Multi-feature fusion endoscope image quality online evaluation method and system
By combining the U-Net semantic segmentation model and the inflation-deformation mapping model, the quantitative problem of gastroscopy image quality assessment was solved, enabling the identification of inflation status and accurate extraction of lesion features, thereby improving the diagnostic accuracy of gastroscopy.
Patent Information
- Application Number
- CN202511793194.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-03
AI Technical Summary
Current gastroscopy image analysis technology lacks quantitative evaluation standards and relies on doctors' subjective judgment. It cannot identify image quality problems caused by abnormal inflation, increasing the risk of misdiagnosis and missed diagnosis. Furthermore, existing lesion feature extraction methods do not take into account the interference of inflation fluctuations.
An online assessment method for endoscopic image quality using multi-feature fusion was adopted. Stable anatomical regions were extracted using the U-Net semantic segmentation model, and the credibility signal was determined by combining gray-level entropy. The gastric inflation status was analyzed, the lesion region was located and two-dimensional feature analysis was performed, and an inflation-deformation mapping model was constructed for dynamic correction to eliminate inflation interference.
This improves the reliability of gastroscopy images and the accuracy of lesion feature extraction, reduces the risk of misdiagnosis and missed diagnosis, and ensures the accuracy and reliability of diagnosis.
Smart Images
Figure CN121600318A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image analysis technology, specifically a method and system for online evaluation of endoscopic image quality through multi-feature fusion. Background Technology
[0002] Gastroscopy is the standard for diagnosing digestive tract diseases. However, the quality of gastroscopy images is easily affected by the gastric inflation state. During gastroscopy, gas needs to be injected into the stomach through the endoscope to relax the stomach wall. However, if the inflation volume is not properly controlled, two abnormal states may occur: over-inflation or under-inflation.
[0003] Current gastroscopy image analysis techniques largely rely on doctors' subjective judgment of the inflation status, lacking quantitative evaluation standards, resulting in strong subjectivity and poor consistency. At the same time, existing lesion feature extraction methods do not fully consider the interference of inflation status fluctuations, directly extracting morphological and texture parameters based solely on the original image, leading to extracted features that cannot truly reflect the inherent attributes of the lesion. Furthermore, existing technologies lack a pre-judgment mechanism for the reliability of gastroscopy images, failing to identify image quality problems caused by abnormal inflation in advance, resulting in distorted images directly entering the diagnostic process, further increasing the risk of misdiagnosis and missed diagnosis.
[0004] Therefore, the present invention provides a method and system for online evaluation of endoscopic image quality through multi-feature fusion. Summary of the Invention
[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.
[0006] The technical solution adopted by this invention to solve its technical problem is: an online evaluation method for endoscopic image quality based on multi-feature fusion, comprising:
[0007] A clean gastroscopy image is acquired and a stable anatomical region is extracted using the U-Net semantic segmentation model. Image confidence analysis is performed based on the stable anatomical region to obtain the gray-level entropy of the stable anatomical region. The gray-level entropy is used to determine whether a confidence signal is triggered.
[0008] If triggered, the gastric cavity inflation state is analyzed on the clean gastroscopy image to obtain the image expansion coefficient; the gastric cavity inflation state is identified based on the image expansion coefficient.
[0009] The target lesion region is located based on pure gastroscopy images, and two-dimensional feature analysis is performed on the target lesion region to obtain morphological regularity and texture fluctuation values; the current morphological parameter set is constructed based on the morphological regularity and texture fluctuation values.
[0010] Deformation attribution analysis is performed based on the current set of morphological parameters to obtain morphological residual values and texture residual values; the morphological residual values and texture residual values are then used to determine whether a mapping signal is triggered.
[0011] If a mapping signal is triggered, an inflation-deformation mapping model is constructed based on the image stretching coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set.
[0012] Furthermore, the process of determining whether the credibility signal has been triggered is as follows:
[0013] Obtain normal gastroscopy images of Q cases under standard gastric cavity inflation state, and grayscale entropy of Q cases of normal gastroscopy images;
[0014] The gray entropy feature value is obtained by arithmetically averaging the gray entropy of Q normal gastroscopy images.
[0015] The gray-level entropy of the stable anatomical region is obtained and compared with the gray-level entropy feature value;
[0016] If the gray entropy of at least one region in the stable anatomical region is less than the gray entropy feature value, then the credibility signal is triggered.
[0017] Furthermore, the process of obtaining the gray-level entropy of the stable anatomical region is as follows:
[0018] Obtain stable anatomical regions, including the gastric angle region, the pyloric sphincter region, and the greater curvature region of the gastric body;
[0019] The gray values of all pixels within the gastric angle region are obtained and divided into L gray levels;
[0020] The frequency of each gray level in the gastric angle region is counted and marked as gray level frequency. The probability distribution value of each gray level in the gastric angle region is obtained by calculating the ratio of the gray level frequency of each gray level to the total number of pixels in the gastric angle region.
[0021] The gray entropy of the gastric angle region is calculated using the gray entropy formula based on the probability distribution value of each gray level.
[0022] Obtain the grayscale entropy of the pyloric sphincter region and the grayscale entropy of the greater curvature region of the gastric body.
[0023] Furthermore, the process of performing the gastric cavity inflation state analysis is as follows:
[0024] Obtain the central coordinates of the stable anatomical regions, namely the central coordinates of the gastric angle region (X1, Y1), the central coordinates of the pyloric sphincter (X2, Y2), and the central coordinates of the greater curvature of the gastric body (X3, Y3).
