A cervical cancer pathological image-based classification system and storage medium

Abstract

The application relates to the technical field of image processing, in particular to a classification system based on cervical cancer pathological images and a storage medium. A computer program in the storage medium executes the following steps: obtaining a cervical cancer pathological whole section image, performing instance segmentation on the cervical cancer pathological whole section image to obtain a plurality of epithelial cell nuclei, center coordinates and comprehensive confidence; extracting morphological features and depth features of each epithelial cell nucleus, and splicing the morphological features and the depth features to form an initial feature vector; determining adjacent relationship nodes between the epithelial cell nuclei based on the center coordinates; constructing an epithelial cell nucleus graph structure data according to the epithelial cell nuclei, the initial feature vector and the adjacent relationship nodes; constructing a cervical cancer classification model, inputting the epithelial cell nucleus graph structure data and the comprehensive confidence into the cervical cancer classification model to obtain a classification result. The problems that the classification based on the cervical cancer pathological images in the prior art is not intelligent and accurate are solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically a classification system and storage medium based on cervical cancer pathological images. Background Technology

[0002] Cervical cancer is a common malignant tumor of the female reproductive system. Accurate classification of whole-slide pathology images is the core basis for early screening and lesion grading diagnosis of cervical cancer. Currently, in clinical practice, pathologists mainly rely on visual observation of hematoxylin-eosin stained whole-slide images, judging the morphology and distribution characteristics of epithelial cell nuclei to grade lesions based on experience. This method is not only time-consuming and labor-intensive, but also easily affected by factors such as the doctor's subjective experience and visual fatigue, resulting in low diagnostic consistency. Furthermore, the large amount of data from high-resolution whole-slide images makes manual screening inefficient for meeting the needs of large-scale clinical diagnosis.

[0003] Therefore, there is an urgent need for a more intelligent and accurate classification method based on cervical cancer pathological images. Summary of the Invention

[0004] To address the shortcomings of existing technologies, one aspect of the present invention provides a classification system and storage medium based on cervical cancer pathological images, which solves the problem that existing classification systems based on cervical cancer pathological images are not intelligent or accurate enough.

[0005] To achieve the above objectives, one aspect of the present invention provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the following steps: acquiring a whole-section image of cervical cancer pathology, and performing instance segmentation on the whole-section image of cervical cancer pathology to obtain multiple epithelial cell nuclei, center coordinates, and comprehensive confidence scores; extracting morphological and depth features of each epithelial cell nucleus, and concatenating the morphological and depth features to form an initial feature vector; determining adjacent relationship nodes between the epithelial cell nuclei based on the center coordinates; constructing epithelial cell nucleus map structure data based on the epithelial cell nuclei, the initial feature vectors, and the adjacent relationship nodes; constructing a cervical cancer classification model, and inputting the epithelial cell nucleus map structure data and the comprehensive confidence scores into the cervical cancer classification model to obtain a classification result.

[0006] This invention obtains epithelial cell nuclei, center coordinates, and comprehensive confidence scores by segmenting whole-slice images of cervical cancer pathology. It accurately locates target regions and quantifies the reliability of results. It extracts the morphological and depth features of epithelial cell nuclei and concatenates them into an initial feature vector. It integrates concrete pathological features and abstract depth features to enrich the dimension of node feature representation. It determines adjacent nodes by center coordinates and constructs epithelial cell nucleus graph structure data, capturing the spatial topological relationships between cell nuclei and restoring the structural features of pathological tissue. By inputting the graph structure data and comprehensive confidence scores into a cervical cancer classification model, it obtains classification results. The model combines structural information and reliability weights, improving the accuracy and intelligence of classification.

[0007] Optionally, the step of segmenting the whole cervical cancer pathology image to obtain multiple epithelial cell nuclei, center coordinates, and overall confidence scores includes: performing color space conversion and staining channel separation on the whole cervical cancer pathology image to obtain a hematoxylin channel image; adjusting the hematoxylin channel image using a pre-acquired standard staining matrix to obtain an adjusted hematoxylin channel image; and performing instance segmentation on the adjusted hematoxylin channel image to obtain epithelial cell nuclei, center coordinates, and overall confidence scores.

[0008] This invention obtains hematoxylin channel images by performing color space conversion and staining channel separation on whole-section images of cervical cancer pathology, focusing on the core region of the cell nucleus and suppressing background interference. The hematoxylin channel images are adjusted using a standard staining matrix to correct image deviations caused by staining processes and equipment, achieving standardization and uniformity of staining style and ensuring consistency of image features. Based on the adjusted images, instance segmentation is performed to obtain information such as epithelial cell nuclei, allowing segmentation to be performed on a high-quality, standardized image basis, thus improving the accuracy of cell nucleus segmentation and recognition.

[0009] Optionally, adjusting the hematoxylin channel image using a pre-acquired standard staining matrix to obtain an adjusted hematoxylin channel image includes: calculating the optical density mapping matrix of the hematoxylin channel image; performing singular value decomposition on the optical density mapping matrix to obtain a target image concentration matrix; obtaining a standard staining matrix; calculating a normalized optical density mapping matrix using the standard staining matrix and the target image concentration matrix; and performing an inverse space transformation on the normalized optical density mapping matrix to obtain the adjusted hematoxylin channel image.

[0010] This invention calculates the optical density mapping matrix of the hematoxylin channel image, converting pixel intensity into a value linearly related to dye concentration, eliminating interference from illumination variations and improving the effectiveness of image features. By performing singular value decomposition on the optical density mapping matrix, the target image concentration matrix is ​​obtained, accurately extracting dye concentration information and removing staining-irrelevant noise. By calculating the standardized optical density mapping matrix using the standard staining matrix and the concentration matrix, a unified correction of staining style is achieved, solving the problem of staining differences between images. The adjusted image is obtained through inverse space transformation and restored to an image format suitable for subsequent processing, ensuring the continuity of processing and improving the standardization and consistency of the hematoxylin channel image.

[0011] Optionally, the step of performing instance segmentation on the adjusted hematoxylin channel image to obtain epithelial cell nuclei, center coordinates, and overall confidence scores includes: inputting the adjusted hematoxylin channel image into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nuclei, center coordinates, and segmentation confidence scores; classifying the cell nuclei to obtain epithelial cell nuclei and classification confidence scores; and calculating an overall confidence score based on the classification confidence scores and the segmentation confidence scores.

