Intelligent esophageal flora image detection method for risk prediction of gastroesophageal reflux disease

Abstract

The invention belongs to the technical field of medical image detection, and particularly relates to an esophageal flora image intelligent detection method for risk prediction of gastroesophageal reflux disease, which comprises the following steps: acquiring flora images and clinical data, and constructing a perception network; quantifying flora displacement characteristics by using image registration; the method comprises the following steps of: extracting flora density, aggregation degree and fluorescence gradient characteristics by combining improved U-Net + + segmentation, fusing flora and epithelial cell lesion characteristics through space-time attention, generating a flora-host association map, constructing a flora dynamic evolution model by utilizing a space-time map convolutional network, predicting Barrett esophageal occurrence probability and flora diffusion trend, and predicting the occurrence probability of the Barrett esophagus. Calculating a lesion risk level through a flora imbalance quantification algorithm; in combination with reflux frequency and pH fluctuation, a flora-environment interaction model is used for simulating a field planting coefficient and correcting parameters, the parameters are transmitted to a diagnosis platform to deduce canceration risks, grading suggestions are triggered, and a visual report is generated. Therefore, the problems of long detection period, weak microscopic analysis capability and the like in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image detection, and particularly relates to an esophageal flora image intelligent detection method for gastroesophageal reflux disease risk prediction. BACKGROUND

[0002] With the development of precise medical technology, the market urgently needs an intelligent detection system for reflux esophagitis based on flora-mucosa correlation analysis, which requires multi-modal data fusion, dynamic feature extraction, and multi-dimensional pathological deduction capabilities. However, the existing technology cannot realize deep fusion of flora and mucosal lesion data, lacks flora-host collaborative reasoning framework and dynamic risk quantification mechanism, and is difficult to support precise detection in complex digestive tract scenarios. There is a significant gap between the existing technology and the technical requirements of constructing flora-host correlation atlas and real-time correction of detection sensitivity.

[0003] However, the traditional flora detection method has the following key defects: the detection method relying on gastroscopy biopsy and culture has a long cycle (24-72 hours), and is insufficient in capturing early microflora changes; image analysis can only identify macroscopic lesions such as esophageal mucosal erosion, and lacks micro-analysis of flora community structure and dominant bacterial species morphology; the data dimension is limited to single pathological images or biochemical indicators, and a multi-modal perception network has not been formed, lacking a spatiotemporal attention mechanism for collaborative processing of flora dynamics and epithelial cell lesions, resulting in a data gap between flora characteristics and mucosal damage. With the advancement of precision medicine, the market urgently needs high-sensitivity, low-latency flora intelligent detection technology. However, the existing technology has limited feature expression capability and insufficient detection robustness, making it difficult to support early warning applications in complex scenarios. SUMMARY

[0004] The present application provides an esophageal flora image intelligent detection method for gastroesophageal reflux disease risk prediction to solve the problems of long detection cycle and weak micro-analysis capability in the prior art.

[0005] The first aspect of this application provides an intelligent detection method for esophageal microbiota images for predicting the risk of gastroesophageal reflux disease (GERD), comprising the following steps: acquiring microscopic images of microbiota and clinical feature data of esophageal mucosal samples; constructing a perceptual network based on the microscopic images and clinical feature data; quantifying the spatial displacement features of microbiota in esophageal sample sequences using an image registration algorithm based on the perceptual network; simultaneously segmenting the image using an improved U-Net++ network to extract microbiota density, aggregation, and fluorescence intensity gradient features; and fusing microbiota dynamics and epithelial cell lesion features through a spatiotemporal attention mechanism to generate a microbiota-host association map; based on... The aforementioned microbiota-host association map utilizes a spatiotemporal graph convolutional network to construct a dynamic microbiota evolution model, predicting the probability of Barrett's esophagus and the microbiota spread trend. Based on the predicted probability of Barrett's esophagus and the microbiota spread trend, the lesion risk level is calculated using a microbiota loss quantification algorithm. Combining the patient's gastric acid reflux frequency and esophageal pH fluctuation data, a microbiota-environment interaction model is used to simulate the microbiota colonization coefficient, and sensitivity parameters are corrected in real time. The corrected sensitivity parameters are transmitted to a clinical diagnostic platform for multi-dimensional cancer risk extrapolation, triggering graded diagnostic recommendations and generating a visual report of a risk heatmap.

[0006] Preferably, generating a visualization report of a risk heatmap includes: constructing a microbiota-carcinogenesis co-diffusion model; based on the microbiota-carcinogenesis co-diffusion model, simulating the invasion trajectory of pathogenic bacteria in the esophageal mucosa layer using an improved particle swarm optimization algorithm, mapping the spatial association between microbiota aggregation areas and epithelial cell dysplasia areas to carcinogenesis probability values ​​using a Gaussian mixture model, and generating a carcinogenesis risk probability cloud map; superimposing the carcinogenesis risk probability cloud map onto a three-dimensional esophageal anatomical model, characterizing the intensity of carcinogenesis risk through color gradients, and generating a heatmap visualization report annotating key path nodes of microbiota pathogenicity, the expected carcinogenesis progression time, and the coordinates of high-risk areas.

[0007] Preferably, the spatiotemporal attention mechanism formula is:

[0008] in Output results for attention; This is the microbial community dynamics feature matrix; A matrix representing the characteristics of epithelial cell lesions; The query vector generated by M; The key vector generated by E; Key vector The transpose of the matrix; The dimension of the key vector; It is a value vector; The scaled score represents the microbiome-host feature association score.

[0009] Preferably, the colonization coefficient of the flora is simulated by using a flora-environment interaction model, and the sensitivity parameter is corrected in real time, comprising: constructing a flora-environment interaction dynamics model; inputting the frequency of gastric acid reflux, the pH value fluctuation data of the esophagus, and the concentration data of the flora metabolites into the flora-environment interaction dynamics model, simulating the change curve of the colonization coefficient of the flora with the environmental factors by differential equations; and correcting the sensitivity parameter in real time based on the change curve to control the error between the predicted value and the clinically measured value.

[0010] Preferably, the flora dynamic evolution model is constructed by using a spatio-temporal graph convolution network to obtain the Barrett's esophagus occurrence probability and the flora diffusion trend, comprising: constructing a flora dynamic evolution model; based on the flora dynamic evolution model, setting the lower esophageal dentate line and the precancerous lesion high-incidence area of the cardia mucosal fold as the core nodes of the graph network, combining the micro features of the flora biofilm thickness, the apoptosis rate of epithelial cells, and the concentration of inflammatory factors, modeling the spatio-temporal correlation between the flora invasion depth and the degree of epithelial dysplasia by using a spatio-temporal graph convolution network, and predicting the Barrett's esophagus occurrence probability and the corresponding flora diffusion trend.

[0011] Preferably, a perception network is constructed, and based on the perception network, an image registration algorithm is used to quantify the spatial displacement features of the flora sequence of the esophageal sample, comprising: constructing a multi-modal perception network; according to the multi-modal perception network, the texture features of the flora microscopic image and the structured information of the clinical feature data are fused to generate a dynamic weight matrix; based on the dynamic weight matrix, an image registration algorithm is used to calculate the three-dimensional spatial displacement vector of the flora in the flora sequence of the esophageal sample, and the migration distance and direction angle parameters of the flora are quantified.

[0012] Preferably, an improved U-Net++ network is used to segment the image, extract the flora density, aggregation degree, and fluorescence intensity gradient features, comprising: constructing an improved U-Net++ network; based on the improved U-Net++ network, the flora microscopic image is segmented at the pixel level to generate a flora region mask; according to the flora region mask, the flora density index is calculated by using a density clustering algorithm, the aggregation degree features are extracted by using a morphological operator, and the gradient amplitude and direction features of the fluorescence intensity are calculated by using a Sobel operator.

[0013] Preferably, the improved U-Net++ network formula is:

[0014] wherein, is the image segmentation accuracy at time t; is the basic segmentation accuracy; is the flora fluorescence contrast at time t; is the standard contrast threshold; is the flora region weight coefficient; is a noise influence factor; is an image noise level at time t.

[0015] Preferably, the trigger hierarchical diagnosis suggestion includes: setting a four-level diagnosis mechanism, when the Barrett's esophagus occurrence probability is less than 20%, the flora has no abnormal diffusion, and the epithelial cells are normal, it is divided into a first-level low risk, and diet intervention and reexamination every 6 months are recommended; when the occurrence probability is 20%-50%, the flora is locally aggregated with mild epithelial hyperplasia, it is divided into a second-level medium-low risk, and probiotic intervention and gastroscopy follow-up are recommended; when the occurrence probability is 50%-80%, the flora diffuses to the upper part of the esophagus and is accompanied by moderate dysplasia, it is divided into a third-level medium-high risk, and endoscopic mucosal resection evaluation and targeted antibacterial treatment are triggered; and when the occurrence probability is greater than 80%, the flora invades the whole layer and is accompanied by severe dysplasia or early cancer, it is divided into a fourth-level high risk, and multidisciplinary consultation and radical surgery evaluation are started.