[0025] The Euclidean distance formula is used to calculate the Euclidean distance between (X1, Y1) and (X2, Y2), the Euclidean distance between (X1, Y1) and (X3, Y3), and the Euclidean distance between (X2, Y2) and (X3, Y3), and they are labeled as R1, R2, and R3 respectively.
[0026] Based on R1, R2, and R3, obtain r1, r2, and r3 under standard gastric inflation conditions;
[0027] The ratios of R1 to r1, R2 to r2, and R3 to r3 are respectively used to obtain D1, D2, and D3;
[0028] The image stretching coefficient is calculated by summing D1, D2, and D3 and taking the average value.
[0029] Furthermore, the process of constructing the current morphological parameter set is as follows:
[0030] Identify the target lesion area based on clean gastroscopy images;
[0031] The current morphological regularity is obtained by performing morphological regularity analysis on the target lesion area, and the current texture fluctuation value is obtained by performing texture fluctuation analysis on the target lesion area.
[0032] The current morphological parameter set is obtained by integrating the current morphological regularity of the target lesion area with the current texture fluctuation value.
[0033] Furthermore, the process of identifying the target lesion region is as follows:
[0034] The grayscale value of each pixel in a clean gastroscopy image is obtained using a brightness perception method.
[0035] The gray value of each pixel is compared with the gray value reference range. If the gray value of a pixel is not within the gray value reference range, the pixel is marked as a gray value outlier.
[0036] Gray-scale outliers are connected using the eight-neighbor connectivity method to form regions, and these regions are marked as the target lesion areas.
[0037] Furthermore, the process of performing the aforementioned morphological rule analysis is as follows:
[0038] The contour of the target lesion region is extracted using a gradient thresholding method based on the target lesion region.
[0039] Statistically analyze the coordinates of all contour points in the target lesion area, and find the maximum and minimum values of the X-coordinate and the Y-coordinate among all contour point coordinates;
[0040] The width value is calculated by taking the difference between the maximum and minimum values of the X coordinate, and the height value is calculated by taking the difference between the maximum and minimum values of the Y coordinate.
[0041] The area of the minimum bounding rectangle is calculated by multiplying the width and height values.
[0042] The total number of pixels within the outline of the target lesion region is the actual area of the lesion.
[0043] The current morphological regularity is obtained by comparing the area of the smallest bounding rectangle with the actual area of the lesion.
[0044] Furthermore, the process of performing texture fluctuation analysis is as follows:
[0045] The gray-level co-occurrence matrix parameters are determined based on the target lesion area. The local coordinates of the target lesion area are used as the reference, and the calculation distance of the gray-level co-occurrence matrix is set to K pixels. The calculation direction covers four directions: 0°, 45°, 90°, and 135°.
[0046] Obtain the grayscale values of all pixels within the target lesion area;
[0047] Based on the gray-level co-occurrence matrix parameters and gray values, the gray-level co-occurrence matrices in four calculation directions are calculated respectively. The gray-level matrices in the four directions are then averaged to obtain the comprehensive gray-level co-occurrence matrix of the target lesion region.
[0048] The current texture fluctuation value is calculated using the contrast calculation formula based on the comprehensive gray-level co-occurrence matrix.
[0049] Furthermore, the process of constructing the inflation-deformation mapping model is as follows:
[0050] Calculate the standard image expansion coefficient under standard gastric cavity inflation state based on r1, r2, and r3;
[0051] The morphology correction coefficient is obtained by calculating the ratio of the standard image stretching coefficient to the image stretching coefficient.
[0052] The standard morphological regularity and standard texture fluctuation value are calculated by multiplying the morphological regularity and texture fluctuation value in the current morphological parameter set with the morphological correction coefficient, respectively.
[0053] The standard morphological parameter set is obtained by integrating the standard morphological regularity and the standard texture fluctuation value.
[0054] An online endoscopic image quality assessment system based on multi-feature fusion includes the following modules:
[0055] The confidence analysis module acquires clean gastroscopy images and extracts stable anatomical regions using the U-Net semantic segmentation model. Based on the stable anatomical regions, image confidence analysis is performed to obtain the gray-level entropy of the stable anatomical regions. The gray-level entropy is used to determine whether a confidence signal is triggered.
[0056] Status recognition module: If triggered, it analyzes the gastric cavity inflation status of the clean gastroscopy image to obtain the image expansion coefficient; and identifies the gastric cavity inflation status based on the image expansion coefficient.
[0057] Two-dimensional analysis module: Based on the pure gastroscopy image, the target lesion area is located, and two-dimensional feature analysis is performed on the target lesion area to obtain the morphological regularity and texture fluctuation value; based on the morphological regularity and texture fluctuation value, the current morphological parameter set is constructed.
[0058] Mapping judgment module: Performs deformation attribution analysis based on the current morphological parameter set to obtain morphological residual values and texture residual values; determines whether to trigger a mapping signal based on the morphological residual values and texture residual values;
[0059] Model building module: If a mapping signal is triggered, an inflation-deformation mapping model is built based on the image stretching coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set.
