CN116434920A - A method and device for predicting the risk of progression of gastrointestinal metaplasia - Google Patents

Abstract

The invention discloses a method and device for predicting the risk of progression of gastrointestinal metaplasia, and relates to the technical field of image processing. The feature quantification value of the gastroscopic image is classified into intestinal grade; the first label attribute of the marker segmentation image includes blood vessel attribute, fold attribute, color attribute, and diffuse attribute; the second label attribute of the atrophic gastritis segmentation image includes Roughness attribute, fuzzy attribute, off-white attribute, bright blue ridge attribute; the third label attribute of the intestinalized segmentation image includes position attribute and shape attribute. The present invention fully considers the influence of multiple characteristic quantification values of different attributes of the gastroscope image on the accuracy and intuition of image processing, and effectively improves the identification efficiency and identification accuracy of intestinal metaplasia risk levels.

Description

A method and device for predicting the risk of progression of gastrointestinal metaplasia

technical field

The invention belongs to the technical field of image processing, and in particular relates to a method and device for predicting the risk of gastrointestinal metaplasia progression.

Background technique

Intestinal metaplasia is the main precancerous disease of intestinal type gastric cancer. A multi-center survey found that the prevalence of gastrointestinal metaplasia in my country is about 23.6%, which is a huge base. At present, patients with intestinal metaplasia are regularly reexamined according to the guidelines in clinical practice. However, only 0.2% of intestinal metaplasia patients have the chance to progress to gastric cancer, and the remaining patients remain stable or even improve. However, there is no effective way to distinguish between high and low risk of intestinal metaplasia. On the one hand, this has led to doctors and patients not paying much attention to intestinal metaplasia, and the standard review rate is less than 50%, which has delayed the treatment of some patients with high risk of progression; A certain amount of medical resources will be wasted.

Contents of the invention

In view of the above problems, the first party of the present invention provides a method for predicting the risk of progression of gastrointestinal metaplasia, which can effectively improve the identification efficiency and accuracy of intestinal metaplasia risk level.

For achieving above object, the technical scheme that the present invention takes is:

A method for predicting the risk of progression of gastrointestinal metaplasia, comprising the following steps:

By obtaining the feature quantification value of the label attribute of the marker segmentation image, the atrophic gastritis segmentation image and the intestinalization segmentation image, the intestinalization level classification of the gastroscopy image is carried out;

The first label attributes of the marker segmentation image include blood vessel attributes, fold attributes, color attributes, and diffuse attributes;

The second label attributes of the atrophic gastritis segmented image include roughness attributes, fluff-like attributes, off-white attributes, and bright blue ridge attributes;

The third label attribute of the intestinalized segmented image includes position attribute and shape attribute.

In some embodiments, the step of classifying the intestinalization level of the gastroscopic image by acquiring the marker segmentation image, the atrophic gastritis segmentation image and the feature quantification value of the label attribute of the intestinalization segmentation image includes the following steps:

Obtaining a gastroscope image, performing marker segmentation on the gastroscope image, and acquiring a marker segmentation image;

Carrying out feature extraction of the first label attribute on the marker segmentation image, obtaining the first feature quantization value corresponding to the first label attribute, inputting the first feature quantization value into a trained machine learning classifier for classification, and obtaining gastroscope image atrophy Classification results of gastritis;

Carrying out feature extraction of the second label attribute on the segmented image of atrophic gastritis, obtaining a second feature quantization value corresponding to the second label attribute, inputting the second feature quantization value into a trained machine learning classifier for classification, and obtaining Gastroscopic image enterochemical classification results;

Carrying out the feature extraction of the third tag attribute on the intestinalization segmented image, obtaining the third feature quantization value corresponding to the third tag attribute, inputting the third feature quantization value into a trained machine learning classifier for classification, and obtaining the gastroscopy Image enterification risk level classification results.

In some embodiments, the feature extraction of the first label attribute is performed on the marker segmentation image, and the first feature quantization value corresponding to the first label attribute is obtained, including the following steps:

Obtain the blood vessel segmentation image in the marker segmentation image;

According to the formula:

Obtain the blood vessel feature quantization value label ₁ , where n _v is the number of blood vessels in the blood vessel segmentation image, s _v is the area of blood vessels in the blood vessel segmentation image, and S is the area of the marker segmentation image;

According to the formula:

Get the wrinkle feature quantization value label ₂ , where s _z is the wrinkle area in the marker region in the marker segmentation image, _nz is the number of folds in the marker region in the marker segmentation image, S is the area of the marker segmentation image, s _f is the wrinkle area in the non-marker region in the marker segmentation image, n _f is the number of folds in the non-marker region in the marker segmentation image;

According to the formula:

Get the color feature quantization value label ₃ , where r _c , g _c , b _c are the average color values of the three channels, S is the area of the marker region in the marker segmentation image, and n _c is the calculation of the remaining list after removing the color close to black average number;

According to the formula:

The diffuse feature quantification value label ₄ is obtained, where M is the number of images collected for the entire gastric inner wall, w _Qi , h _Qi are the width and height of each image, S _i is the marker segmentation of each image and in the connected domain Based on the area to the marker, w _bi , h _bi are the width and height of the marker, and (x _bi , y _bi ) are the centroid coordinates of the marker.

In some embodiments, the obtaining step of the blood vessel segmentation image includes:

Binarize the marker segmentation image into a grayscale image;

performing an erosion operation on the binarized marker segmentation image to obtain a mask image of the marker segmentation image;

performing median filtering on the grayscale image of the marker segmentation image;

performing histogram equalization on the obtained median-filtered marker segmentation image;

performing gamma transformation on the marker segmented image after the histogram equalization;

performing a convolution operation on the obtained gamma-transformed marker segmentation image;

Comparing the convolutional marker segmentation image with the marker segmentation image mask image bit by bit, if a pixel value in the mask image is 0, the convolutional marker segmentation image is here Set the pixel value at 0 to obtain the denoised marker segmentation image;

Contrast stretching is performed on the denoised marker segmented image to obtain a blood vessel segmented image corresponding to the marker segmented image.

