A label allocation method and device based on cervical cancer cell target detection

Abstract

The present invention proposes a label assignment method and device based on cervical cancer cell target detection, which relates to the technical field of cervical cancer cell detection. The label assignment method reasonably assigns labels based on discriminant scores, so that the model assigns labels to samples based on whether the feature points have discriminative power rather than just by distance, which solves the problem of inaccurate label assignment in cervical cancer cell recognition based on target detection algorithms, enables the model to reasonably assign labels according to image content, greatly reduces the irrationality of label assignment caused by differences in target morphology of medical images, and thus improves the detection performance of the model.

Description

Technical Field

[0001] The present application relates to the technical field of cervical cancer cell detection, and in particular to a label allocation method and device based on cervical cancer cell target detection. Background Art

[0002] Cervical cancer is one of the most common cancers causing cancer deaths in women. Statistics show that most cervical cancers occur in developing countries because these regions lack comprehensive medical services and accurate diagnosis. Early screening of cervical cancer and timely treatment can greatly reduce the mortality rate of cancer patients, so cervical cancer screening is crucial for the prevention of cervical cancer.

[0003] In recent years, the detection performance of deep learning-based object detection algorithms on natural images has been breaking through year by year. The use of computer-assisted cervical cancer cell detection algorithms can effectively reduce a lot of time-consuming and labor-intensive manual screening, greatly improve detection efficiency, and reduce subjective errors in manual analysis.

[0004] The target detection algorithm mainly builds an optimization target to generate a detection model through label assignment. However, when the target detection algorithm is used for cervical cancer cell detection, there will be a certain misalignment in label assignment. The main reason for this misalignment is that the morphology of medical cells is more different from that of targets in natural images, making it impossible for the features extracted based on convolutional neural networks to properly match the labels.

[0005] Previous work mainly used IoU thresholds to divide positive and negative samples. However, due to the large differences in the morphology of cervical cancer cells, the positive and negative samples divided according to the IoU threshold often have a certain degree of misalignment, which increases the difficulty of model optimization.

[0006] The current cervical cancer cell detection methods based on target detection mainly have the problem that the label assignment results and feature maps are misaligned. This is specifically manifested in the following two situations:

[0007] In the first case, refer to Figure 1 , anchor box A2 will be classified as a positive sample, while anchor box A1 will be classified as a negative sample or ignored sample. If we look closely Figure 1 It can be found that the most discriminative point is the center p1 of anchor box A1 rather than the center p2 of anchor box A2, because the information such as the nucleus and the outline of the cell is closer to point p1. As the convolutional neural network deepens, the target semantic information contained in point p1 is richer than that of the green point. Therefore, anchor box A1 should be classified as a positive sample, while anchor box A2 should be classified as a negative sample. The main reason for this distribution result is that the morphology of cervical cancer cells is quite different, and the positive and negative samples divided according to the IoU threshold often have a certain degree of misalignment.

[0008] In the second case, refer to Figure 2 , Figure 2 There are two different types of cells (ASCH and HSIL) in the figure. The dotted anchor box will be assigned as the positive sample set of the ASCH class according to the maximum IoU assignment rule. However, the center point p3 of the dotted anchor box has more discriminative information of the HSIL class. Such misaligned assignment directly turns into an incorrect assignment, which will cause ambiguity in model optimization.

[0009] Therefore, since the IoU-based heuristic assignment algorithm is difficult to adapt to the shape diversity of cervical cancer cells, it is urgent to design an adaptive label assignment algorithm that can perceive the target content to improve the detection performance of the model. Summary of the invention

[0010] The technical problem to be solved by the present application is to propose a label allocation method and device based on cervical cancer cell target detection in view of the above-mentioned deficiencies in the prior art.

