A precancerous transformation early warning system and method for oral leukoplakia

Abstract

The present application relates to the technical field of medical image processing, and provides a precancerous lesion of oral leukoplakia early warning system and method, comprising: establishing an association model of oral leukoplakia image features and blood tumor marker time series, when the last sampling points of the two are not synchronized, generating predictive blood risk representation data based on the current leukoplakia image features; constructing a cross-modal time series trajectory according to the leukoplakia image features and blood tumor marker time series collected multiple times, identifying abnormal trajectory patterns; and generating a precancerous lesion risk level of oral leukoplakia accordingly. The present application effectively solves the information fault problem caused by the asynchronization of image data and blood data sampling time points, and improves the real-time performance and accuracy of risk assessment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an early warning system and method for malignant transformation of oral leukoplakia. Background Technology

[0002] Oral leukoplakia is a common clinical phenotype among potential malignant lesions in the oral cavity, and its risk assessment for malignant transformation is of great significance in the field of oral medicine. Current technologies typically employ oral image recognition for automatic detection of leukoplakia, or provide risk alerts based on single blood tumor marker test results. However, these approaches mostly remain at the level of single-modal analysis. Image analysis methods often focus only on lesion area or color classification, and hematological methods often rely solely on whether a single test value exceeds a reference range. This makes it difficult to form a precise, dynamic, and individualized assessment of the risk of malignant transformation in leukoplakia.

[0003] More importantly, the local morphological changes in oral leukoplakia and the changes in blood tumor markers are often asynchronous in clinical practice. Patients may frequently have images of leukoplakia taken during outpatient visits, follow-up examinations, or even home visits, but blood tumor marker tests are usually performed on a fixed schedule of several weeks or months. This results in a significant time gap between image data and blood data at any given decision point. If the most recent time point matching strategy is directly adopted, it is easy to overlook the significant progression of the lesion that may have occurred between the blood test and the current image acquisition, as well as the temporal differences that hematological changes may precede local morphological changes, or that high-risk local morphology may precede hematological abnormalities. Therefore, there is an urgent need for an early warning scheme that can solve the problem of asynchronous sampling of image data and blood data and enable multimodal data correlation analysis. Summary of the Invention

[0004] To address the problem of insufficient accuracy in risk assessment caused by the asynchronous sampling time points of oral leukoplakia image data and blood tumor marker data in existing technologies, this application proposes an oral leukoplakia malignant transformation early warning system and method to achieve cross-modal correlation analysis and risk warning between image features and blood markers.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An early warning system for malignant transformation of oral leukoplakia includes: a cross-modal association modeling module, used to establish an association model between oral leukoplakia image features and time series of blood tumor markers, and to generate predictive blood risk characterization data based on the current leukoplakia image features when the last sampling point of the blood tumor marker time series is not synchronized with the last sampling point of the leukoplakia image features; a trajectory coupling analysis module, used to construct a cross-modal time series trajectory based on the leukoplakia image features and the blood tumor marker time series collected multiple times, and to identify abnormal trajectory patterns; and a risk assessment module, used to generate a risk level for malignant transformation of oral leukoplakia based on the predictive blood risk characterization data and the abnormal trajectory patterns.

[0006] The above scheme effectively bridges the time gap between image and blood data sampling points by establishing a cross-modal correlation model and generating predictive blood risk characterization data when data is asynchronous, thereby improving the real-time performance and accuracy of risk assessment. At the same time, it enhances the ability to capture complex disease evolution processes by identifying abnormal patterns through trajectory coupling analysis.

[0007] Preferably, the cross-modal association modeling module is used to generate the predictive blood risk characterization data based on the current and historical image feature change trends when the last sampling point of the blood tumor marker time series is earlier than the last sampling point of the vitiligo image features.

[0008] Preferably, the cross-modal association modeling module is used to generate the predictive blood risk characterization data based on the vitiligo internal heterogeneity score, edge anomaly score, and the historical image feature change trend, through one or more of the following methods: time extrapolation, risk mapping, or weight estimation.

[0009] Preferably, the abnormal trajectory pattern includes at least one of the following: an image-first pattern in which high-risk features of vitiligo images precede abnormal changes in blood tumor markers; an indicator-first pattern in which abnormal changes in blood tumor markers precede high-risk features of vitiligo images; and a parallel pattern in which high-risk features of vitiligo images and abnormal changes in blood tumor markers deteriorate synchronously.

[0010] The above-mentioned preferred scheme can more accurately capture different paths of disease evolution and improve the pertinence of early warning by identifying various abnormal patterns such as image-first, indicator-first, or synchronous deterioration.

[0011] Preferably, the white spot image features include a white spot internal heterogeneity score and an edge anomaly score; wherein, the white spot internal heterogeneity score is calculated based on the color difference, texture difference and roughness difference between multiple sub-regions within the white spot region; the edge anomaly score is calculated based on the morphological complexity, boundary ambiguity and local outward expansion trend of the white spot edge region.

[0012] Furthermore, this invention also provides a method for early warning of malignant transformation of oral leukoplakia, comprising the following steps: S100, establishing a correlation model between oral leukoplakia image features and time series of blood tumor markers, and generating predictive blood risk characterization data based on the current leukoplakia image features when the last sampling point of the blood tumor marker time series is not synchronized with the last sampling point of the leukoplakia image features; S200, constructing a cross-modal time series trajectory based on the leukoplakia image features and the blood tumor marker time series collected multiple times, and identifying abnormal trajectory patterns; S300, generating a risk level for malignant transformation of oral leukoplakia based on the predictive blood risk characterization data and the abnormal trajectory patterns.

