AI-based endoscopic polyp identification method

Abstract

The invention discloses an AI-based endoscopic polyp identification method, which relates to the technical field of endoscopic polyp identification, is used for solving the problem that the imaging quality fluctuation of an endoscopic image is increased, and comprises the following steps of: acquiring a detection image through polyp detection equipment, processing the image to generate a feature image, detecting the white balance offset and the pixel displacement of the feature image, and identifying the polyp according to the white balance offset and the pixel displacement. Calculating an image quality score by using the two features, performing mucous membrane texture line detection on the feature images to generate texture distribution features, calculating a texture quality index, analyzing an imaging evaluation coefficient in combination with the image quality score, classifying the feature images, calling polyp distribution data of a spatial index database, and generating distribution coverage features according to a classification result; according to the method, the distribution coverage features and the polyp distribution data are integrated to evaluate an anomaly identification ratio, whether an anomaly detection prompt is generated or not is judged accordingly, and the integrity, accuracy and inspection efficiency of polyp identification are improved and omission and misjudgment are reduced by comprehensively evaluating the image quality and the mucous membrane texture features.

Description

Technical Field

[0001] This invention relates to the field of endoscopic polyp identification technology, and more specifically, to an AI-based endoscopic polyp identification method. Background Technology

[0002] With the rapid development of medical imaging technology and artificial intelligence technology, endoscopy has been widely used in clinical diagnosis. Endoscopic equipment uses image acquisition modules to image the internal cavities of the human body in real time. Doctors observe and judge the lesions based on the endoscopic images, which is especially important in the detection of polyps in the digestive system. Traditional polyp identification mainly relies on manual observation and judgment, which has a certain degree of clinical accuracy.

[0003] The existing technology has the following shortcomings: Currently, existing technologies are greatly affected by factors such as lighting conditions, shooting angle, and tissue reflectivity in endoscopic images. These factors can easily lead to phenomena such as white balance shift, uneven brightness, or local blurring in the images, making it impossible to dynamically quantify and evaluate the image quality. This results in increased fluctuations in the imaging quality of endoscopic images, and the texture difference between polyps and the mucosal background is small. Therefore, an AI-based endoscopic polyp identification method is proposed.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AI-based endoscopic polyp identification method, which solves the problems mentioned in the background art by employing an AI intelligent identification model that fuses image quality quantitative assessment with mucosal texture features.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-based method for endoscopic polyp identification, comprising the following steps: Step S1: When the polyp detection device performs polyp identification, it acquires a detection image, performs image processing on the detection image to generate a feature image, and detects the white balance shift and pixel displacement of the feature image; Step S2: Calculate the image quality score of the feature image using white balance offset and pixel displacement, perform mucosal texture line detection on the feature image and generate texture distribution features; Step S3: Calculate the texture quality index based on the texture distribution characteristics, analyze the imaging evaluation coefficient in combination with the image quality score, classify the feature image based on the imaging evaluation coefficient, and retrieve the polyp distribution data from the spatial index library; Step S4: The classification results of the feature image generate the distribution coverage features of the current polyp identification. The distribution coverage features and polyp distribution data are combined to evaluate the proportion of abnormal identification. Based on the proportion of abnormal identification, it is determined whether to generate an abnormal detection prompt.

[0007] In a preferred embodiment, in step S1, a preset acquisition period is set, and the detection image of the detection area is acquired by the endoscopic image acquisition unit in the polyp detection device; The feature image is obtained by performing image processing on the detection image of the detection area through color edge enhancement and brightness normalization. The average values ​​of the red, green, and blue color channels of each pixel in the feature image are calculated separately. The red channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the red channel of each pixel. The blue channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the blue channel of each pixel.