[0012] This invention inputs the adjusted hematoxylin channel image into a cell nucleus instance segmentation model to accurately segment the cell nucleus and obtain its center coordinates and segmentation confidence. This achieves precise localization of the cell nucleus and quantifies the reliability of segmentation. By classifying the cell nucleus, epithelial cell nuclei and classification confidence are obtained, enabling precise screening of target cell nuclei and quantifying the reliability of classification results. By fusing segmentation and classification confidence to calculate a comprehensive confidence score, the overall reliability of cell nucleus analysis results is comprehensively evaluated, thus improving the accuracy and reliability of epithelial cell nucleus extraction.

[0013] Optionally, the step of inputting the adjusted hematoxylin channel image into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nuclei, center coordinates, and segmentation confidence includes: inputting the adjusted hematoxylin channel image into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nucleus masks and segmentation confidence; calculating the centroid of the cell nucleus masks to obtain center coordinates; and extracting multiple cell nuclei from the cervical cancer pathological whole-section image based on the cell nucleus masks.

[0014] This invention obtains a cell nucleus mask and segmentation confidence score by inputting the adjusted hematoxylin channel image into a segmentation model, accurately delineates the cell nucleus outline and quantifies the confidence score of the segmentation result, obtains the center coordinates by calculating the centroid of the cell nucleus mask, achieves accurate quantification of the cell nucleus position, extracts the cell nucleus from the original image through the mask, accurately preserves the complete morphological features of the cell nucleus, eliminates background interference, and improves the accuracy of cell nucleus extraction and the accuracy of position representation.

[0015] Optionally, classifying the cell nuclei to obtain epithelial cell nuclei and classification confidence includes: acquiring statistical features of epithelial cell nuclei, and constructing a morphological feature penalty term based on the statistical features of epithelial cell nuclei; constructing an optimized loss function based on the morphological feature penalty term, and constructing an epithelial cell nuclei classification model based on the optimized loss function; and classifying the cell nuclei based on the epithelial cell nuclei classification model to obtain epithelial cell nuclei and classification confidence.

[0016] This invention obtains statistical features of epithelial cell nuclei and constructs a morphological feature penalty term, enabling the classification model to learn morphological boundaries that conform to pathological patterns, reducing misclassification of non-epithelial cell nuclei samples. By combining the penalty term to construct an optimized loss function and train the classification model, the model can balance the classification results with the conformity of morphological features. Through this model, cell nuclei are classified and confidence scores are output, accurately screening epithelial cell nuclei and quantifying the reliability of the results, thus improving the accuracy and reliability of epithelial cell nuclei classification.

[0017] Optionally, determining the adjacent relationship nodes between the epithelial cell nuclei based on the central coordinates includes: introducing the Delaunay triangulation method, constructing a triangulation mesh for the epithelial cell nuclei based on the central coordinates using the Delaunay triangulation method; extracting triangle edges from the triangulation mesh to obtain initial adjacent relationship nodes; and filtering the initial adjacent relationship nodes to obtain adjacent relationship nodes.

[0018] This invention introduces the Delaunay triangulation method, constructing a triangular mesh based on the central coordinates of the epithelial cell nucleus. This accurately captures the natural adjacency relationships between cell nuclei, conforming to the spatial distribution patterns of pathological tissues. Initial adjacency nodes are obtained by extracting the sides of the triangles, and a preliminary topological connection between cell nuclei is established. By screening the initial adjacency nodes and eliminating invalid and redundant connections, the accuracy of adjacency relationships is optimized, improving the rationality and accuracy of determining the adjacency nodes of epithelial cell nuclei.

[0019] Optionally, the step of filtering the initial adjacent nodes to obtain adjacent nodes includes: calculating the Euclidean distance between the nodes based on the initial adjacent nodes; obtaining a maximum distance threshold; and filtering the nodes based on the maximum distance threshold using the Euclidean distance to obtain adjacent nodes.

[0020] This invention calculates the Euclidean distance between nodes based on initial adjacent nodes, accurately quantifies the spatial distance between cell nuclei, provides an objective basis for adjacency screening, and eliminates invalid connections that exceed reasonable pathological distances by setting a maximum distance threshold and screening nodes by Euclidean distance, thereby avoiding redundant topological interference and improving the accuracy and effectiveness of adjacent nodes of epithelial cell nuclei.

[0021] Optionally, the step of inputting the epithelial cell nuclear map structure data and the comprehensive confidence score into the cervical cancer classification model to obtain the classification result includes: calculating the normalized similarity value of the connecting edges in the epithelial cell nuclear map structure data based on the initial feature vector; calculating the normalized reliability value of the connecting edges based on the comprehensive confidence score; calculating the attention weight of the connecting edges based on the normalized similarity value and the normalized reliability value; aggregating the epithelial cell nuclear map structure data into a graph-level feature vector based on the attention weight; and inputting the graph-level feature vector into the cervical cancer classification model to obtain the classification result.

[0022] This invention calculates the normalized similarity value of connecting edges using initial feature vectors to accurately capture node feature associations. It also calculates the normalized reliability value using comprehensive confidence to quantify the reliability of edges. The two are then combined to obtain attention weights, allowing the model to focus on effective associations. Based on these weights, the graph structure data is aggregated into graph-level feature vectors, condensing global effective information and improving the accuracy of cervical cancer classification models.

[0023] Another aspect of the present invention provides a classification system based on cervical cancer pathological images, including an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes the computer-readable storage medium described in the preceding aspect of the present invention, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions.

[0024] The classification system based on cervical cancer pathological images of the present invention has a compact structure, stable performance, high integration and simple composition. It can stably execute the steps of the program instructions in the computer-readable storage medium provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0025] Figure 1 This is a flowchart of program instructions in a computer-readable storage medium according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a classification system based on cervical cancer pathological images according to an embodiment of the present invention. Detailed Implementation

[0026] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0027] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0028] Please see Figure 1 In one embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the following steps: Step S1: Obtain a whole slide image of cervical cancer pathology, and perform instance segmentation on the whole slide image of cervical cancer pathology to obtain multiple epithelial cell nuclei, center coordinates, and comprehensive confidence scores.

