Model growth method for recognition model of cervical cell slide digital image

Abstract

The present disclosure describes a model growth method for a recognition model of a cervical cell slide digital image, including: recognizing a target pathological slide digital image from a first target source to obtain a plurality of original sub-features, obtaining a plurality of combination sets based on the plurality of original sub-features, each combination set in the plurality of combination sets including at least one original sub-feature, using a preliminary recognition module to extract features from the combination set to obtain data features, and obtaining a preliminary recognition result based on the data features, using a result summary module to obtain a terminal recognition result matching the target pathological slide digital image based on the plurality of preliminary recognition results, the preliminary recognition module is trained by the cervical cell slide digital image from the second target source, and the result summary module is trained by the cervical cell slide digital image from the first target source. In this case, the recognition model can be applied to pathological slide digital images of different hospitals.

Description

Technical Field

[0001] The present disclosure generally relates to a model growing method for a recognition model of a digital image of a cervical cell slide. Background Art

[0002] As the country increases its efforts to popularize cervical screening, the demand for cervical screening in various regions has also increased. However, due to the long training cycle of pathologists, the pathologist resources of various hospitals cannot cope with such a large demand for screening. Therefore, it is of practical significance to use computer algorithm models to assist in the diagnosis of digital images of pathological slides to help pathologists speed up diagnosis.

[0003] Currently, the more widespread recognition process for digital pathology slide images is to first extract features from the digital pathology slide images through an algorithm model, and then obtain recognition results based on the image features.

[0004] However, in different hospitals, the reagents used to make cervical cell slides, the doctors' production processes, and the scanning instruments are different, which will lead to differences in the numerical distribution of contrast, brightness, etc. of cervical cell pathology digital images. The recognition results will also have different feature distributions due to the different sources of the digital images. Therefore, in general, diagnostic recognition models trained using only single-source data or statistical constants are more likely to overfit. Summary of the invention

[0005] The present disclosure is completed in view of the above-mentioned state of the prior art, and its purpose is to provide a model growth method for a recognition model of cervical cell slide digital images that is applicable to different hospitals and reduces the recognition accuracy of the recognition model due to overfitting.

[0006] To this end, the first aspect of the present disclosure provides a model growth method for a recognition model of a cervical cell slide digital image, comprising: recognizing a target pathology slide digital image from a first target source to obtain a plurality of original sub-features, the original sub-features including the type of diseased cells and the number of diseased cells in the target pathology slide digital image, obtaining a plurality of combination sets based on the plurality of original sub-features, each combination set in the plurality of combination sets including at least one original sub-feature, performing feature extraction on the combination sets using a preliminary recognition module to obtain data features, and obtaining preliminary recognition results based on the data features, obtaining a terminal recognition result matching the target pathology slide digital image based on the plurality of preliminary recognition results using a result summary module, the preliminary recognition module being trained using the cervical cell slide digital image from the second target source, and the result summary module being trained using the cervical cell slide digital image from the first target source.

[0007] In this case, multiple preliminary recognition modules can use different feature extraction models and classification models to extract features and recognize the target pathology slide digital image to obtain preliminary recognition results. And the result summary module can be used to process the preliminary recognition results and obtain terminal recognition results. Because the preliminary recognition module is trained with pathology slide digital images from multiple hospitals, the recognition model can have good recognition effects on various images. Because the result summary module is trained with pathology slide digital images from the target hospital, the recognition model can have good recognition effects for the target hospital. Therefore, the recognition model can be applied to the recognition of target pathology slide digital images in different hospitals.

[0008] In addition, in the model growth method involved in the first aspect of the present disclosure, optionally, the number of the preliminary recognition modules is not greater than the first preset number, and the number of the preliminary recognition modules is the same as the number of the combination sets. In this case, the overfitting phenomenon caused by the excessive number of preliminary recognition modules can be avoided, and the number of preliminary recognition modules can be limited based on the actual calculation speed of the recognition model, and one preliminary recognition module can perform feature extraction and recognition for one combination set.

[0009] In addition, in the model growth method involved in the first aspect of the present disclosure, optionally, the multiple preliminary recognition modules are screened based on the feature importance of the preliminary recognition results obtained by each preliminary recognition module. In this case, a preliminary recognition result with higher feature importance can be obtained.

[0010] In addition, in the model growth method involved in the first aspect of the present disclosure, optionally, the data features include image features and statistical features. In this case, different types of data features can be extracted according to different combination sets.

[0011] In addition, in the model growth method involved in the first aspect of the present disclosure, optionally, the preliminary recognition module obtains multiple preliminary recognition results through different feature extraction methods and classification methods. In this case, different feature extractors have different feature extraction effects, thereby being able to extract different types of features in the target pathological slide digital image, thereby improving the accuracy of recognition.

[0012] In addition, in the model growth method involved in the first aspect of the present disclosure, optionally, the preliminary recognition result includes whether a certain type of lesion exists in the target pathological slide digital image. In this case, the type of the preliminary recognition result can be selected according to actual needs.

[0013] In addition, the model growth method involved in the first aspect of the present disclosure may optionally further include at least one augmentation module, and when the preliminary identification module or the result summary module is used, the augmentation module obtains an augmented image and an augmented annotation result matching the augmented image based on the cervical cell slide digital image and the annotation result matching the cervical cell slide digital image. In this case, the sample augmentation system can augment the pathology slide digital image and the annotation result matching the pathology slide digital image, and can use different sample augmentation methods to process the pathology slide digital image to obtain augmented images with different features, or can augment the pathology slide digital image multiple times based on actual needs and obtain an augmented image that meets the actual needs.

[0014] To this end, the second aspect of the present disclosure provides a recognition model for cervical cell slide digital images, wherein the cervical cell slide digital images include multiple target pathology slide digital images with different resolutions, characterized in that the recognition model includes: an acquisition module, a combination module, multiple preliminary recognition modules and a result summary module, the acquisition module is configured to recognize the target pathology slide digital image from a first target source to obtain multiple original sub-features, the original sub-features include the diseased cell type and the number of diseased cells in the target pathology slide digital image; the combination module is configured to obtain multiple combination sets based on the multiple original sub-features, the multiple Each combination set in the combination set includes at least one original sub-feature; the preliminary recognition module includes a feature extraction unit and a recognition unit, the feature extraction unit is configured to extract features from the combination set to obtain data features, and the recognition unit is configured to obtain preliminary recognition results based on the data features; the result summary module is configured to obtain a terminal recognition result matching the target pathology slide digital image based on multiple preliminary recognition results; wherein the multiple preliminary recognition modules are trained by cervical cell slide digital images from the second target source, and the result summary module is trained by cervical cell slide digital images from the first target source. In this case, the multiple preliminary recognition modules can use different feature extraction models and classification models to extract features and recognize the target pathology slide digital image to obtain preliminary recognition results. And the result summary module can be used to process the preliminary recognition results and obtain terminal recognition results. Because the preliminary recognition module is trained by pathology slide digital images from multiple hospitals, the recognition model can have good recognition effects on various images. Because the result summary module is trained by pathology slide digital images from the target hospital, the recognition model can have good recognition effects for the target hospital. Therefore, the recognition model can be applied to the recognition of target pathology slide digital images in different hospitals.