[0013] Preferably, step S100 includes: comparing the time difference between the last sampling point of the blood tumor marker time series and the last sampling point of the vitiligo image features; when the time difference exceeds a preset time synchronization threshold window, generating the predictive blood risk characterization data based on the current vitiligo image features.

[0014] Preferably, the generation of the predictive blood risk characterization data in step S100 includes: generating the predictive blood risk characterization data by one or more of the following methods: time extrapolation, risk mapping, or weight estimation, based on the internal heterogeneity score of vitiligo, the edge abnormality score, and the trend of historical image feature changes.

[0015] Preferably, the abnormal trajectory pattern identified in step S200 includes at least one of the following: an image-first pattern in which high-risk features of vitiligo images precede abnormal changes in blood tumor markers; an indicator-first pattern in which abnormal changes in blood tumor markers precede the high-risk features of vitiligo images; and a parallel pattern in which the high-risk features of vitiligo images and the abnormal changes in blood tumor markers deteriorate synchronously.

[0016] Preferably, the method further includes: obtaining the pathological biopsy results or postoperative histological results corresponding to the risk level of malignant transformation of oral leukoplakia; recording the error between the risk level of malignant transformation of oral leukoplakia and the pathological biopsy results or postoperative histological results; and updating the leukoplakia image feature weights, blood tumor marker weights, trajectory coupling parameters and the time synchronization threshold window according to the error when the number of error samples reaches a preset number.

[0017] The above-mentioned preferred scheme improves the system's generalization ability and clinical reliability by introducing a histological results feedback learning mechanism, which allows pathological biopsy conclusions to serve as true labels to continuously correct model parameters.

[0018] Beneficial effects: This invention establishes a cross-modal association model between oral leukoplakia image features and time series of blood tumor markers, and generates predictive blood risk characterization data when data sampling times are asynchronous. This effectively solves the information gap problem caused by asynchronous sampling times of image and blood data, improving the real-time performance and accuracy of risk assessment. By constructing a cross-modal trajectory coupling analysis mechanism, it can identify different abnormal patterns such as image-leading, marker-leading, or synchronous deterioration, thereby enhancing the ability to capture complex disease evolution processes. By introducing a histological result feedback learning mechanism, pathological biopsy conclusions can be used as true labels to continuously correct model parameters, improving the system's generalization ability and clinical credibility. Furthermore, by performing multi-dimensional heterogeneity analysis on leukoplakia image features, it can identify focal heterogeneous changes that are difficult to stabilize and quantify in traditional manual observation, improving the early identification ability of high-risk leukoplakia lesions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall module structure of the oral leukoplakia malignant transformation early warning system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the cross-modal correlation modeling and predictive blood risk characterization data generation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the cross-modal trajectory coupling analysis process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall processing flow of the oral leukoplakia malignant transformation early warning method according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the calculation process for vitiligo heterogeneity scoring and edge abnormality scoring in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. Example 1:

[0022] like Figure 1As shown in the figure, this embodiment provides an early warning system for the malignant transformation of oral leukoplakia. The system includes a cross-modal association modeling module, a trajectory coupling analysis module, and a risk assessment module.

[0023] The cross-modal association modeling module establishes an association model between oral leukoplakia image features and time series of blood tumor markers. When the last sampling point of the blood tumor marker time series is out of sync with the last sampling point of the leukoplakia image features, it generates predictive blood risk characterization data based on the current leukoplakia image features. This module is crucial for addressing the time gap between high-frequency image acquisition and low-frequency blood testing in clinical practice. Typically, patients may frequently have oral images taken, but blood tests are conducted over a longer period, leading to a mismatch in the timeline of the data. This module, through the established association model, utilizes the inherent relationship between image features and blood indicators to deduce blood risk characterization data matching the current image time point, thereby eliminating the assessment blind spots caused by time asynchrony. For example, if a patient's most recent blood test was two weeks ago, and a new oral leukoplakia image has just been taken, this module can predict the potential risk status of the blood system at the current moment based on the current image features and the association model, rather than simply using blood data from two weeks ago.

[0024] The trajectory coupling analysis module is used to construct cross-modal time-series trajectories based on the features of the vitiligo images acquired multiple times and the time series of blood tumor markers, and to identify abnormal trajectory patterns. This module receives synchronized data processed by the cross-modal association modeling module, or historical synchronized data, to construct dual-track trajectories showing the evolution of image features and blood indicators over time. Its core function is to capture the dynamic process of disease evolution, rather than relying solely on single snapshot data. By analyzing the temporal relationship between the two trajectories—for example, whether the deterioration of image features precedes the abnormality of blood indicators, or vice versa—this module can identify different abnormal trajectory patterns, providing a dynamic dimension for risk assessment.

[0025] The risk assessment module generates a risk level for malignant transformation of oral leukoplakia based on the predictive blood risk characterization data and the abnormal trajectory patterns. This module, as the system's decision output unit, integrates predictive data from the cross-modal association modeling module and pattern recognition results from the trajectory coupling analysis module. Through preset fusion strategies, such as weighted calculation or rule-based determination, multi-dimensional risk factors are transformed into intuitive risk levels (e.g., high risk, medium risk, low risk), thereby assisting clinical decision-making.

