A method for predicting the risk of an adenocarcinoma

Abstract

The application discloses a kind of adenocarcinoma risk prediction methods, by collecting the cell thin layer diagnosis and treatment data of different diagnosis and treatment time nodes of the same patient, constructs patient lesion longitudinal health data set, generates cell timing health data sample set;Based on the cell timing health data sample set, the first health assessment feature, the second health assessment feature and the third health assessment feature are extracted to carry out multi-dimensional health data splicing and cross-dimension weight fusion processing, generate lesion evolution comprehensive health data;The lesion evolution comprehensive health data is input into health data evaluation model, quantitatively calculates the individual health risk index of patient adenocarcinoma long-term progress, generates individual health risk assessment result output, realizes single physiological cell level dynamic evolution analysis, improves the accuracy of adenocarcinoma early identification and progress risk assessment, especially in the prediction field of cervical adenocarcinoma has more accurate prediction effect.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for predicting the risk of glandular epithelial lesions. Background Technology

[0002] Cervical glandular epithelial lesions are atypical proliferations, intraepithelial lesions, and malignant lesions occurring in the glandular epithelial cells of the cervical canal. Early diagnosis is crucial for improving patient prognosis. However, unlike common squamous epithelial lesions, glandular epithelial lesions often exhibit atypical morphological features in routine liquid-based thin-layer cytology, leading to diagnostic difficulties and a high rate of missed diagnoses. To improve the objectivity and accuracy of diagnosis, existing technologies have attempted to incorporate multimodal analysis methods. For example, combining liquid-based thin-layer cytology images with p16 / Ki-67 double-staining immunohistochemical results helps determine the nature of the lesion from both morphological and molecular dimensions. Other studies have constructed spatial relationship maps between cells to analyze abnormal patterns in cell distribution at single time points, or utilized imaging and clinical data from multiple patient follow-ups to predict the risk of lesion progression through time-series models. However, most of these methods are based on static images or macroscopic follow-up records, failing to delve into the individual cell level to track dynamic changes in morphology, polarity, and relationships with surrounding cells and tissues. Because the progression of glandular epithelial lesions is a gradual and heterogeneous process, it is difficult to capture subtle but crucial dynamic signals at the microscopic level, such as nuclear polarity disorder and glandular topological structure destruction, in the early stages of the lesion, relying solely on single or non-cellular level information. This limits the ability to identify early lesions or the risk of progression. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method for predicting the risk of glandular epithelial lesions, which achieves early identification and risk assessment of lesions by tracking the morphological and spatial relationship changes of individual cells in multimodal images over time.

[0004] To achieve the above objectives, this application provides a method for predicting the risk of glandular epithelial lesions, comprising:

[0005] Collect thin-layer cell diagnostic data from different time points of diagnosis and treatment for the same patient to construct a longitudinal multimodal health data set of the patient's lesions;

[0006] The range of cellular lesions is defined and cross-time period health data is matched for longitudinal multimodal health datasets. The temporal correlation of single physiological cells in different treatment periods is established to generate a cellular temporal health data sample set.

[0007] Based on a cellular time-series health data sample set, the cell nuclear polarity consistency health index at each diagnosis and treatment time point is calculated to construct the first health assessment feature;

[0008] Based on a cellular time-series health data sample set, and combined with the Delaunay triangulation algorithm, we statistically analyzed the spatial correlation patterns of cells at different diagnosis and treatment time points, extracted glandular spatial topological health parameters, and constructed a second health assessment feature.

[0009] Based on the cell time-series health data sample set, electronic medical record data or HPV test data at each diagnosis and treatment time point are obtained to construct a third health assessment feature;

[0010] Multi-dimensional health data splicing and cross-dimensional weight fusion processing are performed on the first health assessment feature, the second health assessment feature, and the third health assessment feature to generate comprehensive health data on disease evolution.

[0011] Input comprehensive health data on lesion evolution into the medical data diagnostic model, and output auxiliary diagnostic results of glandular epithelial lesions at the current diagnosis and treatment time point;

[0012] Simultaneously, comprehensive health data on lesion evolution is input into the health data assessment model to quantify the individual health risk index of long-term progression of patients' glandular epithelial lesions, and generate individual health risk assessment results.

[0013] Furthermore, the longitudinal multimodal health dataset is used to define the extent of cellular lesions and match health data across different time periods, establishing temporal correlations of single physiological cells at different treatment stages, and generating a cellular temporal health data sample set, including:

[0014] Cell thin-layer diagnostic data at each time point are input into a segmentation network with an encoder-decoder architecture. The encoder part of the segmentation network extracts image features through layer-by-layer convolution and downsampling operations, while the decoder part restores the image resolution through layer-by-layer deconvolution and upsampling operations. Finally, a probability response map with the same size as the input image is output. The value of each pixel position in the probability response map represents the confidence that the pixel is located in the cell nucleus region.

[0015] The probability response map is binarized, and the mask region and centroid coordinates of each cell nucleus are extracted using connected component analysis.

[0016] Based on the cell thin-layer diagnostic data at adjacent time points, the cell thin-layer diagnostic data at the previous time point is used as a fixed image, and the cell thin-layer diagnostic data at the next time point is used as a floating image. A B-spline free deformation model is constructed to perform spatial transformation on the floating image.

[0017] Using mutual information as a similarity measure, the negative mutual information between the fixed image and the transformed floating image is minimized through an iterative optimization algorithm, and the registered floating image and its corresponding cell nucleus centroid coordinate transformation results are output.

[0018] Based on the registered coordinates, search for the candidate matching target that is closest to the centroid of each cell nucleus at the previous time point in the cell nucleus centroid at the next time point.

[0019] Filter out candidate matching pairs whose distance is less than a preset matching threshold, and construct a candidate matching set;

[0020] The Hungarian algorithm is used to perform a globally optimal allocation of the candidate matching set, eliminate one-to-many conflicts, and establish a one-to-one correspondence between cell nuclei at adjacent time points.

[0021] The correspondence established between adjacent time points is concatenated along the time axis to construct an ordered sequence containing mask data and coordinate data at different time points for each continuously tracked cell, thereby generating a cell time series health data sample set.

[0022] Furthermore, based on the cellular time-series health data sample set, the cell nuclear polarity consistency health index at each diagnosis and treatment time point is calculated to construct the first health assessment feature, including:

[0023] For each time point, the covariance matrix of the mask region of each cell nucleus is calculated, and the covariance matrix is ​​decomposed into eigenvalues. The eigenvector corresponding to the largest eigenvalue after eigenvalue decomposition is determined as the major axis direction vector of the cell nucleus.

[0024] A circular neighborhood is constructed with the centroid of each cell nucleus as the center, and other cell nuclei within the neighborhood except for the nucleus itself are selected as the neighborhood set.

[0025] For each cell nucleus, calculate the absolute value of the cosine of the angle between its major axis direction vector and the major axis direction vector of each cell nucleus in the neighborhood set. Take the average of all the absolute cosine values ​​to obtain the nuclear polarity consistency health index of the cell nucleus at the corresponding time point.

[0026] For all time points, the nuclear polarity consistency health index of all cell nuclei at each time point is statistically analyzed to generate statistics including mean, variance and the proportion below a preset threshold, which constitute the first health assessment sub-feature of the corresponding time point.

[0027] The first health assessment sub-features at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the first health assessment features after global average pooling.

[0028] Furthermore, for each cell nucleus, the absolute value of the cosine of the angle between its major axis direction vector and the major axis direction vector of every cell nucleus in its neighborhood set is calculated. The average of all the absolute cosine values ​​is then used to obtain the nuclear polarity consistency health index of that cell nucleus at the corresponding time point, including:

[0029] Obtain the major axis direction vector of the target cell nucleus. The major axis direction vector is a two-dimensional vector containing a first direction component and a second direction component.

[0030] Obtain the major axis direction vector of each neighboring cell nucleus within the neighborhood set of the target cell nucleus. The major axis direction vector of each neighboring cell nucleus is a two-dimensional vector containing a first direction component and a second direction component.

[0031] Calculate the dot product between the major axis direction vector of the target cell nucleus and the major axis direction vector of each neighboring cell nucleus. The dot product is the product of the first direction component of the target cell nucleus and the first direction component of the neighboring cell nucleus, plus the product of the second direction component of the target cell nucleus and the second direction component of the neighboring cell nucleus.

[0032] Calculate the magnitude of the major axis vector of the target cell nucleus. The magnitude is the square root of the sum of the squares of the first and second directional components of the target cell nucleus.

[0033] Calculate the magnitude of the major axis direction vector of each neighboring cell nucleus. The magnitude is the square root of the sum of the squares of the first direction component and the squares of the second direction component of the neighboring cell nucleus.

[0034] Divide the dot product by the product of the modulus of the target cell nucleus and the modulus of the neighboring cell nuclei to generate the cosine of the angle between the target cell nucleus and each neighboring cell nucleus.

[0035] Take the absolute value of the cosine value to generate the absolute value of the cosine of the angle between the target cell nucleus and each neighboring cell nucleus;

[0036] The number of all neighboring cell nuclei in the neighborhood set is counted, and the sum of all the absolute values ​​of the cosines is divided by the number of all neighboring cell nuclei in the neighborhood set to generate the nuclear polarity consistency health index of the target cell nucleus at the corresponding time point.

[0037] Furthermore, based on the cellular time-series health data sample set, and combined with the Delaunay triangulation algorithm, the spatial correlation patterns of cells at different diagnosis and treatment time points are statistically analyzed, glandular spatial topological health parameters are extracted, and a second health assessment feature is constructed, including:

[0038] For each time point, using the centroid coordinates of all cell nuclei as the node set, Delaunay triangulation is performed to generate an initial edge set, edges with a length greater than a preset distance threshold are deleted, and a cell adjacency graph is constructed.

[0039] For the cell adjacency graph at each time point, calculate the local clustering coefficients of all nodes and take the average value to generate the average clustering coefficient; calculate the shortest path length between all node pairs and take the average value to generate the average shortest path length; calculate the information entropy of the node degree distribution to generate the node degree distribution entropy.