[0015] According to the present disclosure, a model growth method for a recognition model of a cervical cell slide digital image that is applicable to different hospitals and reduces the possibility that the recognition model has low recognition accuracy due to overfitting can be provided. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Embodiments of the present disclosure will now be explained in further detail, by way of example only, with reference to the accompanying drawings, in which:

[0017] Figure 1 is a block diagram showing a recognition model involved in the examples of the present disclosure.

[0018] Figure 2 is a schematic diagram showing a digital image of a pathology slide involved in the examples of the present disclosure.

[0019] Figure 3 Schematic diagram showing a partial area of ​​a digital image of a pathology slide involved in the example of the present disclosure.

[0020] Figure 4 Schematic diagram showing an original feature of a partial area of ​​a digital image of a pathology slide involved in the example of the present disclosure.

[0021] Figure 5 is a flow chart showing an identification method involved in the examples of the present disclosure.

[0022] Figure 6 is a flowchart showing a model growth method of a recognition model involved in an example of the present disclosure.

[0023] Figure 7 is an example diagram showing an augmented image obtained by an augmentation module based on Gaussian horizontal and vertical stripes involved in the example of the present disclosure.

[0024] Figure 8 It is an example diagram showing the working principle of the augmentation module based on image mosaic involved in the example of the present disclosure.

[0025] Fig. 9 is an example diagram showing an augmented image obtained by an augmentation module based on image mosaicking involved in the examples of the present disclosure. DETAILED DESCRIPTION

[0026] Hereinafter, the preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same symbols are assigned to the same components, and repeated descriptions are omitted. In addition, the accompanying drawings are only schematic diagrams, and the ratio of the dimensions of the components or the shapes of the components may be different from the actual ones.

[0027] The present disclosure relates to a recognition model 10 for pathological slide digital images, a recognition method of the recognition model 10 , and a model growth method of the recognition model 10 . Figure 1 is a block diagram showing a recognition model involved in the examples of the present disclosure. Figure 2 is a schematic diagram showing a digital image of a pathology slide involved in the examples of the present disclosure. Figure 3 Schematic diagram showing a partial area of ​​a digital image of a pathology slide involved in the example of the present disclosure. Figure 4 is a schematic diagram showing an original feature of a partial area of ​​a digital image of a pathological slide involved in an example of the present disclosure. Figure 1 The present disclosure provides a recognition model 10 for digital images of pathological slides. In some examples, the recognition model 10 for digital images of pathological slides may also be referred to as a recognition model 10 for digital images of cervical cell slides. In some examples, the recognition model 10 involved in the present disclosure may be configured to recognize digital images of pathological slides of a target source. In the present disclosure, the recognition model 10 for digital images of pathological slides may also be referred to as a growth model 10 or model 10.

[0028] For some examples, see Figure 1 , the recognition model 10 may include an acquisition module 110, a plurality of preliminary recognition modules 130 and a result summary module 140. In this case, the recognition result of the pathological slide digital image can be obtained by using the acquisition module 110, the combination module 120, the preliminary recognition module 130 and the result summary module 140. In some examples, the recognition result can be used for preliminary screening of diseased cells in the pathological slide digital image. In this case, the recognition result of the pathological slide digital image can be used to assist in reducing the diagnosis time of the pathologist.

[0029] In some examples, the acquisition module 110 may be configured to acquire the original features of the target pathology slide digital image. In some examples, the combination module 120 may be configured to receive the original features to obtain a plurality of combination sets. In some examples, the preliminary recognition module 130 may be configured to perform operations such as feature extraction and recognition on the target pathology slide digital image. In this case, the data features of the pathology slide digital image can be obtained, and the pathology slide digital image can be recognized based on the data features. In some examples, the result summary module 140 may be configured to obtain a terminal recognition result based on a plurality of preliminary recognition results. In this case, the preliminary recognition result can be processed using the result summary module 140 and a terminal recognition result can be obtained.

[0030] In some examples, each of the multiple preliminary recognition modules 130 can be trained using multiple groups of pathology slide digital images from different sources. In this case, the recognition model 10 can have a good recognition effect on pathology slide digital images from various sources. In some examples, the result summary module 140 is trained using multiple groups of pathology slide digital images from the same source as the target pathology slide digital images. In this case, the recognition model 10 can have a good recognition effect for the target hospital. Thus, the recognition model 10 can be applied to the recognition of pathology slide digital images from different target sources.

[0031] In some examples, such as Figure 2 and Figure 3 As shown, the pathology slide digital image can be a cervical cell slide digital image. In this case, the cervical cell slide digital image can be identified.

[0032] In some examples, the acquisition module 110 may be configured to acquire original features. In some examples, the acquisition module 110 may be configured to acquire original features of a target pathology slide digital image.

[0033] In some examples, the target pathology slide digital image may come from a pathology slide digital image. In some examples, the pathology slide digital image may have multiple target pathology slide digital images with multiple resolutions. In some examples, the target pathology slide digital image may come from a target source. In some examples, the target source may be a hospital that needs to use the recognition model 10. In this case, the acquisition module 110 can be used to acquire the target pathology slide digital image of the hospital that needs to use the recognition model 10. Pathology slide digital image Pathology slide digital image In some examples, the pathology slide digital image can be obtained by a pathology slide scanner. In some examples, the pathology slide digital image may be a whole slide image (WSIs). In this case, the whole slide image can be recognized.

[0034] In some examples, the original feature refers to the recognition result obtained after recognizing the target pathological slide digital image. In some examples, the original feature can be obtained by recognition. In some examples, the original feature can be obtained by recognition based on a machine learning method. In some examples, the original feature can include multiple original sub-features. In some examples, the original sub-features can be recognition results for different information of diseased cells.

[0035] In some examples, the acquisition module 110 can be configured to identify the target pathology slide digital image from the first target source to obtain multiple original sub-features. In some examples, the pathology slide digital image from the first target source can be a pathology slide digital image from multiple medical institutions. In other words, the first target source can include multiple medical institutions. In some examples, the first target source can include a second target source (described later). In some examples, the first target source may also not include the second target source.

[0036] In some examples, such as Figure 4 As shown, the original feature may include multiple original sub-features, and the original sub-feature may be a diseased cell type. In some examples, the original sub-feature may also be cell information such as the location of the diseased cell, the number of diseased cells, or confidence. For example, the original sub-feature may be a diseased cell type (for example, it is believed that the diseased cell belongs to a high degree of squamous intraepithelial lesion cell), and the original feature may include information such as the number of high degree of squamous intraepithelial lesion cells, the location, size, shape, confidence, and nuclear size of each high degree of squamous intraepithelial lesion cell. However, the present disclosure is not limited thereto, and the original feature may also include information of a recognition frame for selecting the diseased cell, such as information such as the location, size, number, aspect ratio, and cell area ratio of the recognition frame. In some examples, the original feature may also include source information of the target pathological slide digital image. In this case, the preliminary recognition module 130 is capable of processing different types of recognition results.