[0008] In a preferred embodiment, in step S1, the white balance offset of the feature image is calculated by taking the square root of the sum of the squares of the red channel offset and the blue channel offset of the feature image. The feature value of the corner is obtained by acquiring the numerical value of the local brightness change of the pixel in the feature image through FAST corner detection; Pixels with obvious features are obtained based on the feature value judgment of corner points; Two-dimensional coordinates are established by using the preset fixed vertex of the feature image as the origin of the two-dimensional coordinates and the pixel interval of the feature image as the scale of the horizontal and vertical axes of the two-dimensional coordinates. Select the pixel with obvious features in the feature image of adjacent acquisition time and obtain the displacement of the pixel by Euclidean distance calculation method; The pixel displacement of the feature image is obtained by averaging the displacement of all pixels with obvious features.

[0009] In a preferred embodiment, in step S2, the white balance offset and pixel displacement of the feature image are standardized to obtain the offset factor and displacement factor. The image quality score of the feature image is obtained by weighted summation of the offset factor and the displacement factor; Edge response calculation is performed on the feature image using the multi-directional Sobel operator to extract gradient magnitudes in the horizontal, vertical, and diagonal directions; The gradient magnitudes in each direction of the feature image are weighted and superimposed to obtain a comprehensive texture response map.

[0010] In a preferred embodiment, non-maximum suppression processing is performed on the composite texture response map to remove redundant edge lines while retaining texture lines corresponding to local gradient extrema. The composite texture response map after non-maximum suppression is converted to grayscale to obtain a grayscale image; Calculate the texture energy of a grayscale image based on the gray-level co-occurrence matrix; The texture energy distribution of the grayscale image is formed into a two-dimensional matrix and then normalized to generate the texture distribution features of the feature image.

[0011] In a preferred embodiment, in step S3, the average value and standard deviation of the texture distribution features of the feature image are added together to obtain the texture quality index of the feature image; The texture quality index and image quality score of the feature image are processed by the Max-Min normalization method to obtain the texture factor and image factor. The imaging evaluation coefficient is calculated by combining texture factors and image factors. The calculation formula is as follows: ,in, For texture factor, For image factors, This is the imaging evaluation coefficient.

[0012] In a preferred embodiment, in step S3, if the imaging evaluation coefficient is greater than a preset imaging threshold, the feature image is determined to be a high-quality image. If the imaging evaluation coefficient is less than or equal to the preset imaging threshold, the feature image is determined to be a low-quality image. Obtain polyp distribution data, including polyp distribution density, from the spatial index.

[0013] In a preferred embodiment, in step S4, the detection area is divided into several grid cells of equal size, and each grid cell corresponds to a spatial location; Traverse all feature images within the same detection region, and map the spatial location covered by each image to the corresponding grid cell based on the classification result of each image; If a grid cell has high-quality image coverage, then mark the cell as high-quality coverage in the coverage feature; If a grid cell is covered only by low-quality images, then the cell is marked as low-quality coverage in the coverage feature to fill in the blank areas; After all feature images have been traversed, the labeling information is merged to generate the distribution coverage matrix of the entire detection area, forming the current distribution coverage feature for polyp identification.

[0014] In a preferred embodiment, in step S4, the distribution coverage features are mapped to polyp distribution data in the spatial index library to form an input feature matrix; The input feature matrix is ​​fed into a pre-trained convolutional neural network model to obtain the anomaly detection rate. If the proportion of anomaly detections is greater than or equal to a preset proportion threshold, an anomaly detection prompt will be generated. If the percentage of anomaly detections is less than the preset percentage threshold, then no anomaly detection prompt will be generated.