[0040] The average clustering coefficient, average shortest path length and node degree distribution entropy are combined to form the glandular spatial topological health parameters at the corresponding time points.

[0041] The glandular spatial topological health parameters at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the second health assessment features after global average pooling.

[0042] Furthermore, the local clustering coefficients of all nodes are calculated and averaged to generate the average clustering coefficient, including:

[0043] For each node in the cell adjacency graph, obtain all neighboring nodes of the node and count the actual number of connecting edges between neighboring nodes;

[0044] The maximum possible number of connecting edges between neighboring nodes is counted, and the actual number of connecting edges is divided by the maximum number of connecting edges to generate the local clustering coefficient of the nodes.

[0045] The average clustering coefficient is generated by summing the local clustering coefficients of all nodes and then dividing by the total number of nodes.

[0046] Calculate the shortest path length between all node pairs and take the average to generate the average shortest path length, including:

[0047] For each pair of nodes in the cell adjacency graph, a breadth-first search algorithm is used to traverse the cell adjacency graph to determine the number of edges traversed in the shortest path connecting each pair of nodes, and the number of edges is determined as the shortest path length between each pair of nodes.

[0048] The average shortest path length is generated by summing the shortest path lengths between all node pairs and dividing by the total number of node pairs.

[0049] Calculate the information entropy of the node degree distribution to generate the node degree distribution entropy, including:

[0050] Calculate the degree of each node in the cell adjacency graph. The degree of a node is the number of edges connected to that node.

[0051] Count the number of nodes corresponding to each degree value, divide the number of nodes corresponding to each degree value by the total number of nodes, and generate the probability of each degree value occurring.

[0052] Take the natural logarithm of the probability of each degree value, take the negative of the product of the probability and the natural logarithm, and sum the negative values ​​corresponding to all degree values ​​to generate the node degree distribution entropy.

[0053] Furthermore, based on the cellular time-series health data sample set, electronic medical record data or HPV test data at each diagnosis and treatment time point are obtained to construct a third health assessment feature, including:

[0054] For each time point, acquire the electronic medical record data or HPV test data corresponding to that time point;

[0055] Electronic medical record data or HPV test data are used as the third health assessment sub-feature at the corresponding time point;

[0056] The third health assessment sub-features at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the third health assessment features after global average pooling.

[0057] Furthermore, multi-dimensional health data splicing and cross-dimensional weighted fusion processing are performed on the first, second, and third health assessment features to generate comprehensive health data on disease evolution, including:

[0058] The first, second, and third health assessment features are concatenated along the feature dimension to generate initial comprehensive health data.

[0059] The initial comprehensive health data is input into the first linear transformation layer, the second linear transformation layer, and the third linear transformation layer, respectively. The first linear transformation layer generates a query vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term. The second linear transformation layer generates a key vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term. The third linear transformation layer generates a value vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term.

[0060] Calculate the dot product between the transpose of the query vector and the key vector, scale the dot product by dividing it by the square root of the query vector dimension, and input the scaled dot product into the Softmax function for normalization to generate the attention weight matrix.

[0061] Perform matrix multiplication on the attention weight matrix and the value vector to output comprehensive health data on lesion evolution after cross-modal attention weighting.

[0062] Furthermore, by inputting comprehensive health data on lesion evolution into the medical data diagnostic model, the model outputs auxiliary diagnostic results for glandular epithelial lesions at the current treatment time point, including:

[0063] The comprehensive health data of disease evolution is input into the first fully connected layer. The first fully connected layer multiplies the comprehensive health data of disease evolution with the first weight matrix and adds the first bias term to generate the first linear transformation result.

[0064] The result of the first linear transformation is input into the ReLU activation function, which sets negative values ​​to zero and keeps positive values ​​unchanged, and outputs the first hidden feature vector.

[0065] A random deactivation operation is performed on the first hidden feature vector, and some elements in the first hidden feature vector are randomly set to zero with a preset probability to generate a regularized hidden feature vector.

[0066] The regularized hidden feature vector is input into the second fully connected layer. The second fully connected layer multiplies the regularized hidden feature vector with the second weight matrix and adds the second bias term to generate the second linear transformation result.

[0067] The result of the second linear transformation is input into the Softmax activation function. The Softmax activation function maps each element in the result of the second linear transformation to a probability value between zero and one, and the sum of all elements is one. The output is the probability distribution corresponding to each preset lesion category.

[0068] Identify the maximum probability value from the probability distribution and determine the lesion category corresponding to the maximum probability value as the auxiliary diagnostic result of glandular epithelial lesions at the current time point.

[0069] Furthermore, by inputting comprehensive health data on disease evolution into a health data assessment model, the individual health risk index for the long-term progression of glandular epithelial lesions in patients is quantified, generating individual health risk assessment results, including:

[0070] The comprehensive health data on lesion evolution is input into the health data evaluation model, which contains a series of interconnected multi-layered dilated convolutional layers and a global average pooling layer.

[0071] The comprehensive health data of disease evolution is input into the first dilated convolutional layer. The first dilated convolutional layer performs convolution operation along the feature dimension through the first dilated convolutional kernel. The dilation factor is set to the first preset value, and the first temporal feature map is output.

[0072] The first temporal feature map is input into the second dilated convolutional layer. The second dilated convolutional layer performs convolution operation along the feature dimension through the second dilated convolutional kernel. The dilation factor is set to the second preset value, and the second temporal feature map is output.

[0073] The second temporal feature map is input into the third dilated convolutional layer. The third dilated convolutional layer performs convolution operation along the feature dimension through the third dilated convolutional kernel. The dilation factor is set to the third preset value, and the third temporal feature map is output.

[0074] The third temporal feature map is input into the global average pooling layer. The global average pooling layer calculates the average value of all elements in the feature dimension of the third temporal feature map to generate a compressed temporal feature vector.

[0075] The compressed temporal feature vector is input into the linear output layer. The linear output layer generates a linear transformation result by multiplying the compressed temporal feature vector by the output weight matrix and adding the output bias term.

[0076] The linear transformation result is input into the Sigmoid activation function, which maps the linear transformation result to a value between zero and one, and outputs an individual health risk index.

[0077] Unlike existing technologies, the above-mentioned technical solution provides a method for predicting the risk of glandular epithelial lesions. It constructs a longitudinal health data set of the patient's lesions by collecting thin-layer cellular diagnostic data from different treatment time points of the same patient. The longitudinal health data set is then used to define the extent of cellular lesions and match health data across different time periods, establishing a temporal correlation between single physiological cells at different treatment time points, generating a cell-based temporal health data sample set. Based on this sample set, a first, second, and third health assessment features are extracted, and these features are then combined using multi-dimensional health data splicing and cross-dimensional weighted fusion processing to generate comprehensive health data on lesion evolution. This comprehensive health data is input into a medical data diagnostic model to output auxiliary diagnostic results for glandular epithelial lesions at the current treatment time point. Simultaneously, it is input into a health data assessment model to quantify and calculate the individual health risk index for the long-term progression of the patient's glandular epithelial lesions, generating an individual health risk assessment result. This achieves dynamic evolutionary analysis at the single physiological cell level, improving the accuracy of early identification and progression risk assessment of glandular epithelial lesions, particularly demonstrating more precise predictive effects in the field of cervical glandular epithelial lesion prediction.

[0078] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0079] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0080] In the accompanying drawings of the instruction manual:

[0081] Figure 1 This is a schematic diagram illustrating steps S101 to S107 of the method described in the specific implementation embodiment;

[0082] Figure 2 This is a schematic diagram illustrating steps S201 to S208 of the method described in a specific implementation.

[0083] Figure 3This is a schematic diagram illustrating steps S301 to S305 of the method described in a specific implementation.

[0084] Figure 4 This is a schematic diagram illustrating steps S401 to S404 of the method described in a specific embodiment;

[0085] Figure 5 The following is a schematic diagram illustrating steps S501 to S503 of the method described in a specific implementation. Detailed Implementation

[0086] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0087] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0088] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0089] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0090] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0091] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0092] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0093] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0094] Please see Figure 1 This embodiment provides a method for predicting the risk of glandular epithelial lesions, including:

[0095] S101. Collect thin-layer cell diagnostic data from different time points of diagnosis and treatment for the same patient, and construct a longitudinal multimodal health data set of the patient's lesions;

[0096] S102. Define the extent of cellular lesions and match health data across time periods for the longitudinal multimodal health dataset, establish the temporal correlation of a single physiological cell in different treatment periods, and generate a cellular temporal health data sample set.

[0097] S103. Based on the cell time-series health data sample set, calculate the cell nuclear polarity consistency health index at each diagnosis and treatment time node, and construct the first health assessment feature;

[0098] S104. Based on the cell time-series health data sample set, the Delaunay triangulation algorithm is used to statistically analyze the cell spatial correlation patterns at different diagnosis and treatment time nodes, extract gland spatial topological health parameters, and construct a second health assessment feature.

[0099] S105. Based on the cell time-series health data sample set, obtain electronic medical record data or HPV test data at each diagnosis and treatment time node, and construct a third health assessment feature.

[0100] S106. Perform multi-dimensional health data splicing and cross-dimensional weight fusion processing on the first health assessment feature, the second health assessment feature, and the third health assessment feature to generate comprehensive health data on disease evolution.

[0101] S107. Input the comprehensive health data of lesion evolution into the medical data diagnostic model and output the auxiliary diagnostic results of glandular epithelial lesions at the current diagnosis and treatment time point; at the same time, input the comprehensive health data of lesion evolution into the health data assessment model, quantify and calculate the individual health risk index of the long-term progression of the patient's glandular epithelial lesions, and generate the individual health risk assessment results output.

[0102] It should be noted that the glandular epithelial cells shown in this embodiment include the glandular epithelium of the uterus, the glandular epithelium of the breast, the glandular epithelium of the prostate, etc. The steps shown in this embodiment can be applied to most glandular epithelial cells for risk prediction. Preferably, the steps shown in this embodiment are more suitable for risk prediction and analysis of cervical glandular epithelial cells.