[0016] The second aspect embodiment of the present application provides an esophageal flora image intelligent detection system for gastroesophageal reflux disease risk prediction, including: an acquisition module, configured to acquire flora microscopic images and clinical feature data of an esophageal mucosa sample; an extraction module, configured to construct a perception network according to the flora microscopic images and the clinical feature data, to quantize esophageal sample sequence flora spatial displacement features based on the perception network by using an image registration algorithm, to extract flora density, aggregation degree and fluorescence intensity gradient features by using an improved U-Net++ network segmentation image, to generate a flora-host correlation atlas by fusing flora dynamics and epithelial cell lesion features through a spatiotemporal attention mechanism; and a generation module, configured to construct a flora dynamic evolution model by using a spatiotemporal graph convolution network according to the flora-host correlation atlas, to predict Barrett's esophagus occurrence probability and flora diffusion trend, to calculate lesion risk levels by using a flora imbalance quantization algorithm based on the predicted Barrett's esophagus occurrence probability and flora diffusion trend, to simulate flora colonization coefficients by using a flora-environment interaction model in combination with patient gastric acid reflux frequency and esophageal pH value fluctuation data, to real-time correct sensitivity parameters, to transmit the corrected sensitivity parameters to a clinical diagnosis platform, to perform multidimensional cancer risk deduction, to trigger hierarchical diagnosis suggestions, and to generate a visual report of a risk heat map.

[0017] Therefore, the present application includes the following beneficial effects: the embodiment of the present application integrates the esophageal mucosa flora micrograph and the clinical multi-dimensional data, accurately quantifies the spatial displacement characteristics of the flora by means of the image registration algorithm, realizes the efficient extraction of the microcosmic characteristics of the flora by combining the improved U-Net++ network, and then deeply integrates the dynamic changes of the flora and the lesion characteristics of the epithelial cells through the space-time attention mechanism. The flora-host correlation atlas generated breaks the fragmented state of the flora data and the mucosa damage information in the traditional detection, realizes the cross-scale correlation analysis from the microcosmic flora behavior to the macroscopic lesion, and significantly improves the capture ability of the early microcosmic flora abnormalities and the occult mucosa damage. The flora dynamic evolution model constructed relying on the space-time graph convolution network can accurately predict the occurrence probability of Barrett's esophagus and the diffusion trend of the flora. Combined with the environmental data such as the frequency of gastric acid reflux and the pH value fluctuation of the esophagus, the detection sensitivity parameters are corrected in real time through the flora-environment interaction model, effectively solving the problem of poor adaptability to individual differences and environmental fluctuations in the traditional method, making the risk assessment more in line with the actual pathological state of the patient, and greatly improving the accuracy and timeliness of the precancerous lesion warning. The final generated hierarchical diagnosis suggestion and risk thermodynamic map visualization report provide intuitive decision-making basis for the clinic from the flora pathogenic path to the cancer risk, not only shortening the detection period (more than 60% shorter than the traditional culture method), but also reducing the risk of over-treatment and missed diagnosis through accurate stratified diagnosis and treatment suggestions, and comprehensively improving the accuracy, clinical adaptability and diagnosis and treatment efficiency of the detection of reflux esophagitis and related precancerous lesions, providing new technical support for the early intervention and precise management of digestive tract diseases. Therefore, the problems of long detection period and weak microcosmic analysis capability in the prior art are solved.

[0018] Additional aspects and advantages of the application will be set forth in part in the description that follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 A flow chart of an esophageal flora image intelligent detection method for gastroesophageal reflux disease risk prediction according to an embodiment of the present application is provided; Figure 2 A schematic diagram of esophageal flora micrograph sequence analysis of a gastric mucosa sample according to an embodiment of the present application is provided; Figure 3 A schematic diagram of reflux esophagitis flora image intelligent detection according to an embodiment of the present application is provided; Figure 4 A schematic diagram of Barrett's esophagus risk assessment according to an embodiment of the present application is provided; Figure 5This is a schematic diagram of an intelligent detection method for esophageal microbiota images for predicting the risk of gastroesophageal reflux disease according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an intelligent esophageal microbiota image detection system for predicting the risk of gastroesophageal reflux disease according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0021] The following describes, with reference to the accompanying drawings, an intelligent detection method for esophageal microbiota images for predicting the risk of gastroesophageal reflux disease (GERD) based on embodiments of this application. Addressing the issue of long detection cycles mentioned in the background section, this application provides an intelligent detection method for esophageal microbiota images for predicting the risk of GERD. This method integrates esophageal mucosal microscopic images with multi-dimensional clinical data, accurately quantifies the spatial displacement characteristics of the microbiota using image registration algorithms, efficiently extracts microscopic features of the microbiota using an improved U-Net++ network, and deeply fuses dynamic changes in the microbiota with epithelial cell lesion features through a spatiotemporal attention mechanism. The generated microbiota-host association map breaks down the separation between microbiota data and mucosal damage information in traditional detection methods, achieving cross-scale association analysis from microscopic microbiota behavior to macroscopic lesions, significantly improving the ability to capture early microbial abnormalities and occult mucosal damage. A microbial dynamic evolution model constructed based on a spatiotemporal graph convolutional network can accurately predict the probability of Barrett's esophagus and the trend of microbial diffusion. Combined with environmental data such as the frequency of gastric acid reflux and esophageal pH fluctuations, the model uses a microbial-environment interaction model to correct detection sensitivity parameters in real time. This effectively solves the problem of poor adaptability to individual differences and environmental fluctuations in traditional methods, making risk assessment more closely aligned with the patient's actual pathological state and significantly improving the accuracy and timeliness of precancerous lesion early warning. The final generated tiered diagnostic recommendations and risk heatmap visualization report provide clinicians with intuitive decision-making basis from the pathogenesis pathway of microbial flora to the risk of cancer. This not only shortens the detection cycle (by more than 60% compared to traditional culture methods) but also reduces the risk of overtreatment and missed diagnosis through precise stratified treatment recommendations. It comprehensively improves the accuracy, clinical suitability, and diagnostic efficiency of reflux esophagitis and related precancerous lesion detection, providing new technical support for the early intervention and precise management of digestive tract diseases. Thus, it solves the problems of long detection cycles and weak microscopic analysis capabilities in existing technologies.

[0022] Specifically, Figure 1 A flowchart of an esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease provided by the embodiment of the present application.

[0023] As Figure 1 shown, the esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease includes the following steps: In step S101, the flora microscopic image and the clinical feature data of the esophageal mucosa sample are obtained.

[0024] The clinical feature data refers to the collective term of various clinical information such as symptoms, medical history, examination results, and treatment records related to the patient's disease.

[0025] It can be understood that the flora microscopic image of the esophageal mucosa sample can provide micro features such as flora morphology and density, and the clinical feature data can reflect macro information such as patient symptoms, medical history, and examination results, providing a comprehensive data basis for constructing flora-host correlation atlas and analyzing the correlation between flora and lesions, accurately capturing micro flora abnormalities, and integrating patient individual clinical background to improve the comprehensiveness and pertinence of detection.

[0026] In step S102, a perception network is constructed according to the flora microscopic image and the clinical feature data, and based on the perception network, an image registration algorithm is used to quantify the spatial displacement features of the flora sequence sample, and an improved U-Net++ network is used to segment the image to extract flora density, aggregation degree, and fluorescence intensity gradient features, and through a spatiotemporal attention mechanism, flora dynamics and epithelial cell lesion features are fused to generate a flora-host correlation atlas.

[0027] The image registration algorithm is a technical method that finds the spatial transformation relationship to align multiple images obtained under different conditions (such as sample images at different times and angles) in space, eliminate positional differences, and achieve accurate comparative analysis.

[0028] It can be understood that by using the image registration algorithm, the sample images at different times and angles can be spatially aligned to eliminate positional differences and achieve accurate comparison, accurately quantify the spatial displacement features of the flora in the sequence sample, avoid errors in flora dynamic analysis caused by image misalignment, provide a reliable spatial reference for extracting flora density, aggregation degree, and other features, more accurately capture flora migration, diffusion, and other dynamic rules, lay a solid data foundation for generating a flora-host correlation atlas, and improve the accuracy and reliability of flora dynamic analysis in the overall detection.

[0029] It should be noted that the formula of the image registration algorithm is:

[0030] in, The x-coordinate of the transformed point; The ordinate of the transformed point; These are the homogeneous coordinates of the transformed points; These are the homogeneous coordinates of the point before the transformation; The x-coordinate of the point before the transformation; The ordinate of the point before the transformation; This is the core parameter matrix of the affine transformation; This represents the translation in the x-direction; This represents the translation in the y-direction; , , , This is the linear transformation part.

[0031] Microbiome-host association maps are visualization and analysis tools used to systematically analyze the interaction between microbial communities and hosts. Their core function is to reveal the dynamic relationship between the two in physiological metabolism, immune regulation, and disease development through multi-dimensional data association.

[0032] For example, such as Figure 2 As shown, in the microscopic image sequence analysis of gastric mucosal microbiota, when using the image registration algorithm, the first frame of the sequence is used as a reference template to extract the edge contour feature points and specific fluorescent marker anchor points of the microbiota. A scale-invariant feature transformation algorithm is then used to match the homologous features of subsequent frames. To address spatial misalignment caused by sample slice curling and field-of-view shift, the optimal rigid transformation matrix (including rotation angle and translation) is calculated to calibrate each frame to a unified coordinate system. After registration, the spatial coordinate changes of the labeled microbiota are tracked, and the displacement parameters after removing noise interference are quantified. For example, a certain streptococcal community has an average displacement of 15.6 μm over 3 hours, with the displacement direction matching the epithelial cell arrangement direction by 82%. The population displacement dispersion has decreased from 9.3 μm before registration to 3.5 μm. This approach accurately removes spurious displacements caused by physical disturbances in the samples, providing a standardized spatial benchmark for subsequent temporal comparison of microbiota density and spatial distribution analysis of aggregation, ensuring the accuracy of the fusion of microbiota dynamic characteristics and epithelial lesion characteristics, and supporting the quantification of spatial interactions in the microbiota-host association map.