[0060] The beneficial effects of this invention are as follows:
[0061] 1. Acquire clean gastroscopy images and extract stable anatomical regions using the U-Net semantic segmentation model. Perform image confidence analysis on these stable anatomical regions to obtain their grayscale entropy. Determine if a confidence signal is triggered based on the grayscale entropy. Segment the stable anatomical regions using clean gastroscopy images and perform confidence analysis using grayscale entropy to filter reliable images and improve image reliability. If triggered, analyze the gastric cavity inflation state of the clean gastroscopy images to obtain the image expansion coefficient. Identify the gastric cavity inflation state based on the image expansion coefficient. Identify the current gastric cavity inflation state based on reliable clean gastroscopy images. Locate the target lesion region based on the clean gastroscopy images and perform two-dimensional feature analysis on the target lesion region to obtain morphological regularity and texture fluctuation values. Construct the current morphological parameter set based on the morphological regularity and texture fluctuation values. Obtain the corresponding morphological parameter set of the target lesion under the current gastric cavity inflation state to provide a data benchmark for subsequent correction.
[0062] 2. Based on the current morphological parameter set, deformation attribution analysis is performed to obtain morphological residual values and texture residual values; the morphological residual values and texture residual values are used to determine whether a mapping signal is triggered; deformation attribution analysis is used to quantify the difference between the true characteristics of the lesion and interfering factors, providing data support for the correction of lesion feature data in subsequent images; if a mapping signal is triggered, an inflation-deformation mapping model is constructed based on the image expansion coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set; the interference of gastric inflation state on lesion morphology is eliminated, and standardized lesion parameters are obtained through dynamic correction. Attached Figure Description
[0063] The invention will now be further described with reference to the accompanying drawings.
[0064] Figure 1 This is a flowchart illustrating the steps of an online assessment method for endoscopic image quality using multi-feature fusion, as described in an embodiment of the present invention.
[0065] Figure 2 This is a logic diagram for determining whether a mapping signal is triggered, as described in an embodiment of the present invention.
[0066] Figure 3 This is a block diagram of an online endoscopic image quality assessment system based on multi-feature fusion, as described in an embodiment of the present invention. Detailed Implementation
[0067] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0068] Example 1: Please refer to Figure 1 As shown in the embodiments of the present invention, an online assessment method for endoscopic image quality based on multi-feature fusion addresses the core problem of unreliable endoscopic image quality and lesion feature distortion caused by fluctuations in gastric inflation (over-inflation or under-inflation) during gastroscopy. This leads to insufficient image clarity, lesion localization errors, and false variations in lesion morphology regularity and texture features, masking or misjudging the true state of the lesion, ultimately affecting the accuracy and reliability of clinical diagnosis. The method extracts stable anatomical regions such as the gastric angle and pyloric sphincter using a U-Net semantic segmentation model, combines grayscale entropy for image confidence screening, and then uses Euclidean distance to calculate the image expansion coefficient to identify the gastric inflation state. Subsequently, it locates the target lesion and extracts morphology regularity and texture fluctuation values to construct the current morphological parameter set. Deformation attribution analysis determines whether the lesion deformation is caused by inflation. Finally, by constructing an inflation-deformation mapping model and morphological correction coefficients, the current morphological parameters are dynamically corrected to the standard morphological parameter set under standard inflation conditions, greatly reducing the interference of abnormal gastric inflation on image quality and lesion features. Specifically, the method includes the following steps:
[0069] Step 1: Acquire clean gastroscopy images and extract stable anatomical regions using the U-Net semantic segmentation model. Based on the stable anatomical regions, perform image confidence analysis to obtain the grayscale entropy of the stable anatomical regions. Determine whether a confidence signal is triggered based on the grayscale entropy.
[0070] In step one, the process of acquiring clean gastroscopy images and extracting stable anatomical regions using the U-Net semantic segmentation model, followed by image confidence analysis to obtain the gray-level entropy of the stable anatomical regions, is as follows:
[0071] The raw gastroscopy images are acquired, and a two-stage preprocessing purification is performed on the raw gastroscopy images to obtain clean gastroscopy images.
[0072] The preferred method for two-stage pretreatment purification is as follows:
[0073] The first level targets the original gastroscopy image and uses the eight-neighbor average absolute gray difference method to filter out random noise in the image.
[0074] The second-level processing is performed on the original gastroscopy image after the first-level processing. The second-level processing uses the linear contrast stretching method to map the pixel grayscale of the local window to the full grayscale space, highlighting the grayscale difference between the gastric mucosa and the stable anatomical area, and obtaining a pure gastroscopy image.
[0075] Preferably, the size of the local window is 16×16 pixels;
[0076] It should be noted that the stable anatomical regions include the gastric angle region, the pyloric sphincter region, and the greater curvature region of the gastric body.
[0077] It should also be noted that the advantages of selecting the gastric angle, pyloric sphincter, and greater curvature of the gastric body as stable anatomical regions are that all three have strong anatomical stability and are not easily affected by fluctuations in gastric gas volume. The anatomical characteristics of the three regions enable them to escape the interference of fluctuations in gastric gas volume and always provide a stable and repeatable positioning benchmark, providing a reliable data basis for subsequent assessment of gastric gas volume status and correction of lesion characteristics.
[0078] Stable anatomical regions were extracted from clean gastroscopy images using the U-Net semantic segmentation model. Image regions of the gastric angle, pyloric sphincter, and greater curvature of the gastric body were extracted separately, and the image region of each stable anatomical region was defined as 10×10 pixels.