[0103] In step S101, the thin-layer cell diagnostic data is obtained by preparing thin-layer cell sheets from exfoliated cervical cells using liquid-based slide preparation technology and then scanning them, providing cell morphology information. The longitudinal health data set of the patient's lesions provides a data foundation across time points for subsequent analysis.

[0104] In step S102, the extent of cellular lesions is defined by extracting the contour and location information of each cell nucleus from each thin-layer diagnostic data image using image segmentation technology. Cross-temporal health data matching utilizes image registration algorithms to spatially align data from different time points, eliminating geometric distortions introduced by slide preparation, staining, or scanning. Temporal correlations are then established in the registered data using a matching algorithm to create a one-to-one correspondence between the same cell at different time points. The cellular temporal health data sample set is a structured sequence set of information such as masks and coordinates corresponding to each successfully tracked cell at different time points. This step transforms discrete static data into temporally correlated cell-level sequences, providing data support for subsequent extraction of dynamic evolutionary features.

[0105] In step S103, the nuclear polarity consistency health index is calculated based on the degree of consistency between the long axis of the cell nucleus and the long axis of other nuclei in its local neighborhood. The value directly corresponds to the severity of the pathological feature of glandular epithelial cell nuclear polarity disorder. The first health assessment feature, extracted by temporal modeling of the nuclear polarity consistency health index of the same tissue region at different time points, can characterize the evolutionary trend of the degree of nuclear polarity disorder over time, including its dynamic patterns of exacerbation, stabilization, or improvement. This step transforms static nuclear polarity indicators into temporal evolution features, achieving dynamic quantification of the key pathological feature of glandular epithelial cell nuclear polarity disorder.

[0106] In step S104, Delaunay triangulation uses the centroid coordinates of all cell nuclei as nodes and generates a triangular mesh covering the entire cell region through geometric calculation. The edges of this triangular mesh naturally reflect the spatial adjacency relationships between cells. The cell adjacency graph is constructed using cell nuclei as nodes and the edges generated by triangulation as connections. Some long edges can be filtered according to preset conditions to retain adjacency relationships with actual biological significance. Glandular spatial topological health parameters are quantitative indicators extracted from the cell adjacency graph, reflecting the degree of damage to glandular structures, such as the tightness of cell connections or the distribution of path lengths. The second health assessment feature is extracted by temporally modeling the values ​​of these glandular spatial topological health parameters at different time points, which can characterize the gradual destruction process of glandular structures over time, such as pathological changes like glandular crowding, common wall formation, or the appearance of cribriform structures. This step extends the static analysis of glandular topology to temporal evolution analysis, realizing the dynamic tracking of glandular structural damage.

[0107] It should be noted that the liquid-based slide preparation process disrupts the macroscopic tissue structure of the gland, resulting in a random distribution of cells on the slide. However, diseased cells and normal cells differ significantly in physical properties (such as cell size, nucleocytoplasmic ratio, and cell membrane adhesion). These differences cause diseased cells to tend to form non-random spatial distribution patterns at the microscopic level during slide preparation, such as local clustering or sparse distribution. Based on this, this step uses the centroid coordinates of all cell nuclei as nodes and constructs a cell adjacency graph using the Delaunay triangulation algorithm. It then extracts graph topological features such as the average clustering coefficient, average shortest path length, and node degree distribution entropy as parameters of cell spatial distribution heterogeneity. These parameters statistically quantify the non-randomness of cell spatial distribution, capturing subtle spatial anomalies that are difficult to detect with conventional morphological analysis. Unlike existing techniques that rely on intact tissue sections for glandular structure analysis, the method shown in this step does not depend on whether cells are arranged in a regular morphology. Instead, it provides supplementary information for lesion identification from the perspective of spatial distribution patterns by analyzing the spatial adjacency relationships between the centroids of discrete cell nuclei.

[0108] In step S105, electronic medical record data or HPV test data, as indicators reflecting the patient's clinical status or viral infection status, are used to provide clinical or molecular-level information related to disease progression. The third health assessment feature, extracted by temporal modeling of the values ​​of electronic medical record data or HPV test data at different time points, can characterize the evolutionary trend of the patient's clinical status or viral infection status over time. This step, by introducing the temporal dimension and changes of clinical or virological information, supplements the quantitative basis for disease evolution from a clinical risk or etiological perspective, complementing nuclear polarity and glandular topological features.

[0109] Specifically, the retrieval and collection of electronic medical record information may include the following content, as shown in Table 1:

[0110] The hospital's electronic medical record system was used to retrieve and collect demographic information, present illness history, menstrual history, personal history, past medical history, family history, clinical diagnosis, surgical records, and diagnostic information from the pathology information system, totaling 43 data elements (see Table 1). Present illness history recorded high-risk information such as vaginal bleeding and abnormal discharge; menstrual history recorded information such as the age of menarche, menstrual cycle, and menstrual patterns; personal history covered information such as lifestyle habits, sexual behavior, and reproductive history; past medical history recorded information such as past medical history, surgical history, and drug allergies; family history recorded information on hereditary diseases in the immediate family members; clinical diagnosis was the disease judgment made by clinicians based on the patient's symptoms, signs, and auxiliary examination results; surgical records included the process, surgical method, and extent of surgical resection for cervical lesions; and the pathology information system recorded quality control information and diagnostic results for cytopathology and histopathology in a structured data format, which could verify the consistency between cytological and histological diagnoses. HPV testing is another technology for cervical cancer screening, which can indicate whether the person being screened is infected with the HPV virus and the type of strain.

[0111] Table 1. Summary of data related to cervical cancer screening in the electronic medical record system.

[0112]

[0113] In step S106, the multi-dimensional health data concatenation connects the first, second, and third health assessment features along their respective feature dimensions to generate initial comprehensive health data containing all three types of health assessment information. Cross-dimensional weighted fusion processing automatically learns the contribution weights of different health assessment features to subsequent diagnostic and assessment tasks through an attention mechanism, weighting and optimizing the initial comprehensive health data to enable the model to adaptively highlight key dimension information and suppress redundant information. The disease evolution comprehensive health data is the comprehensive feature data after attention-weighted optimization. This feature data integrates the temporal evolution information of the three dimensions, providing a unified feature representation for subsequent disease diagnosis and risk assessment.

[0114] In step S107, the medical data diagnostic model receives comprehensive health data on lesion evolution as input. Preferably, it maps this data to a probability distribution of preset lesion categories through multi-layer nonlinear transformation, selecting the category with the highest probability as the auxiliary diagnostic result for glandular epithelial lesions at the current treatment time point. The health data assessment model receives the same comprehensive health data on lesion evolution as input and maps it to a value between 0 and 1 through time-series modeling. This value is the individual health risk index; a higher value indicates a greater risk of developing a higher-level lesion in the future. This step, through a dual-task output architecture, simultaneously achieves accurate diagnosis of the current state and quantifies the risk of future trends, providing end-to-end information support for clinical decision-making from current diagnosis to prognostic assessment.

[0115] This embodiment introduces a time dimension, transforming discrete static data into temporally correlated sequence data through cross-time periodic tracking at the single physiological cell level. Based on this, health assessment features with clear pathological significance are extracted from three dimensions: cell nuclear polarity, glandular spatial topology, and electronic medical record data or HPV detection data. Adaptive fusion is then performed through a cross-dimensional weighted fusion mechanism, enabling the model to comprehensively utilize the dynamic changes of multiple pathological features. Finally, through the collaborative output of the medical data diagnostic model and the health data assessment model, both current auxiliary diagnostic results for glandular epithelial lesions and individual health risk indices are provided, offering technical support for the early diagnosis and follow-up management of glandular epithelial lesions and achieving dynamic evolution analysis and accurate identification of glandular epithelial lesions.

[0116] Please see Figure 2 In some embodiments, the longitudinal multimodal health data set is used to define the extent of cellular lesions and match health data across time periods, establishing the temporal correlation of single physiological cells at different treatment periods, and generating a cellular temporal health data sample set, including:

[0117] S201. Input the cell thin-layer diagnostic data at each time point into the segmentation network of the encoder-decoder architecture. The encoder part of the segmentation network extracts image features through layer-by-layer convolution and downsampling operations, and the decoder part restores the image resolution through layer-by-layer deconvolution and upsampling operations. Finally, output a probability response map with the same size as the input image. The value of each pixel position in the probability response map represents the confidence that the pixel is located in the cell nucleus region.

[0118] S202. The probability response map is binarized, and the mask region and centroid coordinates of each cell nucleus are extracted using connected component analysis.

[0119] S203. Based on the cell thin-layer diagnostic data at adjacent time points, the cell thin-layer diagnostic data at the previous time point is used as a fixed image, and the cell thin-layer diagnostic data at the next time point is used as a floating image. A B-spline free deformation model is constructed to perform spatial transformation on the floating image.

[0120] S204. Using mutual information as a similarity measure, minimize the negative mutual information between the fixed image and the transformed floating image through an iterative optimization algorithm, and output the registered floating image and its corresponding cell nucleus centroid coordinate transformation results.

[0121] S205. Based on the registered coordinates, search for the candidate matching target that is closest to the centroid of each cell nucleus at the previous time point in the cell nucleus centroid at the next time point.

[0122] S206. Filter out candidate matching pairs whose distance is less than the preset matching threshold and construct a candidate matching set;

[0123] S207. Use the Hungarian algorithm to perform a global optimal allocation of the candidate matching set, eliminate one-to-many conflicts, and establish a one-to-one correspondence between cell nuclei at adjacent time points.

[0124] S208. Connect the correspondence established at each adjacent time point along the time axis to construct an ordered sequence containing the mask data and coordinate data at different time points for each continuously tracked cell, and generate a cell time series health data sample set.

[0125] In step S201, the encoder-decoder architecture segmentation network progressively compresses the image spatial resolution and extracts high-level semantic features through layer-by-layer convolution and downsampling operations, and then progressively restores the image spatial resolution through layer-by-layer deconvolution and upsampling operations, finally outputting a probability response map with the same size as the input image. This segmentation network is trained on a large amount of thin-layer diagnostic data of cells with labeled cell nucleus contours. During training, the network parameters are optimized by minimizing the cross-entropy loss between the predicted probability response map and the labeled mask. The value of each pixel in the probability response map is between 0 and 1. The closer the value is to 1, the higher the confidence that the pixel is located in the cell nucleus region; the closer the value is to 0, the higher the confidence that the pixel is located in the background region.