[0015] The technical effects and advantages of this invention are as follows: This invention acquires detection images during polyp identification using a polyp detection device, processes these images to generate feature images, and simultaneously detects the white balance shift and pixel displacement of the feature images. The white balance shift and pixel displacement are used to calculate the image quality score of the feature images, and mucosal texture lines are detected to generate texture distribution features. A texture quality index is calculated based on the texture distribution features, and an imaging evaluation coefficient is analyzed in conjunction with the image quality score. The feature images are then classified based on the imaging evaluation coefficient, and polyp distribution data from a spatial index database is retrieved. The classification results of the feature images are used to generate the current polyp identification distribution coverage features. The proportion of abnormal identifications is evaluated by combining the distribution coverage features and polyp distribution data, and an abnormal detection prompt is generated based on this proportion. By comprehensively evaluating image quality and mucosal texture features, the classification and coverage information of high- and low-quality images are integrated. Through intelligent evaluation of the abnormal identification proportion, timely abnormal detection prompts are generated, ensuring the completeness and accuracy of polyp identification, improving the efficiency and reliability of endoscopic examinations, and reducing the risk of omissions and misjudgments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of an AI-based endoscopic polyp identification method according to the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the steps of an AI-based endoscopic polyp identification method according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention acquires detection images during polyp identification using a polyp detection device, processes these images to generate feature images, and simultaneously detects the white balance shift and pixel displacement of the feature images. The white balance shift and pixel displacement are used to calculate the image quality score of the feature images, and mucosal texture lines are detected to generate texture distribution features. A texture quality index is calculated based on the texture distribution features, and an imaging evaluation coefficient is analyzed in conjunction with the image quality score. The feature images are then classified based on the imaging evaluation coefficient, and polyp distribution data from a spatial index database is retrieved. The classification results of the feature images are used to generate the current polyp identification distribution coverage features. The proportion of abnormal identifications is evaluated by combining the distribution coverage features and polyp distribution data, and an abnormal detection prompt is generated based on this proportion. By comprehensively evaluating image quality and mucosal texture features, the classification and coverage information of high- and low-quality images are integrated. Through intelligent evaluation of the abnormal identification proportion, timely abnormal detection prompts are generated, ensuring the completeness and accuracy of polyp identification and improving the efficiency and reliability of endoscopic examinations.

[0020] Example 1: An AI-based method for endoscopic polyp identification, such as... Figures 1 to 2 As shown, it includes the following steps: Step S1: When the polyp detection device performs polyp identification, it acquires a detection image, performs image processing on the detection image to generate a feature image, and detects the white balance shift and pixel displacement of the feature image; Step S2: Calculate the image quality score of the feature image using white balance offset and pixel displacement, perform mucosal texture line detection on the feature image and generate texture distribution features; Step S3: Calculate the texture quality index based on the texture distribution characteristics, analyze the imaging evaluation coefficient in combination with the image quality score, classify the feature image based on the imaging evaluation coefficient, and retrieve the polyp distribution data from the spatial index library; Step S4: The classification results of the feature image generate the distribution coverage features of the current polyp identification. The distribution coverage features and polyp distribution data are combined to evaluate the proportion of abnormal identification. Based on the proportion of abnormal identification, it is determined whether to generate an abnormal detection prompt.

[0021] The specific implementation is as follows: In step S1, a preset acquisition period is set, and the detection image of the detection area is acquired through the endoscopic image acquisition unit in the polyp detection device; The feature image is obtained by performing image processing on the detection image of the detection area through color edge enhancement and brightness normalization. It should be noted that the preset acquisition cycle is a fixed time or frame interval for image acquisition, set according to the endoscope's movement speed. If the endoscope moves quickly, the acquisition cycle should be shortened to ensure that polyp areas are not missed; if the endoscope moves slowly, the acquisition cycle should be extended to reduce redundant data. The endoscope image acquisition unit in the polyp detection device is a device that acquires high-resolution color images in real time to obtain detection images. Color edge enhancement extracts edge information from each color channel of the color image using the Canny operator. Brightness normalization scales the brightness values ​​of image pixels to a fixed range using linear or non-linear methods to ensure uniform brightness levels across different images, and is used to process the detection images to obtain feature images.

[0022] Each pixel in the feature image contains three color channel values: red, green, and blue. The average value of each pixel in the feature image is calculated for the three color channel values: red, green, and blue. The red channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the red channel of each pixel. The blue channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the blue channel of each pixel. The white balance offset of the feature image is calculated by taking the square root of the sum of the squares of the red channel offset and the blue channel offset of the feature image. The feature values ​​of corner points in the feature image are obtained through FAST corner detection; A corner point is a point in an image where the local brightness changes. The characteristic value of a corner point is the numerical value of the degree of local brightness change of that pixel. The feature values ​​of corner points in the feature image are compared with a preset feature threshold for judgment. If the feature value of a corner point in the feature image is greater than the preset feature threshold, then the pixel point is determined to be a pixel point with obvious features. Conversely, if the pixel does not have obvious features, it is determined that the pixel is not a pixel with obvious features. It should be noted that the preset feature threshold is an important parameter for determining whether a pixel in a feature image is a pixel with obvious features. The feature values ​​of pixels in historical feature images are statistically analyzed, and the mean and standard deviation are calculated. The sum of the mean and standard deviation is used as the preset feature threshold.

