Semi-supervised cervix uteri abnormal cell long tail resampling detection method based on diagnosis result guiding pseudo tag and unified data enhancement
By adopting a semi-supervised method based on diagnostic results-guided pseudo-labeling and unified data enhancement in cervical cancer screening, the complexity of the cervical cytology image reading task and the low utilization efficiency of unlabeled data are solved, and cervical abnormalities are improved The performance of cell detection is suitable for automated screening of cervical cancer.
Patent Information
- Application Number
- CN202410902564.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-07
- Publication Date
- 2024-11-26
AI Technical Summary
In early screening of cervical cancer, the cervical cytology image reading task is labor-intensive, complex and cumbersome. Due to the difficulty in obtaining annotated data, the utilization efficiency of unlabeled images is low, which affects the computer-assisted detection performance.
A semi-supervised long-tail resampling detection method for cervical abnormal cells based on diagnostic results guidance and unified data enhancement is proposed. The pseudo-tagging detection method is proposed through tail category labeling resampling, unified data enhancement and diagnostic results guidance is conducted through teacher-student framework. Generate, improve tail category detection performance and mitigate noise impact.
Effectively utilize a large number of unmarked panoramic images of the cervical abnormality detection, improve the detection performance, and is suitable for automated cervical cancer screening systems.
Smart Images

Figure CN119027365A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to semi-supervised abnormal cervical cell detection. Background Art
[0002] Cervical cancer is one of the most common malignant tumors in women and poses a major threat to women's health worldwide. In 2020, 604,000 women worldwide were diagnosed with cervical cancer and approximately 342,000 died from the disease. However, cervical cancer is preventable because it develops slowly, and effective early screening can significantly reduce the incidence of cervical cancer by at least 60%. The current effective cervical cancer screening method is cervical exfoliated cell smear pathological examination, which requires experienced pathologists to observe the diseased cells under a microscope before making a diagnosis. This process faces some obstacles, such as the shortage of trained pathologists. In addition, reading cervical cytology images is an extremely labor-intensive, complex and tedious task for pathologists.
[0003] In recent years, with the significant progress of image processing and deep learning technology, computer-assisted methods based on deep learning have effectively solved the above challenges. Although it is challenging to obtain annotated data, the collection of unlabeled images is relatively simple. A large number of cervical cytology sections in hospitals can be digitized into unlabeled images without the participation of pathologists. How to make full use of these unlabeled data to improve the performance of computer-assisted. In order to improve this problem, the present invention proposes a semi-supervised cervical abnormal cell long-tail resampling detection method based on pseudo-labels guided by diagnostic results and unified data enhancement. Specifically, the present invention proposes a label resampling method for the tail category under the teacher-student semi-supervised framework, which effectively improves the detection performance of the tail category while ensuring the stability of the head category performance; then, the present invention proposes a unified data enhancement method, which sets strong and weak enhancements with better adaptability according to the characteristics of cervical cell images, and alleviates the noise impact caused by excessive and inappropriate enhancement; finally, the present invention proposes a pseudo-label generation method based on panoramic image diagnosis results. The method uses different pseudo-label generation methods for different categories of lesion-positive samples, which greatly improves the reliability of pseudo-labels. Summary of the invention
[0004] The purpose of the present invention is to solve the problem of the difference between the convenience of obtaining diagnostic results of cervical panoramic images and the time-consuming and difficult instance annotation. How to use a large number of unlabeled cervical panoramic images with diagnostic results to assist model training, and propose a semi-supervised cervical abnormal cell long-tail resampling detection method based on pseudo-labels guided by diagnostic results and unified data enhancement.