[0126] In step S202, the binarization process compares the value of each pixel in the probability response image with a preset threshold. Pixels with values ​​greater than the threshold are set to 1, and pixels with values ​​less than the threshold are set to 0, generating a binary mask image. Preferably, the preset threshold can be determined by optimizing the segmentation performance index on a validation set. Connected component analysis marks and extracts connected regions in the binary mask image. Each connected region corresponds to a cell nucleus within a lesion area. The centroid coordinates of the cell nucleus are obtained by calculating the average of the coordinates of all pixels within the connected region.

[0127] In step S203, the B-spline free deformation model describes the nonlinear deformation of the image using a grid of control points distributed in the image space. Each control point carries a displacement vector, and the displacements between control points are obtained through B-spline basis function interpolation. The thin-layer cell diagnostic data from the previous time point is used as a fixed image, and the thin-layer cell diagnostic data from the next time point is used as a floating image. The floating image is spatially transformed by adjusting the displacement vectors of the control points, so that the transformed floating image is geometrically aligned with the fixed image. Preferably, the grid spacing of the control points can be set according to the image resolution and the average size of the cell nuclei.

[0128] In step S204, mutual information, as a similarity measure, is used to measure the degree of information overlap between two images by statistically analyzing the joint probability distribution of pixel grayscale values ​​in the fixed image and the transformed floating image. The larger the mutual information value, the more similar the two images are. The iterative optimization algorithm repeatedly adjusts the displacement vectors of the control points in the B-spline free deformation model to minimize the negative mutual information between the fixed image and the transformed floating image. When the optimization converges, it outputs the registered floating image and the spatially transformed coordinates of the centroid of each cell nucleus.

[0129] In step S205, based on the registered coordinates, among all nuclear centroids at the later time point, the nuclear centroid with the closest Euclidean distance for each nuclear centroid at the previous time point is searched, and the nuclear centroid corresponding to this closest distance is taken as a candidate matching target. The Euclidean distance is obtained by calculating the square root of the sum of the squares of the differences between the coordinates of the two centroids.

[0130] In step S206, a preset matching threshold is used to limit the spatial proximity range of candidate matching pairs. Only when the Euclidean distance between the cell nucleus centroid at the previous time point and the candidate matching target at the next time point is less than the threshold is the pair included in the candidate matching set. The preset matching threshold can be determined based on the average diameter of the cell nucleus or the image resolution to ensure the biological rationality of cross-time period health data matching.

[0131] In step S207, the Hungarian algorithm constructs a cost matrix to perform a globally optimal allocation of the candidate matching set. The rows of the cost matrix correspond to cell nuclei at the previous time point, the columns correspond to cell nuclei at the next time point, and the matrix elements are the distance values ​​of the corresponding pairings. The Hungarian algorithm searches for the matching scheme with the minimum total cost in the cost matrix, ensuring that each cell nucleus at the previous time point matches at most one cell nucleus at the next time point, and vice versa, thereby eliminating one-to-many conflicts and establishing a one-to-one correspondence between cell nuclei at adjacent time points.

[0132] In step S208, the correspondence established between adjacent time points is concatenated along the time axis to construct an ordered sequence containing mask data and coordinate data at different time points for each continuously tracked cell. For cells that appear or disappear at a certain time point, the time point of their appearance or disappearance can be marked in the sequence. This ordered sequence is the cell time-series health data sample set, with each successfully tracked cell corresponding to an independent time series. The data at each time point in the sequence includes the mask region and centroid coordinates of the cell at that time point.

[0133] This embodiment achieves precise segmentation and localization of cell nuclei in defining the extent of cellular lesions through a segmentation network with an encoder-decoder architecture. It employs a B-spline free-form model and mutual information metric to achieve non-rigid registration in cross-time-period health data matching, and utilizes the Hungarian algorithm to achieve globally optimal matching and tracking of cell nuclei, thereby establishing the temporal correlation between single physiological cells at different treatment stages. This embodiment transforms discrete static data into a temporally correlated set of cellular time-series health data samples, providing a precise cellular-level data foundation for subsequent extraction of dynamic evolutionary features such as the first, second, and third health assessment features.

[0134] Please see Figure 3In some embodiments, based on a cellular time-series health data sample set, a cell nuclear polarity consistency health index is calculated at each diagnosis and treatment time point to construct a first health assessment feature, including:

[0135] S301. For each time point, calculate the covariance matrix of the mask region of each cell nucleus, perform eigenvalue decomposition on the covariance matrix, and determine the eigenvector corresponding to the largest eigenvalue after eigenvalue decomposition as the major axis direction vector of the cell nucleus.

[0136] S302. Construct a circular neighborhood with the centroid of each cell nucleus as the center, and select other cell nuclei within the neighborhood except for the nucleus itself as the neighborhood set.

[0137] S303. For each cell nucleus, calculate the absolute value of the cosine of the angle between its major axis direction vector and the major axis direction vector of each cell nucleus in the neighborhood set. Take the average value of all the absolute cosine values ​​to obtain the nuclear polarity consistency health index of the cell nucleus at the corresponding time point.

[0138] S304. For all time points, perform statistical analysis on the nuclear polarity consistency health index of all cell nuclei at each time point to generate statistics including mean, variance and the proportion below the preset threshold, which constitute the first health assessment sub-feature of the corresponding time point.

[0139] S305. Arrange the first health assessment sub-features of all time points in chronological order and input them into the temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the first health assessment features after global average pooling.

[0140] In step S301, the covariance matrix is ​​calculated from the two-dimensional coordinates of all pixels within the cell nucleus mask region. The diagonal elements of this matrix reflect the distribution variance of pixels in the horizontal and vertical directions, while the off-diagonal elements reflect the distribution covariance of pixels in the horizontal and vertical directions. After eigenvalue decomposition of the covariance matrix, two eigenvalues ​​and their corresponding eigenvectors are obtained. The eigenvector corresponding to the largest eigenvalue points in the direction of the largest pixel distribution variance, which is the long axis direction of the cell nucleus, reflecting the spatial extension orientation of the cell nucleus. The mask region and centroid coordinates are provided by the results of cell lesion extent definition and connected component analysis in step S202.

[0141] In step S302, a circular neighborhood is defined with the centroid of each cell nucleus as the center and a preset radius. This radius is set based on the average size of the cell nucleus and the intercellular distance, ensuring that the neighborhood covers the adjacent cell nuclei surrounding the current nucleus. Other cell nuclei within the circular neighborhood, excluding the current nucleus itself, are selected to form the neighborhood set of that cell nucleus. For cell nuclei located at the image edge or in sparsely populated areas, resulting in a small number of nuclei in the neighborhood set, they can be marked as isolated cell nuclei and assigned a default nuclear polarity consistency health index.

[0142] In step S303, the absolute cosine value is obtained by calculating the absolute value of the cosine of the angle between the long axis direction vector of the target cell nucleus and the long axis direction vector of each neighboring cell nucleus in the neighborhood set. This value is between 0 and 1. The arithmetic mean of all the absolute cosine values ​​in the neighborhood set is taken to obtain the nuclear polarity consistency health index of the cell nucleus at the corresponding time point. The closer the index is to 1, the more consistent the long axis direction of the cell nucleus is with the long axis direction of the neighboring cell nuclei. The closer the index is to 0, the more the direction tends to be perpendicular or randomly distributed.

[0143] In step S304, statistical analysis summarizes the nuclear polarity consistency health index of all effective cell nuclei at the same time point, generating three statistics: mean, variance, and the proportion below a preset threshold. The mean reflects the overall level of nuclear polarity consistency at that time point, the variance reflects the degree of difference in polarity consistency between different cell nuclei, and the proportion below the preset threshold reflects the percentage of cells with disordered nuclear polarity. The preset threshold can be set based on pathological diagnostic experience and is used to define whether nuclear polarity is in a disordered state. The combination of the three statistics forms the first health assessment sub-feature for the corresponding time point, which characterizes the nuclear polarity distribution characteristics of the tissue region at that time point using three-dimensional numerical values.

[0144] In step S305, the temporal convolutional network takes as input a feature sequence consisting of the first health assessment sub-features at all time points arranged chronologically, and extracts features layer by layer along the time dimension through multiple dilated convolutions. Dilated convolutions expand the receptive field by inserting holes between convolutional kernel elements, enabling the network to simultaneously capture short-term changes at nearby time points and long-term trends at more distant time points. The output of the multiple dilated convolutions is compressed along the time dimension by a global average pooling layer, aggregating the temporal information of each feature channel into a scalar value, and finally outputting the first health assessment feature. This temporal convolutional network is trained based on the first health assessment sub-feature sequence of historical patients and the corresponding pathological diagnosis labels. During training, the network parameters are optimized by minimizing the difference between the predicted output and the label.

[0145] This embodiment quantifies the long axis direction of each cell nucleus through covariance matrix and eigenvalue decomposition. A nuclear polarity consistency health index is calculated based on circular neighborhoods to reflect the degree of nuclear polarity disorder. Statistical analysis aggregates single-cell indicators into tissue-level first health assessment sub-features. A temporal convolutional network is then used to extract the dynamic evolution pattern of nuclear polarity along the time dimension, ultimately generating the first health assessment feature. This embodiment transforms nuclear polarity, a morphological feature with clear pathological significance, from a static indicator into a temporal evolution feature, providing crucial information reflecting the dynamic changes of nuclear polarity disorder for subsequent multi-dimensional fusion and lesion identification.

[0146] In some embodiments, for each cell nucleus, the absolute value of the cosine of the angle between its major axis direction vector and the major axis direction vector of each cell nucleus in the neighborhood set is calculated. The average of all the absolute cosine values ​​is taken to obtain the nuclear polarity consistency health index of the cell nucleus at the corresponding time point, including:

[0147] Obtain the major axis direction vector of the target cell nucleus. The major axis direction vector is a two-dimensional vector containing a first direction component and a second direction component.