In some embodiments, the feature extraction of the second label attribute is performed on the segmented image of atrophic gastritis, and the second feature quantization value corresponding to the second label attribute is obtained, including the following steps:

According to the formula:

Get the roughness feature quantization value label ₅ , where W _W , H _W are the width and height of the segmented image of atrophic gastritis respectively, P _mean is the average pixel value of the segmented image of atrophic gastritis, img _W is the segmented image of atrophic gastritis, W ₀ is the row with the largest variance after binarizing the atrophic gastritis segmentation map according to a set threshold, 0<W ₀ <W _W ;

According to the formula:

Get the fluff-like feature quantization value label ₆ , where N _r is the number of fluff-like regions, s _ri is the area of the fluff-like regions, w _ri , h _ri are the width and height of the smallest circumscribed rectangle of the fluff-like regions, and n _ri is the fluff-like regions and break at the corners to obtain the number of fluff segments;

According to the formula:

Obtain the partial white feature quantization value label ₇ , wherein n _b is the maximum number of pixel points, and (r _bi , g _bi , b _bi ) the pixel value of each pixel point of the category with the largest number of colors;

According to the formula:

是标准亮蓝脊三通道平均像素值，PL是亮蓝脊分割图像平均像素值。Get the bright blue ridge feature quantization value label ₈ , where W _L , _HL are the width and height of the bright blue ridge segmented image, img _L is the bright blue ridge segmented image, (x _RSL , y _RSL ) is the bright blue ridge segmented image Centroid coordinates, S _RSL is the bright blue ridge segmented image area, (x _RS , y _RS ) is the centroid coordinates of the stained and enlarged atrophic gastritis segmented image, S _RS is the area of the stained and enlarged atrophic gastritis,

is the average pixel value of the three channels of the standard bright blue ridge, and PL is the average pixel value of the bright blue ridge segmented image.

In some embodiments, the feature extraction of the third label attribute is performed on the intestinalization segmentation image, and the third feature quantization value corresponding to the third label attribute is obtained, including the following steps:

According to the formula:

Get the location feature quantization value label9, where S _C is the intestinalized segmented image area, [(x _FDX ,y _FDX ),(x _FDD ,y _FDD ),(x _FJ ,y _FJ ),(x _FTX ,y _FTX ) ,(x _FTD ,y _FTD )] is the centroid coordinates of the gastric risk part segmentation image, [S _FDX , S _FDD , S _FJ , S _FTX , S _FTD ] is the area of the risk part, list _d = [d _FDX , d _FDD , d _FJ , d _FTX , d _FTD ] are the distances between the intestinalization segmentation image and each risk site;

According to the formula:

The morphological feature quantification value label10 is obtained, where W _C , H _C are the width and height of the smallest circumscribed rectangle of the intestinalization segmentation image, (x _C , y _C ) are the centroid coordinates of the intestinalization segmentation image, r _C is the intestinalization Radius of the minimum circumscribed circle of the segmented image, n _f is the number of non-intestinalized segmentation regions in the minimum circumscribed circle, s _fi is the area of non-intestinalized segmented regions, (x _fi , y _fi ) is the center of the minimum circumscribed circle of the intestinalized segmented image.

In some embodiments, the third feature quantization value is input into a trained machine learning classifier for classification to obtain the classification result of gastroscopic image intestinalization risk level, wherein the classifier includes a feature fitting sub-network and a classification sub-network;

The obtaining the classification result of gastroscopic image intestinal metaplasia risk level comprises the following steps:

Using the feature fitting sub-network to perform fitting processing on the feature quantization value of the tag attribute to obtain a determination coefficient;

Based on the determination coefficient, the classification sub-network is used for analysis to obtain the identification result.

In view of the above problems, the second aspect of the present invention provides a device for predicting the risk of progression of gastrointestinal metaplasia, which can effectively improve the identification efficiency and accuracy of intestinal metaplasia risk level.

For achieving above object, the technical scheme that the present invention takes is:

A device for predicting the risk of progression of gastrointestinal metaplasia, used for:

By obtaining the feature quantification value of the label attribute of the marker segmentation image, the atrophic gastritis segmentation image and the intestinalization segmentation image, the intestinalization level classification of the gastroscopy image is carried out;

The first label attributes of the marker segmentation image include blood vessel attributes, fold attributes, color attributes, and diffuse attributes;

The second label attributes of the atrophic gastritis segmented image include roughness attributes, fluff-like attributes, off-white attributes, and bright blue ridge attributes;

The third label attribute of the intestinalized segmented image includes position attribute and shape attribute.

Some examples include:

Acquisition module, it is used for obtaining gastroscope image;

A segmentation module, which is used to perform marker segmentation on the gastroscope image, and obtain a marker segmentation image;

A feature extraction module, which is used to perform feature extraction of a first label attribute on the marker segmented image, and obtain a first feature quantization value corresponding to the first label attribute;

A feature extraction module, which is also used to perform feature extraction of the second label attribute on the atrophic gastritis segmented image, obtain a second feature quantization value corresponding to the second label attribute, and perform a third label attribute on the intestinalization segmented image feature extraction, and obtain the third feature quantization value corresponding to the third tag attribute;

A generation module, which is used to input the first feature quantization value into the trained machine learning classifier for classification, obtain the classification result of gastroscopic image atrophic gastritis, and input the second feature quantization value into the trained machine learning classifier for classification , to obtain the classification result of intestinalization of the gastroscopic image; input the third feature quantization value into the trained machine learning classifier for classification, and obtain the classification result of the risk level of intestinalization of the gastroscopic image.

In some embodiments, the feature extraction module is used for:

Carrying out the feature extraction of the first label attribute on the segmentation image of the marker, and obtaining the first feature quantization value corresponding to the first label attribute, comprising the following steps:

Carrying out the feature extraction of the first label attribute on the segmentation image of the marker, and obtaining the first feature quantization value corresponding to the first label attribute, comprising the following steps:

Obtain the blood vessel segmentation image in the marker segmentation image;

According to the formula:

Obtain the blood vessel feature quantization value label ₁ , where n _v is the number of blood vessels in the blood vessel segmentation image, s _v is the area of blood vessels in the blood vessel segmentation image, and S is the area of the marker segmentation image;

According to the formula:

Get the wrinkle feature quantization value label ₂ , where s _z is the wrinkle area in the marker region in the marker segmentation image, _nz is the number of folds in the marker region in the marker segmentation image, S is the area of the marker segmentation image, s _f is the wrinkle area in the non-marker region in the marker segmentation image, n _f is the number of folds in the non-marker region in the marker segmentation image;

According to the formula:

Get the color feature quantization value label ₃ , where r _c , g _c , b _c are the average color values of the three channels, S is the area of the marker region in the marker segmentation image, and n _c is the calculation of the remaining list after removing the color close to black average number;

According to the formula:

The diffuse feature quantification value label ₄ is obtained, where M is the number of images collected for the entire gastric inner wall, w _Qi , h _Qi are the width and height of each image, S _i is the marker segmentation of each image and in the connected domain Based on the area to the marker, w _bi , h _bi are the width and height of the marker, and (x _bi , y _bi ) are the centroid coordinates of the marker.

The method for predicting the risk of progression of gastrointestinal metaplasia in the present invention includes the following steps: classifying the intestinal metaplasia grade of the gastroscopic image by acquiring the feature quantification value of the label attribute of the marker segmented image, the atrophic gastritis segmented image and the intestinal metaplasia segmented image; The first label attribute of the marker segmentation image includes blood vessel attribute, fold attribute, color attribute, and diffuse attribute; the second label attribute of the atrophic gastritis segmentation image includes roughness attribute, fluffy attribute, off-white attribute, bright Blue ridge attribute; the third label attribute of the intestinalized segmentation image includes position attribute and shape attribute. The present invention fully considers the influence of multiple characteristic quantification values of different attributes of the gastroscope image on the accuracy and intuition of image processing, and effectively improves the identification efficiency and identification accuracy of intestinal metaplasia risk levels.