[0037] In some examples, the original feature may include the type of diseased cells in the target pathological slide digital image. In some examples, when the target pathological slide digital image is a cervical cell slide digital image, the type of diseased cells may include low-grade squamous intraepithelial lesions (LSIL), high-grade squamous intraepithelial lesions (HSIL), and atypical squamous cells of undetermined significance (ASC-US) and other diseased cell types of cervical cells. In some examples, the original feature may include the position of the diseased cells in the target pathological slide digital image. In some examples, the diseased cells may be framed by an identification box. In some examples, the identification box may be a rectangle. In some examples, the position of the diseased cells may be indicated by the following coordinates (uppermost position (ymax), lowermost position (ymin), leftmost position (xmin), rightmost position (xmax)). In some examples, the original sub-feature may also be a heat map based on the target pathological slide digital image. In some examples, the heat map may reflect the data information in a two-dimensional matrix or table by color changes, thereby being able to intuitively represent the size of the pixel value with a preset color depth. In some examples, the pixel values ​​in the heat map may correspond to the degree of diseased cells.

[0038] In some examples, the pathology slide digital image may include multiple target pathology slide digital images with different resolutions, and the original feature may include recognition results obtained for multiple target pathology slide digital images with different resolutions. In some examples, the original feature can be embodied by the recognition image. In addition, the cervical cells of low-grade squamous intraepithelial lesions can be framed by a rectangular recognition frame. In some examples, the original feature can be embodied as multiple recognition frames (not shown) on the target pathology slide digital image for framing diseased cells. For example, diseased cells of different lesion types of the target pathology slide digital image can be framed, and diseased cells of different lesion types can be framed using recognition frames of different colors.

[0039] In some examples, the original features may also include confidence, which is used to determine the accuracy of the diseased cell type. In some examples, the confidence may range from 0 to 1. In some examples, when a classification result is obtained using a random forest, the confidence may be the ratio of the number of classification results to the number of decision trees.

[0040] In some examples, the combination module 120 can be configured to receive the original features to obtain multiple combination sets. In some examples, each combination set can include at least one of each of the multiple original sub-features. In some examples, each combination set can be named as a first combination set, a second combination set, a third combination set, etc.

[0041] In some examples, the original sub-features included in each combination set may be the same. For example, the first combination set may be a diseased cell type and the second combination set may also be a diseased cell type. In some examples, the original sub-features included in the combination set may be different from each other, for example, the first combination set may include a diseased cell type and the second combination set may include a diseased cell type and a confidence level. In this case, the original features containing a variety of data information can be recombined into combination sets for different original sub-features. Since different medical institutions have different reagents, methods, and operation levels when processing pathological slide digital images, these differences will affect the original sub-features to varying degrees. Splitting and recombining the original features can better perform subsequent recognition of the target pathological slide digital image.

[0042] In some examples, the original sub-features of the combination set may come from different medical institutions. For example, the first combination set may be the lesion cell type and the corresponding confidence obtained after identification of the target pathology slide digital image from the same medical institution, and the second combination set may be the lesion cell type and the corresponding confidence obtained after identification of the target pathology slide digital image from another medical institution. In some examples, the combination set may be any combination of multiple original sub-features. For example, the first combination set may be the lesion cell type, the second combination set may be the lesion cell type and the confidence, and the third combination set may be the lesion cell type and the number of lesion cells.

[0043] In some examples, multiple preliminary recognition modules 130 can receive multiple combination sets from the combination module 120 and perform feature extraction and recognition. In some examples, one preliminary recognition module 130 can receive one combination set from the combination module 120 and perform feature extraction and recognition. In some examples, the number of preliminary recognition modules 130 can be the same as the number of combination sets. In this case, one preliminary recognition module 130 can perform feature extraction and recognition for one combination set.

[0044] In some examples, the preliminary recognition module 130 may include a feature extraction unit and a recognition unit. In this case, feature extraction can be performed on the target pathology slide digital image to obtain data features, and recognition can be performed based on the data features to obtain a preliminary recognition result. In some examples, the feature extraction unit can be configured to perform feature extraction on a combination set of the pathology slide digital image and the target pathology slide digital image to extract data features. In this case, data features of the combination set of the target pathology slide digital image can be extracted.

[0045] In some examples, the target pathology slide digital image may be a pathology slide digital image of maximum resolution. In this case, data features of a combination set of target pathology slide digital images at maximum resolution can be obtained. In some examples, different target pathology slide digital images may have different resolutions.

[0046] In some examples, the data features may be features of the original sub-features in the combination set. In some examples, the data features may include image features and statistical features. In some examples, the data features may also be features of other types. In this case, different types of data features can be extracted according to different combination sets.

[0047] In some examples, multiple preliminary recognition modules 130 may include feature extraction units using different feature extraction methods. In this case, multiple preliminary recognition modules can use different feature extraction models and classification models to perform feature extraction and recognition on the target pathology slide digital image to obtain preliminary recognition results. For example, the feature extraction unit can be a feature extractor based on a histogram of oriented gradients, a local binary pattern, a neural network, or a convolutional neural network. In some examples, the feature extractor for extracting image features can be used to process the original sub-features that can be represented by the image in the combination set and extract the image features of the original sub-features. In some examples, the feature extraction unit can also be a feature extractor based on other models. In this case, different feature extractors have different feature extraction effects, thereby being able to extract different types of features in the target pathology slide digital image, thereby improving the accuracy of recognition.

[0048] In some examples, when a combination set includes the type of diseased cells, the number of diseased cells, and the confidence level, the feature extraction unit can extract the number of diseased cells of high-grade squamous intraepithelial lesion cells, the number of diseased cells in high-grade squamous intraepithelial lesion cells with a confidence level of 0 to 0.5, and the number of diseased cells in high-grade squamous intraepithelial lesion cells with a confidence level of 0.5 to 1. In this case, the relative relationship between the confidence level and the number of high-grade squamous intraepithelial lesion cells can be obtained, and the correlation between the low-confidence recognition result and the type of diseased cells can be obtained.

[0049] In some examples, the feature extraction unit can also be used to count the proportion of various diseased cells. In some examples, the feature extraction unit can also be used for statistics. In some examples, the type of feature extraction unit can be selected according to the original sub-features contained in the combination set. When the combination set includes the diseased cell type, the diseased cell shape or the heat map, a feature extractor for extracting image features can be selected.

