CN107169556A - stem cell automatic counting method based on deep learning - Google Patents

Abstract

An automatic stem cell counting method based on deep learning, in the field of cell counting technology, comprising the following steps: S11: using a phase-contrast microscope to photograph cells to be counted to generate a stem cell image. S12: image preprocessing; performing noise reduction processing and illumination equalization processing on the cell image to obtain a stem cell image with uniform illumination. S13: Remove cell artifacts generated during phase contrast microscopy. S14: Segment the stem cell image after artifact removal, and acquire multiple candidate stem cell images. S21: Manually mark the plurality of segmented candidate cell images to establish a training set. S22: Input the training set into CNN for training. S23: Counting the stem cell counting results. The invention solves the shortcomings of traditional manual cell counting that consumes a lot of manpower, and also overcomes the defect that the flow counting method will destroy the cell growth environment, and counts through the captured cell images, which has the characteristics of stability, high efficiency, automaticity, and non-destructiveness.

Description

Automatic counting method of stem cells based on deep learning

technical field

The invention relates to the technical field of cell counting, in particular to an automatic stem cell technology method based on deep learning.

Background technique

With the advancement of human science, more and more researchers are devoted to revealing the mysteries of life. Stem cells are a type of cell that can replicate and proliferate by itself, and under certain circumstances, can differentiate into other types of cells. Therefore, research on the growth, proliferation and differentiation of stem cells is an important research direction in cell biology.

Accurate cell counting is of great significance to the research work of cell growth and division. The existing traditional cell research method mainly uses researchers to directly observe cell samples through a microscope, which not only consumes a lot of time, but also has complicated procedures. There is serious subjectivity. In addition, in order to study the division, proliferation and differentiation process of stem cells, the cells are often stained and fluorescent technology is added, which will have a certain impact on cell activity. Therefore, how to maintain the natural active state of pluripotent stem cells in an artificial culture environment and make them divide without intervention has become a key research issue. Using microscopic observation technology and computer vision processing technology, it is possible to quantitatively analyze stem cell images.

Therefore, generally speaking, there are three methods for cell counting in the prior art, and the disadvantages of these three types of cell counting methods will be described below.

The first: Coulter counting method. The Coulter counting method is to put the electrolyte solution into the measurement tube, suspend the particle group in the electrolyte solution, there is a fine hole on the wall of the measurement tube, and there is a certain voltage between the holes and electrodes, and the particles suspended in the electrolyte will move with the electrolyte solution. When passing through the small hole tube, the same volume of electrolyte is replaced, and the current changes due to the change of resistance and is recorded on the recorder. Finally, the electrical signal can be converted into particle size. This method can be used to obtain the particle size distribution. The Coulter counting method can treat cells as particles, so it is widely used in automated blood cell counters and has great significance in medicine. However, this method is not suitable for stem cell research. First, the stem cells on the culture medium cannot pass through the counter, otherwise the stable environment for stem cell growth will be destroyed and the results of the research will be affected. Second, this method is to obtain the total number of cells in the solution by statistical means, but for stem cell research, the cell count needs to be accurate to the single digit, so this method is not applicable.

The second method: artificial statistical method; by manually counting the cell images under the microscope, the counting accuracy is relatively high. The workload of this method is huge and consumes a lot of valuable human resources. When the amount of data is large, this method is almost impossible to realize. In addition, due to the subjectivity of manual work, although the counting is relatively accurate, due to the inconsistency of the professional level of the observers, it is possible to The state of the cell was misjudged.

The third method: counting cells by flow cytometer. Flow cytometry is an earlier instrument used for quantitative analysis of cell numbers. Flow cytometry counts in the state of cell flow. First, these flowing cells pass through the detection area of the flow cytometer. The excited fluorescent antibody emits fluorescence of a certain wavelength, and the cytometer detects the excitation light and converts the light signal into an electrical signal, and also measures some physical and biochemical characteristic parameters of the cell. The characteristic parameter distinguishes different types of cells, and counts the number of different types of cells. However, the actual spatial position of the sample entering the flow cytometer has been destroyed, so the measurement of the cytometer lacks spatial position information, and the actual spatial position information of the sample is very important for the study of cells, because in the tissue In the microenvironment, the movement of a class of cells from one location in the microenvironment to another is likely to seriously affect the state of the tissue or organ. At the same time, the sample preparation before entering the cytometer requires a lot of manual participation. Therefore, the loss cell counter is not suitable for counting stem cells.