Description of drawings

Fig. 1 is a flow chart of a method for predicting the risk of gastrointestinal metaplasia progression in an embodiment of the present invention;

Fig. 2 is an image effect diagram of marker segmentation in the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 3 is a display diagram of non-atrophic gastritis/atrophic gastritis in the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 4 is a display diagram of atrophic gastritis/intestinal metaplasia in the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 5 is an effect diagram of blood vessel extraction of the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 6 is an extraction effect diagram of color principal components of the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 7 is a villi-like segmentation effect diagram of the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present invention;

Fig. 8 is a diagram showing bright blue ridges of the method for predicting the risk of progression of gastrointestinal metaplasia in the embodiment of the present application.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments It is a part of the embodiments of this application, but not all of them. Based on the embodiments in the present application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present application.

Referring to Figure 1, the first party of the embodiment of the present invention provides a method for predicting the risk of gastrointestinal metaplasia progression, which includes the following steps:

S1. Obtain a gastroscope image, perform marker segmentation on the gastroscope image, and acquire a marker segmented image;

Wherein, the gastroscope image is acquired, and the gastroscope image can be a white light image, a stained image, or a stained magnified image, and the gastroscope image is an electronic gastroscope image of the stomach captured by an electronic endoscope. Specifically, the gastroscope image may be collected through a gastroscope, or the gastroscope image may be acquired from an image library pre-stored in the memory of the computer device.

Marker segmentation is performed on the gastroscope image to obtain a marker segmented image, as shown in FIG. 2 .

Specifically, the gastroscope image and the marker region are used as sample images, and the marker segmentation model is pre-trained, for example, image segmentation network models such as Unet++ and Mask-RCNN are selected. In a specific embodiment, the gastroscope image is used as the segmentation after training. The input of the model, the output of the segmentation model is the landmark segmentation image. It can be understood that in this embodiment, the segmentation map of the marker is obtained, so as to subsequently perform multiple label attribute feature quantification on the segmented image of the marker, so as to improve the accuracy of quantification.

S2. Perform feature extraction of the first label attribute on the marker segmented image, obtain the first feature quantization value corresponding to the first label attribute, input the first feature quantization value into the trained machine learning classifier for classification, and obtain a gastroscope Image classification results of atrophic gastritis;

Perform feature extraction of the first label attribute on the marker segmentation image, and obtain a first feature quantization value corresponding to the first label attribute, wherein the first label attribute refers to multiple attributes of the marker segmentation image, for example, marker segmentation For the blood vessel attribute, fold attribute, color attribute, diffuse attribute, etc. of the image, the first feature quantization value refers to the quantization value corresponding to the feature of each first label attribute.

Specifically, the feature extraction method is used to extract the features of the image of the target area to obtain the first feature quantization value, wherein the feature extraction method can be an artificial feature extraction method combined with an algorithm based on image feature analysis, such as pixel neighborhood mean calculation, maximum pixel value Extraction, etc., to calculate the first feature quantization value, can also be a feature extraction method of deep learning, such as convolutional neural network CNN, UNet++, etc., which can be selected according to the characteristics of the first label attribute, and there is no limitation here. In this embodiment, by performing feature extraction on the marker segmented image and obtaining the corresponding first feature quantization value, the quantitative calculation of the features of each first label attribute of the marker segmented image is realized, making the feature quantization value more comprehensive and rich , so that subsequent accurate and intuitive image analysis and recognition can be performed based on the plurality of first characteristic quantification values, and the accuracy rate of recognition of atrophic gastritis is improved.

In some embodiments, each first feature quantization value is input into a trained machine learning classifier for classification, and a classification result of atrophic gastritis is obtained. Wherein, the trained machine learning classifier can be realized by learning a machine learning algorithm model with classification ability through samples, and the machine learning classifier in this embodiment is used to divide different first feature value sets into non-atrophic gastritis or atrophic gastritis A class in the result. Specifically, at least one machine learning model can be used to classify the classifier. The machine learning model can be one or more of the following: neural network (for example, convolutional neural network, BP neural network, etc.), logistic regression model, support vector machine, decision tree, random forest, perceptron and other machine learning Model. As part of the training of such a machine learning model, the training input is each first feature quantization value, for example, blood vessel attributes, fold attributes, color attributes, diffuse attributes, etc., through training, the first feature value set and the gastroscope to be identified are established. The classifier of the image atrophic gastritis response relationship enables the preset classifier to have the ability to judge whether the classification result corresponding to the gastroscope image to be recognized is the result of non-atrophic gastritis or atrophic gastritis. In this embodiment, the classifier is a binary classifier, that is, two classification results are obtained, that is, the result of non-atrophic gastritis or the result of atrophic gastritis, as shown in FIG. 3 . It can be understood that in this embodiment, the influence of the quantified values of the multiple attributes of the atrophic gastritis segmented image on the accuracy and intuitiveness of the image processing is fully considered. Extracting features with richer information and quantifying and comprehensively processing multiple features with different attributes improves the rationality of feature value quantification. Compared with traditional methods that only consider a single feature information and a single statistical comparison method, it greatly improves Atrophic gastritis identification efficiency.

S3. Perform feature extraction of the second label attribute on the segmented image of atrophic gastritis, obtain a second feature quantization value corresponding to the second label attribute, and input the second feature quantization value into a trained machine learning classifier for classification , to obtain the intestinalization classification result of the gastroscope image;

In some embodiments, the feature extraction of multiple second label attributes is performed on the segmented image of atrophic gastritis, and the second feature quantization values corresponding to the multiple second label attributes are obtained. Wherein, the second label attribute refers to multiple attributes of the atrophic gastritis segmented image, for example, the roughness attribute, fluffy attribute, off-white attribute, bright blue ridge attribute, etc. of the atrophic gastritis segmented image, and the second feature quantization value is Refers to the quantized value corresponding to the feature of each second label attribute.

Specifically, the feature extraction method is used to extract the features of the image of the target area to obtain the second feature quantization value, wherein the feature extraction method can be an artificial feature extraction method combined with an algorithm based on image feature analysis, such as pixel neighborhood mean value calculation, maximum pixel value Extraction, etc., to calculate the second feature quantization value, can also be a feature extraction method of deep learning, such as convolutional neural network CNN, UNet++, etc., which can be selected according to the characteristics of the second label attribute, which is not limited here. In this embodiment, by performing feature extraction on the segmented image of atrophic gastritis and obtaining the corresponding second feature quantization value, the quantitative calculation of the features of each second label attribute of the segmented image of atrophic gastritis is realized, so that the feature quantization value is more accurate. It is comprehensive and enriched, so that accurate and intuitive image analysis and recognition can be performed subsequently based on the plurality of second characteristic quantification values, and the accuracy rate of intestinalization recognition is improved.