[0026] Through the collaborative work of the three modules mentioned above, the system in this embodiment establishes a complete risk assessment closed loop. First, the cross-modal correlation modeling module addresses data asynchrony issues and generates predictive representations. Then, the trajectory coupling analysis module uncovers dynamic anomaly patterns in the time series. Finally, the risk assessment module outputs a comprehensive judgment result. This architecture not only automates the assessment of the risk of malignant transformation of oral leukoplakia, but more importantly, by introducing a predictive blood risk representation data generation mechanism, it effectively bridges the time gap in clinical data collection, improving the real-time nature and accuracy of early warnings. Example 2:

[0027] This embodiment, based on Embodiment 1, provides a detailed explanation of the specific processing logic of the cross-modal association modeling module in the case of data asynchrony. In actual clinical scenarios, patients often have images of oral leukoplakia taken more frequently than blood tumor markers are detected, resulting in a time difference between the last image sampling point and the last blood sampling point. If this time difference is ignored and the lagging blood data is used for risk assessment, information on the progression of lesions between the blood test and the current image acquisition may be missed.

[0028] Therefore, this embodiment provides a method for generating predictive blood risk characterization data. For example... Figure 2 As shown, the cross-modal association modeling module first performs a time synchronization determination step. Specifically, the module compares the time difference between the last sampling point of the blood tumor marker time series and the last sampling point of the vitiligo image features. This time difference can be calculated by directly calculating the absolute duration between the two timestamps, such as days or hours. The module has a preset time synchronization threshold window, which can be set according to clinical follow-up guidelines or data update frequency, for example, 7 days or 14 days. When the calculated time difference does not exceed the preset time synchronization threshold window, the system determines that the data is basically synchronized and can directly call the most recent blood test data for subsequent analysis; when the time difference exceeds the preset time synchronization threshold window, the system determines that the data is not synchronized, and at this time, the predictive blood risk characterization data generation process is triggered.

[0029] In cases of data asynchrony, specifically when the last sampling point of the blood tumor marker time series is earlier than the last sampling point of the vitiligo image features, the cross-modal association modeling module generates predictive blood risk characterization data based on the current vitiligo image features and the changing trends of historical image features. This predictive data is not a simple replication of historical blood data, but rather an inferred value based on the intrinsic correlation between image features and blood indicators, used to characterize the potential risk status of the blood system at the current moment (i.e., the last sampling point of the image).

[0030] Furthermore, this embodiment provides a specific implementation method for generating predictive blood risk characterization data. The module generates predictive blood risk characterization data based on vitiligo internal heterogeneity scores, edge anomaly scores, and historical image feature change trends, through one or more methods including time extrapolation, risk mapping, or weight estimation.

[0031] The first method is time extrapolation. This method infers values ​​based on the rate of change in historical blood data and the increment of current image features. Specifically, the system first extracts the rate of change of historical blood tumor markers (such as SCC antigen or CEA), which reflects the inherent trend of changes in the patient's blood indicators over time. Simultaneously, the system calculates the increment of vitiligo image features (such as heterogeneity score) from the last blood sampling point to the current image sampling point. If the increment of image features indicates a worsening trend in the lesion (e.g., an increased heterogeneity score), the system uses a linear or nonlinear regression model to correlate the historical rate of change in blood data with the increment of image features, extrapolating the possible values ​​of the blood indicators at the current moment. For example, if historical data shows that SCC antigen increases by an average of 0.5 ng / ml per month, while current image features show an accelerated rate of lesion deterioration, the system can extrapolate, based on a preset correlation coefficient, that the current SCC antigen may have increased to a higher level. This method is suitable for scenarios where changes in blood indicators are relatively regular and strongly correlated with image features.

[0032] The second approach is the risk mapping method. This method establishes a mapping matrix between image feature increments and blood indicator increments. Specifically, the system pre-constructs a mapping knowledge base based on a large amount of historical case data, recording the changes in blood indicators (e.g., a 10% increase in CYFRA21-1) corresponding to different image feature change patterns (e.g., a 20% increase in edge blurring accompanied by a 15% increase in color heterogeneity). When a current image feature increment is detected, the system retrieves the best-matching mapping relationship from the mapping knowledge base, directly outputs the corresponding predicted increment of the blood indicator, and superimposes it onto the previous blood test value, thereby generating predictive blood risk characterization data. This method eliminates the need for complex regression calculations, has high computational efficiency, and is suitable for scenarios where there is a clear correspondence between image features and blood indicators.

[0033] The third approach is the weighted estimation method. This method comprehensively considers the weighted contributions of multiple image features. Specifically, the system assigns dynamic weights to the intralesional heterogeneity score, edge abnormality score, and historical trend features. The weight allocation can be based on machine learning model training or on clinical expert experience. For example, in certain case types with a higher risk of malignancy, the weight of the edge abnormality score may be set higher. The system calculates a comprehensive risk index by weighted summation or weighted fusion of each feature value and its corresponding weight, and then converts this index into a specific blood risk representation value. This method is highly flexible and can adapt to differences in individual patients and lesion types.