[0023] Two-dimensional coordinates are established by using the preset fixed vertex of the feature image as the origin of the two-dimensional coordinates and the pixel interval of the feature image as the scale of the horizontal and vertical axes of the two-dimensional coordinates. Select the pixel with obvious features in the feature image of adjacent acquisition time and obtain the displacement of the pixel by Euclidean distance calculation method; The pixel displacement of the feature image is obtained by averaging the displacement of all pixels with obvious features.

[0024] It needs to be explained that FAST corner detection is an algorithm that quickly detects corners or feature points in an image. These points are usually locations in the image that have significant changes and rich structures, and are used to obtain the feature values ​​of corners in the feature image. The preset fixed vertex angle refers to a corner position that is manually or systematically specified in advance on the feature image and used as the origin of the two-dimensional coordinate system. One of the four vertices of the feature image can be selected as the preset fixed vertex angle. The Euclidean distance calculation method refers to calculating the straight-line distance between two points in two-dimensional or multi-dimensional space, and is used to obtain the displacement of a pixel.

[0025] In step S2, the white balance offset and pixel displacement of the feature image are standardized to obtain the offset factor and displacement factor. The image quality score of the feature image is calculated by combining the offset factor and the displacement factor. The calculation formula is as follows: ,in, As the offset factor, It is the displacement factor. and To preset the weighting coefficients, Score the image quality of the feature image; It should be noted that the standardization processing methods include, but are not limited to, standard linear transformation based on interval scaling, statistical Z-Score standardization, or normalization based on nonlinear mapping functions. The specific methods of standardization processing will not be elaborated upon here. A larger offset factor and a larger displacement factor result in greater image color deviation, more unstable motion, and a higher image quality score for the feature image. Conversely, a smaller offset factor and a smaller displacement factor result in less image color deviation, more stable motion, and a lower image quality score for the feature image. The preset weighting coefficient measures the influence of the offset factor and displacement factor on the image quality score of the feature image. Multiple sets of polyp detection images can be collected, the corresponding offset factor and displacement factor can be calculated, and the contribution ratio of the offset factor and displacement factor to the image quality score can be analyzed using manually evaluated image quality scores to determine the preset weighting coefficient.

[0026] Edge response calculation is performed on the feature image using the multi-directional Sobel operator to extract gradient magnitudes in the horizontal, vertical, and diagonal directions; The gradient magnitudes in each direction of the feature image are weighted and superimposed to obtain a comprehensive texture response map; It should be noted that the gradient magnitude of each direction in the feature image is calculated using each pixel in the image as the unit. The gradient magnitude of the pixel in different directions is used as the new pixel information to replace it, thereby forming a multi-directional gradient magnitude map. At the same pixel position, the gradient magnitudes in each direction are weighted and summed according to preset weights to obtain a new pixel response value. After the gradient magnitudes of all pixel positions in the feature image are replaced with the corresponding new pixel response values, the resulting two-dimensional response matrix constitutes a new image in space, namely the comprehensive texture response map.

[0027] Non-maximum suppression is performed on the composite texture response map to remove redundant edge lines, while retaining the texture lines corresponding to local gradient extrema. The composite texture response map after non-maximum suppression is converted to grayscale to obtain a grayscale image; Calculate the texture energy of a grayscale image based on the gray-level co-occurrence matrix; The texture energy distribution of the grayscale image is formed into a two-dimensional matrix and then normalized to generate the texture distribution features of the feature image.