In some embodiments, each second feature quantization value is input into a trained machine learning classifier for classification, and an intestinalized classification result is obtained. Wherein, the trained machine learning classifier can be realized by learning a machine learning algorithm model capable of classification through samples, and the machine learning classifier in this embodiment is used to divide different second feature value sets into non-intestinalized or intestinalized results of a class. Specifically, at least one machine learning model can be used to classify the classifier. The machine learning model can be one or more of the following: neural network (for example, convolutional neural network, BP neural network, etc.), logistic regression model, support vector machine, decision tree, random forest, perceptron and other machine learning Model. As part of the training of such a machine learning model, the training input is each second feature quantization value, for example, roughness attribute, fluffy attribute, off-white attribute, bright blue ridge attribute, etc. Through training, a second feature value set is established The classifier corresponding to the intestinalization of the gastroscope image to be recognized enables the preset classifier to have the ability to determine whether the classification result corresponding to the gastroscope image to be recognized is a result of non-intestinalization or intestinalization. In this embodiment, the classifier is a binary classifier, that is, two classification results are obtained, that is, non-intestinalized or intestinalized results, as shown in FIG. 4 . It can be understood that in this embodiment, the influence of the quantified values of the multiple attributes of the atrophic gastritis segmented image on the accuracy and intuitiveness of the image processing is fully considered. Extracting features with richer information and quantifying and comprehensively processing multiple features with different attributes improves the rationality of feature value quantification. Compared with traditional methods that only consider a single feature information and a single statistical comparison method, it greatly improves Enterochemical recognition efficiency.

S4. Perform feature extraction of the third label attribute on the intestinalized segmented image, obtain a third feature quantization value corresponding to the third label attribute, and input the third feature quantization value into a trained machine learning classifier for classification, The risk classification results of intestinal metaplasia in gastroscopy images were obtained.

In some embodiments, feature extraction of multiple third label attributes is performed on the intestinalization segmentation image, and third feature quantization values corresponding to the multiple third label attributes are obtained.

Wherein, the third label attribute refers to a plurality of attributes of the intestinalized segmented image, for example, the position attribute, shape attribute, etc. of the intestinalized segmented image, and the third feature quantization value refers to the quantized value corresponding to the feature of each third label attribute.

Specifically, the feature extraction method is used to extract the features of the image of the target area to obtain the third feature quantization value, wherein the feature extraction method can be an artificial feature extraction method combined with an algorithm based on image feature analysis, such as the calculation of the mean value of the pixel neighborhood, the maximum pixel value Extraction, etc., to calculate the third feature quantization value, can also be a feature extraction method of deep learning, such as convolutional neural network CNN, UNet++, etc., which can be selected according to the characteristics of the third label attribute, and there is no limitation here. In this embodiment, by performing feature extraction on the intestinalization segmentation image and obtaining the corresponding third feature quantization value, the quantitative calculation of the features of each third label attribute of the intestinalization segmentation image is realized, making the feature quantization value more comprehensive and rich , so that subsequent accurate and intuitive image analysis and identification can be performed based on the plurality of third characteristic quantification values, and the accuracy rate of identification of intestinalization risk level is improved.

Each third feature quantization value is input into a trained machine learning classifier for classification, and an intestinalized classification result is obtained.

Among them, the trained machine learning classifier can be realized by learning a machine learning algorithm model with classification capabilities through samples. The machine learning classifier in this embodiment is used to divide different feature value sets into low-risk or high-risk results. . Specifically, at least one machine learning model can be used to classify the classifier. The machine learning model can be one or more of the following: neural network (for example, convolutional neural network, BP neural network, etc.), logistic regression model, support vector machine, decision tree, random forest, perceptron and other machine learning Model. As part of the training of such a machine learning model, the training input is each third feature quantization value, for example, position attribute, morphological attribute, etc., and through training, the third feature value set is established to correspond to the intestinalization risk level of the gastroscope image to be identified The classifier of the relationship makes the preset classifier capable of judging whether the classification result corresponding to the gastroscope image to be recognized is a low-risk or high-risk result. In this embodiment, the classifier is a binary classifier, that is, two classification results are obtained, that is, low-risk or high-risk results. It can be understood that in this embodiment, the quantified values of the multiple attributes of the intestinalized segmented image and the influence of the quantized values of the multiple attributes of the segmented image with the intestinalized segmented image on the accuracy and intuitiveness of the image processing are fully considered. By extracting the information Quantify and comprehensively process multiple features with different attributes, which improves the rationality of feature value quantification. Compared with the traditional method of only considering a single feature information and a single statistical comparison method, it greatly improves the quality Risk level identification efficiency.

The above embodiment provides a method for predicting the risk of gastrointestinal metaplasia progression. First, obtain and perform marker segmentation based on the gastroscope image to be recognized to obtain a marker segmentation image, and then acquire multiple first feature quantifications corresponding to the marker segmentation image. value, and then input multiple first feature quantization values into the trained machine learning classifier to obtain the recognition result of atrophic gastritis, obtain multiple second feature quantization values of atrophic gastritis segmented image again, and again multiple second feature quantization values Input the trained machine learning classifier to obtain the intestinalization identification result, obtain multiple third feature quantization values of the intestinalization segmentation image again, and input multiple third feature quantization values into the trained machine learning classifier again to obtain the intestinalization risk level Recognition results; this embodiment fully considers the impact of the quantified values of multiple attributes of the gastroscope image on the accuracy and intuition of image processing, effectively improving the identification efficiency and accuracy of the risk level of intestinal metaplasia.

In one embodiment, the multiple first label attributes include blood vessel attributes, fold attributes, color attributes, and diffuse attributes; the feature extraction of multiple first label attributes is performed on the landmark segmentation image, and the first label attributes corresponding to each first label attribute are obtained. A step of feature quantification, comprising: using a preset blood vessel quantification method to determine the blood vessel quantification result of the marker segmentation image; using a preset wrinkle quantification method to determine the marker segmentation image wrinkle quantification result; using a preset color quantification method to determine the marker The color quantification result of the object segmentation image; the diffuse quantification result of the marker segmentation image is determined using a preset diffuse quantification method.

Among them, microvessels can hardly be observed on the surface of normal gastric mucosa under white light, but when diseases such as inflammation occur on the surface of gastric mucosa, submucosal blood vessels can be clearly displayed.