[0050] In some examples, multiple preliminary recognition modules 130 may include feature extraction units using the same feature extraction method. In some examples, the network structures of the multiple feature extraction units may be the same. In some examples, multiple feature extraction units with the same network structure may have different network parameters.

[0051] In some examples, the recognition unit can be configured to receive data features and identify based on data features to obtain preliminary recognition results. In some examples, the recognition unit of preliminary recognition module 130 can use different recognition methods. For example, the recognition method can be based on classifiers such as random forest (randomforest, RF), support vector machine (Support VectorMachine, SVM), decision tree (tree), gradient descent tree (Gradient Boosting Decision Tree, GBDT) or extreme gradient descent tree (eXtreme Gradient Boosting, XGBoost). In this case, different classifiers have different classification effects, which can improve the accuracy of recognition.

[0052] In some examples, the number of preliminary recognition modules 130 is not greater than the first preset number. In some examples, the first preset number is 2 to 10. The first preset number may also be any natural number, and the first preset number may be set based on the calculation speed of the recognition model 10. In some examples, the calculation speed of the recognition model 10 may be positively correlated with the first preset number. In this case, the overfitting phenomenon caused by the excessive number of preliminary recognition modules 130 can be avoided, and the number of preliminary recognition modules 130 can be limited based on the actual calculation speed of the recognition model 10.

[0053] In some examples, the preliminary recognition module 130 can be replaced at any time. In this case, a suitable preliminary recognition module 130 can be selected based on the actual situation. In some examples, multiple preliminary recognition modules 130 are screened based on the feature importance (described later) of the preliminary recognition results obtained by each preliminary recognition module 130. In this case, a preliminary recognition result with higher feature importance can be obtained.

[0054] In some examples, the preliminary recognition results may be obtained by processing different combination sets using different feature extraction methods and classification methods. In some examples, the preliminary recognition results may be obtained by processing different combination sets using the same feature extraction method and classification method. In some examples, the preliminary recognition results may be obtained by processing the same combination set using different feature extraction methods and classification methods. Different combination sets may be combination sets with different sources (medical institutions) or combination sets whose original sub-features are not completely the same. Under such conditions, different data features can be extracted in a variety of ways, and preliminary recognition results for the features can be obtained.

[0055] In some examples, the preliminary recognition result may include whether the target pathology slide digital image has one or more of multiple lesion types. In some examples, the preliminary recognition result may include whether the target pathology slide digital image has a certain lesion type. In some examples, the preliminary recognition result may include whether the target pathology slide digital image has multiple lesion types at the same time. In this case, the type of preliminary recognition result can be selected according to actual needs.

[0056] In some examples, the lesion type of the target pathology slide digital image may include squamous cell carcinoma, adenocarcinoma, or adenosquamous carcinoma. In some examples, the lesion type of the target pathology slide digital image may include invasive squamous cell carcinoma, endophytic squamous cell carcinoma, ulcerative squamous cell carcinoma, common adenocarcinoma, and mucinous adenocarcinoma. For example, the preliminary recognition result may be for whether squamous cell carcinoma exists in the target pathology slide digital image or for whether squamous cell carcinoma and adenocarcinoma exist simultaneously in the target pathology slide digital image.

[0057] In some examples, the preliminary recognition result may include part of the original sub-features. For example, the preliminary recognition result may include one or more of the lesion type, shape, quantity, quantity ratio, position, and confidence of the lesion cell. In some examples, the preliminary recognition result may also include one or more of the identification box for selecting the lesion cell and the size, position, aspect ratio, and other information of the identification box.

[0058] In some examples, the preliminary recognition result obtained by one preliminary recognition module 130 can be used as the input of another preliminary recognition module 130. That is, after the combination set is input into the preliminary recognition module 130 and the preliminary recognition result is obtained, the preliminary recognition result can be used to form a new combination set and input into another preliminary recognition module 130 to obtain the preliminary recognition result. For example, the original features include cell information such as the lesion type, cell position and confidence of the lesion cells in the digital image of the target pathological slide. The lesion type, cell position and confidence of the lesion cells can be used as the first combination set and input into a preliminary recognition module 130. The preliminary recognition module 130 is used to obtain the first combination set to sort the lesion cells according to the size of the confidence from large to small, and obtain the cell positions of the first 50 lesion cells and the images of the lesion cells as the preliminary recognition result. The image of the lesion cell is used as the second combination set and input into another preliminary recognition module 130, and the image is re-recognized to obtain the lesion type matching the image. The lesion type obtained this time can be used as the preliminary recognition result or as a new combination set to continue to be input into other preliminary recognition modules 130.

[0059] In some examples, the result summary module 140 can be configured to obtain a terminal recognition result based on multiple preliminary recognition results. In some examples, the result summary module 140 can be a classifier based on a neural network. In some examples, the result summary module 140 can be a classifier based on a random forest (RF), a gradient descent tree (GradientBoosting Decision Tree, GBDT), or an extreme gradient descent tree (eXtreme Gradient Boosting, XGBoost). In this case, the classifier can be used to classify the preliminary recognition results and obtain the terminal recognition results, and the feature importance of each preliminary recognition result can be calculated when training the result summary module 140.

[0060] In some examples, the presentation form of the terminal recognition result may be the same as the primary recognition result. Specifically, the terminal recognition result may be for one or more of the multiple lesion types for the target pathology slide digital image. In some examples, the terminal recognition result may be for whether the target pathology slide digital image has a certain lesion type. In some examples, the terminal recognition result may be for whether the target pathology slide digital image has multiple lesion types at the same time. In some examples, the lesion type of the target pathology slide digital image may include squamous cell carcinoma, adenocarcinoma, or adenosquamous carcinoma. In some examples, the lesion type of the target pathology slide digital image may further include invasive squamous cell carcinoma, endogenous squamous cell carcinoma, ulcerative squamous cell carcinoma, common adenocarcinoma, and mucinous adenocarcinoma. For example, the terminal recognition result may be for whether the target pathology slide digital image has squamous cell carcinoma or for whether the target pathology slide digital image has squamous cell carcinoma and adenocarcinoma at the same time. In some examples, the terminal recognition result may include the degree of lesion of a certain lesion type. In some examples, the terminal recognition result may include the grade of cervical intraepithelial neoplasia. In some examples, the terminal recognition result may include part of the original sub-features. For example, the terminal recognition result may include one or more of the lesion type, shape, quantity, quantity ratio, location, and confidence of the lesion cell. In some examples, the terminal recognition result may also include one or more of the identification box for selecting the lesion cell and the size, location, aspect ratio, and other information of the identification box.

[0061] Figure 5 1 is a flow chart showing a recognition method involved in an example of the present disclosure. The present disclosure provides a recognition method for a pathology slide digital image, which uses the recognition model 10 for pathology slide digital images described above for recognition. In this case, the recognition model 10 can be used to recognize the pathology slide digital image. In some examples, such as Figure 5As shown, the recognition method may include: acquiring original features of a target pathology slide digital image in multiple groups of pathology slide digital images (step S110), combining the original sub-features and obtaining a combination set (step S120), performing feature extraction on the combination set to extract data features and performing recognition based on the data features to obtain multiple preliminary recognition results (step S130), and obtaining a terminal recognition result using the multiple preliminary recognition results (step S140).