[0011] A label assignment method based on cervical cancer cell target detection, comprising:

[0012] Step S101, for each label g, construct a candidate sample set according to the intersection over union (IoU);

[0013] Step S102: In each layer of the feature graph pyramid network FPN, k samples with the largest discriminant scores are selected from the candidate sample set to obtain a sample set T g ; Among them, the calculation formula of the discriminant score Score is:

[0014] Score(C g , g)=-L cls (P g , g)-L bbox (B g , g),

[0015]

[0016] Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch F b The parameter, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively;

[0017] Step S103, calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g ;

[0018] Step S104, according to the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor;

[0019] Step S105: The sample set T g The discriminant score Score of each sample in is compared with the positive sample threshold t one by one, and the samples with discriminant scores Score greater than the positive sample threshold t are added to the positive sample set;

[0020] Step S106, subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

[0021] In some improved technical solutions, step S101 is specifically as follows:

[0022] Set an IoU threshold τ;

[0023] For each label g, calculate the intersection over union (IoU) value of each sample;

[0024] Determine the label g corresponding to the maximum intersection-over-union (IoU) value of each sample;

[0025] The maximum intersection-over-union (IoU) value of each sample is compared with the IoU threshold τ. If the maximum intersection-over-union (IoU) value of the sample is greater than the IoU threshold τ, the sample is included in the candidate sample set of the label g corresponding to the maximum intersection-over-union (IoU) value.

[0026] In some improved technical solutions, the value range of the balance factor λ is between 0.5 and 1.5.

[0027] In some improved technical solutions, the value of the balance factor λ is 1.0.

[0028] On the other hand, the present application also proposes a label allocation device based on cervical cancer cell target detection, comprising:

[0029] The construction module is used to construct a candidate sample set for each label g according to the intersection over union (IoU);

[0030] The selection module is used to select k samples with the largest discriminant scores in each layer of the feature graph pyramid network FPN from the candidate sample set to obtain a sample set T g; Among them, the calculation formula of the discriminant score Score is:

[0031] Score(C g , g)=-L cls (P g , g)-L bbox (B g , g)

[0032]

[0033]

[0034] Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch F b The parameter, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively;

[0035] The first calculation module is used to calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g ;

[0036] The second calculation module is used to calculate the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor;

[0037] The comparison module is used to compare the sample set T g The discriminant score Score of each sample in is compared with the positive sample threshold t one by one, and the samples with discriminant scores Score greater than the positive sample threshold t are added to the positive sample set;

[0038] The production module is used to subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

[0039] In some improved technical solutions, a building block is used to:

[0040] Setting submodule, used to set an IoU threshold τ;

[0041] For each label g, calculate the intersection over union (IoU) value of each sample;

[0042] Determine the label g corresponding to the maximum intersection-over-union (IoU) value of each sample;

[0043] The maximum intersection-over-union (IoU) value of each sample is compared with the IoU threshold τ. If the maximum intersection-over-union (IoU) value of the sample is greater than the IoU threshold τ, the sample is included in the candidate sample set of the label g corresponding to the maximum intersection-over-union (IoU) value.

[0044] In some improved technical solutions, in the second calculation module, the value range of the balance factor λ is between 0.5 and 1.5.

[0045] In some improved technical solutions, in the second calculation module, the value of the balance factor λ is 1.0.

[0046] On the other hand, the present application also proposes an electronic device having a processor and a memory, wherein the memory is used to store computer instructions, and when the computer instructions are executed by the processor, the label allocation method proposed in the above section is implemented.

[0047] On the other hand, the present application also proposes a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the label allocation method proposed in the above section is implemented.

[0048] This application proposes an effective dynamic mean-standard deviation assignment (DMSA) algorithm, which reasonably assigns labels based on discriminant scores, so that the model assigns labels to samples based on whether the feature point has discriminative power rather than just by distance, solving the problem of inaccurate label assignment in cervical cancer cell recognition based on target detection algorithms, enabling the model to reasonably assign labels based on image content, greatly reducing the irrationality of label assignment caused by differences in target morphology in medical images, thereby improving the detection performance of the model. For specific descriptions of the technical effects, please refer to the specific implementation method of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is one of the schematic diagrams of sample label allocation in cervical cancer cell images in the background technology of this application.

[0050] Figure 2 This is the second schematic diagram of sample label allocation in cervical cancer cell images in the background technology of this application.

[0051] Figure 3 This is one of the flow charts of the label assignment method based on cervical cancer cell target detection in the embodiment of the present application.

[0052] Figure 4 This is the second flowchart of the label assignment method based on cervical cancer cell target detection in the embodiment of the present application.