[0029] In this embodiment, a complete pathological slide image of cervical cancer is first acquired. This image is a digital pathological slide stained with hematoxylin and eosin, typically stored in high resolution in a multi-level pyramid format, containing the complete cervical tissue region. In the complete pathological slide image of cervical cancer, the cell nucleus is stained blue-purple by hematoxylin, and the cytoplasm and stroma are stained pink by eosin, clearly showing the morphology and structure of the epithelial cell nucleus, chromatin distribution, nucleocytoplasmic ratio, and the spatial arrangement between cell nuclei.

[0030] The process of segmenting the whole cervical cancer pathology image into multiple epithelial cell nuclei, center coordinates, and overall confidence scores specifically includes the following sub-steps: Step S101: Perform color space conversion and staining channel separation processing on the whole cervical cancer pathological slide image to obtain a hematoxylin channel image.

[0031] In this embodiment, the whole-section image of cervical cancer pathology is converted from the RGB color space to the HED color space, where the H channel corresponds to hematoxylin staining, the E channel corresponds to eosin staining, and the D channel corresponds to diaminobenzidine staining. A linear transformation is performed on the three color channels using a color space transformation matrix to separate the H channel as the hematoxylin channel image. In this channel, the nucleus region exhibits deep staining due to hematoxylin staining, the nucleus boundary is clear, the chromatin distribution pattern within the nucleus is preserved, and the cytoplasmic background is significantly suppressed due to the separation by eosin staining, making the morphological features of the nucleus and the chromatin texture more prominent.

[0032] Step S102: Adjust the hematoxylin channel image using a pre-acquired standard staining matrix to obtain the adjusted hematoxylin channel image.

[0033] The process of adjusting the hematoxylin channel image using a pre-acquired standard staining matrix to obtain the adjusted hematoxylin channel image specifically includes the following sub-steps: Step S10201: Calculate the optical density mapping matrix of the hematoxylin channel image.

[0034] In this embodiment, the grayscale value of each pixel in the hematoxylin channel image is converted into an optical density value through a negative logarithmic transformation, i.e. ,in The pixel grayscale values ​​are used to form a density mapping matrix with the same size as the original image. According to the Beer-Lambert law, tissue absorption of light is proportional to dye concentration. This transformation converts the pixel intensity of the original image into density values ​​that are linearly related to dye concentration, eliminating the influence of light intensity variations on subsequent analysis. The nucleus region, due to its deep hematoxylin staining and strong absorption, corresponds to higher values ​​in the density mapping matrix, while the background region corresponds to lower values.

[0035] Step S10202: Perform singular value decomposition on the optical density mapping matrix to obtain the target image concentration matrix.

[0036] In this embodiment, singular value decomposition (SVD) is performed on the optical density mapping matrix, decomposing it into a product of a staining matrix and a concentration matrix. SVD yields a left singular vector matrix, a singular value matrix, and a right singular vector matrix, where the left singular vector corresponds to the staining direction and the right singular vector corresponds to the concentration component. Analyzing the distribution of singular values, the first two principal components are selected as effective staining components. The target image concentration matrix is ​​extracted from the decomposition results. Each element in this concentration matrix represents the relative concentration of dye at the corresponding pixel, reflecting the degree of chromatin aggregation and distribution pattern within the cell nucleus.

[0037] Step S10203: Obtain the standard staining matrix, and calculate the normalized optical density mapping matrix using the standard staining matrix and the target image density matrix.

[0038] In this embodiment, a pre-calculated and stored standard staining matrix is ​​obtained. This matrix originates from a pre-selected standard hematoxylin channel image, representing a uniform ideal staining style. The target image density matrix is ​​multiplied by the standard staining matrix to reconstruct a standardized optical density mapping matrix. Each column of the standard staining matrix represents a standardized dye color vector, and the elements in the density matrix represent the relative dye concentration at each pixel. This operation corrects color shifts in the original image caused by differences in staining processes, production batches, and scanning equipment to the standard staining style, aligning images from different sources in the color space.

[0039] Step S10204: Perform inverse spatial transformation on the standardized optical density mapping matrix to obtain the adjusted hematoxylin channel image.

[0040] In this embodiment, an inverse spatial transformation is performed on the standardized optical density mapping matrix. Each pixel value in the matrix is ​​then exponentially plotted to a base-natural constant to obtain its corresponding pixel intensity value. This intensity value is then linearly mapped to a grayscale range of 0 to 255, forming an adjusted hematoxylin channel image. In this adjusted hematoxylin channel image, the contrast between the nucleus region and the background region remains consistent with the standard reference image. Color shifts in the original image caused by differences in staining processes, slide batches, and scanning equipment are effectively corrected, providing a high-quality input image with a consistent style and suitable contrast for subsequent nucleus instance segmentation.

[0041] Step S103: Perform instance segmentation on the adjusted hematoxylin channel image to obtain the epithelial cell nucleus, center coordinates, and overall confidence level.

[0042] Specifically, the process of segmenting the adjusted hematoxylin channel image to obtain the epithelial cell nucleus, center coordinates, and overall confidence score includes the following sub-steps: Step S10301: Input the adjusted hematoxylin channel image into the pre-constructed cell nucleus instance segmentation model to obtain multiple cell nuclei, center coordinates, and segmentation confidence.

[0043] Specifically, inputting the adjusted hematoxylin channel image into a pre-constructed cell nucleus instance segmentation model yields multiple cell nuclei, center coordinates, and segmentation confidence scores, including: Step S1030101: Input the adjusted hematoxylin channel image into the pre-constructed cell nucleus instance segmentation model to obtain multiple cell nucleus masks and segmentation confidence scores.

