Simulation model construction method, identification method and system for identifying lung cancer metastatic cells

Abstract

The invention discloses a construction method, a recognition method and a system for simulating lung cancer metastasis cells in cerebrospinal fluid based on diffraction imaging and a deep learning technology. The construction method comprises the following steps: culturing a plurality of lung cancer cell lines which are easy to cause brain metastasis, collecting whole blood of a healthy person and splitting red blood cells; the method comprises the following steps: respectively acquiring diffraction images of cultured cells through a polarization diffraction imaging flow cytometry with a TDI camera to obtain a pair of diffraction image pairs with S polarization and P polarization; performing image processing on the obtained diffraction cell image pair, and establishing a training image data set and a test image data set; and establishing a convolutional neural network, and training the convolutional neural network according to the training image data set to obtain a model for identifying the lung cancer metastatic cells in the cerebrospinal fluid. When in use, a cerebrospinal fluid sample of a patient is collected, S polarization and P polarization image pairs are obtained, and the trained model is input to respectively extract S polarization and P polarization image features of cells for feature recognition, so that whether the current sample contains lung cancer cells or not is predicted.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and cell recognition, and in particular to a simulation model construction method, recognition method and system for identifying lung cancer metastasis cells in cerebrospinal fluid based on diffraction imaging and deep learning technology. Background Art

[0002] Leptomeningeal metastasis (LM) is common in lung cancer, breast cancer, and malignant melanoma. The prognosis of this disease is extremely poor, with a median survival of 4 to 6 weeks without intervention. Most LMs cannot be pathologically obtained through surgery or tissue biopsy. Therefore, establishing appropriate methods, early diagnosis, and selecting specific chemotherapy drugs for primary cancer will be of great significance in prolonging patient survival and improving quality of life.

[0003] Cerebrospinal fluid cytology (CSF cytology) is the gold standard for diagnosing leptomeningeal metastasis of tumors. This method uses cerebrospinal fluid extracted by biopsy as the analysis object, obtains cells after centrifugation to make cell smears, and then stains them to find tumor cells under a microscope. This method has extremely high specificity, but low sensitivity. In order to promote the early diagnosis of LM, some cell detection technologies have been extended to this field, such as CellSearch and immune flow cytometry technology with fluorescent staining or specific antibody labeling. Other technologies, such as PCR and RT-PCR, examine free gene fragments in cerebrospinal fluid. The above technologies have high technical requirements for operators, high time and economic costs, and some medical institutions cannot complete the detection independently. Moreover, as the gold standard for control, CSF cytology has a low sensitivity, which limits the development of detection technology in this field. Therefore, a high-throughput cell recognition and enrichment technology that does not require staining or antibody labeling is needed to be combined with CSF cytplogy to improve its detection rate and benefit patients with LM. Summary of the invention

[0004] The purpose of the present invention is to provide a simulation model construction method, identification method and system for identifying lung cancer metastasis cells in cerebrospinal fluid based on diffraction imaging and deep learning technology. The construction method and identification method establish a model simulating the formed components of cerebrospinal fluid of patients with brain metastasis of lung cancer by collecting whole blood of healthy people and lysing red blood cells to obtain other blood formed components including white blood cells and cultured lung cancer cells, and realize the identification of lung cancer cells and other blood formed components except red blood cells (including blood cells and platelets, and possibly desquamated epithelial cells) by collecting diffraction images of cells and other formed components and combining deep learning technology.

[0005] The first aspect of the present invention is to disclose a method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid based on diffraction imaging and deep learning technology. The construction method is based on the principle that conventional cerebrospinal fluid contains a small amount of white blood cells and desquamated epithelial cells, which are basically consistent with other cell components in blood except red blood cells. The method uses other blood formed components except red blood cells to simulate formed components in cerebrospinal fluid, including:

[0006] Cultivate a variety of lung cancer cell lines that are prone to brain metastasis and collect whole blood from healthy people and lyse red blood cells to simulate the formed elements in the cerebrospinal fluid of patients with lung cancer brain metastasis;

[0007] The diffraction images of the cells are collected by a polarization diffraction imaging flow cytometer with a time delay integration (TDI) camera to obtain a pair of diffraction images in the S polarization direction and the P polarization direction; thereby obtaining a pair of diffraction images of the cultured lung cancer cells and a pair of diffraction images of other blood formed elements in the collected blood except red blood cells;

[0008] After preprocessing of screening, grayscale normalization and scaling to a preset size, the obtained diffraction image pairs are divided into a training image data set and an independent test image data set by random sampling; a convolutional neural network is established, and the convolutional neural network is trained according to the training image data set to obtain a model for identifying lung cancer metastatic cells in cerebrospinal fluid.

