Artificial intelligence-based endoscopic stomach early cancer risk grading system

Abstract

The invention relates to the technical field of endoscopic diagnosis, and discloses an endoscopic stomach early cancer risk grading system based on artificial intelligence. The system comprises an image acquisition module, a first feature extraction module, a second feature extraction module, a dynamic classification module and a time sequence optimization module. The image acquisition module acquires a gastroscope video stream in real time and extracts a single-frame image; a first feature extraction module identifies mucous membrane surface microvessel shape distribution and gland opening structure abnormal features in the single frame image; a second feature extraction module quantifies color saturation offset and surface texture roughness features of the lesion area; the dynamic grading module generates an initial canceration probability value according to a capillary and gland tube opening feature fusion result, and corrects the value in combination with a color and texture feature dynamic association relationship; and the time sequence optimization module analyzes the change rate of the initial canceration probability value of the continuous frames, and shortens the frame sampling interval of the image acquisition module when the change rate exceeds a preset fluctuation threshold. The system can optimize early gastric cancer risk grading and improve diagnosis accuracy and objectivity.

Description

Technical Field

[0001] This invention relates to the field of endoscopic diagnostic technology, specifically to an endoscopic early gastric cancer risk grading system based on artificial intelligence. Background Technology

[0002] Early gastric cancer, as an early stage of gastric cancer development, directly impacts the patient's treatment options and prognosis through accurate diagnosis. Currently, gastroscopy is the primary method for detecting early gastric cancer in clinical practice, relying on endoscopists to visually observe the morphology, color, and texture of the gastric mucosa to determine the nature of the lesion. However, this traditional diagnostic method has significant limitations. The abnormalities in the distribution of microvessels and the structure of glandular openings in the superficial layer of the gastric mucosa are subtle, and different physicians have varying abilities to recognize these features. Especially for inexperienced physicians, missed or misdiagnosed cases are common. The color saturation variations and surface texture roughness of lesion areas typically exhibit non-standardized characteristics, making quantitative analysis difficult based solely on subjective human judgment. This leads to a lack of unified standards for assessing lesion severity, consequently affecting the consistency of subsequent treatment decisions. Furthermore, during gastroscopy, the video stream data is dynamically changing. Traditional examination methods use fixed frame sampling intervals, which cannot flexibly adjust the sampling frequency according to changes in lesion probability. When the probability of cancerous changes in a lesion area fluctuates rapidly, a fixed sampling interval may fail to capture crucial image information in a timely manner, further increasing diagnostic uncertainty. Current technologies lack a mechanism for effectively fusing multi-dimensional image features and dynamically correcting the probability of cancer. During diagnosis, physicians often need to analyze features such as microvessels, glandular openings, color, and texture separately before making a comprehensive judgment. This approach is not only inefficient but also prone to diagnostic bias due to insufficient analysis of the correlation between features. With the increasing clinical demand for early diagnosis of gastric cancer, these technological deficiencies have become significant factors restricting diagnostic accuracy and efficiency, necessitating an intelligent diagnostic system capable of multi-feature quantitative analysis, dynamic probability correction, and adaptive frame sampling. Summary of the Invention

[0003] The purpose of this invention is to provide an endoscopic early gastric cancer risk grading system based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides an endoscopic early gastric cancer risk stratification system based on artificial intelligence, the system comprising: The image acquisition module is used to acquire the video stream of the gastroscopy examination in real time and extract single-frame images; The first feature extraction module is used to identify the morphological distribution features of microvessels on the mucosal surface and the abnormal features of glandular opening structures from the single frame image; The second feature extraction module is used to quantify the color saturation shift features and surface texture roughness features of the lesion area from the single frame image; The dynamic grading module is used to generate an initial cancer probability value based on the fusion result of the microvascular morphological distribution characteristics and the abnormal features of the glandular opening structure, and to correct the initial cancer probability value based on the dynamic correlation between the color saturation shift feature and the surface texture roughness feature. The timing optimization module is used to analyze the rate of change of the initial cancer probability value in consecutive frame images. When the rate of change exceeds a preset fluctuation threshold, the frame sampling interval of the image acquisition module is shortened.

[0005] Preferably, the dynamic hierarchical module includes: A probability generation unit is used to input the microvascular morphological distribution features and the abnormal features of the glandular opening structure into a convolutional kernel size adaptive network and output an initial cancer probability value. The probability correction unit is used to calculate the covariance matrix between the color saturation shift feature and the surface texture roughness feature, and adjust the confidence weight of the initial cancer probability value according to the distribution range of the eigenvalues ​​of the covariance matrix. A grading decision unit is used to multiply the confidence weight by the initial cancer probability value to generate a final grading index.

