A multi-scale dual-stream cervical abnormal cell detection method integrating pathological knowledge
Through the multi-scale double-flow cervical abnormal cell detection method, combined with pathological knowledge and deep learning models, the problem of abnormal cell detection in cervical cell pathological images is solved, the accuracy and efficiency of the detection are improved, and pathologists can support more accurate diagnosis.
Patent Information
- Application Number
- CN202310375318.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-10
- Publication Date
- 2025-09-02
- Estimated Expiration
- 2043-04-10
AI Technical Summary
The characteristics in the panoramic image of cervical cell pathology are sparse, and it is difficult to detect abnormal cells of different sizes. It is difficult for the existing technology to effectively identify and locate abnormal cells.
A multi-scale double-flow cervical abnormal cell detection method with fusion pathological knowledge was used to construct detection models through multi-scale ideas combined with Faster R-CNN and DenseNet, and a double-flow model based on cross-visual Transformer and CNN were used to extract global and local features of abnormal cells and perform feature fusion.
It significantly improves the accuracy, sensitivity and specificity of cervical cell detection, reduces the workload of pathologists, and improves the accuracy and popularity of cervical cancer screening.
Smart Images
Figure BDA0004170232950000021 
Figure BDA0004170232950000022 
Figure BDA0004170232950000031
Abstract
Description
Technical Field
[0001] The present invention relates to the detection of abnormal cervical cells. Background Art
[0002] In recent years, the continued outbreak of cancer has become a serious social problem in my country. Early diagnosis and treatment are effective means of addressing this issue. Statistics from the International Agency for Research on Cancer indicate that in 2020, there were 19.3 million new cancer cases and 10 million deaths worldwide. my country accounted for 4.57 million new cancer cases, ranking first globally. Cervical cancer remains one of the most deadly cancers threatening women's health, with one woman dying from it every two minutes. Due to my country's large population and limited diagnostic and treatment resources, it has become a high-incidence and hardest-hit area for cervical cancer. Cytopathology, in which pathologists screen cervical smears under a microscope, is a widely used routine examination method. This method consumes a significant amount of time and effort and has low accuracy. With the development of artificial intelligence, computer-assisted interpretation has been developed and implemented, meeting this urgent need. This has reduced the workload of pathologists, reduced potential misdiagnosis rates, and increased the accuracy and prevalence of cervical cancer screening in my country.
[0003] Computer-assisted image interpretation technology captures panoramic images (Whole Slide Images, WSIs) of cervical cell slides, then uses image processing and recognition techniques to analyze and measure them. Finally, abnormal cells are identified and classified for review by the physician. However, due to issues such as high image resolution, microlesions, sparse panoramic image features, insufficient application of pathological knowledge, and difficulty identifying cervical cells of varying sizes, automated analysis of cervical images using digital WSIs remains a significant challenge. Therefore, this paper proposes a multi-scale, dual-stream method for abnormal cervical cell detection that integrates pathological knowledge. This method effectively integrates pathological knowledge using multi-scale concepts to identify and locate abnormal cells. A dual-stream model based on a cross-vision Transformer and CNN is used to extract global and local features of abnormal cells, respectively. This method improves the accuracy of cell classification in panoramic cervical pathology images and effectively reduces the workload of pathologists in routine practice. Summary of the Invention
[0004] The purpose of this invention is to solve the problem of sparse features and difficulty in detecting abnormal cells of different sizes in the abnormal cell detection technology of cervical cell pathology panoramic images, and thus propose a multi-scale dual-stream cervical abnormal cell detection method that integrates pathological knowledge.
[0005] The above invention objectives are mainly achieved through the following technical solutions:
[0006] S1. Data preparation: Acquire panoramic images of cervical cell pathology and divide them into multiple non-overlapping image blocks, and remove image blocks with low information content for subsequent training and testing;
[0007] First, a large number of cervical cytopathology smears are collected and digitized into cervical cytopathology panoramic images. The panoramic images are then cropped into multiple non-overlapping fixed-size image blocks. Finally, image blocks with low information content, i.e., image blocks covered by background or blurred, are removed, and the remaining image blocks are used for subsequent training and testing.