[0028] It needs to be explained that the Sobel operator is an image edge detection operator that calculates the gradient values ​​in each direction of an image using a convolution kernel. It is used to calculate the edge response of feature images, extracting the gradient magnitudes in the horizontal, vertical, and diagonal directions. Weighted stacking refers to multiplying each pixel value of multiple images or feature matrices by their corresponding weights and then summing them to obtain the final pixel value. Historical data analysis is used to determine which directions are more sensitive to polyp or texture detection, assigning them higher weights. Cross-validation is then used to adjust the weighted stacking result to best match human evaluation or a standard reference. Non-maximum suppression is the operation of preserving local maxima and suppressing non-maximums, used to remove redundant edge lines and retain texture lines corresponding to local gradient extrema. Grayscale conversion is the operation in image processing that converts a color image to a grayscale image. The grayscale co-occurrence matrix is ​​a square matrix where each element... Indicates the grayscale value in the image The pixel and grayscale values The number of times or probability that a pixel appears in a specific spatial relationship is used as the texture energy of that pixel in the feature image.

[0029] In step S3, the average value and standard deviation of the texture energy of each pixel in the texture distribution features of the feature image are calculated to obtain the average value and standard deviation of the texture distribution features of the feature image. The texture quality index of the feature image is obtained by adding the mean and standard deviation of the texture distribution features of the feature image. Texture factor and image factor are obtained by processing the texture quality index and image quality score of the feature image using the Max-Min normalization method. The calculation formula is as follows: , ,in, and These represent the texture quality index and image quality score of the feature image, respectively. and These are the maximum and minimum values ​​of the texture quality index of the feature image, respectively. and These represent the maximum and minimum image quality scores of the feature images, respectively. and These are texture factor and image factor, respectively; The imaging evaluation coefficient is calculated by combining texture factors and image factors. The calculation formula is as follows: ,in, For texture factor, For image factors, For imaging evaluation coefficients; It should be noted that the larger the texture factor and the smaller the image factor, the clearer the texture features, the smaller the image deviation, and the larger the imaging evaluation coefficient; conversely, the smaller the texture factor and the larger the image factor, the blurrier the texture features, the greater the image deviation, and the smaller the imaging evaluation coefficient.

[0030] The imaging evaluation coefficient is compared with the preset imaging threshold for judgment. If the imaging evaluation coefficient is greater than the preset imaging threshold, the feature image is determined to be a high-quality image. If the imaging evaluation coefficient is less than or equal to the preset imaging threshold, the feature image is determined to be a low-quality image. Retrieve polyp distribution data from the spatial index library; Polyp distribution data refers to a structured dataset that records the location, quantity density, and morphological characteristics of polyps in a spatial region during medical imaging or endoscopic examination, including polyp distribution density.

[0031] It should be explained that the preset imaging threshold is an important parameter for determining whether a feature image is a high-quality or low-quality image. A large number of polyp detection images are collected, and the imaging evaluation coefficient of each image is calculated. Based on manual or expert annotation, the images are divided into two categories: high-quality and low-quality. The mean of the imaging evaluation coefficient of high-quality images is calculated, and this mean is selected as the initial threshold. The initial threshold is applied to the classification of new images to evaluate the accuracy of high-quality and low-quality judgment. Based on the classification results, the threshold is adjusted through cross-validation to maximize the recognition rate of high-quality images and minimize the false recognition rate of low-quality images. The spatial index library is a database for storing and quickly accessing polyp distribution information, used to obtain polyp distribution data.