[0033] In this embodiment of the application, the spatiotemporal attention mechanism formula is as follows:

[0034] in, Output results for attention; This is the dynamic feature matrix of the microbial community; A matrix representing the characteristics of epithelial cell lesions; The query vector generated by M; The key vector generated by E; is a key vector is a transpose matrix of the key vector; is a dimension of the key vector; is a value vector; is a scaled flora-host feature correlation score.

[0035] It can be understood that, when the spatiotemporal attention mechanism of the embodiment of the present application fuses the flora dynamics and the epithelial lesion features, the spatial dimension focuses on the key area where the flora aggregation area and the lesion are closely related, and weakens irrelevant background interference; the time dimension captures the time sequence correlation between the flora displacement rhythm and the lesion development, preferentially strengthens the dynamic features with clinical significance, improves the targeting and precision of feature fusion, reduces the interference of noise information on correlation analysis, and makes the generated flora-host correlation atlas more clearly reflect the core law, providing visual data for the clinical interpretation of the causal relationship between flora abnormalities and disease progression, and enhancing the adaptability of the model to complex pathological scenarios.

[0036] For example, when processing esophageal sample sequences, the spatiotemporal attention mechanism first focuses on the time dimension, analyzes the rate and direction change of flora spatial displacement, and gives higher attention weight to time periods with displacement amplitude exceeding the threshold, so as to highlight the key time nodes of flora migration, such as the time when the flora in the lesion area suddenly aggregates or spreads. Then, in the spatial dimension, the flora high-density area segmented by the improved U-Net++ network, the fluorescence intensity change area, and the morphological features of the epithelial lesion area are compared, and the area with high overlap degree is paid more attention to, and the spatial weight value is dynamically adjusted. Through the synergistic optimization of the time and space weights, irrelevant noise information can be effectively filtered, and the core correlation information between the flora dynamics and the lesion features can be significantly strengthened in the fusion process, so as to accurately outline the spatiotemporal linkage mode of the two in different time periods and different areas, and provide more accurate feature support for the construction of the flora-host correlation atlas.

[0037] It should be noted that the displacement amplitude threshold is not a fixed value and needs to be dynamically set according to the specific scene. Generally, it depends on the sample type (such as intestinal, skin samples), the flora type (such as the motion characteristics difference between cocci and bacilli), and the imaging resolution (such as the actual spatial scale corresponding to micrometer-level pixels). For the coccus flora of intestinal mucosa samples, if the imaging resolution is 0.5 μm / pixel, the threshold is set to 5-10 pixels (i.e. 2.5-5 μm), because this range can effectively distinguish the random displacement of flora due to Brownian motion from the directional migration driven by the lesion microenvironment; and for the more motile bacilli, the threshold is increased to 15-20 pixels (7.5-10 μm) to capture their significant aggregation or diffusion behavior.

[0038] In this embodiment, a perceptual network is constructed, and based on the perceptual network, an image registration algorithm is used to quantify the spatial displacement characteristics of the microbial community in the esophageal sample sequence. This includes: constructing a multimodal perceptual network; based on the multimodal perceptual network, fusing the texture features of the microscopic image of the microbial community with the structured information of clinical feature data to generate a dynamic weight matrix; and based on the dynamic weight matrix, using an image registration algorithm to calculate the three-dimensional spatial displacement vector of the microbial community in the esophageal sample sequence, and quantifying the migration distance and orientation angle parameters of the microbial community.

[0039] Among them, the multimodal perception network is a network model that integrates multiple modal data such as microscopic images of microbial communities and clinical features, and extracts key cross-modal information through collaborative perception, providing multi-source data integration support for microbial community-host association analysis.

[0040] It is understood that the embodiments of this application utilize a multimodal sensing network to effectively integrate the texture features of microscopic images of microbiota with the structured information of clinical feature data. Through cross-modal collaborative sensing, more comprehensive key information is extracted, and the generated dynamic weight matrix can accurately provide data for image registration algorithms. This makes the calculation of the three-dimensional spatial displacement vector of microbiota in esophageal sample sequences, the quantification of migration distance and orientation angle parameters more in line with clinical practice, improves the completeness of feature extraction, and enhances the accuracy of subsequent quantitative analysis.

[0041] For example, such as Figure 3 As shown, in the intelligent detection of esophageal microbiota images in reflux esophagitis, the multimodal perception network deeply integrates multidimensional texture features of esophageal microscopic images (such as gradient changes in microbiota aggregation density, differences in attachment morphology to damaged epithelium, and intensity distribution of fluorescently labeled microbiota) with multidimensional clinical feature data (such as acid exposure time from 24-hour esophageal pH monitoring, endoscopic inflammation scores, patient reflux frequency, and symptom duration). Through cross-modal association mining, it accurately captures the potential correlation between microbiota colonization patterns and disease progression, generating a dynamic weight matrix—for example, assigning higher weights to microbiota adhesion texture in erosion areas and the proportion of total time with pH < 4 during the acute phase, while strengthening the weights of microbiota biofilm formation texture and tissue fibrosis degree during the chronic phase. Based on this matrix, combined with image registration algorithms, the three-dimensional spatial displacement of the microbiota in the inflammatory area (such as the migration distance to the erosion site and the diffusion angle along the longitudinal axis of the esophagus) can be accurately quantified, enabling precise identification of the abnormal colonization patterns of the characteristic microbiota in reflux esophagitis (such as the abnormal adhesion of Helicobacter pylori to the squamous epithelium), improving the detection specificity by 15%-20% and greatly enhancing the fit with clinical diagnosis and treatment needs.

[0042] In this embodiment, an improved U-Net++ network is used to segment images and extract bacterial community density, aggregation degree, and fluorescence intensity gradient features. This includes: constructing an improved U-Net++ network; performing pixel-level segmentation on the bacterial community microscopic image based on the improved U-Net++ network to generate a bacterial community region mask; calculating the bacterial community density index using a density clustering algorithm based on the bacterial community region mask; extracting aggregation degree features using morphological operators; and calculating the gradient magnitude and direction features of fluorescence intensity using the Sobe operator.

[0043] Density clustering algorithm is a method of clustering based on the spatial density of sample points. It divides samples with connected densities into the same cluster and identifies noise points in low-density areas, thereby achieving cluster analysis of data with complex distribution patterns.

[0044] It is understood that the embodiments of this application employ density clustering algorithms to accurately process the complex patterns of non-uniform bacterial distribution based on bacterial community region masks. By identifying low-density noise points to remove interfering information, and dividing densely connected bacterial communities into clusters to calculate density indices, the accuracy of bacterial community density quantification is improved. This provides a reliable basis for subsequent extraction of aggregation features using morphological operators and calculation of fluorescence intensity gradient features using the Sobel operator, accurately capturing the distribution characteristics of bacterial communities in lesion areas and enhancing the effectiveness of bacterial community-lesion correlation analysis.

[0045] It should be noted that the microbial community region mask is a binary image generated after pixel-level segmentation of the microbial community image. The region where the microbial community is located is marked (e.g., the pixel value is 1), while the background region is marked with other values ​​(e.g., the pixel value is 0), which is used to accurately define the spatial range of the microbial community.

[0046] Density clustering algorithm formula:

[0047] in, The data points to be determined; With p as the center and radius as The set of all data points contained within the area; The minimum number of neighborhood points specified by the user.

[0048] Morphological operator formulas:

[0049]

[0050]

[0051]

[0052] in, a set of pixels corresponding to the colony region mask; a structure element; a coordinate vector of the pixel; a relative coordinate of the pixel within the structure element B.

[0053] Sobe operator formula:

[0054]

[0055]

[0056]

[0057] wherein, a colony microscopic image; a gradient response of the image in the x direction; a gradient response of the image in the y direction; a gradient magnitude; a gradient direction.

[0058] For example, in the analysis of a colony microscopic image, a density clustering algorithm such as DBSCAN can accurately identify colony aggregation regions. For example, for a colony region mask generated by improved U-Net++ network segmentation, first set the neighborhood radius ε (such as 3 pixels) and the minimum number of points MinPts (such as 5 pixels), and calculate the number of pixels in the ε neighborhood of each colony pixel. If the number of pixels in the neighborhood of a certain pixel is ≥ MinPts, it is determined to be a core point; the pixels in the ε neighborhood of the core point form a "cluster" (dense colony area) through density reachability, and the pixels that cannot be covered by any core point are marked as noise points (scattered colonies). By calculating the ratio of the total number of pixels in the cluster to the area of the cluster, the colony density index can be quantified to assist in analyzing the colonization pattern of the colony in scenarios such as reflux esophagitis. This method does not need to pre-set the number of clusters, can automatically discover aggregation areas of any shape, and is robust to noise, significantly better than traditional K-means algorithms.

[0059] In the embodiments of the present application, the improved U-Net++ network formula is:

[0060] wherein, image segmentation accuracy at time t; base segmentation accuracy; colony fluorescence contrast at time t; standard contrast threshold; colony region weight coefficient; noise influence factor; is the image noise level at time t.