[0034] The three methods described above can be used individually or in combination. For example, the system can first obtain a preliminary predicted value through time extrapolation, and then verify and correct it through risk mapping to finally generate more accurate predictive blood risk characterization data. It should be understood that the specific algorithms listed in this embodiment are only illustrative examples. In practical applications, other algorithmic models that can establish the correlation between image features and blood indicators can also be used, such as neural network models, support vector machine models, etc. As long as they can achieve the function of inferring blood risk status based on image features, they should be covered within the protection scope of this invention.

[0035] Through the above technical solution, this embodiment effectively solves the information gap problem caused by asynchronous clinical data collection. The generated predictive blood risk characterization data can fill the risk information gaps during blood testing gaps, enabling the risk assessment module to make comprehensive judgments based on complete information at the "current moment," thereby improving the real-time performance and accuracy of early warnings. Example 3:

[0036] This embodiment, based on the above embodiments, provides a detailed explanation of the specific judgment logic for the trajectory coupling analysis module to identify abnormal trajectory patterns. For example... Figure 3 As shown, the trajectory coupling analysis module essentially constructs a dual-time-axis analysis framework when building cross-modal time-series trajectories. One time axis records the evolution trajectory of vitiligo image features, and the other time axis records the change trajectory of blood tumor markers. By comparing the time nodes of key events on the two trajectories, the module identifies abnormal trajectory patterns, which include at least one of the following: an image-priority pattern where high-risk features of vitiligo images precede abnormal changes in blood tumor markers; an indicator-priority pattern where abnormal changes in blood tumor markers precede high-risk features of vitiligo images; and a parallel pattern where high-risk features of vitiligo images and abnormal changes in blood tumor markers deteriorate synchronously.

[0037] Specifically, in image-first mode, the judgment logic focuses on capturing the rapid progression of local lesions. The system has a preset first threshold, such as 14 days. When the trajectory coupling analysis module detects that the time point of appearance of high-risk features in the vitiligo image is earlier than the time point of abnormal changes in blood tumor markers, and the time difference between these two time points exceeds the first threshold, the system determines it to be image-first mode. Clinically, this mode often indicates strong local invasiveness of the lesion, with the rate of malignant transformation of local tissue exceeding the response rate of the systemic system. In this mode, the system assigns a higher risk weight to image features during subsequent risk assessment, prompting physicians to prioritize the necessity of biopsy of the local lesion rather than solely relying on blood indicators.

[0038] In the indicator-first mode, the judgment logic reflects the systemic system's early warning of lesions. The system has a preset second threshold, such as 21 days. When the time point at which abnormal changes in blood tumor markers are detected is earlier than the time point at which high-risk features of vitiligo images appear, and the time difference exceeds this second threshold, the system classifies it as indicator-first mode. This situation may mean that although the lesion has not yet shown typical high-risk features in local morphology, a significant tumor-related immune or metabolic response has already occurred within the body. For this type of mode, the system will increase the weight of predictive blood risk characterization data and prompt physicians to shorten the follow-up period or conduct more in-depth hematological re-examinations even if the current image features are not yet typical, to prevent missed diagnosis of early occult lesions.

[0039] For the parallel deterioration mode, the judgment logic reflects the full-blown progression of the lesion. When the time point of the appearance of high-risk features in the vitiligo image and the time point of abnormal changes in blood tumor markers largely coincide, and the time difference between the two is less than a preset synchronous judgment threshold (e.g., 7 days), the system judges it as a parallel deterioration mode. This usually indicates that the lesion is in a rapid progression phase, with both local lesions and systemic systems significantly affected. In parallel mode, the risk weights of both image features and blood indicators are at a high level, and the risk level generated by the system is usually high. The system will directly recommend that a histopathological examination be performed as soon as possible for diagnosis.

[0040] By further subdividing and identifying the three patterns described above, this embodiment can more accurately depict the dynamic evolution path of malignant transformation of oral leukoplakia. It should be understood that the specific values ​​of the first threshold, the second threshold, and the simultaneous judgment threshold can be adaptively adjusted based on clinical big data statistics or the individual baseline characteristics of different patients; this embodiment does not impose any restrictions on this. This time-series logic-based pattern recognition overcomes the limitations of traditional static evaluation methods that only focus on data at a single point in time, providing a more dynamic and valuable reference for clinical decision-making. Example 4:

[0041] This embodiment provides a method for early warning of malignant transformation of oral leukoplakia. This method shares the same concept as the system embodiment described above, but differs in that this embodiment describes the technical solution from the perspective of the method flow. Figure 4 As shown, the method includes the following steps: Step S100: Establish a correlation model between oral leukoplakia image features and blood tumor marker time series, and when the last sampling point of the blood tumor marker time series is not synchronized with the last sampling point of the leukoplakia image features, generate predictive blood risk characterization data based on the current leukoplakia image features.

