Artificial intelligence-based colposcope collection image comparison method and system

Abstract

The invention provides a colposcope collection image comparison method and system based on artificial intelligence, and the method comprises the steps: carrying out the mixed noise reduction and multi-scale enhancement processing of a 4K original image collected by a colposcope, and obtaining an enhanced noise reduction image; adaptively adjusting the scaling of the enhanced and denoised image based on image feature distribution, and focusing columnar epithelium and squamous epithelium cell features by using an attention convolutional network to obtain a feature focusing image; cross-modal feature fusion segmentation processing is carried out on the feature focusing image, columnar epithelium and squamous epithelium regions are segmented, and segmented region images are obtained; image convolution and topological feature extraction are carried out on the segmented region image, an image adjacent image is constructed, and an image feature vector is obtained; and performing fuzzy similarity measurement and comparison decision processing on the image feature vector to obtain a colposcope acquisition image comparison result. According to the invention, the defect that the consistency and the accuracy are difficult to guarantee when the colposcope collection image is identified manually at present is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly relates to a method and system for comparing colposcopy-acquired images based on artificial intelligence. Background Art

[0002] A colposcope can magnify the images of parts such as the cervix and vagina, enabling doctors to directly observe the epithelial and vascular morphology of these parts, and thus assisting in judging whether there are lesions.

[0003] Traditional colposcopy image analysis methods mainly rely on doctors' naked-eye observation and subjective experience. Doctors need to rely on their professional knowledge to identify tissue characteristics such as columnar epithelium and squamous epithelium in the images to judge the lesion situation. However, this method has significant limitations.

[0004] On the one hand, due to limited image clarity, subtle cellular-level changes are difficult to detect, and the sensitivity for detecting early lesions is relatively low. On the other hand, there are differences in the experience and judgment criteria of different doctors, resulting in difficulties in ensuring the consistency and accuracy of diagnostic results. According to relevant clinical research statistics, the consistency of the recognition results of different doctors for the same batch of colposcopy images is only about 60%-70%. Summary of the Invention

[0005] The main objective of the present invention is to provide a method and system for comparing colposcopy-acquired images based on artificial intelligence, aiming to overcome the defects that it is difficult to ensure consistency and accuracy in the recognition of colposcopy-acquired images relying on manual work currently.

[0006] To achieve the above objective, the present invention provides a method for comparing colposcopy-acquired images based on artificial intelligence, including the following steps:

[0007] Perform hybrid noise reduction and multi-scale enhancement processing on the 4K original images acquired by the colposcope to obtain enhanced and noise-reduced images;

[0008] Based on the image feature distribution, adaptively adjust the scaling ratio of the enhanced and noise-reduced images, and use an attention convolutional network to focus on the columnar epithelium and squamous epithelium cell features to obtain feature-focused images;

[0009] Perform cross-modal feature fusion and segmentation processing on the feature-focused images to segment the columnar epithelium and squamous epithelium regions to obtain segmented region images;

[0010] Perform graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph to obtain image feature vectors;

[0011] Perform fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain the comparison result of the colposcopy-acquired images.

[0012] Further, perform hybrid noise reduction and multi-scale enhancement processing on the 4K original images collected by the colposcope, including:

[0013] Use a combination of non-local means filtering and wavelet transform to reduce noise in the 4K original images, and adopt the multi-scale Retinex algorithm to enhance image details.

[0014] Further, perform cross-modal feature fusion segmentation processing on the feature-focused images, segment the columnar epithelium and squamous epithelium regions, and obtain the segmented region images, including:

[0015] Use qubits to encode the color and texture features of the feature-focused images to generate a quantum feature map;

[0016] Extract the simulated bioelectrical characteristic features of the feature-focused images to generate bioelectrical features;

[0017] Based on the symbiotic evolution algorithm, fuse the quantum feature map and the bioelectrical features to generate a symbiotic fusion feature matrix;

[0018] Divide the symbiotic fusion feature matrix into multiple cells, and simulate the cell growth and boundary formation processes according to the local rules and state transitions between cells, and adaptively segment the columnar epithelium and squamous epithelium regions to obtain the segmented region images.

[0019] Further, perform graph convolution and topological feature extraction on the segmented region images, construct an image adjacency graph, and obtain image feature vectors, including:

[0020] Adopt a graph cut-based superpixel segmentation algorithm to divide the segmented region images into superpixel blocks with semantic consistency to obtain a refined superpixel segmentation image;

[0021] Calculate the color moments, texture features, and shape features of each superpixel block, and combine them as node features to obtain a set of feature nodes; where each superpixel block corresponds to a node in the graph;

[0022] Determine the edges between nodes according to the spatial adjacency and feature similarity between the superpixel blocks; where, for nodes that are spatially adjacent and have a feature similarity higher than the set threshold, establish an edge connection and use the similarity value as the weight of the edge to obtain a preliminary image adjacency graph;

[0023] Apply the graph attention mechanism to assign different attention weights to the adjacent edges of each node to obtain an optimized image adjacency graph;

[0024] Adopt a multi-layer graph convolutional network to extract features from the optimized image adjacency graph to obtain graph convolutional features;

[0025] Perform a global pooling operation on the graph convolution features to obtain the image feature vector.

[0026] Furthermore, perform fuzzy similarity measurement and contrast decision processing on the image feature vector to obtain the colposcopy acquisition image contrast result, including:

[0027] Divide the feature space where the image feature vectors are located into multiple fuzzy clustering centers based on the clustering algorithm, calculate the fuzzy membership degrees of each image feature vector with respect to each clustering center to obtain a fuzzy membership degree matrix;

[0028] Adopt the cosine similarity algorithm to calculate the similarity between different image feature vectors to obtain a set of fuzzy similarity values;

[0029] Obtain a fuzzy decision rule base; the fuzzy decision rule base contains qualitative descriptions of image comparison corresponding to different similarity intervals;

[0030] Based on the Mamdani fuzzy inference method, perform inference according to the fuzzy similarity values and the fuzzy decision rule base to obtain a fuzzy contrast decision result;

[0031] Adopt the centroid method to defuzzify the fuzzy contrast decision result to obtain a clear colposcopy acquisition image contrast result.

[0032] Furthermore, after obtaining the colposcopy acquisition image contrast result, including:

[0033] Obtain the user's account information, and construct a character array based on the account information;

[0034] Extract keywords from the colposcopy acquisition image contrast result, and perform character mapping based on the extracted keywords to obtain key characters;

[0035] Deform the preset coding table based on the key characters to obtain a deformed coding table;

[0036] Construct a text matrix based on the text other than the keywords in the colposcopy acquisition image contrast result;

[0037] Adjust the character array based on the text matrix to obtain an adjusted array;

[0038] Generate a communication key based on the deformed coding table and the adjusted array to encrypt the 4K original image and then send it to the corresponding user account.

[0039] Furthermore, generating a communication key based on the deformed coding table and the adjusted array includes:

[0040] Perform attribute recognition on each character in the adjustment array, and construct a first curve based on the attributes of the recognized characters.

[0041] Obtain the table attributes of the deformation coding table, and generate a second curve based on the table attributes.

[0042] Superimpose the first curve and the second curve onto the deformation coding table according to the rules.

[0043] Obtain the area enclosed by the first curve and the second curve in the deformation coding table as the target area.

[0044] Combine the characters in the target area to obtain the communication key.

[0045] The present invention also provides a comparison system for colposcopy-acquired images based on artificial intelligence, including:

[0046] An enhancement unit for performing hybrid noise reduction and multi-scale enhancement processing on the 4K original image acquired by the colposcope to obtain an enhanced and noise-reduced image.

[0047] A focusing unit for adaptively adjusting the scaling ratio of the enhanced and noise-reduced image based on the image feature distribution, and using an attention convolutional network to focus on the characteristics of columnar epithelium and squamous epithelial cells to obtain a feature-focused image.

[0048] A segmentation unit for performing cross-modal feature fusion segmentation processing on the feature-focused image to segment the columnar epithelium and squamous epithelial regions to obtain a segmented region image.

[0049] An extraction unit for performing graph convolution and topological feature extraction on the segmented region image to construct an image adjacency graph and obtain an image feature vector.

[0050] A comparison unit for performing fuzzy similarity measurement and comparison decision processing on the image feature vector to obtain the comparison result of the colposcopy-acquired image.