Contents of the invention

The purpose of the present invention is: according to the background technology, there are following technical problems in the prior art: (1) the Walter counting method can destroy the stable environment of stem cell growth on the culture medium, and affect the results of the research; (2) artificial The statistical method has a huge workload and consumes a lot of valuable human resources. When the amount of data is large, this method is almost impossible to realize, and the manual counting method has a certain degree of subjectivity. Although the counting is relatively accurate, due to the inconsistent professional level of the observers, there are Misjudgment of the state of cells may occur; (3) counting by flow cytometer will destroy the actual spatial position of the sample, and the sample preparation before entering the cytometer requires a lot of manual participation. In order to solve these three problems, the present invention provides a method for counting stem cells based on deep learning for stem cell digital images.

Technical scheme of the present invention is as follows:

A method for automatically counting stem cells based on deep learning, comprising the steps of:

S1: Use a phase-contrast microscope to photograph the cells to be counted to generate stem cell images;

S2: Segment the stem cell image by cell segmentation technology to obtain multiple potential candidate stem cell images; specifically, S2 includes the following steps:

S21: Image preprocessing; performing noise reduction processing and illumination equalization processing on the stem cell image to obtain a stem cell image with uniform illumination;

S22: removing cell artifacts generated during phase contrast microscopy;

S23: Segment the stem cell image after artifact removal, and acquire multiple candidate stem cell images;

S3: cell identification; specifically, S3 includes the following steps:

S31: Manually mark the plurality of segmented candidate stem cell images to establish a training set;

S32: Input the training set into CNN for training, and for each candidate stem cell image, CNN will output a result indicating whether it is a cell;

S32 comprises the following steps:

S33: Counting the number of stem cells in all the plurality of candidate stem cell images to obtain a stem cell counting result.

Preferably, a Gaussian filter is used for the denoising processing of the stem cell image, and a background subtraction method is used for the illumination equalization processing.

Specifically, the specific process of S22 is:

Suppose g is the imaging image, f is the original image, H is the imaging matrix, and C is the background model. According to the imaging principle of the phase contrast microscope, the imaging process is defined as a linear model:

g≈Hf+C

If the background has been removed, the model can be simplified as:

g≈Hf

The process of removing artifacts is the process of restoring f from g, using a constrained quadratic function to restore f:

Among them, L is the Laplacian matrix that defines the similarity of neighbors of spatial pixels, Λ is a positive diagonal matrix, w _s and w _r are regularization terms with different weights obtained through grid search learning;

The above formula has no closed solution and can only be approximated numerically. Therefore, the constraint restoration function f has a non-negative solution, and the formula can be transformed into an optimization problem:

O(f)＝f ^T Qf+2b ^T f+c stf≥0

Q＝H ^T H+w _s L+w _t ∑

b＝-H ^T gw _t ∑ ^T f ^(t) +w _r diag(Λ)/2

Among them, c is a constant term w _t is the weighting factor of time consistency regularization, f ^(t) is f corresponding to time t, s, r, Q, b are the customary expressions of weighting factors in constrained quadratic programming problems ; Use the non-negative multiplication update algorithm to get f, and get the stem cell image after artifact removal:

Specifically, the specific process of S23 is:

S231: Perform binarization processing on the stem cell image after artifact removal to obtain a stem cell image with clear edges;

S232: Recalculate the area of each connected domain in the stem cell image;

S233: Discard the part whose area is lower than the threshold, and use the rest as the candidate cell area, and the threshold is set according to the resolution of the image cell image;

S234: Cutting candidate cell regions to obtain candidate stem cell images.

Specifically, the specific steps of S31 are:

S311: Set classification labels on part of the segmented stem cell images, and divide them into two types: cells and non-cells;

S312: Taking samples classified as cells as positive samples and samples classified as non-cells as negative samples, rotate and expand the positive samples, and rotate each picture by 90°, 180°, and 270°;

S313: Perform mirror transformation on each original image and rotated image in the positive sample, expand one positive sample image into 8 images, and form the training set together with the positive sample, its expanded image and the negative sample.