[0042] Specifically, this step is the core of the entire early warning method's initialization and data preprocessing. In actual clinical scenarios, patients often have oral leukoplakia images taken during outpatient follow-ups, while blood tumor markers may have been detected weeks or even months ago. The system first acquires the time-stamped image feature sequence of oral leukoplakia and the time series of blood tumor markers, and attempts to establish a correlation model between the two. This correlation model is used to describe the intrinsic mapping relationship between changes in image features and changes in blood indicators. Subsequently, the system executes synchronization determination logic: comparing the position of the last sampling point of the blood tumor marker time series with the last sampling point of the leukoplakia image features on the time axis. If the two are synchronized, the measured blood data is directly called; if the two are not synchronized, especially when the blood sampling point is earlier than the image sampling point, the system will trigger a prediction mechanism. This prediction mechanism utilizes the principle detailed in Example 2 above, based on the latest leukoplakia image features (such as heterogeneity score, edge anomaly score) and historical evolution trends, to calculate predictive blood risk characterization data that matches the current image time point. This step effectively bridges the time gap in multimodal data acquisition, providing time-aligned input data for subsequent analysis.

[0043] Step S200: Construct a cross-modal time series trajectory based on the features of the vitiligo images collected multiple times and the time series of the blood tumor markers, and identify abnormal trajectory patterns.

[0044] After obtaining synchronized data, this step focuses on mining the dynamic characteristics of disease evolution. The system arranges the features of vitiligo images collected multiple times and blood tumor markers (including measured values ​​or predicted values ​​generated in step S100) in chronological order to construct a cross-modal time series trajectory. This essentially transforms discrete sampling points into continuous dynamic trajectory curves. Based on this, the system uses the trajectory coupling analysis logic described in Example 3 above to compare feature changes on dual time axes. Specifically, the system monitors the time points when high-risk features appear on the image feature trajectory and the time points when abnormal changes occur on the blood marker trajectory. By calculating the time difference between the two and comparing it with a preset threshold, specific abnormal trajectory patterns are identified. For example, it identifies an "image-first pattern" where image feature deterioration significantly precedes changes in blood markers, or an "indicator-first pattern" where abnormal blood markers precede changes in image features, or a "parallel pattern" where both deteriorate simultaneously. Through this step, the system can capture disease evolution patterns that cannot be discovered through static analysis at a single time point.

[0045] Step S300: Based on the predictive blood risk characterization data and the abnormal trajectory pattern, generate the risk level of malignant transformation of oral leukoplakia.

[0046] This step is the decision output stage. The system integrates the immediate risk intensity represented by the predictive blood risk characterization data generated in step S100, and the dynamic evolution risk implied by the abnormal trajectory patterns identified in step S200, to conduct a multi-dimensional risk assessment. In specific implementation, the system can adopt a weighted fusion strategy to assign corresponding weights to different types of input data. For example, for the "image-first mode," the system can appropriately increase the weight of image feature-related data; for the "indicator-first mode," the weight of blood-related data is increased. Finally, the system maps the calculated comprehensive risk score to a specific risk level of malignant transformation of oral leukoplakia, such as high risk, medium risk, or low risk, thereby providing a quantitative basis for clinical decision-making. Through the time-series processing of steps S100 to S300, this embodiment achieves a complete closed loop from multimodal data synchronization and dynamic trajectory analysis to comprehensive risk assessment, effectively improving the accuracy and timeliness of early warning. Example 5:

[0047] This embodiment, based on the above embodiments, provides a detailed explanation of the specific calculation process for leukoplakia image features. The leukoplakia image features include internal heterogeneity scoring and edge anomaly scoring. These two scoring parameters are not abstract numerical values, but rather quantitative indicators calculated based on specific image processing algorithms. They objectively reflect the microscopic morphological characteristics of oral leukoplakia lesions, providing a quantifiable data foundation for subsequent cross-modal correlation modeling.

[0048] Specifically, for calculating the heterogeneity score within vitiligo lesions, the system first divides the segmented vitiligo lesion areas into sub-regions. For example... Figure 5 As shown, the partitioning method can employ a grid-based approach, dividing the white patch region into multiple rectangular sub-regions of equal size; alternatively, a superpixel-based partitioning method can be used, dividing the white patch region into several irregular superpixel sub-regions based on the local color and texture similarity of the image. It should be understood that the granularity of sub-region partitioning can be adjusted according to image resolution and analysis requirements, for example, dividing it into 16, 25, or more sub-regions; this embodiment does not impose such limitations.

[0049] After dividing the area into sub-regions, the system calculates the color features, texture features, and surface roughness features of each sub-region. The heterogeneity score within the white patch is calculated based on the color differences, texture differences, and roughness differences among multiple sub-regions within the white patch area.

[0050] To calculate color differences, the system can calculate the average color difference between different sub-regions. Specifically, the image is converted from the RGB color space to the Lab color space, the Euclidean distance between the Lab values ​​of any two sub-regions is calculated, and the average of the Euclidean distances between all pairs of sub-regions is taken as the color difference index. The larger the average color difference, the more uneven the color distribution within the vitiligo lesion, and the higher the possibility of red and white mixing or varying shades of color. Clinically, this usually corresponds to uneven angiogenesis or keratinization of the lesion tissue.

[0051] For calculating texture differences, the system can extract gray-level co-occurrence matrix (GLCM) feature parameters of each sub-region, such as contrast, homogeneity, energy, and entropy. It calculates the difference in contrast between different sub-regions' GLCMs and statistically analyzes the distribution of these differences. Greater texture differences often indicate a higher degree of cell disorder within the diseased tissue, and its clinical significance usually corresponds to tissue structural atypia.