[0053] Figure 5 It is a structural schematic diagram of a label distribution device based on cervical cancer cell target detection in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The following is the specific embodiment of the present application and in conjunction with the accompanying drawings, the technical scheme of the present application is further described, but the present application is not limited to these embodiments. In the following description, specific details such as specific configuration and assembly are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures has been omitted.

[0055] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other.

[0056] refer to Figure 3 The present application proposes a label assignment method based on cervical cancer cell target detection, which reasonably assigns labels based on discriminant scores, so that the model assigns labels to samples based on whether the feature points have discriminative power. The method includes steps S101 to S106. The technical solution of the present application is described in detail below in conjunction with the accompanying drawings.

[0057] Step S101: for each label g, a candidate sample set is constructed according to the intersection over union (IoU).

[0058] Step S102: In each layer of the feature graph pyramid network FPN, k samples with the largest discriminant scores are selected from the candidate sample set to obtain a sample set T g ; Among them, the calculation formula of the discriminant score Score is:

[0059] Score(C g , g)=-L cls (P g , g)-L bbox (B g , g),

[0060]

[0061] Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch Fb The parameter, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively.

[0062] Step S103, calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g .

[0063] Step S104, according to the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor.

[0064] Step S105: The sample set T g The discriminant score Score of each sample in is compared with the positive sample threshold t one by one, and the samples whose discriminant score Score is greater than the positive sample threshold t are added to the positive sample set.

[0065] Step S106, subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

[0066] In step S101, refer to Figure 4 Specifically, step S101 includes:

[0067] Step S101a, setting an IoU threshold τ.

[0068] Step S101b, for each label g, calculate the intersection over union (IoU) value of each sample;

[0069] Step S101c, determining the label g corresponding to the maximum intersection-over-union (IoU) value of each sample;

[0070] Step S101d, compare the maximum intersection over union (IoU) value of each sample with the IoU threshold τ. If the maximum intersection over union (IoU) value of the sample is greater than the IoU threshold τ, the sample is classified into the candidate sample set of the label g corresponding to the maximum intersection over union (IoU) value.

[0071] For each label g, a candidate sample set is constructed according to the maximum intersection over union (IoU) principle. Specifically, an IoU threshold τ is first defined, and the IoU value of each anchor box sample and all labels is calculated. For each anchor box sample, the label with the corresponding maximum IoU value is found. If the value is greater than the given IoU threshold τ, then the anchor box sample is added to the candidate sample set corresponding to the label g. Step S101 is used to perform the first screening of the samples, and the subsequent steps perform the second screening based on the screening results of step S101.

[0072] In step S102, the negative of the sum of the losses of the classification branch and the regression branch is used to represent the discriminability score of the elements in the sample set. The larger the score, the richer the semantic information of the sample. g The candidate sample set obtained in step S101 is used to calculate the discriminant score Score based on the candidate sample set, and remove several samples with smaller discriminant scores Score, and retain the k samples with the largest discriminant scores, thereby obtaining the sample set T g .

[0073] In step S103, firstly, the sample set T is calculated g The discriminant score Score of each sample in the calculation formula of the discriminant score Score is g is the sample set T obtained in step S102 g Next, calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g .

[0074] In step S104, according to the mean m obtained in step S103 g and standard deviation v g , calculate the positive sample threshold t, where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor.

[0075] In step S105, the sample set T g The discriminant score Score of each sample in the positive sample set is compared with the positive sample threshold t calculated in step S104 one by one, and the samples whose discriminant score Score is greater than the positive sample threshold t are added to the positive sample set. Step S105 g The samples in the dataset are further screened and the screening results are used as the positive sample set.

[0076] In step S106, positive samples are removed from all samples, and the remaining samples constitute a negative sample set.

[0077] In the technical solution composed of the above steps, first in step S101, for each label g, a candidate sample set is constructed according to the intersection over union ratio IoU, and the samples are screened for the first time. Then, in step S102, several samples with smaller discriminant scores Score are eliminated, and k samples with the largest discriminant scores are retained, thereby obtaining a sample set T g , based on the first screening, the second screening is performed using the discrimination score Score as the basis. In step S103 to step S105, in the sample set T g Based on this, the positive sample threshold t generated by the discriminant score Score is used to g Conduct a third screening.