Specifically, the step of obtaining blood vessels in the marker segmentation image includes:

Convert the marker segmentation image into a grayscale image and perform binarization. The binarization threshold is τ, and τ∈(0,255) can be determined according to the actual situation, and there is no specific limitation here;

Call the opencv toolkit cv2.erode() to corrode the binarized marker segmentation image to obtain the marker segmentation image mask image;

Call the opencv toolkit cv2.medianBlur() to perform median filtering on the grayscale image of the marker segmentation image;

Call the opencv toolkit cv2.createCLAHE() to perform histogram equalization on the marker segmentation image after the median filtering;

Perform gamma transformation on the marker segmented image after the histogram equalization, the gamma transformation formula is O(x, y)=I(x, y) ^γ , where O(x, y) is the gamma transformation After the figure, I (x, y) is the original figure, and γ=0.5 among the present invention;

Call the opencv toolkit cv2.filter2D() to perform a convolution operation on the gamma-transformed marker segmentation image;

Comparing the convolutional marker segmentation image with the marker segmentation image mask image bit by bit, if a pixel value in the mask image is 0, the convolutional marker segmentation image is here Set the pixel value at 0 to obtain the denoised marker segmentation image;

Contrast stretching is performed on the denoised landmark segmentation image to obtain the blood vessel segmentation image corresponding to the landmark segmentation image, as shown in Figure 5.

其中S为标志物分割图像面积，标志物分割时候可得到。Specifically, the blood vessel area in the blood vessel segmentation image is obtained on the basis of the connected domain as s _v and the number of blood vessels n _v , then the blood vessel quantization value is

Among them, S is the area of the marker segmentation image, which can be obtained when the marker is segmented.

Among them, there are folds in the normal gastric mucosa, but when there are diseases such as inflammation on the surface of the gastric mucosa, the mucosal folds will flatten or even disappear.

Specifically, the folds in the gastroscope image are segmented by using the trained fold segmentation model. The fold segmentation model can be an image segmentation network model such as Unet++, Mask-RCNN, etc., and on the basis of the connected domain, calculate the If the fold area is s _z , the number of folds is n _z , the area of the marker area is S, the area of folds in the non-marker area is s _f , and the number of folds is n _f , then the quantitative value of the fold is

Among them, the gastric mucosa is red or white or red and white, indicating that the gastric mucosa is abnormal, and there is a high probability of atrophic gastritis.

其中，S为标志物区域面积。Specifically, the color principal component of the marker segmentation image is extracted through the getcolors() method that comes with PIL, as shown in Figure 6, in a specific embodiment, when the marker segmentation image is used for color principal component extraction, the output color-number Corresponding list [(14757,(193,87,73)),(12541,(176,85,69)),(25072,(175,65,58)),(11435,(211,110,91)),( 16153,(195,96,79)),(15781,(179,76,64)),(3488,(116,48,42)),(4308,(5,0,2)),(1920, (0,0,3)),(156689,(0,0,0))], calculate the average number of remaining lists after removing colors close to black (color _r ≤20, color _g ≤20, color _b ≤20) is n _c , the average color of the three channels is r _c , g _c , b _c , then the color quantization value is

Among them, S is the area of the marker area.

Among them, atrophic gastritis spreads over the entire surface of the gastric mucosa, and may spread to the entire inner surface of the stomach over time.

Specifically, a total of M images are collected for the entire gastric inner wall, and the width and height of each image are w _Qi , h _Qi , respectively. Each image is segmented by markers and the area of the markers is calculated on the basis of the connected domain is S _i , the width and height of the marker are w _bi , h _bi , and the centroid coordinates of the marker are (x _bi , y _bi ), then the quantitative value of the diffuse degree is

In one embodiment, the trained machine learning classifier includes a feature fitting sub-network and a classification sub-network; each first feature quantization value is input into the trained machine learning classifier for classification, and the result of the classification of gastroscopic image atrophic gastritis is obtained. The steps include: using a feature fitting sub-network to perform fitting processing on each first feature quantization value to obtain a determination coefficient; based on the determination coefficient, use a classification sub-network to analyze to obtain a classification result.

分类结果为非萎缩性胃炎、萎缩性胃炎结果。Specifically, the fitting process is carried out through the first quantized value of the feature fitting sub-network, and the corresponding weight of the fitting process is performed according to the first quantized value of the fitting result. Take a feature quantization value label ₁ ~ label ₄ as an example, use decision tree, random forest, etc. to determine label ₁ ~ label ₄ , and the corresponding weights are λ ₁ , λ ₂ , λ ₃ , λ ₄ , then the fusion feature value at this time is :

The classification results are the results of non-atrophic gastritis and atrophic gastritis.

In this embodiment, the information features of the gastroscope image are enriched and the quantification is more accurate by fusion calculation of each first feature quantization value, which improves the recognition accuracy and recognition efficiency of atrophic gastritis in the gastroscope image.

In one embodiment, the plurality of second label attributes include roughness attribute, fuzzy attribute, off-white attribute, and bright blue ridge attribute; the feature extraction of a plurality of second label attributes is performed on the segmented image of atrophic gastritis, and each first label attribute is obtained. The step of the second feature quantification value corresponding to the two-label attribute includes: using a preset roughness feature quantification method to determine the roughness feature quantification result of the atrophic gastritis segmentation image; using a preset villi-like feature quantization method to determine the atrophic gastritis segmentation Quantification results of fluff-like features in the image; using the preset whitening quantification method to determine the whitening quantification result of the segmented image of atrophic gastritis; using the preset bright blue ridge quantification method to determine the bright blue ridge quantification result of the segmented image of atrophic gastritis.

Among them, from atrophic gastritis to intestinal metaplasia, the surface of the gastric mucosa will gradually become rough.

其中，W_W，H_W分别为萎缩性胃炎分割图像的宽和高，P_mean为萎缩性胃炎分割图像的平均像素值，

img_W为萎缩性胃炎分割图像，W₀为按照某一设定的阈值将萎缩性胃炎分割图进行二值化后方差最大所在行，0＜W₀＜W_W。Specifically, the surface roughness feature quantization value is

Among them, W _W , H _W are the width and height of the segmented image of atrophic gastritis respectively, P _mean is the average pixel value of the segmented image of atrophic gastritis,

img _W is the segmented image of atrophic gastritis, W ₀ is the row where the variance is the largest after binarizing the segmented image of atrophic gastritis according to a set threshold, 0<W ₀ <W _W .

Among them, the gastric mucosa after intestinalization produced villi, as shown in FIG. 7 .

Specifically, the villi-like segmentation model in the atrophic gastritis segmentation image is segmented using the villi-like segmentation model that has been trained. The villi-like segmentation model can be an image segmentation network model such as Unet++, Mask-RCNN, etc. Separately extract the fluff-like region and obtain the area s _ri of the fluff-like region, the number N _r of the fluff-like region, and the width and height w ri and _height of the smallest circumscribed rectangle of the fluff-like region, _and then use corner detection to obtain the value of the fluff-like region corner point and interrupt at the corner point to obtain multiple fluff segments, the number of fluff segments is n _ri , then the fluff-like feature quantization value is

Among them, the color of gastric mucosa after intestinal transformation is white.