[0044] In this embodiment, a pre-constructed cell nucleus instance segmentation model is obtained. This model is based on a deep convolutional neural network architecture, using Mask R-CNN as the base network, and is adapted for training based on the characteristics of hematoxylin channel images. Before training, single-channel hematoxylin channel images are copied into three-channel images to match the input requirements of Mask R-CNN, or the number of input channels of the first layer convolutional kernel of the network is modified to 1 and the corresponding parameters are initialized. During the training phase, multiple hematoxylin channel images after staining and normalization are collected as training data. Pathologists perform pixel-level annotation of cell nuclei in each image to generate corresponding cell nucleus mask labels. During training, the adjusted hematoxylin channel images are used as input. Multi-scale feature maps are extracted by a feature extraction network, candidate cell nucleus regions are generated by a region proposal network, features of each candidate region are extracted by a ROI alignment layer, and finally, the category confidence of each candidate region is output through a classification branch, and the pixel-level segmentation mask of each candidate region is output through a mask branch. The model is optimized using a joint loss function, including classification loss, bounding box regression loss, and mask loss. After training, the adjusted hematoxylin channel image obtained in the current step is input into the model. After forward propagation, the model outputs multiple nucleus masks and corresponding segmentation confidence scores. Each nucleus mask is a binary image that marks the precise contour of the nucleus. The segmentation confidence score indicates the reliability of the model for the detection result.

[0045] Step S1030102: Calculate the centroid of the cell nucleus mask to obtain the center coordinates.

[0046] In this embodiment, each cell nucleus mask is treated as a binary image, where the pixel value within the mask area is 1 and the pixel value within the background area is 0. The center coordinates of the cell nucleus are obtained by summing and averaging the coordinates of all pixels within the mask area, i.e., by calculating the average of the x-coordinates and y-coordinates of all pixels within the mask area, which are used as the x-coordinates and y-coordinates of the cell nucleus. For each cell nucleus mask, the centroid calculation is performed independently to obtain the center coordinates of the corresponding cell nucleus.

[0047] Step S1030103: Extract multiple cell nuclei from the whole cervical cancer pathology slide image according to the cell nucleus mask.

[0048] In this embodiment, for each cell nucleus mask, the mask is multiplied pixel-by-pixel by the original cervical cancer pathology whole-slice image. Pixel values ​​within the masked area remain unchanged, while pixel values ​​outside the masked area are set to zero, resulting in a precise image region of the cell nucleus. This image region retains only the complete morphology of the cell nucleus, including detailed information such as the nuclear boundary and the distribution of chromatin within the nucleus, while the background area is completely eliminated. The above pixel-by-pixel multiplication operation is performed independently for each cell nucleus mask, resulting in multiple cell nucleus sub-images.

[0049] Step S10302: Classify the cell nuclei to obtain epithelial cell nuclei and classification confidence levels.

[0050] The classification of cell nuclei to obtain epithelial cell nuclei and the classification confidence level specifically includes the following sub-steps: Step S1030201: Obtain the statistical features of epithelial cell nuclei, and construct a morphological feature penalty term based on the statistical features of epithelial cell nuclei.

[0051] In this embodiment, all samples labeled as epithelial cell nuclei are collected from pre-obtained classification training samples. These samples include normal epithelial cell nuclei and cancerous epithelial cell nuclei. Pre-defined morphological features are extracted for each sample, including nucleus area, roundness, and aspect ratio. For each morphological feature, its value distribution is statistically analyzed to obtain the lower and upper limits of that morphological feature, forming the statistical feature range of epithelial cell nuclei. After obtaining the statistical feature range of epithelial cell nuclei, a morphological feature penalty term is constructed. This penalty term is used during the classification model training process. When a sample is predicted as an epithelial cell nucleus and any of its morphological features exceeds the corresponding statistical feature range, a penalty proportional to the prediction probability is applied to the sample, thereby guiding the model to learn a classification decision boundary that conforms to the morphological rules of epithelial cell nuclei.

[0052] The morphological feature penalty term satisfies the following formula: in, This is a penalty term for morphological features. The number of training samples for classification. For the first The probability that a training sample is predicted to be an epithelial cell nucleus. For indicator functions, The number of morphological features. For the first The nth classification training sample morphological characteristics, For the first The lower limit of a morphological characteristic. For the first The upper limit of a morphological feature.

[0053] The above formula is used to impose constraints on predictions that do not conform to the morphological patterns of epithelial cell nuclei during the training of the classification model. For each classification training sample, it checks whether each morphological feature of the sample exceeds the range of epithelial cell nucleus morphological features, that is, to determine whether there is at least one morphological feature. Less than the lower limit of this feature Or greater than the upper limit of this feature If any feature is out of range, the indicator function... A return value of 1 indicates that the sample has abnormal morphological features; otherwise, a return value of 0 indicates that all morphological features are within the normal range. The model then predicts the probability that the sample belongs to an epithelial cell nucleus. When the probability is greater than 0.5, it indicates that the model classifies the sample as an abnormal epithelial cell nucleus and should be penalized. If it is less than 0.5, the indicator function is 0, indicating no penalty. This is because when the probability is less than 0.5, it essentially predicts a non-epithelial cell nucleus, so no penalty is needed. When penalty is applied, the predicted probability is... Multiplying by the aforementioned indicator function ensures that a penalty is applied only when a sample is predicted to be an epithelial cell nucleus and exhibits abnormal morphological features. The penalty strength is proportional to the model's prediction confidence; the more confident the model is that the sample is an epithelial cell nucleus, the greater the penalty. The summation is then applied to all samples and divided by the total number of samples. Normalization is performed to obtain the final morphological feature penalty term. This penalty term, as part of the optimized loss function, participates in model training, prompting the model to not only focus on the predicted results during classification, but also to consider whether the morphological features of the predicted results conform to the pathological patterns of epithelial cell nuclei, thereby improving the accuracy and reliability of the classification model.

[0054] Step S1030202: Construct an optimized loss function based on the morphological feature penalty term, and construct an epithelial cell nuclear classification model based on the optimized loss function.

[0055] In this embodiment, a convolutional neural network is used as the backbone network when constructing the epithelial cell nucleus classification model. This network takes cell nucleus images as input and extracts image features step-by-step through multiple convolutional layers, pooling layers, and fully connected layers. Finally, the probability that the cell nucleus is an epithelial cell nucleus is output through a sigmoid activation function. The network structure can adopt lightweight architectures such as ResNet-18 or MobileNetV3 to adapt to the rapid processing of a large number of cell nuclei in pathological images. During the training phase, each cell nucleus image and its corresponding ground truth label are obtained from the training dataset. (1 represents the epithelial cell nucleus, 0 represents the non-epithelial cell nucleus), the image is input into the network to obtain the predicted probability. Define the basic loss function. The cross-entropy term is used for binary classification, measuring the difference between the predicted and the true labels, while a morphological feature penalty term is constructed. Samples predicted as epithelial cell nuclei but exhibiting abnormal morphological features are penalized. The weighted sum of these two factors yields the optimized loss function, which is then trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. The training is iterated multiple times until the loss value converges. During training, backpropagation continuously updates the network parameters, enabling the model to learn decision boundaries that conform to the morphological patterns of epithelial cell nuclei while correctly classifying them. The resulting epithelial cell nucleus classification model is then obtained.