[0005] The above invention objectives are mainly achieved through the following technical solutions:
[0006] S1. Prepare a dataset for detecting abnormal cervical cells, which is a dataset for detecting abnormal cervical cells using semi-supervised targets, by cropping labeled and unlabeled panoramic images into tile images;
[0007] First, positive cervical cell pathology panoramic images from different institutions and with different diagnostic results were collected; then, four types of abnormal cells were marked on the cervical cell pathology panoramic images using ASAP software; then, the above cervical cell pathology panoramic images were divided into a training set and a validation set of panoramic images in a ratio of 8:2; finally, the training set and validation set of panoramic images were cut into tile images with a length and width of 1024; as the final training set and validation set, 1%, 2%, 5%, and 10% of the training set were extracted as labeled data, and the rest of the training set was used as unlabeled data;
[0008] S2. Build a semi-supervised end-to-end detection network framework. This framework adds a label resampling mechanism for tail cell categories and a unified data enhancement mechanism based on the teacher-student approach, as well as a pseudo-label generation mechanism guided by diagnostic results.
[0009] S2-1 Marker resampling mechanism of tail cell classes;
[0010] The label resampling mechanism for tail categories is the first module of the framework. The sampler samples from the labeled and unlabeled datasets in a ratio of 1:4, where the tail categories are oversampled from the labeled dataset and enough head categories are mined from the pseudo-labels in S5 to balance the undersampling of the head categories in the labeled data. First, the label frequency of each category c is calculated: f(c). Then, the repetition factor r(c) of category c is determined by the frequency f(c) of the category in the image and the threshold t. The specific formula is as follows:
[0011]
[0012] Among them, t is a hyperparameter given by humans, and its value is 1e-3. n c is the number of samples of category i and N is the total number of samples. Finally, the image repetition factor of each image I is calculated:
[0013] S2-2 Unified data enhancement mechanism;
[0014] Unify the data enhancement mechanism for each unlabeled image; first, weakly enhance the unlabeled image; apply color transformation, select a transformation from color transformation (brightness adjustment, contrast adjustment, saturation adjustment and hue adjustment), and the transformation is based on P color = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M colorRandomly select within the allowed range [0,0.4]. If a transformation is decided, the image is transformed according to the selected magnitude, otherwise the image remains unchanged. Apply a shape transformation (horizontal flip, vertical flip, translation (left and right, up and down), scaling), and randomly select a transformation from the shape transformation category. The transformation is based on P shape = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M shape Randomly select within the allowed range [0,0.4]. If it is decided to apply the transformation, the image is transformed according to the selected magnitude, otherwise the image remains unchanged and the weakly enhanced image is input into the teacher model; then for the strong enhancement of the unlabeled image; apply color transformation. For each unlabeled image, first select two color transformation operations, each color transformation operation has a probability parameter P color = 0.8 to decide whether to apply this operation. If you choose to apply the transformation, then color Randomly select an amplitude value in the amplitude range [0.5, 1.0] for transformation; randomly select and apply shape transformation, and then select two shape transformation operations, each of which also has a probability parameter P shape =0.8, if you choose to apply the transformation, then M shape Randomly select an amplitude value from the allowed amplitude range [0.5, 1.0] for transformation. Since the shape transformation will affect the bounding box in the image, the coordinates of the bounding box need to be adjusted accordingly. Randomly select and apply a hybrid transformation, and finally select a hybrid transformation operation. The hybrid transformation operation also has a probability parameter P. mix =0.8, if you choose to apply the transformation, then M mix A magnitude value is randomly selected from the allowed magnitude range [0.5, 1.0] for transformation. In particular, for the Mixup operation, the parameter range is set to [0, 0.2], indicating that the influence of other images on the current image is relatively low, and is input into the student model for training;
[0015] S2-3 Pseudo-label generation mechanism guided by diagnostic results;