[0062] In step S110, the original features of the target pathology slide digital images in the multiple groups of pathology slide digital images can be obtained. In some examples, the pathology slide digital images can be full-view digital slide images. In this case, the full-view digital slide images can be identified. In some examples, the pathology slide digital images can include target pathology slide digital images of multiple resolutions. In some examples, the pathology slide digital images can be pathology slide digital images from a target hospital (i.e., a second target source), and the original features can include multiple original sub-features.

[0063] In step S120, the original sub-features may be combined to obtain a combination set. In some examples, the combination set may be any combination of multiple original sub-features. In some examples, the original sub-features may be combined to obtain multiple combination sets. In some examples, each combination set in the multiple combination sets may include at least one original sub-feature. In some examples, the contents of some combination sets may be the same. In some examples, each combination set in the multiple combination sets may be different.

[0064] In step S130, feature extraction can be performed on the combination set to extract data features and identification can be performed based on the data features to obtain multiple preliminary identification results. The data features can include image features and data features. In some examples, the data features can also be other types of features. In some examples, a feature extraction unit can be used to extract data features. In some examples, the types of feature extraction units and identification units can be selected according to the content of the combination set. In some examples, the preliminary identification result can be an identification result of whether a certain type of lesion exists in the target pathology slide digital image.

[0065] In step S140, a terminal recognition result may be obtained using a plurality of preliminary recognition results. In some examples, the terminal recognition result may be obtained using a classification model such as random forest, GBDT, NN, SVM, logistic regression, Bayesian network, or XGBoost.

[0066] Figure 61 is a flowchart showing a model growth method of the recognition model 10 involved in the examples of the present disclosure. In some examples, the model growth method of the recognition model 10 may include: obtaining multiple groups of pathology slide digital images as training pathology slide digital images, annotating the training pathology slide digital images and identifying the training pathology slide digital images to obtain original features (step S210), combining original sub-features in the original features of the training pathology slide digital images to form a combination set (step S220), using the preliminary recognition module 130 to identify the combination set of the training pathology slide digital images to obtain multiple first preliminary recognition results of the training pathology slide digital images (step S230), using the annotation results of the training pathology slide digital images and the first preliminary recognition results to obtain a first loss function to perform a preliminary recognition on the preliminary recognition module 130. Training (step S240), using the trained preliminary recognition module 130 to recognize the training pathology slide digital images from the target hospital (i.e., the second target source) to obtain multiple second preliminary recognition results of the training pathology slide digital images from the target hospital (step S250), using the result summary module 140 to obtain a terminal recognition result based on the multiple second preliminary recognition results (step S260), using the result summary module 140 to obtain a terminal recognition result based on the multiple second preliminary recognition results (step S270), calculating the feature importance of each second preliminary recognition result (step S280), and screening the trained preliminary recognition modules 130 based on the feature importance of each second preliminary recognition result (step S290).

[0067] In some steps S210, multiple groups of pathology slide digital images may be obtained as training pathology slide digital images, the training pathology slide digital images may be annotated, and the pathology slide digital images may be recognized to obtain original features. In some examples, multiple groups of pathology slide digital images from a first target source may be obtained as training pathology slide digital images.

[0068] In some examples, the pathology slide digital images may include training pathology slide digital images of multiple resolutions. In some examples, the pathology slide digital images may have different sources, for example, in some examples, the pathology slide digital images may be pathology slide digital images from different hospitals or research institutions. In some examples, the pathology slide digital images may be pathology slide digital images from different periods of the target hospital. In this case, multiple groups of pathology slide digital images can be obtained, so that the training data can be increased, thereby improving the recognition effect of the recognition model 10.

[0069] In some examples, the training pathology slide digital images may include training pathology slide digital images from other hospitals and training pathology slide digital images from the target hospital. In this case, the training pathology slide digital images from different sources can be used to train each module of the recognition model 10, thereby improving the versatility of the recognition model 10 and the recognition accuracy of the recognition model 10 for the target hospital.

[0070] In some examples, the training pathology slide digital image can be annotated to obtain the annotation result. In some examples, the annotation result can be obtained by the pathologist annotating the training pathology slide digital image. In some examples, the annotation result can include the pathological cell type and the pathological cell position of the pathological cell. In some examples, the annotation result can be embodied by the annotated image. In some examples, the pathological cell can be framed and the pathological type of the pathological cell can be given by using a rectangular recognition frame. In some examples, the shape of the pathological cell can be outlined and the pathological type of the pathological cell can be given. In some examples, the annotation result can include the pathological type and / or pathological degree of the training pathology slide digital image. In some examples, the annotation result can include the pathological type and / or pathological degree of each block image in the training pathology slide digital image. In this case, the annotation result can be used as the true value of the recognition result to train the recognition model 10.

[0071] In some examples, the block image can be obtained by processing the training pathology slide digital image using the block module. In some examples, the block image can also be called a patch image. In some examples, the block module can use a rectangular sliding window to crop the pathology slide digital image to obtain the block image. In some examples, after obtaining a block image, the position of the rectangular sliding window can be changed and cropped to obtain other block images. In this case, the block module can be used to obtain multiple block images of smaller sizes.

[0072] In some examples, the plurality of block images may have overlapping areas, in which case the recognition accuracy of the recognition model 10 can be improved.

[0073] In some steps S220, the original sub-features in the original features of the training pathology slide digital image may be combined to form a combination set. In some examples, the combination set may be any combination of the original sub-features. The method for obtaining the combination set has been described in detail above and will not be repeated here.

[0074] In some steps S230, the combination set may be identified using the preliminary identification module 130 to obtain a plurality of first preliminary identification results of the training pathology slide digital image. This process is the same as the process of obtaining the preliminary identification results and the terminal identification results of the target pathology slide cells using the preliminary identification module 130 and the terminal identification results as described above, and will not be repeated here.

[0075] In some examples, each of the multiple preliminary recognition modules 130 can be trained using multiple groups of pathology slide digital images from different sources. In this case, the recognition model 10 can have good recognition effects on pathology slide digital images from various sources. In some examples, the training pathology slide digital images can have multiple first preliminary recognition results.

[0076] In some examples, different preliminary recognition modules 130 can receive different combination sets of training pathology slide digital images from the same hospital. In some examples, different preliminary recognition modules 130 can receive combination sets of training pathology slide digital images from different hospitals. In some examples, different preliminary recognition modules 130 can receive the same combination set of training pathology slide digital images from the same hospital, but use different feature extraction methods or classification methods, so different types of data features can be extracted. In this case, multiple preliminary recognition modules 130 can be used to extract data features of different medical institutions, different combination sets, or different time periods. Since the importance of different data features in different situations is different, the use of multiple preliminary recognition modules 130 can improve the versatility of the recognition model 10.