[0056] The optimized loss function satisfies the following formula: in, To optimize the loss function, Based on the loss function, As a penalty weight, This is a penalty term for morphological features.

[0057] in, Based on the loss function, The number of training samples for classification. For the first The true labels of each classification training sample Represents the epithelial cell nucleus. Indicates non-epithelial cell nuclei. For the first The probability that a training sample is predicted to be an epithelial cell nucleus.

[0058] Step S1030203: Based on the epithelial cell nucleus classification model, classify the cell nuclei to obtain the epithelial cell nuclei and classification confidence.

[0059] In this embodiment, for each cell nucleus sub-image, it is input into an epithelial cell nucleus classification model. This model is a binary classification model, outputting the probability of epithelial cell nucleus category and the probability of non-epithelial cell nucleus category. The sum of the two probabilities is 1. The two probability values ​​are compared. If the probability of epithelial cell nucleus category is greater than the probability of non-epithelial cell nucleus category, the cell nucleus is classified as an epithelial cell nucleus, and the probability value of epithelial cell nucleus category is taken as the classification confidence value and output. If the probability of epithelial cell nucleus category is less than or equal to the probability of non-epithelial cell nucleus category, it is classified as a non-epithelial cell nucleus, and the cell nucleus is not output. The above determination is performed on all cell nucleus sub-images, and only those classified as epithelial cell nuclei and their corresponding classification confidence values ​​are retained.

[0060] Step S10303: Calculate the overall confidence level based on the classification confidence level and the segmentation confidence level.

[0061] In this embodiment, for each nucleus identified as an epithelial cell nucleus, its segmentation confidence score and classification confidence score are obtained. The segmentation confidence score and classification confidence score are multiplied to obtain a product value. This product value reflects the comprehensive reliability of the cell nucleus in terms of both segmentation quality and classification reliability, i.e., the comprehensive confidence score.

[0062] To prevent the introduction of noise, for each nucleus identified as an epithelial cell, if its overall confidence level is lower than a preset threshold (e.g., 0.1), the nucleus can be removed and not included in the subsequent graph structure construction.

[0063] Step S2: Extract the morphological and depth features of each epithelial cell nucleus, and concatenate the morphological and depth features to form an initial feature vector.

[0064] In this embodiment, morphological features of each epithelial cell nucleus are first extracted. Multiple morphological indicators are calculated based on the nucleus mask, including nucleus area, roundness, aspect ratio (ratio of the longer to shorter side of the circumscribed rectangle), and chromatin distribution uniformity (variance of pixel grayscale values ​​within the mask). Simultaneously, depth features are extracted. The nucleus sub-image is scaled to a fixed size (e.g., 64×64 pixels) and input into a pre-trained convolutional neural network (e.g., ResNet-18) on ImageNet or pathological image datasets. The feature vector obtained by global average pooling of the feature map before the fully connected layers is taken as the depth feature. The extracted morphological feature vector and the depth feature vector are concatenated end-to-end to form the initial feature vector of the cell nucleus. This initial feature vector integrates interpretable pathological morphological information and abstract features automatically learned by the deep network, providing rich representations for node attributes in the subsequent graph structure.

[0065] Step S3: Determine the adjacent relationship nodes between the epithelial cell nuclei based on the central coordinates.

[0066] The determination of the adjacent relationship nodes between the epithelial cell nuclei based on the central coordinates specifically includes the following sub-steps: Step S301: Delaunay triangulation method is introduced, and a triangulation mesh is constructed for the epithelial cell nucleus based on the center coordinates using the Delaunay triangulation method.

[0067] In this embodiment, each epithelial cell nucleus is first mapped one-to-one with its central coordinates. These central coordinates are then used as a set of planar points to construct a triangulated mesh for these nuclei. The Delaunay triangulation method is employed to ensure that the circumcircle of each triangle does not contain any other points. By connecting adjacent points to form a set of non-overlapping triangles, the convex hull region of all point sets is covered, resulting in the triangulated mesh. In this triangulated mesh, each edge connects two spatially adjacent epithelial cell nuclei.

[0068] Delaunay triangulation was used to construct triangular meshes for epithelial cell nuclei because the spatial arrangement of cell nuclei is a crucial factor in cervical cancer pathology diagnosis, as cancerous epithelial cell nuclei often exhibit spatial distribution characteristics such as aggregation, crowding, and loss of normal polarity. Delaunay triangulation, a classic computational geometry method, automatically constructs a non-overlapping triangular mesh connecting adjacent points based on the spatial location of a point set. The circumcircle of each triangle contains no other points, ensuring that each edge connects to the spatially nearest cell nucleus. This proximity-based construction aligns with the natural perception of cell nuclei spatial arrangement by pathologists, effectively capturing local adjacency relationships between nuclei and providing a reasonable topological foundation for subsequent construction of graphical data reflecting tissue structure.

[0069] Step S302: Extract the triangle edges from the triangular mesh to obtain the initial adjacent relationship nodes.

[0070] In this embodiment, each triangle in the triangulation mesh is traversed, and the three edges of the triangle are extracted. All the extracted edges are merged into an edge set, and duplicate edges in the set are deduplicated (i.e., the same edge is only kept once) to obtain the initial adjacency node. Each edge in the initial adjacency node connects two spatially adjacent epithelial cell nuclei, reflecting the initial spatial adjacency relationship between cell nuclei.

[0071] Step S303: Filter the initial adjacent relationship nodes to obtain adjacent relationship nodes.

[0072] The process of filtering the initial adjacent nodes to obtain adjacent nodes specifically includes the following sub-steps: Step S30301: Calculate the Euclidean distance between nodes based on the initial adjacent relationship nodes.