[0148] Obtain the major axis direction vector of each neighboring cell nucleus within the neighborhood set of the target cell nucleus. The major axis direction vector of each neighboring cell nucleus is a two-dimensional vector containing a first direction component and a second direction component.

[0149] Calculate the dot product between the major axis direction vector of the target cell nucleus and the major axis direction vector of each neighboring cell nucleus. The dot product is the product of the first direction component of the target cell nucleus and the first direction component of the neighboring cell nucleus, plus the product of the second direction component of the target cell nucleus and the second direction component of the neighboring cell nucleus.

[0150] Calculate the magnitude of the major axis vector of the target cell nucleus. The magnitude is the square root of the sum of the squares of the first and second directional components of the target cell nucleus.

[0151] Calculate the magnitude of the major axis direction vector of each neighboring cell nucleus. The magnitude is the square root of the sum of the squares of the first direction component and the squares of the second direction component of the neighboring cell nucleus.

[0152] Divide the dot product by the product of the modulus of the target cell nucleus and the modulus of the neighboring cell nuclei to generate the cosine of the angle between the target cell nucleus and each neighboring cell nucleus.

[0153] Take the absolute value of the cosine value to generate the absolute value of the cosine of the angle between the target cell nucleus and each neighboring cell nucleus;

[0154] The number of all neighboring cell nuclei in the neighborhood set is counted, and the sum of all the absolute values ​​of the cosines is divided by the number of all neighboring cell nuclei in the neighborhood set to generate the nuclear polarity consistency health index of the target cell nucleus at the corresponding time point.

[0155] In this embodiment, the major axis direction vector of the target cell nucleus is obtained by the eigenvalue decomposition of the covariance matrix in step S301. This vector is a two-dimensional vector, where the first directional component represents the projection component in the horizontal direction of the image, and the second directional component represents the projection component in the vertical direction of the image. The major axis direction vector of each neighboring cell nucleus is also obtained by step S301, and it also contains the first directional component and the second directional component. These vectors together constitute the data basis for subsequent calculations.

[0156] The dot product is obtained by multiplying the first directional component of the target cell nucleus by the first directional components of neighboring cell nuclei, and then multiplying the second directional component of the target cell nucleus by the second directional components of neighboring cell nuclei. This result reflects the degree of alignment between the two vectors in direction; a larger value indicates a more consistent orientation. The magnitude of the target cell nucleus is obtained by taking the square root of the sum of the squares of its first and second directional components. The magnitude of each neighboring cell nucleus is obtained using the same calculation method. The magnitude is used to eliminate the influence of vector length differences on the angle calculation.

[0157] The cosine value is generated by dividing the dot product by the product of the modulus of the target cell nucleus and the modulus of the neighboring cell nuclei. This value, between -1 and +1, represents the cosine function value of the angle between the two vectors. Taking the absolute value of the cosine value compresses the range to between zero and one, eliminating the possibility of a negative cosine value due to opposite directions. This ensures that cell nuclei with parallel but opposite directions are also identified as having a consistent polarity, which aligns with the pathological judgment logic of nuclear polarity analysis that focuses on directional consistency rather than specific orientation.

[0158] The target cell nucleus's nuclear polarity consistency health index at the corresponding time point is generated by summing the absolute values ​​of all neighboring cell nuclei within the neighborhood set and dividing by the total number of nuclei. If the number of nuclei in the neighborhood set is too small, this step can be skipped, and a default value can be directly assigned to the nucleus to avoid bias caused by small sample statistics. This nuclear polarity consistency health index comprehensively reflects the degree of consistency between the long axis of the target cell nucleus and the long axis of surrounding cell nuclei using a single value. The closer the value is to one, the more consistent the polarity; the closer the value is to zero, the more disordered the polarity.

[0159] This embodiment precisely quantifies the angular relationship between the direction vectors of the long axis of the cell nucleus through the calculation of dot product and modulus. Absolute value operations are used to match the calculation results with the pathological significance of nuclear polarity analysis, and the average value is used to aggregate directional consistency information within a local neighborhood into a single index. This embodiment provides a clear and reproducible calculation process for the nuclear polarity consistency health index, ensuring the quantitative accuracy and biological interpretability of this indicator.

[0160] Please see Figure 4In some embodiments, based on a cellular time-series health data sample set, and combined with the Delaunay triangulation algorithm to statistically analyze the spatial correlation patterns of cells at different diagnosis and treatment time points, glandular spatial topological health parameters are extracted to construct a second health assessment feature, including:

[0161] S401. For each time point, using the centroid coordinates of all cell nuclei as the node set, perform Delaunay triangulation to generate an initial edge set, delete edges with edge lengths greater than a preset distance threshold, and construct a cell adjacency graph.

[0162] S402. For the cell adjacency graph at each time point, calculate the local clustering coefficients of all nodes and take the average value to generate the average clustering coefficient; calculate the shortest path length between all node pairs and take the average value to generate the average shortest path length; calculate the information entropy of the node degree distribution and generate the node degree distribution entropy.

[0163] S403. Combine the average clustering coefficient, average shortest path length and node degree distribution entropy to form the glandular spatial topological health parameters at the corresponding time point.

[0164] S404. Arrange the gland spatial topology health parameters of all time points in chronological order and input them into the temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the second health assessment features after global average pooling.

[0165] In step S401, Delaunay triangulation uses the centroid coordinates of all cell nuclei as the node set. A set of non-overlapping triangular meshes is generated using a point-by-point insertion algorithm or a divide-and-conquer algorithm. Each edge of this triangular mesh connects two spatially adjacent cell nuclei centroids, and each triangle's circumcircle is an empty circle containing no other nodes, thus naturally reflecting the spatial adjacency relationships between cells. In the initial edge set generated by Delaunay triangulation, some edges with excessively long lengths may connect spatially non-adjacent cell nuclei. These edges can be filtered and deleted using a preset distance threshold. The preset distance threshold is set based on the average internuclear spacing and cell density; preferably, it can be determined by referring to a multiple of the average cell nucleus diameter to retain adjacency relationships with practical biological significance and eliminate noisy connections. The filtered edges and all nodes together constitute a cell adjacency graph, which is an undirected graph where nodes correspond to cell nuclei and edges correspond to the spatial adjacency relationships between cell nuclei.

[0166] In step S402, the average clustering coefficient is obtained by calculating the arithmetic mean of the local clustering coefficients of each node in the cell adjacency graph. The local clustering coefficient measures the degree of interconnection between a node's neighboring nodes. A higher coefficient indicates a more compact connection between the cell nuclei surrounding the node, corresponding to cell crowding or shared-wall phenomena in the glandular structure. The average shortest path length is obtained by calculating the arithmetic mean of the shortest path lengths between all pairs of nodes in the cell adjacency graph. The shortest path length measures the minimum number of edges traversed between two nodes. A smaller value indicates higher information transmission efficiency between cell nuclei, corresponding to the compactness of the glandular structure. The node degree distribution entropy is obtained by statistically analyzing the degree of each node in the cell adjacency graph (i.e., the number of edges connected to that node) and calculating the information entropy of the probability of degree values ​​occurring. A larger entropy value indicates a more dispersed distribution of node degrees, corresponding to a higher degree of heterogeneity in the glandular structure. These three graph topological features quantify the degree of disruption of the glandular spatial topology from different dimensions: the average clustering coefficient focuses on local clustering, the average shortest path length focuses on global connectivity, and the node degree distribution entropy focuses on structural heterogeneity.

[0167] In step S403, the average clustering coefficient, average shortest path length, and node degree distribution entropy are concatenated in a fixed order to form a three-dimensional vector, which is the glandular spatial topological health parameter at the corresponding time point. This three-dimensional numerical representation comprehensively characterizes the glandular topological structure features of the tissue region at that time point. The three components of this glandular spatial topological health parameter correspond to a quantitative description of the glandular structure in terms of local clustering, global connectivity, and structural heterogeneity.

[0168] In step S404, the temporal convolutional network takes as input a feature sequence composed of glandular spatial topological health parameters arranged chronologically at all time points. It extracts features layer by layer along the time dimension through multi-layer dilated convolutions. Dilated convolutions expand the receptive field by inserting holes between convolutional kernel elements, enabling the network to simultaneously capture short-term changes at nearby time points and long-term trends at more distant time points. The output of the multi-layer dilated convolutions is compressed along the time dimension by a global average pooling layer, aggregating the temporal information of each feature channel into a scalar value, ultimately outputting the second health assessment feature. This temporal convolutional network is trained based on the glandular spatial topological health parameter sequence of historical patients and corresponding pathological diagnostic labels. During training, the network parameters are optimized by minimizing the difference between the predicted output and the label.

[0169] This embodiment constructs a cell adjacency graph reflecting the spatial adjacency relationships of cells using Delaunay triangulation. It extracts three graph topological features with clear pathological significance: average clustering coefficient, average shortest path length, and node degree distribution entropy. Furthermore, it employs a temporal convolutional network to extract the dynamic evolution pattern of the glandular spatial topology along the time dimension, ultimately generating a second health assessment feature. This embodiment transforms the glandular spatial topology—a key pathological feature reflecting the degree of glandular damage—from a static indicator into a temporal evolution feature, providing crucial information reflecting the dynamic changes in glandular structure for subsequent multi-dimensional fusion and lesion identification.

[0170] In some embodiments, the local clustering coefficients of all nodes are calculated and averaged to generate an average clustering coefficient, including:

[0171] For each node in the cell adjacency graph, obtain all neighboring nodes of the node and count the actual number of connecting edges between neighboring nodes;

[0172] The maximum possible number of connecting edges between neighboring nodes is counted, and the actual number of connecting edges is divided by the maximum number of connecting edges to generate the local clustering coefficient of the nodes.

[0173] The average clustering coefficient is generated by summing the local clustering coefficients of all nodes and then dividing by the total number of nodes.

[0174] Calculate the shortest path length between all node pairs and take the average to generate the average shortest path length, including:

[0175] For each pair of nodes in the cell adjacency graph, a breadth-first search algorithm is used to traverse the cell adjacency graph to determine the number of edges traversed in the shortest path connecting each pair of nodes, and the number of edges is determined as the shortest path length between each pair of nodes.