[0078] Based on the above steps, a new label assignment algorithm based on dynamic mean-standard deviation (Dynamic Mean-Std Assignment, DMSA algorithm) is generated.

[0079]

[0080] In the above algorithm, step S101 corresponds to rows 2-3, step S102 corresponds to rows 4-9, step S103 corresponds to rows 10-12, step S104 corresponds to row 13, step S105 corresponds to rows 14-15, and step S106 corresponds to row 17.

[0081] It should be noted that in the prior art, when the target detection algorithm is used for cervical cancer cell detection, there will be a certain misalignment in label allocation. The main reason for this misalignment is that the morphology of medical cells is more different from that of targets in natural images, making it impossible for features extracted based on convolutional neural networks to properly match labels. The label allocation method proposed in the embodiment of the present application is used to improve the label allocation misalignment problem existing in cervical cancer cell recognition based on the existing target detection algorithm. The positive and negative samples of the same label are more accurately divided and the labels are more accurately allocated between different types of cells.

[0082] The dynamic mean-standard deviation assignment (DMSA) algorithm proposed in this application solves the problem of inaccurate label assignment in cervical cancer cell recognition based on target detection algorithms, enabling the model to reasonably assign labels according to image content, greatly reducing the irrationality of label assignment caused by differences in target morphology in medical images, thereby improving the detection performance of the model. Compared with the heuristic assignment algorithm based on IoU, the technical solution of this application is more adaptable to the shape variation diversity of cervical cancer cells and effectively improves the detection performance of the model.

[0083] In some implementations, the balance factor λ has a value range of 0.5-1.5. Further, the balance factor λ has a value of 1.0.

[0084] The label allocation algorithm proposed in the above section is verified by setting up a specific experiment to illustrate the advantages brought by the label allocation algorithm.

[0085] 1. Data introduction. We validated the proposed method on the cervical cell detection dataset (CCD) produced by liquid-based cytology technology. The CCD dataset contains 40k images with a resolution of about 4000×3000, which are divided into 10 categories by experienced doctors, namely normal cells (NORMAL), atypical squamous cells-unclear significance (ASCUS), atypical squamous epithelial cells-cannot rule out HSIL (ASCH), low-grade squamous intraepithelial lesions (LSIL), high-grade squamous intraepithelial lesions (HSIL), atypical glandular cells (AGC), vaginal trichomoniasis (VAG), Candida (MON), dysbacteriosis (DYS), and endocervical cells (EC). The CCD dataset is divided into 30,000 images for training and 10,000 images for testing.

[0086] 2. Evaluation indicators and implementation details. We mainly use the mAP (mean Average Precision) of the Pascal VOC challenge as the evaluation indicator, and the default IoU threshold is 0.5. In addition, for more accurate positioning, we also report the mAP under the IoU threshold of 0.75. At the same time, the performance on small and large objects is evaluated separately. In terms of implementation, for a fair comparison, we first enhance the basic residual RetinaNet with the IoU branch, and then build our model on this basis. The residual network ResNet-50 is the default backbone network, and its parameters are initialized from the pre-trained model on the ImageNet dataset. The initial learning rate is 0.02, and the mini-batch size is 16. Focal loss, GloU loss and Cross-entropy loss are used for classification, regression and IoU branches, respectively, and their weights are 1.0, 1.4 and 0.5, respectively.

[0087] 3. Experimental Results

[0088] Results As shown in the table below, our model is mainly based on the improved RetinaNet, replacing the basic Max-IoU label assignment algorithm with our DMSA, and introducing GCB to better align with DMSA. As can be seen from Table 2, the DMSA strategy proposed in this paper can improve the mAP by about 1.5% compared with the Max-IoU assignment, so it is more reasonable to assign labels according to the discriminant score than to assign labels according to IoU. In addition, it can be seen that the use of the global context module GCB can indeed enhance the feature representation by aligning the assigned results. By using these two components together, the baseline model finally improved the mAP by about 2.0%. These results verify the effectiveness of this method in alleviating the problem of deviation between feature and label assignment results.

[0089] Table 2: The contribution of each component on CCD.