[0052] For calculating roughness differences, the system can extract surface roughness features of each sub-region based on Local Binary Pattern (LBP) or high-frequency wavelet coefficients. The system also calculates the dispersion of roughness features between different sub-regions. Greater roughness differences suggest uneven distribution of granular or nodular texture on the lesion surface, potentially indicating focal hyperplasia or abnormal keratinization.

[0053] Finally, the system performs weighted fusion or normalization on color differences, texture differences, and roughness differences to generate a heterogeneity score within the vitiligo lesions. A higher score indicates more complex histological characteristics within the vitiligo lesions and a higher potential risk of malignant transformation.

[0054] For calculating the edge anomaly score, the system focuses on the morphological characteristics of the vitiligo edge region. The edge anomaly score is calculated based on the morphological complexity, boundary ambiguity, and local outward expansion trend of the vitiligo edge region.

[0055] To calculate morphological complexity, the system can determine the fractal dimension of the vitiligo lesion's edge. Specifically, it uses box counting or the perimeter-area method to calculate the fractal dimension of the edge curve. A higher fractal dimension indicates a more irregular and complex edge morphology, with a higher likelihood of serrated, pseudopodia-like, or map-like changes. Clinically, this morphological characteristic often suggests invasive growth of the lesion.

[0056] To calculate boundary ambiguity, the system can calculate the gradient entropy of the edge region. Specifically, it extracts the gray-level gradient information of the edge region and calculates the entropy value of the gradient distribution. The higher the gradient entropy, the more chaotic the gray-level transition at the boundary, and the more blurred the boundary. Blurred boundaries often mean that the boundary between the lesion and the surrounding normal tissue is unclear, and there may be microscopic infiltration.

[0057] To calculate the local expansion trend, the system can calculate the radial derivative of the edge. Specifically, taking the centroid of the white spot region as the origin, radial rays are drawn to each point on the edge, and the change of the derivative at the edge point along the radial direction is calculated. If the radial derivative of a certain edge region is consistently positive and has a large value, it indicates that the region has a tendency to expand outward. The system can statistically analyze the proportion or intensity of edge segments with significant expansion trends as a quantitative indicator of the local expansion trend.

[0058] Finally, the system performs weighted fusion or normalization of morphological complexity, boundary ambiguity, and local expansion trend to generate an edge anomaly score. The higher the score, the more obvious the abnormal morphological features of the vitiligo edge, and the higher the risk of the lesion invading or spreading outward.

[0059] Through the specific image processing algorithms described above, this embodiment transforms the subjective judgment of clinicians through visual observation into quantifiable objective parameters. The internal heterogeneity score and edge abnormality score of vitiligo not only accurately capture the microscopic morphological features of the lesion but also serve as input data for the cross-modal correlation modeling module, enabling correlation analysis with time series of blood tumor markers, thereby improving the objectivity and accuracy of the early warning system. Example 6:

[0060] This embodiment, based on the above embodiments, provides a detailed explanation of the system's feedback learning mechanism. In actual clinical applications, the risk level of malignant transformation of oral leukoplakia generated by the system is inferred solely based on current image features and blood data, and its accuracy depends on the initial parameter settings of the model. To enable the system to continuously optimize with the accumulation of clinical data, this embodiment introduces a feedback learning mechanism based on pathological histological results.

[0061] Specifically, the system also includes a feedback learning step. First, the system obtains the pathological biopsy results or postoperative histological results corresponding to the risk level of malignant transformation of oral leukoplakia. This result serves as the "gold standard" for clinical diagnosis and is a true label for verifying the accuracy of the system's predictions. Pathological biopsy results or postoperative histological results are usually obtained through the hospital's pathology information system interface or manually entered into the system by the doctor. The specific form of the results can include pathological grading such as mild dysplasia, moderate dysplasia, severe dysplasia, carcinoma in situ, or invasive carcinoma.

[0062] Secondly, the system records the error between the risk level of malignant transformation of oral leukoplakia and the pathological biopsy results or postoperative histological results. Here, "error" is a broad concept, specifically including risk level misjudgment and feature weight bias. Risk level misjudgment refers to the situation where the risk level generated by the system does not match the pathological results. For example, if the system determines it as "low risk," but the pathological result shows "severe dysplasia" or "carcinoma in situ," it is judged as a missed diagnosis error; if the system determines it as "high risk," but the pathological result shows "benign lesion" or "mild dysplasia," it is judged as a false alarm error. Feature weight bias refers to the situation where, in some error samples, the contribution of a specific feature (such as a marginal anomaly score) to the final risk level does not match its correlation with the actual pathology, leading to model judgment bias. The system records the feature vector, prediction result, actual pathological result, and error type for each error sample in detail, forming an error sample library.

[0063] Finally, when the number of error samples reaches a preset number, the system updates the weights of vitiligo image features, blood tumor markers, trajectory coupling parameters, and time synchronization threshold window based on the error. The preset number can be set according to the rate of clinical data accumulation and model update needs, for example, 50 cases. When the number of samples in the error sample library reaches 50, the system automatically triggers the parameter update process. The specific logic of parameter update can use the backpropagation algorithm or statistical learning methods. Taking the backpropagation algorithm as an example, the system takes the pathological result as the target output and the current image feature weights, blood marker weights, and other parameters as adjustable variables. By calculating the loss function between the predicted output and the true result, the system adjusts the weight values ​​of each parameter along the gradient descent direction. For example, if the error analysis shows that the "marginal anomaly score" is generally low in missed cases, the system will correspondingly increase the weight coefficient of this feature in the risk assessment model; if it is found that the "time synchronization threshold window" is set too wide, causing some lagging blood data to be incorrectly regarded as synchronous data, the system will automatically reduce the value of the window. For trajectory coupling parameters, if the judgment threshold of "image-first mode" is found to be too lenient, leading to an increase in false alarms, the system will adjust the size of the first threshold to improve the judgment accuracy.

