s denotes a first pixel threshold value, w₀、h₀The width and height of the image at the first magnification are shown, respectively. The second pixel threshold may be set according to actual needs, for example, the second pixel threshold may be 2000, 2500, 3000, or the like.

As one implementation manner, a region having a pixel value greater than 2500 pixels and smaller than the first pixel threshold S in the first binarized image may be used as an impurity region, and the impurity region obtained thereby may be presented in a rectangular frame form.

After the impurity region is obtained, for S22, histogram statistics may be performed on the image in the impurity region represented in the rectangular frame format, and the statistical range is 256 gray levels from 0 to 255, so as to obtain the pixel amount at each gray level in the image in the same impurity region. And accumulating the pixel quantity corresponding to each gray level according to the sequence of the gray levels from low to high according to the statistical result of the histogram until the ratio of the pixel quantity of the accumulated sum to the total pixel quantity reaches a first percentage, and taking the value of the maximum gray level corresponding to the pixel quantity of the accumulated sum meeting the first percentage as the gray value of the rectangular frame. The gray scale value corresponding to each impurity region can be obtained according to the implementation mode. The first percentage is illustrated as 90% in this example. In other embodiments, the first percentage may be a set value of 85%, 88%, 92%, 95%, etc.

After obtaining the gray scale values of the plurality of impurity regions, in S23, all the gray scale values of the plurality of impurity regions may be counted to obtain a median of all the gray scale values, and the gray scale values of all the impurity regions involved in the counting may be reset to the median to obtain a first gray scale image.

Optionally, as shown in fig. 5, the S3 in the cell mass identification method may specifically include the following sub-steps: S31-S33.

S31: and calculating the cell ratio of the cervical liquid-based cells to be detected according to the images at the third time in the scanning images.

S32: and calculating the cell density of the cervical liquid-based cells to be detected according to the images at the third time in the scanning images.

S33: and calculating a target binary threshold value according to the cell density, the cell ratio and the first gray level image.

Wherein, S31, S32 can be executed simultaneously. For example, the cell fraction can be determined in the course of determining the cell density.

For S33, as an embodiment, histogram statistics may be performed on the first grayscale image obtained in step S2. The statistical range includes 256 gray levels of 0-255, wherein the gray level 0 corresponds to the pixel value A₀The gray level 1 corresponds to the pixel value A₁By analogy, the gray level 255 corresponds to the pixel value A₂₅₅. Accumulating the pixel values from the pixel value corresponding to the low gray level to the pixel value corresponding to the high gray level according to the gray level until the accumulated sum of the pixel values is larger than the cell nucleus pixel quantity, recording the maximum gray level corresponding to the accumulated sum larger than the cell nucleus pixel quantity as the current gray level GR, and continuing to accumulate until the accumulated sum is larger than the cell nucleus pixel quantity, wherein the gray level corresponding to the cell nucleus pixel quantity just larger than the cell nucleus pixel quantity is used as the current gray level GR, so that the updating of the current gray level GR is realized. Wherein, the current gray scale GR is the target binarization threshold value.

The cell nucleus pixel amount is calculated based on the cell density and the cell ratio in substeps S31 and S32. The cell nucleus pixel quantity can be calculated according to the following second expression:

in the second expression, N represents the amount of the cell nucleus pixel found, and w₀、h₀Respectively representing the width and height of the image at a first magnification, e.g. w₀、h₀The width and height of the first gray image can be expressed respectively, P represents the cell ratio, and X represents a parameter related to the cell density level. When the density levels are 0, 1, 2, 3, 4, 5, the corresponding X may be 100, 70, 45, 30, 25, 12, respectively.

In the process of S31-S33, in the process of obtaining the target binarization threshold value by calculation based on the cell ratio, the cell density and the first gray level image, the color corresponding to the cell nucleus pixel quantity can be fully considered by the participation of the cell ratio and the cell density in the calculation. The target binarization threshold determined based on the cell nucleus pixel quantity and the first gray level image without the impurity region can be used in the subsequent data filtering process, and determination of a more reasonable candidate region is facilitated.

Alternatively, the value of cell density may be one of a plurality of density levels. In one example, there may be 6 density levels, and if there are 6 density levels with increasing density represented by cell density levels 0-5, density level 0 may represent a very sparse distribution of cells, density level 1 may represent a relatively sparse distribution, and so on, density level 4 may represent a very dense distribution of cells,density level 5 may represent a very dense distribution of cells, and so on.