[0051] The present invention also provides a computer device, including a memory and a processor, where a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described in any one of the above are implemented.

[0052] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0053] The present invention provides a method and system for comparing colposcopy-acquired images based on artificial intelligence, including: performing hybrid noise reduction and multi-scale enhancement processing on the 4K original images acquired by the colposcope to obtain enhanced and noise-reduced images; adaptively adjusting the scaling ratio of the enhanced and noise-reduced images based on the image feature distribution, and using an attention convolutional network to focus on the features of columnar epithelium and squamous epithelial cells to obtain feature-focused images; performing cross-modal feature fusion segmentation processing on the feature-focused images to segment the columnar epithelium and squamous epithelial regions and obtain segmented region images; performing graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph and obtain image feature vectors; and performing fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain the comparison result of the colposcopy-acquired images. In the present invention, through image enhancement, it is easier to identify the corresponding tissues; and then, relying on cross-modal feature fusion segmentation processing, graph convolution and topological feature extraction, an image adjacency graph is constructed, and then fuzzy similarity measurement and comparison decision processing are performed to obtain the comparison result of the colposcopy-acquired images, making the comparison results unified and improving the comparison accuracy; overcoming the defect that it is difficult to ensure consistency and accuracy in the current identification of colposcopy-acquired images relying on manual work. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a schematic diagram of the steps of a method for comparing colposcopy-acquired images based on artificial intelligence according to an embodiment of the present invention;

[0055] Figure 2 is a block diagram of the structure of a system for comparing colposcopy-acquired images based on artificial intelligence according to an embodiment of the present invention;

[0056] Figure 3 is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0057] The implementation, functional features and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0059] Refer to Figure 1 , an embodiment of the present invention provides a method for comparing colposcopy-acquired images based on artificial intelligence, including the following steps:

[0060] Step S1, performing hybrid noise reduction and multi-scale enhancement processing on the 4K original images acquired by the colposcope to obtain enhanced and noise-reduced images;

[0061] Step S2, adaptively adjust the scaling ratio of the enhanced and denoised image based on the image feature distribution, and use an attention convolutional network to focus on the features of columnar epithelium and squamous epithelial cells to obtain a feature-focused image;

[0062] Step S3, perform cross-modal feature fusion and segmentation processing on the feature-focused image to segment the columnar epithelium and squamous epithelium regions to obtain a segmented region image;

[0063] Step S4, perform graph convolution and topological feature extraction on the segmented region image to construct an image adjacency graph to obtain an image feature vector;

[0064] Step S5, perform fuzzy similarity measurement and comparison decision processing on the image feature vector to obtain a comparison result of the colposcopy-acquired image.

[0065] In this embodiment, as described in step S1 above, hybrid denoising adopts a method of combining multiple denoising algorithms. For example, Gaussian filtering can be used to remove Gaussian noise in the image, and median filtering can be used to eliminate salt-and-pepper noise, etc. Multi-scale enhancement is based on the analysis of image features at different scales. By decomposing and reconstructing the image at different scales, such as using wavelet transform, the details and contours of the image are enhanced in different frequency sub-bands. During the acquisition of the 4K raw image collected by the colposcope, it is prone to be affected by device noise, environmental interference, etc., resulting in noise. At the same time, since the cell and tissue features present different details at different scales, through multi-scale enhancement processing, the cell structure, tissue texture and other features in the image can be comprehensively highlighted, providing a clearer and more accurate data basis for subsequent analysis, improving the image quality, and facilitating the subsequent extraction and analysis of image features.

[0066] As described in step S2 above, the adaptive scaling based on the image feature distribution determines which parts of the image need to be enlarged or reduced by analyzing the feature richness, edge information, etc. in different regions of the image. The attention convolutional network simulates the attention mechanism of human vision. During the convolution process, it automatically learns the importance weights of different regions in the image and pays more attention to important regions (such as the regions where columnar epithelium and squamous epithelial cells are located). In different colposcopy images, the distribution and presentation of columnar epithelium and squamous epithelial cells vary. Adaptive adjustment of the scaling ratio can make the key cell regions in the image be more appropriately displayed, avoiding the loss or unclear of key features caused by a fixed scaling ratio. Using the attention convolutional network to focus on these cell features can highlight the morphology, arrangement and other features of the cells, reduce the interference of irrelevant background information, and provide strong support for the subsequent accurate segmentation of the cell regions.

[0067] As described in step S3 above, cross-modal feature fusion integrates various different types of image features, such as optical features (color, brightness, etc.), texture features (local binary pattern, gray-level co-occurrence matrix, etc.), and deep learning-based semantic features. Through a specific fusion algorithm, these features of different modalities are organically combined, and then the fused features are processed using a segmentation model (such as a segmentation model based on a neural network) to achieve the segmentation of columnar epithelium and squamous epithelium regions. Single-modal image features often cannot comprehensively and accurately describe cell regions. Cross-modal feature fusion can integrate the advantages of multiple features and more comprehensively characterize the characteristics of columnar epithelium and squamous epithelium cell regions. By accurately segmenting these cell regions, complex images are simplified into clear cell region information, providing pure data for subsequent analysis of the topological structure and features of cells, facilitating further in-depth study of cell features and states.

[0068] As described in step S4 above, the graph convolutional network performs convolutional operations on graph-structured data and extracts node features through information transmission and aggregation between nodes. In this step, the segmented region image is constructed into a graph structure, with cell regions as nodes and the spatial relationship or feature similarity between cells as edges. Topological feature extraction focuses on topological attributes such as the shape and connectivity of cell regions. By processing the graph structure with the graph convolutional network, these topological features are extracted and transformed into image feature vectors. The topological structure and interrelationships of cells are of great significance for understanding cell functions and pathological states. By constructing an image adjacency graph and extracting topological features, characteristic information of cells can be deeply mined from the perspective of the spatial layout and interconnections of cells. Transforming these features into image feature vectors facilitates subsequent quantitative analysis and comparison, providing more representative and discriminative feature data for image comparison.

[0069] As described in step S5 above, fuzzy similarity measurement takes into account the uncertainty and fuzziness of image features and uses methods of fuzzy mathematics, such as fuzzy membership functions and fuzzy distance calculations, to measure the similarity degree between different image feature vectors. The comparison decision is based on set fuzzy rules and thresholds, and according to the calculation results of fuzzy similarity, the difference degree between images is judged to obtain a comparison result. In this embodiment, due to the certain fuzziness and uncertainty of colposcopy image features, traditional precise similarity measurement methods are difficult to accurately reflect the true relationship between images. Fuzzy similarity measurement can more reasonably handle this fuzziness and more accurately measure the similarity degree between different colposcopy-acquired images. Through comparison decision processing, the similarity result is transformed into a specific diagnostic conclusion, such as judging whether the tissue corresponding to the image has a lesion and the degree of the lesion, providing an intuitive and accurate basis for image comparison diagnosis for doctors.

[0070] In one embodiment, the 4K original images collected by the colposcope are subjected to hybrid noise reduction and multi-scale enhancement processing, including:

[0071] The 4K original images are denoised by combining non-local mean filtering and wavelet transform, and the multi-scale Retinex algorithm is used to enhance the image details.

[0072] In this embodiment, non-local mean filtering: Its core principle is to use similar local image patches in the image to filter the current pixel. In the 4K original image, other neighborhood patches similar to the neighborhood where the current pixel is located are searched through traversal, and the filtered value of the current pixel is determined by calculating the weighted average of these similar neighborhood patches and the current neighborhood patch. The calculation of the weight is based on the similarity between the neighborhood patches. The higher the similarity, the greater the weight. This method can effectively retain the texture and detail information of the image because it does not only rely on the local neighborhood of the current pixel, but searches for similar structures from the entire image.

[0073] Wavelet transform: It is a time-frequency analysis method that decomposes the image into sub-bands of different frequencies. In the noise reduction process, the noise of the image is usually concentrated in the high-frequency sub-bands, while the main information of the signal is located in the low-frequency sub-bands. By performing threshold processing on the high-frequency sub-bands, the high-frequency coefficients corresponding to the noise are removed or weakened, and then the image is reconstructed through inverse wavelet transform, so as to achieve the purpose of noise reduction. When combined with non-local mean filtering, wavelet transform can further remove the high-frequency noise in the image, while non-local mean filtering can make up for the loss of texture details that may be caused by wavelet transform.