[0073] In this embodiment, for each edge in the initial adjacent relationship node, the edge connects two epithelial cell nuclei. The center coordinates of the two epithelial cell nuclei are obtained respectively, and the Euclidean distance between the two center coordinates is calculated. The calculated Euclidean distance value is used as the Euclidean distance of the edge. The above calculation is performed on each edge in the initial adjacent relationship node to obtain the Euclidean distance corresponding to each edge, that is, the node Euclidean distance.

[0074] Step S30302: Obtain the maximum distance threshold, and filter the nodes by Euclidean distance based on the maximum distance threshold to obtain adjacent relationship nodes.

[0075] In this embodiment, the average diameter of cell nuclei is calculated from all cell nucleus masks in the current image, and a maximum distance threshold of 2 to 3 times this average diameter is used. Cell nuclei in pathological images typically exhibit an approximately elliptical shape, and their diameter reflects their size. The distance between adjacent cell nuclei generally does not exceed 2 to 3 times the diameter of the cell nucleus. Two cell nuclei exceeding this range no longer have a direct biological correlation in space (e.g., belonging to different glandular structures or tissue regions). The Euclidean distance of each edge is compared with this maximum distance threshold. If the Euclidean distance is greater than the maximum distance threshold, the edge is discarded; if the Euclidean distance is less than or equal to the maximum distance threshold, the edge is retained. This filtering operation is performed on all edges in the initial adjacent relationship nodes. The retained edges constitute adjacent relationship nodes. Each edge in these adjacent relationship nodes connects two epithelial cell nuclei that are spatially adjacent and within a reasonable distance, effectively eliminating invalid connections caused by excessively large internuclear distances, making the constructed graph structure more compact and meeting the needs of pathological diagnosis.

[0076] Delaunay triangulation constructs spatial adjacency relationships based solely on the coordinates of the cell nuclei's centers. This method is prone to errors due to uneven cell nucleus distribution, image noise, or isolated points, which can lead to the misgenerating of connecting edges between two epithelial cell nuclei whose spatial distance exceeds the reasonable range for pathological cervical epithelial tissue. Such long-distance connections fail to reflect the true spatial arrangement of epithelial cell nuclei, introducing invalid and redundant topological information and reducing the reliability of the graph structure data. Therefore, by setting a maximum distance threshold, invalid edges with Euclidean distances exceeding the threshold are removed, retaining only adjacency relationships with spatial distances conforming to physiological distribution. This ensures that the constructed epithelial cell nucleus graph structure accurately matches the true topological features of the pathological tissue.

[0077] Step S4: Construct epithelial cell nucleus map structure data based on the epithelial cell nucleus, the initial feature vector, and the adjacent relationship nodes.

[0078] In this embodiment, each epithelial cell nucleus is treated as a node in the graph structure, and the initial feature vector of each node is used as the feature attribute of that node. Each edge in the adjacent relationship nodes is used as the edge connecting the corresponding two nodes in the graph structure. This edge represents the spatial adjacency relationship between the two epithelial cell nuclei. All nodes and all edges are combined to form the epithelial cell nucleus graph structure data. This graph structure data uses cell nuclei as nodes and spatial adjacency relationships as edges, which fully preserves the individual characteristics and local topological relationships of the epithelial cell nuclei, providing structured input data for subsequent graph neural network models.

[0079] Step S5: Construct a cervical cancer classification model by inputting the epithelial cell nuclear structure data and the comprehensive confidence score into the cervical cancer classification model to obtain the classification result.

[0080] In this embodiment, a graph attention network is used as the core architecture of the cervical cancer classification model to process epithelial cell nuclear map structure data. The input to the cervical cancer classification model is the epithelial cell nuclear map structure data and the comprehensive confidence score of each node. The graph structure data includes the initial feature vector of each node and the set of edges. The cervical cancer classification model first updates the node features through multiple graph attention layers. In each layer, the normalized similarity value of the edges is calculated based on the initial feature vector of the node, and the normalized reliability value of the edges is calculated based on the comprehensive confidence score. The two are combined to obtain the attention weight of the edges, and the features of neighboring nodes are weighted and aggregated based on this attention weight. After passing through a nonlinear activation function, the updated node features are obtained. After completing the multi-layer feature update, all node features are aggregated into a graph-level feature vector through global average pooling. Finally, the graph-level feature vector is input into a classification head composed of a fully connected layer and a Softmax activation function, and the probabilities of five categories are output, corresponding to normal (no lesions), CIN1 (cervical intraepithelial neoplasia grade 1, mild atypia), CIN2 (cervical intraepithelial neoplasia grade 2, moderate atypia), CIN3 (cervical intraepithelial neoplasia grade 3, severe atypia / carcinoma in situ), and invasive carcinoma (cancer cells break through the basement membrane).

[0081] Training samples were constructed from an annotated dataset: for each cervical cancer pathology whole-slice image, the corresponding epithelial cell nuclear structure data was obtained through the aforementioned steps, and pathologists labeled the image with one of five categories (normal, CIN1, CIN2, CIN3, invasive carcinoma). During model training, a multi-class cross-entropy loss function was used as the basic classification loss, with the Adam optimizer employed. The initial learning rate was set to 0.001, the batch size to 16, and the training epochs to 100. An early stopping strategy was used on the validation set to prevent overfitting. After training, the model outputs the probabilities of the five categories for the input graph structure data, and the category with the highest probability is taken as the classification result to determine the lesion grade of the cervical cancer pathology whole-slice image.

[0082] The specific steps for inputting the epithelial cell nuclear structure data and the comprehensive confidence score into the cervical cancer classification model to obtain the classification result include the following: Step S501: Calculate the normalized similarity value of the connecting edges in the epithelial cell nuclear map structure data based on the initial feature vector.

[0083] In this embodiment, the similarity value is calculated using a learnable module defined internally within the graph attention network model, rather than an externally preset fixed function. For connected nodes... and nodes The model first sets the initial feature vectors of the two nodes to the edge. and Each through learnable feature transformation matrix Perform a linear transformation to obtain the transformed features. and Then, the two transformed feature vectors are concatenated and passed through a learnable attention vector. A linear mapping is performed to obtain a scalar value; finally, this scalar value is processed by the LeakyReLU (negative_slope=0.2) activation function to obtain the similarity value. Throughout the calculation process, the characteristic transformation matrix... and attention vector These are all model parameters, which are continuously optimized through backpropagation during training, enabling the model to learn the most suitable similarity metric for the current classification task. Therefore, the calculation of similarity values ​​is learned autonomously by the model, rather than by fixed rules preset by humans.