[0009] Furthermore, the obtained diffraction image pair is preprocessed, specifically comprising the following steps:

[0010] Set the ratio of the number of too dark pixels and the threshold of too dark pixels in the image, and set the ratio of too bright pixels;

[0011] When the average pixel value of each image in a pair of cell diffraction images is less than the set dark pixel threshold, the image is considered underexposed and the current cell image pair is deleted;

[0012] If the saturated pixels in any image of a pair of images exceed the set over-bright pixel ratio, the image is considered overexposed and the current cell image pair is deleted. Further, the cultured lung cancer cell lines include: human lung adenocarcinoma cell lines A549 and H1975, and human small cell lung cancer cell line H196.

[0013] Furthermore, the constructed convolutional neural network includes: an input layer, an intermediate layer and an output layer, the input layer is a preprocessed three-channel cell diffraction image, the intermediate layer has a total of 5 modules, each of the first four modules includes two convolutional layers and a maximum pooling layer, and the last module includes two convolutional layers, the convolutional layer uses a convolution kernel of size 3×3, the pooling layer uses a convolution kernel of size 1×1, and the convolution step is 1; the output layer includes a fully connected layer and an output, the activation function of the fully connected layer is the normalization function sigmoid that maps the output to between [0,1], and uses 0.5 as the threshold to divide it into blood cells or lung cancer cells.

[0014] Furthermore, the training of the convolutional neural network includes:

[0015] Expand the input feature data to one dimension;

[0016] Binary cross entropy is used as the cost function for training, and the optimizer chooses stochastic gradient descent with a learning rate of 0.01;

[0017] Through random inactivation dropout, randomly reduce 50% of parameters;

[0018] When the cost function no longer decreases, the model is evaluated based on the classification accuracy and the model is modified. The batch_size of the model training is 24.

[0019] Based on the above settings, the convolutional neural network framework was trained through cross-validation to obtain simulation models of lung cancer cells and other blood components except red blood cells;

[0020] The pre-trained simulation model is then tested using an additional independent test image dataset to verify the performance of the model;

[0021] If the accuracy reaches the set value, it means that the automatic recognition model training is completed; otherwise, the diffraction image is added or replaced in the training image data set and the test image data set, and re-training is performed.

[0022] The second aspect of the present invention is to disclose a method for automatically identifying lung cancer metastatic cells in cerebrospinal fluid. The identification method is based on the application of the model established in the first aspect of the invention in cerebrospinal fluid samples, and specifically includes:

[0023] A certain amount of cerebrospinal fluid samples were collected from patients with a high suspicion of lung cancer leptomeningeal metastasis in clinical evaluation, and the samples were centrifuged and resuspended to 300-500 μl. The S-polarized diffraction image and P-polarized diffraction image of cerebrospinal fluid cells were obtained by a polarized diffraction imaging flow cytometer with a TDI camera to form an image pair.

[0024] After preprocessing by screening, grayscale normalization and scaling to a preset size, the image pairs are input into a trained model for identifying lung cancer metastatic cells in cerebrospinal fluid to extract the image features of the cell S polarization and P polarization for feature recognition, thereby predicting whether the current sample contains lung cancer cells.