[0079] The process of obtaining grayscale entropy for stable anatomical regions is as follows:
[0080] Divide the gray values of all pixels in the gastric angle region into L gray levels;
[0081] Preferably, L is 256, which is the full grayscale space;
[0082] The frequency of each gray level in the gastric angle region is counted and marked as gray level frequency. The probability distribution value of each gray level in the gastric angle region is obtained by calculating the ratio of the gray level frequency of each gray level to the total number of pixels in the gastric angle region.
[0083] The gray entropy of the gastric angle region is calculated using the gray entropy formula based on the probability distribution value of each gray level in the gastric angle region.
[0084] Based on the above method of obtaining grayscale entropy, the grayscale entropy of the pyloric sphincter region and the grayscale entropy of the greater curvature region of the gastric body are obtained.
[0085] In step one, the process of determining whether to trigger a credibility signal based on grayscale entropy is as follows:
[0086] Based on the clinical gastroscopy image database, normal gastroscopy images of Q cases under standard gastric cavity inflation were obtained, and the grayscale entropy of the Q normal gastroscopy images was obtained.
[0087] Preferably, Q is 1000; the standard gastric cavity inflation state is the standardized examination state in which an appropriate amount of gas is injected into the stomach through the endoscope during gastroscopy, so that the gastric cavity is moderately expanded, the gastric mucosa is fully stretched without wrinkles, and the gastric wall structure and mucosal details can be clearly presented.
[0088] The gray entropy feature value is obtained by arithmetically averaging the gray entropy of Q normal gastroscopy images.
[0089] For example, the gray-level entropy of the acquired stable anatomical region is compared with the gray-level entropy feature value;
[0090] If the gray-level entropy of the stable anatomical region is greater than or equal to the gray-level entropy feature value, it indicates that the image of the stable anatomical region has high clarity and no processing is required.
[0091] If the gray entropy of at least one region in a stable anatomical region is less than the gray entropy feature value, it indicates that the stable anatomical region may have abnormal gray distribution due to abnormal gastric inflation, triggering a confidence signal.
[0092] Step 2: If triggered, analyze the gastric cavity inflation state of the clean gastroscopy image to obtain the image expansion coefficient; identify the gastric cavity inflation state based on the image expansion coefficient;
[0093] In step two, if triggered, the process of analyzing the gastric cavity inflation state of the gastroscopy image to obtain the image expansion coefficient is as follows:
[0094] Based on the pure gastroscopy images, the center coordinates of three stable anatomical regions were obtained, namely the center coordinates of the gastric angle region (X1, Y1), the center coordinates of the pyloric sphincter (X2, Y2), and the center coordinates of the greater curvature of the gastric body (X3, Y3).
[0095] The Euclidean distance formula was used to calculate the Euclidean distance between the center coordinates of the gastric angle region and the center coordinates of the pyloric sphincter, the Euclidean distance between the center coordinates of the gastric angle region and the center coordinates of the greater curvature of the gastric body, and the Euclidean distance between the center coordinates of the pyloric sphincter and the center coordinates of the greater curvature of the gastric body, and these distances were labeled as R1, R2, and R3, respectively.
[0096] Based on R1, R2, and R3, we introduce the Euclidean distance r1 between the center coordinates of the gastric angle region and the center coordinates of the pyloric sphincter under standard gastric inflation conditions, the Euclidean distance r2 between the center coordinates of the gastric angle region and the center coordinates of the greater curvature of the gastric body, and the Euclidean distance r3 between the center coordinates of the pyloric sphincter and the center coordinates of the greater curvature of the gastric body.
[0097] Preferably, r1, r2, and r3 are the average values obtained by statistical analysis of normal gastroscopy images under the standard gastric cavity inflation state in Q cases;
[0098] The ratios of R1 to r1, R2 to r2, and R3 to r3 are respectively used to obtain D1, D2, and D3;
[0099] The image stretching coefficient is calculated by summing D1, D2, and D3 and taking the average value.
[0100] For example, the image stretching factor is compared with a preset standard factor range;
[0101] If the image stretching coefficient is greater than the maximum value of the preset standard coefficient range, it indicates that the current gastric cavity inflation state is an overinflation state.
[0102] It should be noted that the preset standard coefficient range was obtained by those skilled in the art through comparative analysis of gastroscopy images under different gastric inflation states;
[0103] If the image expansion coefficient is within the preset standard coefficient range, it indicates that the current gastric cavity inflation state is the standard inflation state.
[0104] If the image stretching coefficient is less than the minimum value of the preset standard coefficient range, it indicates that the current gastric cavity is underfilled.
[0105] It should be noted that the purpose of obtaining the image expansion coefficient is to: fully assess the gastric cavity's inflation state by quantifying the spatial expansion state of the gastric cavity's anatomical structures, thereby providing a basis for quality control of gastroscopy images; by comparing the image expansion coefficient with the mean distance under standard inflation conditions, it can clearly distinguish between over-inflation, standard inflation, and under-inflation states: if the coefficient exceeds the upper limit of the standard range, it suggests that excessive gastric expansion may obscure lesions or cause discomfort during observation; if it is below the lower limit, it means that the gastric cavity is not fully expanded, which may lead to the omission of mucosal details; this ensures the standardization of gastroscopy image acquisition and lays a data foundation for subsequent lesion feature correction and improved diagnostic accuracy;
[0106] Step 3: Locate the target lesion region based on the clean gastroscopy image, perform two-dimensional feature analysis on the target lesion region to obtain morphological regularity and texture fluctuation value; construct the current morphological parameter set of the target lesion region based on the morphological regularity and texture fluctuation value;
[0107] In step three, the process of locating the target lesion area based on the clean gastroscopy image is as follows:
[0108] Based on the coordinates of the centers of the three stable anatomical regions that have been obtained, a planar rectangular reference system is constructed.