[0074] The multi-scale Retinex algorithm is used to enhance image details. The multi-scale Retinex algorithm is based on the Retinex theory, which holds that the color perception of an object depends not only on the reflection characteristics of the object's surface but also on the lighting environment in which the object is located. The multi-scale Retinex algorithm simulates the lighting components at different scales by performing Gaussian filtering on the image at different scales. At each scale, the original image is divided by the corresponding Gaussian filtering result to obtain the reflection component at that scale. Then, the reflection components at multiple scales are weighted and fused to obtain the final enhanced image. By adjusting the weights of different scales, the detail information of different sizes in the image can be highlighted, and the details from small-scale cellular details to larger-scale tissue region details can be effectively enhanced. The cell and tissue features present rich details at different scales, and these details are crucial for accurately judging the lesion situation. However, these details in the original image may not be clear enough due to uneven lighting, insufficient contrast, etc. The multi-scale Retinex algorithm can effectively enhance the detail information of the image at different scales, improve the contrast and clarity of the image, make the features such as the morphology and texture of cells and the boundaries of tissues more obvious, provide clearer and more distinguishable data for subsequent image analysis, such as feature extraction and region segmentation, and help improve the accuracy and reliability of diagnosis.

[0075] In one embodiment, the feature-focused image is subjected to cross-modal feature fusion segmentation processing to segment the columnar epithelium and squamous epithelium regions, and a segmented region image is obtained, including:

[0076] Using qubits to encode the color and texture features of the feature-focused image to generate a quantum feature map;

[0077] Performing bioelectric characteristic simulation feature extraction on the feature-focused image to generate bioelectric features;

[0078] Based on the co-evolution algorithm, fusing the quantum feature map and the bioelectric features to generate a co-evolution fusion feature matrix;

[0079] Dividing the co-evolution fusion feature matrix into multiple cells, and simulating the cell growth and boundary formation processes according to the local rules and state transitions between cells to adaptively segment the columnar epithelium and squamous epithelium regions to obtain the segmented region image.

[0080] In this embodiment, a qubit is the basic information unit in quantum computing. Different from traditional bits, it can simultaneously be in a superposition state of 0 and 1. In this step, by utilizing the superposition property of qubits, the color information (such as RGB values or numerical values in other color spaces) and texture features (such as features extracted by methods like local binary pattern) in the feature-focused image are encoded. Through a specific quantum algorithm, these classical image features are mapped into the quantum state space. The state change of each qubit reflects different values or combinations of the image features, and finally a quantum feature map that can characterize the color and texture features of the image is generated. When dealing with complex colposcopy images, traditional image feature representation methods may not be able to fully capture the subtle feature differences and potential feature correlations. Qubit encoding can utilize the superposition and entanglement properties of quantum states to encode image features in a richer and more efficient way, providing a new feature representation form with high-dimensional information expression ability for subsequent feature fusion and analysis, which helps to uncover cell feature details that are difficult to discover by traditional methods and improve the description accuracy of the features of columnar epithelium and squamous epithelium regions.

[0081] Human cells generate bioelectric activities, and the bioelectric properties of columnar epithelium and squamous epithelium cells are different in physiological and pathological states. A mathematical model is established to simulate the generation and propagation processes of the bioelectricity of these cells, such as based on the cell membrane ion channel model, intercellular electrical coupling model, etc. From the feature-focused image, information related to cell morphology and distribution is extracted as input parameters of the model, and the bioelectric properties of the cells, such as transmembrane potential, intercellular conductivity, etc., are simulated and calculated. These simulated bioelectric parameters are used as bioelectric features. Introducing this unique modal feature of bioelectric properties can describe columnar epithelium and squamous epithelium cells from a brand-new perspective. Bioelectric features are closely related to the physiological functions of cells, and the bioelectric properties of diseased cells often change. By extracting bioelectric features, an additional information dimension with biological significance can be provided for cell region segmentation. Combining with traditional image features, different types of cell regions can be identified and distinguished more comprehensively and accurately, improving the accuracy and reliability of segmentation, and helping to discover potential diseased regions.

[0082] The co - evolutionary algorithm simulates the process of mutual dependence and co - evolution among different species in the biological world. In feature fusion, the quantum feature map and bio - electrical features are regarded as two different "species". Through continuous iteration, the algorithm gradually optimizes these two features during their interaction. For example, by defining an appropriate fitness function, the effectiveness of the fused features for the cell region segmentation task is evaluated. In each iteration, according to the fitness value, the weight distribution of the quantum feature map and bio - electrical features, as well as their combination method, are adjusted. Eventually, a co - symbiotic fusion feature matrix that can fully integrate the advantages of both is generated. This matrix contains complementary information from different modality features. Single - modality features often cannot comprehensively describe the characteristics of cell regions. The quantum feature map focuses on the physical feature representation of images, while the bio - electrical features reflect the physiological characteristics of cells. Through fusion by the co - evolutionary algorithm, the boundaries between modalities can be broken, and the synergistic effect between different features can be fully exploited, making the fused features more representative and discriminative. Such a co - symbiotic fusion feature matrix provides a more powerful feature basis for accurately segmenting columnar and squamous epithelial regions subsequently, and can better handle the complex and variable cell feature patterns in colposcopy images.

[0083] Cellular automata is a mathematical model with discrete time, space, and state. The co - symbiotic fusion feature matrix is divided into multiple cells, and each cell represents a local region in the image. The state of the cell is determined by the corresponding element value in the co - symbiotic fusion feature matrix. According to pre - set local rules, such as the relationship between the cell state and the states of its adjacent cells (such as based on similarity, gradient, etc.), and the transition rule of the cell state over time (simulating the state changes during cell growth, division, differentiation, etc.), during the iteration process, the state of the cell is continuously updated. As the iteration progresses, the state distribution of the cells gradually simulates the cell growth and boundary formation process. Eventually, based on the clustering result of the cell states, the boundaries of the columnar and squamous epithelial regions are adaptively determined, thereby obtaining the segmented region image. This segmentation method based on cellular automata can fully utilize the information in the co - symbiotic fusion feature matrix and perform region segmentation in a dynamic and adaptive manner. Compared with traditional fixed - threshold or specific - model - based segmentation methods, cellular automata can better adapt to the complex shapes and irregular boundaries of cell regions in the image. By simulating the natural process of cell growth and boundary formation, it is more in line with the actual situation of cell biology and can more accurately segment the columnar and squamous epithelial regions, providing accurate basic data for further analysis and diagnosis of cell regions subsequently.

[0084] In one embodiment, graph convolution and topological feature extraction are performed on the segmented region image to construct an image adjacency graph and obtain an image feature vector, including:

[0085] Adopt a graph-cut based superpixel segmentation algorithm to divide the segmented region image into superpixel blocks with semantic consistency, obtaining a refined superpixel segmentation image;

[0086] Calculate the color moment, texture feature, and shape feature of each superpixel block, and combine them as node features to obtain a set of feature nodes; among them, each superpixel block corresponds to a node in the graph;

[0087] Determine the edges between nodes according to the spatial adjacency and feature similarity between the superpixel blocks; among them, for nodes that are adjacent in space and have a feature similarity higher than the set threshold, establish an edge connection, and use the similarity value as the weight of the edge to obtain a preliminary image adjacency graph;

[0088] Apply the graph attention mechanism to assign different attention weights to the adjacent edges of each node to obtain an optimized image adjacency graph;

[0089] Use a multi-layer graph convolutional network to extract features from the optimized image adjacency graph to obtain graph convolutional features;

[0090] Perform a global pooling operation on the graph convolutional features to obtain the image feature vector.

[0091] In this embodiment, the graph-cut based superpixel segmentation algorithm regards the image as a graph structure, where pixel points are nodes, and the edge weights between adjacent pixel points reflect their similarity, such as the similarity of features like color and texture. By finding the optimal cut set of the graph, the image is segmented into several superpixel blocks. The pixels within these superpixel blocks have high consistency in features. Compared with traditional pixel-level processing, superpixel blocks, as a whole, can better retain the semantic information of the image and reduce the data volume for subsequent processing. The structure of cells and tissues in the segmented region image is complex, and directly processing pixel-level data will face problems of large computational complexity and susceptibility to noise interference. Superpixel segmentation divides the image into superpixel blocks with semantic consistency, which not only simplifies the representation of the image but also retains the important features of the cell region. Such a refined superpixel segmentation image provides a more reasonable and efficient data basis for subsequent extraction of the features of superpixel blocks and construction of the graph structure, contributing to more accurate analysis of the characteristics of the cell region.