[0077] In some steps S240, the first loss function may be obtained by using the annotation results of the training pathology slide digital image and the first preliminary recognition result to train the preliminary recognition module 130. In this case, the recognition accuracy of the preliminary recognition module 130 can be improved.

[0078] In some examples, the first loss function is obtained by calculating the difference between the annotation result and the first preliminary recognition result. In some examples, the difference between the annotation result and the first preliminary recognition result may include a difference in the diseased cell type of the diseased cell. In some examples, the difference between the annotation result and the first preliminary recognition result may include a difference in the location of the diseased cell. In some examples, the difference between the annotation result and the first preliminary recognition result may include a difference in the diseased cell type of the training pathology slide digital image.

[0079] In some steps S250, the trained preliminary recognition module 130 can be used to recognize the training pathology slide digital images from the target hospital to obtain a plurality of second preliminary recognition results of the training pathology slide digital images. In this case, the trained preliminary recognition module 130 can be used to extract features from the training pathology slide digital images from the target hospital to obtain different types of image features.

[0080] In some examples, the result summary module 140 can be trained using multiple groups of pathology slide digital images from the same source as the target pathology slide digital images. In this case, the recognition model 10 can have a good recognition effect for the target hospital. Thus, the recognition model 10 can be applied to the recognition of pathology slide digital images from different target sources.

[0081] In some steps S260, the terminal identification result can be obtained based on the multiple second preliminary identification results by using the result aggregation module 140. In this case, since the multiple second preliminary identification results are not completely the same, the result aggregation module 140 can be used to obtain a unified terminal identification result.

[0082] In some steps S270, the second loss function can be obtained by using the annotation results of the training pathology slide digital images from the target hospital and the terminal recognition results to train the result summary module 140. In this case, the classification effect of the result summary module 140 can be improved, thereby improving the recognition effect of the recognition model 10.

[0083] In some steps S280, the feature importance of each second preliminary recognition result is calculated. In this case, the importance of each second preliminary recognition result to the terminal recognition result can be judged according to the feature importance of each second preliminary recognition result. In some examples, the feature importance can be calculated using the Gini index or the out-of-bag data error rate. In some steps S290, the trained preliminary recognition module 130 can be screened based on the feature importance of each second preliminary recognition result. In some examples, each preliminary recognition module 130 can be sorted from large to small based on the feature importance, and a preset number of preliminary recognition modules 130 with larger feature importance can be selected, while other preliminary recognition modules 130 are eliminated. In this case, the number of preliminary recognition results can be reduced and the versatility of the recognition model 10 can be improved. In some examples, the feature importance of each preliminary recognition result can be calculated using the result summary module 140. In this case, the importance of each preliminary recognition result to the terminal recognition result can be obtained.

[0084] The present disclosure also relates to a sample augmentation system and augmentation method for pathology slide digital images. Before training the recognition model 10 described above, the sample augmentation system or augmentation method can be used to obtain an augmented image and an augmented annotation result matching the augmented image based on the cervical cell slide digital image and the annotation result matching the cervical cell slide digital image, and the augmented image and the augmented annotation result matching the augmented image are input into the recognition model 10 for training. In this case, a richer training set can be obtained.

[0085] In some examples, the sample augmentation system involved in the present disclosure can be used to obtain a training data set for training the recognition model 10 of the pathology slide digital image. In some examples, the sample augmentation system or augmentation method of this article can augment the pathology slide digital image (cervical cell slide digital image) described above, and can also augment the block image described above, which is explained below by taking the pathology slide digital image as an example.

[0086] In some examples, the training data set may include a digital image of a pathology slide and a labeling result that matches the digital image of the pathology slide. In some examples, the training data set may include an augmented image obtained by a sample augmentation system and an augmented labeling result that matches the augmented image. In this case, a high-quality training data set can be generated, thereby improving the effect of the recognition model 10.

[0087] In some examples, the sample augmentation system may include an acquisition module, a plurality of augmentation modules, and an allocation module. In some examples, the acquisition module may be configured to acquire a pathology slide digital image and a labeling result matching the pathology slide digital image. In this case, the sample augmentation system can augment or store the pathology slide digital image and the labeling result matching the pathology slide digital image. In some examples, each augmentation module may be configured to process the pathology slide digital image and the labeling result based on different sample augmentation methods to obtain an augmented image and an augmented labeling result matching the augmented image. In this case, different sample augmentation methods can be used to process the pathology slide digital image to obtain an augmented image with different features. In some examples, the allocation module may be configured to select at least one augmentation module for processing the pathology slide digital image according to a predetermined setting. In this case, the pathology slide digital image can be augmented multiple times based on actual needs, and an augmented image that meets the actual needs can be obtained. In some examples, the sample augmentation system may also include a storage module. In this case, the pathology slide digital image and the annotation result matching the pathology slide digital image can be stored, and the required pathology slide digital image and the annotation result can be read when needed.

[0088] In some examples, the sample augmentation method may include inversion, rotation, brightness change, contrast change, and scaling, etc. In some examples, the sample augmentation method of the augmentation module may also include noise augmentation, such as adding random salt and pepper noise and / or random RGB noise.

[0089] In some examples, the sample augmentation method of the augmentation module may further include pixel perturbation. Specifically, the mask image and the pathology slide digital image may be fused to obtain the augmented image. In some examples, the fusion method may be to superimpose the corresponding pixels in the mask image and the pathology slide digital image. In this case, noise can be conveniently added to the pathology slide digital image. In some examples, the size of the mask image may be the same as the size of the pathology slide digital image. In some examples, the number of color channels of the mask image may be the same as the number of color channels of the pathology slide digital image. In this case, different color channels may be processed separately. In some examples, the pixel values ​​of each pixel point in the mask image may be randomly distributed within a certain range. In some examples, the distribution range of the pixel values ​​may be random. In some examples, the distribution range of the pixel values ​​may be modified according to actual conditions. In this case, the pixel value range of each pixel point in the mask image may be selected according to actual conditions. In some examples, each pixel point of the augmented image after the mask image and the pathology slide digital image are fused may be corrected so that the pixel values ​​of each pixel point of the augmented image are within a preset range.

[0090] In some examples, the augmentation module may include an image processing unit and a labeling processing unit. In some examples, the image processing unit may be configured to augment the pathology slide digital image based on the sample augmentation method and obtain the augmented image. In some examples, the labeling processing unit is configured to obtain an augmented labeling result matching the augmented image based on the sample augmentation method. In this case, the labeling processing unit can be used to obtain an augmented labeling result matching the augmented image, thereby reducing the workload of re-labeling the augmented image.