[0109] The process of constructing a Cartesian reference system is as follows:
[0110] Set the center coordinates of the gastric angle region as the origin, and mark the X-axis along the line connecting the center coordinates of the gastric angle region and the center coordinates of the pyloric sphincter region.
[0111] Take the unit vector perpendicular to the X-axis as the Y-axis, with its direction pointing from the center coordinates of the gastric angle region to the greater curvature region of the gastric body;
[0112] Convert all pixel coordinates in a clean gastroscopy image to local coordinates in a Cartesian reference system;
[0113] The grayscale value of each pixel in a clean gastroscopy image is obtained using a brightness perception method.
[0114] For example, the way the brightness perception method calculates grayscale values is as follows: if the RGB value of a pixel is (100, 200, 50), then the grayscale value of this pixel is (0.29×100) + (0.587×200) + (0.114×50).
[0115] The gray value of each pixel is compared with the gray value reference range. If the gray value of a pixel is not within the gray value reference range, the pixel is marked as a gray value outlier. Otherwise, no processing is performed on the pixel.
[0116] It should be noted that the grayscale reference range was set by those skilled in the art based on the grayscale values of lesion areas in historical gastroscopy images;
[0117] Gray-level outliers are connected using the eight-neighbor connectivity method to form regions, and these regions are marked as the target lesion regions. The location of the target lesion region is then determined based on a Cartesian reference system.
[0118] It should be noted that the eight-neighbor connectivity method is a method for identifying continuous regions in an image by determining whether a pixel has common features (such as gray values) with its eight neighboring pixels (upper, lower, left, right, and four diagonals). It is used for region segmentation or target extraction.
[0119] In step three, two-dimensional feature analysis is performed on the target lesion region to obtain morphological regularity and texture fluctuation values; the process of constructing the current morphological parameter set based on morphological regularity and texture fluctuation values is as follows:
[0120] Morphological regularity of the target lesion area:
[0121] The contour of the target lesion region is extracted using a gradient thresholding method based on the target lesion region.
[0122] Based on the outline of the target lesion area, count the coordinates of all outline points; find the maximum and minimum values of the X coordinate and the maximum and minimum values of the Y coordinate among all outline point coordinates;
[0123] The width value is calculated by taking the difference between the maximum and minimum values of the X coordinate, and the height value is calculated by taking the difference between the maximum and minimum values of the Y coordinate.
[0124] It should be noted that the width and height values are the width and height values of the smallest bounding rectangle of the target lesion area;
[0125] The area of the minimum bounding rectangle is calculated by multiplying the width and height values.
[0126] The total number of pixels within the outline of the target lesion region is the actual area of the lesion.
[0127] The current morphological regularity is obtained by comparing the area of the smallest bounding rectangle with the actual area of the lesion.
[0128] Texture fluctuation values of the target lesion area:
[0129] The gray-level co-occurrence matrix parameters are determined based on the target lesion area. The local coordinates of the target lesion area are used as the reference, and the calculation distance of the gray-level co-occurrence matrix is set to K pixels. The calculation direction covers four directions: 0°, 45°, 90°, and 135°.
[0130] Preferably, K is 2;
[0131] Obtain the grayscale values of all pixels within the target lesion area;
[0132] Based on the determined gray-level co-occurrence matrix parameters and the gray values corresponding to all pixels in the target lesion area, the gray-level co-occurrence matrices in four directions are calculated respectively. The gray-level matrices in the four directions are then averaged to obtain the comprehensive gray-level co-occurrence matrix of the target lesion area.
[0133] The current texture fluctuation value is calculated using the contrast calculation formula based on the comprehensive gray-level co-occurrence matrix.
[0134] The current morphological parameter set is obtained by integrating the current morphological regularity of the target lesion area with the current texture fluctuation value;
[0135] It should be noted that the advantage of obtaining the current morphological parameter set is that, by quantifying the morphological regularity and texture fluctuation value of the target lesion, the constructed morphological parameter set can provide quantifiable data support for the control of gastroscopy image quality; the current morphological parameter set formed by combining the current morphological regularity and the current texture fluctuation value not only realizes the digital representation of lesion features, but also further corrects the lesion deformation error caused by changes in gastric cavity morphology through correlation analysis with the gastric cavity inflation state (such as over-inflation, standard inflation, or under-inflation), ensuring the accuracy and reliability of lesion feature extraction;
[0136] The technical solution of this invention is as follows: A clean gastroscopy image is acquired, and a stable anatomical region is extracted using the U-Net semantic segmentation model. Image confidence analysis is performed on the stable anatomical region to obtain its grayscale entropy. The grayscale entropy is used to determine whether a confidence signal is triggered. The stable anatomical region is segmented using the clean gastroscopy image, and confidence analysis is performed using the grayscale entropy to filter reliable images and improve image reliability. If triggered, the gastric cavity inflation state is analyzed on the clean gastroscopy image to obtain the image expansion coefficient. The gastric cavity inflation state is identified based on the image expansion coefficient. The current gastric cavity inflation state is identified based on the reliable clean gastroscopy image. The target lesion region is located based on the clean gastroscopy image, and two-dimensional feature analysis is performed on the target lesion region to obtain its morphological regularity and texture fluctuation value. A current morphological parameter set is constructed based on the morphological regularity and texture fluctuation value. The corresponding morphological parameter set of the target lesion under the current gastric cavity inflation state is obtained to provide a data benchmark for subsequent correction.