In order to calculate the cell density, the step S32 may specifically include the following sub-steps: S321-S325.

S321: and obtaining a third gray image according to the image at a third multiplying factor in the scanned image, wherein the third multiplying factor is lower than the first multiplying factor and the second multiplying factor.

S322: and obtaining a second binary image according to the third gray level image.

Taking the third magnification of 0.15625 times as an example, the scanned image of 0.15625 times can be grayed out to obtain a third grayscale image. The width and height of the third gray scale image may be within the pixel range of 250-500 pixels.

In order to accurately calculate the cell density, the width and the height of the third grayscale image may be normalized to obtain a size-normalized third grayscale image, where the size includes the width and the height. For example, it may be normalized to a pixel range of 300-400 pixels. If the size of the third grayscale image is a set uniform size, the normalization step can be omitted.

After the third gray level image with uniform size is obtained, a Gaussian filtering mode can be adopted to reduce noise, and an Otsu algorithm is adopted to calculate a binarization threshold value for calculating cell density. Note that this binarization threshold value for calculating the cell density is not the aforementioned target binarization threshold value. And performing binarization processing on the third gray level image according to the binarization threshold value for calculating the cell density to obtain a second binarization image. After the second binarized image is obtained, S323 is performed to determine a value of the cell density.

S323: the envelope boundary of the second binarized image is detected by the current detection rule, and step S324 or S325 is performed.

S324: when the envelope boundary of the second binary image is detected, the current cell density is obtained by updating, and the step S323 is skipped to continue to execute.

S325: and when the envelope boundary of the second binary image is not detected, taking the current cell density as the cell density corresponding to the second binary image.

With respect to the detection rule in S323, each current detection rule includes: and processing the current second binary image to obtain a new second binary image, carrying out edge detection on the new second binary image by adopting a Canny detection algorithm, and detecting the envelope boundary of the new second binary image by adopting Hough transform.

In one embodiment, the current second binarized image is expanded, eroded and expanded according to the current iteration coefficient in the current detection rule to obtain a new second binarized image.

In one example, four sets of iteration coefficients are used to determine the lower density level cell density. The four sets of iteration coefficients are: [11, 14, 3], [7, 10, 3], [3, 6, 3], [0, 3, 3], the size of the kernel used was 3 × 3, and each set of iterative coefficients was used for the dilation-erosion-dilation operation on the second binarized image. 0 in the iteration coefficient means that no corresponding operation is performed, e.g. [0, 3, 3] this iteration coefficient means that only the erosion-dilation operation is performed.

A complete iterative detection process will be described below with reference to the four sets of iterative coefficients.

The default cell density is 0 under the initial condition, expansion-corrosion-expansion operation is carried out by adopting [11, 14, 3], then edge detection is carried out by adopting a Canny detection algorithm, and envelope boundary of the second binary image is detected by adopting Hough transform. If the envelope boundary is not detected through Hough transformation, keeping the current density level unchanged and equal to 0, if the envelope boundary is detected, adding 1 to the density level to update the current density level, performing expansion-corrosion-expansion operation by adopting a next group of iteration coefficients [7, 10, 3], performing edge detection, detecting the envelope boundary through Hough transformation, updating the current density level when the envelope boundary is detected, and the like, and stopping iteration until the envelope boundary is not detected any more. Where the envelope boundaries may be circular outlines as shown in the three sub-graphs in fig. 4.

With the foregoing example, 5 cell density levels labeled 0-4 can be determined, and since the density labeled 5 corresponds to a very large density, in order to confirm an image with a cell density of 5, it can be determined whether the density level corresponding to the second binarized image is 5 or not by the following two conditions: (1) the current density grade obtained by updating the detection rule is 4; (2) the target binarization threshold GR calculated in the case where the density level is 4 is within a preset interval range.

As an embodiment, when the density level is determined to be 4 by the detection rule, the cell nucleus pixel amount N is calculated according to the parameter X associated with the density level 4, a temporary GR value is obtained based on N, the temporary GR value is determined, and when the temporary GR value is in an interval range of 30 or more and less than 70, the current density level is updated to 5, that is, the cell density value is 5.

After the cell density value is determined to be 5, a new cell nucleus pixel quantity N can be calculated according to the parameter X associated with the density level of 5, and a new GR value is obtained based on the new cell nucleus pixel quantity N and serves as a new target binarization threshold.