[0176] The average shortest path length is generated by summing the shortest path lengths between all node pairs and dividing by the total number of node pairs.

[0177] Calculate the information entropy of the node degree distribution to generate the node degree distribution entropy, including:

[0178] Calculate the degree of each node in the cell adjacency graph. The degree of a node is the number of edges connected to that node.

[0179] Count the number of nodes corresponding to each degree value, divide the number of nodes corresponding to each degree value by the total number of nodes, and generate the probability of each degree value occurring.

[0180] Take the natural logarithm of the probability of each degree value, take the negative of the product of the probability and the natural logarithm, and sum the negative values ​​corresponding to all degree values ​​to generate the node degree distribution entropy.

[0181] In this embodiment, the average clustering coefficient is calculated by traversing the adjacency relationships of each node in the cell adjacency graph. If the number of neighboring nodes of a node is less than 2, the local clustering coefficient of that node is directly defined as zero to avoid calculation errors due to a zero denominator. The actual number of connecting edges is obtained by iterating through neighboring node pairs in a double loop and querying the edge set of the cell adjacency graph. The count is incremented by one for each pair of neighboring nodes found to have an edge connection. The average clustering coefficient is generated by summing the local clustering coefficients of all nodes and dividing by the total number of nodes. The closer this coefficient is to one, the higher the degree of local clustering of nodes in the cell adjacency graph, corresponding to the severity of cell crowding or shared-wall phenomena in the glandular structure.

[0182] To calculate the average shortest path length, a breadth-first search is performed starting from each node, expanding outwards layer by layer to visit neighboring nodes, and recording the shortest path length from the starting point to all other nodes. If there are disconnected regions in the cell adjacency graph, i.e., there is no path between two nodes, the shortest path length between these two nodes is defined as the total number of nodes to ensure that subsequent average calculations are not affected by infinity. The total number of node pairs is obtained by multiplying the total number of nodes by one and dividing by two. The average shortest path length is generated by summing the shortest path lengths between all node pairs and dividing by the total number of node pairs. The smaller this value, the higher the connectivity efficiency between nodes in the cell adjacency graph, and the more compact the corresponding glandular structure.

[0183] To calculate the node degree distribution entropy, the degree of each node is obtained by counting the number of its neighboring nodes in the cell adjacency graph. The possible range of the degree value is zero to the total number of nodes minus one. When counting the number of nodes corresponding to each degree value, if a certain degree value does not appear, its probability is zero, and the corresponding entropy contribution term is directly defined as zero to avoid calculation errors caused by taking the natural logarithm of zero. The node degree distribution entropy is generated by summing the entropy contribution terms corresponding to all appearing degree values. The larger the entropy value, the more dispersed the distribution of node degrees. The presence of both low-degree and high-degree nodes in the cell adjacency graph reflects a higher degree of heterogeneity in the glandular structure.

[0184] This embodiment quantifies the topological features of the cell adjacency graph from three dimensions—local clustering, global connectivity, and structural heterogeneity—by precisely calculating the local clustering coefficient, shortest path length, and node degree distribution entropy. The average clustering coefficient reflects the degree of clustering of glandular cells, the average shortest path length reflects the connectivity efficiency of the glandular structure, and the node degree distribution entropy reflects the degree of heterogeneity of the glandular structure. These three features complement each other, providing a precise quantitative basis for the subsequent extraction of glandular topological temporal evolution features.

[0185] Please see Figure 5In some embodiments, based on a cellular time-series health data sample set, electronic medical record data or HPV test data at each diagnosis and treatment time point are obtained to construct a third health assessment feature, including:

[0186] S501. For each time point, obtain the electronic medical record data or HPV test data corresponding to that time point;

[0187] S502. Use electronic medical record data or HPV test data as the third health assessment sub-feature at the corresponding time point.

[0188] S503. Arrange the third health assessment sub-features of all time points in chronological order and input them into the temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the third health assessment features after global average pooling.

[0189] In step S501, electronic medical record data is obtained by structured extraction of clinical records during the patient's diagnosis and treatment process, including but not limited to textual or numerical information such as the patient's age, past medical history, cervical lesion screening history, and treatment history; HPV test data is obtained by HPV nucleic acid testing on cervical exfoliated cell samples from the patient, including HPV genotype information (such as high-risk types such as HPV16 and HPV18) and viral load information. The electronic medical record data or HPV test data at each time point serves as the original feature data corresponding to that time point.

[0190] In step S502, the electronic medical record data or HPV test data acquired at each time point are directly used as the third health assessment sub-feature for the corresponding time point. This sub-feature, in the form of raw data or a numerical vector after being formatted according to a basic format, represents the patient's clinical status or viral infection status at that time point.

[0191] In step S503, the third health assessment sub-feature vectors of all time points are arranged chronologically and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the third health assessment features after global average pooling. This temporal convolutional network is trained based on the third health assessment sub-feature sequence of historical patients and the corresponding pathological diagnosis labels. During training, the network parameters are optimized by minimizing the difference between the predicted output and the label.

[0192] This embodiment acquires electronic medical record data or HPV test data at each diagnosis and treatment time point, uses them as the third health assessment sub-feature at the corresponding time point, and uses a temporal convolutional network to extract the dynamic evolution pattern of clinical status or viral infection status along the time dimension, and finally generates the third health assessment feature, which provides key information reflecting the dynamic changes of the patient's clinical risk or viral infection status for subsequent multi-dimensional fusion and lesion identification.

[0193] In some embodiments, multi-dimensional health data splicing and cross-dimensional weighted fusion processing are performed on the first health assessment feature, the second health assessment feature, and the third health assessment feature to generate comprehensive health data on disease evolution, including:

[0194] The first, second, and third health assessment features are concatenated along the feature dimension to generate initial comprehensive health data.

[0195] The initial comprehensive health data is input into the first linear transformation layer, the second linear transformation layer, and the third linear transformation layer, respectively. The first linear transformation layer generates a query vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term. The second linear transformation layer generates a key vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term. The third linear transformation layer generates a value vector by multiplying the initial comprehensive health data with the weight matrix and adding a bias term.

[0196] Calculate the dot product between the transpose of the query vector and the key vector, scale the dot product by dividing it by the square root of the query vector dimension, and input the scaled dot product into the Softmax function for normalization to generate the attention weight matrix.

[0197] Perform matrix multiplication on the attention weight matrix and the value vector to output comprehensive health data on lesion evolution after cross-modal attention weighting.

[0198] In this embodiment, the first health assessment feature, the second health assessment feature, and the third health assessment feature are independently output by temporal convolutional networks from different sources. The dimensions of the three features are preferably set to the same value. If the dimensions are inconsistent, an additional linear projection layer can be used to unify the dimensions before performing the concatenation operation, ensuring that the information of each dimension has a balanced representational capability during the fusion process. The concatenation operation itself does not introduce learnable parameters; it simply combines multiple feature vectors into a single vector.

[0199] The first, second, and third linear transformation layers each contain independent weight matrices and bias terms. These parameters are learned through an end-to-end training process based on historical patient data and corresponding pathological diagnostic labels. During training, gradients are backpropagated to the three linear transformation layers via attention calculations and subsequent diagnostic or evaluation losses, enabling them to gradually learn the optimal mapping method for the initial comprehensive health data into query vectors, key vectors, and value vectors.

[0200] The dot product between the transpose of the query vector and the key vector reflects the correlation strength between features in each dimension. The scaling factor is determined by the square root of the query vector's dimension. When the query vector has a high dimension, the variance of the dot product increases with the dimension, causing the Softmax function to enter the gradient saturation region. The scaling operation can control the variance of the dot product within a stable range independent of dimension, thus maintaining the discriminability of the attention weight distribution. The scaled dot product is normalized by the Softmax function to generate an attention weight matrix. Each element in this matrix represents the degree of attention from one dimension to another, with rows corresponding to the query vector's dimension and columns corresponding to the key vector's dimension.

[0201] The comprehensive health data on disease evolution, output by matrix multiplication of the attention weight matrix and the value vector, will be simultaneously input into both the subsequent medical data diagnosis model and the health data assessment model. The two models share the same feature representation, achieving feature reuse in multi-task learning. The dimension of the comprehensive health data on disease evolution is the same as the dimension of the value vector. This dimension is determined as a hyperparameter before model training and can be selected through cross-validation to balance feature representation capability and computational efficiency.

[0202] This embodiment maps the initial comprehensive health data into query vectors, key vectors, and value vectors through a linear transformation layer. It uses a scaling dot product attention mechanism to calculate the correlation weights between each dimension, enabling the model to adaptively adjust the contribution of the first, second, and third health assessment features to subsequent tasks. This solves the problem that simple concatenation of multi-dimensional features cannot distinguish the importance of each dimension, and provides a feature representation that highlights key points and fully integrates information for subsequent auxiliary diagnosis of glandular epithelial lesions and individual health risk assessment.

[0203] In some embodiments, comprehensive health data on lesion evolution is input into a medical data diagnostic model, which outputs auxiliary diagnostic results for glandular epithelial lesions at the current treatment time point, including:

[0204] The comprehensive health data of disease evolution is input into the first fully connected layer. The first fully connected layer multiplies the comprehensive health data of disease evolution with the first weight matrix and adds the first bias term to generate the first linear transformation result.

[0205] The result of the first linear transformation is input into the ReLU activation function, which sets negative values ​​to zero and keeps positive values ​​unchanged, and outputs the first hidden feature vector.

[0206] A random deactivation operation is performed on the first hidden feature vector, and some elements in the first hidden feature vector are randomly set to zero with a preset probability to generate a regularized hidden feature vector.

[0207] The regularized hidden feature vector is input into the second fully connected layer. The second fully connected layer multiplies the regularized hidden feature vector with the second weight matrix and adds the second bias term to generate the second linear transformation result.

[0208] The result of the second linear transformation is input into the Softmax activation function. The Softmax activation function maps each element in the result of the second linear transformation to a probability value between zero and one, and the sum of all elements is one. The output is the probability distribution corresponding to each preset lesion category.

