A method for detecting embryo pronuclear area

Abstract

The present invention discloses a method for detecting the pronuclear area of ​​an embryo, comprising: constructing a pronuclear detection model using a preprocessed embryo source file, wherein the preprocessed embryo source file includes multiple embryo images; the pronuclear detection model outputs an embryo pronuclear area detection result, wherein the embryo pronuclear area detection result includes one or more groups of pronuclear coordinates, target classification, and confidence; performing a comprehensive area calculation on the embryo pronuclear area detection result, and outputting a pronuclear area A; the comprehensive area calculation includes a balanced calculation method, a denoising calculation method, and an enhanced calculation method.

Description

Technical Field

[0001] The present invention relates to the field of embryo development detection based on time-lapse photography, and in particular to a method for embryo pronucleus detection and area calculation. Background Art

[0002] Currently, time-lapse imaging is widely used to assess embryo quality because it can determine the timing of embryonic developmental events and measure embryonic morphological and dynamic parameters. For example, Chinese patent CN101331500B discloses using pixel differences in a series of embryo images acquired through time-lapse photography to generate a difference image. A time series graph formed from the parameter sets of the difference images is used to obtain information on the amount of cell rearrangement, thereby determining embryo quality.

[0003] In order to use images obtained by time-lapse imaging to detect embryo quality, it is necessary to first solve the problem of image detection of various elements in the embryo growth process. However, there is no good technical solution in the existing technology.

[0004] One of the reasons is that compared with other target detection, the pronuclear area required for detection is smaller and the difference in pixel values ​​between the pronuclear area and the surrounding area is not obvious, which increases the difficulty of target detection. In addition, the pronuclear morphology changes dynamically during development, which increases the diversity of pronuclear features required for identification.

[0005] The second reason is that the pronucleus changes dynamically during development. Therefore, the pronuclear area varies visually from different time points and angles. Existing technical solutions rely on a single method for area calculation, making it difficult to accurately measure the area of ​​a dynamic pronucleus. Therefore, to improve the efficiency and accuracy of embryonic pronuclear area detection, a method for pronuclear area detection is needed. Summary of the Invention

[0006] The present invention aims to improve the efficiency and accuracy of embryo pronuclear area detection and provide an embryo pronuclear area detection method.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A method for detecting the pronuclear area of ​​an embryo, comprising:

[0009] Building a pronuclear detection model using an embryo source file, wherein the embryo source file includes a plurality of embryo images;

[0010] The pronuclear detection model outputs an embryonic pronuclear region detection result, wherein the embryonic pronuclear region detection result includes one or more sets of pronuclear coordinates, target classification, and confidence level;

[0011] Performing comprehensive area calculation on the embryo pronuclear region detection results and outputting the pronuclear area A;

[0012] The comprehensive area calculation includes a balance calculation method, a noise removal calculation method and an enhancement calculation method.

[0013] Furthermore, the prokaryotic detection model includes:

[0014] A feature network is used to extract features from the embryo image, with the embryo image as input and the embryo pronuclear feature map as output.

[0015] A candidate region network is used to extract regions of interest from the embryonic pronuclear feature map, the input of which is the embryonic pronuclear feature map, and the output of which is an embryonic region of interest, wherein the embryonic region of interest is composed of multiple sets of pronuclear coordinates and target classifications;

[0016] The classification and regression network takes as input the embryo pronuclear feature map and the embryo region of interest, fine-tunes the embryo region of interest, and outputs classification loss, coordinate loss, and the confidence level.

[0017] Furthermore, the candidate regional networks include:

[0018] In the first step, the embryonic pronuclear feature map is convolved with a 3*3 convolution kernel, where each convolution kernel corresponds to k basic candidate boxes.

[0019] The second step includes two parallel 1*1 convolutional layers.

[0020] The first convolutional layer is used to classify the basic candidate boxes, each of which corresponds to a 2-dimensional target classification;

[0021] The second convolutional layer is used to obtain the border position information, and each basic candidate box corresponds to a 4-dimensional original kernel coordinate.

[0022] Furthermore, the k is 9.

[0023] Furthermore, the classification and regression network includes a classification loss algorithm performed on the first convolutional layer and a regression loss algorithm performed on the second convolutional layer.