Specifically, the colors in the segmented image of atrophic gastritis are clustered by the color clustering method. In a specific embodiment, the k-means clustering method can be used, and there is no specific limitation here, and each pixel of the category with the largest number of colors is obtained pixel value (r _bi , g _bi , b _bi ) and the maximum number of pixel points n _b , then the off-white quantization value is

Among them, the gastric mucosa after intestinalization may appear bright blue ridges in the state of staining and amplification, as shown in Figure 8.

W_L，H_L分别为亮蓝脊分割图像的宽和高，img_L为亮蓝脊分割图像，则亮蓝脊量化值为/>

其中W_W，H_W分别为萎缩性胃炎分割图像最小外接矩形的宽和高，/>

为标准亮蓝脊三通道平均像素值，在一个具体实施例中/>

Specifically, perform atrophic gastritis segmentation on the stained and enlarged image of the gastroscope to obtain the segmented image of the stained and enlarged atrophic gastritis, and obtain the area of the stained and enlarged atrophic gastritis S _RS and the centroid of the stained and enlarged atrophic gastritis segmentation image (x _RS , y _RS ), Use the trained bright blue ridge segmentation model to segment the villi in the atrophic gastritis segmentation image and obtain the bright blue ridge segmentation image and the area S _RSL of the bright blue ridge segmentation image and the centroid of the bright blue ridge segmentation image (x _RSL , y _RSL ), the bright blue ridge segmentation model can be image segmentation network models such as Unet++, Mask-RCNN, obtains the average pixel value of the bright blue ridge segmentation image

W _L , H _L are the width and height of the bright blue ridge segmented image respectively, img _L is the bright blue ridge segmented image, then the quantized value of the bright blue ridge is />

Where W _W , H _W are the width and height of the smallest circumscribed rectangle of the segmented image of atrophic gastritis respectively, />

is the average pixel value of the standard bright blue ridge three channels, in a specific embodiment />

In one embodiment, the trained machine learning classifier includes a feature fitting sub-network and a classification sub-network; each second feature quantization value is input into the trained machine learning classifier for classification, and the step of obtaining the intestinalization classification result of the gastroscope image , including: using a feature fitting sub-network to perform fitting processing on each second feature quantization value to obtain a determination coefficient; based on the determination coefficient, using a classification sub-network to perform analysis to obtain a classification result.

分类结果为非肠化、肠化结果。Specifically, the fitting process is performed by performing the fitting process on each second feature quantization value of the feature fitting sub-network, and the corresponding weights of the fitting process are performed on each second feature quantization value according to the fitting result, and continue to use the first Take two-feature quantization values label ₅ ~ label ₈ as an example, use decision tree, random forest, etc. to determine label ₅ ~ label ₈ , and the corresponding weights are λ ₅ , λ ₆ , λ ₇ , λ ₈ , then the fusion feature value at this time is :

The classification result is the result of non-intestinalization and intestinalization.

In this embodiment, the information features of the gastroscope image are enriched and the quantification is more accurate by fusion calculation of each second feature quantization value, which improves the recognition accuracy and recognition efficiency of gastroscope image intestinalization.

In one embodiment, the plurality of third tag attributes include position attribute and morphological attribute; the feature extraction of a plurality of third tag attributes is performed on the segmented image of atrophic gastritis, and the third feature quantization value corresponding to each third tag attribute is obtained. The steps include: using a preset position quantification method to determine the position quantification result of the intestinalization segmentation image; using a preset morphological quantification method to determine the morphology quantification result of the intestinalization segmentation image.

Among them, when intestinal metaplasia occurs in the lesser curvature of the gastric antrum, the greater curvature of the gastric antrum, the gastric angle, the lesser curvature of the gastric body, and the greater curvature of the gastric body, the risk factor is higher.

Specifically, use the trained gastric risk part segmentation model to segment the gastric risk part in the gastroscope image and obtain the centroid of the gastric risk part segmentation image [(x _FDX ,y _FDX ),(x _FDD ,y _FDD ),(x _FJ , y _FJ ), (x _FTX , y _FTX ), (x _FTD , y _FTD )] and risk area [S _FDX , S _FDD , S _FJ , S _FTX , S _FTD ], the gastric risk area segmentation model can be If it is an image segmentation network model such as Unet++, Mask-RCNN, etc., the distance between the intestinalization segmentation image and each risk part is list _d = [d _FDX ,d _FDD ,d _FJ ,d _FTX ,d _FTD ], where the distance calculation formula is intestinalization The Euclidean distance between the centroid of the segmented image and the centroid of the segmented image of each risk part, then the intersection-area ratio of the distance between the segmented image and each risk part is

Where S _C is the area of the intestinalized segmented image, and the position quantization value is

Among them, the more complex the shape of the intestinalization segmentation image, the greater the risk factor.

其中W_C，H_C分别为肠化分割图像最小外接矩形的宽和高。Specifically, the centroid (x _C , y _C ) of the intestinalization segmentation image is obtained on the basis of the connected domain, the minimum circumscribed circle and the minimum circumscribed circle radius r _C of the intestinalization segmentation image are obtained, and the non-intestinalization segmentation in the minimum circumscribed circle is obtained The number of regions n _f and the area of non-intestinalized segmentation region s _fi and centroid (x _fi , y _fi ), then the morphological quantization value is

Among them, W _C and H _C are the width and height of the smallest circumscribed rectangle of the intestinalization segmentation image, respectively.

In one embodiment, the trained machine learning classifier includes a feature fitting sub-network and a classification sub-network; each of the third feature quantization values is input into the trained machine learning classifier for classification, and the gastroscope image intestinalization risk level classification result is obtained The step includes: using a feature fitting sub-network to perform fitting processing on each third feature quantization value to obtain a determination coefficient; based on the determination coefficient, using a classification sub-network to perform analysis to obtain a classification result.

分类结果为低风险、高风险结果。Specifically, the fitting process is performed through each third feature quantization value of the feature fitting sub-network, and the corresponding weight of the fitting process is performed according to each third feature quantization value of the fitting result, continuing to use the first Take the three-feature quantization value label ₉ ~ label ₁₀ as an example, use decision tree, random forest, etc. to determine label ₉ ~ label ₁₀ , and the corresponding weights are λ ₉ , λ ₁₀ respectively, then the fusion feature value at this time is:

Classification results are low risk and high risk results.

The method for predicting the risk of progression of gastrointestinal metaplasia in the present invention includes the following steps: classifying the intestinal metaplasia grade of the gastroscopic image by acquiring the feature quantification value of the label attribute of the marker segmented image, the atrophic gastritis segmented image and the intestinal metaplasia segmented image; The first label attribute of the marker segmentation image includes blood vessel attribute, fold attribute, color attribute, and diffuse attribute; the second label attribute of the atrophic gastritis segmentation image includes roughness attribute, fluffy attribute, off-white attribute, bright Blue ridge attribute; the third label attribute of the intestinalized segmentation image includes position attribute and shape attribute. The present invention fully considers the influence of multiple characteristic quantification values of different attributes of the gastroscope image on the accuracy and intuition of image processing, and effectively improves the identification efficiency and identification accuracy of intestinal metaplasia risk levels.