[0091] In some examples, the annotation result may include an annotation box that selects the detection target in the pathology slide digital image. In some examples, the detection target may be a pathological cell in the pathology slide digital image. In some examples, the augmented annotation result of the augmented image after augmentation may be the same as the annotation result of the original pathology slide digital image. In some examples, the annotation processing unit may process the annotation result based on the sample augmentation method to obtain the augmented annotation result.

[0092] In some examples, the augmented image and augmented annotation results generated by one augmentation module can be used as input to another augmentation module, and a new augmented image and augmented annotation results can be output. In this case, the sample augmentation method of the digital image of the pathological slide can be more abundant.

[0093] Figure 7 is an example diagram showing an augmented image obtained by an augmentation module based on Gaussian horizontal and vertical stripes according to an example of the present disclosure. In some examples, among the multiple augmentation modules, an augmentation module based on Gaussian horizontal and vertical stripes is included. In some examples, the augmentation module based on Gaussian horizontal and vertical stripes may include a stripe image generation unit and a calculation unit. In some examples, the stripe image generation unit may be configured to generate a stripe image that conforms to a Gaussian distribution. In some examples, the stripe image may include a plurality of stripes. In some examples, see Figure 7 The grayscale values ​​of multiple stripes in the stripe image can satisfy Gaussian distribution. In some examples, the stripe image may have stripes with higher grayscale values, and the farther the stripes are from the stripes, the lower the grayscale value. In some examples, the stripe image may have stripes with lower grayscale values, and the farther the stripes are from the stripes, the higher the grayscale value. Since the digital image of the pathological slide is generally obtained by strip scanning, it may happen that each strip area has a color gamut shift. In this case, the stripe image that conforms to Gaussian distribution can simulate the color gamut shift of the digital image of the pathological slide during the scanning process.

[0094] In some examples, the widths of the stripes in the stripe image may be the same. In this case, it is possible to simulate a situation where the scanning distance of the pathology slide scanner remains unchanged during the scanning process of the pathology slide digital image. In some examples, the widths of the stripes in the stripe image may be different. In this case, it is possible to simulate a situation where the scanning distance of the pathology slide scanner changes during the scanning process. In some examples, the stripe image generation unit may have adjustable parameters, and the adjustable parameters may include an expected value and a standard deviation. In this case, the distribution state of the grayscale in the stripe image can be changed by adjusting the parameters.

[0095] In some examples, the stripe image may be a Gaussian grayscale image in the form of horizontal stripes, oblique stripes, or vertical stripes. In this case, it is possible to simulate the color gamut shift caused by different scanning directions. In some examples, the computing unit may be configured to generate an augmented image based on the stripe image and the pathology slide digital image. In some examples, the pathology slide digital image and the augmented image may be represented by RGB channels. In some examples, the relationship between the augmented image, the stripe image, and the pathology slide digital image may be represented by the following formula: Y=X+b*Ma. Among them, Y represents the matrix formed by the pixel values ​​of each pixel point in the augmented image, X represents the matrix formed by the pixel values ​​of each pixel point in the pathology slide digital image, M represents the matrix formed by the pixel values ​​of each pixel point in the stripe image, a is a preset parameter, and b is a preset parameter.

[0096] In some examples, the augmentation module based on Gaussian horizontal and vertical stripes may include a correction unit, which may be configured to receive the augmented image generated by the calculation unit and correct the pixels of the augmented image to obtain an augmented image with pixel values ​​within a preset range. In this case, the pixels of the augmented image can be corrected to obtain an augmented image with pixel values ​​within a preset range. In some examples, the preset range may be 0 to 255. In some examples, the correction unit may proportionally adjust the pixel values ​​of the augmented image to limit the pixel values ​​within a preset range. In some examples, the correction unit may limit the pixel values ​​of the augmented image within a preset range using a normalization method or a standardization method. In some examples, the correction unit may geometrically scale the values ​​in the matrix according to the difference between the maximum and minimum values ​​in the matrix formed by the pixel values ​​of the augmented image. In this case, each value in the matrix of the augmented image can be within a reasonable range. In some examples, the grayscale of the augmented image obtained by the augmentation module based on Gaussian horizontal and vertical stripes may be related to the Gaussian distribution.

[0097] Figure 8 It is an example diagram showing the working principle of the augmentation module based on image mosaic involved in the example of the present disclosure. Fig. 9 is an example diagram showing an augmented image obtained by an augmentation module based on image mosaicking involved in the examples of the present disclosure.

[0098] In some examples, the plurality of augmentation modules may include an augmentation module based on image mosaicking. In some examples, the augmentation module based on image mosaicking includes an image screening unit, a mosaicking unit, and a cropping unit. In this case, the augmentation module based on image mosaicking can be used to process the pathology slide digital image to obtain an augmented image. In some examples, the augmentation module based on image mosaicking can receive multiple pathology slide digital images. In this case, the received multiple pathology slide digital images are spliced ​​or mosaicked.

[0099] In some examples, the image screening unit may be configured to screen a first preset number of pathology slide digital images as target images 31, and screen a second preset number of pathology slide digital images as background images 32. In some examples, the target images 31 may be pathology slide digital images including the detection target. In some examples, the background images 32 may be random pathology slide digital images. In some examples, the background images 32 may be pathology slide digital images that do not include the detection target (see Figure 8 ). In this case, different digital images of pathology slides can be obtained and mosaicked or spliced ​​with subsequent images. In some examples, the first preset number and the second preset number can be 1 to 10, for example, the first preset number and the second preset number can be 2, 3, 4 or 5, etc. In some examples, the sum of the first preset number and the second preset number can be less than the third preset number. In some examples, the third preset number can be 2, 3, 4, 5, 6, 7, 8, 9 or 10, etc. In some examples, the resolution of the target image 31 and the background image 32 can be 1024*1024, 512*512 or 2048*2048, etc. In some examples, the resolution of the target image 31 and the background image 32 can be different. In some examples, the resolution of multiple target images 31 can be different. In some examples, the mosaicking unit can be configured to receive the target image 31 and the background image 32 and perform mosaicking and / or mosaicking processing to obtain a mosaicked image. In this case, a mosaicked image can be obtained by mosaicking and / or mosaicking. In some examples, see Figure 8 , two target images 31 with a resolution of 1024*1024 and two background images 32 with a resolution of 1024*1024 can be spliced. In this case, a mosaic image of 2048*2048 can be obtained. In some examples, the target image 31 can be mosaicked on the background image 32, for example, the target image 31 is superimposed on the background image 32, and a mosaic image with irregular edges is obtained.

[0100] In some examples, the cropping unit can be configured to receive the mosaic image and crop it to obtain an augmented image. In this case, the mosaic image can be cropped into an augmented image with a resolution equal to or similar to that of the digital image of the pathology slide. In some examples, the cropping area A (see Fig. 9 ) can be a rectangle of 1024*1024. In this case, an augmented image with the same resolution as the pathology slide digital image can be obtained. In some examples, the image mosaic-based augmentation module can also include an edge filling unit. The edge filling unit can be configured to fill the surroundings of the mosaic image to expand the background of the mosaic image. In this case, the invalid area of ​​the augmented image obtained by the cropping unit through cropping can be reduced.