[0006] Preferably, the timing optimization module includes: The rate of change monitoring unit is used to calculate the standard deviation of the final grading index corresponding to three adjacent frames of images; The sampling control unit is used to shorten the frame sampling interval to a preset ratio of the original value when the standard deviation is greater than the first threshold and less than the second threshold; and to activate the full frame rate sampling mode when the standard deviation is greater than the second threshold.

[0007] Preferably, the sampling control unit performs the following synchronously when shortening the frame sampling interval: The resource monitoring unit is used to detect the real-time memory usage of the image processing hardware platform. The attenuation adjustment unit is used to dynamically attenuate the preset ratio by using a reduction factor when the memory occupancy rate exceeds the preset load threshold.

[0008] Preferably, the dynamic hierarchical module further includes: The regional division unit is used to divide the lesion area into a core lesion area and a peripheral transition area according to the spatial distribution gradient of the final grading index. The weight redistribution unit is used to increase the weight of the final grading index of the core lesion area in the overall assessment.

[0009] Preferably, the weight redistribution unit performs: Calculate the difference between the mean final grading index of the core lesion area and the peripheral transition area; The weight increase of the core lesion area is determined based on the magnitude of the mean difference. When the difference between the mean and the mean exceeds the preset difference threshold, the weight increase of the core lesion area is set to the preset upper limit.

[0010] Preferably, the system further includes: A multi-source input module is used to receive time-series data of serum tumor marker concentrations from the patient's electronic medical record; The cross-modal fusion module is used to correlate the fluctuation range of the serum tumor marker concentration time series data with the rate of change of the final grading index. When the correlation coefficient exceeds the preset correlation threshold, a cross-modal verification flag is generated.

[0011] Preferably, when performing correlation matching, the cross-modal fusion module: The maximum fluctuation range of serum tumor marker concentrations within the image acquisition time window was extracted. Calculate the Pearson correlation coefficient between the maximum volatility and the rate of change of the final graded index; When the Pearson correlation coefficient is greater than the first correlation threshold and less than the second correlation threshold, the initial state of the cross-modal verification flag is activated; When the Pearson correlation coefficient is greater than the second association threshold, the advanced state of the cross-modal verification flag is activated.

[0012] Preferably, the dynamic grading module is in the advanced state when the cross-modal verification flag is in the advanced state: The fluctuation range of serum tumor marker concentrations was normalized to a correction factor; The correction coefficients are weighted and superimposed with the final grading index to generate the enhanced grading result.

[0013] Preferably, the timing optimization module further includes: The phase calibration unit is used to detect the time deviation of the final graded index generation between the current frame and the previous frame after shortening the frame sampling interval. The timing compensation unit is used to adjust the start time point of feature extraction for the next frame image based on the time deviation value, so that the feature extraction operation is synchronized with the frame output time point of the image acquisition module.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By acquiring the video stream of the gastroscopy examination in real time and extracting single-frame images through the image acquisition module, a stable and continuous source of image data is provided for subsequent feature analysis. This avoids the randomness and incompleteness that may occur when acquiring images manually in the traditional way, and ensures the reliability of the data on which the diagnostic analysis depends. The first feature extraction module can identify the morphological distribution characteristics of microvessels on the mucosal surface and abnormal features of glandular opening structures from a single frame image. With the precise recognition capabilities of artificial intelligence technology, it can effectively capture subtle feature changes that are difficult to detect with the naked eye, reduce missed diagnoses or misdiagnoses caused by differences in physicians' subjective judgment, and make the analysis of the basic structural features of lesions more objective and comprehensive. The second feature extraction module performs quantitative analysis on the color saturation shift features and surface texture roughness features of the lesion area. This changes the traditional diagnostic approach that relies on subjective manual evaluation of these non-standardized features. Through quantitative processing, it provides comparable and referable objective data for lesion assessment, enabling the feature differences of different lesion areas to be clearly presented, which helps to more accurately determine the nature of the lesion.