[0024] Furthermore, the balance calculation method comprises the steps of:

[0025] If the width and height of the pronucleus given by the pronucleus coordinates are the same, the pronucleus area a1 is calculated based on the width and height;

[0026] If the width and height of the pronucleus given by the pronucleus coordinates are different, the pronucleus area a1 is calculated by weighted averaging the pronucleus areas calculated using the width and height as the radius.

[0027] Furthermore, the denoising calculation method comprises the steps of:

[0028] Obtaining a pronuclear image based on the pronuclear coordinates, and performing a binarization operation using the OTSU method;

[0029] Remove noise and small connected areas;

[0030] Morphological operations are used to eliminate holes by dilating and then corroding.

[0031] Fill the interior of the pronucleus with white;

[0032] The area of ​​the pronuclear image is calculated based on the width and height of the pronuclear given by the pronuclear coordinates.

[0033] The number of white pixels is obtained, that is, the original nucleus area a2.

[0034] Furthermore, the enhanced calculation method comprises the steps of:

[0035] Obtaining a pronuclear image based on the pronuclear coordinates, and performing Gaussian filtering on the pronuclear image;

[0036] By formula forming a detection operator, wherein the detection operator traverses the entire pronuclear image to obtain pronuclear derivative images in the x and y directions;

[0037] Get the original kernel derivative image in the x and y directions and calculate the gradient of each pixel;

[0038] Non-maximum suppression is used to refine the pronuclear edge;

[0039] The pronuclear image after non-maximum suppression is processed by hysteresis thresholding to remove pseudo edges;

[0040] The least squares method is used to fit the pronuclear region circle.

[0041] The best function matching of the circle is obtained by calculating the minimum sum of squares of the errors,

[0042] The original nucleus area a3 is calculated based on the best function match of the circle.

[0043] Furthermore, the pronuclear area A=0.2*a1+0.3*a2+0.5*a3.

[0044] The present invention can perform pronucleus detection on complex embryo images through the detection model, and the pronucleus position in the output result is accurate, with low missed detection rate and false detection rate.

[0045] The present invention obtains reliable pronuclear area parameters in complex embryo images through a comprehensive area algorithm, providing reliable decision parameters for embryo development prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above technical content of the present invention and the following detailed description will be better understood when read in conjunction with the accompanying drawings. It should be noted that the accompanying drawings are only examples of the technical solutions claimed. In the accompanying drawings, the same reference numerals represent the same or similar elements.

[0047] Figure 1 is a schematic diagram of a basic candidate frame in one embodiment of the present invention;

[0048] Figure 2 is a schematic diagram of the first step of the candidate area network operation in one embodiment of the present invention;

[0049] Figure 3 1 is a schematic diagram of the detection results of the embryonic pronuclear region in one embodiment of the present invention;

[0050] Figure 4 Schematic diagram of a decision tree for determining the success rate of embryo formation into blastocysts in one embodiment of the present invention. DETAILED DESCRIPTION

[0051] The detailed features and advantages of the present invention are described below in the specific embodiments, and the content is sufficient to enable those skilled in the art to understand the technical content of the present invention and implement it accordingly. Based on this specification, claims and drawings, those skilled in the art can easily understand the relevant purposes and advantages of the present invention.

[0052] In this article, the term "IOU" stands for Intersection over Union, a measure of the accuracy of detecting corresponding objects in a specific dataset. The IoU score is a standard performance metric for object classification and segmentation problems. Given a set of images, the IOU measure gives the similarity between the predicted region and the ground truth region of an object present in the set of images and is defined by the following equation, where TP, FP, and FN represent the true positive, false positive, and false negative counts, respectively:

[0053]

[0054] In this article, the term "ground truth box" refers to the real annotation box, which represents the original image label information, that is, the manual annotation, and is generally regarded as the "true value".

[0055] In this article, the term "OTSU" refers to an efficient algorithm for binarization of images proposed by Japanese scholar OTSU in 1979, which uses a threshold to divide the original image into two images: foreground and background.

[0056] The present invention discloses a method for detecting the pronuclear area of ​​an embryo, comprising:

[0057] Constructing a pronuclear detection model using embryo source files, which include multiple embryo images;

[0058] The pronuclear detection model outputs the embryonic pronuclear region detection results, which include one or more sets of pronuclear coordinates, target classification, and confidence level.

[0059] Perform comprehensive area calculation on the embryo pronuclear region detection results and output the pronuclear area A;

[0060] Comprehensive area calculation includes balance calculation method, noise removal calculation method and enhancement calculation method.