[0008] S2. Data annotation: manually draw bounding boxes around abnormal cells using annotation tools;
[0009] To obtain true annotations, three expert pathologists participated in the preparation of the annotations for the image blocks obtained in S1. They performed detailed annotations of abnormal cervical cells in each image block. Normal and abnormal cervical cells were defined according to the American College of Obstetricians and Gynecologists (ACOG) guidelines. Each pathologist used tools such as Colabeler to annotate bounding boxes around the abnormal cells in the image blocks. This annotation process was performed in three steps: initial annotation, verification, and final review.
[0010] S3. Design region proposals in a multi-scale network based on the pathological knowledge of the size and shape distribution of real cervical abnormal cells;
[0011] According to pathological knowledge, most cervical exfoliated cells are approximately circular or elliptical and rarely expand in one direction. Statistical calculations are performed on abnormal cervical cells in the collected data set to obtain size-related features and indicators of abnormal cells in pixels, including nuclear-cytoplasmic ratio, cell circular fit, cell area ratio, Iod, etc. Based on these relevant indicators, region proposals and anchor points in the multi-scale network are determined. The calculation formula is as follows:
[0012]
[0013] Where R is the nucleus-cytoplasm ratio, A is the n is the cell nucleus area, A c The cell area is calculated in a similar way to the cell nucleus area. Taking the cell nucleus area as an example, the definition is as follows:
[0014]
[0015] Where S n is the nuclear region, p i is a pixel point in the cell nuclear region, and the cell nuclear area is the sum of the pixels in the cell nuclear region;
[0016]
[0017] Where N polygon is the cell circle fitting degree, A n is the area of the cell nucleus, d1 and d2 are the long axes of the cell nucleus, and P is the perimeter of the cell nucleus;
[0018]
[0019] Where K area is the cell area ratio, A normal is the normal cell area, A abnormal is the area of abnormal cells;
[0020]
[0021] Where Gx i To change the gray value of the pixel, GA background is the gray value of the sample background;
[0022] To ensure that the size and aspect ratio of the region proposal are consistent with the actual cervical cell distribution, an anchor box with one optional size and two optional aspect ratios is generated at each feature map position. The size (i.e., height in pixels) and aspect ratio are defined at different scales, which can cover abnormal cervical cells of different sizes in the current image.
[0023] S4. Build and train a multi-scale abnormal cervical cell detection model that integrates pathological knowledge;
[0024] First, the processed image block training dataset is used and input into the constructed multi-scale cervical abnormal cell detection model for training; the main structure of the model is the multi-scale idea combined with Faster R-CNN for abnormal cell detection. First, a region proposal network that integrates pathological knowledge is used to generate proposals, and then the Fast R-CNN module is used to predict the final ROI; the multi-scale idea is combined with Faster R-CNN to solve the problem that it cannot detect cells of different sizes due to the use of only single-scale feature representation. The network uses feature maps extracted from images of different scales to detect cells of different sizes; among them, the network framework uses DenseNet as the backbone network of the detection network, which contains a total of four dense blocks. There is a transition layer at the end of the first three dense blocks, and then the dense block feature maps are downsampled according to the scale factor. The trunk forms a bottom-up path, and the output of the convolution layer of each transition layer is selected as the multi-scale feature; after the multi-scale network, feature maps of five scales are obtained, and region proposals are generated at different scales through the region proposal network; finally, the region proposals and corresponding features are combined Figure 1 The images are fed into the Fast R-CNN block to detect abnormal cells of different sizes.