[0137] Example 2: Please refer to Figures 1-2 As shown in the embodiment of the present invention, an online evaluation method for endoscopic image quality based on multi-feature fusion includes:
[0138] Step 4: Perform deformation attribution analysis based on the current morphological parameter set to obtain morphological residual values and texture residual values; determine whether to trigger a mapping signal based on the morphological residual values and texture residual values.
[0139] In step four, the process of obtaining the morphological regularity residual and texture fluctuation residual by performing deformation attribution analysis on the target lesion area based on the constructed inflation-deformation mapping model is as follows:
[0140] Based on the clinical gastroscopy image database, normal gastroscopy images under standard gastric cavity inflation were obtained, and the morphological regularity benchmark value and texture fluctuation benchmark value of similar target lesion areas were obtained.
[0141] The morphological residual value is obtained by performing absolute difference processing between the current morphological regularity and the current morphological regularity benchmark value; the texture residual value is obtained by performing absolute difference processing between the current texture fluctuation value and the texture fluctuation benchmark value.
[0142] Q gastroscopy images without lesions but with different gastric inflation states were obtained from the clinical gastroscopy image database. The morphological regularity and texture fluctuation values of each gastroscopy image were obtained, and the control morphological residual value and control texture residual value of each gastroscopy image were calculated.
[0143] In step four, the process of determining whether to trigger the mapping signal based on the morphological residual value and the texture residual value is as follows:
[0144] The confidence intervals for morphological residuals and texture residuals were obtained by statistically analyzing all control morphological residuals and control texture residuals, respectively.
[0145] The upper limit of the confidence interval for morphological residuals is set to the morphological regularity threshold, and the upper limit of the confidence interval for texture residuals is set to the texture fluctuation threshold.
[0146] For example, the morphological residual value and the texture residual value are compared with the morphological regularity threshold and the texture fluctuation threshold, respectively;
[0147] If the morphological residual value is less than or equal to the morphological regularity threshold and the texture residual value is less than or equal to the texture fluctuation threshold, it indicates that the actual lesion morphology is within a reasonable range from the morphological deviation caused by pure gas filling, and there is no morphological change caused by additional lesions. It is determined to be a lesion deformation caused by gastric gas filling, triggering the mapping signal.
[0148] Conversely, it indicates that there is an additional shape deviation beyond the influence of inflation, which is caused by other uncontrollable factors;
[0149] Step 5: If the mapping signal is triggered, construct an inflation-deformation mapping model based on the image stretching coefficient and the current morphological parameter set; dynamically correct the current morphological parameters based on the inflation-deformation mapping model to obtain a standard morphological parameter set;
[0150] In step five, the process of constructing the inflation-deformation mapping model is as follows:
[0151] An inflation-deformation mapping model is constructed. The image stretching coefficient and the current morphological parameter set are input into the inflation-deformation mapping model to obtain the standard morphological parameter set of the target lesion area under the standard gastric cavity inflation state.
[0152] The process of constructing the inflation-deformation mapping model is as follows:
[0153] S301. Obtain the Euclidean distance r1 between the center coordinates of the gastric angle region and the center coordinates of the pyloric sphincter under standard gastric inflation conditions, the Euclidean distance r2 between the center coordinates of the gastric angle region and the center coordinates of the pyloric sphincter, and the Euclidean distance r3 between the center coordinates of the pyloric sphincter and the center coordinates of the greater curvature of the gastric body.
[0154] Calculate the standard image expansion coefficient under standard gastric cavity inflation state based on r1, r2, and r3;
[0155] S302. Calculate the morphology correction coefficient by the ratio of the standard image stretching coefficient to the image stretching coefficient.
[0156] It should be noted that the physical meaning of the morphological correction coefficient is: based on the ratio of the standard image expansion coefficient under standard inflation state to the current image expansion coefficient, it quantifies the degree of deviation of the current inflation state from the standard state; the morphological correction coefficient is essentially to provide a targeted compensation benchmark for the deviation of lesion features caused by abnormal inflation, ensuring that the regularity of lesion morphology and texture fluctuation value after correction can truly reflect the inherent characteristics of lesions under standard physiological state.
[0157] S303. Multiply the morphological regularity and texture fluctuation value in the current morphological parameter set with the morphological correction coefficient to obtain the standard morphological regularity and standard texture fluctuation value.
[0158] S304. Integrate the standard morphological regularity with the standard texture fluctuation value to obtain the standard morphological parameter set;
[0159] Understandably, the benefits of constructing an inflation-deformation mapping model are as follows: by quantifying the deviation between the current gastric inflation state and the standard inflation state, a morphological correction coefficient is established as a compensation benchmark, achieving standardized conversion of lesion morphological parameters under different gastric inflation states; the inflation-deformation mapping model dynamically correlates the gastric inflation state with lesion morphological features, forming a state-feature pair binding analysis framework, enabling lesion features to be compared and evaluated under a unified physiological benchmark; and it provides quantifiable standards for the quality control of gastroscopy images and the correction of lesion features.