[0084] Then, for each node Softmax normalization is applied to all neighboring edges to obtain normalized similarity values. This makes the node The sum of the normalized similarity values ​​of all adjacent edges is 1. The normalized similarity value satisfies the following formula: in, For connecting nodes and nodes The normalized similarity value of the edges, For connecting nodes and nodes The similarity value of the edges, For nodes The set of neighboring nodes, for Neighbor nodes in, For learnable attention vectors, The feature transformation matrix is ​​a learnable feature. For nodes The initial feature vector, For nodes The initial eigenvectors. The eigentransformation matrix. Attention vector Joint optimization is performed through backpropagation of the entire classification model.

[0085] Step S502: Calculate the normalized reliability value of the connection edge based on the comprehensive confidence level.

[0086] In this embodiment, the reliability value is used to measure the overall trustworthiness of the two cell nuclei connected by an edge during the segmentation and classification process. For connecting nodes... and nodes The edge, which combines the confidence scores of the two nodes. and Multiply by the learnable reliability coefficient. The original reliability value is obtained. The overall confidence level reflects the overall reliability of the cell nucleus from instance segmentation to classification results. The product form means that a lower confidence level in either node will lead to a decrease in edge reliability, consistent with the intuition that "both nodes connected by an edge must be reliable." Learnable coefficients. This allows the model to adaptively adjust the weight of reliability in the overall attention calculation. Then, for each node... Softmax normalization is performed on all adjacent edges to obtain the normalized reliability value. This makes the node The sum of the normalized reliability values ​​of all neighboring edges is 1, which is on the same order of magnitude as the normalized similarity value.

[0087] The overall confidence level indirectly affects the aggregation of node features by adjusting the edge weights, so that the cell nuclei with high segmentation quality and classification confidence can make a greater contribution to the update of neighbor features, thereby realizing the dominant role of reliable nodes in graph-level features.

[0088] The normalized reliability value satisfies the following formula: in, For connecting nodes and nodes The normalized reliability value of the edge, For connecting nodes and nodes The reliability value of the edge, For nodes The set of neighboring nodes, for Neighbor nodes in, The reliability coefficient is a learnable value. For nodes The overall confidence level, For nodes The overall confidence level. Learnable coefficients. and , Synchronous updates and adaptive adjustment of reliability weights.

[0089] Step S503: Calculate the attention weight of the connection edge based on the normalized similarity value and the normalized reliability value.

[0090] In this embodiment, the attention weight integrates information from two dimensions: feature similarity and segmentation / classification reliability. For each edge, its similarity value is normalized. With normalized reliability value Add them together to get the merged value. Then, for each node... The fused values ​​of all neighboring edges are Softmax normalized to obtain the final edge attention weights. This makes the node The sum of the attention weights of all neighboring edges is 1. This design ensures that an edge receives a high attention weight only when the features of the two nodes are similar and the combined confidence of the two nodes is high. The addition operation is simple and effective, avoiding complex parameterized fusion, and since the two components have already been normalized, the sum can still be naturally processed by Softmax. This attention weight is directly used in the subsequent node feature aggregation process, guiding the model to pay more attention to reliable and similar neighboring nodes when aggregating neighbor information.

[0091] Attention weights satisfy the following formula: in, For connecting nodes and nodes Attention weights of edges For nodes The set of neighboring nodes, for Neighbor nodes in, For connecting nodes and nodes The normalized similarity value of the edges, For connecting nodes and nodes The normalized reliability value of the edge.

[0092] It should be noted that similarity and reliability values ​​must be normalized before they can be added together; otherwise, the inconsistency in the units of measurement will cause one to be suppressed, thereby weakening its influence on attention weight.

[0093] Step S504: Based on the attention weights, aggregate the epithelial cell nuclear map structure data into a graph-level feature vector.

[0094] In this embodiment, after obtaining the attention weights for each edge, the features of each node are first updated. , all its neighbor nodes initial feature vector Through the characteristic transformation matrix Perform a linear transformation, then apply the edge attention weights. Perform weighted summation, and finally pass through the activation function. (e.g., ELU) obtains nodes Updated features This aggregation method allows each node to adaptively collect information from its neighbors, with neighbors having higher attention weights contributing more to node updates. After all nodes have completed their updates, global average pooling (or global max pooling) aggregates the updated node features into a graph-level feature vector. This graph-level feature vector condenses the global information of the entire cell nucleus graph structure, and finally, inputting it into a classification head yields the classification results for the whole-section pathological image of cervical cancer.

[0095] The node update formula satisfies the following formula: in: It is a node Updated features It is a node The set of neighboring nodes, For connecting nodes and nodes Attention weights of edges It is the characteristic transformation matrix.

[0096] The overall confidence level is primarily used to adjust the contribution of a node to the aggregation of neighbor features, rather than restricting its ability to receive information from neighbors. This design allows low-confidence nodes to absorb information from high-confidence neighbors in subsequent aggregations, thus providing an opportunity to correct their own features. This aligns with the pathological logic that cell nucleus segmentation results can be corrected through the features of surrounding cell nuclei.

[0097] Step S505: Input the graph-level feature vector into the cervical cancer classification model to obtain the classification result.

[0098] In this embodiment, the graph-level feature vector is input into the classification head of the cervical cancer classification model. This classification head consists of a fully connected layer and a Softmax activation function. The fully connected layer maps the graph-level feature vector to raw scores for five categories. The Softmax function converts these scores into probability distributions for the five categories, corresponding to the probabilities of normal, CIN1, CIN2, CIN3, and invasive carcinoma, respectively. The category corresponding to the maximum value among the five probabilities is taken as the final classification result of the cervical cancer pathology whole-slice image, and the lesion level is output. This classification result can be directly used to assist pathologists in diagnosis.