Wherein the cell fraction P can be calculated each time the envelope boundary is detected using the hough transform. If the cell density is 0, the cell fraction may be a default value, for example, 0.9, 0.8, etc. When the value of the cell density is other than 0, the cell ratio P can be calculated by the following third expression.

The third expression is:

in a third expression, r denotes the radius of the detected envelope boundary, e.g. the radius of the detected circular contour, w₁、h₁The width and height of the image at the third magnification are shown, respectively, and for example, the width and height of the second binarized image are shown, respectively.

The cell density determined by the method is high in accuracy and high in cell proportion reliability, the cell nucleus pixel quantity calculated according to the cell density and the cell proportion determined by the method is more reasonable, and the high-reliability target binarization threshold value can be obtained.

Alternatively, in order to avoid the missing of part of the cell mass region, the target binarization threshold GR obtained in the above method may be finely adjusted.

As an implementation manner for fine tuning the target binarization threshold, the target binarization threshold GR may be fine-tuned through the following fine-tuning strategy, where the fine-tuning strategy includes:

when the GR is less than or equal to 35 and the density levels are 0, 1 and 2, after the density levels are adjusted up by two levels, S33 is executed again to calculate a new target binarization threshold value; when GR is 35 or less and the density level is 3, the level is raised one by one, and S33 is executed again.

When GR is greater than 35 and less than 70 and the density level is 0, the density level is up-regulated by two levels, and S33 is executed again; when GR is greater than 35 and less than 70,density levels 1, 2, 3, 4, one level is raised, and S33 is re-executed.

After the final target binarization threshold is determined, S4 is executed.

Alternatively, for the above S4, the method may include the sub-steps of: S41-S42.

S41: and performing primary filtering treatment on the second gray scale image based on the cell density of the cervical fluid-based cells to be detected to obtain a primary filtering image.

Wherein, the primary filtering treatment may include: expansion operation, corrosion operation, Gaussian filtering and bilateral filtering. The parameters used in the specific implementation of the primary filtration treatment may be different for different cell densities.

S42: and carrying out binarization processing on the preliminary filtering image according to a target binarization threshold value to obtain a target binary image.

The pixel value in the preliminary filtered image may be compared with a target binarization threshold GR, and the preliminary filtered image is binarized based on the comparison result, so as to obtain a target binary image containing 0 and 1. S42 may be regarded as a re-filtering process for the preliminary filtered image, which can remove non-clustered regions in the de-noised image.

By the method, the images with different cell densities can be denoised respectively through a primary filtering processing mode, non-agglomerated areas such as non-cell aggregates, single cells and fine impurities can be removed through binarization processing of the primary filtering image, the non-agglomerated areas can be deleted under the condition that more suspicious cell aggregate areas are reserved as far as possible, and identification interference caused by the non-agglomerated areas is avoided.

Optionally, for the above S41, in one example, different primary filtering processing modes are set for the three types of cell density grades. In the three primary filtering processing modes, parameters such as the diameter of the adopted filtering field, the standard deviation of the spatial Gaussian function, the similarity standard deviation of the gray value and the like are different.

In the first category, for images with a high density level, for example, atdensity levels 2, 3, and 4, the second gray scale image is subjected to two iterations of the dilation-erosion operation, and the kernel size used in each iteration of the dilation-erosion operation is 3 × 3. And performing the next expansion-corrosion operation after the first expansion-corrosion operation is performed. And then, carrying out smooth denoising in a Gaussian filtering mode, wherein the kernel size adopted in the filtering process is 5 multiplied by 5, and filtering redundant fine impurities in a bilateral filtering mode.

In the second category, for an image with very sparse density, for example, when the density level is 1, the second gray image is subjected to iterative processing of three expansion-corrosion operations, and gaussian filtering and bilateral filtering are performed after the three iterative processing processes of expansion-corrosion are finished.

In the third category, for an extremely sparse image, for example, when the density level is 0, the second grayscale image is subjected to iterative processing of three dilations, and then gaussian filtering and bilateral filtering are performed. Wherein, for the third class of images, the target binary image B obtained after the first-stage filtering processing and the binarization processing can represent a temporary binary image B_tFor the temporary binary image B_tAnd after the iteration treatment of the three times of corrosion operation, obtaining a new target binary image B.