[0092] The above color moments concisely and effectively describe the color distribution characteristics by calculating the mean, variance, and third-order moments of the color within the superpixel blocks. Texture features can be extracted using methods such as Local Binary Pattern (LBP), which reflects the texture patterns within the superpixel blocks. Shape features are characterized by calculating parameters such as the perimeter, area, and compactness of the superpixel blocks. Combining these different types of features can comprehensively describe the characteristics of each superpixel block. Each superpixel block corresponds to a node in the graph, and its combined features serve as the feature vector of the node, which is used for subsequent information transmission and analysis in the graph structure.

[0093] To accurately represent and distinguish different superpixel blocks in the graph structure, each node needs to be endowed with rich and representative features. Color moments, texture features, and shape features describe the characteristics of superpixel blocks from different perspectives. Combining these features can more comprehensively depict the characteristics of the cell regions represented by the superpixel blocks. The establishment of the feature node set provides specific data content for constructing the image adjacency graph and subsequent extraction of topological features, enabling the graph structure to carry rich image information and laying a foundation for in-depth analysis of cell region relationships.

[0094] Spatial adjacency ensures that superpixel blocks (i.e., nodes) adjacent in physical space have potential connection relationships. Feature similarity is measured by calculating the distance (such as Euclidean distance, cosine similarity, etc.) between the feature vectors of two superpixel block nodes. When nodes are adjacent in space and their feature similarity is higher than a pre-set threshold, an edge connection is established between them. The weight of the edge is set to the similarity value. The higher the similarity, the greater the weight, which means the closer the connection between the two nodes. The preliminary image adjacency graph constructed in this way can reflect the spatial and feature correlation relationships between superpixel blocks.

[0095] Constructing the image adjacency graph is to reflect the mutual relationships between superpixel blocks in the graph structure, so as to mine the topological information of cell regions. The combination of spatial adjacency and feature similarity enables the adjacency graph to not only reflect the continuity of cell regions in space but also embody their similarity in features. The preliminary image adjacency graph provides the basic graph structure for subsequent feature extraction using graph convolutional networks. The setting of edge weights helps to highlight the connections between similar nodes during information transmission and better capture the topological features between cell regions.

[0096] The graph attention mechanism simulates the attention process of human vision. In the graph structure, for each adjacent edge of a node, according to the node features and the features of the adjacent edge, through a specific attention calculation method (such as weighted summation based on attention scores), different attention weights are assigned to different adjacent edges. The higher the attention score, the higher the importance of the corresponding adjacent edge in information transmission. In this way, the graph attention mechanism can automatically learn the importance degree of each node's adjacent edges, highlighting the adjacent information that is more valuable for node feature expression. In the preliminary image adjacency graph, not all adjacent edges contribute equally to the node features. Using the graph attention mechanism can dynamically adjust the weights of the adjacent edges, making the graph structure pay more attention to the adjacent information closely related to the node features. This helps in the subsequent graph convolution process to more effectively aggregate the features of the node and its adjacent nodes, enhancing the graph convolution network's ability to extract the topological features of cell regions. The optimized image adjacency graph can more accurately reflect the important relationships between cell regions, providing a guarantee for extracting more representative graph convolution features.

[0097] The graph convolution network realizes the aggregation and update of node features by performing convolution operations on the graph structure. In each layer of graph convolution, the features of a node are weighted and aggregated with the features of its adjacent nodes, and the weights are determined by the weights of the adjacent edges (after being adjusted by the graph attention mechanism) and the convolution kernel parameters. After being processed layer by layer by the multi-layer graph convolution network, the node features can gradually fuse the topological information of their neighborhoods, thereby extracting higher-level graph convolution features. These features not only contain the features of the node itself but also fuse the structural and feature information of its surrounding neighborhoods. The multi-layer structure of the graph convolution network can gradually mine the topological features in the image adjacency graph. By continuously aggregating the information of adjacent nodes, the spatial relationships and feature associations between cell regions are transformed into more representative feature representations. The graph convolution features can describe the topological structure of cell regions from both global and local levels, providing strong feature support for subsequent image analysis and comparison, helping to discover the topological differences between different cell regions, and improving the accuracy and depth of the analysis of cell regions in colposcopy images.

[0098] The global pooling operation aggregates the graph convolution features over the entire graph structure. Common global pooling methods include average pooling and max pooling. Average pooling calculates the average value of the graph convolution features over all nodes, while max pooling selects the maximum value among all node features. Through global pooling, the graph convolution features are compressed into a fixed-length vector, which synthesizes the feature information of the entire graph structure, removes the spatial location information in the graph structure, and only retains the statistical information of the features. To facilitate subsequent operations such as comparing and classifying images, it is necessary to transform the graph convolution features into a fixed-length image feature vector. The global pooling operation can effectively integrate and compress the complex feature information of the graph structure to obtain a concise vector that represents the topological features of the entire image. This image feature vector can be used as a compact representation of the image for subsequent operations such as fuzzy similarity measurement, facilitating comparison with the feature vectors of other images, thereby achieving accurate comparative analysis of the images collected by the colposcope.

[0099] In one embodiment, performing fuzzy similarity measurement and contrast decision processing on the image feature vector to obtain the contrast result of the images collected by the colposcope, including:

[0100] Dividing the feature space where the image feature vector is located into multiple fuzzy clustering centers based on the clustering algorithm, and calculating the fuzzy membership degree of each image feature vector with respect to each clustering center to obtain a fuzzy membership degree matrix;

[0101] Using the cosine similarity algorithm to calculate the similarity between different image feature vectors to obtain a set of fuzzy similarity values;

[0102] Obtaining a fuzzy decision rule base; the fuzzy decision rule base contains qualitative descriptions of image comparison corresponding to different similarity intervals;

[0103] Based on the Mamdani fuzzy inference method, performing inference according to the fuzzy similarity value and the fuzzy decision rule base to obtain a fuzzy contrast decision result;

[0104] Using the centroid method to defuzzify the fuzzy contrast decision result to obtain a clear contrast result of the images collected by the colposcope.

[0105] In this embodiment, a clustering algorithm, such as the fuzzy C-means clustering (FCM), assigns image feature vectors to different clusters by iteratively optimizing the objective function, and these cluster centers are the fuzzy cluster centers. After determining the cluster centers, the fuzzy membership function is used to calculate the degree to which each image feature vector belongs to each cluster center. For example, using a Gaussian membership function, the membership degree is determined according to the distance between the feature vector and the cluster center. The closer the distance, the higher the membership degree. Finally, a matrix is formed, where each element represents the fuzzy membership degree of an image feature vector to a certain cluster center, that is, the fuzzy membership matrix. The image feature vector space is complex and diverse, and it is difficult to effectively distinguish the similarity degree of different feature vectors by directly performing similarity measurement. By dividing the fuzzy cluster centers and calculating the fuzzy membership degree, the feature vectors can be grouped fuzzily according to similarity. The fuzzy membership matrix provides richer information for subsequent similarity calculation. It reflects the belonging degree of each feature vector in different fuzzy categories, helps to more accurately measure the similarity relationship between feature vectors, and provides a more detailed feature analysis basis for image comparison.

[0106] The cosine similarity algorithm measures the similarity between two image feature vectors by calculating the cosine value of the included angle between them. The elements in different rows (representing different image feature vectors) of the fuzzy membership matrix are used as vectors, and the cosine value between them is calculated. The closer the cosine value is to 1, the more similar the directions of the two vectors are, that is, the higher the similarity of the corresponding image feature vectors in the feature space; the closer the cosine value is to 0, the lower the similarity. By calculating all possible pairs of image feature vectors, a series of similarity values are obtained, forming a set of fuzzy similarity values. The cosine similarity algorithm can effectively measure the similarity degree between different image feature vectors on the basis of considering the overall distribution of feature vectors. Combining the fuzzy membership information can more comprehensively reflect the fuzzy similarity relationship between image features. The set of fuzzy similarity values provides a quantitative data basis for subsequent comparison decisions. These similarity values can intuitively show the similarity degree of different colposcopy-acquired images at the feature level, facilitating further analysis and judgment of the differences between images.