[0061] It can be understood that the embodiments of the present application can accurately segment the microbial community microscopic image by using the improved U-Net++ network, effectively extract key features such as microbial community density, aggregation degree and fluorescence intensity gradient, and enhance the multi-scale feature fusion capability through the optimized network structure, acutely capture the subtle morphological and spatial distribution differences of the microbial community, and improve the segmentation accuracy and the integrity of feature extraction. At the same time, for the complex interweaving scene of the microbial community and the epithelial cells in the sample, the improved network has stronger context information integration capability, which provides high-quality basic data support for the subsequent spatio-temporal attention mechanism to integrate the microbial community dynamics and lesion characteristics, generate accurate microbial community-host correlation atlas, and improve the reliability of clinical feature analysis.

[0062] For example, in the fluorescence microscopic image analysis of the intestinal flora sample, the improved U-Net++ network significantly improves the segmentation accuracy of dense microbial community colonies by introducing a multi-scale feature fusion module and a spatial attention mechanism. For example, when dealing with images with weak fluorescence signals and complex interweaving of microbial communities and epithelial cells, the network dynamically fuses feature information at different levels through an optimized dense skip connection structure, not only accurately segments the boundaries of the microbial community (the Dice coefficient is above 0.95), but also accurately extracts the density gradient, aggregation morphology and fluorescence intensity distribution characteristics of the microbial community. Experiments show that the improved network has an accuracy of 95% in images containing more than 10,000 bacterial cells, and can effectively distinguish complex clusters formed by microbial community aggregation from the surrounding host tissue. The high-resolution segmentation results output by the network provide pixel-level positioning basis for subsequent quantitative analysis of the spatial displacement of the microbial community relative to the diseased epithelial cells (such as detecting the migration path of the microbial community through an image registration algorithm), and lay a foundation for generating a microbial community-host correlation atlas, helping to reveal the potential relationship between microbial community dynamics and host lesions.

[0063] In step S103, according to the microbial community-host correlation atlas, a microbial community dynamic evolution model is constructed using a spatio-temporal graph convolution network to predict the probability of Barrett's esophagus occurrence and the diffusion trend of the microbial community. Based on the predicted probability of Barrett's esophagus occurrence and the diffusion trend of the microbial community, a lesion risk grade is calculated by a microbial imbalance quantification algorithm, combined with the patient's gastric acid reflux frequency and esophageal pH value fluctuation data, a microbial community-environment interaction model is used to simulate the microbial colonization coefficient, and the sensitivity parameter is corrected in real time. The corrected sensitivity parameter is transmitted to the clinical diagnosis platform for multi-dimensional cancer risk deduction, triggering a hierarchical diagnosis suggestion and generating a visual report of the risk heat map.

[0064] Among them, the microbial community dynamic evolution model is an analysis model that describes or predicts the dynamic changes of the composition, structure and function of the microbial community under the influence of time lapse and environmental and host state factors, and integrates multi-dimensional data.

[0065] It can be understood that the bacterial community dynamic evolution model of the embodiment of the application can accurately predict the Barrett's esophagus occurrence probability and the bacterial community diffusion trend by integrating the bacterial community-host correlation graph and the spatiotemporal data, and provide dynamic basis for lesion risk assessment; the integration capability of multi-dimensional data can support the effective operation of the bacterial community imbalance quantification algorithm and the bacterial community-environment interaction model, simulate the bacterial community colonization coefficient and real-time correction sensitivity parameters, improve the accuracy and timeliness of disease risk prediction, provide a dynamically updated key parameter for the clinical diagnosis platform, promote the multi-dimensional cancer risk deduction to be more in line with the actual situation of patients, and enhance the scientificity and pertinence of clinical decision-making through hierarchical diagnosis suggestion and visual report.

[0066] It should be noted that the bacterial community dynamic evolution model:

[0067]

[0068] wherein, is the bacterial community state vector at time t; is the adjacency matrix of the bacterial community-host correlation graph; is the host state parameter; is the environmental factor vector; is the network parameter of the ST-GCN; is the random disturbance term; is the bacterial community state vector at time t+1; is the Barrett's esophagus occurrence probability at time t+1; is the sigmoid activation function; , is the probability prediction layer parameter.

[0069] Spatiotemporal graph convolution network formula:

[0070] wherein, is the graph node feature matrix at time t+1; is the kth graph adjacency matrix; is the convolution weight corresponding to the dth feature dimension; is the time convolution kernel; is the number of spatial convolution kernels; is the feature dimension; is the activation function; is the time convolution operation; is the dth-dimensional feature matrix of the node in the graph at time t.

[0071] Bacterial community-environment interaction model formula:

[0072] wherein, is the colonization coefficient of the microbiota at time t; is the state vector of the microbiota; is the vector of environmental factors; is the state parameter of the host; is the weight matrix of the microbiota; is the weight matrix of the environment; is the weight matrix of the microbiota-environment interaction; is the weight matrix of the host factors; is the activation function; b is the bias term, is the random disturbance term.

[0073] Quantification algorithm formula of microbiota imbalance:

[0074] wherein, is the risk level of microbiota imbalance quantification at time t; is the characteristic vector of the microbiota at time t; is the characteristic baseline vector of the healthy population; is the difference measure between the two; is the current microbiota diversity index; is the average diversity index of the healthy population; is the degree of diversity imbalance; is the predicted probability of Barrett's esophagus occurrence; is the microbiota spread trend index; , , is the weight coefficient, respectively measuring the contribution degree of the microbiota composition deviation, diversity anomaly, and disease correlation risk.

[0075] For example, as Figure 4As shown, in the Barrett's esophagus risk assessment, the flora dynamic evolution model is based on the flora-host correlation atlas, integrates the esophageal flora abundance, spatial distribution and host epithelial cell lesion characteristics through the spatiotemporal graph convolution network, predicts the probability of Barrett's esophagus occurrence within 6 months (such as 32%) and the diffusion path of specific pathogenic bacteria (such as Helicobacter pylori) of the patient; combined with the prediction results, the flora imbalance quantitative algorithm calculates the risk level of moderate lesions, and then links the patient's gastric acid reflux frequency (5 times per week) and esophageal pH value fluctuation data, simulates the flora colonization coefficient (0.68) through the flora-environment interaction model, and real-time corrects the sensitivity parameters (such as adjusting the pH value influence weight from 0.2 to 0.35), finally provides dynamic parameters for the clinical diagnosis platform, deduces the grading suggestion of "endoscopy review", and generates the heat map report of the superposition of the flora dense area in the upper esophagus and the lesion risk.

[0076] In the embodiments of the present application, the visualization report of the risk heat map is generated, including: constructing a flora-cancer synergistic diffusion model; based on the flora-cancer synergistic diffusion model, combining an improved particle swarm algorithm to simulate the invasion trajectory of pathogenic bacteria in the esophageal mucosa layer, mapping the spatial correlation between the flora aggregation area and the epithelial cell dysplasia area into a cancer risk probability value through a Gaussian mixture model, and generating a cancer risk probability cloud chart; superimposing the cancer risk probability cloud chart on the esophageal three-dimensional anatomical model, representing the cancer risk intensity through a color gradient, and generating a heat map visualization report labeled with key path nodes of flora pathogenesis, predicted cancer progression time and high-risk area coordinates.

[0077] The improved particle swarm algorithm is an intelligent optimization algorithm that optimizes the particle update strategy, enhances the local search ability or introduces an adaptive parameter adjustment mechanism on the basis of the standard particle swarm algorithm to more accurately simulate the invasion trajectory of pathogenic bacteria in the esophageal mucosa layer and other complex optimization processes.

[0078] It can be understood that the embodiments of the present application accurately simulate the invasion trajectory of pathogenic bacteria in the esophageal mucosa layer by using the improved particle swarm algorithm, capture the subtle path changes in flora diffusion, and avoid the defect that the standard algorithm is easy to fall into local optimum; at the same time, the adaptive parameter adjustment mechanism can dynamically adapt to the heterogeneity of the esophageal mucosa microenvironment, improve the authenticity and accuracy of the trajectory simulation. It provides reliable trajectory data for the subsequent Gaussian mixture model to map the spatial correlation between flora aggregation and epithelial cell dysplasia, makes the generated cancer risk probability cloud chart more in line with physiological reality, ensures that the heat map superimposed on the esophageal three-dimensional anatomical model is more accurate in labeling the key path nodes of flora pathogenesis, the predicted cancer progression time and the high-risk area coordinates, enhances the reference value of the risk heat map visualization report for clinical diagnosis, and enables doctors to efficiently identify high-risk areas of cancer.

[0079] It should be noted that the formula of the improved particle swarm algorithm is:

[0080]

[0081]

[0082] wherein, is the velocity of the i-th particle at time t; is the position of the i-th particle at time t; is the inertia weight at the t-th iteration; is the velocity of the i-th particle at time t+1; , is the learning factor; , is the random number; is the best position found so far by the i-th particle; is the global optimal position; is the local search correction term; is the correction coefficient; is the position of the i-th particle at time t+1; is the upper limit of the inertia weight; is the lower limit of the inertia weight; is the maximum number of iterations; is the time.