[0160] Based on the established inflation-deformation mapping model, the current morphological parameter set of the target lesion area is obtained;
[0161] The current morphological parameter set is input into the inflation-deformation mapping model and deformed to obtain the standard morphological parameter set of the target lesion area.
[0162] It should be noted that the standard morphological parameter set consists of the morphological parameters of the target lesion area under standard gastric cavity inflation conditions, eliminating the interference of abnormal gastric cavity inflation on the morphology of the target lesion area and restoring the inherent characteristics of the target lesion.
[0163] The technical solution of this invention is as follows: Deformation attribution analysis is performed based on the current morphological parameter set to obtain morphological residual values and texture residual values; it is determined whether a mapping signal is triggered based on the morphological residual values and texture residual values; the difference between the true features of the lesion and interfering factors is quantified through deformation attribution analysis, providing data support for the correction of lesion feature data in subsequent images; if a mapping signal is triggered, an inflation-deformation mapping model is constructed based on the image expansion coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set; the interference of the gastric cavity inflation state on the lesion morphology is eliminated, and standardized lesion parameters are obtained through dynamic correction.
[0164] Example 3: Please refer to Figure 3 As shown in the embodiment of the present invention, an online assessment system for endoscopic image quality based on multi-feature fusion includes the following modules:
[0165] The confidence analysis module acquires clean gastroscopy images and extracts stable anatomical regions using the U-Net semantic segmentation model. Based on the stable anatomical regions, image confidence analysis is performed to obtain the gray-level entropy of the stable anatomical regions. The gray-level entropy is used to determine whether a confidence signal is triggered.
[0166] Status recognition module: If triggered, it analyzes the gastric cavity inflation status of the clean gastroscopy image to obtain the image expansion coefficient; and identifies the gastric cavity inflation status based on the image expansion coefficient.
[0167] Two-dimensional analysis module: Based on the pure gastroscopy image, the target lesion area is located, and two-dimensional feature analysis is performed on the target lesion area to obtain the morphological regularity and texture fluctuation value; based on the morphological regularity and texture fluctuation value, the current morphological parameter set is constructed.
[0168] Mapping judgment module: Performs deformation attribution analysis based on the current morphological parameter set to obtain morphological residual values and texture residual values; determines whether to trigger a mapping signal based on the morphological residual values and texture residual values;
[0169] Model building module: If a mapping signal is triggered, an inflation-deformation mapping model is built based on the image stretching coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set.
[0170] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for online evaluation of endoscopic image quality through multi-feature fusion, characterized in that: include: A clean gastroscopy image is acquired and a stable anatomical region is extracted using the U-Net semantic segmentation model. Image confidence analysis is performed based on the stable anatomical region to obtain the gray-level entropy of the stable anatomical region. The gray-level entropy is used to determine whether a confidence signal is triggered. If triggered, the gastric cavity inflation state is analyzed on the clean gastroscopy image to obtain the image expansion coefficient; the gastric cavity inflation state is identified based on the image expansion coefficient. Based on the pure gastroscopy image, the target lesion area is located, and two-dimensional feature analysis is performed on the target lesion area to obtain the morphological regularity and texture fluctuation value; The current morphological parameter set is constructed based on morphological regularity and texture fluctuation value; Deformation attribution analysis is performed based on the current set of morphological parameters to obtain morphological residual values and texture residual values; Determine whether to trigger the mapping signal based on the morphological residual value and the texture residual value; If the mapping signal is triggered, an inflation-deformation mapping model is constructed based on the image stretching coefficient and the current morphological parameter set. The standard morphological parameter set is obtained by dynamically correcting the current morphological parameters based on the inflation-deformation mapping model.
2. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 1, characterized in that: The process of determining whether the credibility signal has been triggered is as follows: Obtain normal gastroscopy images of Q cases under standard gastric cavity inflation state, and grayscale entropy of Q cases of normal gastroscopy images; The gray entropy feature value is obtained by arithmetically averaging the gray entropy of Q normal gastroscopy images. The gray-level entropy of the stable anatomical region is obtained and compared with the gray-level entropy feature value; If the gray entropy of at least one region in the stable anatomical region is less than the gray entropy feature value, then the credibility signal is triggered.
3. The online evaluation method for endoscopic image quality based on multi-feature fusion according to claim 2, characterized in that: The process of obtaining the gray-level entropy of the stable anatomical region is as follows: Obtain stable anatomical regions, including the gastric angle region, the pyloric sphincter region, and the greater curvature region of the gastric body; The gray values of all pixels within the gastric angle region are obtained and divided into L gray levels; The frequency of each gray level in the gastric angle region is counted and marked as gray level frequency. The probability distribution value of each gray level in the gastric angle region is obtained by calculating the ratio of the gray level frequency of each gray level to the total number of pixels in the gastric angle region. The gray entropy of the gastric angle region is calculated using the gray entropy formula based on the probability distribution value of each gray level. Obtain the grayscale entropy of the pyloric sphincter region and the grayscale entropy of the greater curvature region of the gastric body.
4. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 1, characterized in that: The process of performing the gastric cavity inflation state analysis is as follows: Obtain the central coordinates of the stable anatomical regions, namely the central coordinates of the gastric angle region (X1, Y1), the central coordinates of the pyloric sphincter (X2, Y2), and the central coordinates of the greater curvature of the gastric body (X3, Y3). The Euclidean distance formula is used to calculate the Euclidean distance between (X1, Y1) and (X2, Y2), the Euclidean distance between (X1, Y1) and (X3, Y3), and the Euclidean distance between (X2, Y2) and (X3, Y3), and they are labeled as R1, R2, and R3 respectively. Based on R1, R2, and R3, obtain r1, r2, and r3 under standard gastric inflation conditions; The ratios of R1 to r1, R2 to r2, and R3 to r3 are respectively used to obtain D1, D2, and D3; The image stretching coefficient is calculated by summing D1, D2, and D3 and taking the average value.
5. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 1, characterized in that: The process of constructing the current morphological parameter set is as follows: Identify the target lesion area based on clean gastroscopy images; The current morphological regularity is obtained by performing morphological regularity analysis on the target lesion area, and the current texture fluctuation value is obtained by performing texture fluctuation analysis on the target lesion area. The current morphological parameter set is obtained by integrating the current morphological regularity of the target lesion area with the current texture fluctuation value.
6. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 5, characterized in that: The process of identifying the target lesion area is as follows: The grayscale value of each pixel in a clean gastroscopy image is obtained using a brightness perception method. The gray value of each pixel is compared with the gray value reference range. If the gray value of a pixel is not within the gray value reference range, the pixel is marked as a gray value outlier. Gray-scale outliers are connected using the eight-neighbor connectivity method to form regions, and these regions are marked as the target lesion areas.
7. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 5, characterized in that: The process of performing the aforementioned morphological rule analysis is as follows: The contour of the target lesion region is extracted using a gradient thresholding method based on the target lesion region. Statistically analyze the coordinates of all contour points in the target lesion area, and find the maximum and minimum values of the X-coordinate and the Y-coordinate among all contour point coordinates; The width value is calculated by taking the difference between the maximum and minimum values of the X coordinate, and the height value is calculated by taking the difference between the maximum and minimum values of the Y coordinate. The area of the minimum bounding rectangle is calculated by multiplying the width and height values. The total number of pixels within the outline of the target lesion region is the actual area of the lesion. The current morphological regularity is obtained by comparing the area of the smallest bounding rectangle with the actual area of the lesion.
8. The online evaluation method for endoscopic image quality based on multi-feature fusion according to claim 5, characterized in that: The process of performing texture fluctuation analysis is as follows: The gray-level co-occurrence matrix parameters are determined based on the target lesion area. The local coordinates of the target lesion area are used as the reference, and the calculation distance of the gray-level co-occurrence matrix is set to K pixels. The calculation direction covers four directions: 0°, 45°, 90°, and 135°. Obtain the grayscale values of all pixels within the target lesion area; Based on the gray-level co-occurrence matrix parameters and gray values, the gray-level co-occurrence matrices in four calculation directions are calculated respectively. The gray-level matrices in the four directions are then averaged to obtain the comprehensive gray-level co-occurrence matrix of the target lesion region. The current texture fluctuation value is calculated using the contrast calculation formula based on the comprehensive gray-level co-occurrence matrix.
9. The online assessment method for endoscopic image quality based on multi-feature fusion according to claim 1, characterized in that: The process of constructing the inflation-deformation mapping model is as follows: Calculate the standard image expansion coefficient under standard gastric cavity inflation state based on r1, r2, and r3; The morphology correction coefficient is obtained by calculating the ratio of the standard image stretching coefficient to the image stretching coefficient. The standard morphological regularity and standard texture fluctuation value are calculated by multiplying the morphological regularity and texture fluctuation value in the current morphological parameter set with the morphological correction coefficient, respectively. The standard morphological parameter set is obtained by integrating the standard morphological regularity and the standard texture fluctuation value.
10. A multi-feature fusion online assessment system for endoscopic image quality, used to implement the multi-feature fusion online assessment method for endoscopic image quality as described in any one of claims 1-9, characterized in that, Includes the following modules: The confidence analysis module acquires clean gastroscopy images and extracts stable anatomical regions using the U-Net semantic segmentation model. Based on the stable anatomical regions, image confidence analysis is performed to obtain the gray-level entropy of the stable anatomical regions. The gray-level entropy is used to determine whether a confidence signal is triggered. Status recognition module: If triggered, it analyzes the gastric cavity inflation status of the clean gastroscopy image to obtain the image expansion coefficient; and identifies the gastric cavity inflation status based on the image expansion coefficient. Two-dimensional analysis module: Based on the pure gastroscopy image, the target lesion area is located, and two-dimensional feature analysis is performed on the target lesion area to obtain the morphological regularity and texture fluctuation value; The current morphological parameter set is constructed based on morphological regularity and texture fluctuation value; Mapping judgment module: Based on the current morphological parameter set, it performs deformation attribution analysis to obtain morphological residual values and texture residual values; Determine whether to trigger the mapping signal based on the morphological residual value and the texture residual value; Model building module: If a mapping signal is triggered, an inflation-deformation mapping model is built based on the image stretching coefficient and the current morphological parameter set; the current morphological parameters are dynamically corrected based on the inflation-deformation mapping model to obtain a standard morphological parameter set.