[0015] The dynamic grading module uses the fusion results of microvascular morphological distribution features and glandular opening structural abnormalities to generate an initial cancer probability value. Simultaneously, it refines the initial cancer probability value by incorporating the dynamic correlation between color saturation shift features and surface texture roughness features, achieving deep fusion and collaborative analysis of multi-dimensional image features. This dynamic correction mechanism fully utilizes the inherent correlations between various features, making the calculation of cancer probability more closely reflect the actual lesion situation, avoiding the limitations of single-feature analysis, and improving the accuracy of probability assessment. The timing optimization module analyzes the rate of change of the initial cancer probability value in consecutive frames of images. When the rate of change exceeds a preset fluctuation threshold, it shortens the frame sampling interval of the image acquisition module, achieving adaptive adjustment of the sampling frequency. When the lesion probability changes rapidly, denser frame sampling can capture key image information in a timely manner, ensuring that important details that may reflect changes in the nature of the lesion are not missed. When the lesion probability is relatively stable, maintaining a normal sampling interval can avoid unnecessary data redundancy, ensuring both diagnostic accuracy and data processing efficiency. Attached Figure Description

[0016] Figure 1 This is a timeline diagram of the artificial intelligence-based endoscopic early gastric cancer risk grading system described in this invention. Figure 2 A flowchart illustrating the operation of the dynamic hierarchical module; Figure 3 A flowchart for resource-aware sampling adjustments; Figure 4 Flowchart for weight redistribution execution; Figure 5 This is a flowchart for multi-source data fusion verification. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides an endoscopic early gastric cancer risk grading system based on artificial intelligence, the system comprising: By acquiring real-time video streams from gastroscopy examinations, extracting single-frame images, and performing multi-dimensional feature analysis, combined with dynamic probability calculation and temporal optimization mechanisms, accurate cancer risk grading is achieved. The image acquisition module uses high-definition endoscopic equipment to capture video streams at a rate of 30 frames per second and extracts single-frame images with a resolution of 1920×1080 pixels through a frame buffer. The first feature extraction module applies a pre-trained deep learning model to identify the morphological distribution characteristics of microvessels and abnormal features of glandular opening structures in the mucosal surface from single-frame images. Microvessel morphological distribution characteristics include vessel density, branching angle, and distribution uniformity; abnormal features of glandular opening structures include opening shape variability and arrangement disorder index. The second feature extraction module, based on a color space conversion algorithm, converts the image from RGB mode to HSV mode, quantifies the color saturation shift characteristics of the lesion area, and calculates its saturation difference with normal mucosa; simultaneously, it uses the gray-level co-occurrence matrix to calculate surface texture roughness features and extracts contrast and entropy values. The dynamic grading module fuses the outputs of the first and second feature extraction modules. First, it generates an initial cancer probability value based on the microvascular morphology distribution characteristics and abnormal glandular opening structure characteristics. Then, it corrects this probability value based on the dynamic correlation between color saturation shift characteristics and surface texture roughness characteristics. The temporal optimization module monitors the rate of change of the initial cancer probability value in consecutive frames of images. When the rate of change exceeds a preset fluctuation threshold, it optimizes the system response speed by adjusting the frame sampling interval.

[0019] Example 1: See Figure 2The image acquisition module obtains a continuous video stream from the endoscopy device and deframes it into a series of static images. These images are fed into the first feature extraction module, which uses a deep convolutional neural network to identify the microstructures of the mucosal surface. The network first preprocesses the input images, including brightness normalization and contrast enhancement, to optimize feature visibility. Subsequently, the core branch of the network focuses on vascular morphology analysis, extracting the topological features of blood vessels through a series of separable convolutional layers, and outputting a quantified microvascular morphology distribution feature vector. This vector contains information on multiple dimensions such as vascular density, branch complexity, and distribution uniformity. Simultaneously, another parallel branch of the network processes glandular opening structures, using an attention mechanism to focus on the tiny indentations on the mucosal surface, calculating their shape regularity, arrangement order, and size variability, ultimately generating an abnormal feature vector of glandular opening structures. These two feature vectors together constitute a digital description of the mucosal surface state.

[0020] The second feature extraction module employs a different image processing method. It first converts the RGB image to the HSV color space, directly calculates the saturation channel values ​​for the lesion area, and compares them with the saturation values ​​of standard normal mucosa in the database to obtain a quantified color saturation shift feature. This feature reflects the changes in biochemical components that may be caused by cancer in the tissue. Furthermore, this module evaluates texture by analyzing the spatial relationships of image grayscale values, using the gray-level co-occurrence matrix to calculate contrast and entropy values, thereby generating a surface texture roughness feature. This feature characterizes the physical structural changes of the mucosal surface.

[0021] The probability generation unit receives the two feature vectors from the first feature extraction module, which are then input into a uniquely designed convolutional kernel size adaptive network. Instead of using fixed-size convolutional kernels, this network dynamically selects the most suitable kernel size based on the scale of the input feature map. For feature maps requiring a large receptive field, such as blood vessel density, the network automatically selects a larger 7x7 convolutional kernel for information aggregation; while for fine features like glandular opening shape, it may switch to a smaller 3x3 convolutional kernel to capture details. The network ultimately outputs a scalar value between 0 and 1 through fully connected layers and a sigmoid activation function, representing the initial cancer probability. This value initially reflects the cancer risk based on morphological features.