[0083] Formula of the synergistic diffusion model of flora-carcinogenesis:

[0084] wherein, is the time rate of change of the pathogenic bacteria concentration at spatial position x, time t; is the diffusion coefficient of the flora; is the Laplacian operator; is the basic growth rate of the flora; is the environmental carrying capacity of the flora; is the synergistic coefficient of flora-carcinogenesis; is the pathogenic bacteria concentration at spatial position x, time t; is the density of epithelial dysplasia cells at spatial position x, time t; is the natural decay rate of the flora; is the time rate of change of the density of epithelial dysplasia cells at spatial position x, time t; is the number of tumor cells at spatial position x, time t; is the diffusion coefficient of the carcinogenic cells; is the basic growth rate of the carcinogenic cells; is the synergistic coefficient of flora-carcinogenesis; is the natural decay rate of cancer cells; is the self-interaction coefficient of cancer cells; is the density-dependent term of cancer cells.

[0085] Gaussian mixture model formula:

[0086] wherein, is the cancer probability density corresponding to a spatial position x of the esophageal mucosa layer; is the number of mixed Gaussian distribution components; is the kth Gaussian distribution component; is the weight of the kth Gaussian component; is the probability density function of the kth Gaussian distribution; is the mean vector of the kth Gaussian distribution; is the covariance matrix of the kth Gaussian distribution.

[0087] For example, in the intelligent detection of intestinal flora fluorescence microscopic images, the improved particle swarm algorithm realizes accurate positioning and feature extraction of dense flora regions by fusing adaptive inertia weight and neighborhood particle information correction mechanism. In view of the fact that the edges of flora and epithelial cells in the image are blurred, the fluorescence signal is uneven, and the local aggregation forms a block interference, the algorithm dynamically adjusts the particle search strategy: in the global stage, a larger inertia weight (0.9→0.4 linear attenuation) is used to quickly traverse the image and capture the candidate area where flora may be distributed; in the local stage, the neighborhood particle cooperation mechanism is introduced, the speed parameter is corrected by calculating the Euclidean distance between particles, and the detail recognition of small flora outlines (such as cocci with a diameter of less than 5 μm) is strengthened. Experimental data show that, compared with the standard particle swarm algorithm, the Dice coefficient of flora region segmentation is improved by 8.3%, the detection rate of low fluorescence intensity flora is improved by 12.7%, and the time-consuming for processing a single 4K resolution image is shortened to 0.8 seconds. The accurate flora spatial coordinates and morphological parameters output by the algorithm provide reliable data support for subsequent calculation of flora density gradient and analysis of spatial correlation with host cells, significantly improving the detection efficiency of abnormal flora aggregation patterns in clinical samples.

[0088] In the embodiments of the present application, the flora-environment interaction model is used to simulate the flora colonization coefficient and real-time correct the sensitivity parameter, including: constructing a flora-environment interaction dynamics model; inputting the frequency of gastric acid reflux, esophageal pH value fluctuation data and flora metabolite concentration data into the flora-environment interaction dynamics model, simulating the change curve of the flora colonization coefficient with environmental factors through differential equations; based on the change curve, real-time correct the sensitivity parameter to control the error between the predicted value and the clinically measured value.

[0089] Among them, the colonization coefficient of the flora is an index for quantifying the ability of the flora to adapt, reproduce and maintain stable existence in a specific host environment (such as esophageal mucosa, intestinal tissue, etc.), reflecting the interaction strength with environmental factors and host status.

[0090] It can be understood that the embodiments of the present application integrate the data of gastric acid reflux frequency, pH fluctuation and metabolite concentration through the flora-environment interaction dynamics model, dynamically present the law of change with the environment by using differential equations, provide quantitative basis for analyzing the interaction mechanism of flora and host microenvironment, correct the sensitivity parameters in real time through the Bayesian optimization algorithm, reduce the error between the predicted value and the clinically measured value, improve the prediction accuracy of the model for the colonization state of the flora, and enhance the scientificity and accuracy of medical decision-making.

[0091] It should be noted that the differential equation formula is:

[0092] Among them, is the rate of change of the colonization coefficient of the flora with time; is the intrinsic growth rate of the flora in an ideal environment; is the maximum carrying capacity of the environment for the colonization of the flora; is the colonization coefficient of the flora; is the logistic growth term; is the inhibition coefficient of gastric acid reflux on the colonization of the flora; is the frequency of gastric acid reflux; is the inhibition term of gastric acid reflux; is the inhibition coefficient of pH deviation on the colonization of the flora; is the real-time pH value in the esophagus; is the optimal pH value for the growth of the flora; is the inhibition term of pH deviation; is the promotion coefficient of low-concentration metabolites on the colonization of the flora; is the concentration of metabolites of the flora; is the promotion term of low-concentration metabolites; is the inhibition coefficient of high-concentration metabolites on the colonization of the flora; is the inhibition term of high-concentration metabolites.

[0093] Based on the curve of the colonization coefficient of the flora output by the flora-environment interaction dynamics model changing with environmental factors such as the frequency of gastric acid reflux, the pH fluctuation of the esophagus, and the concentration of flora metabolites (i.e., the dynamic trend of the colonization coefficient predicted by the model), the sensitivity parameters (such as the inhibition coefficient of gastric acid reflux on colonization and the inhibition coefficient of pH deviation from the optimal value) reflecting the influence of environmental factors on the colonization of flora in the model are continuously and dynamically adjusted by means of algorithms such as Bayesian optimization; through this real-time parameter correction, the deviation between the colonization coefficient of flora predicted by the model (the predicted value obtained from the change curve) and the colonization coefficient of flora actually detected in the clinic (the actual value) can be continuously reduced, making the simulation of the colonization state of flora in the host environment by the model more close to the real physiological scenario, and thus providing more accurate quantitative support for the risk assessment of flora-related diseases, clinical diagnosis and treatment decisions, etc.

[0094] For example, in the intelligent detection of the flora image of reflux esophagitis, the flora-environment interaction dynamics model is first used to integrate the patient's gastric acid reflux frequency (such as 3 times a day), the pH fluctuation range of the esophagus (2.1-6.8), and the concentration data of flora metabolites, and a differential equation is used to simulate the change curve of the colonization coefficient of flora in the lower esophagus (0.32-0.78) with acid exposure time; then, based on this curve, the image detection algorithm parameters are optimized to accurately identify the flora aggregation clusters (such as Streptococcus clumps with a diameter > 20 μm) in the fluorescence microscopic image that match the high-value area of the colonization coefficient, and to associate with the mucosal hyperemia and edema area, thereby realizing intelligent quantitative detection of the abnormal colonization state of flora in patients with reflux esophagitis and providing microbiological basis for inflammation severity assessment.

[0095] In the embodiments of the present application, a spatiotemporal graph convolution network is used to construct a flora dynamic evolution model to obtain the probability of Barrett's esophagus occurrence and the flora diffusion trend, including: constructing a flora dynamic evolution model; based on the flora dynamic evolution model, setting the dentate line of the lower esophagus and the precancerous lesion high-incidence area of the cardiac mucosal folds as the core nodes of the graph network, combining the micro features of the flora biofilm thickness, the apoptosis rate of epithelial cells, and the concentration of inflammatory factors, and modeling the spatiotemporal correlation between the flora invasion depth and the degree of epithelial dysplasia by the spatiotemporal graph convolution network to predict the probability of Barrett's esophagus occurrence and the corresponding flora diffusion trend.

[0096] The flora dynamic evolution model is a model that describes the dynamic process of the structure and quantity of flora changing with time, space, and environmental factors in the host environment, reveals the interaction rules of flora and host and environment, and predicts the evolution trend.

[0097] It can be understood that, by using the flora dynamic evolution model, the spatial features of key anatomical regions (such as the dentate line and the cardia mucosal folds) in the lower esophagus and the micro-dynamic information such as the thickness of the bacterial biofilm and the apoptosis rate of epithelial cells are integrated by the spatio-temporal graph convolution network, the spatio-temporal correlation between flora invasion and epithelial dysplasia is accurately modeled, the occurrence probability of Barrett's esophagus is quantitatively predicted, and the flora diffusion trend is clearly presented, which provides a visual dynamic basis for analyzing flora-driven precancerous lesions, and provides scientific support for early screening of high-risk populations and developing targeted intervention strategies, thereby improving the accuracy and forward-looking nature of Barrett's esophagus prevention and treatment.

[0098] In the embodiments of the present application, the trigger hierarchical diagnosis suggestion includes: setting a four-level diagnosis mechanism, when the occurrence probability of Barrett's esophagus is less than 20%, the flora is not abnormally diffused and the epithelial cells are normal, it is divided into a low-risk level one, and diet intervention and reexamination every 6 months are recommended; when the occurrence probability is 20%-50%, the flora is locally aggregated with mild epithelial hyperplasia, it is divided into a middle-low risk level two, and probiotic intervention and gastroscopy follow-up are recommended; when the occurrence probability is 50%-80%, the flora diffuses to the upper esophagus and is accompanied by moderate dysplasia, it is divided into a middle-high risk level three, and endoscopic mucosal resection evaluation and targeted antibacterial therapy are triggered; and when the occurrence probability is greater than 80%, the flora invades the whole layer and is accompanied by severe dysplasia or early cancer, it is divided into a high-risk level four, and multidisciplinary consultation and radical surgery evaluation are started.