[0025] The third aspect of the present invention is to disclose an automatic recognition system for identifying lung cancer metastasis cells in cerebrospinal fluid, comprising a data preprocessing module, a data training module and an automatic recognition module; the data preprocessing module is used to screen, grayscale normalize and scale the collected image pairs containing S-polarized diffraction images and P-polarized diffraction images to a preset size; the data training module is used to train and test a convolutional neural network according to a pre-established training image data set and an independent test image data set to obtain a model for identifying lung cancer metastasis cells in cerebrospinal fluid; the data training module pre-collects 2000-4000 lung cancer cell lines that are prone to brain metastasis and collects diffraction images of formed elements in the blood of healthy people other than red blood cells, and performs recognition training based on the convolutional neural network pattern, thereby obtaining a trained model for identifying lung cancer metastasis cells in cerebrospinal fluid; the automatic recognition module is used to input the image pairs after the data preprocessing module into the trained simulation model to extract the image features of S polarization and P polarization respectively and perform feature splicing processing and then perform feature recognition, thereby identifying the type of the current cell. The training data set accounts for 80% of the total data, and the test data set accounts for 20% of the total data. The training method uses pattern recognition with feature extraction, and labels are assigned to cultured lung cancer cells and other formed elements in the blood except red blood cells. "1" represents lung cancer cells and "0" represents other formed elements in the blood except red blood cells. The neural network is used to perform image recognition, thereby obtaining a simulated model for identifying lung cancer metastasis cells in cerebrospinal fluid after training. A large number of sample trainings are required for the image source in the learning stage to ensure the selection of weights.

[0026] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0027] The present invention proposes a model that uses whole blood cells and cultured lung cancer cells to simulate the cerebrospinal fluid cell environment of lung cancer brain metastasis. Whole blood contains white blood cells in various cell cycles and states, including B lymphocytes, T lymphocytes, neutrophils, eosinophils, basophils, monocytes, platelets and other cells and formed elements. Compared with the use of cultured immortalized white blood cells, the present invention increases the complexity of the cell model and improves the analysis and prediction capabilities of real cerebrospinal fluid samples and blood samples.

[0028] Another advantage of the present invention is that the diffraction images of blood cells and lung cancer cells are collected by non-staining labeled polarized diffraction imaging flow cytometry detection technology, and the images are recognized and classified by constructing and training a small convolutional neural network and a densely connected classifier, so as to achieve the purpose of predicting blood cells and lung cancer cells. Compared with the cell detection technology using fluorescent dye staining and antibody labeling, the experimental operation process is simplified, and the time cost of operation and the economic cost of purchasing reagents are saved. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of the method for constructing a simulation model and the method for identifying lung cancer metastasis cells in cerebrospinal fluid according to the present invention;

[0030] Figure 2 is a schematic diagram of the convolutional neural network involved;

[0031] Figure 3 This is an example of a diffraction image of culturing and collecting cells during the culturing and collecting cell phase;

[0032] Figure 4 This is an example of removing underexposed and overexposed images during the image processing stage in step three. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The described specific embodiments are only used to explain the present invention and are not intended to limit the present invention.

[0034] like Figure 1 As shown, a method for constructing lung cancer metastasis cells in simulated cerebrospinal fluid based on diffraction imaging and deep learning technology includes:

[0035] Step 1: Cultivation and cell collection

[0036] Step 1.1: Cultivate a variety of lung cancer cell lines that are prone to brain metastasis, including lung adenocarcinoma A549, H1975, and small cell lung cancer H196. Use F1640 medium plus 10% fetal bovine serum and subculture in a 37° constant temperature incubator with 5% CO2; digest the cells with 0.25% trypsin before use, centrifuge at 1000RPM for 5 minutes, collect the cells, resuspend the cells in phosphate buffered saline (PBS), and adjust the cell concentration to 1×10 5 With 1×10 7 / ml;

[0037] Step 1.2: Collect whole blood from healthy individuals, lyse the whole blood with red blood cell lysis buffer to remove red blood cells, centrifuge at 450 g for 10 min at 4°C, collect the precipitate, and resuspend the cells in PBS to 1×10 5 With 1×107 / ml.