[0101] In some examples, the edge filling unit can use part of the background image 32 to fill the four sides of the mosaic image. In some examples, the size of the filling image 33 can be smaller than the size of the mosaic image. In some examples, the filling image 33 can be randomly filled around the mosaic image. In some examples, the filling image 33 can be spliced ​​with the four sides of the mosaic image to expand the size of the mosaic image. In some examples, the filling image 33 can be spliced ​​with the four sides of the mosaic image to expand the size of the mosaic image. In some examples, the mosaic image can be overlaid on multiple filling images 33. In this case, the filling image 33 can be used to expand the size of the mosaic image.

[0102] In some examples, the size of the filling image 33 may be related to the cropping area A of the cropping unit, for example, the area of ​​the filling image 33 may not be less than the area of ​​the cropping area A. For example, when the filling image 33 and the cropping area A are both square, the side length of the filling image 33 may not be less than the side length of the cropping area A. In this case, the invalid area of ​​the augmented image obtained by the cropping unit through cropping can be effectively reduced.

[0103] In some examples, the augmentation module based on image mosaicking may further include a free transformation unit, which may be configured to randomly rotate and / or flip the mosaicked image. In this case, the diversity of the augmented image can be improved. In some examples, the annotation processing unit of the augmentation module based on image mosaicking may calculate the annotation result based on the rotation, flipping and cropping area A. In some examples, the augmented image obtained by the augmentation module based on image mosaicking may be obtained by splicing and / or mosaicking and cropping multiple target images 31 and background images 32.

[0104] In some examples, the sample augmentation system may include an allocation module. In some examples, the allocation module may be configured to select at least one augmentation module for processing the digital image of the pathology slide according to a predetermined setting. For example, the allocation module may select an augmentation module based on image mosaics to augment the digital image of the pathology slide to obtain an augmented image. The allocation module may also select an augmentation module based on image mosaics and an augmentation module based on light and darkness changes to augment the digital image of the pathology slide to obtain an augmented image. In this case, the augmentation module can be selected according to actual needs. In some examples, the predetermined setting may set the number and type of augmented images generated by the sample augmentation system.

[0105] In some examples, the allocation module can use multiple augmentation modules to process a specified pathology slide digital image to obtain multiple augmented images. In some examples, the allocation module can use one augmentation module to process a specified pathology slide digital image to obtain multiple augmented images. In some examples, the allocation module can arrange the order in which each augmentation module processes the pathology slide digital image based on a predetermined setting. For example, it can be set to use an augmentation module based on inversion, an augmentation module based on rotation, and an augmentation module based on light and darkness changes in sequence to augment the pathology slide digital image. In this case, augmented images with different characteristics can be obtained through different augmentation modules or different processing orders.

[0106] The present disclosure also provides a sample augmentation method for a digital image of a pathological glass slide. In some examples, the sample augmentation method can process the digital image of a pathological glass slide through the sample augmentation system described above to obtain an augmented image and an augmented annotation result matching the augmented image. For details, see the relevant description of the sample augmentation system.

[0107] Although the present disclosure is specifically described above in conjunction with the accompanying drawings and embodiments, it is to be understood that the above description does not limit the present disclosure in any form. Those skilled in the art may modify and change the present disclosure as needed without departing from the essential spirit and scope of the present disclosure, and these modifications and changes all fall within the protection scope of the present disclosure.

Claims

1. A model growth method for a recognition model of a cervical cell slide digital image, characterized in that: include: Recognize the target pathology slide digital image from the first target source to obtain a plurality of original sub-features, wherein the original sub-features include the type and number of diseased cells in the target pathology slide digital image, Based on the plurality of original sub-features, a plurality of combination sets are obtained, each combination set in the plurality of combination sets includes at least one original sub-feature, Using a preliminary recognition module to extract features from the combination set to obtain data features, and obtaining a preliminary recognition result based on the data features, Using a result summary module to obtain a terminal recognition result that matches the target pathology slide digital image based on the multiple preliminary recognition results, The preliminary recognition module is trained by digital images of cervical cell slides from a second target source, The result aggregation module is trained by using digital images of cervical cell slides from the first target source, The preliminary recognition result includes whether the target pathological slide digital image has one or more of a plurality of lesion types, and the terminal recognition result is expressed in the same form as the preliminary recognition result.

2. The model growth method according to claim 1, characterized in that: The number of the preliminary identification modules is not greater than a first preset number.

3. The model growth method according to claim 1, characterized in that: The number of the preliminary recognition modules is the same as the number of the combination sets.

4. The model growth method according to claim 1, characterized in that: The preliminary recognition module performs screening based on the feature importance of the preliminary recognition results obtained by each preliminary recognition module.

5. The model growth method according to claim 1, characterized in that: The data features include image features and statistical features.

6. The model growth method according to claim 1, characterized in that: The preliminary recognition module obtains a plurality of preliminary recognition results through different feature extraction methods and classification methods.

7. The model growth method according to claim 1, characterized in that: The preliminary recognition result includes whether a certain lesion type exists in the target pathological slide digital image.

8. The model growth method according to claim 1, characterized in that The cervical cell slide digital image is a full-field digital slide image.

9. The model growth method according to claim 1, characterized in that: Before training, an augmented image and an augmented annotation result matching the augmented image are obtained based on the cervical cell slide digital image and the annotation result matching the cervical cell slide digital image, and the augmented image and the augmented annotation result matching the augmented image are input into the recognition model for training.

10. A recognition model for a cervical cell slide digital image, wherein the cervical cell slide digital image comprises a plurality of target pathology slide digital images with different resolutions, characterized in that: The recognition model includes: an acquisition module, a combination module, multiple preliminary recognition modules and a result summary module. The acquisition module is configured to identify the target pathology slide digital image from the first target source to obtain a plurality of original sub-features, wherein the original sub-features include the type and number of diseased cells in the target pathology slide digital image; The combination module is configured to obtain a plurality of combination sets based on the plurality of original sub-features, each combination set in the plurality of combination sets including at least one original sub-feature; The preliminary recognition module includes a feature extraction unit and a recognition unit, wherein the feature extraction unit is configured to perform feature extraction on the combination set to obtain data features, and the recognition unit is configured to obtain a preliminary recognition result based on the data features; The result aggregation module is configured to obtain a terminal recognition result matching the target pathology slide digital image based on the plurality of preliminary recognition results; The multiple preliminary recognition modules are trained by digital images of cervical cell slides from a second target source. The result aggregation module is trained by using digital images of cervical cell slides from a first target source, The preliminary recognition result includes whether the target pathological slide digital image has one or more of a plurality of lesion types, and the terminal recognition result is expressed in the same form as the preliminary recognition result.

Images