[0061] Specifically, the embryo source file can be formed by time-difference imaging, and each file represents the embryonic development process detected at a certain layer (the camera intermittently shoots multiple layers of embryonic development).

[0062] The detection model consists of three modules: feature network, candidate region network, and classification and regression network.

[0063] The feature network is used to extract features from the embryo image, with the embryo image as input and the embryo pronuclear feature map as output.

[0064] A CNN convolutional neural network is used to extract features from the original embryo input image, outputting a 13*13*256 embryo pronuclear feature map. Alternatively, existing pre-trained models such as VGG16 can be used to extract features from embryo pronuclear images.

[0065] The candidate region network is used to extract the region of interest (pronuclear region) in the embryo pronuclear feature map, the input is the embryo pronuclear feature map, and the output is the embryo region of interest, which consists of multiple sets of pronuclear coordinates and target classifications.

[0066] The region proposal network consists of two steps.

[0067] The first step of the candidate region network operation is to convolve the 13*13*256 embryonic pronuclear feature map with a 3*3 convolution kernel. 3*3 refers to a sliding window. The output is an 11*11*256 embryonic pronuclear feature map.

[0068] Each of the convolution kernels corresponds to k basic candidate boxes.

[0069] Specifically, such as Figure 1 As shown in Figure 1, three rectangular boxes with different scales and aspect ratios can be selected as basic candidate boxes: three scales: {128, 256, 512}, and three ratios: {1:1, 1:2, 2:1}. In this case, k is 9.

[0070] Each embryonic pronuclear feature map output by the feature network has 13*13=169 pixels and uses boundary padding. When performing a 3*3 convolution, each pixel can be used as the center point of a 3*3 convolution kernel. The entire convolution is equivalent to 169 convolution centers. Each of the 169 convolution centers has 9 basic candidate boxes. Therefore, relative to the original input embryonic pronuclear image, there are a total of 169*9=1521 candidate boxes. These candidate boxes cover the entire area of ​​the original embryonic image and have a large amount of overlapping areas.

[0071] like Figure 2 As shown, the left side shows that the first step of the candidate box network operation uses a 3*3 convolution kernel to convolve the 13*13 embryo pronuclear feature map. One convolution kernel corresponds to 9 basic candidate boxes (not all candidate boxes are shown in the figure).

[0072] The second step consists of two parallel 1*1 convolutional layers:

[0073] The first convolutional layer is used to classify the basic candidate frames. Specifically, the target classification is divided into two categories: presence and absence, that is, whether there is a target in the embryo candidate frame or not.

[0074] According to the 169 convolution centers obtained in the first step, there are 9 candidate boxes. Each embryo candidate box has two values ​​(i.e., yes or no, expressed by probability). Each of the basic candidate boxes corresponds to a target classification, which is a 2-dimensional vector.

[0075] The second convolutional layer is used to obtain the border position information. Each candidate box has 4 position values ​​(x, y, w, h). Each of the basic candidate boxes corresponds to a protokernel coordinate, which is a 4-dimensional vector.

[0076] In one embodiment, 1521 embryo candidate frames can be obtained through the first two steps of the candidate region network operation, and each embryo candidate frame has a target classification and pronuclear coordinates. The non-maximum suppression algorithm is used to remove some redundant embryo candidate frames from the 1521 embryo candidate frames, and 300 embryo regions of interest with the highest probability are selected.

[0077] The candidate box network defines embryo training samples as follows: Positive samples are selected from the 300 embryo ROIs: the candidate box with the largest IOU with the ground truth box or an IOU with the ground truth box greater than 0.5 is considered a positive sample. Negative samples are samples with an IOU with the ground truth box less than 0.1. Ultimately, this patent uses 128 embryo ROIs as a single embryo image for training, and all 128 positive and negative samples are normalized (subtracting the mean and dividing by the standard deviation).

[0078] The classification and regression network takes as input the embryonic pronuclear feature map and the embryonic region of interest (ROI), fine-tunes the embryonic region of interest, and outputs classification loss, coordinate loss, and confidence. Confidence includes two aspects: the confidence that the object is within the predicted coordinate box and the confidence that the predicted coordinate box includes the entire object. The classification and regression network includes a classification loss algorithm for the first convolutional layer and a regression loss algorithm for the second convolutional layer. The classification loss algorithm can use softmax, and the coordinate regression loss algorithm can use smooth L1.