[0022] The probability correction unit provides a confidence assessment for the initial probability values. This unit receives two features: color saturation shift and surface texture roughness. It first calculates the covariance matrix of these two features to analyze the strength of the statistical correlation between color and texture information. The eigenvalues ​​of the covariance matrix are calculated, and their distribution range is used for evaluation. If the two features show high consistency—for example, a significant color saturation shift accompanied by abnormally rough texture—the eigenvalues ​​are concentrated, indicating strong evidence, and the confidence weight increases. Conversely, if the evidence provided by the color and texture features contradicts each other, the eigenvalues ​​are dispersed, and the confidence weight decreases. This dynamically calculated confidence weight is a multiplier factor between 0.5 and 1.5.

[0023] The grading decision unit performs the final calculation, multiplying the initial cancer probability value output by the probability generation unit by the confidence weight output by the probability correction unit to obtain the final grading index. This design means that even if the initial probability value is high, the final index will be lowered if the supporting color and texture evidence is inconsistent; conversely, if multiple pieces of evidence are highly consistent, the final index may be raised even if the initial probability value is moderate. This final grading index provides clinicians with a comprehensive and reliability-calibrated quantitative indicator for risk assessment.

[0024] Example 2: See Figure 3 This involves a mechanism that dynamically adjusts the processing pace during continuous image analysis. This mechanism optimizes resource allocation and response speed by monitoring the stability of risk assessment results in real time. When the endoscope moves within the patient's digestive tract, the illumination conditions on the mucosal surface, the lens distance, and tissue morphology may change rapidly, causing fluctuations in the risk assessment results of consecutive frames. The core component of the temporal optimization module is the rate of change monitoring unit. This unit continuously tracks the final grading index output by the dynamic grading module. For each newly processed image frame, this unit caches the final grading index values ​​corresponding to the three most recent frames. The standard deviation of these three values ​​is then calculated; this statistic reflects the dispersion of risk assessment results in the short term. The standard deviation is calculated based on standard statistical methods, using the sum of squared deviations of three data points from their arithmetic mean. For example, when the final grading indices for three consecutive frames are 0.45, 0.62, and 0.53, their standard deviation will be calculated as a specific value reflecting the amplitude of fluctuation.

[0025] The sampling control unit makes decisions based on the standard deviation value provided by the rate of change monitoring unit, which presets two key thresholds: a first threshold of 0.05 and a second threshold of 0.15. These thresholds divide the standard deviation range into three response intervals. When the calculated standard deviation is less than or equal to the first threshold, it indicates that the risk assessment result is relatively stable, and the system maintains the default frame sampling interval, for example, processing one frame of image every 33 milliseconds (corresponding to the original sampling rate of 30 frames per second). When the standard deviation value is greater than the first threshold but less than the second threshold, for example, a standard deviation of 0.12, the system determines that the risk status has a moderate degree of fluctuation. At this time, the sampling control unit activates the response mechanism, shortening the frame sampling interval to a preset proportion of the original value. The preset proportion can be configured to 60%, which means that the original 33-millisecond interval will be shortened to about 20 milliseconds, equivalent to increasing the sampling rate to about 50 frames per second. This adjustment allows the system to capture organizational details at a higher frequency during critical periods when the risk status may change. When the standard deviation exceeds the second threshold, for example, reaching 0.18, it indicates that the risk assessment result has fluctuated drastically. At this time, the sampling control unit activates the full frame rate sampling mode. In this mode, the system will process images at the highest frame rate supported by the image acquisition hardware, such as the original upper limit of 30 frames per second. Full frame rate mode ensures that the system does not miss critical information due to insufficient sampling when the risk state evolves rapidly.

[0026] When the sampling control unit performs frame sampling interval shortening, it simultaneously triggers a resource monitoring and adaptive adjustment mechanism. The resource monitoring unit obtains real-time memory usage data from the image processing hardware platform by calling the application programming interface provided by the operating system. This data is expressed as a percentage of current memory usage. The system presets a load threshold, such as 80%, as a resource warning line. When the memory usage is below this threshold, the system shortens the sampling interval normally according to a preset ratio (e.g., 60%). However, when the resource monitoring unit detects that the memory usage exceeds the preset load threshold, for example, reaching 85%, the attenuation adjustment unit immediately intervenes. This unit uses a reduction factor, such as 0.8, to dynamically calculate the attenuation of the preset ratio. Specifically, the original 60% shortening ratio will be adjusted to 60% × 0.8 = 48%, meaning the actual shortened sampling interval is approximately 52% of the original interval. This dynamic attenuation mechanism automatically reduces the increase in processing load when system resources are strained, preventing memory overflow or system crashes caused by excessive pursuit of high-frequency sampling. The reduction factor can be configured to different values ​​according to hardware performance, providing flexible adjustment capabilities in resource-constrained environments.