[0038] Step 2: Image acquisition stage:

[0039] The diffraction images of the cultured cells in the above steps 1.1 and 1.2 are respectively collected using polarization diffraction imaging flow cytometry detection technology. The polarization diffraction imaging flow cytometry detection technology is an existing device independently developed by the laboratory, which is a polarization diffraction imaging flow cytometer (p-DIFC) using a TDI camera. The polarization diffraction imaging flow cytometer detection technology uses TDI cameras in two directions to obtain a pair of orthogonal polarization diffraction images, such as Figure 3 As shown. The image acquisition conditions are as follows: the rated power of the coherent light source is 100 mW, the wavelength of the coherent light is 532 nm, and the incident angle with the fluid chamber is 90 degrees; the side scattered light is divided into S polarized light and P polarized light by a 45-degree polarizer, and a cell diffraction image (S polarized diffraction image and P polarized diffraction image) is acquired by two TDI (Time Delay Integration) cameras in each direction, with an image size of 508 × 512 pixels, a bit depth of 12 bits, and saved as 16 bits. A pair of cell diffraction images represents a cell, thereby obtaining the diffraction image pair of the cancer cells cultured in step 1.1 and the diffraction image pair of other formed elements of the blood except red blood cells acquired in step 1.2.

[0040] Step 3: Image preprocessing

[0041] Step 3.1: Screen the cell diffraction image pairs obtained in step 2 to remove underexposed and overexposed images. Examples of underexposed and overexposed images are as follows: Figure 4 As shown;

[0042] In the algorithm, the ratio of the number of dark pixels in the image is set to under_rate 0.02, and the threshold of dark pixels is set to (2 12 -1)×under_rate;

[0043] When the average pixel value of each image in a pair of cell diffraction images is less than the above-set threshold, the image is considered underexposed and the current cell image pair is deleted;

[0044] Set the ratio of over-bright pixels in the image to over_rate to 0.08. If the saturated pixels in any image of a pair of images exceed this ratio, the image is considered to be overexposed; delete the current cell image pair.

[0045] Step 4: Convolutional Neural Network Training

[0046] Step 4.1: The image pairs obtained in step 3 are divided into a training image data set and an independent test image data set by random sampling; wherein the training data set accounts for 80% of the total data, and the test data set accounts for 20% of the total data. The data distribution in the embodiment is shown in Table 1. The lung cancer cell class contains images of three types of lung cancer cells, A549, H196 and H1975.

[0047] Table 1: Example data distribution

[0048]

[0049] Each dataset is divided into blood cell class (blood dataset, label is 0) and lung cancer cell class (tumor dataset, label is 1).

[0050] Step 4.2: The cell image pair in step 4.1 is grayscale normalized and scaled to a preset size. The image is normalized to 8 bits according to [0, 255] and the image size is scaled to (224, 224, 3) to form a 224×224 three-channel image.

[0051] A small neural network was established to extract the image features of cell S polarization and P polarization and perform identification and classification. Figure 2 As shown, the convolutional neural network includes an input layer, an intermediate layer and an output layer; the input layer is a three-channel cell diffraction image, and the intermediate layer has a total of 10 convolutional layers, which are divided into 5 modules (blocks). In the first 4 modules, each module includes two convolutional layers (CONV1 and CONV 2) and a maximum pooling layer (MaxPooling). The last module only contains two convolutional layers (CONV1 and CONV 2) and no pooling layer. Since the input image is three-channel, the convolution kernel is also three-channel, and the convolution stride is set to 1, the activation function is ReLU, and the kernel size is a compact convolution operation of 3×3. The pooling layer uses a convolution kernel of size 1×1. After that is the output layer, which includes a fully connected layer and an output. In the output layer, the incoming feature data is expanded to one dimension, followed by a densely connected layer with an output of 256 units to transform the data, and a random inactivation (dropout) function is added after the convolution layer to suppress overfitting. The activation function of the fully connected layer is the normalization function sigmoid, which maps the output to [0,1], and the output is the cell category, tumor cell or blood cell. The training of the convolutional neural network includes: using binary cross entropy as the cost function (loss function) of training, selecting stochastic gradient descent (sgd) as the optimizer, and the learning rate is 0.01;

[0052] When the cost function no longer decreases, the model is evaluated based on the classification accuracy and the model is modified. The batch_size of the model training is 24.

[0053] This completes the training of the convolutional neural network framework and obtains a simulation model for identifying lung cancer metastasis cells in cerebrospinal fluid.

