with small samples;

S4: Use the BP algorithm to fine-tune the weights and thresholds of the trained model to further optimize the model;

The specific implementation measures are as follows:

(1) Model pre-training uses the transferred weights as the initial weights of the new network, which will be changed by the gradient descent algorithm during the training process.

Gradient descent algorithm:

1) Starting from 0 and ending with the number of training set data:

① Calculate the gradient of the weight w and bias b of the i-th training data relative to the loss function. So we end up with the gradient values of the weights and biases for each training data.

② Calculate the sum of the gradients of all training data weights w.

③ Calculate the sum of the gradients of all training data deviations b.

2) After completing the above calculations, we start to perform the following calculations:

① Using the results obtained in steps ② and ③ above, calculate the average value of the gradients of the weights and biases of all samples.

②Using the following formula, update the weight value and bias value of each sample.

Repeat the above process until the loss function converges.

(2) Reverse fine-tuning is to perform supervised training on the ResNet network to reduce training errors and improve the accuracy of the classification model. BP algorithm steps:

1) Input the training set;

2) For each sample x in the training set, set the activation value a ¹ corresponding to the input layer:

Forward propagation:

3) Since there is an error between the output result and the actual result, calculate the error generated by the output layer:

δ ^L =Δ _a Ceσ'(z ^L ) (6)

4) Backpropagate the error obtained in the previous step from the output layer to the hidden layer:

δ ^l =((w ^l+1 ) ^T δ ^l+1 )eδ'(z ^l ) (7)

5) Use gradient descent, train parameters, and iterate until convergence:

S5: Input the test set into the model, and the output result is the cerebrospinal fluid cell identification result.

Input the test set data into the trained classification model, after multi-layer ResNet mapping, the number of output layer nodes is the number of recognition states, the input vector successfully activates the corresponding category node in the output layer, and the monocyte in the category node is node 0 , lymphocytes are node 1, and neutrophils are node 2.

In the present invention, the problem of difficulty in feature extraction caused by the existence of background impurities in the collected images can be effectively solved, and the recognition of cerebrospinal fluid cells under complex backgrounds can be well adapted; since the trained model parameters are used in the present invention The initial training parameters reduce part of the training time to a certain extent, and the trained model parameters are generally more reliable than randomly selected parameters, so they are suitable for small sample learning and training; for samples with cell adhesion, this The invention also makes use of the near-circularity of its cells to predict the center point for segmentation, so that the application of the invention is diversified.

In the present invention, the automatic identification technology of cerebrospinal fluid cells based on deep learning can help neurologists to quickly establish a more scientific differential diagnosis model, can reduce the adverse effects of doctors on microscopic examination results caused by subjective factors, and help doctors to count and classify cells , greatly improve the diagnosis rate; and can be integrated with the resources of high-level medical institutions, so that the overall diagnosis model tends to be standardized and unified, greatly improving the radiation effect of high-quality medical resources to primary medical institutions, and improving the level of differential diagnosis in primary hospitals. Therefore, The construction of an automatic identification system of cerebrospinal fluid cells based on deep learning is of great significance for improving the diagnosis rate of infectious diseases in the central nervous system, solving the problems of regional medical differences, low seniority and misdiagnosis by primary doctors, thus ultimately benefiting the majority of patients.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. The equivalent replacement or change of the inventive concept thereof shall be included within the protection scope of the present invention.

Claims (7)

1. a complex small sample cerebrospinal fluid cell identification and classification method, is characterized in that, comprises the following steps: S1: use the microscope automatic scanning platform to acquire the image of the sample slide to obtain a complete image set of the cerebrospinal fluid cell slide with multiple cells; S2: Preprocess the obtained image set, filter and denoise the image, remove irrelevant factors in the picture, and separate the cells that adhere to each other, and form the obtained sample image set into training set and test set in batches; S3: Carry out model transfer training for small sample sets, and use deep learning networks trained in similar fields to perform transfer learning on small sample data sets; S4: Use the BP algorithm to fine-tune the weights and thresholds of the trained model to further optimize the model; S5: Input the test set into the model, and the output result is the cerebrospinal fluid cell identification result. 2. The method for identifying and classifying cerebrospinal fluid cells in a complex small sample according to claim 1, wherein in the S1, the cerebrospinal fluid cell glass slide is placed on an electric translation stage of a microscope, and a software system is used to quantify the cerebrospinal fluid cell slide. The scanning range is determined by the diagonal coordinate point positioning, the scanning range of the image is determined, and the size of the scanned image range is recorded. The software system platform is used to stitch the collected images to obtain a complete picture of the cell sample. Repeat this step for image acquisition. 3. a kind of complex small sample cerebrospinal fluid cell identification and classification method according to claim 1, is characterized in that, in described S2, the concrete step of preprocessing the image set is: Step 1: For the irrelevant impurities in the background of the sample, firstly separate the background of the image, obtain the binary image by the maximum inter-class variance method, and use the morphological opening operation to smooth the contour of the target in the binary image. In the open operation of learning, the impurities in the background that are not the target can be removed, and finally the Canny boundary detection algorithm is used to obtain the contour edge information of the target; Step 2: Use the concave point detection to segment the adhesion cells. The concave point of the adhesion cell refers to the local maximum curvature point in the concave area formed by the adhesion of two or more quasi-circular objects due to overlapping with each other. For near-circular images, there is no sudden change in curvature unless there are two or more cells; Step 3: Ellipse fitting. In order to obtain the outline boundary of the sticking target lost due to sticking, the algorithm uses the prior knowledge that the target generally appears to be a circle, and uses the ellipse fitting method based on the least squares method for fitting to complete Adhesive segmentation. 4. a kind of complexity small sample cerebrospinal fluid cell identification and classification method according to claim 1, is characterized in that, adopts the multi-layer ResNet model trained in other fields in the described S3, intercepts the front of the fully connected layer of the model The output part sets three output nodes according to the required classification types, and then uses the pre-training migration method to take the parameters of the current multi-layer ResNet as the initial parameters of the present invention, and then use the image data processed in S2 to carry out The training process of the network is as follows: 1) Determine the number of nodes in the first layer of the network according to the dimension of the input data, that is, the number of nodes in the input layer; 2) Input data to the residual network unit. According to the characteristics of the ResNet network, that is, the identity mapping function of the residual network, the output of each module is the current input plus the residual, and the training data is used to train the network layer by layer; 3) Use the trained ResNet network for network transfer learning, and use its well-trained parameters as the initial training parameters of this model, which saves part of the training time and training samples, which is very suitable for training and learning with small samples. 5. a kind of complex small sample cerebrospinal fluid cell identification and classification method according to claim 1, is characterized in that, the concrete steps of optimizing model in described S4 are: 1) When the training is completed, the model is supervised by adding label data to the top layer of ResNet, that is, using the back propagation algorithm (BP) to fine-tune the relevant parameters of the network; 2) Input the classified labeled data into the top layer of ResNet, fine-tune the weight and threshold of ResNet through BP algorithm, and further reduce the training error and improve the accuracy of the transfer learning recognition model through supervised training. 6. a kind of complexity small sample cerebrospinal fluid cell identification and classification method according to claim 1, is characterized in that, in described S5, test set data is input in the classification model that is trained, after multi-layer ResNet mapping, The number of output layer nodes is the number of recognition states, and the input vector successfully activates the corresponding category nodes in the output layer. 7 . The method for identifying and classifying cerebrospinal fluid cells in a complex small sample according to claim 6 , wherein the monocytes in the class nodes in S5 are node 0, lymphocytes are node 1, and neutrophils are node 1. 8 . for node 2.

Images