[0107] The establishment of the fuzzy decision rule base is based on the experience of medical experts and a large amount of clinical data. By analyzing and summarizing cases of colposcopy image comparisons with different similarity values, the similarity values are divided into different intervals, such as high similarity interval, medium similarity interval, low similarity interval, etc., and corresponding qualitative descriptions of image comparisons are assigned to each interval, such as "The images are highly similar and may be normal tissues", "The images are moderately similar, and some feature differences need further attention", "The images have low similarity and there is a suspicion of lesions", etc. These rules are stored in the fuzzy decision rule base in the form of conditional statements. For example, "If the similarity value is in the high similarity interval, then the qualitative description of the image comparison is that the images are highly similar and may be normal tissues". The fuzzy decision rule base provides a conversion basis from fuzzy similarity values to clear image comparison conclusions. It combines the quantified similarity values with professional knowledge and experience in the medical field, enabling the computer to give qualitative descriptions with practical diagnostic significance according to the calculated fuzzy similarity values based on the established rules. Such a rule base is a key link in realizing automated and intelligent image comparison analysis, which helps doctors quickly understand the differences between images and the potential possibility of lesions.

[0108] The Mamdani fuzzy inference method is a reasoning method based on fuzzy logic. First, the fuzzy similarity value is used as the input, and through the fuzzification operation, it is transformed into a fuzzy set, that is, the membership degree in different fuzzy linguistic variables (such as high, medium, low similarity) is determined. Then, according to the rules in the fuzzy decision rule base, matching and reasoning are carried out. For example, if there is a rule "If the similarity is high, then the qualitative description of the image comparison is that the images are highly similar and may be normal tissues", and the currently input fuzzy similarity value has a high membership degree in the "high similarity" fuzzy set, then this rule is activated. By synthesizing the outputs of all activated rules, a fuzzy output set is obtained, and this set is the fuzzy comparison decision result, which describes the possible conclusions of the image comparison in the form of fuzzy language. The Mamdani fuzzy inference method can make full use of the knowledge in the fuzzy decision rule base to transform the fuzzy similarity value into a meaningful fuzzy decision result. It takes into account the fuzziness of the similarity value and the mutual relationship between rules, making the reasoning process more in line with the human thinking mode and the actual situation of medical diagnosis. Although the fuzzy comparison decision result is still fuzzy, it synthesizes the similarity information and professional rules, provides a transition for finally obtaining a clear comparison conclusion, and can more comprehensively and flexibly reflect the complex situation of image comparison.

[0109] The centroid method is a commonly used defuzzification method. For the fuzzy set formed by the fuzzy comparison decision results, an exact value is determined by calculating the centroid position of the set. During specific calculation, the membership degree of each element in the fuzzy set is used as the weight, and weighted summation is performed with its corresponding value (in the qualitative description of image comparison, a representative value can be assigned to each qualitative description), and then divided by the total membership degree. The result obtained is the defuzzified value. According to this value, the corresponding image comparison conclusion is searched in the pre-set clear conclusion mapping table, such as mapping the value to clear descriptions such as "normal image", "suspected lesion image", "confirmed lesion image", etc., so as to obtain the clear colposcopy acquisition image comparison result. The defuzzification operation converts the fuzzy comparison decision result into a clear, easy-to-understand and apply conclusion. Doctors need clearer image comparison results in clinical diagnosis to assist in judging the condition. Through defuzzification by the centroid method, the result obtained by fuzzy reasoning can be converted into clear diagnostic information, providing doctors with an intuitive and accurate image comparison analysis report to help doctors make more accurate diagnostic decisions.

[0110] In one embodiment, after obtaining the colposcopy acquisition image comparison result, it includes:

[0111] Obtain the user's account information, and based on the account information, construct a character array;

[0112] Extract keywords from the colposcopy acquisition image comparison result, and perform character mapping based on the extracted keywords to obtain key characters;

[0113] Deform the pre-set coding table based on the key characters to obtain a deformed coding table;

[0114] Based on the text other than the keywords in the colposcopy acquisition image comparison result, construct a text matrix;

[0115] Adjust the character array based on the text matrix to obtain an adjusted array;

[0116] Generate a communication key based on the deformed coding table and the adjusted array to encrypt the 4K original image and then send it to the corresponding user account.

[0117] In this embodiment, user account information related to the current operation is retrieved and extracted from the user database. This information includes the username, user ID, registered email, etc. These information are characterized according to certain rules, for example, each character is converted into the corresponding ASCII code value or other character encoding forms. Then, these encoded characters are constructed into a two-dimensional or multi-dimensional character array according to a specific arrangement. The dimension and arrangement rules of the array can be designed according to the requirements of subsequent processing. For example, the category of user information is used as one dimension, and the specific information content is used as another dimension.

[0118] The construction of the character array is to convert the user account information into a data structure convenient for subsequent algorithm processing. Through this structured representation, effective association and operation can be carried out with other processing results in subsequent steps, laying a foundation for generating an encryption key related to the user, ensuring that the encryption process is closely bound to the user identity, and enhancing the pertinence and security of encryption.

[0119] Using keyword extraction algorithms in natural language processing, such as the term frequency-inverse document frequency (TF-IDF) algorithm or the TextRank algorithm, representative keywords are extracted from the text description of the colposcopy image comparison result. These keywords can usually summarize the core content of the comparison result, such as "normal", "lesion", "suspected", etc. Then, a unique character or character combination is assigned to each keyword for character mapping. For example, "normal" is mapped to the character "A", "lesion" is mapped to the character "B", etc., and finally the key characters representing the keywords are obtained. Keyword extraction can refine key information from the complex text of the image comparison result, while character mapping further simplifies this key information into an easily processed character form. The key characters are used as important identifiers in subsequent steps for specific deformation operations on the preset coding table, enabling the coding table to be associated with the key content of the image comparison result and providing key elements for generating a more targeted and secure encryption key.

[0120] The preset coding table is a pre - defined character coding correspondence table. For example, it has a structure similar to the ASCII code table, but is customized according to specific encryption requirements. In this embodiment, based on the extracted key characters, deformation operations are performed on the coding table. There can be various deformation methods. For example, according to the number of key characters, the order of characters, or specific combinations of characters, the order of characters in the coding table is rearranged, or the coding values of specific characters are modified. For example, if the key characters are "A" and "B", the positions of the characters with coding values 65 (corresponding to the character "A") and 66 (corresponding to the character "B") in the coding table can be swapped to obtain a deformed coding table. By performing deformation on the preset coding table based on key characters, a close connection is established between the coding table and the comparison result of the colposcopy - acquired image. The deformed coding table can incorporate the key information of the image comparison result into the key - generation mechanism during the subsequent generation of communication keys, increasing the complexity and security of the key, preventing the key from being easily cracked, and at the same time endowing the encryption process with uniqueness related to the image content.

[0121] Organize the remaining text content after removing keywords from the text of the colposcopy - acquired image comparison result. First, perform word - segmentation on this text, splitting the continuous text into individual words. Then, according to the position and order of the words in the text, arrange them into a matrix structure according to rules. For example, the sentences can be used as the rows of the matrix in the order of the sentences, and the words in each sentence can be used as the columns to construct a two - dimensional text matrix. Constructing the text matrix is to structure the non - key text information in the image comparison result for subsequent interactive processing with the character array. The text matrix can retain other important detailed information in the comparison result except for the key information. By combining with the character array, it further enriches the information source in the key - generation process, enabling the generated communication key to more comprehensively reflect the content of the image comparison result and enhancing the reliability and security of encryption.

[0122] Adjust the character array constructed previously according to the structure and content of the text matrix. The adjustment method can be based on the number of words in the text matrix, the characteristics of the words, or the relationships between the words. For example, if the length of a word at a certain position in the text matrix is odd, a certain transformation can be performed on the corresponding character in the character array, such as increasing the ASCII code value of the character by a fixed offset; or according to the semantic categories of the words in the text matrix (such as descriptive words, judgment words, etc.), rearrange the characters in the corresponding area of the character array. Through a series of such rules, adjust the character array to obtain an adjusted array. Adjusting the character array through the text matrix realizes the integration of non-critical text information in the user account information and the image comparison result. The adjusted array combines the characteristics of both, providing a more complex and diverse information basis for generating the communication key. This fusion method makes the key not only related to the user identity but also closely connected to the specific content of the image comparison result, further improving the security and pertinence of encryption and increasing the difficulty of key cracking.