[0025] S5. Build and train a dual-stream abnormal cervical cell classification model;
[0026] First, the cell image is input into two branches for processing. One branch is based on the Cross Vision Transformer (Cross Vision Transformer, CrossViT) model, which is used to extract the global features of the cell image. Usually, the ViT model consists of an encoder (Encoder)-decoder (Decoder) structure, in which the encoder is to convert the input image into a feature vector; the CrossViT model is divided into two independent and parallel Large branches and Small branches. The cell image is input into the model, and the input image is divided into blocks in two different segmentation methods based on pixels. In this way, effective input sequences of different sizes and independent parallel are obtained. The two image block sequences of different sizes correspond to the input Large branch and the Small branch for extracting image block features; the image block sequence is mapped into a fixed-dimensional vector after a learnable linear projection, namely the image patch encoding (Image Patch Token), in order to clarify the position information of the image block in the original image, a position code is added to the image block encoding to form an encoding vector; after receiving the encoding vector, the Transformer encoder runs iteratively, obtaining image semantic and structural features at different levels through multiple layers of encoders, and the output vector sequence of the previous layer encoder is the input of the next layer encoder; the output of the last layer encoder is a feature vector sequence combined with the classification flag CLS, which is input into the Cross-Attention module for multi-scale feature vector fusion, and the classification flag CLS in the fused vector of the Cross-Attention module is extracted separately as information for image classification prediction; the other branch is the CNN model, which is used to extract local features of the image; the cell image is input into the CrossViT model and the CNN model respectively to extract different types of features, and then the output vectors of the two feature extractors are spliced and fused to obtain the final feature vector to predict cell abnormalities.
[0027] S6. Use the trained cervical abnormal cell detection model and classification model to detect abnormal cells in the cervical cell pathology panoramic image;
[0028] First, the panoramic image of cervical cell pathology is cropped to generate non-overlapping image blocks of fixed size. Then, the multi-scale cervical abnormal cell detection model that integrates pathological knowledge described in S4 is used to detect the location and probability information of abnormal cells in the image blocks. This information can provide valuable diagnostic suggestions for pathologists during the screening process. Then, the two-stream cervical abnormal cell classification model that has been trained in S5 is used to calculate the abnormal cell classification results, and then the image block level classification is completed, that is, to determine whether the image block is identified as normal or abnormal. After the image block level classification, each image block contained in each panoramic image has its corresponding abnormal probability. The 10 image blocks with the highest abnormal probability of each panoramic image are collected, and their average value is calculated to obtain the abnormal probability value at the panoramic image level.
[0029] Effects of the Invention
[0030] This paper provides a multi-scale, two-stream method for detecting abnormal cervical cells that integrates pathological knowledge. This algorithm mimics the way pathologists read cervical cytology images. It first acquires a panoramic cervical cytology image and crops it to generate fixed-size, non-overlapping image blocks. It then uses a multi-scale, region-based feature fusion network guided by pathological knowledge to detect fine features of abnormal cervical cells at different scales. A two-stream model based on a cross-vision Transformer and a CNN is then combined to extract global and local features of the abnormal cells, respectively. The output vectors of the two feature extractors are then fused, and the resulting feature vector is used for class prediction. Finally, the predictions are aggregated to determine a panoramic image-level diagnosis. Experiments demonstrate that this method effectively integrates pathological knowledge, significantly improving the quality of multi-scale detection region proposals, robustly detecting cells of varying sizes, and improving the accuracy, sensitivity, and specificity of cervical cell detection. This method provides effective and efficient technical support for cervical cancer screening and improving pathology workflows, helping pathologists make more accurate diagnoses. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 Schematic diagram of the main process of the algorithm
[0032] Figure 2 Schematic diagram of a multi-scale abnormal cervical cell detection model integrating pathological knowledge;
[0033] Figure 3 Schematic diagram of the dual-stream abnormal cervical cell classification model; Specific implementation methods Specific implementation method one:
[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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 making creative efforts shall fall within the scope of protection of the present invention.