[0016] The pseudo-label results of the teacher model processing the weakly enhanced data are used to supervise the student model training of the strongly enhanced data; different pseudo-label generation methods are applied to the panoramic image diagnosis results of different lesion levels, and the number of ASC-US cells is defined as N ASC-US , define the number of LSIL cells as N LSIL , define the number of ASC-H cells as N ASC-H , define the number of HSIL cells as N HSIL , T low is the lower threshold of abnormal cell number, and the total number of cells N total =N ASC-US +NLSIL +N ASC-H +N HSIL ;
[0017] For panoramic images with ASC-US diagnosis results, the label generation condition 1 is N ASC-US >N ASC-H And N ASC-US >N HSIL And N ASC-US >N LSIL , condition 2 is N total <T low , the judgment formula is as follows:
[0018] Label ASC-US =IF N ASC-US >MAX(N ASC-H ,N HSIL ,N LSIL )AND(N total <T low ) (2)
[0019] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic image diagnosis of ASC-US is given as: Threshold ASC-US =0.4, Threshold LSIL =0.4, Threshold ASC-H =0.7, Threshold HSIL =0.7;
[0020] For panoramic image diagnosis results with LSIL, the label generation condition 1 is N LSIL >N ASC-US And N LSIL >N ASC-H And N LSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows:
[0021] Label LSIL =IF N LSIL >MAX(N ASC-H ,N HSIL ,N ASC-US )AND(N total <T low ) (3)
[0022] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as LSIL is given as: ASC-US =0.5,Threshold LSIL =0.5,ThresholdASC-H =0.7, Threshold HSIL =0.6;
[0023] For panoramic image diagnosis result ASC-H, label generation condition 1 is And N ASC-H >N ASC-US or N ASC-H >N LSIL or N HSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows:
[0024]
[0025] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as ASC-H is given as: Threshold ASC-US =0.6, Threshold LSIL =0.6, Threshold ASC-H =0.5,Threshold HSIL =0.5;
[0026] For panoramic images with HSIL diagnostic results, the label generation condition 1 is N HSIL >N ASC-US or N HSIL >N ASC-H or N HSIL >N LSIL Condition 2 is N total >=T low , the judgment formula is as follows:
[0027]
[0028] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as HSIL is given as: Threshold ASC-US =0.6, Threshold LSIL =0.6, Threshold ASC-H =0.5,Threshold HSIL =0.4.
[0029] Effects of the Invention
[0030] The present invention provides a semi-supervised cervical abnormal cell long-tail resampling detection method based on pseudo-labels guided by diagnostic results and unified data enhancement. The algorithm first prepares cervical cell abnormality detection data, which is divided into labeled and unlabeled semi-supervised detection data according to pathological panoramic image lesions. Then, a semi-supervised end-to-end detection network framework is constructed, which expands the traditional teacher-student architecture, adds a label resampling mechanism and a unified data enhancement mechanism for the tail cell category, and a pseudo-label generation mechanism guided by diagnostic results. In the framework, the detection performance of the tail category is improved by label resampling of the tail cell category, and the ratio of sampled labeled data to unlabeled data is 1:4. Then, these data are uniformly enhanced with strong and weak data to prevent excessive enhancement noise from entering. Finally, the proposal generated by the teacher model is sent to the pseudo-label generation mechanism guided by diagnostic results to generate pseudo-labels for unlabeled data for supervising the training of the student model. The weight of the teacher model is updated by the weight exponential average moving of the student model to improve performance. Experiments show that the present invention can make full use of a large number of unlabeled pathological panoramic images with lesion diagnosis results, improve the effect of semi-supervised cervical abnormality detection, and can be effectively applied to the automated cervical cancer screening system. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 Main flow chart of the algorithm
[0032] Figure 2 Overall framework structure diagram Specific implementation methods Specific implementation method one:
[0035] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] like Figure 1 The semi-supervised cervical abnormal cell long-tail resampling detection method provided in this paper based on pseudo-label guided by diagnosis results and unified data enhancement mainly includes the following steps:
[0037] S1. Prepare a dataset for detecting abnormal cervical cells, which is a dataset for detecting abnormal cervical cells using semi-supervised targets, by cropping labeled and unlabeled panoramic images into tile images;
[0038] S2. Build a semi-supervised end-to-end detection network framework. This framework adds a label resampling mechanism for tail cell categories and a unified data enhancement mechanism based on the teacher-student approach, as well as a pseudo-label generation mechanism guided by diagnostic results.