[0123] Use the deformed coding table and the adjusted array as inputs, and apply a specific key generation algorithm to generate the communication key. Take the information in the deformed coding table and the adjusted array as the initial key material, and through a series of transformations and operations, generate the final communication key. The generated communication key is used to encrypt the 4K original image, and the encrypted image is sent to the storage location or display terminal associated with the corresponding user account through a secure network transmission channel. Generating the communication key and encrypting the 4K original image are to ensure the security of the image during transmission and storage. By integrating the information of the deformed coding table and the adjusted array into the key generation process, the generated key has a high degree of uniqueness and complexity, making it difficult to be illegally obtained and cracked. The encrypted image can only be decrypted and viewed under the user account with the correct key, effectively protecting the privacy of the patient and the security of the image data, meeting the strict requirements for the secure transmission and storage of medical data.

[0124] In one embodiment, generating the communication key based on the deformed coding table and the adjusted array includes:

[0125] Identify the attributes of each word in the adjusted array, and construct a first curve based on the identified attributes of each word;

[0126] Obtain the table attributes of the deformed coding table, and generate a second curve based on the table attributes;

[0127] Superimpose the first curve and the second curve onto the deformed coding table according to the rules;

[0128] Obtain the area enclosed by the first curve and the second curve in the deformed coding table as the target area;

[0129] Combine the characters in the target area to obtain the communication key.

[0130] In this embodiment, the recognition of text attributes involves a variety of natural language processing and text analysis techniques. For example, part-of-speech tagging is used to determine whether a word is a noun, verb, adjective, etc.; semantic analysis is used to judge the semantic category to which a word belongs, such as medical terms, descriptive words, etc.; the attributes of a word can also be comprehensively judged according to information such as its position and frequency of occurrence in the text. After the attributes of each word are recognized, these attribute values are quantified, for example, mapping part-of-speech, semantic category, etc. to specific numerical values. Then, taking the order of the words in the adjustment array as the abscissa and the corresponding quantified attribute values as the ordinate, a continuous curve, that is, the first curve, is constructed through mathematical methods such as interpolation and fitting. For example, if the quantified part-of-speech value of a certain word is 3 and it is in the 5th position in the adjustment array, then when constructing the curve, the coordinates of this point are (5, 3), and many such points are connected to form the first curve.

[0131] The words in the adjustment array contain relevant information in the comparison result of the colposcopy-acquired images and features after fusion with the user account information. By recognizing the text attributes and constructing the first curve, these discrete text information are transformed into a continuous mathematical representation with specific features. The first curve can comprehensively reflect the overall attribute features of the words in the adjustment array, providing a new information dimension for subsequent fusion with the deformation coding table and generating a unique communication key, increasing the complexity and uniqueness of the key generation process, and helping to improve the security of the key.

[0132] The tabular attributes of the deformation coding table cover multiple aspects, such as the number of rows and columns in the table, the character distribution density, the positions of specific characters, etc. Quantitatively analyze these attributes. For example, directly use the number of rows and columns as quantitative values. The character distribution density can be obtained by calculating the number of characters per unit area. For the positions of specific characters, convert their row and column coordinates in the table into corresponding numerical values. Arrange these quantified tabular attribute values in a certain logical order. Similarly, use a certain numerical value as the abscissa (such as the category serial number of the attribute), and the quantified attribute value as the ordinate to construct a second curve. For example, if the quantified value of the number of rows in the deformation coding table is 10 and it is used as the attribute value with the category serial number 1, then when constructing the curve, the corresponding point coordinates are (1, 10). Connect the points corresponding to all relevant attribute values in a similar way to form a second curve. The deformation coding table itself carries the coding rules related to the key information of the comparison result of the colposcopy-acquired images. By generating the second curve, present the structure and characteristics of the deformation coding table in the form of a continuous curve. This curve can reflect the unique attributes of the deformation coding table. Combined with the first curve, it can further enrich the information source for key generation, making the generated communication key not only related to the content of the image comparison result but also closely connected to the characteristics of the coding table, enhancing the complexity and security of the key and preventing the key from being easily cracked.

[0133] Pre-determine the overlay rules. For example, correspond according to the coordinate values of the curve and the row and column indices of the deformation coding table. For the first curve, perform a certain operation on the ordinate value (quantified attribute value) of each point on the curve and the character at the corresponding abscissa position in the deformation coding table (the order position of the text in the adjusted array can be mapped to a certain dimension of the coding table, such as the column index), such as adding the ASCII code value of the character to the ordinate value. Similarly, for the second curve, perform a similar operation on the ordinate value (quantified tabular attribute value) and the character at the corresponding abscissa position in the deformation coding table (the category serial number of the tabular attribute can be mapped to another dimension of the coding table, such as the row index). In this way, integrate the information carried by the two curves into the characters in the deformation coding table and change the original characteristics of the characters in the coding table.

[0134] The overlay operation enables the deformation coding table to be deeply integrated with the information represented by the first curve and the second curve. This integration further confuses the information in the coding table, intertwining the user account information, the image comparison result information, and the coding table's own attribute information. Through this complex integration method, it creates a more random and unique basis for subsequent determination of the target area and generation of communication keys, greatly increasing the difficulty of reverse-deriving the key from the coding table and improving the security of encryption.

[0135] After completing the curve superposition, based on the coordinate positions of the curves in the deformation coding table, determine the region enclosed by the two curves. This involves the determination and calculation of geometric figures. For example, by comparing the coordinates of points on the curve with the coordinates of character positions in the coding table to determine the region boundary. If the ordinate value of the first curve at a certain abscissa position is y1, and the ordinate value of the second curve at the same abscissa position is y2, and y1 < y2, then in the column of the coding table corresponding to this abscissa, the region formed by the characters from the position of y1 to the position of y2 is a part of the region enclosed by the two curves in this column. By traversing all relevant dimensions (such as rows and columns) of the coding table, determine the complete target region. The determination of the target region is a key step in extracting information directly related to key generation from the deformation coding table that integrates multiple types of information. This region concentrates the characteristics formed by the combined action of user account information, image comparison result information, and the properties of the coding table itself. By defining this region, a set of representative and unique characters can be screened out, providing an accurate data source for generating the final communication key, ensuring that the key is closely related to the encryption object (colposcopy image) and the user, and has a high level of security.

[0136] Arrange and combine the characters in the target region in sequence. The sequence can be based on the geometric shape of the target region (such as from left to right, from top to bottom), or according to a pre-set complex rule, such as arranging according to a certain relative position relationship of the characters in the original adjustment array or the deformation coding table. Connect the arranged characters to form a string, and this string is the generated communication key. For example, if the target region is a rectangular region, connect the characters in the region in sequence from left to right and from top to bottom. Assuming these characters are "a", "b", "c", then the generated communication key is "abc". Combining the characters in the target region into a communication key completes the key conversion from a complex information processing process to the generation of the final encryption key. Through this unique key generation method, the generated communication key has a high degree of randomness and complexity and is difficult to crack. Since the generation of the key integrates information from multiple aspects and is closely connected to the colposcopy image comparison result and user account information, it can provide strong security protection for the encryption of 4K original images, ensure the confidentiality of the images during transmission and storage, and effectively protect patient privacy and medical data security.

[0137] In one embodiment, generating a communication key based on the deformation coding table and the adjustment array includes:

[0138] For the adjustment array, use the quantum neural network algorithm to calculate the quantum semantic potential field values of each character in the multi-dimensional semantic space, construct a quantum semantic potential field matrix, introduce the fractal interpolation theory for feature enhancement, and generate a first feature tensor;

[0139] For the deformed coding table, based on the fractal Brownian motion model, calculate the fractal Brownian motion parameters of characters, combine the encryption rules of the coding table and the implicit logical relationships between characters, construct the dynamic fractal topological structure of the coding table, extract the time-varying topological features from the fractal topological structure, and generate the second feature tensor.

[0140] Fuse the first feature tensor and the second feature tensor to obtain a fused tensor, and map it to a communication key.