[0090]

[0091] We compare our method with the state-of-the-art general object detector and cervical cell detector. To make a fair comparison, we use the same training strategy to train all models. In addition, to further improve the performance of the model without sacrificing the inference time, we also apply a multi-scale training strategy to our method.

[0092] As shown in Table 3, our method achieves a mAP of 52.2%, outperforming all other methods. Even under the stricter IoU=0.75 constraint, our method still performs well. It is worth noting that the method has good performance in small-scale cell detection, with an APS of 25.4%. This is because small-scale objects lack contextual semantic information and are more sensitive to misalignment problems. In addition, through a multi-scale training strategy, the performance of our method can be further improved to 52.8% mAP.

[0093] We also report the inference speed of various models calculated in frames per second (FPS) on a single Nvidia RTX 2080Ti GPU. It can be seen that our method is efficient, achieving 52.2% mAP at 13.9FPS, achieving the best accuracy and efficiency trade-off.

[0094] Table 3: Performance comparison with state-of-the-art detec-tors onCCD. denotes multi-scale training[6].

[0095]

[0096] This paper alleviates the misalignment problem in cervical cell detection from two aspects. First, we propose an effective dynamic mean-standard deviation assignment (DMSA) algorithm, which reasonably assigns labels based on the discriminant scores. Second, we utilize global context information at all locations in the feature map to align the results of label assignment. The effectiveness of this method is verified in cervical cell detection experiments.

[0097] refer to Figure 5 The embodiment of the present application also proposes a label allocation device based on cervical cancer cell target detection, including a construction module 201, a selection module 202, a first calculation module 203, a second calculation module 204, a comparison module 205, and a production module 206. The label allocation device proposed above is specifically described below in conjunction with the accompanying drawings.

[0098] A construction module 201 is used to construct a candidate sample set for each label g according to the intersection over union (IoU);

[0099] The selection module 202 is used to select k samples with the largest discriminant scores in each layer of the feature graph pyramid network FPN from the candidate sample set to obtain a sample set T g ; Among them, the calculation formula of the discriminant score Score is:

[0100] Score(C g , g)=-L cls (P g , g)-L bbox (B g , g),

[0101]

[0102] Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch F b The parameter, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively;

[0103] The first calculation module 203 is used to calculate the sample set T gThe mean value m of the discriminant score Score of the samples g and standard deviation v g ;

[0104] The second calculation module 204 is used to calculate the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor;

[0105] Comparison module 205, used to compare the sample set T g The discriminant score Score of each sample in is compared with the positive sample threshold t one by one, and the samples with discriminant scores Score greater than the positive sample threshold t are added to the positive sample set;

[0106] The production module 206 is used to subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

[0107] In some embodiments, the construction module 201 is used to:

[0108] Setting submodule, used to set an IoU threshold τ;

[0109] For each label g, calculate the intersection over union (IoU) value of each sample;

[0110] Determine the label g corresponding to the maximum intersection-over-union (IoU) value of each sample;

[0111] The maximum intersection-over-union (IoU) value of each sample is compared with the IoU threshold τ. If the maximum intersection-over-union (IoU) value of the sample is greater than the IoU threshold τ, the sample is included in the candidate sample set of the label g corresponding to the maximum intersection-over-union (IoU) value.

[0112] In some implementations, in the second calculation module 204, the balance factor λ has a value range of 0.5-1.5.

[0113] In some implementations, in the second calculation module 204, the value of the balance factor λ is 1.0.

[0114] Since the label allocation device based on cervical cancer cell target detection corresponds to the label allocation proposed in the previous part of the embodiment of the present application, the relevant content can be found in the description of the above-mentioned label allocation method, which will not be repeated here.

[0115] The embodiment of the present application further proposes an electronic device having a processor and a memory, wherein the memory is used to store computer instructions, and when the computer instructions are executed by the processor, the label allocation method proposed in the above part is implemented.

[0116] The embodiment of the present application further proposes a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the label allocation method proposed in the above part is implemented.

[0117] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0118] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0119] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0120] The specific embodiments described herein are merely examples of the spirit of the present application. A person skilled in the art of the present application may make various modifications or additions to the specific embodiments described or replace them in a similar manner, but this will not deviate from the spirit of the present application or exceed the scope defined by the attached claims.