[0039] S3, the ratio of labeled data to unlabeled data is 1:4 through the label resampling mechanism of the tail cell category of S2-1;
[0040] S4, feeds the data of S3 into the unified data augmentation mechanism of S2-2, where the unlabeled data is fed into the student and teacher models of S2 using strong and weak augmentation respectively;
[0041] S5, the proposals generated by the teacher model of S4 are fed into the pseudo-label generation mechanism guided by the diagnosis results of S2-3 to generate pseudo-labels for unlabeled data to supervise the training of the student model of S4, where the weight of the teacher model is generated by the exponential moving average of the student model;
[0042] S6. Repeat S3-S5 until the preset number of training iterations is reached or the evaluation performance of the teacher-student learning model meets the requirements.
[0043] In the embodiment of the present invention, first, positive cervical cell pathology panoramic images of different institutions and different diagnostic results are collected for abnormal cell annotation, and then the data set is divided according to the panoramic image level and sliced at the same time; then, under the teacher-student semi-supervised learning framework, this study innovatively proposed a label resampling method for the tail cell category, which significantly improved the detection accuracy of the tail category by ensuring the stability of the head category performance; then, a unified data enhancement strategy was studied, and more adaptable strong enhancement and weak enhancement methods were designed specifically for the characteristics of cervical cell images, which effectively alleviated the noise problem that may be caused by inappropriate enhancement; finally, a pseudo-label generation technology guided by diagnostic results was developed, and a personalized pseudo-label generation method was adopted for positive samples of different lesion levels, which greatly improved the reliability of pseudo-labels and model detection performance.
[0044] The embodiments of the present invention are described in detail below:
[0045] like Figure 1 The implementation of the shown algorithm consists of the steps:
[0046] S1. Prepare a dataset for detecting abnormal cervical cells, which is a dataset for detecting abnormal cervical cells using semi-supervised targets, by cropping labeled and unlabeled panoramic images into tile images;
[0047] First, positive cervical cell pathology panoramic images from different institutions and with different diagnostic results were collected; then, four types of abnormal cells were marked on the cervical cell pathology panoramic images using ASAP software; then, the above cervical cell pathology panoramic images were divided into a training set and a validation set of panoramic images in a ratio of 8:2; finally, the training set and validation set of panoramic images were cut into tile images with a length and width of 1024; as the final training set and validation set, 1%, 2%, 5%, and 10% of the training set were extracted as labeled data, and the rest of the training set was used as unlabeled data;
[0048] S2. Build a semi-supervised end-to-end detection network framework. This framework adds a label resampling mechanism for tail cell categories and a unified data enhancement mechanism based on the teacher-student approach, as well as a pseudo-label generation mechanism guided by diagnostic results.
[0049] S2-1 Marker resampling mechanism of tail cell classes;
[0050] The label resampling mechanism for tail categories is the first module of the framework. The sampler samples from the labeled and unlabeled datasets in a ratio of 1:4, where the tail categories are oversampled from the labeled dataset and enough head categories are mined from the pseudo-labels in S5 to balance the undersampling of the head categories in the labeled data. First, the label frequency of each category c is calculated: f(c). Then, the repetition factor r(c) of category c is determined by the frequency f(c) of the category in the image and the threshold t. The specific formula is as follows:
[0051]
[0052] Among them, t is a hyperparameter given by humans, and its value is 1e-3. n c is the number of samples of category i and N is the total number of samples. Finally, the image repetition factor of each image I is calculated:
[0053] S2-2 Unified data enhancement mechanism;