[0064] Through the aforementioned feedback learning mechanism, this embodiment achieves a closed-loop optimization system. The system is no longer a static evaluation tool, but a dynamic system capable of continuous self-evolution based on clinical feedback. This mechanism effectively solves the problem of lacking reverse correction based on pathological results in existing technologies. As clinical application time increases, the error sample library is continuously enriched, and the model parameters continuously approach the optimal solution, thereby continuously improving the accuracy and clinical reliability of the system in predicting the risk of malignant transformation of oral leukoplakia. Example 7:

[0065] To more intuitively demonstrate the application effect of the technical solution of the present invention in a real clinical environment, this embodiment combines specific clinical cases to provide a detailed explanation of the entire operation process of the system.

[0066] A 52-year-old male patient presented to the outpatient clinic with recurrent leukoplakia on his right buccal mucosa for approximately three months. During the initial visit, the doctor used an intraoral camera to capture images of the patient's oral leukoplakia. The system received the image through the image data acquisition module and performed standardized preprocessing such as color correction and brightness normalization. Subsequently, the leukoplakia feature extraction module automatically segmented the lesion area and extracted multidimensional features. Based on the segmentation results, the heterogeneity analysis module calculated the internal heterogeneity score and edge anomaly score of the leukoplakia. Specifically, the system divided the leukoplakia area into multiple sub-regions and calculated the color and texture differences between each sub-region. It was found that the patient's leukoplakia contained obvious red-white mixed areas with a high color difference index, resulting in a high level of internal heterogeneity score. At the same time, the system analyzed the edge region, calculating morphological complexity and boundary ambiguity. The results showed that the edges exhibited local jagged changes, and the edge anomaly score was also higher than the preset median level.

[0067] Seven days later, the patient returned for a follow-up visit as instructed by the doctor, and the system collected images of the vitiligo again. By comparing the features of the two images, the system found that although the total area of ​​the vitiligo changed only slightly, the range of local high-risk heterogeneous areas had expanded, and the feathering-like changes at the edges had worsened, indicating an upward trend in local morphological risk. At this point, the system attempted to obtain the patient's blood tumor marker data. A search revealed that the patient's most recent blood test (including SCC antigen, CEA, etc.) occurred 15 days prior. The cross-modal association modeling module compared the time difference between the last sampling point of the blood tumor marker time series and the last sampling point of the vitiligo image features, determining that this time difference exceeded a preset time synchronization threshold window (e.g., set to 7 or 14 days), thus confirming data asynchrony.

[0068] To address this asynchrony, the cross-modal association modeling module triggers a prediction mechanism. Since the blood sampling point precedes the image sampling point, the module, based on the current vitiligo internal heterogeneity score, edge anomaly score, and the historical image feature change trends over the past 7 days, uses a combination of time extrapolation and risk mapping to generate predictive blood risk characterization data synchronized with the current image time point. Specifically, the system uses the historical blood data change rate, combined with the lesion deterioration speed shown by the current image features, to extrapolate the current predicted blood index values; simultaneously, it retrieves blood index increments that match the current image feature increments through a mapping knowledge base, verifies and corrects the predicted values, and ultimately generates data representing the current blood risk status.

[0069] Subsequently, the trajectory coupling analysis module intervenes. Based on the features of multiple acquired vitiligo images and the time series of blood tumor markers (including measured and predicted values), the module constructs a cross-modal time series trajectory. During trajectory analysis, the module identifies time points where high-risk features in the vitiligo images (such as elevated heterogeneity scores) appear significantly earlier than time points of abnormal changes in blood tumor markers, and the time difference between the two exceeds a preset first threshold (e.g., 14 days). Based on this, the system determines that the patient's disease progression conforms to an image-first pattern. This pattern clinically typically indicates strong local invasiveness of the lesion, with the rate of malignant transformation of local tissues exceeding the systemic response rate.

[0070] Based on the above analysis, the risk assessment module integrates predictive blood risk characterization data and the trajectory anomaly confidence level of the image-based pattern to generate a risk level for malignant transformation of oral leukoplakia. Due to the high risk of the image features and the trajectory pattern indicating local invasiveness, the system ultimately generates a high-risk determination. Based on this, the system outputs a warning message to the physician, recommending priority biopsy of the rough, thickened area behind the leukoplakia, and repeat testing of SCC antigen and related blood indicators within one week, while shortening the follow-up interval.

[0071] The doctor performed a biopsy on the patient based on the system's recommendations. Subsequent pathological examination revealed moderate epithelial dysplasia. This pathological result was entered into the system as the "true value." The feedback learning module recorded the correspondence between the system's previously generated "high-risk" judgment and the pathological result of "moderate epithelial dysplasia," and calculated the error. Although the risk level determined by the system generally trended in line with the severity of the pathological result, there were slight deviations in the specific risk score calculation. As clinical cases accumulate, when the number of such error samples reaches a preset number (e.g., 50 cases), the system will use the backpropagation algorithm to adjust the weights of vitiligo image features, blood biomarkers, and trajectory coupling parameters based on this error data, thereby achieving continuous model optimization.