Different primary filtering modes are set for different cell density grades, so that efficient processing can be performed on different density grades to quickly remove non-agglomerated areas. Compared with a uniform processing mode, the method can avoid excessive processing time waste on the images of partial density levels, and is beneficial to improving the identification efficiency.

Optionally, after obtaining the target binary image B, regions in the target binary image that satisfy any one of the following conditions may be removed: (1) the width and the height of the region are lower than 14 pixels; (2) width below 8 pixels or height below 8 pixels; (3) the ratio between the width and the height or the inverse of the ratio exceeds 15. Thus, invalid regions such as a fine region having an excessively small number of pixels and a region having a severely unbalanced aspect ratio can be eliminated, and a clustered region can be retained.

For convenience of description, the image subjected to the above-described thinning-out operation is still referred to as a target binary image B.

After the final target binary image is obtained, S5 is executed. There may be a plurality of irregular regions in the final target binary image B. These irregular regions are regions in the target binary image B that are set to 1.

Alternatively, as an implementation manner of S5, the target binary image may be subjected to three iterative processes of erosion operation, that is, erosion-erosion operation. For the target binary image after the corrosion is finished, according to the position of each region in the image and the position mapping relation between the images under each scanning magnification, each related image M related to each region in the target binary image is found out from the first gray image_i. Then calculating the associated image M_iOf each associated image M_iAn average gray value can be calculated, and all the related images M in the same first gray image are used_iAfter all the average gray values are sorted, a specified number of associated images M are selected from the sorting result_iAs a candidate region. The number of the appointments can be set according to actual needs, for example, the number of the appointments can be 800, 1000, 1500, 2000, etc., and the example is explained with the number of the appointments being 1000.

Wherein for the actual associated image M_iWhen the number is less than the specified number, all the related images M are processed_iAre all taken as candidate regions.

After the candidate region is determined, S6 is performed.

For S6, since the image scan magnification corresponding to the candidate region in S5 is low, image data matching the candidate region can be mapped from the images at the selected scan magnification in the scanned images as the original data set for the recognition stage by the coordinate mapping relationship between the images at the respective scan magnifications. The selected scan magnification may be the highest scan magnification of the scanned image, in one example, the selected scan magnification is 20 times.

In order to improve the identification accuracy and speed, the original data set can be input into a pre-established neural network model for identification so as to obtain a cell mass identification result.

The cell mass recognition result may include a plurality of cell mass categories including: background, normal glandular cell mass, normal squamous epithelial cell mass, abnormal glandular cell mass, abnormal squamous epithelial cell mass, neutrophil aggregation, impurities, squamous epithelial cell overlap, mucus thread, fuzzy acellular mass, fuzzy cell mass.

The cell mass recognition result may further include a cell mass location for identifying coordinates of the recognized cell mass.

In order to make the sizes of the images in the original data set uniform for inputting the neural network model, the images in the original data set may be subjected to size processing and then input into the neural network model. For example, with 512 × 512 as a fixed input size, an area smaller than 512 pixels on any one side of each image in the original data set is expanded, and a sliding block fetching is performed on an area larger than 512 pixels on any one side, and the sliding step size may be 128 pixels.

By inputting the original data set into the convolutional neural network model, the predicted coordinates of the detection box and the cell mass category can be obtained. In practical application, elements in the same image may be identified by frames with different sizes, so that in order to avoid repeated identification of the same element, a non-maximum value suppression method can be adopted to remove repeated detection frames in the same area, and a final detection result can be obtained.

In one example, the convolutional neural network occupies 3G of video memory, one video card can run three processes simultaneously, 55 images can be tested in one second, and the time required for testing the original data set corresponding to the candidate regions of one slice in the identification stage is less than 20 seconds. From step S1 to the completion of step S6, the total operation time is about 40 seconds, the operation time is short, the recognition efficiency is high, and the reliability is high.

In the embodiment of the present application, the convolutional neural network architecture is shown in fig. 6. The arrows in fig. 6 may represent the data flow in the neural network.

After the raw data set is prepared, the raw data set is first fed into the multi-level convolutional layers in the base network for a series of convolution operations.