[0036] like Figure 1 This article provides a multi-scale dual-stream cervical abnormal cell detection method that integrates pathological knowledge, including data preparation, data annotation, model training, and panoramic image abnormal cell detection:
[0037] The data preparation and data labeling include the following steps:
[0038] S1. Data preparation: Acquire panoramic images of cervical cell pathology and divide them into multiple non-overlapping image blocks, and remove image blocks with low information content for subsequent training and testing;
[0039] S2. Data annotation: manually draw bounding boxes around abnormal cells using annotation tools;
[0040] The model training includes the following steps:
[0041] S3. Design region proposals in a multi-scale network based on the pathological knowledge of the size and shape distribution of real cervical abnormal cells;
[0042] S4. Build and train a multi-scale abnormal cervical cell detection model that integrates pathological knowledge;
[0043] S5. Build and train a dual-stream abnormal cervical cell classification model;
[0044] The panoramic image abnormal cell detection comprises the following steps:
[0045] S6. Use the trained cervical abnormal cell detection model and classification model to detect abnormal cells in the cervical cell pathology panoramic image.
[0046] The embodiment of the present invention first prepares cervical cell pathology panoramic image data, divides it into multiple non-overlapping image blocks for training and testing, and uses annotation tools to manually draw bounding boxes around abnormal cells; next, pathology knowledge is fully utilized to construct and train an abnormal cell detection and classification model; then a multi-scale region-based feature fusion network guided by pathology knowledge is used to detect the fine features of abnormal cervical cells at different scales; a dual-stream cervical abnormal cell classification model is used to classify abnormal cells by extracting global and local information of the cells; finally, the classification results are aggregated to obtain the cervical cell pathology panoramic image classification results.
[0047] The embodiments of the present invention are described in detail below:
[0048] Data preparation and data labeling include the following steps:
[0049] S1. Data preparation: Acquire panoramic images of cervical cell pathology and divide them into multiple non-overlapping image blocks, and remove image blocks with low information content for subsequent training and testing;
[0050] First, a large number of cervical cytopathology smears are collected and digitized into cervical cytopathology panoramic images. The panoramic images are then cropped into multiple non-overlapping fixed-size image blocks. Finally, image blocks with low information content, i.e., image blocks covered by background or blurred, are removed, and the remaining image blocks are used for subsequent training and testing.
[0051] S2. Data annotation: manually draw bounding boxes around abnormal cells using annotation tools;
[0052] To obtain true annotations, three expert pathologists participated in the preparation of the annotations for the image blocks obtained in S1. They performed detailed annotations of abnormal cervical cells in each image block. Normal and abnormal cervical cells were defined according to the American College of Obstetricians and Gynecologists (ACOG) guidelines. Each pathologist used tools such as Colabeler to annotate bounding boxes around the abnormal cells in the image blocks. This annotation process was performed in three steps: initial annotation, verification, and final review.
[0053] The model training includes the following steps:
[0054] S3. Design region proposals in a multi-scale network based on the pathological knowledge of the size and shape distribution of real cervical abnormal cells;
[0055] According to pathological knowledge, most cervical exfoliated cells are approximately circular or elliptical and rarely expand in one direction. Statistical calculations are performed on abnormal cervical cells in the collected data set to obtain size-related features and indicators of abnormal cells in pixels, including nuclear-cytoplasmic ratio, cell circular fit, cell area ratio, Iod, etc. Based on these relevant indicators, region proposals and anchor points in the multi-scale network are determined. The calculation formula is as follows:
[0056]
[0057] Where R is the nucleus-cytoplasm ratio, A is the n is the cell nucleus area, A c The cell area is calculated in a similar way to the cell nucleus area. Taking the cell nucleus area as an example, the definition is as follows:
[0058]
[0059] Where Sn is the nuclear region, p i is a pixel point in the cell nuclear region, and the cell nuclear area is the sum of the pixels in the cell nuclear region;
[0060]
[0061] Where N polygon is the cell circle fitting degree, A n is the area of the cell nucleus, d1 and d2 are the long axes of the cell nucleus, and P is the perimeter of the cell nucleus;
[0062]
[0063] Where K area is the cell area ratio, A normal is the normal cell area, A abnormal is the area of abnormal cells;
[0064]
[0065] Where Gx i To change the gray value of the pixel, GA background is the gray value of the sample background;
[0066] To ensure that the size and aspect ratio of the region proposal are consistent with the actual cervical cell distribution, an anchor box with one optional size and two optional aspect ratios is generated at each feature map position. The size (i.e., height in pixels) and aspect ratio are defined at different scales, which can cover abnormal cervical cells of different sizes in the current image.