[0072] As demonstrated in this embodiment, the system provided by this invention can effectively identify high-risk lesions by capturing local heterogeneity features and abnormal patterns in cross-modal time series trajectories, even when the area of ​​vitiligo does not change significantly. Simultaneously, by generating predictive hematological risk characterization data, it solves the assessment blind spot problem caused by asynchronous clinical data collection, providing a dynamically valuable reference for clinical decision-making.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A system for early warning of malignant transformation of oral leukoplakia, characterized in that, include: The cross-modal association modeling module is used to establish an association model between oral leukoplakia image features and blood tumor marker time series, and to generate predictive blood risk characterization data based on the current leukoplakia image features when the last sampling point of the blood tumor marker time series is not synchronized with the last sampling point of the leukoplakia image features. The trajectory coupling analysis module is used to construct cross-modal time series trajectories based on the features of the vitiligo images collected multiple times and the time series of blood tumor markers, and to identify abnormal trajectory patterns; The risk assessment module is used to generate a risk level for malignant transformation of oral leukoplakia based on the predictive blood risk characterization data and the abnormal trajectory pattern.

2. The oral leukoplakia malignant transformation early warning system according to claim 1, characterized in that, The cross-modal association modeling module is used for: When the last sampling point of the blood tumor marker time series is earlier than the last sampling point of the vitiligo image features, the predictive blood risk characterization data is generated based on the current vitiligo image features and the historical image feature change trends.

3. The oral leukoplakia malignant transformation early warning system according to claim 2, characterized in that, The cross-modal association modeling module is used for: Based on the internal heterogeneity score of vitiligo, the edge abnormality score, and the trend of historical image feature changes, the predictive blood risk characterization data is generated through one or more of the following methods: time extrapolation, risk mapping, or weight estimation.

4. The oral leukoplakia malignant transformation early warning system according to claim 1, characterized in that, The abnormal trajectory pattern includes at least one of the following: High-risk features of vitiligo images precede abnormal changes in blood tumor markers in an image-prior pattern. Abnormal changes in blood tumor markers precede the indicator-leading pattern of high-risk features in the vitiligo images; The high-risk features of vitiligo images and the abnormal changes in the blood tumor markers showed a parallel pattern of deterioration.

5. The oral leukoplakia malignant transformation early warning system according to claim 1, characterized in that, The image features of the vitiligo include internal heterogeneity scores and edge anomaly scores. The internal heterogeneity score of the white spot is calculated based on the color difference, texture difference and roughness difference between multiple sub-regions within the white spot area; The edge anomaly score is calculated based on the morphological complexity of the white spot edge region, boundary ambiguity, and local outward expansion trend.

6. A method for early warning of malignant transformation of oral leukoplakia, characterized in that, Includes the following steps: S100, Establish a correlation model between oral leukoplakia image features and blood tumor marker time series, and when the last sampling point of the blood tumor marker time series is not synchronized with the last sampling point of the leukoplakia image features, generate predictive blood risk characterization data based on the current leukoplakia image features; S200, construct a cross-modal time series trajectory based on the features of the vitiligo images collected multiple times and the time series of the blood tumor markers, and identify abnormal trajectory patterns; S300, Based on the predictive blood risk characterization data and the abnormal trajectory pattern, generate the risk level of malignant transformation of oral leukoplakia.

7. The method for early warning of malignant transformation of oral leukoplakia according to claim 6, characterized in that, Step S100 includes: Compare the time difference between the last sampling point of the blood tumor marker time series and the last sampling point of the vitiligo image features; When the time difference exceeds a preset time synchronization threshold window, the predictive blood risk characterization data is generated based on the current vitiligo image features.

8. The method for early warning of malignant transformation of oral leukoplakia according to claim 7, characterized in that, Step S100 generates the predictive blood risk characterization data, including: Based on the internal heterogeneity score of vitiligo, the edge abnormality score, and the trend of historical image feature changes, the predictive blood risk characterization data is generated through one or more of the following methods: time extrapolation, risk mapping, or weight estimation.

9. The method for early warning of malignant transformation of oral leukoplakia according to claim 6, characterized in that, The abnormal trajectory pattern identified in step S200 includes at least one of the following: Image-leading patterns where high-risk features of vitiligo precede abnormal changes in blood tumor markers; Abnormal changes in blood tumor markers precede the indicator-leading pattern of high-risk features in the vitiligo images; The high-risk features of the vitiligo image and the abnormal changes in the blood tumor markers deteriorate in a parallel pattern.

10. The method for early warning of malignant transformation of oral leukoplakia according to claim 7, characterized in that, Also includes: Obtain the pathological biopsy results or postoperative histological results corresponding to the risk level of malignant transformation of oral leukoplakia; Record the error between the risk level of malignant transformation of the oral leukoplakia and the pathological biopsy results or postoperative histological results; When the number of error samples reaches a preset number, the white spot image feature weights, blood tumor marker weights, trajectory coupling parameters, and time synchronization threshold window are updated according to the error.

Images