Specifically, the specific steps of S32 are:

S321: The input data is a matrix composed of 24×24 pixels, the first feature layer uses a 5×5 window to convolve the input image, and the size of each feature map is obtained as 20×20;

S322: Perform a downsampling operation on the feature map obtained by convolution in the first downsampling layer to obtain the same number of feature maps with a size reduced to 10×10;

S323: Carry out another convolution operation and downsampling operation, and finally obtain a 1×2 output result, that is, the candidate image is a cell or a non-cell;

S324: Compare the output result with the corresponding label in the training set, feed back the comparison result to CNN, use the backpropagation algorithm to update the weights of neurons in CNN, repeat this process, and the recognition accuracy will gradually increase until Relatively stable, and finally a CNN that can be used for candidate image classification.

Specifically, the specific process of S33 is: input all candidate cell images of a microscopic image into CNN for classification, and for each candidate image, CNN will output a result indicating whether it is a stem cell, and then count all cell images judged by CNN The number of candidate images of , to get the total number of stem cells in the input microscopic image.

After adopting the above scheme, the beneficial effects of the present invention are as follows: the present invention mainly utilizes the digital image processing technology to automatically count the stem cells taken by the phase contrast microscope, which completely solves the shortcomings of traditional manual cell counting that consumes a lot of manpower, and also overcomes the Overcoming the defect that flow cytometry will destroy the cell growth environment, counting through the captured cell images has the characteristics of stability, high efficiency, automatic, and non-destructive.

Description of drawings

Fig. 1 is the image before removing artifacts after noise reduction and background removal in the present invention;

Fig. 2 is the image after artifact removal among the present invention;

Fig. 3 is a schematic diagram of the structure of the convolutional neural network model;

Fig. 4 is a schematic diagram of full connection and partial connection of the local perception principle in a specific embodiment;

Fig. 5 is a schematic diagram of weight sharing in a specific embodiment;

Figure 6 is a structural diagram of a convolutional neural network.

detailed description

The solutions of the present invention will be further described in detail below in conjunction with specific embodiments.

The method for automatically counting stem cells based on deep learning comprises the following steps:

S1: Use a phase-contrast microscope to photograph stem cells that need to be counted, and generate stem cell images;

S2: Segment the stem cell image by cell segmentation technology to obtain multiple potential candidate stem cell images; specifically, S2 includes the following steps:

S21: image preprocessing; performing noise reduction and illumination equalization on the stem cell image to obtain a stem cell image with uniform illumination; avoiding the impact on the accurate segmentation of cells due to noise and uneven illumination.

In this embodiment, a Gaussian filter is used to denoise the stem cell image. The Gaussian filter is widely used in the denoising process of image processing, and replaces the pixels in the image with the weighted average of itself and other pixel values in the field. Very effective for suppressing noise that follows a normal distribution.

The one-dimensional zero-mean Gaussian function is:

Among them, σ is the parameter of the Gaussian function, and its physical meaning is the variance of the Gaussian function, which needs to be set manually in actual use, and the width of the Gaussian function is determined by σ. The two-dimensional zero-mean discrete Gaussian function is often used as a smoothing filter for image processing. The two-dimensional Gaussian function is:

Then, the image is de-illuminated by the background subtraction method to achieve the effect of light balance. Suppose the background C is a binomial polynomial model, namely

C(u,v)＝k ₀ +k ₁ u+k ₂ v+k ₃ u ² +k ₄ uv+k ₅ v ²

Among them, u, v are the horizontal and vertical coordinates of the pixel, and k0-k5 are the coefficients of the model. If we want to estimate k, we must know the background pixel g _c . To get g _c , the background must be segmented. To do this, first consider all pixels as the background, and then estimate the value of k based on the least squares method:

k ^* ＝(A ^T A) ^-1 A ^T g

Among them, A is the image matrix. After obtaining the restoration image and the segmentation result, the estimated value of the background can be improved through repeated iterations, and the background C=Ak ^* can be calculated for a picture, and then the background can be removed according to g←gC to obtain a stem cell image with balanced illumination .