[0054] The established simulation model is then tested against the pre-trained network model using an additional independent test image data set to verify the performance of the model and obtain the classification accuracy. The classification accuracy and model evaluation of this embodiment are shown in Table 2.

[0055] Table 2: Classification accuracy and model evaluation

[0056]

[0057] If the accuracy reaches the above result or reaches the set value, it means that the simulation model training of the convolutional neural network is completed; otherwise, return to step 4.1 to add the cell diffraction image and retrain.

[0058] The automated identification method using the trained model for identifying lung cancer metastatic cells in cerebrospinal fluid includes:

[0059] Cerebrospinal fluid samples were collected from patients with clinical high suspicion of leptomeningeal metastasis of lung cancer and centrifuged at 450 g for 10 min at 4°C.

[0060] The S-polarized diffraction image and P-polarized diffraction image of cerebrospinal fluid cells were collected by a polarized diffraction imaging flow cytometer with a TDI camera to form a diffraction image pair; the image acquisition conditions were: the coherent light source had a rated power of 100 mW, a coherent light wavelength of 532 nm, and an incident angle of 90 degrees with the fluid chamber; the side scattered light was divided into S-polarized light and P-polarized light by a 45-degree polarizer, and a pair of diffraction images were acquired by TDI (Time Delay Integration) cameras in two directions, with an image size of 508×512 pixels, a bit depth of 12 bits, and saved as 16 bits. A pair of image pairs represented a cell, thereby obtaining a cell image pair of the collected cerebrospinal fluid cells.

[0061] According to step 3, the grayscale of the image is normalized to 8 bits according to [0,255], and the image size is adjusted to (224,224,3) to form a 224×224 three-channel image, which is input into the trained model for identifying lung cancer metastasis cells in cerebrospinal fluid;

[0062] A 224×224 three-channel image is used as input. Since the input image is three-channel, the convolution kernel is also three-channel. The convolution stride is set to 1, the activation function is ReLU, and the kernel size is 3×3 for compact convolution operations. The pooling layer uses a 1×1 convolution kernel. After the input image is calculated by the convolution layer, the output is a high-dimensional array representing the image features. The features of a cell in the S-polarization and P-polarization diffraction images are extracted respectively, and the two high-order arrays are spliced ​​in one dimension. The data dimension after splicing is (1, 6, 6, 1024).

[0063] The extracted features are input into the trained model; finally, the output of each pair of images is mapped to [0,1] through a fully connected layer with a sigmoid activation function, and divided into blood cells and lung cancer cells with a threshold of 0.5; thus, the type of current cells is identified. The prediction results of a cerebrospinal fluid sample of a patient with lung cancer leptomeningeal metastasis with a positive pathological report of heterogeneous cells and a blood sample of a healthy volunteer are shown in the following table:

[0064] Table 3. Predictions of the simulation model on samples of patients and healthy volunteers

[0065] Patient cerebrospinal fluid samples Blood samples from healthy volunteers Diffraction image logarithm 530 445 Predicting the positive rate of lung cancer cells 0.18％ 0.0％

[0066] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which fall within the scope of protection of the present invention.

Claims

1. A method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid, characterized in that: The simulation model utilizes other blood formed elements in addition to red blood cells to simulate formed elements in cerebrospinal fluid, including: Cultivate a variety of lung cancer cell lines that are prone to brain metastasis and collect whole blood from healthy people and lyse red blood cells to simulate the formed elements in the cerebrospinal fluid of patients with lung cancer brain metastasis; The diffraction images of the cells are collected by a polarization diffraction imaging flow cytometer with a TDI camera to obtain a pair of diffraction images in the S polarization direction and the P polarization direction; thereby obtaining a pair of diffraction images of the cultured lung cancer cells and a pair of diffraction images of other blood formed elements in the collected blood except red blood cells; After preprocessing of screening, grayscale normalization and scaling to a preset size, the obtained diffraction image pairs are divided into a training image data set and an independent test image data set by random sampling; a convolutional neural network is established, and the convolutional neural network is trained according to the training image data set to obtain a model for identifying lung cancer metastatic cells in cerebrospinal fluid.