[0209] Identify the maximum probability value from the probability distribution and determine the lesion category corresponding to the maximum probability value as the auxiliary diagnostic result of glandular epithelial lesions at the current time point.

[0210] In this embodiment, the first weight matrix and the first bias term are obtained through model training based on historical patient disease evolution comprehensive health data and corresponding pathological diagnosis gold standard labels. During training, the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label, and the network parameters are optimized through the backpropagation algorithm. The output dimension of the first fully connected layer is used as a hyperparameter, and preferably, the dimension value that optimizes the classification performance is selected on the validation set through cross-validation.

[0211] The ReLU activation function sets negative elements to zero and leaves positive elements unchanged in the first linear transformation result, introducing nonlinear transformation capability that allows the network to fit the complex nonlinear decision boundary between features and lesion categories. In the random deactivation operation, the specific value of the preset probability is determined through validation set performance tuning. This operation is only enabled during the training phase and disabled during the inference phase. By randomly disabling some neurons, it forces the network to learn redundant feature representations, thereby suppressing overfitting and improving the model's generalization ability on unseen data.

[0212] The second weight matrix and the second bias term are also obtained through model training. The output dimension of the second fully connected layer is equal to the number of preset lesion categories. The preset lesion categories are determined according to clinical diagnostic criteria and include three categories: normal glandular epithelial cells, glandular epithelial dysplasia, and glandular epithelial lesions, corresponding to the main classifications in the diagnosis of glandular epithelial lesions.

[0213] The Softmax activation function maps each element in the second linear transformation result to a probability value between zero and one, where the sum of all elements is one. This allows the output to be directly interpreted as the confidence level of the comprehensive health data on lesion evolution belonging to each preset lesion category. Identifying the lesion category corresponding to the highest probability value from the probability distribution yields the auxiliary diagnostic result for glandular epithelial lesions at the current time point.

[0214] This embodiment uses a two-layer fully connected network to perform nonlinear mapping and classification decisions on comprehensive health data of lesion evolution. It uses random inactivation to suppress the risk of overfitting and outputs an interpretable diagnostic result through the probability distribution of the Softmax function, providing end-to-end classification output for the automatic identification of glandular epithelial lesions.

[0215] In some embodiments, comprehensive health data on disease evolution is input into a health data assessment model to quantify and calculate an individual health risk index for the long-term progression of glandular epithelial lesions in patients, generating individual health risk assessment results output, including:

[0216] The comprehensive health data on lesion evolution is input into the health data evaluation model, which contains a series of interconnected multi-layered dilated convolutional layers and a global average pooling layer.

[0217] The comprehensive health data of disease evolution is input into the first dilated convolutional layer. The first dilated convolutional layer performs convolution operation along the feature dimension through the first dilated convolutional kernel. The dilation factor is set to the first preset value, and the first temporal feature map is output.

[0218] The first temporal feature map is input into the second dilated convolutional layer. The second dilated convolutional layer performs convolution operation along the feature dimension through the second dilated convolutional kernel. The dilation factor is set to the second preset value, and the second temporal feature map is output.

[0219] The second temporal feature map is input into the third dilated convolutional layer. The third dilated convolutional layer performs convolution operation along the feature dimension through the third dilated convolutional kernel. The dilation factor is set to the third preset value, and the third temporal feature map is output.

[0220] The third temporal feature map is input into the global average pooling layer. The global average pooling layer calculates the average value of all elements in the feature dimension of the third temporal feature map to generate a compressed temporal feature vector.

[0221] The compressed temporal feature vector is input into the linear output layer. The linear output layer generates a linear transformation result by multiplying the compressed temporal feature vector by the output weight matrix and adding the output bias term.

[0222] The linear transformation result is input into the Sigmoid activation function, which maps the linear transformation result to a value between zero and one, and outputs an individual health risk index.

[0223] In this embodiment, the health data assessment model receives comprehensive health data on disease evolution as input. The structure of this model is similar to the temporal convolutional network used to extract temporal evolution features, but here it is used to map the fused features to a single individual health risk index rather than extracting temporal evolution patterns.

[0224] Preferably, the first, second, and third preset values ​​are set as a progressively increasing numerical sequence, which expands the receptive field of the first, second, and third dilated convolutional layers sequentially, enabling the capture of dependencies at different time scales from comprehensive health data on disease evolution. The specific values ​​of the three dilation factors are determined through validation set performance tuning to balance the ability to capture short-term fluctuations and long-term trends. The sizes of the first, second, and third dilated convolutional kernels are preferably set to the same value to simplify network design and ensure that changes in the dilation factors effectively adjust the receptive field.

[0225] The global average pooling layer calculates the average of all elements in the third temporal feature map along the feature dimension, compressing the multi-channel temporal feature map into a fixed-length compressed temporal feature vector. This operation reduces the number of network parameters and suppresses the risk of overfitting. The output weight matrix and output bias term in the linear output layer are learned through model training based on historical patient disease evolution comprehensive health data and corresponding progression outcome labels. During training, the binary cross-entropy loss function is used to measure the difference between the predicted individual health risk index and the actual progression label. The Sigmoid activation function maps the linear transformation result to a value between zero and one, which is the individual health risk index. The closer the value is to one, the higher the risk of developing a higher-level disease in the future; the closer the value is to zero, the lower the risk of progression.

[0226] This embodiment expands the receptive field layer by layer through multiple dilated convolutional layers to capture the dependencies between different time scales in the comprehensive health data of disease evolution. After compressing the feature dimension through a global average pooling layer, an individual health risk index between zero and one is output through a linear output layer and a Sigmoid activation function, providing a basis for quantitatively assessing the risk of disease progression in clinical follow-up.

[0227] The above technical solution provides a method for predicting the risk of glandular epithelial lesions. It constructs a longitudinal health data set of the patient's lesions by collecting thin-layer cellular diagnostic data from different treatment time points of the same patient. The longitudinal health data set is then used to define the extent of cellular lesions and match health data across different time periods, establishing a temporal correlation between single physiological cells at different treatment time points, generating a cell-based temporal health data sample set. Based on this sample set, a first, second, and third health assessment features are extracted, and these features are then combined using multi-dimensional health data splicing and cross-dimensional weighted fusion processing to generate comprehensive health data on lesion evolution. This comprehensive health data is input into a medical data diagnostic model to output the auxiliary diagnostic result of glandular epithelial lesions at the current treatment time point. Simultaneously, the comprehensive health data is input into a health data assessment model to quantify and calculate the individual health risk index for the long-term progression of the patient's glandular epithelial lesions, generating an individual health risk assessment result. This achieves dynamic evolutionary analysis at the single physiological cell level, improving the accuracy of early identification and progression risk assessment of glandular epithelial lesions, particularly demonstrating more accurate predictive effects in the field of cervical glandular epithelial lesion prediction.

[0228] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A method for predicting the risk of glandular epithelial lesions, characterized in that, include: Collect thin-layer cell diagnostic data from different time points of diagnosis and treatment for the same patient to construct a longitudinal multimodal health data set of the patient's lesions; The longitudinal multimodal health data set is used to define the range of cellular lesions and match health data across time periods, establish the temporal correlation of a single physiological cell in different treatment periods, and generate a cellular temporal health data sample set; Based on the cell time-series health data sample set, the cell nuclear polarity consistency health index at each diagnosis and treatment time point is calculated to construct the first health assessment feature; Based on the aforementioned cellular time-series health data sample set, and combined with the Delaunay triangulation algorithm, the spatial correlation patterns of cells at different diagnosis and treatment time nodes are statistically analyzed, glandular spatial topological health parameters are extracted, and a second health assessment feature is constructed. Based on the cell time-series health data sample set, electronic medical record data or HPV test data at each diagnosis and treatment time point are obtained to construct a third health assessment feature; Multi-dimensional health data splicing and cross-dimensional weight fusion processing are performed on the first health assessment feature, the second health assessment feature, and the third health assessment feature to generate comprehensive health data on disease evolution. The comprehensive health data on the evolution of the disease is input into the medical data diagnostic model, and the auxiliary diagnostic results of glandular epithelial lesions at the current diagnosis and treatment time point are output. Simultaneously, the comprehensive health data on the disease evolution is input into the health data assessment model to quantify and calculate the individual health risk index of the long-term progression of the patient's glandular epithelial lesions, and generate individual health risk assessment results.

2. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, The longitudinal multimodal health data set is used to define the extent of cellular lesions and match health data across different time periods, establishing the temporal correlation of single physiological cells at different treatment stages, and generating a cellular temporal health data sample set, including: Cell thin-layer diagnostic data at each time point are input into a segmentation network with an encoder-decoder architecture. The encoder part of the segmentation network extracts image features through layer-by-layer convolution and downsampling operations, and the decoder part restores the image resolution through layer-by-layer deconvolution and upsampling operations. Finally, a probability response map with the same size as the input image is output. The value of each pixel position in the probability response map represents the confidence that the pixel is located in the cell nucleus region. The probability response map is binarized, and the mask region and centroid coordinates of each cell nucleus are extracted using connected component analysis. Based on the cell thin-layer diagnostic data at adjacent time points, the cell thin-layer diagnostic data at the previous time point is used as a fixed image, and the cell thin-layer diagnostic data at the next time point is used as a floating image. A B-spline free deformation model is constructed to perform spatial transformation on the floating image. Using mutual information as a similarity measure, the negative mutual information between the fixed image and the transformed floating image is minimized through an iterative optimization algorithm, and the registered floating image and its corresponding cell nucleus centroid coordinate transformation results are output. Based on the registered coordinates, search for the candidate matching target that is closest to the centroid of each cell nucleus at the previous time point in the cell nucleus centroid at the next time point. Filter out candidate matching pairs whose distance is less than a preset matching threshold, and construct a candidate matching set; The Hungarian algorithm is used to perform a globally optimal allocation of the candidate matching set, eliminating one-to-many conflicts and establishing a one-to-one correspondence between cell nuclei at adjacent time points; The correspondence established at each adjacent time point is concatenated along the time axis to construct an ordered sequence containing mask data and coordinate data at different time points for each continuously tracked cell, thereby generating the cell time-series health data sample set.

3. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, Based on the aforementioned cellular time-series health data sample set, the cell nuclear polarity consistency health index at each diagnosis and treatment time point is calculated to construct a first health assessment feature, including: For each time point, the covariance matrix of the mask region of each cell nucleus is calculated, and the covariance matrix is ​​decomposed into eigenvalues. The eigenvector corresponding to the largest eigenvalue after eigenvalue decomposition is determined as the major axis direction vector of the cell nucleus. A circular neighborhood is constructed with the centroid of each cell nucleus as the center, and other cell nuclei within the neighborhood except for the nucleus itself are selected as the neighborhood set. For each cell nucleus, calculate the absolute value of the cosine of the angle between its major axis direction vector and the major axis direction vector of each cell nucleus in the neighborhood set, and take the average of all the said absolute cosine values ​​to obtain the nuclear polarity consistency health index of the cell nucleus at the corresponding time point; For all time points, the nuclear polarity consistency health index of all cell nuclei at each time point is statistically analyzed to generate statistics including mean, variance and the proportion below a preset threshold, which constitute the first health assessment sub-feature of the corresponding time point. The first health assessment sub-features at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the first health assessment features after global average pooling.

4. The method for predicting the risk of glandular epithelial lesions according to claim 3, characterized in that, For each cell nucleus, the absolute cosine of the angle between its major axis direction vector and the major axis direction vector of every cell nucleus in its neighborhood set is calculated. The average of all these absolute cosine values ​​is used to obtain the nuclear polarity consistency health index of that cell nucleus at the corresponding time point, including: Obtain the major axis direction vector of the target cell nucleus, wherein the major axis direction vector is a two-dimensional vector containing a first direction component and a second direction component; Obtain the major axis direction vector of each neighboring cell nucleus within the neighborhood set of the target cell nucleus. The major axis direction vector of each neighboring cell nucleus is a two-dimensional vector containing a first direction component and a second direction component. Calculate the dot product between the major axis direction vector of the target cell nucleus and the major axis direction vector of each of the neighboring cell nuclei. The dot product is the product of the first direction component of the target cell nucleus and the first direction component of the neighboring cell nuclei, plus the product of the second direction component of the target cell nucleus and the second direction component of the neighboring cell nuclei. Calculate the magnitude of the major axis vector of the target cell nucleus, where the magnitude is the square root of the sum of the squares of the first and second directional components of the target cell nucleus. Calculate the magnitude of the major axis direction vector of each of the neighboring cell nuclei, where the magnitude is the square root of the sum of the squares of the first direction component and the squares of the second direction component of the neighboring cell nuclei; Divide the dot product by the product of the modulus of the target cell nucleus and the modulus of the neighboring cell nuclei to generate the cosine of the angle between the target cell nucleus and each of the neighboring cell nuclei. The absolute value of the cosine is taken to generate the absolute value of the cosine of the angle between the target cell nucleus and each of the neighboring cell nuclei; The number of all neighboring cell nuclei in the neighborhood set is counted, and the sum of all the absolute values ​​of the cosines is divided by the number of all neighboring cell nuclei in the neighborhood set to generate the nuclear polarity consistency health index of the target cell nucleus at the corresponding time point.

5. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, Based on the aforementioned cellular time-series health data sample set, and combined with the Delaunay triangulation algorithm to statistically analyze the spatial correlation patterns of cells at different treatment time points, glandular spatial topological health parameters are extracted to construct a second health assessment feature, including: For each time point, using the centroid coordinates of all cell nuclei as the node set, Delaunay triangulation is performed to generate an initial edge set, edges with a length greater than a preset distance threshold are deleted, and a cell adjacency graph is constructed. For the cell adjacency graph at each time point, calculate the local clustering coefficients of all nodes and take the average value to generate the average clustering coefficient; calculate the shortest path length between all node pairs and take the average value to generate the average shortest path length; calculate the information entropy of the node degree distribution to generate the node degree distribution entropy. The average clustering coefficient, the average shortest path length, and the node degree distribution entropy are combined to form the glandular spatial topological health parameters at the corresponding time point. The glandular spatial topological health parameters at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the second health assessment features after global average pooling.

6. The method for predicting the risk of glandular epithelial lesions according to claim 5, characterized in that, Calculate the local clustering coefficients of all nodes and take the average to generate the average clustering coefficient, including: For each node in the cell adjacency graph, obtain all neighboring nodes of the node and count the actual number of connection edges between the neighboring nodes; The maximum possible number of connection edges between the neighboring nodes is counted, and the actual number of connection edges is divided by the maximum number of connection edges to generate the local clustering coefficient of the node. The average clustering coefficient is generated by summing the local clustering coefficients of all nodes and dividing by the total number of nodes. Calculate the shortest path length between all node pairs and take the average to generate the average shortest path length, including: For each pair of nodes in the cell adjacency graph, a breadth-first search algorithm is used to traverse the cell adjacency graph to determine the number of edges traversed by the shortest path connecting each pair of nodes, and the number of edges is determined as the shortest path length between each pair of nodes. The average shortest path length is generated by summing the shortest path lengths between all node pairs and dividing by the total number of node pairs. Calculate the information entropy of the node degree distribution to generate the node degree distribution entropy, including: The degree of each node in the cell adjacency graph is calculated, where the degree of a node is the number of edges connected to that node. Count the number of nodes corresponding to each degree value, divide the number of nodes corresponding to each degree value by the total number of nodes, and generate the probability of each degree value occurring. Take the natural logarithm of the probability of each degree value, take the negative value of the product of the probability and the natural logarithm, and sum the negative values ​​corresponding to all degree values ​​to generate the node degree distribution entropy.

7. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, Based on the aforementioned cellular time-series health data sample set, electronic medical record data or HPV test data at each diagnosis and treatment time point are obtained to construct a third health assessment feature, including: For each time point, acquire the electronic medical record data or HPV test data corresponding to that time point; The electronic medical record data or HPV test data are used as the third health assessment sub-feature at the corresponding time point; The third health assessment sub-features at all time points are arranged in chronological order and input into a temporal convolutional network. The temporal convolutional network extracts features along the time dimension through multiple dilated convolutions and outputs the third health assessment features after global average pooling.

8. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, The first, second, and third health assessment features are subjected to multi-dimensional health data splicing and cross-dimensional weighted fusion processing to generate comprehensive health data on disease evolution, including: The first health assessment feature, the second health assessment feature, and the third health assessment feature are concatenated along the feature dimension to generate initial comprehensive health data. The initial comprehensive health data is input into the first linear transformation layer, the second linear transformation layer, and the third linear transformation layer, respectively. The first linear transformation layer generates a query vector by multiplying the initial comprehensive health data with a weight matrix and adding a bias term. The second linear transformation layer generates a key vector by multiplying the initial comprehensive health data with a weight matrix and adding a bias term. The third linear transformation layer generates a value vector by multiplying the initial comprehensive health data with a weight matrix and adding a bias term. Calculate the dot product between the query vector and the transpose of the key vector, scale the dot product by dividing it by the square root of the query vector dimension, and input the scaled dot product result into the Softmax function for normalization to generate the attention weight matrix. Perform matrix multiplication between the attention weight matrix and the value vector to output cross-modal attention-weighted comprehensive health data on lesion evolution.

9. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, The comprehensive health data on the disease progression is input into the medical data diagnostic model, which outputs the auxiliary diagnostic results of glandular epithelial lesions at the current treatment time point, including: The comprehensive health data of disease evolution is input into the first fully connected layer. The first fully connected layer multiplies the comprehensive health data of disease evolution with the first weight matrix and adds a first bias term to generate a first linear transformation result. The first linear transformation result is input into the ReLU activation function, which sets negative values ​​to zero and keeps positive values ​​unchanged, and outputs the first hidden feature vector. A random deactivation operation is performed on the first hidden feature vector, and some elements in the first hidden feature vector are randomly set to zero with a preset probability to generate a regularized hidden feature vector. The regularized hidden feature vector is input into the second fully connected layer. The second fully connected layer multiplies the regularized hidden feature vector with the second weight matrix and adds a second bias term to generate the second linear transformation result. The result of the second linear transformation is input into the Softmax activation function, which maps each element in the result of the second linear transformation to a probability value between zero and one, and the sum of all elements is one, and outputs the probability distribution corresponding to each preset lesion category. The maximum probability value is identified from the probability distribution, and the lesion category corresponding to the maximum probability value is determined as the auxiliary diagnostic result of glandular epithelial lesions at the current time point.

10. The method for predicting the risk of glandular epithelial lesions according to claim 1, characterized in that, The comprehensive health data on the disease progression is input into a health data assessment model to quantify and calculate the individual health risk index for the long-term progression of the patient's glandular epithelial lesions, generating individual health risk assessment results, including: The disease evolution integrated health data is input into the health data evaluation model, which includes a series of interconnected multi-layered dilated convolutional layers and a global average pooling layer. The disease evolution comprehensive health data is input into the first dilated convolutional layer. The first dilated convolutional layer performs convolution operation along the feature dimension through the first dilated convolutional kernel. The dilation factor is set to the first preset value, and the first temporal feature map is output. The first temporal feature map is input into the second dilated convolutional layer. The second dilated convolutional layer performs convolution operation along the feature dimension through the second dilated convolutional kernel. The dilation factor is set to the second preset value, and the second temporal feature map is output. The second temporal feature map is input into the third dilated convolutional layer. The third dilated convolutional layer performs convolution operation along the feature dimension through the third dilated convolutional kernel. The dilation factor is set to the third preset value, and the third temporal feature map is output. The third temporal feature map is input into a global average pooling layer, which calculates the average value of all elements in the feature dimension of the third temporal feature map to generate a compressed temporal feature vector. The compressed temporal feature vector is input into the linear output layer. The linear output layer generates a linear transformation result by multiplying the compressed temporal feature vector by the output weight matrix and superimposing the output bias term. The linear transformation result is input into the Sigmoid activation function, which maps the linear transformation result to a value between zero and one, and outputs an individual health risk index.

Images