[0032] In step S4, the detection area is divided into several grid cells of equal size, and each grid cell corresponds to a spatial location; Traverse all feature images within the same detection region, and map the spatial location covered by each image to the corresponding grid cell based on the classification result of each image; If a grid cell has high-quality image coverage, then mark the cell as high-quality coverage in the coverage feature; If a grid cell is covered only by low-quality images, then the cell is marked as low-quality coverage in the coverage feature to fill in the blank areas; After all feature images have been traversed, the labeling information is merged to generate the distribution coverage matrix of the entire detection area, forming the distribution coverage features for the current polyp identification. The distribution coverage features are mapped to polyp distribution data in the spatial index library to form the input feature matrix; The input feature matrix is ​​fed into a pre-trained convolutional neural network model to obtain the anomaly detection rate. Convolutional neural network models consist of multiple convolutional layers, pooling layers, and fully connected layers, and can automatically extract spatial and local pattern information of covering features; The model training process includes the following steps: Data preparation: Collect a large number of historical feature images and corresponding polyp spatial distribution data, and match the distribution coverage features of each image with the actual polyp distribution density to form input label pairs; Data preprocessing: The input feature matrix is ​​standardized to normalize the image coverage features and polyp distribution density to the same scale; Training strategy: Supervised learning method is used, with the input feature matrix as the model input, the polyp distribution density label as the label, and the mean squared error loss function used to measure the difference between the model prediction and the true distribution. Model optimization: The weight parameters of the convolutional neural network are updated through backpropagation to continuously reduce the loss function value; a learning rate decay strategy is adopted during training to improve training effect and generalization ability. Model validation: Evaluate model performance on an independent validation dataset to ensure that the model can accurately predict the proportion of anomaly identifications; Model update: Retraining or fine-tuning by continuously adding historical detection data to optimize the ability to identify low-quality coverage areas and polyp distribution deviations; The judgment is made by comparing the percentage of anomalies with a preset percentage threshold. If the proportion of anomaly detections is greater than or equal to a preset proportion threshold, an anomaly detection prompt will be generated. If the percentage of anomaly detections is less than the preset percentage threshold, then no anomaly detection prompt will be generated.

[0033] It's important to explain that supervised learning is a machine learning method used to train a model using existing input-output samples, enabling the model to predict or classify new data. Mean squared error (MSE) is a commonly used error metric used to assess the difference between predicted and true values. Backpropagation is the most frequently used optimization algorithm for training artificial neural networks, used to calculate the gradient of the loss function relative to the network weights and update the weights to reduce prediction error. Learning rate decay is a common optimization method for training neural networks, used to dynamically adjust the learning rate during training to improve convergence stability and final model accuracy. The preset percentage threshold is a crucial parameter for determining whether to generate anomaly detection alerts. This involves collecting a large amount of historical endoscopic examination data, statistically analyzing the percentage of anomalies identified in each detection region, and classifying the percentages requiring anomaly detection alerts based on manual or expert annotations. The median of these annotated data is then used as the preset percentage threshold.

[0034] Finally, it should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0035] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] In this document, the singular forms “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that terms such as “comprising / including” or “having” specify the presence of the stated features, integrals, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, integrals, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0037] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0038] The above description of the disclosed embodiments will enable those skilled in the art to make or use various modifications to these embodiments. It will be readily apparent to those skilled in the art that the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AI-based method for endoscopic polyp identification, characterized in that: Includes the following steps: Step S1: When the polyp detection device performs polyp identification, it acquires a detection image, performs image processing on the detection image to generate a feature image, and detects the white balance shift and pixel displacement of the feature image; Step S2: Calculate the image quality score of the feature image using white balance offset and pixel displacement, perform mucosal texture line detection on the feature image and generate texture distribution features; Step S3: Calculate the texture quality index based on the texture distribution characteristics, analyze the imaging evaluation coefficient in combination with the image quality score, classify the feature image based on the imaging evaluation coefficient, and retrieve the polyp distribution data from the spatial index library; Step S4: The classification results of the feature image generate the distribution coverage features of the current polyp identification. The distribution coverage features and polyp distribution data are combined to evaluate the proportion of abnormal identification. Based on the proportion of abnormal identification, it is determined whether to generate an abnormal detection prompt.

2. The AI-based endoscopic polyp identification method according to claim 1, characterized in that: In step S1, a preset acquisition period is set, and the detection image of the detection area is acquired through the endoscopic image acquisition unit in the polyp detection device; The feature image is obtained by performing image processing on the detection image of the detection area through color edge enhancement and brightness normalization. The average values ​​of the red, green, and blue color channels of each pixel in the feature image are calculated separately. The red channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the red channel of each pixel. The blue channel offset of the feature image is obtained by subtracting the average value of the green channel from the average value of the blue channel of each pixel.