[0027] The entire timing optimization process runs continuously within clinical examination scenarios. For example, when an endoscopist steadily aims the lens at a flat lesion, the risk assessment index may remain relatively stable, with the standard deviation below the first threshold, and the system maintains a regular sampling frequency. However, as the lens moves towards the lesion's edge or encounters mucosal folds, the surface microvascular morphology may change significantly across consecutive frames, causing the standard deviation to rise to 0.1. At this point, the system automatically shortens the sampling interval to capture detailed tissue feature changes at a higher frequency. If, during this process, system memory usage rises to 82% due to increased parallel tasks, exceeding the 80% load threshold, the actual reduction in the sampling interval will be adjusted by a reduction factor to avoid resource overload. When the endoscope rapidly scans a widespread area of ​​mucosal abnormality, the risk index may jump significantly between consecutive frames, with the standard deviation exceeding 0.15, triggering full-frame-rate sampling mode to ensure the capture of rapidly changing tissue states. This dynamic, resource-aware sampling adjustment allows the system to maintain a balance between computational resource constraints and clinical information needs.

[0028] Example 3: See Figure 4 This approach involves refined zoning management of lesion areas and adjusting the contribution weight of different areas in the overall assessment based on their significant differences in risk, thereby improving the spatial resolution and clinical relevance of risk stratification. This implementation is based on the observation that early gastric cancer lesions often exhibit spatial heterogeneity, typically presenting as a core area with a high risk index surrounded by transitional areas with gradually decreasing risk indices. Simply homogenizing the entire suspicious area may dilute the signal intensity of high-risk areas.

[0029] After generating the final grading index, the dynamic grading module's region partitioning unit begins operation. This unit receives input not only of the final scalar index value, but more importantly, the original distribution information of that index in the image space—that is, the risk value calculated for each pixel or image patch. The region partitioning unit first constructs the spatial distribution gradient field of the final grading index. Gradient calculation is based on standard image processing methods, quantifying the rate of risk change at a given location by examining the difference in risk indices between each pixel and its neighboring pixels. High gradient values ​​indicate a sharp spatial change in risk level, typically corresponding to the boundary between normal mucosa and diseased tissue; while low gradient value regions indicate a relatively homogeneous risk level.

[0030] Based on the calculated gradient field, the region segmentation unit performs a partitioning operation. It divides the lesion area into two sub-regions with different clinical significance: the core lesion area and the peripheral transition area. The segmentation criterion directly depends on the gradient amplitude. Specifically, the core lesion area is defined as a continuous region where the spatial gradient amplitude of the final grading index exceeds a preset threshold (e.g., a gradient value greater than 0.1 perpixel). The risk index within these regions remains at a high and relatively stable level, typically corresponding to the most severe and typical central part of the lesion. The peripheral transition area is defined as the region surrounding the core lesion area where the gradient amplitude is below the aforementioned preset threshold. This region is characterized by a gradual transition of the risk index from high values ​​in the core area to low values ​​in the surrounding normal mucosa, reflecting the extent of lesion infiltration or inflammation.

[0031] After partitioning, the weight redistribution unit begins operation. Its core task is to increase the assessment weight of the core lesion area, as it has higher diagnostic value in pathology. The weight redistribution unit first calculates the arithmetic mean of the final grading indices of all pixels within the core lesion area and the arithmetic mean of the final grading indices of all pixels within the edge transition area. Then, it calculates the absolute difference between these two means. This difference quantifies the degree of separation in risk level between the core and edge regions.

[0032] The magnitude of the mean difference directly determines the weight increase of the core lesion area. The system presets a difference threshold (e.g., a mean difference greater than 0.2). When the calculated mean difference is less than or equal to this threshold, it indicates that the risk level difference between the core and peripheral areas is not significant enough, and the weight increase is set to a low value that has a linear or non-linear relationship with the difference. When the mean difference is greater than the preset difference threshold, it indicates that the core lesion area is significantly different from the surrounding tissues in terms of risk level, and its pathological significance is more prominent. At this time, the weight redistribution unit sets the weight increase of the core lesion area to a preset upper limit (e.g., increasing its weight in the overall assessment to 1.5 times the original value).

[0033] The mathematical expression of this weight adjustment process can be summarized by the following formula:

[0034] in: This indicates the new weighting coefficient assigned to the core lesion area; It is a scaling factor used to map the mean difference to the weight increment; This represents the difference between the calculated mean values ​​of the final classification indices for the core and edge regions. This represents a preset difference threshold; This represents the upper limit of the weight increase allowed by the system.