[0067] S4. Build and train a multi-scale abnormal cervical cell detection model that integrates pathological knowledge;
[0068] The detection model structure is as follows Figure 2As shown in the figure; first, the processed image block training dataset is used and input into the multi-scale cervical abnormal cell detection model for training; the main structure of the model is the multi-scale idea combined with Faster R-CNN for abnormal cell detection. First, the region proposal network that integrates pathological knowledge is used to generate proposals, and then the Fast R-CNN module is used to predict the final ROI; the multi-scale idea is combined with Faster R-CNN to solve the problem that it cannot detect cells of different sizes due to the use of only single-scale feature representation. The network uses feature maps extracted from images of different scales to detect cells of different sizes; among them, the network framework uses DenseNet as the backbone network of the detection network, which contains a total of four dense blocks. There is a transition layer at the end of the first three dense blocks, and then the dense block feature map is downsampled according to the scale factor. The trunk forms a bottom-up path, and the output of the convolution layer of each transition layer is selected as the multi-scale feature; after the multi-scale network, feature maps of five scales are obtained, and region proposals are generated at different scales through the region proposal network; finally, the region proposals and corresponding features are combined Figure 1 The images are fed into the Fast R-CNN block to detect abnormal cells of different sizes.
[0069] S5. Build and train a dual-stream abnormal cervical cell classification model;
[0070] The classification model structure is as follows Figure 3As shown in the figure, first, the cell image is input into two branches for processing. One branch is based on the Cross Vision Transformer (Cross Vision Transformer, CrossViT) model, which is used to extract the global features of the cell image. Usually, the ViT model consists of an encoder (Encoder)-decoder (Decoder) structure, in which the encoder is used to convert the input image into a feature vector. The CrossViT model is divided into two independent and parallel Large branches and Small branches. The cell image is input into the model, and the input image is divided into blocks in two different segmentation methods based on pixels. In this way, effective input sequences of different sizes and independent parallel are obtained. The two image block sequences of different sizes correspond to the input Large branch and the Small branch for extracting image block features. After a learnable linear projection, the image block sequence is mapped into a fixed-dimensional vector, namely the image patch encoding (Image Patch Token), in order to clarify the position information of the image block in the original image, a position code is added to the image block encoding to form an encoding vector; after receiving the encoding vector, the Transformer encoder runs iteratively, obtaining image semantic and structural features at different levels through multiple layers of encoders, and the output vector sequence of the previous layer encoder is the input of the next layer encoder; the output of the last layer encoder is a feature vector sequence combined with the classification flag CLS, which is input into the Cross-Attention module for multi-scale feature vector fusion, and the classification flag CLS in the fused vector of the Cross-Attention module is extracted separately as information for image classification prediction; the other branch is the CNN model, which is used to extract local features of the image; the cell image is input into the CrossViT model and the CNN model respectively to extract different types of features, and then the output vectors of the two feature extractors are spliced and fused to obtain the final feature vector to predict cell abnormalities.
[0071] The panoramic image abnormal cell detection comprises the following steps:
[0072] S6. Use the trained cervical abnormal cell detection model and classification model to detect abnormal cells in the cervical cell pathology panoramic image.
[0073] First, the panoramic image of cervical cell pathology is cropped to generate non-overlapping image blocks of fixed size. Then, the multi-scale cervical abnormal cell detection model that integrates pathological knowledge described in S4 is used to detect the location and probability information of abnormal cells in the image blocks. This information can provide valuable diagnostic suggestions for pathologists during the screening process. Then, the two-stream cervical abnormal cell classification model that has been trained in S5 is used to calculate the abnormal cell classification results, and then the image block level classification is completed, that is, to determine whether the image block is identified as normal or abnormal. After the image block level classification, each image block contained in each panoramic image has its corresponding abnormal probability. The 10 image blocks with the highest abnormal probability of each panoramic image are collected, and their average value is calculated to obtain the abnormal probability value at the panoramic image level.