3. The AI-based endoscopic polyp identification method according to claim 2, characterized in that: In step S1, the white balance offset of the feature image is calculated by taking the square root of the sum of the squares of the red channel offset and the blue channel offset of the feature image. The feature value of the corner is obtained by acquiring the numerical value of the local brightness change of the pixel in the feature image through FAST corner detection; Pixels with obvious features are obtained based on the feature value judgment of corner points; Two-dimensional coordinates are established by using the preset fixed vertex of the feature image as the origin of the two-dimensional coordinates and the pixel interval of the feature image as the scale of the horizontal and vertical axes of the two-dimensional coordinates. Select the pixel with obvious features in the feature image of adjacent acquisition time and obtain the displacement of the pixel by Euclidean distance calculation method; The pixel displacement of the feature image is obtained by averaging the displacement of all pixels with obvious features.

4. The AI-based endoscopic polyp identification method according to claim 3, characterized in that: In step S2, the white balance offset and pixel displacement of the feature image are standardized to obtain the offset factor and displacement factor. The image quality score of the feature image is obtained by weighted summation of the offset factor and the displacement factor; Edge response calculation is performed on the feature image using the multi-directional Sobel operator to extract gradient magnitudes in the horizontal, vertical, and diagonal directions; The gradient magnitudes in each direction of the feature image are weighted and superimposed to obtain a comprehensive texture response map.

5. The AI-based endoscopic polyp identification method according to claim 4, characterized in that: In step S2, non-maximum suppression processing is performed on the comprehensive texture response map to remove redundant edge lines and retain the texture lines corresponding to local gradient extrema. The composite texture response map after non-maximum suppression is converted to grayscale to obtain a grayscale image; Calculate the texture energy of a grayscale image based on the gray-level co-occurrence matrix; The texture energy distribution of the grayscale image is formed into a two-dimensional matrix and then normalized to generate the texture distribution features of the feature image.

6. The AI-based endoscopic polyp identification method according to claim 1, characterized in that: In step S3, the average value and standard deviation of the texture distribution features of the feature image are added together to obtain the texture quality index of the feature image; The texture quality index and image quality score of the feature image are processed by the Max-Min normalization method to obtain the texture factor and image factor. The imaging evaluation coefficient is calculated by combining texture factors and image factors. The calculation formula is as follows: ,in, For texture factor, For image factors, This is the imaging evaluation coefficient.

7. The AI-based endoscopic polyp identification method according to claim 6, characterized in that: In step S3, if the imaging evaluation coefficient is greater than the preset imaging threshold, the feature image is determined to be a high-quality image. If the imaging evaluation coefficient is less than or equal to the preset imaging threshold, the feature image is determined to be a low-quality image. Obtain polyp distribution data, including polyp distribution density, from the spatial index.

8. The AI-based endoscopic polyp identification method according to claim 1, characterized in that: In step S4, the detection area is divided into several grid cells of equal size, and each grid cell corresponds to a spatial location; Traverse all feature images within the same detection region, and map the spatial location covered by each image to the corresponding grid cell based on the classification result of each image; If a grid cell has high-quality image coverage, then mark the cell as high-quality coverage in the coverage feature; If a grid cell is covered only by low-quality images, then the cell is marked as low-quality coverage in the coverage feature to fill in the blank areas; After all feature images have been traversed, the labeling information is merged to generate the distribution coverage matrix of the entire detection area, forming the current distribution coverage feature for polyp identification.

9. The AI-based endoscopic polyp identification method according to claim 8, characterized in that: In step S4, the distribution coverage features are mapped to the polyp distribution data in the spatial index library to form the input feature matrix; The input feature matrix is ​​fed into a pre-trained convolutional neural network model to obtain the anomaly detection rate. If the proportion of anomaly detections is greater than or equal to a preset proportion threshold, an anomaly detection prompt will be generated. If the percentage of anomaly detections is less than the preset percentage threshold, then no anomaly detection prompt will be generated.

Images