[0079] The total losses are as follows:

[0080]

[0081] Lcls: softmax loss multi-classification loss function;

[0082] Where pi represents the probability that the i-th box predicted by the network is the target. pi* represents the original embryo label information. If the iou between the i-th box and the original label box is greater than 0.7, then pi*=1, and if the iou is less than 0.3, then pi*=0;

[0083] Lreg: smooth L1 loss;

[0084] Where ti represents the parameter coordinate vector, which indicates the offset between the predicted box and the original label box, and ti* represents the corresponding original label box (ground truth);

[0085] Ncls is the size of a training data block during training;

[0086] Nreg represents the feature map size;

[0087] The calculation formula is as follows:

[0088]

[0089]

[0090] Test results such as Figure 3 shown.

[0091] Comprehensive area calculation, that is, multiple prokaryotic area calculation methods are used for weighted summation to calculate the final prokaryotic area value:

[0092] (1) Prokaryotic area detection method 1 (balance calculation method):

[0093] Based on the width and height coordinates output by the pronucleus detection model, the pronucleus area a1 can be directly calculated based on the width and height output. If the detection model outputs different widths and heights, a strategy is used to first calculate the pronucleus area using the width as the radius, then using the height as the radius, and then calculating the pronucleus area a1 by weighted averaging the pronucleus areas calculated using the width and height as the radii. Specifically, the weighting coefficient can be 0.5.

[0094] (2) Prokaryotic area detection method 2 (denoising calculation method):

[0095] ① Obtain the pronuclear image based on the pronuclear coordinates and use the OTSU method to perform binarization operation;

[0096] ② After the binarization operation, there are some scattered noises in the original nuclear image. Here, the noise is removed, small connected areas are removed, and small connected areas with an area of ​​less than 30 and isolated point noise are directly removed;

[0097] ③ There will be some holes in the denoised image, and the morphological operation of first dilation and then corrosion (closing operation) is used to eliminate the holes;

[0098] ④ Fill the interior of the pronucleus with white (RGB 255,255,255);

[0099] ⑤ Calculate the area of ​​the entire image based on the width and height of the pronucleus given by the pronucleus coordinates. In the image, white (RGB 255,255,255) represents the pronucleus area, and black (RGB 0,0,0) represents the background area;

[0100] ⑥ Get the number of white pixels in the pronuclear area, which is the pronuclear area size a2.

[0101] (3) Prokaryotic area detection method three (enhanced calculation method):

[0102] ① Gaussian filtering is performed on the prokaryotic image output by the prokaryotic detection model. The Gaussian kernel size is set to 3*3 to reduce the interference of noise in the prokaryotic image on the detection results.

[0103] ② The pronuclear gradient image is a rough pronuclear edge image. The characteristic of the pronuclear image edge is the mutation of the pixel value at that point. The following formula is used to reflect the degree of mutation:

[0104]

[0105] Since the image coordinates x and y are discrete, the approximate formula is:

[0106]

[0107] A simple detection operator [-1,1] is formed, which traverses the entire pronuclear image to obtain the edge of the pronuclear image in the x direction, and the same method is used in the y direction;

[0108] ③ Get the original nuclear derivative image in the x and y directions and calculate the gradient of each pixel. The formula is as follows:

[0109]

[0110] It represents a vector perpendicular to the edge of the original nucleus pixel. The direction of the vector is the gradient direction, and the magnitude of the vector is the intensity of the original nucleus edge.

[0111] ④ Use non-maximum suppression to refine the edges of the pronucleus, refining the broad edges in the original rough edge detection image into true edges, thereby better highlighting the pronucleus outline. Non-maximum suppression detects whether the gradient direction of each pixel in the pronucleus is a maximum pixel. If it is, it is retained. If not, the pixel is set to 0. If the gradient direction of the pixel is not at a specific angle (0 degrees, 45 degrees), a soft pixel (the weighted sum of the pixels on both sides) is used.

[0112] ⑤ Apply hysteresis threshold processing to the pronuclear image after non-maximum suppression to remove pseudo edges.

[0113] Hysteresis thresholding uses dual thresholding, first using a high threshold to filter out pixels that are definitely edges, and then using a low threshold to extend the edge. It traverses the eight surrounding pixels. If there is a pixel with a value greater than the high threshold, then the pixel is considered an extended edge, its pixel value is increased, and this change is reflected in the image. The final edge is the true original core edge.