The multi-level convolutional layers in the base network include Conv1_1, Conv1_2, Conv2_1, Conv2_2, Conv3_1, Conv3_2, Conv3_3, Conv4_1, Conv4_2, Conv4_3, Conv5_1, Conv5_2,Conv5_ 3. The convolution kernel size of each convolution layer in the basic network is 3 multiplied by 3, the step length is 1, the padding is 1, and only the number of channels of each convolution layer is different. The number of lanes for Conv1_1 and Conv1_2 was 32, the number of lanes for Conv2_1 and Conv2_2 was 64, the number of lanes for Conv3_1, Conv3_2 and Conv3_3 was 128, the number of lanes for Conv4_1, Conv4_2 and Conv4_3 was 256, and the number of lanes for Conv5_1, Conv5_2 and Conv5_3 was 256.

Data that is convolved by the base network is fed into the test network. The test network has a special pooling layer, a multi-level convolutional layer and a multi-level anti-convolutional layer.

The multi-level convolutional layers in the test network include Conv5_4, Conv5_5, Conv6_1, Conv6_2, Conv7_1, Conv7_2, Conv8_1, Conv8_2, Conv9_1, Conv9_2, Conv10_1, Conv10_ 2. The special pooling layer may be disposed between any two of the multi-level convolutional layers of the test network.

The multilevel deconvolution layers in the test network include Deconv9_1, Deconv9_2, Deconv8_1, Deconv8_2, Deconv7_1, Deconv7_2, Deconv6_1, Deconv6_2, Deconv5_4, Deconv5_5, Deconv4_3_1, Deconv4_3_ 2.

After the image data set with the highest scanning magnification in the scanning image corresponding to the candidate area is input into the multi-stage convolution layer in the neural network, a first convolution result corresponding to part of convolution layers is obtained. The partial convolutional layer refers to the convolutional layer of which the Conv4_3-Conv10_2 part corresponds to the subsequent deconvolution layer.

And sending the first convolution result to a multi-stage deconvolution layer in the neural network to obtain a second convolution result, wherein the second convolution result is obtained after deconvolution operation is carried out on the deconvolution layer.

After obtaining the second convolution result according to the first convolution result, the confidence value group may be determined according to the first convolution result and the second convolution result. For example, the output results of part of the convolutional layers and the deconvolution layers are cascaded to realize the cascade of the feature maps, and the output of the multilayer is synthesized after the cascade of the multilayer to obtain the confidence value group.

And determining the cell mass recognition result according to each confidence value in the confidence value group.

If there are 11 classifications, a confidence array containing at least 11 elements may be obtained. For multiple confidence elements within the same set of confidence values, the class corresponding to the element with the highest confidence value may be a cell mass class in the object under test. Any set of confidence values may correspond to a detection box for identifying the subject, the detection box identifying the location of the cell mass under test.

The following describes the special pooling layers, convolution layers, and deconvolution layers in the test network.

The special pooling layer uses maximal pooling, the kernel size is 3 × 3, the step size is 1, and the padding is 1. The feature map obtained after the basic network and the special pooling layer can be sent to Conv5_4, wherein Conv5_4 is an expansion convolutional layer, the convolutional kernel is 3 × 3, padding is 6, the expansion coefficient is 6, and the number of channels is 512. Then, the data output by the Conv5_4 is sent to Conv5_5, the convolution kernel used by the Conv5_5 is 1 × 1, and the number of channels is 512.

Convolution operations of convolution layers Conv6_1 and Conv6_2 are the same as those of subsequent Conv7_1, Conv7_2, Conv8_1, Conv8_2, Conv9_1 and Conv9_2, and are described by taking Conv6_1 and Conv6_2 as examples. The convolution kernel size of Conv6_1 is 1 × 1, the number of channels is 64, the convolution kernel size of Conv6_2 is 3 × 3, the step size is 2, the padding is 1, and the number of channels is 128.

The convolution layer Conv10_1 uses convolution kernels of 1 × 1 size and 128 channels. The convolution kernel size of Conv10_2 is 4 × 4, step size is 1, padding is 1, and the number of channels is 256.

After the convolution operation of Conv10_2, the size of the feature map is 1 × 1, at this time, the feature map is sent to Deconv9_1 to be subjected to deconvolution operation, the convolution kernel size is 3 × 3, the step size is 2, the number of channels is 128, and the feature map size becomes 2 × 2. Then cascade the characteristic diagrams of Deconv9_1 and Conv9_2, and then perform convolution operation of Deconv9_2, with the convolution kernel size of 3 × 3, the step size of 1, the padding of 1, and the number of channels of 128. The processes of Deconv9_1 and Deconv9_2 are the same as the subsequent processes of Deconv8_1, Deconv8_2, Deconv7_1, Deconv7_2, Deconv6_1 and Deconv6_2, and are different from Deconv5_4, Deconv5_5, Deconv4_3_1 and Deconv4_3_2 only in the number of output channels, wherein the number of channels corresponding to Deconv5_4 and Deconv5_5 is 512, and the number of channels corresponding to Deconv4_3_1 and Deconv4_3_2 is 256.