[0099] like Figure 2As shown, in another aspect, the present invention also provides a classification system based on cervical cancer pathological images, including an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes the computer-readable storage medium described above, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions.

[0100] In this embodiment, the input device is used to provide the system with relevant input data or instructions. In a classification system based on cervical cancer pathology images, the input device may include common human-computer interaction interface devices such as keyboards, mice, and touchscreens. Through the input device, doctors or researchers can input necessary parameters.

[0101] The processor is the core component of the system, responsible for executing computer program instructions and performing data processing and analysis. In a classification system based on cervical cancer pathology images, the processor analyzes and interprets the input test data by running pre-programmed algorithms and models. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other dedicated processing unit.

[0102] The memory is used to store computer programs, data, and parameters required by the system. It may include random access memory (RAM) for temporary data storage and processing, and persistent memory (such as hard disks or solid-state drives) for long-term data storage and preservation.

[0103] The output device is used to present the results of system processing and analysis to users or external devices. The output device can be a monitor, printer, charting device, etc. Through the output device, the system can display the prediction results, which can be used as a reference for doctors, researchers, or patients to assist in decision-making and communication.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the following steps: Obtain whole-section images of cervical cancer pathology, and perform instance segmentation on the whole-section images of cervical cancer pathology to obtain multiple epithelial cell nuclei, center coordinates, and comprehensive confidence scores; The morphological and depth features of each epithelial cell nucleus are extracted, and the morphological and depth features are concatenated to form an initial feature vector; The adjacent relationship nodes between the epithelial cell nuclei are determined based on the central coordinates; Construct epithelial cell nucleus map structure data based on the epithelial cell nucleus, the initial feature vector, and the adjacent relationship nodes; A cervical cancer classification model is constructed by inputting the epithelial cell nuclear structure data and the comprehensive confidence score into the cervical cancer classification model to obtain the classification result.

2. The computer-readable storage medium according to claim 1, characterized in that, The step of segmenting the whole cervical cancer pathological image to obtain multiple epithelial cell nuclei, center coordinates, and overall confidence scores includes: The cervical cancer pathological whole slide image was subjected to color space conversion and staining channel separation processing to obtain the hematoxylin channel image; The hematoxylin channel image is adjusted using a pre-acquired standard staining matrix to obtain the adjusted hematoxylin channel image; The adjusted hematoxylin channel image was segmented to obtain the epithelial cell nucleus, center coordinates, and overall confidence level.

3. The computer-readable storage medium according to claim 2, characterized in that, The adjustment of the hematoxylin channel image using a pre-acquired standard staining matrix to obtain the adjusted hematoxylin channel image includes: Calculate the optical density mapping matrix of the hematoxylin channel image; The target image density matrix is ​​obtained by performing singular value decomposition on the optical density mapping matrix; Obtain a standard staining matrix, and use the standard staining matrix and the target image density matrix to calculate a normalized optical density mapping matrix; The adjusted hematoxylin channel image is obtained by performing an inverse spatial transformation on the standardized optical density mapping matrix.

4. The computer-readable storage medium according to claim 2, characterized in that, The step of performing instance segmentation on the adjusted hematoxylin channel image to obtain the epithelial cell nucleus, center coordinates, and overall confidence score includes: The adjusted hematoxylin channel image is input into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nuclei, center coordinates, and segmentation confidence. The cell nuclei were classified to obtain epithelial cell nuclei and classification confidence levels; The overall confidence level is calculated based on the classification confidence level and the segmentation confidence level.

5. A computer-readable storage medium according to claim 4, characterized in that, The step of inputting the adjusted hematoxylin channel image into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nuclei, center coordinates, and segmentation confidence includes: The adjusted hematoxylin channel image is input into a pre-constructed cell nucleus instance segmentation model to obtain multiple cell nucleus masks and segmentation confidence scores; The centroid coordinates of the cell nucleus mask are obtained by centroid calculation; Multiple cell nuclei were extracted from the whole slide image of cervical cancer pathology based on the described nuclear mask.

6. The computer-readable storage medium according to claim 4, characterized in that, The classification of the cell nuclei to obtain epithelial cell nuclei and the classification confidence level include: Statistical features of epithelial cell nuclei are obtained, and morphological feature penalty terms are constructed based on the statistical features of epithelial cell nuclei. An optimized loss function is constructed based on the morphological feature penalty term, and an epithelial cell nucleus classification model is constructed based on the optimized loss function; The cell nuclei are classified based on the epithelial cell nucleus classification model to obtain the epithelial cell nuclei and classification confidence levels.

7. The computer-readable storage medium according to claim 1, characterized in that, The step of determining the adjacent relationship nodes between the epithelial cell nuclei based on the central coordinates includes: The Delaunay triangulation method is introduced, and a triangulation grid is constructed for the epithelial cell nucleus based on the central coordinates using the Delaunay triangulation method; Extract the triangle edges from the triangular mesh to obtain the initial adjacent relationship nodes; The initial adjacent relationship nodes are filtered to obtain adjacent relationship nodes.

8. The computer-readable storage medium according to claim 7, characterized in that, The process of filtering the initial adjacent relationship nodes to obtain adjacent relationship nodes includes: Calculate the Euclidean distance between nodes based on the initial adjacent node relationships; Obtain the maximum distance threshold, and filter the nodes by Euclidean distance based on the maximum distance threshold to obtain adjacent nodes.

9. A computer-readable storage medium according to claim 1, characterized in that, The step of inputting the epithelial cell nuclear structure data and the comprehensive confidence score into the cervical cancer classification model to obtain the classification result includes: The normalized similarity value of the connecting edges in the epithelial cell nucleus map structure data is calculated based on the initial feature vector; The normalized reliability value of the connecting edge is calculated based on the overall confidence level; The attention weights of the connecting edges are calculated based on the normalized similarity values ​​and the normalized reliability values. Based on the attention weights, the epithelial cell nuclear map structure data is aggregated into a graph-level feature vector; The graph-level feature vector is input into the cervical cancer classification model to obtain the classification result.

10. A classification system based on whole-section images of cervical cancer pathology, characterized in that, The device includes an input device, a processor, an output device, and a memory, wherein the input device, processor, output device, and memory are interconnected, the memory includes a computer-readable storage medium as described in any one of claims 1 to 9, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions.

Images