[0035] The system's overall risk assessment output is no longer a simple average of the entire lesion area, but rather an integration of the weighted average of the core lesion area and the average of the peripheral transition area (or values ​​calculated according to their original weights). This spatial partitioning-based weight redistribution mechanism allows the risk assessment results to focus more on the clinically significant core of the lesion, reducing assessment interference that may be caused by peripheral transitional or inflammatory changes, thus providing more targeted risk grading information. The entire process is automated, eliminating the need for manual area delineation, enhancing the system's practicality and objectivity in real-time endoscopic examination.

[0036] Example 4: See Figure 5 By integrating biochemical data from electronic medical records with endoscopic image analysis results, a cross-modal validation mechanism is established to enhance the comprehensiveness and reliability of system risk assessment. This implementation is based on clinical observations: the dynamic changes of serum tumor markers are potentially associated with the progression of gastric mucosal carcinogenesis, and joint analysis with endoscopic features can provide complementary evidence. A multi-source input module serves as the data access port, acquiring time-series data of serum tumor marker concentrations from patients' electronic medical records through the hospital information system's standard interface. This module is configured to periodically query the database to extract tumor marker detection records for specified patients within a specific time window before and after gastroscopy. Typical extracted markers include carcinoembryonic antigen (CEA) and carbohydrate antigens (CAAs). This data is stored in a structured format, with each record containing a detection timestamp, marker name, and concentration value. The system's preset time window is 72 hours before the examination to 24 hours after, ensuring coverage of the biological change cycle related to endoscopic examination. After data extraction, the module performs preprocessing operations, including unit standardization, outlier filtering, and timestamp standardization, forming a well-organized time-series dataset.

[0037] The core task of the cross-modal fusion module is to correlate and match the fluctuation patterns of serum biomarkers with the results of endoscopic image analysis. This module first receives preprocessed time-series data of serum biomarkers from the multi-source input module. For a selected target biomarker, the module calculates its fluctuation amplitude within the image acquisition time window. The fluctuation amplitude is defined as the absolute difference between the maximum and minimum concentrations of the biomarker within that time window, quantifying the intensity of the biomarker's variation during the examination. For example, if a patient's lowest concentration within the time window is 5.2 ng / mL and the highest is 8.7 ng / mL, the fluctuation amplitude is calculated to be 3.5 ng / mL. This module then obtains the rate of change data of the final grading index from the time-series optimization module. The rate of change is obtained by dividing the difference in the final grading index between adjacent frames by the time interval, reflecting the dynamic speed of change in the endoscopic risk assessment results. For example, if the final grading index changes from 0.58 to 0.72 in two consecutive images with a 0.5-second interval, the rate of change is 0.28 per second.

[0038] The matching process employs statistical correlation analysis, with the module calculating the Pearson correlation coefficient between the fluctuation range of serum biomarkers and the rate of change of the final grading index. The correlation coefficient ranges from -1 to 1, with a positive value indicating a positive correlation between the two trends. The system presets two correlation thresholds: a first threshold of 0.5 and a second threshold of 0.8. When the calculated correlation coefficient is greater than the first threshold but less than the second threshold, the module activates the primary state of the cross-modal validation flag. When the correlation coefficient is greater than the second threshold, the advanced state of the flag is activated. This flag serves as an internal system state variable, indicating the strength of the co-validation between biochemical indicators and endoscopic features.

[0039] Table 1: Cross-modal data matching analysis table.

[0040]

[0041] Referring to Table 1, the analysis results for the three hypothetical patients are presented. Patient P-1026's CEA marker fluctuation amplitude and the rate of change of the grading index had a correlation coefficient of 0.63, falling between the two thresholds, triggering the primary validation state. Patient P-1027's CEA correlation coefficient reached 0.87, exceeding the second threshold, triggering the advanced validation state. Simultaneously, the CA19-9 correlation coefficient of 0.79 triggered the primary state, and the system adopted the highest state as the output. Patient P-1028's correlation coefficient was below the first threshold, and the validation flag remained inactive.

[0042] The entire analysis process runs in real time during clinical examinations. When an endoscopist completes an examination of a specific anatomical region, the system automatically extracts the corresponding serum biomarker data and grading index changes for that time period, performs the aforementioned correlation calculations, and updates the validation marker status. This status information is displayed in real time on the physician's interface as a reference for decision support. For example, when the system detects an advanced validation status, the interface may display a special prompt symbol, indicating a strong correlation between the current endoscopic findings and serological evidence, increasing the clinician's confidence in risk assessment. This cross-modal validation mechanism provides multi-dimensional evidence support for early gastric cancer risk assessment, overcoming the limitations of a single data source.