2. The method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid according to claim 1, characterized in that: The preprocessing of the obtained diffraction image pair specifically includes the following steps: Set the ratio of the number of too dark pixels and the threshold of too dark pixels in the image, and set the ratio of too bright pixels; When the average pixel value of each of the two images in a pair of cell diffraction images is less than the set dark pixel threshold, the image is considered underexposed and the current cell image pair is deleted; If the saturated pixels in any image of a pair of images exceed the set over-bright pixel ratio, the image is considered overexposed and the current cell image pair is deleted.

3. The method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid according to claim 1, characterized in that: The cultured lung cancer cell lines include: human lung adenocarcinoma cell lines A549 and H1975, and human small cell lung cancer cell line H196.

4. The method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid according to claim 1, characterized in that: The constructed convolutional neural network includes: an input layer, an intermediate layer and an output layer. The input layer is a preprocessed three-channel cell diffraction image. The intermediate layer has a total of 5 modules. Each of the first four modules includes two convolutional layers and a maximum pooling layer. The last module includes two convolutional layers. The convolutional layer uses a convolution kernel of size 3×3, and the pooling layer uses a convolution kernel of size 1×1, and the convolution step is 1; the output layer includes a fully connected layer and an output. The activation function of the fully connected layer is the normalization function sigmoid, which maps the output to between [0,1], and uses 0.5 as the threshold to divide it into blood cells or lung cancer cells.

5. The method for constructing a simulation model for identifying lung cancer metastatic cells in cerebrospinal fluid according to claim 4, characterized in that: The training of the convolutional neural network includes: Expand the input feature data to one dimension; Binary cross entropy is used as the cost function for training, and the optimizer chooses stochastic gradient descent with a learning rate of 0.01; Through random inactivation dropout, randomly reduce 50% of parameters; When the cost function no longer decreases, the model is evaluated based on the classification accuracy and the model is modified. The batch_size of the model training is 24. The convolutional neural network framework was trained through cross-validation to obtain models of lung cancer cells and other blood components except red blood cells; Then, the pre-trained model is tested with an additional independent test image dataset to verify the performance of the model; If the accuracy reaches the set value, it means that the automatic recognition model training is completed; otherwise, the diffraction image is added or replaced in the training image data set and the test image data set, and re-training is performed.

6. An automatic identification method for identifying lung cancer metastasis cells in cerebrospinal fluid, characterized in that: include: Cerebrospinal fluid samples were collected from patients with a high suspicion of leptomeningeal metastasis of lung cancer during clinical evaluation, and the samples were centrifuged and resuspended to 300-500 μl. The S-polarized diffraction image and P-polarized diffraction image of cerebrospinal fluid cells were obtained by a polarized diffraction imaging flow cytometer with a TDI camera to form an image pair. After preprocessing by screening, grayscale normalization and scaling to a preset size, the image pairs are input into the model obtained according to the method for constructing a simulation model for identifying lung cancer metastasis cells in cerebrospinal fluid according to claim 1 for feature recognition, thereby predicting whether the current sample contains lung cancer cells.

7. An automatic identification system for identifying lung cancer metastasis cells in cerebrospinal fluid, characterized in that: It includes data preprocessing module, data training module and automatic recognition module; The data preprocessing module is used for preprocessing the collected image pairs including the S-polarized diffraction image and the P-polarized diffraction image by screening, grayscale normalization and scaling to a preset size; The data training module is used to train and independently test the convolutional neural network according to the pre-established training image data set and the independent test image data set, so as to obtain a model for identifying lung cancer metastasis cells in cerebrospinal fluid; the data training module pre-collects 2000-4000 lung cancer cell lines that are prone to brain metastasis and collects diffraction images of formed elements in the blood of healthy people other than red blood cells, and performs recognition training based on the pattern of the convolutional neural network, so as to obtain a trained model for identifying lung cancer metastasis cells in cerebrospinal fluid; The automatic identification module is used to input the image pairs after the data preprocessing module into the trained simulation model to extract the image features of S polarization and P polarization respectively and perform feature splicing processing and then perform feature recognition, so as to identify the type of current cells; wherein, the training data set accounts for 80% of the total data, and the test data set accounts for 20% of the total data.

Images