[0114] ⑥ Calculate the pronuclear area by fitting the pronuclear with the least squares method;

[0115] Based on the edge coordinate information output by the protokeratin edge detection, the least squares method is used to fit the protokeratin area circle. The optimal function matching of the circle is obtained by calculating the sum of squares of the minimization error, so that the sum of squares of the error between the obtained data and the actual data is minimized.

[0116] The least squares circle fitting model formula is derived as follows:

[0117] The formula for calculating the area of ​​a circle is as follows:

[0118] (x-x0) 2 +(y-y0) 2 =r 2

[0119] For the circle fitting of the least squares method, the optimization objective function of the square error is:

[0120]

[0121] In the formula, (xi,yi) represents the coordinates of the feature points on the edge of the protocore; (x0,y0) represents the coordinates of the center of the circle, and n is the number of feature points involved in the fitting.

[0122] On the premise of maintaining the characteristics of this optimization objective function, we use a slightly different improved method to define the square of the error, which can obtain a direct solution to the minimization problem, defined as follows:

[0123]

[0124] ⑦ Make E get the minimum value, finally get the coordinates of the center of the pronucleus and the radius, and calculate the area of ​​the pronucleus area3;

[0125] The pronucleus area values ​​are calculated using the three area calculation methods mentioned above. In the pronucleus area detection, we can see that the first area calculation method has a flaw. When the pronucleus is not circular, it will affect the actual pronucleus area value. The second area calculation method has a flaw. When the foreground and background of the pronucleus in the image are not obvious (when the edge pixel value is similar to the surrounding pixel value), the surrounding pixels will be calculated into the pronucleus area. Therefore, the weights of the three area values ​​are assigned to 0.2, 0.3, and 0.5 respectively. The final pronucleus area is:

[0126] Area=0.2*area1+0.3*area2+0.5*area3

[0127] like Figure 4 As shown, this is a decision-making method for determining the success rate of embryo formation into blastocyst using decision rules.

[0128] It is necessary to input one embryo image per second for processing according to the pronucleus detection model. The first result output by the detection model containing the pronucleus image can be used to calculate the time when the pronucleus appears during the embryonic development process (pronucleus formation time) and obtain the position coordinates of the two pronuclei (relative pronuclei position); then, according to the pronucleus area calculation method, calculate the area difference of the two pronuclei when the pronucleus is formed (area difference when pronucleus is formed); according to the pronucleus detection model, input one embryo image per second for processing, the last result output by the detection model containing the pronucleus image can be used to calculate the time when the pronucleus ends during the embryonic development process (pronucleus disappearance time), and obtain the position coordinates of the two pronuclei (relative pronuclei position); then, according to the pronucleus area calculation method, calculate the area difference of the two pronuclei eight hours before the pronucleus disappears and when the pronucleus disappears (pronucleus area difference eight hours before the pronucleus disappears, pronucleus area difference when the pronucleus disappears).

[0129] Specifically, the decision-making methods include:

[0130] S1. Judge based on the prokaryotic formation time, retain if it is less than 15 hours, and discard if it is more than 15 hours;

[0131] S2. Determine the number of pronuclei according to the output of the pronucleus detection model. If the number of pronuclei is equal to 2, the embryo is retained. If the number of pronuclei is 0, 1, 3, etc., it is discarded (number of pronuclei);

[0132] S3. Judge based on the disappearance time of the pronucleus, if it is less than 20 hours, keep it, if it is more than 20 hours, discard it;

[0133] S4-1. Determine the relative positions of the two pronuclei when they are formed. The closer one is output as B, and the farther one is output as A.

[0134] S4-2, judge based on the difference in pronuclear area when the pronucleus is formed, output A if there is a difference, and B if there is no difference;

[0135] S4-3, judge based on the difference in prokaryotic area 8 hours before prokaryotic disappearance, output A if there is a difference, and B if there is no difference;

[0136] S4-4, judge based on the difference in the pronucleus area when the pronucleus disappears, output B if there is a difference, and A if there is no difference;

[0137] S4-5. Judge the relative positions of the two pronuclei when the pronuclei disappear. The closer one is output as A, and the farther one is output as B.

[0138] The above S4 are all parallel steps.

[0139] According to the above judgment rules, A represents a 20% probability and B represents a 10% probability, so the success rate of blastocyst formation can be finally obtained based on the relevant parameters of the pronucleus.