[0074] The present invention may have many other 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 multi-scale dual-stream cervical abnormal cell detection method integrating pathological knowledge, characterized in that: The following steps are included: S1. Data preparation: Collect panoramic images of cervical cell pathology and divide them into multiple non-overlapping image blocks, and remove image blocks with low information content for subsequent training and testing; S2. Data annotation: manually draw bounding boxes around abnormal cells using annotation tools; S3. Design region proposals in a multi-scale network based on the pathological knowledge of the size and shape distribution of real cervical abnormal cells; S4. Build and train a multi-scale abnormal cervical cell detection model that integrates pathological knowledge; S5. Construct and train a dual-stream abnormal cervical cell classification model. First, the cell image is input into two branches for processing. One branch is the CrossViT model based on the Cross Visual Transformer, which is used to extract the global features of the cell image. Usually, the ViT model consists of an encoder-decoder structure, where the encoder converts the input image into a feature vector. The CrossViT model is divided into two independent and parallel Large branches and Small branches. The cell image is input into the model, and the input image is divided into blocks in two different segmentation methods based on pixels. In this way, effective input sequences of different sizes and independent parallelism are obtained. The two image block sequences of different sizes correspond to the input Large branch and the Small branch for extracting image block features. After a learnable linear projection, the image block sequence is mapped into a vector of fixed dimension, namely, the image patch encoding (Image Patch Token), in order to clarify the position information of the image block in the original image, a position code is added to the image block encoding to form an encoding vector; after receiving the encoding vector, the Transformer encoder runs in an iterative manner, obtaining image semantic and structural features at different levels through multi-layer encoders, and the output vector sequence of the previous layer encoder is the input of the next layer encoder; the output of the last layer encoder is a feature vector sequence combined with the classification flag CLS, which is input into the Cross-Attention module for multi-scale feature vector fusion, and the classification flag CLS in the fused vector of the Cross-Attention module is extracted separately as information for image classification prediction; the other branch is the CNN model, which is used to extract local features of the image; the cell image is input into the CrossViT model and the CNN model branches respectively to extract different types of features, and then the output vectors of the two feature extractors are spliced and fused to obtain the final feature vector to predict cell abnormalities; S6. Use the trained cervical abnormal cell detection model and classification model to detect abnormal cells in the cervical cell pathology panoramic image.
2. The multi-scale dual-stream cervical abnormal cell detection method integrating pathological knowledge as claimed in claim 1, characterized in that: The data preparation described in step S1 is to obtain a panoramic image of cervical cell pathology, divide it into multiple non-overlapping image blocks, and remove image blocks with low information content for subsequent training and testing; the specific steps are as follows: First, a large number of cervical cytopathology smears are collected and digitized into cervical cytopathology panoramic images using an automatic scanner. The panoramic images are then cropped into multiple non-overlapping image blocks of fixed size. Finally, image blocks with low information content, i.e., image blocks covered by background or blurred, are removed, and the remaining image blocks are used for subsequent training and testing.
3. The multi-scale dual-stream cervical abnormal cell detection method integrating pathological knowledge as claimed in claim 1, characterized in that: For the data annotation described in step S2, manually draw bounding boxes around abnormal cells using the annotation tool. The specific steps are as follows: To obtain realistic annotations, three pathologists participated in the preparation of the annotations for the image blocks obtained in S1. They performed detailed annotations on the abnormal cervical cells in each image block. The definitions of positive and abnormal cervical cells were based on the American College of Obstetricians and Gynecologists (ACOG) guidelines. Each pathologist used tools such as Colabeler to annotate bounding boxes around the abnormal cells in the image blocks. The following three-step annotation process was performed: an initial annotation step, a verification step, and a final inspection step.