[0099] It can be understood that, by using the four-level mechanism, the Barrett's esophagus risk is accurately stratified, and the differentiated intervention scheme is matched according to the occurrence probability, flora diffusion state and epithelial lesion degree: the low-risk level focuses on diet regulation and regular monitoring, the middle-low risk level strengthens probiotic intervention and gastroscopy follow-up, the middle-high risk level triggers endoscopic resection evaluation and targeted antibacterial therapy, and the high-risk level starts multidisciplinary consultation and radical surgery evaluation, thereby avoiding excessive medical treatment for low-risk persons and intervention delay for high-risk persons, improving the pertinence and effectiveness of intervention through ladder measures, standardizing the follow-up frequency and diagnosis and treatment path, helping to efficiently allocate medical resources in clinic, and reducing the risk of Barrett's esophagus progressing to esophageal cancer.

[0100] According to the esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease proposed in the embodiment of the present application, by integrating the esophageal mucosa flora microscopic image and the clinical multi-dimensional data, with the help of image registration algorithm, the spatial displacement characteristics of the flora are accurately quantified, the improved U-Net++ network is combined to realize efficient extraction of the micro features of the flora, and then through the spatio-temporal attention mechanism, the dynamic changes of the flora and the epithelial cell lesion characteristics are deeply fused, the flora-host correlation atlas generated breaks the separation state of the flora data and the mucosa damage information in the traditional detection, realizes the cross-scale correlation analysis from the micro flora behavior to the macro lesion, and significantly improves the capture ability of the early micro flora abnormalities and the occult mucosa damage. Relying on the flora dynamic evolution model constructed by the spatio-temporal graph convolution network, the occurrence probability of Barrett's esophagus and the flora diffusion trend can be accurately predicted, combined with the environmental data such as the frequency of gastric acid reflux of the patient and the pH value fluctuation of the esophagus, through the flora-environment interaction model, the detection sensitivity parameters are corrected in real time, effectively solving the problem of poor adaptability of the traditional method to individual differences and environmental fluctuations, making the risk assessment more in line with the actual pathological state of the patient, and greatly improving the accuracy and timeliness of the precancerous lesion warning. The final generated hierarchical diagnosis suggestion and risk heat map visualization report provide intuitive decision-making basis for the clinic from the flora pathogenic path to the cancer risk, not only shortening the detection cycle (more than 60% shorter than the traditional culture method), but also reducing the risk of over-treatment and missed diagnosis through accurate stratified diagnosis and treatment suggestions, and comprehensively improving the accuracy, clinical adaptability and diagnosis and treatment efficiency of the detection of reflux esophagitis and related precancerous lesions, providing new technical support for the early intervention and precise management of digestive tract diseases. As a result, the problems of long detection cycle and weak microanalysis ability in the prior art are solved.

[0101] A specific embodiment of an esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease will be described below, as shown in Figure 5 , including: A total of 120 patients with reflux esophagitis (72 males and 48 females, aged 23-72 years, mean 46.3±9.8 years) and 60 healthy controls (32 males and 28 females, aged 22-69 years, mean 43.5±10.1 years) were enrolled in the study. The informed consents were obtained from all the participants. The samples were collected by 2.8 mm sterile disposable biopsy forceps at 2-5 cm from the lower esophageal sphincter. Six to eight typical lesions (hyperemia and erosion) were selected from the patients and three to four normal mucosa were selected from the controls. The esophageal tissues were immediately placed in sterile cryopreservation tubes containing RNA protection solution, labeled and stored at -80℃. The temperature was continuously monitored during transportation (fluctuation ≤±2℃). After 24 h fixation with 4% paraformaldehyde, the samples were dehydrated with gradient ethanol (70%, 80%, 95% and 100%) for 15 min each, cleared with xylene for 10 min twice and embedded in paraffin to prepare 5 μm serial sections. The sections were deparaffinated, incubated with 3% H2O2 at room temperature for 10 min to inactivate endogenous peroxidase, and subjected to high-temperature antigen retrieval with 0.1 mol / L citric acid buffer (pH 6.0) for 15 min. The mixed solution containing EUB338-Cy3 (total bacteria), Lacto-FITC (Lactobacillus) and Strepto-Cy5 (Streptococcus) probes (10 ng / μL for each probe) was added, and the sections were hybridized in a wet box at 37℃ for 16 h. The sections were washed with 2×SSC solution at 37℃ for 10 min twice, stained with DAPI for 5 min and mounted with an anti-fluorescence quencher. The images were collected using a Zeiss LSM900 confocal microscope. The DAPI (nucleus, blue), FITC (Lactobacillus, green), Cy3 (total bacteria, red) and Cy5 (Streptococcus, magenta) were excited by 405 nm, 488 nm, 555 nm and 640 nm lasers, respectively. The images were collected at 40-fold oil immersion lens with a step of 0.3 μm Z-stack scanning. Fifteen to twenty non-overlapping fields were collected for each sample. The images were in 1024×1024 pixel TIFF format, and the acquisition parameters (laser intensity, gain, etc.) were attached during saving. The clinical data were extracted from the electronic medical record system, including basic information (age, gender, BMI, etc.), symptom data (frequency of heartburn, degree of acid reflux VAS score), 24 h esophageal pH monitoring results (Medtronic Digitrapper MKIV device, recording the number of reflux, the proportion of total time with pH<4, the longest reflux duration), Los Angeles classification (A-D grade) and previous medication history (PPI use duration, dose, etc.).

[0102] Systematic preprocessing of bacterial colony microscopic images: First, non-local mean filtering was used for denoising, with a search window of 7x7, a similar window of 3x3, and a filtering coefficient of 0.1, which reduced the loss of details by 30% compared to traditional Gaussian filtering. The contrast was enhanced by the CLAHE algorithm, with a size of 8x8 and a clipping threshold of 0.02, which increased the gray scale difference between the bacterial colonies and the background to 1.5 times that of the original image. Finally, the pixel values were linearly normalized to the [0, 1] interval to eliminate the brightness differences between devices. Clinical data preprocessing included: continuous variables (such as reflux frequency) were mapped to [0, 1] by Min-Max standardization, categorical variables (such as the Los Angeles classification) were converted into 4-dimensional vectors using one-hot encoding, missing values (accounting for <5%) were filled using the KNN algorithm (k=5) to ensure data integrity. Perception network was built based on PyTorch1.13 framework: the input layer used a double-branch structure to receive 512x512 resolution images (processed by downsampling) and 12-dimensional clinical feature vectors; the feature extraction layer included 4 residual blocks, each consisting of 2 3x3 convolution layers (step 2, padding=1), BatchNorm layers, and ReLU activation functions, which gradually compressed the image features to 32x32x512 dimensions; the fusion layer combined the image features and clinical features (mapped to 512 dimensions by a fully connected layer) into a 32x32x1024 tensor. Image registration used the SIFT+RANSAC algorithm: for sequence images of different biopsy sites from the same patient, first extract feature points using the SIFT algorithm (detection threshold 0.04, edge threshold 10), with an average of 300-500 stable feature points per image; then use the FLANN matcher (kd-tree index, matching distance ratio 0.7) for rough matching, remove false matching points (retention rate about 60%) by the RANSAC algorithm (iteration number 1000, threshold 1.5 pixels), calculate the affine transformation matrix, and output the translation amount (μm) and rotation angle (°) of the bacterial colonies in the x / y axis, quantifying the spatial displacement features. The improved U-Net++ network was optimized based on the original version: a coordinate attention module was added to the skip connection (by 1x1 convolution to compress the channels and generate x / y axis attention weights), enhancing the focus on dense bacterial colony areas; the deep network used a dilated convolution (rate 2), which expanded the receptive field to 64x64 pixels without reducing the resolution. During model training, the dataset was divided into training set (108 patients + 54 controls), validation set (13 patients + 6 controls), and test set (13 patients + 6 controls) at a ratio of 8:1:1, with Dice loss (weight 0.3) + cross-entropy loss (weight 0.7) as the optimization objective, Adam optimizer with an initial learning rate of 0.0001, decay of 50% every 20 rounds, and a Dice coefficient of 0.92±0.03 on the validation set after 150 rounds of training.Three types of features were extracted from the segmentation results: bacterial density (number of bacterial colonies per mm2, converted by pixel counting), aggregation degree (Moran's I index, range [-1, 1], > 0 indicates aggregated distribution), and fluorescence intensity gradient (Sobel operator calculates x / y direction gradient, takes the mean value to represent the rate of fluorescence change). Spatiotemporal attention mechanism integrates features: time dimension uses GRU network (hidden layer dimension 128) to model sequence features, generating time attention weights (highlighting time points with dramatic changes in bacterial dynamics); spatial dimension compresses features to a single channel through 1x1 convolution, and Sigmoid activation generates a spatial attention map (enhances epithelial lesion area); after weighted fusion, the output is a 256x256x3 bacterial-host interaction atlas (red channel: bacterial density, green channel: host cell lesion degree, blue channel: interaction intensity).