[0140] The terms and expressions used herein are for descriptive purposes only, and the present invention is not limited to these terms and expressions. The use of these terms and expressions does not exclude any equivalent features shown and described (or portions thereof), and it should be recognized that various modifications that may exist are also intended to be included within the scope of the claims. Other modifications, variations, and substitutions are also possible. Accordingly, the claims should be deemed to cover all such equivalents.

[0141] Similarly, it should be pointed out that although the present invention has been described with reference to the current specific embodiments, ordinary technicians in this technical field should realize that the above embodiments are only used to illustrate the present invention, and various equivalent changes or substitutions can be made without departing from the spirit of the present invention. Therefore, as long as the changes and modifications to the above embodiments are within the scope of the essential spirit of the present invention, they will fall within the scope of the claims of the present invention.

Claims

1. A method for detecting the pronuclear area of ​​an embryo, characterized in that: include: Building a pronuclear detection model using an embryo source file, wherein the embryo source file includes a plurality of embryo images; The pronuclear detection model outputs an embryo pronuclear region detection result, wherein the embryo pronuclear region detection result includes one or more sets of pronuclear coordinates, target classification, and confidence level; Performing comprehensive area calculation on the embryo pronuclear region detection results and outputting the pronuclear area A; The comprehensive area calculation includes a balance calculation method, a noise removal calculation method and an enhancement calculation method, wherein: - The equilibrium calculation method comprises the steps of: if the width and height of the pronucleus given by the pronucleus coordinates are the same, calculating the pronucleus area a1 based on the width and height; If the width and height of the pronuclear given by the pronuclear coordinates are different, the pronuclear area a1 is calculated by weighted averaging the pronuclear areas calculated using the width and height as the radius respectively; - The denoising algorithm comprises the steps of: obtaining a pronuclear image based on the pronuclear coordinates, and performing a binarization operation using the OTSU method; Remove noise and small connected areas; Morphological operations are used to eliminate holes by dilating and then corroding. Fill the interior of the pronucleus with white; The area of ​​the pronuclear image is calculated based on the width and height of the pronuclear given by the pronuclear coordinates. Get the number of white pixels, that is, the original nucleus area a2; - The enhanced algorithm comprises the steps of: Obtaining a pronuclear image based on the pronuclear coordinates, and performing Gaussian filtering on the pronuclear image; By formula forming a detection operator, wherein the detection operator traverses the entire pronuclear image to obtain pronuclear derivative images in the x and y directions; Get the original kernel derivative image in the x and y directions and calculate the gradient of each pixel; Non-maximum suppression is used to refine the pronuclear edge; The pronuclear image after non-maximum suppression is processed by hysteresis thresholding to remove pseudo edges; The least squares method is used to fit the pronuclear region circle. The best function matching of the circle is obtained by calculating the minimum sum of squares of the errors, The original nucleus area a3 is calculated based on the best function match of the circle.

2. The method for detecting the embryo pronuclear area according to claim 1, wherein: The prokaryotic detection model includes: A feature network is used to extract features from the embryo image, with the embryo image as input and the embryo pronuclear feature map as output. A candidate region network is used to extract regions of interest from the embryonic pronuclear feature map, the input of which is the embryonic pronuclear feature map, and the output of which is an embryonic region of interest, wherein the embryonic region of interest is composed of multiple sets of pronuclear coordinates and target classifications; The classification and regression network takes as input the embryo pronuclear feature map and the embryo region of interest, fine-tunes the embryo region of interest, and outputs classification loss, coordinate loss, and the confidence level.

3. The method for detecting the embryo pronuclear area according to claim 2, wherein: The candidate area networks include: In the first step, the embryonic pronuclear feature map is convolved with a 3*3 convolution kernel, where each convolution kernel corresponds to k basic candidate boxes. The second step includes two parallel 1*1 convolutional layers. The first convolutional layer is used to classify the basic candidate boxes, and each basic candidate box corresponds to a 2-dimensional target classification; The second convolutional layer is used to obtain the border position information, and each basic candidate box corresponds to a 4-dimensional original kernel coordinate.

4. The method for detecting the embryo pronuclear area according to claim 3, wherein: The k is 9.

5. The method for detecting the embryo pronuclear area according to claim 3, wherein: The classification and regression network includes a classification loss algorithm performed on the first convolutional layer and a regression loss algorithm performed on the second convolutional layer.

6. The method for detecting the embryo pronuclear area according to any one of claims 1 to 5, wherein: The pronuclear area A = 0.2 * a1 + 0.3 * a2 + 0.5 * a3.

Images