4. The multi-scale dual-stream cervical abnormal cell detection method integrating pathological knowledge as claimed in claim 1, characterized in that: The region proposal in the multi-scale network is designed based on the pathological knowledge of the size and shape distribution of abnormal cervical cells described in step S3. The specific steps are as follows: According to pathological knowledge, most cervical exfoliated cells are approximately circular or elliptical and rarely expand in one direction. Statistical calculations are performed on abnormal cervical cells in the collected dataset to obtain size-related features and indicators of abnormal cells in pixels, including nuclear-cytoplasmic ratio, cell circularity fit, cell area ratio, and Iod. Based on these indicators, region proposals and anchor points in the multi-scale network are determined. The calculation formula is as follows: Where R is the nucleus-cytoplasm ratio, A is the n is the cell nucleus area, A c The calculation methods of cell area and cell nuclear area are similar. Taking the calculation of cell nuclear area as an example, the formula is as follows: Where S n is the nuclear region, p i is a pixel point in the cell nuclear region, and the cell nuclear area is the sum of the pixels in the cell nuclear region; Where N polygon is the cell circle fitting degree, A n is the area of the cell nucleus, d1 and d2 are the long axes of the cell nucleus, and P is the perimeter of the cell nucleus; Where K area is the cell area ratio, A normal is the normal cell area, A abnormal is the area of abnormal cells; Where Gx i To change the gray value of the pixel, GA background is the gray value of the sample background; To ensure that the size and aspect ratio of the region proposal are consistent with the actual cervical cell distribution, an anchor box with one optional size and two optional aspect ratios is generated at each feature map position. The size (i.e., height in pixels) and aspect ratio are defined at different scales, which covers abnormal cervical cells of different sizes in the current image.
5. The multi-scale dual-stream abnormal cervical cell detection method integrating pathological knowledge as claimed in claim 1, characterized in that: The specific steps of constructing and training the multi-scale abnormal cervical cell detection model integrating pathological knowledge described in step S4 are as follows: The model first uses a processed image block training dataset and then inputs it into the constructed multi-scale abnormal cervical cell detection model for training. The main structure of the model combines multi-scale thinking with Faster R-CNN for abnormal cell detection. First, a region proposal network that integrates pathological knowledge is used to generate proposals, and then the Fast R-CNN module is used to predict the final region of interest. The multi-scale thinking is combined with Faster R-CNN to address the problem that Faster R-CNN cannot detect cells of different sizes due to its use of only single-scale feature representation. The network uses feature maps extracted from images of different scales to detect cells of different sizes. The network framework uses DenseNet as the backbone of the detection network, which consists of four dense blocks. The first three dense blocks are each tailed by a transition layer. The dense block feature maps are then downsampled according to the scale factor. The backbone forms a bottom-up path, and the output of the convolutional layer of each transition layer is selected as the multi-scale feature. After the multi-scale network, feature maps at five scales are obtained. The region proposal network generates region proposals at different scales. Finally, the region proposals and corresponding feature maps are input into the Fast R-CNN block to detect abnormal cells of different sizes.
6. The multi-scale dual-stream abnormal cervical cell detection method integrating pathological knowledge as claimed in claim 1, characterized in that: The specific steps of using the trained abnormal cervical cell detection model and classification model to detect abnormal cells in the cervical cell pathology panoramic image in step S6 are as follows: First, the panoramic image of cervical cell pathology is cropped to generate non-overlapping image blocks of fixed size. Then, the multi-scale cervical abnormal cell detection model that integrates pathological knowledge described in S4 is used to detect the location and probability information of abnormal cells in the image blocks. This information provides valuable diagnostic suggestions for pathologists during the screening process. Then, the two-stream cervical abnormal cell classification model that has been trained in S5 is used to calculate the abnormal cell classification results, and then the image block level classification is completed, that is, to determine whether the image block is identified as normal or abnormal. After the image block level classification, each image block contained in each panoramic image has its corresponding abnormal probability. The 10 image blocks with the highest abnormal probability of each panoramic image are collected, and their average value is calculated to obtain the abnormal probability value at the panoramic image level.