The second aspect of the embodiment of the present invention provides a risk prediction device for the progression of gastrointestinal metaplasia, which is used for:

By obtaining the feature quantification value of the label attribute of the marker segmentation image, the atrophic gastritis segmentation image and the intestinalization segmentation image, the intestinalization level classification of the gastroscopy image is carried out;

The first label attributes of the marker segmentation image include blood vessel attributes, fold attributes, color attributes, and diffuse attributes;

The second label attributes of the atrophic gastritis segmented image include roughness attributes, fluff-like attributes, off-white attributes, and bright blue ridge attributes;

The third label attribute of the intestinalized segmented image includes position attribute and shape attribute.

Some examples include:

Acquisition module, it is used for obtaining gastroscope image;

A segmentation module, which is used to perform marker segmentation on the gastroscope image, and obtain a marker segmentation image;

A feature extraction module, which is used to perform feature extraction of a first label attribute on the marker segmented image, and obtain a first feature quantization value corresponding to the first label attribute;

A feature extraction module, which is also used to perform feature extraction of the second label attribute on the atrophic gastritis segmented image, obtain a second feature quantization value corresponding to the second label attribute, and perform a third label attribute on the intestinalization segmented image feature extraction, and obtain the third feature quantization value corresponding to the third tag attribute;

A generation module, which is used to input the first feature quantization value into the trained machine learning classifier for classification, obtain the classification result of gastroscopic image atrophic gastritis, and input the second feature quantization value into the trained machine learning classifier for classification , to obtain the classification result of intestinalization of the gastroscopic image; input the third feature quantization value into the trained machine learning classifier for classification, and obtain the classification result of the risk level of intestinalization of the gastroscopic image.

In some embodiments, the feature extraction module is used for:

Carrying out the feature extraction of the first label attribute on the segmentation image of the marker, and obtaining the first feature quantization value corresponding to the first label attribute, comprising the following steps:

Obtain the blood vessel segmentation image in the marker segmentation image;

According to the formula:

Obtain the blood vessel feature quantization value label ₁ , where n _v is the number of blood vessels in the blood vessel segmentation image, s _v is the area of blood vessels in the blood vessel segmentation image, and S is the area of the marker segmentation image;

According to the formula:

Get the wrinkle feature quantization value label ₂ , where s _z is the wrinkle area in the marker region in the marker segmentation image, _nz is the number of folds in the marker region in the marker segmentation image, S is the area of the marker segmentation image, s _f is the wrinkle area in the non-marker region in the marker segmentation image, n _f is the number of folds in the non-marker region in the marker segmentation image;

According to the formula:

Get the color feature quantization value label ₃ , where r _c , g _c , b _c are the average color values of the three channels, S is the area of the marker region in the marker segmentation image, and n _c is the calculation of the remaining list after removing the color close to black average number;

According to the formula:

The diffuse feature quantification value label ₄ is obtained, where M is the number of images collected for the entire gastric inner wall, w _Qi , h _Qi are the width and height of each image, S _i is the marker segmentation of each image and in the connected domain Based on the area to the marker, w _bi , h _bi are the width and height of the marker, and (x _bi , y _bi ) are the centroid coordinates of the marker.

The blood vessel segmentation image, the obtaining step includes:

Binarize the marker segmentation image into a grayscale image;

Call the opencv toolkit cv2.erode() to corrode the binarized marker segmentation image to obtain the marker segmentation image mask image;

Call the opencv toolkit cv2.medianBlur() to perform median filtering on the grayscale image of the marker segmentation image;

Call the opencv toolkit cv2.createCLAHE() to perform histogram equalization on the obtained median-filtered marker segmentation image;

performing gamma transformation on the marker segmented image after the histogram equalization;

Call the opencv toolkit cv2.filter2D() to perform a convolution operation on the obtained gamma-transformed marker segmentation image;

Comparing the convolutional marker segmentation image with the marker segmentation image mask image bit by bit, if a pixel value in the mask image is 0, the convolutional marker segmentation image is here Set the pixel value at 0 to obtain the denoised marker segmentation image;

Contrast stretching is performed on the denoised marker segmented image to obtain a blood vessel segmented image corresponding to the marker segmented image.

In some embodiments, the feature extraction module is also used to: perform feature extraction of the second label attribute on the atrophic gastritis segmented image, and obtain a second feature quantization value corresponding to the second label attribute, including the following steps:

According to the formula:

Get the roughness feature quantization value label ₅ , where W _W , H _W are the width and height of the segmented image of atrophic gastritis respectively, P _mean is the average pixel value of the segmented image of atrophic gastritis, img _W is the segmented image of atrophic gastritis, W ₀ is the row with the largest variance after binarizing the atrophic gastritis segmentation map according to a set threshold, 0<W ₀ <W _W ;

According to the formula:

Get the fluff-like feature quantization value label ₆ , where N _r is the number of fluff-like regions, s _ri is the area of the fluff-like regions, w _ri , h _ri are the width and height of the smallest circumscribed rectangle of the fluff-like regions, and n _ri is the fluff-like regions and break at the corners to obtain the number of fluff segments;

According to the formula:

Obtain the partial white feature quantization value label ₇ , wherein n _b is the maximum number of pixel points, and (r _bi , g _bi , b _bi ) the pixel value of each pixel point of the category with the largest number of colors;

According to the formula:

是标准亮蓝脊三通道平均像素值，PL是亮蓝脊分割图像平均像素值。Get the bright blue ridge feature quantization value label ₈ , where W _L , _HL are the width and height of the bright blue ridge segmented image, img _L is the bright blue ridge segmented image, (x _RSL , y _RSL ) is the bright blue ridge segmented image Centroid coordinates, S _RSL is the bright blue ridge segmented image area, (x _RS , y _RS ) is the centroid coordinates of the stained and enlarged atrophic gastritis segmented image, S _RS is the area of the stained and enlarged atrophic gastritis,

is the average pixel value of the three channels of the standard bright blue ridge, and PL is the average pixel value of the bright blue ridge segmented image.