[0054] Unify the data enhancement mechanism for each unlabeled image; first, weakly enhance the unlabeled image; apply color transformation, select a transformation from color transformation (brightness adjustment, contrast adjustment, saturation adjustment and hue adjustment), and the transformation is based on P color = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M color Randomly select within the allowed range [0,0.4]. If a transformation is decided, the image is transformed according to the selected magnitude, otherwise the image remains unchanged. Apply a shape transformation (horizontal flip, vertical flip, translation (left and right, up and down), scaling), and randomly select a transformation from the shape transformation category. The transformation is based on Pshape = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M shape Randomly select within the allowed range [0,0.4]. If it is decided to apply the transformation, the image is transformed according to the selected magnitude, otherwise the image remains unchanged and the weakly enhanced image is input into the teacher model; then for the strong enhancement of the unlabeled image; apply color transformation. For each unlabeled image, first select two color transformation operations, each color transformation operation has a probability parameter P color = 0.8 to decide whether to apply this operation. If you choose to apply the transformation, then color Randomly select an amplitude value in the amplitude range [0.5, 1.0] for transformation; randomly select and apply shape transformation, and then select two shape transformation operations, each of which also has a probability parameter P shape =0.8, if you choose to apply the transformation, then M shape Randomly select an amplitude value from the allowed amplitude range [0.5, 1.0] for transformation. Since the shape transformation will affect the bounding box in the image, the coordinates of the bounding box need to be adjusted accordingly. Randomly select and apply a hybrid transformation, and finally select a hybrid transformation operation. The hybrid transformation operation also has a probability parameter P. mix =0.8, if you choose to apply the transformation, then M mix A magnitude value is randomly selected from the allowed magnitude range [0.5, 1.0] for transformation. In particular, for the Mixup operation, the parameter range is set to [0, 0.2], indicating that the influence of other images on the current image is relatively low, and is input into the student model for training;
[0055] S2-3 Pseudo-label generation mechanism guided by diagnostic results;
[0056] The pseudo-label results of the teacher model processing the weakly enhanced data are used to supervise the student model training of the strongly enhanced data; different pseudo-label generation methods are applied to the panoramic image diagnosis results of different lesion levels, and the number of ASC-US cells is defined as N ASC-US , define the number of LSIL cells as N LSIL , define the number of ASC-H cells as N ASC-H , define the number of HSIL cells as N HSIL , T low is the lower threshold of abnormal cell number, and the total number of cells N total =N ASC-US +N LSIL +N ASC-H +N HSIL ;
[0057] For panoramic images with ASC-US diagnosis results, the label generation condition 1 is NASC-US >N ASC-H And N ASC-US >N HSIL And N ASC-US >N LSIL , condition 2 is N total <T low , the judgment formula is as follows:
[0058] Label ASC-US =IF N ASC-US >MAX(N ASC-H ,N HSIL ,N LSIL )AND(N total <T low ) (2)
[0059] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic image diagnosis of ASC-US is given as: Threshold ASC-US =0.4, Threshold LSIL =0.4, Threshold ASC-H =0.7, Threshold HSIL =0.7;
[0060] For panoramic image diagnosis results with LSIL, the label generation condition 1 is N LSIL >N ASC-US And N LSIL >N ASC-H And N LSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows:
[0061] Label LSIL =IF N LSIL >MAX(N ASC-H ,N HSIL ,N ASC-US )AND(N total <T low ) (3)
[0062] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as LSIL is given as: ASC-US =0.5,Threshold LSIL =0.5,Threshold ASC-H =0.7, Threshold HSIL =0.6;
[0063] For panoramic image diagnosis result ASC-H, label generation condition 1 is And N ASC-H >N ASC-US or N ASC-H >N LSIL or N HSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows:
[0064]
[0065] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as ASC-H is given as: Threshold ASC-US =0.6, Threshold LSIL =0.6, Threshold ASC-H =0.5,Threshold HSIL =0.5;
[0066] For panoramic images with HSIL diagnostic results, the label generation condition 1 is N HSIL >N ASC-US or N HSIL >N ASC-H or N HSIL >N LSIL Condition 2 is N total >=T low , the judgment formula is as follows:
[0067]
[0068] Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as HSIL is given as: Threshold ASC-US =0.6, Threshold LSIL =0.6, Threshold ASC-H =0.5,Threshold HSIL =0.4;
[0069] S3, the ratio of labeled data to unlabeled data is 1:4 through the label resampling mechanism of the tail cell category of S2-1;
[0070] S4, feeds the data of S3 into the unified data augmentation mechanism of S2-2, where the unlabeled data is fed into the student and teacher models of S2 using strong and weak augmentation respectively;
[0071] S5, the proposals generated by the teacher model of S4 are fed into the pseudo-label generation mechanism guided by the diagnosis results of S2-3 to generate pseudo-labels for unlabeled data to supervise the training of the student model of S4, where the weight of the teacher model is generated by the exponential moving average of the student model;
[0072] S6. Repeat S3-S5 until the preset number of training iterations is reached or the evaluation performance of the teacher-student learning model meets the requirements.