[0043] Example 5: This example relates to an enhanced grading strategy when the cross-modal validation mechanism triggers an advanced state, and a compensation mechanism to maintain temporal accuracy after adjusting the sampling frequency. This implementation addresses two technical scenarios: first, how to integrate biochemical evidence to improve the biological credibility of risk assessment when serological indicators and endoscopic features show a high correlation; and second, how to resolve the processing timeline shift problem when the system accelerates sampling in response to risk fluctuations. When the cross-modal fusion module detects that the Pearson correlation coefficient between the fluctuation amplitude of serum tumor marker concentration and the rate of change of the final grading index exceeds the second correlation threshold, the cross-modal validation flag enters an advanced activation state. At this time, the dynamic grading module initiates the enhanced grading process. The system first normalizes the fluctuation amplitude of serum markers. The normalization operation is based on preset marker type-related parameters; for example, the reference fluctuation range for carcinoembryonic antigen is set to 0 to 15 ng / mL. The specific normalization process is as follows: subtract the historical minimum value of the marker from the actual fluctuation amplitude, then divide by the difference between its historical maximum and minimum values, and finally linearly map the result to the range of 0.1 to 1.0. This processing eliminates the dimensional differences between different biomarkers, making data from multiple biomarkers comparable. The normalized output value is called the correction coefficient, and its magnitude directly reflects the significance of the serological abnormality.

[0044] After obtaining the correction coefficient, the dynamic grading module performs a weighted summation calculation. This operation does not replace the original final grading index but generates a supplementary enhanced grading result. The specific calculation uses a fixed weight allocation strategy: the final grading index accounts for 70% of the weight, and the correction coefficient accounts for 30%. The new value obtained by weighted summation of the two is the enhanced grading result. For example, if the final grading index of a certain image frame is 0.65, and the normalized correction coefficient for serum carcinoembryonic antigen fluctuation is 0.8, then the enhanced grading result is 0.65 × 0.7 + 0.8 × 0.3 = 0.695. This result is displayed on the endoscopist's interface with a special identifier distinct from the basic assessment, providing cross-modal evidence for high-risk determination. The entire enhanced grading process is executed automatically during the advanced validation state and automatically terminates when the correlation coefficient falls below the threshold.

[0045] In another dimension of system operation, the sampling control unit of the timing optimization module may shorten the frame sampling interval to increase the monitoring frequency when it detects increased fluctuations in the risk index. This dynamic adjustment may cause misalignment issues in processing timing. For example, when the system shortens the sampling interval from 33 milliseconds to 20 milliseconds, the start time of the feature extraction module processing the current frame image may overlap with the end time of processing the previous frame image, causing the feature analysis of the current frame to be delayed. This accumulated delay will cause the generation time of the final grading index to gradually deviate from the actual image acquisition time, disrupting the real-time correspondence between risk assessment and dissection location.

[0046] To address this issue, the phase calibration unit of the timing optimization module immediately initiates time deviation detection after each shortening of the sampling interval. This unit records the system's high-precision clock timestamp when the final classification index of the current frame image is generated, and simultaneously obtains the hardware acquisition timestamp of the same frame image from the image acquisition module. The difference between the two timestamps represents the processing pipeline delay. The phase calibration unit continuously tracks the delay data of the three most recent frames and calculates their average value as the current system's time deviation benchmark.

[0047] The timing compensation unit dynamically adjusts the processing start time of subsequent frames based on the deviation value provided by the phase calibration unit. This unit maintains a timer whose trigger time is determined by the following factors: the expected acquisition time of the next frame (calculated based on the current sampling interval), minus the currently measured average time deviation, and then minus the estimated average processing time for feature extraction. This proactive adjustment ensures that the start time of feature extraction is precisely aligned with the available time of the image frame. For example, when the measured average time deviation is 8 milliseconds, the system advances the feature extraction start time of the next frame by 8 milliseconds to compensate for the processing delay. This mechanism controls the timing deviation within the millisecond range, ensuring that even at a high sampling rate of 60 frames per second, the risk assessment result of each frame accurately corresponds to the anatomical location of the endoscope lens.

[0048] The entire implementation process forms a closed loop in clinical practice. When an endoscopist observes a suspicious lesion area, the system may shorten the sampling interval due to fluctuations in the risk index. Simultaneously, if the patient's serum biomarkers show synchronous fluctuations, an advanced validation state is triggered to generate enhanced grading results. The temporal compensation mechanism ensures that these rapidly generated risk data remain strictly synchronized with the video stream. This multi-dimensional collaborative working mechanism enables the system to maintain the spatiotemporal consistency and multimodal complementarity of assessment results in complex and ever-changing clinical environments.