S22: Remove the cell artifacts generated during the phase-contrast microscope shooting; first, the cause of the artifacts is related to the imaging principle of the phase-contrast microscope. The cells in the culture dish are almost completely transparent because they have not been dyed. It is almost impossible to obtain effective image information by direct observation or shooting with the naked eye using ordinary optical microscopy methods. Therefore, a special optical instrument is required. : phase contrast microscopy. The phase contrast microscope is composed of two special devices, the annular grating and the phase plate, which are added to the ordinary optical microscope. Since the cells have a certain thickness, when the light passes through the cells, a phase difference will be generated. The phase contrast microscope can convert this phase difference into human The amplitude difference (light intensity difference) visible to the eye will produce halo-shaped cell artifacts around the cells during the shooting process, which will interfere with the segmentation work.

The specific steps for removing cell artifacts are:

Suppose g is the imaging image, f is the original image, H is the imaging matrix, and C is the background model. According to the imaging principle of the phase contrast microscope, the imaging process is defined as a linear model:

g≈Hf+C

If the background has been removed, the model can be simplified as:

g≈Hf

The process of removing artifacts is the process of restoring f from g, using a constrained quadratic function to restore f:

Among them, L is the Laplacian matrix that defines the similarity of neighbors of spatial pixels, Λ is a positive diagonal matrix, w _s and w _r are regularization terms with different weights obtained through grid search learning;

The above formula has no closed solution and can only be approximated numerically. Therefore, the constraint restoration function f has a non-negative solution, and the formula can be transformed into an optimization problem:

O(f)＝f ^T Qf+2b ^T f+c stf≥0

Q＝H ^T H+w _s L+w _t ∑

b＝-H ^T gw _t ∑ ^T f ^(t) +w _r diag(Λ)/2

Use the non-negative multiplication update algorithm to obtain f, and obtain the cell image after artifact removal. Figure 1 is the original image, and Figure 2 is the image after artifact removal.

S23: Segment the stem cell image after artifact removal to obtain candidate stem cell images;

The specific process of S23 is:

S231: Perform binarization processing on the stem cell image after artifact removal to obtain a stem cell image with clear edges;

S232: Recalculate the area of each connected domain in the stem cell image;

S233: Discard the part whose area is lower than the threshold, and use the rest as the candidate cell area, and the threshold is set according to the resolution of the stem cell image. In this embodiment, the minimum area threshold is set to 10 pixels;

S234: Cutting candidate cell regions to obtain candidate stem cell images.

Before describing step S2, first explain the convolutional neural network. The convolutional neural network CNN has become a research hotspot in the field of artificial intelligence such as speech recognition, image recognition, and video analysis. It has strong applicability, feature extraction and classification at the same time, Global optimization with fewer training parameters has become a research hotspot in the field of machine learning. The basic network structure of convolutional neural network can be divided into five parts: input layer, feature extraction layer (convolutional layer), feature mapping layer (downsampling layer), full connection layer and output layer. Each neuron in the feature extraction layer is connected to the corresponding local receptive field of the upper layer, and the features of the local receptive field are extracted through filters and nonlinear transformations. The downsampling layer reduces the dimensionality of the feature vector while increasing the anti-distortion ability of the model. A simple convolutional neural network structure consists of two convolutional layers (C ₁ , C ₃ ) and two downsampling layers (S ₂ , S ₄ ) alternately, as shown in Figure 3. In order to reduce the number of parameters in the convolutional neural network and reduce the complexity of the system, the present invention adopts the principle of local perception for the following reasons: for an image of 1000×1000 pixels, it needs to contain 1 million hidden layer neurons. Connection structure, there are 1000×1000×1000000= ¹⁰¹² connections, corresponding to ¹⁰¹² weight parameters, if the local connection structure is used, the local receptive field is set to 10×10, and each receptive field of the hidden layer only needs to be compared with these 10 ×10 local images are connected, only 10 ⁸ parameters are needed, which can be reduced by 4 orders of magnitude, which can significantly simplify the training process. The difference between fully connected and partially connected structures is shown in Figure 4.