[0141] In this embodiment, the quantum neural network uses quantum bits to process information. Compared with the traditional neural network, it can utilize the superposition state and entanglement characteristics of quantum bits to achieve parallel computing, greatly improving the processing efficiency. In this step, by constructing a quantum neural network model, the text information in the adjustment array is input into it. The quantum neurons in the model encode and decode the text, simulate the signal transmission between neurons using quantum gate operations, and continuously adjust the network parameters through the quantum backpropagation algorithm to calculate the quantum semantic potential field values of each text in the multi-dimensional semantic space. These values reflect the energy distribution and interaction relationships of the text in the complex semantic environment. Based on these values, construct a quantum semantic potential field matrix, and the elements in the matrix are the quantum semantic potential field values of each text. Subsequently, introduce the fractal interpolation theory. Fractal interpolation can effectively enhance the feature details in the matrix without adding a large amount of data by constructing self-similar interpolation functions between known data points. Process the quantum semantic potential field matrix through fractal interpolation to generate the first feature tensor containing rich text semantic details. The dimension and structure of the tensor can more comprehensively describe the text semantic features. The first feature tensor not only contains the basic semantic information of the text but also enhances the semantic details through fractal interpolation, providing a rich semantic information basis for subsequent feature fusion with the deformed coding table and key generation, helping to generate keys closely related to the image content and user information, and enhancing the uniqueness and security of the keys.

[0142] The fractal Brownian motion model is used to describe a stochastic process with self - similarity and long - range dependence. When analyzing the deformation coding table, the character distribution is regarded as a stochastic process. By calculating fractal Brownian motion parameters such as the Hurst exponent, etc., the self - similar characteristics and changing trends of the character distribution at different time scales are characterized. Combining the encryption rules of the coding table, such as character replacement rules, shift rules, etc., and the implicit logical relationships between characters, such as specific character combinations representing specific meanings, a dynamic fractal topological structure of the coding table is constructed. In this structure, nodes can represent characters, edges represent the logical connections or encryption relationships between characters, and the weights and directions of the edges reflect the strength and nature of the relationships between characters. Using graph signal processing techniques, the dynamic fractal topological structure is analyzed. By designing appropriate filters and signal propagation models, time - varying topological features that can reflect the structural changes of the coding table at different stages are extracted. These features are organized into a second feature tensor, and the tensor form can better store and process this complex topological feature information. The deformation coding table is an important element in the key generation process, and its own structure and features are closely related to the security of encryption. By using the fractal Brownian motion model and graph signal processing techniques, the dynamic characteristics and topological structure of the coding table are deeply explored to generate a second feature tensor. This tensor contains key information such as the encryption rules of the coding table, character relationships, and structural changes, providing feature data closely related to encryption for the subsequent fusion with the first feature tensor, enabling the generated key to fully consider the characteristics of the coding table and enhancing the encryption strength and anti - cracking ability of the key.

[0143] Tensor fusion adopts a specific fusion algorithm, such as a fusion method based on weighted summation. According to the importance of the information represented by the first feature tensor and the second feature tensor, different weights are assigned to the elements of each tensor, and then the corresponding elements are weighted and summed to obtain a fusion tensor. The fusion tensor synthesizes the text semantic features and the coding table topological features, and contains information from multiple aspects such as the image comparison results, user account information, and coding table characteristics. In order to map the fusion tensor into a communication key, a mapping function is designed. This function can map the high - dimensional data in the fusion tensor into a fixed - length string based on techniques such as hash algorithms and encryption transformations. This string is the communication key. For example, by arranging the elements of the fusion tensor in a specific order and then inputting them into a hash function (such as SHA - 256), the hash value is calculated, and the hash value is converted into a string in a specific format as the communication key.

[0144] In one embodiment, deforming a preset coding table based on the key characters to obtain a deformed coding table includes:

[0145] Obtain a preset coding table; the preset coding table includes an original data column and a coding data column that is mapped to it one - to - one;

[0146] Fill the characters in the encoded data column into each node of a preset directed graph structure one by one in sequence to generate a character directed graph; each node is connected by a directed edge;

[0147] In the character directed graph, find the node corresponding to the key character; adjust the direction of the directed edge associated with the node where the key character is located to the opposite direction to obtain an adjusted directed graph;

[0148] Based on the adjusted directed graph, reorder the characters on the nodes according to the rules based on the direction of the directed edge to obtain a character rearrangement sequence;

[0149] Replace the characters in the character rearrangement sequence into the encoded data column one by one in sequence to obtain a deformed encoding table.

[0150] In this embodiment, read the pre-set encoding table data from a storage device (such as a hard disk, a database). The encoding table exists in a structured form, usually in a table form, and includes an original data column and an encoded data column that is mapped to it one by one. For example, the original data column is common characters, numbers or symbols, and the encoded data column is a sequence of encrypted or transformed characters generated according to specific encoding rules.

[0151] The preset directed graph structure is a data structure in graph theory, consisting of nodes and directed edges. Fill the characters in the encoded data column into each node of the directed graph one by one in sequence, and each character occupies one node. The connection method of the directed edges can be determined according to various rules. For example, according to the order of the characters in the encoded data column, a directed edge is established between the nodes of adjacent characters, and the direction is from the previous character node to the next character node; converting the encoded data column into a character directed graph is to utilize the powerful expression ability and analysis tools of the directed graph. The directed graph can intuitively display the relationship between characters, and through operations on the graph, the character sequence can be processed more flexibly. Compared with directly operating on the encoded data column, the directed graph provides a more abstract, easier to understand and process complex relationships way, providing a convenient data structure basis for subsequent operations related to key characters and deformation of the encoding table.

[0152] In the constructed character directed graph, find the node corresponding to the key character by traversing the nodes. The traversal algorithm can adopt common graph search algorithms such as depth-first search (DFS) or breadth-first search (BFS), and match according to the character information stored in the nodes and the key character. Once the node where the key character is located is found, reverse the direction of the directed edge associated with the node. For example, the directed edge originally pointing from node A to node B becomes pointing from node B to node A after adjustment.

[0153] Adjusting the direction of the directed edges associated with the nodes where the key characters are located is a specific change to the character directed graph structure, aiming to break the original character relationship order and introduce new logical connections. Key characters are often associated with important information. By changing the direction of the directed edges around them, different traversal paths and orders are provided for subsequent character rearrangement based on the directed graph structure, thereby realizing the deformation of the coding table and increasing the complexity and security of the coding.

[0154] Based on the adjusted directed graph, the characters on the nodes are reordered according to the direction of the directed edges according to a preset rule. A common rule is to start from a certain starting node and perform a depth-first traversal or a breadth-first traversal according to the direction of the directed edges. During the traversal process, the characters on the nodes passed through are recorded in sequence to form a character rearrangement sequence. For example, using depth-first traversal, starting from the root node of the graph, continuously visit unvisited nodes along the directed edges, and add the characters on the visited nodes to the sequence in turn. If there are multiple starting nodes or traversal path selection rules, they can be set according to specific requirements, such as determining the starting node or selecting the traversal path according to certain attributes of the nodes (such as the frequency of characters, importance scores, etc.).

[0155] Through this character rearrangement method based on the directed graph structure and the direction of the directed edges, a new sequence different from the character order in the original coding data column is generated. This new sequence reflects the new character relationships and order contained in the adjusted directed graph structure, laying a foundation for subsequent replacement back into the coding data column to realize the deformation of the coding table. The rearranged character sequence may have different statistical characteristics and logical relationships, thus changing the coding rules of the original coding table and enhancing the confidentiality and anti-cracking ability of the coding. By generating a character rearrangement sequence, the external coding form of the data can be changed without changing the essence of the data content. The receiving party can restore the data when knowing the same key characters and rearrangement rules, while it becomes more difficult for third-party attackers who do not know this information to crack the data.

[0156] The characters in the character rearrangement sequence are sequentially replaced one by one into the coding data column. In programming implementation, by looping through the character rearrangement sequence, each character is written into the corresponding position of the coding data column in order, overwriting the original coding data. Since there is a one-to-one mapping relationship between the coding data column and the original data column, the change of the coding data column will indirectly affect the coding rules and mapping relationship of the entire coding table, thereby obtaining a deformed coding table. This step is the ultimate goal of the entire technical solution, that is, applying the new character sequence obtained through directed graph operations and character rearrangement to the coding data column to realize the deformation of the preset coding table. The deformed coding table has different coding rules from the original coding table and can better meet the requirements for data encryption, processing, or analysis in specific scenarios, such as improving data security and adapting to the input requirements of specific algorithms.