[0073] The present invention may also have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of the present invention.
Claims
1. A semi-supervised cervical abnormal cell long-tail resampling detection method based on diagnosis result-guided pseudo-labeling and unified data enhancement, characterized in that: The following steps are involved: S1. Prepare a dataset for detecting abnormal cervical cells, which is a dataset for detecting abnormal cervical cells using semi-supervised targets, by cropping labeled and unlabeled panoramic images into tile images; S2. Build a semi-supervised end-to-end detection network framework. This framework adds a label resampling mechanism for tail cell categories and a unified data enhancement mechanism based on the teacher-student approach, as well as a pseudo-label generation mechanism guided by diagnostic results. S3, the ratio of labeled data to unlabeled data is 1:4 through the label resampling mechanism of the tail cell category of S2-1; S4, feeds the data of S3 into the unified data augmentation mechanism of S2-2, where the unlabeled data is fed into the student and teacher models of S2 using strong and weak augmentation respectively; S5, the proposals generated by the teacher model of S4 are fed into the pseudo-label generation mechanism guided by the diagnosis results of S2-3 to generate pseudo-labels for unlabeled data to supervise the training of the student model of S4, where the weight of the teacher model is generated by the exponential moving average of the student model; S6. Repeat S3-S5 until the preset number of training iterations is reached or the evaluation performance of the teacher-student learning model meets the requirements.
2. The semi-supervised cervical abnormal cell long-tail resampling detection method based on diagnosis result-guided pseudo-labeling and unified data enhancement as described in claim 1, characterized in that: The preparation of the abnormal cervical cell target detection dataset described in step S1 is a semi-supervised target detection abnormal cervical cell dataset that is cropped from labeled and unlabeled panoramic images into tile images. The specific steps are as follows: Firstly, positive cervical cell pathology panoramic images from different institutions and with different diagnostic results were collected. Then, four types of abnormal cells were marked on the cervical cell pathology panoramic images using ASAP software. Next, the panoramic images were divided into training and validation sets of panoramic images in a ratio of 8:
2. Finally, the training and validation sets of panoramic images were divided into tiles with a length and width of 1024 as the final training and validation sets, of which 1%, 2%, 5% and 10% were extracted from the training sets as labeled data, and the rest of the training sets were used as unlabeled data.
3. The semi-supervised cervical abnormal cell long-tail resampling detection method based on diagnosis result-guided pseudo-labeling and unified data enhancement as described in claim 1, characterized in that: The construction described in step S2 is based on a semi-supervised end-to-end detection network framework, which adds a label resampling mechanism for tail cell categories and a unified data enhancement mechanism on the basis of the teacher-student method, as well as a pseudo-label generation mechanism guided by diagnosis results as follows: S2-1 Marker resampling mechanism of tail cell classes; The label resampling mechanism for tail categories is the first module of the framework. The sampler samples from the labeled and unlabeled datasets in a ratio of 1:4, where the tail categories are oversampled from the labeled dataset and enough head categories are mined from the pseudo-labels in S5 to balance the undersampling of the head categories in the labeled data. First, the label frequency of each category c is calculated: f(c). Then, the repetition factor r(c) of category c is determined by the frequency f(c) of the category in the image and the threshold t. The specific formula is as follows: Among them, t is a hyperparameter given by humans, and its value is 1e-3. n c is the number of samples of category i and N is the total number of samples. Finally, the image repetition factor of each image I is calculated: S2-2 Unified data enhancement mechanism; Unify the data enhancement mechanism for each unlabeled image; first, weakly enhance the unlabeled image; apply color transformation, select a transformation from color transformation (brightness adjustment, contrast adjustment, saturation adjustment and hue adjustment), and the transformation is based on P color = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M