In addition to using local connection structure to reduce parameters, weight sharing can also be used to reduce the number of training parameters. Weight sharing, that is, neurons on the same feature map use the same weight parameters. The schematic diagram of weight sharing is shown in Figure 5. Suppose, in the local connection network, each neuron is connected to a 10×10 image area, that is, each neuron corresponds to 100 connection parameters. If these 100 connection parameters are set to be the same, it is equivalent to letting each neuron use the same convolution kernel for image convolution, then only 100 parameters can cover the entire image visual field, ensuring the validity of information Under the premise of greatly reducing the number of parameters.

Next, continue with the step description above:

S3: cell identification; specifically, S2 includes the following steps:

S31: Manually mark the segmented stem cell images to establish a training set; the specific steps of S31 are:

S311: Set classification labels on part of the segmented stem cell images, and divide them into two types: cells and non-cells;

S312: Take the samples labeled as cells as positive samples, and use the samples labeled as non-cells as negative samples, rotate and expand the positive samples, and rotate each picture by 90°, 180°, and 270°;

S313: Perform mirror transformation on each original image and rotated image in the positive sample, expand one positive sample image into 8 images, and form the training set together with the positive sample, its expanded image and the negative sample.

S32: Input the training set into CNN for training; as shown in Figure 6, the specific steps of S22 are:

S321: The input data is a matrix composed of 24×24 pixels, the first feature layer uses a 5×5 window to convolve the input image, and the size of each feature map is obtained as 20×20;

S322: Perform a downsampling operation on the feature map obtained by convolution in the first downsampling layer to obtain the same number of feature maps with a size reduced to 10×10;

S323: Carry out another convolution operation and downsampling operation, and finally obtain a 1×2 output result, that is, the candidate image is a cell or a non-cell;

S324: Compare the output result with the corresponding label in the training set, feed back the comparison result (error) to the CNN, use the backpropagation algorithm to update the weights of the neurons in the CNN, repeat this process, and the recognition accuracy will gradually increase Increase until relatively stable, and finally get a CNN that can be used for candidate image classification.

S33: Statistical cell counting results. Input all candidate stem cell images of a microscopic image into CNN for classification. For each candidate stem cell image, CNN will output a result indicating whether it is a cell, and count it to obtain the number of cells.

After adopting the method of this embodiment, experiments were carried out on the stem cell images provided by the Institute of Biomedicine and Health, Chinese Academy of Sciences, and the results are shown in the following table:

picture number manual counting number of candidate cells Redundancy rate% CNN count Correct rate% 200 15 14 93.33 14 93.33 250 37 39 105.41 36 97.30 300 81 90 111.11 77 95.06 350 138 151 109.42 130 94.20

The invention completely solves the shortcomings of the traditional manual cell counting that consumes a lot of manpower, and also overcomes the defect that the flow counting method will destroy the cell growth environment, and counts through the captured cell images, which has the characteristics of stability, high efficiency, automatic, non-destructive, etc., from As can be seen from the above table, the accuracy rate is relatively high.

The present invention is not limited to the above specific embodiments, and it should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative work. In short, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art shall be within the scope of protection defined by the claims.

Claims (7)

1. the stem cell automatic counting method based on deep learning, it is characterised in that comprise the following steps：

S1：The cell for needing to count is shot using phase contrast microscope, stem cell image is generated；

S2：Stem cell image is split by cell segmentation technology, potential multiple candidate stem cells images are obtained；Specifically Ground, S2 comprises the following steps：

S21：Image preprocessing；Noise reduction process is carried out to stem cell image and illumination equalization processing obtains the dry thin of uniform illumination Born of the same parents' image；

S22：Remove the cell artifact produced in phase contrast microscope shooting process；

S23：The stem cell image for removing pseudo- movie queen is split, multiple candidate stem cells images are obtained；

S3：Cell recognition；Specifically, S3 comprises the following steps：

S31：Manual markings are carried out to the multiple candidate stem cells images being partitioned into, training set is set up；

S32：Training set input CNN is trained, to each candidate stem cells image, CNN can export a result table Whether show it is cell；

S33：Stem cell population in all multiple candidate stem cells images of statistics, draws stem cell count result.

2. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that right in S21 The noise reduction process of stem cell image uses Gaussian filter, and illumination equalization processing uses background subtraction method.

3. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that S22 tool Body process is：

If g is image, f is original image, and H is imaging array, and C is background model, according to phase contrast microscope image-forming principle, Imaging process is defined as a linear model：

g≈Hf+C

If background has been removed, model can be reduced to：

g≈Hf

The process for going artifact is exactly to recover f process from g, and f is recovered using a constraint quadratic function：

<mrow> <mi>O</mi> <mrow> <mo>(</mo> <mi>f</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>|</mo> <mo>|</mo> <mi>H</mi> <mi>f</mi> <mo>-</mo> <mi>g</mi> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>+</mo> <msub> <mi>w</mi> <mi>s</mi> </msub> <msup> <mi>f</mi> <mi>T</mi> </msup> <mi>L</mi> <mi>f</mi> <mo>+</mo> <msub> <mi>w</mi> <mi>r</mi> </msub> <mo>|</mo> <mo>|</mo> <mi>&amp;Lambda;</mi> <mi>f</mi> <mo>|</mo> <msub> <mo>|</mo> <mn>1</mn> </msub> </mrow>

Wherein, L is the Laplacian Matrix of definition space pixel neighbours' similarity, and Λ is a Positive diagonal matrix, w_s、w_rBe The lower weight factor for learning to obtain by grid search of different regularization rules (takes w in testing herein_s=1, w_r=0.01)；

Above-mentioned formula be not closed solution, can only numerical radius, therefore constraint reconstruction f have non-negative solution, formula is converted into and asked Optimization problem：

O (f)=f^TQf+2b^Tf+c s.t.f≥0

Q=H^TH+w_sL+w_t∑

B=-H^Tg-w_t∑^Tf^(t)+w_rdiag(Λ)/2

Wherein, c is constant term, w_t(w is taken for the weighted factor of time consistency regularization in testing herein_t=0.1), f^(t)For t Moment corresponding f；F is obtained using non-negative multiplication more new algorithm, obtains removing the stem cell image of pseudo- movie queen.

4. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that S23 tool Body process is：

S231：Binary conversion treatment is carried out to the stem cell image for removing pseudo- movie queen, sharp-edged stem cell image is obtained；

S232：The area of each connected domain in stem cell image is calculated again；

S233；Give up the part that area is less than threshold value, remaining as candidate cell region, threshold value is differentiated according to figure cell image Rate is set；

S234：Candidate cell region is cut, candidate stem cells image is obtained.

5. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that S31 tool Body step is：

S311：The stem cell image that a part is partitioned into sets tag along sort, is divided into cell and the class of acellular two；

S312：To be categorized as the sample of cell as positive sample, to be categorized as the sample of acellular as negative sample, to positive sample Rotation expansion is carried out, every pictures are carried out with 90 °, 180 °, 270 ° of rotations；

S313：Minute surface conversion is carried out to each original image in positive sample and postrotational image, a positive sample image is expanded Fill for 8, positive sample and its expansion image and negative sample are collectively constituted into training set.

6. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that S32 tool Body step is：

S321：Input data is the matrix that 24 × 24 pixels are constituted, and first feature figure layer is using 5 × 5 window to input Image carries out convolution, and the size for obtaining each characteristic pattern is 20 × 20；

S322：The characteristic pattern obtained in first down-sampling layer to convolution carries out down-sampling operation, obtains same amount of feature Figure, size is reduced to 10 × 10；

S323：A convolution operation and down-sampling operation are carried out again, finally give the output result of one 1 × 2, i.e. candidate image It is cell or acellular；

S324：Output result label corresponding with training set is contrasted, comparing result CNN is fed back into, using reverse Propagation algorithm is updated to the weights of neuron in CNN, repeats this process, and the accuracy of identification can be gradually risen until phase To stable, a CNN that can be used for candidate image to classify is finally obtained.

7. the stem cell automatic counting method according to claim 1 based on deep learning, it is characterised in that S33 tool Body process is：All candidate cell images input CNN of one micro-image is classified, to each candidate image, CNN A result will be exported and represent whether it is stem cell, all candidate images for being determined as cell image by CNN are then counted Quantity, obtains inputting the stem cell sum of micro-image.