The feature maps corresponding to Conv10_2, Deconv9_2, Deconv8_2, Deconv7_2, Deconv6_2, Deconv5_5 and Deconv4_3_2 all output the coordinates of the detection frames and the confidence values of the detection frame types, the detection frames with the confidence values higher than 0.3 are integrated, then the detection frames with high overlapping are removed by using a non-maximum suppression method, and the rest detection frames and the detection frames are used as recognition results.

Through the structural design of the convolutional neural network, the characteristics of a shallow layer and a deep layer can be fully fused, and the position and the category of the measured object can be more accurately predicted. Wherein the features of Conv4_3, Conv5_3 to Conv10_1 are considered shallow features and the features of Conv10_2, Deconv9_1 to Deconv4_3_2 are considered deep features. Because the detection frame coordinates and the confidence values of the detection frame types can be output in the feature maps of a plurality of scales through the convolutional neural network, the detection frames can better cover the detected objects under various sizes, and the detection effect on small targets is remarkable.

Referring to fig. 7, a functional block diagram of a cellmass recognition device 200 according to an embodiment of the present disclosure is shown. The cellmass recognition device 200 may be stored in thememory 102 of theelectronic device 100, and the cellmass recognition device 200 is configured to perform each step and any implementation of the cell mass recognition method.

As shown in fig. 7, the cellmass recognition apparatus 200 includes a scanimage acquisition module 201, apreprocessing module 202, athreshold calculation module 203, afiltering module 204, a candidateregion determination module 205, and arecognition module 206.

And a scannedimage obtaining module 201, configured to obtain scanned images of the cervical fluid-based cells to be detected at different scanning magnifications.

Thepreprocessing module 202 is configured to obtain a first grayscale image and a second grayscale image based on a scanned image, where a scanning magnification corresponding to the first grayscale image is a first magnification, and a scanning magnification corresponding to the second grayscale image is a second magnification.

And athreshold calculation module 203, configured to obtain a target binarization threshold based on the first grayscale image.

And thefiltering module 204 is configured to process the second grayscale image according to the target binarization threshold to obtain a target binary image.

A candidateregion determining module 205, configured to determine a candidate region according to the target binary image and the first grayscale image.

And theidentification module 206 is configured to identify the image with the highest scanning magnification in the scanned image according to the candidate region, so as to obtain a cell mass identification result.

The cellmass recognition apparatus 200 can perform the cell mass recognition method, and thus can rapidly recognize a cell mass.

Optionally, thepreprocessing module 202 may be further configured to perform steps S21-S23 in the cell mass identification method.

Optionally, thepreprocessing module 202 may also be used to perform substeps S211-S213 of the cell mass identification method described above.

Optionally, thethreshold calculation module 203 may be further configured to perform steps S31-S33 in the cell mass identification method.

Optionally, thethreshold calculation module 203 may also be configured to perform substeps S321-S325 of the cell mass identification method described above.

Optionally, thefiltering module 204 may be specifically configured to perform steps S41-S42 in the cell mass identification method.

For other details of the cellmass recognition device 200 provided in the embodiment of the present application, please further refer to the related description of the cell mass recognition method, which is not repeated herein.

In addition to the above embodiments, the present application further provides a readable storage medium, on which a computer program is stored, and the computer program is executed by theprocessor 101 to perform the steps in the cell mass identification method.