[0049] It should be noted that, in this document, relational 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 such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An endoscopic early gastric cancer risk grading system based on artificial intelligence, characterized in that, include: The image acquisition module is used to acquire the video stream of the gastroscopy examination in real time and extract single-frame images; The first feature extraction module is used to identify the morphological distribution features of microvessels on the mucosal surface and the abnormal features of glandular opening structures from the single frame image; The second feature extraction module is used to quantify the color saturation shift features and surface texture roughness features of the lesion area from the single frame image; The dynamic grading module is used to generate an initial cancer probability value based on the fusion result of the microvascular morphological distribution characteristics and the abnormal features of the glandular opening structure, and to correct the initial cancer probability value based on the dynamic correlation between the color saturation shift feature and the surface texture roughness feature. The timing optimization module is used to analyze the rate of change of the initial cancer probability value in consecutive frame images. When the rate of change exceeds a preset fluctuation threshold, the frame sampling interval of the image acquisition module is shortened.

2. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 1, characterized in that, The dynamic hierarchical module includes: A probability generation unit is used to input the microvascular morphological distribution features and the abnormal features of the glandular opening structure into a convolutional kernel size adaptive network and output an initial cancer probability value. The probability correction unit is used to calculate the covariance matrix between the color saturation shift feature and the surface texture roughness feature, and adjust the confidence weight of the initial cancer probability value according to the distribution range of the eigenvalues ​​of the covariance matrix. A grading decision unit is used to multiply the confidence weight by the initial cancer probability value to generate a final grading index.

3. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 2, characterized in that, The timing optimization module includes: The rate of change monitoring unit is used to calculate the standard deviation of the final grading index corresponding to three adjacent frames of images; The sampling control unit is used to shorten the frame sampling interval to a preset ratio of the original value when the standard deviation is greater than the first threshold and less than the second threshold; and to activate the full frame rate sampling mode when the standard deviation is greater than the second threshold.

4. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 3, characterized in that, The sampling control unit performs the following synchronously when shortening the frame sampling interval: The resource monitoring unit is used to detect the real-time memory usage of the image processing hardware platform. The attenuation adjustment unit is used to dynamically attenuate the preset ratio using a reduction factor when the memory occupancy rate exceeds a preset load threshold.

5. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 2, characterized in that, The dynamic hierarchical module also includes: The regional division unit is used to divide the lesion area into a core lesion area and a peripheral transition area according to the spatial distribution gradient of the final grading index. The weight redistribution unit is used to increase the weight of the final grading index of the core lesion area in the overall assessment.

6. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 5, characterized in that, The weight redistribution unit performs the following: Calculate the difference between the mean final grading index of the core lesion area and the peripheral transition area; The weight increase of the core lesion area is determined based on the magnitude of the mean difference. When the difference between the mean and the mean exceeds the preset difference threshold, the weight increase of the core lesion area is set to the preset upper limit.

7. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 1, characterized in that, Also includes: A multi-source input module is used to receive time-series data of serum tumor marker concentrations from the patient's electronic medical record; The cross-modal fusion module is used to correlate the fluctuation range of the serum tumor marker concentration time series data with the rate of change of the final grading index. When the correlation coefficient exceeds the preset correlation threshold, a cross-modal verification flag is generated.

8. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 7, characterized in that, When performing correlation matching, the cross-modal fusion module: The maximum fluctuation range of serum tumor marker concentrations within the image acquisition time window was extracted. Calculate the Pearson correlation coefficient between the maximum volatility and the rate of change of the final graded index; When the Pearson correlation coefficient is greater than the first correlation threshold and less than the second correlation threshold, the initial state of the cross-modal verification flag is activated; When the Pearson correlation coefficient is greater than the second association threshold, the advanced state of the cross-modal verification flag is activated.

9. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 8, characterized in that, When the cross-modal verification flag is in an advanced state, the dynamic grading module: The fluctuation range of serum tumor marker concentrations was normalized to a correction factor; The correction coefficients are weighted and superimposed with the final grading index to generate the enhanced grading result.

10. The endoscopic early gastric cancer risk grading system based on artificial intelligence according to claim 3, characterized in that, The timing optimization module also includes: The phase calibration unit is used to detect the time deviation of the final graded index generation between the current frame and the previous frame after shortening the frame sampling interval. The timing compensation unit is used to adjust the start time point of feature extraction for the next frame image based on the time deviation value, so that the feature extraction operation is synchronized with the frame output time point of the image acquisition module.

Images