color Randomly select within the allowed range [0,0.4]. If a transformation is decided, the image is transformed according to the selected magnitude, otherwise the image remains unchanged. Apply a shape transformation (horizontal flip, vertical flip, translation (left and right, up and down), scaling), and randomly select a transformation from the shape transformation category. The transformation is based on P shape = 0.5 probability is applied, if you choose to apply the transformation, then the amplitude M shape Randomly select within the allowed range [0,0.4]. If it is decided to apply the transformation, the image is transformed according to the selected magnitude, otherwise the image remains unchanged and the weakly enhanced image is input into the teacher model; then for the strong enhancement of the unlabeled image; apply color transformation. For each unlabeled image, first select two color transformation operations, each color transformation operation has a probability parameter P color = 0.8 to decide whether to apply this operation. If you choose to apply the transformation, then color Randomly select an amplitude value in the amplitude range [0.5, 1.0] for transformation; randomly select and apply shape transformation, and then select two shape transformation operations, each of which also has a probability parameter P shape =0.8, if you choose to apply the transformation, then M shape Randomly select an amplitude value from the allowed amplitude range [0.5, 1.0] for transformation. Since the shape transformation will affect the bounding box in the image, the coordinates of the bounding box need to be adjusted accordingly. Randomly select and apply a hybrid transformation, and finally select a hybrid transformation operation. The hybrid transformation operation also has a probability parameter P. mix =0.8, if you choose to apply the transformation, then M mix A magnitude value is randomly selected from the allowed magnitude range [0.5, 1.0] for transformation. In particular, for the Mixup operation, the parameter range is set to [0, 0.2], indicating that the influence of other images on the current image is relatively low, and is input into the student model for training; S2-3 Pseudo-label generation mechanism guided by diagnostic results; The pseudo-label results of the teacher model processing the weakly enhanced data are used to supervise the student model training of the strongly enhanced data; different pseudo-label generation methods are applied to the panoramic image diagnosis results of different lesion levels, and the number of ASC-US cells is defined as N ASC-US , define the number of LSIL cells as N LSIL , define the number of ASC-H cells as N ASC-H , define the number of HSIL cells as N HSIL , T low is the lower threshold of abnormal cell number, and the total number of cells N total =N ASC-US +N LSIL +N ASC-H +N HSIL ; For panoramic images with ASC-US diagnosis results, the label generation condition 1 is N ASC-US >N ASC-H And N ASC-US >N HSIL And N ASC-US >N LSIL , condition 2 is N total <T low , the judgment formula is as follows: Label ASC-US =IF N ASC-US >MAX(N ASC-H ,N HSIL ,N LSIL )AND(N total <T low ) (2) Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic image diagnosis of ASC-US is given as: Threshold ASC-US =0.4, Threshold LSIL =0.4, Threshold ASC-H =0.7, Threshold HSIL =0.7; For panoramic image diagnosis results with LSIL, the label generation condition 1 is N LSIL >N ASC-US And N LSIL >N ASC-H And N LSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows: Label LSIL =IF N LSIL >MAX(N ASC-H ,N HSIL ,N ASC-US )AND(N total <T low ) (3) Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as LSIL is given as: Threshold ASC-US =0.5,Threshold LSIL =0.5,Threshold ASC-H =0.7,Threshold HSIL =0.6; For panoramic image diagnosis result ASC-H, label generation condition 1 is And N ASC-H >N ASC-US or N ASC-H >N LSIL or N HSIL >N HSIL , condition 2 is N total >=T low , the judgment formula is as follows: Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as ASC-H is: Threshold ASC-US =0.6,Threshold LSIL =0.6,Threshold ASC-H =0.5,Threshold HSIL =0.5; For panoramic images with HSIL diagnostic results, the label generation condition 1 is N HSIL >N ASC-US or N HSIL >N ASC-H or N HSIL >N LSIL Condition 2 is N total >=T low , the judgment formula is as follows: Based on the above formula, the pseudo-label dynamic filtering threshold for panoramic images diagnosed as HSIL is given as: Threshold ASC-US =0.6,Threshold LSIL =0.6,Threshold ASC-H =0.5,Threshold HSIL =0.4。
Citation Information
Cited By
Patrol robot real-time abnormal target detection method based on semi-supervised learning
CN121236566A