[0103] ST-GCN based on the graph of flora-host association: The nodes were defined as flora clusters (contiguous regions containing >5 bacteria) and epithelial cell clusters (cell aggregation areas with an area >100 μm²), and each node feature vector contained 16-dimensional parameters (flora characteristics: density, aggregation, etc. 8; host characteristics: cell morphological characteristics, nuclear-cytoplasmic ratio, etc. 8). The edge weight was determined by the Euclidean distance (<50 μm set to 1, otherwise 0) and the interaction strength (based on the spatial adjacency relationship between flora and host cells, >0.6 set to 1), and the adjacency matrix was constructed. The network contained 3 spatio-temporal blocks: each block was composed of a graph convolution layer (using ChebNet operator, K=2 order polynomial) and a 1D convolution layer (kernel size 3, step 1), which captured spatial dependence and temporal dynamics, respectively; the output layer predicted the Barrett's esophagus occurrence probability (0-1) and the flora diffusion vector (speed μm / h, direction angle °) through a fully connected layer. The model was trained with the Adam optimizer (learning rate 0.001), and the loss function was MSE (probability prediction) + cross-entropy (diffusion direction classification). After 80 rounds of training, the prediction accuracy on the test set was 89.6%, and the AUC was 0.91. The calculation steps of the flora imbalance quantification algorithm: ① Shannon diversity index (H=-Σpi×lnpi, pi is the proportion of the i-th flora); ② Dominant bacteria proportion (the relative abundance of the top 1 flora, >30% is defined as imbalance); ③ Host interaction deviation (Euclidean distance of the current sample and healthy group interaction strength); the three are weighted by 0.4:0.3:0.3 to get the imbalance index, which is divided into risk levels (<0.25 low risk, 0.25-0.65 medium risk, >0.65 high risk), combined with the Barrett's esophagus prediction probability (>50% when the level is increased by 1). The flora-environment interaction model is fitted by multiple linear regression: colonization coefficient = 0.32 × reflux frequency (times / 24h) + 0.28 × pH fluctuation amplitude (pH range in 24h) + 0.15 × Los Angeles grading score (A=1, B=2, C=3, D=4) -0.42 (R²=0.78, P<0.001), and the sensitivity (true positive rate) of the model increases by 5% for every 0.1 increase in the coefficient. The sensitivity parameter after modification is transmitted to the clinical diagnosis platform based on SpringBoot through RESTfulAPI, which includes the data layer (MySQL database stores images and clinical data), model layer (deploys ST-GCN algorithm), and application layer (visual interface). The multi-dimensional cancer risk deduction uses a random forest model (100 decision trees), with 15 input features including imbalance index, Barrett probability, colonization coefficient, etc., and outputs a risk value (0-100), triggering a grading recommendation: low risk (<30 points): 6-month follow-up, dietary intervention; medium risk (30-70 points): PPI standard dose treatment, 3-month review; high risk (>70 points): endoscopic radiofrequency ablation, 1-month review.The risk thermogram is drawn by Matplotlib, and the lower esophagus is divided into 6 regions (2-3 cm, 3-4 cm, 4-5 cm from the dentate line, anterior wall / posterior wall), and the risk values of each region are displayed with RGB pseudo-color (red = high risk, yellow = medium risk, green = low risk), and the report is added with the flora distribution thermogram, the dynamic evolution curve (comparison of the last 3 detections) and the personalized intervention suggestion (such as adjusting the PPI dose, the probiotic supplement scheme). After 200 independent sample verifications (150 patients + 50 controls), the risk classification and the pathological diagnosis coincidence rate is 87.5%, the thermogram AUROC is 0.89, which is significantly higher than that of the traditional pathological diagnosis (76.3%, P<0.01), and the average diagnosis time is shortened from 48 h to 2 h.

[0104] In summary, the present application guarantees data quality through a standardized sample collection and processing process, accurately extracts flora characteristics by combining advanced algorithms such as non-local mean filtering and improved U-Net++, efficiently predicts the Barrett esophagus occurrence probability and flora diffusion trend by means of the ST-GCN model, optimizes risk assessment through flora imbalance quantification and environmental interaction model, and finally generates precise grading suggestions and visual thermograms through a clinical diagnosis platform. Not only does it improve the risk classification and pathological diagnosis coincidence rate, which is better than traditional pathological diagnosis, but it also shortens the average diagnosis time, while providing personalized intervention programs, providing strong support for efficient and accurate diagnosis and clinical decision-making of reflux esophagitis and cancer risk.

[0105] Secondly, a kind of esophageal flora image intelligent detection system for gastroesophageal reflux disease risk prediction according to the embodiment of the present application is described with reference to the drawings.

[0106] Figure 6 It is a kind of esophageal flora image intelligent detection system for gastroesophageal reflux disease risk prediction according to the embodiment of the present application.

[0107] As Figure 6 shown, the esophageal flora image intelligent detection system 10 for gastroesophageal reflux disease risk prediction includes an acquisition module 100, an extraction module 200 and a generation module 300.

[0108] The acquisition module 100 is configured to acquire the microbial colony microscopic image and the clinical feature data of the esophageal mucosa sample; the extraction module 200 is configured to construct a perception network according to the microbial colony microscopic image and the clinical feature data, quantize the spatial displacement characteristics of the esophageal sample sequence microbial colony based on the perception network by using an image registration algorithm, extract the microbial colony density, aggregation degree and fluorescence intensity gradient characteristics by using an improved U-Net++ network to segment the image, and generate a microbial colony-host correlation graph by fusing the microbial colony dynamics and the epithelial cell lesion characteristics through a spatiotemporal attention mechanism; the generation module 300 is configured to construct a microbial colony dynamic evolution model by using a spatiotemporal graph convolution network according to the microbial colony-host correlation graph, predict the Barrett's esophagus occurrence probability and the microbial colony diffusion trend, calculate the lesion risk grade by using a microbial colony imbalance quantization algorithm based on the predicted Barrett's esophagus occurrence probability and the microbial colony diffusion trend, simulate the microbial colony colonization coefficient by using a microbial colony-environment interaction model in combination with the patient's gastric acid reflux frequency and the esophageal pH value fluctuation data, real-time correct the sensitivity parameter, transmit the corrected sensitivity parameter to a clinical diagnosis platform, perform multi-dimensional canceration risk deduction, trigger a hierarchical diagnosis suggestion, and generate a risk heat map visual report.

[0109] It should be noted that the foregoing explanation and description of the esophageal microbial colony image intelligent detection method for predicting the risk of gastroesophageal reflux disease also applies to the esophageal microbial colony image intelligent detection system for predicting the risk of gastroesophageal reflux disease, which will not be described here.

[0110] According to the esophageal flora image intelligent detection system for predicting the risk of gastroesophageal reflux disease provided by the embodiment of the application, by integrating the esophageal mucosa flora microscopic image and the clinical multi-dimensional data, with the help of the image registration algorithm, the spatial displacement characteristics of the flora are accurately quantified, the improved U-Net++ network is combined to realize the efficient extraction of the microcosmic characteristics of the flora, and then the spatiotemporal attention mechanism is used to deeply integrate the dynamic changes of the flora and the epithelial cell lesion characteristics, so that the generated flora-host correlation atlas breaks the split state of the flora data and the mucosa damage information in the traditional detection, realizes the cross-scale correlation analysis from the microcosmic flora behavior to the macroscopic lesion, and significantly improves the capture ability of the early microcosmic flora abnormalities and the occult mucosa damage. Relying on the flora dynamic evolution model constructed by the spatiotemporal graph convolution network, the occurrence probability of Barrett esophagus and the flora diffusion trend can be accurately predicted, the flora-environment interaction model is combined with the environmental data such as the frequency of gastric acid reflux and the pH value fluctuation of the esophagus, the detection sensitivity parameters are corrected in real time, the problems of poor adaptability to individual differences and environmental fluctuations in the traditional method are effectively solved, the risk assessment is more in line with the actual pathological state of the patient, and the accuracy and timeliness of the precancerous lesion warning are greatly improved. The finally generated grading diagnosis suggestion and risk heat map visualization report provide intuitive decision-making basis for the clinic from the flora pathogenic path to the cancer risk, not only shorten the detection period (more than 60% shorter than the traditional culture method), but also reduce the risk of over-treatment and missed diagnosis through accurate stratified diagnosis and treatment suggestions, comprehensively improve the accuracy, clinical adaptability and diagnosis and treatment efficiency of the detection of reflux esophagitis and related precancerous lesions, and provide new technical support for the early intervention and precise management of digestive tract diseases. Thus, the problems of long detection period and weak microcosmic analysis capability in the prior art are solved.

[0111] Figure 7 The structure schematic diagram of the electronic device provided by the embodiment of the application is provided. The electronic device can include: The memory 701, the processor 702, and the computer program stored in the memory 701 and executable on the processor 702.

[0112] The processor 702 implements the esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease provided in the above embodiment when executing the program.

[0113] Further, the electronic device further includes: The communication interface 703 is used for communication between the memory 701 and the processor 702.

[0114] The memory 701 is used to store the computer program executable on the processor 702.

[0115] The memory 701 can include a high-speed RAM memory, and can also include a non-volatile memory, such as at least one disk memory.

[0116] If the memory 701, the processor 702 and the communication interface 703 are implemented independently, the communication interface 703, the memory 701 and the processor 702 can be connected to each other through a bus and complete communication between each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 7 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0117] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 703 are integrated on a chip, the memory 701, the processor 702 and the communication interface 703 can complete communication between each other through an internal interface.

[0118] The processor 702 can be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0119] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0120] Furthermore, the terms "first", "second", etc. are used herein only to describe different instances, and do not imply or suggest relative importance or a number of the indicated technical features. Thus, the features defined with "first", "second", etc. can include at least one of the features explicitly or implicitly. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly and specifically limited.

[0121] Any process or method descriptions or blocks in flow charts described herein and elsewhere can be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process. Alternate implementations are included within the scope of the preferred embodiments of the present application in which additional functionality can be added or some functionality can be removed, by, for example, adding one or more steps performing a similar or reciprocal function, combining two or more steps, omitting one or more steps, etc.