Claims

1. A label assignment method based on cervical cancer cell target detection, characterized in that: include: Step S101, for each label g, construct a candidate sample set according to the intersection over union (IoU); Step S102: In each layer of the feature graph pyramid network FPN, k samples with the largest discriminant scores are selected from the candidate sample set to obtain a sample set T g ; Among them, the calculation formula of the discriminant score Score is: Score(C g ,g)=-L cls (P g ,g)-L bbox (B g ,g), Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch F b Parameters, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively; Step S103, calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g ; Step S104, according to the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor; Step S105: The sample set T g The discriminant score Score of each sample in is compared with the positive sample threshold t one by one, and the samples with discriminant scores Score greater than the positive sample threshold t are added to the positive sample set; Step S106, subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

2. The label assignment method based on cervical cancer cell target detection according to claim 1, characterized in that: Step S101 is specifically as follows: Set an IoU threshold τ; For each label g, calculate the intersection over union (IoU) value of each sample; Determine the label g corresponding to the maximum intersection-over-union (IoU) value of each sample; The maximum intersection-over-union (IoU) value of each sample is compared with the IoU threshold T. If the maximum intersection-over-union (IoU) value of the sample is greater than the IoU threshold T, the sample is included in the candidate sample set of the label g corresponding to the maximum intersection-over-union (IoU) value.

3. The label assignment method based on cervical cancer cell target detection according to claim 1, characterized in that: The value range of the balance factor λ is between 0.5 and 1.

5.

4. The label assignment method based on cervical cancer cell target detection according to claim 3 is characterized in that: The balance factor λ is set to 1.

0.

5. A label distribution device based on cervical cancer cell target detection, characterized in that: include: The construction module is used to construct a candidate sample set for each label g according to the intersection over union (IoU); The selection module is used to select k samples with the largest discriminant scores in each layer of the feature graph pyramid network FPN from the candidate sample set to obtain a sample set T g ; Among them, the calculation formula of the discriminant score Score is: Score(C g ,g)=-L cls (P g ,g)-L bbox (B g ,g), Where x is the input image; σ is the sigmoid function; C g is the candidate sample set; θ c ,θ b It is the classification branch F c and regression branch F b Parameters, P g and B g They are classification branches F c and regression branch F b The output, L cls and L bbox They are the classification branch loss function and the regression branch loss function respectively; The first calculation module is used to calculate the sample set T g The mean value m of the discriminant score Score of the samples g and standard deviation v g ; The second calculation module is used to calculate the mean m g and standard deviation v g Calculate the positive sample threshold t; where the calculation formula for the positive sample threshold t is: t = m g +λv g , λ is the balance factor; The comparison module is used to compare the sample set T g Zhongmei indivual The discriminant score of the sample Scor e indivual Compare with the positive sample threshold t, and add samples whose discriminant score Score is greater than the positive sample threshold t to the positive sample set; The production module is used to subtract all positive samples from the total sample set consisting of all samples to generate a negative sample set.

6. The label distribution device based on cervical cancer cell target detection according to claim 5, characterized in that: Building blocks for: Setting submodule, used to set an IoU threshold τ; For each label g, calculate the intersection over union (IoU) value of each sample; Determine the label g corresponding to the maximum intersection-over-union (IoU) value of each sample; The maximum intersection-over-union (IoU) value of each sample is compared with the IoU threshold τ. If the maximum intersection-over-union (IoU) value of the sample is greater than the IoU threshold τ, the sample is included in the candidate sample set of the label g corresponding to the maximum intersection-over-union (IoU) value.

7. The label distribution device based on cervical cancer cell target detection according to claim 5, characterized in that: In the second calculation module, the value range of the balance factor λ is between 0.5 and 1.

5.

8. The label distribution device based on cervical cancer cell target detection according to claim 5, characterized in that: In the second calculation module, the value of the balance factor λ is 1.

0.

9. An electronic device comprising a processor and a memory, wherein the memory is used to store computer instructions, and when the computer instructions are executed by the processor, the label allocation method according to any one of claims 1 to 4 is implemented.

10. A computer-readable storage medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the label allocation method according to any one of claims 1 to 4 is implemented.

Images