In summary, with the cell mass identification method, the cell mass identification device, and theelectronic device 100 provided in the embodiments of the present application, before identification by using a neural network, a target binary threshold for identification is obtained by calculation according to processing scanned images at different magnifications, and in the calculation process of the target binary threshold, the cell nucleus color is considered to be darker than the surrounding quality, so that a highly reliable target binary threshold is determined according to the cell density, the cell proportion parameter, and the grayscale image. The target binary image obtained based on the target binary threshold value is obtained by removing noise and non-cluster regions under the condition that suspicious cell cluster regions are kept as much as possible. The candidate region determined according to the target binary image has high reliability, and does not need to go through complicated feature extraction and fusion processes before entering the identification stage, so that the identification difficulty is reduced, the extraction efficiency of the candidate region is improved, and the whole identification process is accelerated. The convolutional neural network provided by the embodiment of the application is combined to identify the output result of the candidate region extraction stage, so that the processing time of the whole identification process can be shortened, and the accuracy and the reliability of the identification result are high.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of modules is only one division of logical functions, and there may be other divisions in actual implementation, and for example, multiple units or components may be combined or may be integrated, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

In addition, functional modules in the embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

It should be noted that the functions, if implemented in the form of software functional modules and sold or used as independent products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application or portions thereof that substantially contribute to the prior art may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device to perform all or part of the steps of the method of the embodiments of the present application. And the aforementioned storage medium includes: various media that can store program codes, such as a U disk, a removable hard disk, a memory, a magnetic disk, or an optical disk.

In this document, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.

The above embodiments are merely examples of the present application and are not intended to limit the scope of the present application, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying a cell mass, the method comprising:

calculating the cell proportion of the cervical liquid-based cells to be detected according to the images at the third time in the scanning images, and calculating the cell density of the cervical liquid-based cells to be detected according to the images at the third time in the scanning images; calculating a target binarization threshold value according to the cell density, the cell proportion and the first gray level image; the third magnification is lower than the first magnification and the second magnification;

2. The method of claim 1, wherein said deriving a first grayscale image based on the scanned image comprises:

identifying an impurity region of an image at a first magnification in the scanned image;

performing histogram statistics on the identified image of each impurity region to obtain a gray value corresponding to each impurity region;

and processing the gray values of the impurity regions according to the median of all the gray values of the impurity regions to obtain a first gray image.

3. The method of claim 2, wherein the step of identifying an impurity region of an image at a first magnification in the scanned image comprises:

graying an image under a first multiplying power in the scanned image to obtain first grayed data;

performing binarization processing on the first grayscale data according to a set first threshold value to obtain a first binarized image;

and screening the first binary image by using a set pixel threshold value to obtain an impurity region.

4. The method of claim 1, wherein the step of calculating the cell density of the cervical fluid-based cells to be tested from the image at the third time in the scan image comprises:

obtaining a third gray image according to the image at the third time in the scanned image;

obtaining a second binary image according to the third gray level image;

detecting an envelope boundary of the second binary image through a current detection rule;

when the envelope boundary of the second binary image is detected, updating to obtain the current cell density, and skipping to the step of detecting the envelope boundary of the second binary image through the current detection rule to continue to execute;

and when the envelope boundary of the second binary image is not detected, taking the current cell density as the cell density corresponding to the second binary image.

5. The method according to claim 1, wherein the step of processing the second gray scale image according to the target binarization threshold to obtain a target binary image comprises:

performing primary filtering treatment on the second gray level image based on the cell density of the cervical fluid-based cells to be detected to obtain a primary filtering image;

and carrying out binarization processing on the preliminary filtering image according to the target binarization threshold value to obtain a target binary image.

6. The method of claim 1, wherein the cell mass recognition result comprises a plurality of cell mass categories;

7. The method according to claim 1, wherein the step of identifying the image with the highest scanning magnification in the scanned image according to the candidate region to obtain the cell mass identification result comprises:

inputting the image data set with the highest scanning magnification in the scanning image corresponding to the candidate area into the multi-level convolution layer in the neural network to obtain a first convolution result corresponding to part of convolution layers;

sending the first convolution result to a multi-stage deconvolution layer in a neural network to obtain a second convolution result;

determining a confidence value set according to the first convolution result and the second convolution result;

8. A cell mass identification device, the device comprising:

a threshold calculation module, configured to calculate a cell proportion of the cervical fluid-based cell to be detected according to the image at the third time in the scanned image, and calculate a cell density of the cervical fluid-based cell to be detected according to the image at the third time in the scanned image; calculating a target binarization threshold value according to the cell density, the cell proportion and the first gray level image; the third magnification is lower than the first magnification and the second magnification;

9. An electronic device, characterized in that the electronic device comprises: a processor, a memory storing machine-readable instructions executable by the processor, the machine-readable instructions when executed by the processor performing the steps of the method of any of claims 1-7 when the electronic device is run.