[0122] It should be understood that portions of the present application can be implemented in hardware, software, firmware, or combinations thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. As with the hardware implementation, if implemented in software or firmware, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Also, functional computer- readable media that carry instructions or data to perform tasks can be accessed by a computer. For example, wireless media such as acoustic, RF, infrared and other wireless media can all be used to carry such computer-readable instructions or data structures. Computer-readable media also includes communication media including any medium that facilitates transfer of a computer program from one place to another.

[0123] While embodiments of the present application have been shown and described above, it is to be understood that the above-described embodiments are merely illustrative of and not restrictive on the present application, and changes, modifications, substitutions and variations can be made by those of ordinary skill in the art within the scope of the present application.

Claims

1. An esophageal flora image intelligent detection method for predicting the risk of gastroesophageal reflux disease, characterized in that, The method comprises the following steps: obtaining a microbial colony microscopic image and clinical feature data of an esophageal mucosa sample; constructing a perception network according to the microbial colony microscopic image and the clinical feature data, quantifying the spatial displacement characteristics of the esophageal sample sequence microbial colony by using an image registration algorithm, extracting the microbial colony density, aggregation degree and fluorescence intensity gradient characteristics by using an improved U-Net++ network segmentation image, fusing the microbial colony dynamics and the epithelial cell lesion characteristics through a spatio-temporal attention mechanism, and generating a microbial colony-host correlation atlas; constructing a microbial colony dynamic evolution model by using a spatio-temporal graph convolution network according to the microbial colony-host correlation atlas, predicting the Barrett's esophagus occurrence probability and the microbial colony diffusion trend, calculating the lesion risk grade by using a microbial colony imbalance quantization algorithm based on the predicted Barrett's esophagus occurrence probability and the microbial colony diffusion trend, combining the patient's gastric acid reflux frequency and the esophageal pH value fluctuation data, simulating the microbial colony colonization coefficient by using a microbial colony-environment interaction model, real-time correcting the sensitivity parameter, transmitting the corrected sensitivity parameter to a clinical diagnosis platform, performing multi-dimensional cancer risk deduction, triggering a hierarchical diagnosis suggestion and generating a risk heat map visualization report.

2. The esophageal microbiota image intelligent detection method for predicting the risk of gastroesophageal reflux disease according to claim 1, characterized in that, The method for generating the risk heat map visualization report comprises the following steps: constructing a microbial colony-cancer collaborative diffusion model; based on the microbial colony-cancer collaborative diffusion model, combining an improved particle swarm algorithm to simulate the invasion trajectory of pathogenic bacteria in the esophageal mucosa layer, mapping the spatial correlation between the microbial colony aggregation area and the epithelial cell dysplasia area into a cancer probability value by using a Gaussian mixture model, and generating a cancer risk probability cloud map; superimposing the cancer risk probability cloud map on an esophageal three-dimensional anatomical model, representing the cancer risk intensity through a color gradient, and generating a heat map visualization report labeled with the key path node of the pathogenic microbial colony, the predicted cancer progression time and the high-risk area coordinates. 3.The intelligent detection method of esophageal flora image for predicting the risk of gastroesophageal reflux disease according to claim 1, characterized in that, The spatio-temporal attention mechanism formula: ; wherein, is the attention output result; is the microbiota dynamic feature matrix; is the epithelial lesion feature matrix; is the query vector generated from M; is the key vector generated from E; is the key vector is the transpose matrix of the key vector; is the key vector dimension; is the value vector; is the scaled microbiota-host feature association score. 4.The intelligent detection method of esophageal microbiota image for predicting the risk of gastroesophageal reflux disease according to claim 1, characterized in that, simulating the microbial colony colonization coefficient by using a microbial colony-environment interaction model, and real-time correcting the sensitivity parameter, comprises the following steps: constructing a microbial colony-environment interaction dynamics model; inputting the gastric acid reflux frequency, the esophageal pH value fluctuation data and the microbial colony metabolite concentration data into the microbial colony-environment interaction dynamics model, simulating the change curve of the microbial colony colonization coefficient with the environmental factors by using a differential equation; based on the change curve, real-time correcting the sensitivity parameter to control the error between the predicted value and the clinically measured value.

5. The esophageal microbiota image intelligent detection method for predicting the risk of gastroesophageal reflux disease according to claim 1, characterized in that, constructing a microbial colony dynamic evolution model by using a spatio-temporal graph convolution network to obtain the Barrett's esophagus occurrence probability and the microbial colony diffusion trend, comprises the following steps: constructing a microbial colony dynamic evolution model; based on the microbial colony dynamic evolution model, setting the lower esophageal dentate line and the cardia mucosal fold precancerous lesion high-risk area as the core nodes of the graph network, combining the microbial biofilm thickness, the epithelial cell apoptosis rate and the inflammatory factor concentration micro features, modeling the spatio-temporal correlation between the microbial colony invasion depth and the epithelial dysplasia degree by using a spatio-temporal graph convolution network, predicting the Barrett's esophagus occurrence probability and the corresponding microbial colony diffusion trend. 6.The intelligent detection method of esophageal microbiota image for predicting the risk of gastroesophageal reflux disease according to claim 1, wherein, constructing a perception network, and quantifying the spatial displacement characteristics of the esophageal sample sequence microbial colony by using an image registration algorithm based on the perception network, comprises the following steps: constructing a multi-modal perception network; According to the multi-modal perception network, the texture features of the microbial colony microscopic image and the structured information of the clinical feature data are fused to generate a dynamic weight matrix; Based on the dynamic weight matrix, an image registration algorithm is used to calculate the three-dimensional spatial displacement vector of the microbial colony in the esophageal sample sequence, and the migration distance and direction angle parameters of the microbial colony are quantified.

7. The esophageal microbiota image intelligent detection method for predicting the risk of gastroesophageal reflux disease according to claim 1, characterized in that, An improved U-Net++ network is used to segment the image and extract the microbial colony density, aggregation degree and fluorescence intensity gradient features, including: An improved U-Net++ network is constructed; Based on the improved U-Net++ network, the microbial colony microscopic image is segmented at the pixel level to generate a microbial colony region mask; According to the microbial colony region mask, a density clustering algorithm is used to calculate the microbial colony density index, a morphological operator is used to extract the aggregation degree feature, and a Sobel operator is used to calculate the gradient amplitude and direction features of the fluorescence intensity. 8.The intelligent detection method of esophageal microbiota image for predicting the risk of gastroesophageal reflux disease according to claim 1, wherein, The improved U-Net++ network formula is: ; wherein, is the image segmentation accuracy at time t; is the base segmentation accuracy; is the fluorescence contrast of the bacterial population at time t; is the standard contrast threshold; is the bacterial population region weight coefficient; is the noise influence factor; is the image noise level at time t. 9.The intelligent detection method of esophageal microbiota image for predicting the risk of gastroesophageal reflux disease according to claim 1, wherein, Triggering a hierarchical diagnosis suggestion, including: Setting a four-level diagnosis mechanism, when the Barrett's esophagus occurrence probability is less than 20%, the microbial colony has no abnormal diffusion and the epithelial cells are normal, it is divided into a first-level low risk, and diet intervention and review every 6 months are recommended; when the occurrence probability is 20%-50%, the microbial colony is locally aggregated with mild epithelial hyperplasia, it is divided into a second-level medium-low risk, and probiotic intervention and gastroscopy follow-up are recommended; when the occurrence probability is 50%-80%, the microbial colony spreads to the upper esophagus and is accompanied by moderate dysplasia, it is divided into a third-level medium-high risk, and endoscopic mucosal resection evaluation and targeted antibacterial therapy are triggered; when the occurrence probability is greater than 80%, the microbial colony invades the whole layer and is accompanied by severe dysplasia or early cancer, it is divided into a fourth-level high risk, and multidisciplinary consultation and radical surgery evaluation are started.

10. An esophageal microbiota image intelligent detection system for predicting the risk of gastroesophageal reflux disease, characterized in that, Including: An acquisition module is configured to acquire a microbial colony microscopic image of an esophageal mucosa sample and clinical feature data; An extraction module is configured to construct a perception network based on the microbial colony microscopic image and the clinical feature data, use an image registration algorithm to quantify the spatial displacement features of the microbial colony in the esophageal sample sequence, use an improved U-Net++ network to segment the image and extract the microbial colony density, aggregation degree and fluorescence intensity gradient features, and use a spatio-temporal attention mechanism to fuse the microbial colony dynamics and epithelial lesion features to generate a microbial-host correlation graph; A generation module is configured to use a spatio-temporal graph convolution network to construct a microbial colony dynamic evolution model based on the microbial-host correlation graph, predict the Barrett's esophagus occurrence probability and the microbial colony diffusion trend, calculate the lesion risk level by a microbial imbalance quantification algorithm based on the predicted Barrett's esophagus occurrence probability and the microbial colony diffusion trend, use a microbial-environment interaction model to simulate the microbial colonization coefficient by combining the patient's gastric acid reflux frequency and esophageal pH value fluctuation data, real-time correct the sensitivity parameters, transmit the corrected sensitivity parameters to a clinical diagnosis platform, perform multidimensional cancer risk deduction, trigger a hierarchical diagnosis suggestion and generate a visual report of a risk heat map.

Images