[0157] Referring to Figure 2 , in another embodiment of the present invention, a comparison system for colposcopy-acquired images based on artificial intelligence is further provided, including:

[0158] An enhancement unit for performing hybrid noise reduction and multi-scale enhancement processing on the 4K original images acquired by the colposcope to obtain enhanced and noise-reduced images;

[0159] A focusing unit for adaptively adjusting the scaling ratio of the enhanced and noise-reduced images based on the image feature distribution and using an attention convolutional network to focus on the characteristics of columnar epithelium and squamous epithelial cells to obtain feature-focused images;

[0160] A segmentation unit for performing cross-modal feature fusion segmentation processing on the feature-focused images to segment the columnar epithelium and squamous epithelial regions to obtain segmented region images;

[0161] An extraction unit for performing graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph and obtain image feature vectors;

[0162] A comparison unit for performing fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain the comparison result of the colposcopy-acquired images.

[0163] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to that described in the above method embodiment, and details will not be repeated here.

[0164] Referring to Figure 3 , an embodiment of the present invention also provides a computer device, which may be a server, and its internal structure may be as Figure 3 shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0165] Those skilled in the art can understand that Figure 3 the structure shown in

[0166] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0167] In summary, the present invention provides a method and system for comparing colposcopy-acquired images based on artificial intelligence, including: performing hybrid noise reduction and multi-scale enhancement processing on 4K original images acquired by a colposcope to obtain enhanced and noise-reduced images; adaptively adjusting the scaling ratio of the enhanced and noise-reduced images based on the image feature distribution, and using an attention convolutional network to focus on columnar epithelial and squamous epithelial cell features to obtain feature-focused images; performing cross-modal feature fusion segmentation processing on the feature-focused images to segment columnar epithelial and squamous epithelial regions to obtain segmented region images; performing graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph to obtain image feature vectors; performing fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain a comparison result of colposcopy-acquired images. In the present invention, through image enhancement, it is easier to identify the corresponding tissues; furthermore, relying on cross-modal feature fusion segmentation processing, graph convolution and topological feature extraction, an image adjacency graph is constructed, and then fuzzy similarity measurement and comparison decision processing are performed to obtain a comparison result of colposcopy-acquired images, making the comparison result unified and improving the comparison accuracy; overcoming the defect that it is difficult to ensure consistency and accuracy in the current identification of colposcopy-acquired images relying on manual operation.

[0168] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0169] It should be noted that in this article, the terms "include", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, apparatus, article, or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, apparatus, article, or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, apparatus, article, or method including that element.

[0170] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall equally be included in the patent protection scope of the present invention.

Claims

1. A comparison method for colposcopy-acquired images based on artificial intelligence, characterized in that, Including the following steps: Perform hybrid noise reduction and multi-scale enhancement processing on the 4K original images collected by the colposcope to obtain enhanced and denoised images; Based on the image feature distribution, adaptively adjust the scaling ratio of the enhanced and denoised images, and use the attention convolutional network to focus on the characteristics of columnar epithelium and squamous epithelial cells to obtain feature-focused images; Perform cross-modal feature fusion segmentation processing on the feature-focused images to segment the columnar epithelium and squamous epithelial regions and obtain segmented region images; Perform graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph and obtain image feature vectors; Perform fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain the comparison result of the images collected by the colposcope.

2. The method for comparing colposcopy-acquired images based on artificial intelligence according to claim 1, wherein Performing hybrid noise reduction and multi-scale enhancement processing on the 4K original images collected by the colposcope includes: Use a combination of non-local mean filtering and wavelet transform to denoise the 4K original images, and use the multi-scale Retinex algorithm to enhance the image details.

3. The method for comparing colposcope-acquired images based on artificial intelligence according to claim 1, wherein Performing cross-modal feature fusion segmentation processing on the feature-focused images to segment the columnar epithelium and squamous epithelial regions and obtain segmented region images includes: Use quantum bits to encode the color and texture features of the feature-focused images to generate quantum feature maps; Extract the simulated bioelectrical characteristic features of the feature-focused images to generate bioelectrical features; Based on the co-evolution algorithm, fuse the quantum feature maps and bioelectrical features to generate a co-evolution fusion feature matrix; Divide the co-evolution fusion feature matrix into multiple cells, and according to the local rules and state transitions between cells, simulate the cell growth and boundary formation processes, and adaptively segment the columnar epithelium and squamous epithelial regions to obtain the segmented region images.

4. The method for comparing colposcopy-acquired images based on artificial intelligence according to claim 1, wherein Performing graph convolution and topological feature extraction on the segmented region images to construct an image adjacency graph and obtain image feature vectors includes: Use the graph cut-based superpixel segmentation algorithm to divide the segmented region images into superpixel blocks with semantic consistency to obtain refined superpixel segmentation images; Calculate the color moments, texture features, and shape features of each superpixel block and combine them as node features to obtain a set of feature nodes; where each superpixel block corresponds to a node in the graph; Determine the edges between nodes according to the spatial adjacency and feature similarity between the superpixel blocks; where for nodes that are adjacent in space and have a feature similarity higher than the set threshold, establish an edge connection and use the similarity value as the weight of the edge to obtain a preliminary image adjacency graph; Use the graph attention mechanism to assign different attention weights to the adjacent edges of each node to obtain an optimized image adjacency graph; Use a multi-layer graph convolutional network to extract features from the optimized image adjacency graph to obtain graph convolution features; Perform a global pooling operation on the graph convolution features to obtain the image feature vectors.

5. The method for comparing colposcopy-acquired images based on artificial intelligence according to claim 1, wherein Performing fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain the comparison result of the images collected by the colposcope includes: Based on the clustering algorithm, the feature space where the image feature vectors are located is divided into multiple fuzzy clustering centers, and the fuzzy membership degrees of each of the image feature vectors relative to each clustering center are calculated to obtain a fuzzy membership degree matrix; The cosine similarity algorithm is used to calculate the similarity between different image feature vectors to obtain a set of fuzzy similarity values; Obtain a fuzzy decision rule base; the fuzzy decision rule base contains qualitative descriptions of image comparison corresponding to different similarity intervals; Based on the Mamdani fuzzy inference method, reasoning is performed according to the fuzzy similarity values and the fuzzy decision rule base to obtain a fuzzy comparison decision result; The centroid method is used to defuzzify the fuzzy comparison decision result to obtain a clear comparison result of the colposcopy-acquired images.

6. The method for comparing colposcopy-acquired images based on artificial intelligence according to claim 1, wherein After obtaining the comparison result of the colposcopy-acquired images, it includes: Obtain the user's account information, and based on the account information, construct a character array; Extract keywords from the comparison result of the colposcopy-acquired images, and perform character mapping based on the extracted keywords to obtain key characters; Based on the key characters, deform a preset coding table to obtain a deformed coding table; Based on the text other than the keywords in the comparison result of the colposcopy-acquired images, construct a text matrix; Based on the text matrix, adjust the character array to obtain an adjusted array; Based on the deformed coding table and the adjusted array, generate a communication key to encrypt the 4K original image and then send it to the corresponding user account.

7. The method for comparing colposcopy-acquired images based on artificial intelligence according to claim 6, wherein Generating a communication key based on the deformed coding table and the adjusted array includes: Identify the attributes of each text in the adjusted array, and construct a first curve based on the identified attributes of each text; Obtain the table attributes of the deformed coding table, and generate a second curve based on the table attributes; Superimpose the first curve and the second curve onto the deformed coding table according to the rules; In the deformed coding table, obtain the area enclosed by the first curve and the second curve as the target area; Combine the characters in the target area to obtain the communication key.

8. A comparison system for colposcopy-acquired images based on artificial intelligence, characterized in that, It includes: An enhancement unit for performing hybrid noise reduction and multi-scale enhancement processing on the 4K original image acquired by colposcopy to obtain an enhanced and noise-reduced image; A focusing unit for adaptively adjusting the scaling ratio of the enhanced and noise-reduced image based on the image feature distribution, and using an attention convolutional network to focus on the characteristics of columnar epithelium and squamous epithelial cells to obtain a feature-focused image; A segmentation unit for performing cross-modal feature fusion segmentation processing on the feature-focused image to segment the columnar epithelium and squamous epithelial regions to obtain a segmented region image; An extraction unit for performing graph convolution and topological feature extraction on the segmented region image to construct an image adjacency graph to obtain image feature vectors; A comparison unit for performing fuzzy similarity measurement and comparison decision processing on the image feature vectors to obtain a comparison result of the colposcopy-acquired images.

Images