In some embodiments, the feature extraction module is also used to: perform feature extraction of the third label attribute on the intestinalized segmented image, and obtain a third feature quantization value corresponding to the third label attribute, including the following steps:

According to the formula:

Get the location feature quantization value label ₉ , where S _C is the intestinalization segmented image area, [(x _FDX ,y _FDX ),(x _FDD ,y _FDD ),(x _FJ ,y _FJ ),(x _FTX ,y _FTX ), (x _FTD , y _FTD )] are the centroid coordinates of the gastric risk part segmentation image, [S _FDX , S _FDD , S _FJ , S _FTX , S _FTD ] are the risk part area, list _d = [d _FDX , d _FDD ,d _FJ ,d _FTX ,d _FTD ] are the distances between the intestinalization segmentation image and each risk site;

According to the formula:

The quantitative value label ₁₀ of the morphological features is obtained, where W _C , H _C are the width and height of the smallest circumscribed rectangle of the intestinalization segmentation image, (x _C , y _C ) are the centroid coordinates of the intestinalization segmentation image, and r _C is the intestinalization segmentation image The radius of the minimum circumscribed circle of the segmented image, n _f is the number of non-intestinalized segmentation regions in the minimum circumscribed circle, s _fi is the area of the non-intestinalized segmented region, (x _fi , y _fi ) is the center of the minimum circumscribed circle of the intestinalized segmented image.

In some embodiments, the production module is used to input the third feature quantization value into the trained machine learning classifier for classification, and obtain the classification result of gastroscopic image intestinalization risk level, wherein the classifier includes a feature fitting sub-network and classification subnetwork;

The obtaining the classification result of gastroscopic image intestinal metaplasia risk level comprises the following steps:

Using the feature fitting sub-network to perform fitting processing on the feature quantization value of the tag attribute to obtain a determination coefficient;

Based on the determination coefficient, the classification sub-network is used for analysis to obtain the identification result.

In the description of the present application, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower" and so on is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description. It is not intended to indicate or imply that the device or element referred to must have a particular orientation, be constructed, or operate in a particular orientation, and thus should not be construed as limiting the application. Unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be interpreted in a broad sense, for example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection, It can also be an electrical connection; it can be a direct connection, or an indirect connection through an intermediary, or an internal communication between two components. Those of ordinary skill in the art can understand the specific meanings of the above terms in this application according to specific situations.

It should be noted that in this application, relative terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply There is no such actual relationship or order between these entities or operations. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

The above descriptions are only specific implementation manners of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the present application will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims (10)

1. A method of predicting risk of progression of gastrointestinal epithelial metaplasia, comprising the steps of:

performing gastroscopic image intestinal grade classification by acquiring characteristic quantification values of tag attributes of the marker segmentation image, the atrophic gastritis segmentation image and the intestinal segmentation image;

the first label attribute of the marker segmentation image comprises a blood vessel attribute, a fold attribute, a color attribute and a diffuse attribute;

the second label attribute of the atrophic gastritis segmented image comprises a roughness attribute, a villus attribute, a off-white attribute and a bright blue ridge attribute;

the third tag attribute of the enteric segmented image comprises a position attribute and a morphology attribute.

2. The method for predicting the risk of progression of gastrointestinal epithelial metaplasia according to claim 1, wherein the step of classifying the gastroscopic image enteron grade by obtaining the feature quantization value of the tag attribute of the marker segmented image, the atrophic gastritis segmented image and the enteron segmented image comprises the steps of:

obtaining a gastroscope image, and carrying out marker segmentation on the gastroscope image to obtain a marker segmentation image;

extracting the features of the first tag attributes from the marker segmented image, obtaining a first feature quantized value corresponding to the first tag attributes, inputting the first feature quantized value into a trained machine learning classifier for classification, and obtaining a classification result of the gastroscopic image atrophic gastritis;

extracting features of second tag attributes from the atrophic gastritis segmented image, obtaining second feature quantized values corresponding to the second tag attributes, inputting the second feature quantized values into a trained machine learning classifier for classification, and obtaining gastroscope image enteric classification results;

and carrying out feature extraction of a third tag attribute on the enteron segmented image, obtaining a third feature quantized value corresponding to the third tag attribute, inputting the third feature quantized value into a trained machine learning classifier for classification, and obtaining a gastroscope image enteron risk level classification result.

3. The method for predicting the risk of progression of gastrointestinal epithelial metaplasia according to claim 2, wherein the feature extraction of the first tag attribute is performed on the marker-segmented image, and the first feature quantification value corresponding to the first tag attribute is obtained, comprising the steps of:

acquiring a blood vessel segmentation image in the marker segmentation image;

according to the formula:

obtaining a blood vessel characteristic quantification value label ₁ Wherein n is _v Is the number of blood vessels in the blood vessel segmentation image, s _v The blood vessel area in the blood vessel segmentation image is that of the marker segmentation image;

according to the formula:

obtaining a fold characteristic quantization value label ₂ Wherein s is _z Is the fold area, n, in the marker region in the marker-segmented image _z Is the number of folds in the marker region in the marker-segmented image, S is the marker-segmented image area, S _f Is the fold area, n, in the non-marker region in the marker-segmented image _f Is the number of folds in the non-marker region in the marker-segmented image;

according to the formula:

obtaining a color characteristic quantization value label ₃ Wherein r is _c ，g _c ，b _c Is the three-channel average color value, S is the area of the marker region in the marker-partitioned image, n _c The average number is calculated for the rest list after the colors close to black are removed;

According to the formula:

obtaining diffuse characteristic quantized value label ₄ Wherein M is the number of images taken of the entire stomach wall, w _Qi ，h _Qi Is the width and height of each image, S _i The area from each image to the marker is divided by the marker and is based on the connected domain, w _bi ，h _bi Is the width and height of the marker, (x) _bi ,y _bi ) Is the centroid coordinates of the marker.

4. A method of predicting the risk of progression of gastrointestinal epithelial metaplasia as set forth in claim 3 wherein said acquiring a vessel segmentation image from a marker segmentation image comprises:

converting the marker segmentation image into a gray level image and then binarizing the gray level image;

performing corrosion operation on the binarized marker segmentation image to obtain a marker segmentation image mask image;

carrying out median filtering on the gray level image of the marker segmentation image;

performing histogram equalization on the obtained marker segmentation image after median filtering;

performing gamma conversion on the marker segmentation image subjected to histogram equalization;

carrying out convolution operation on the obtained marker segmentation image after gamma conversion;

bit-by-bit comparison is carried out on the convolved marker segmentation image and the marker segmentation image mask image, if a pixel value at a certain position in the mask image is 0, the pixel value of the convolved marker segmentation image at the certain position is 0, and a denoised marker segmentation image is obtained;

And carrying out contrast stretching on the denoised marker segmentation image to obtain a blood vessel segmentation image corresponding to the marker segmentation image.

5. The method for predicting the risk of progression of gastrointestinal epithelial metaplasia according to claim 2, wherein the feature extraction of the second tag attribute is performed on the segmented image of atrophic gastritis to obtain a second feature quantification value corresponding to the second tag attribute, comprising the steps of:

according to the formula:

obtaining a roughness characteristic quantization value label ₅ Wherein W is _W ，H _W Width and height of divided images of atrophic gastritis, P _mean Is the average pixel value of the divided images of atrophic gastritis, img _W Segmenting an image for atrophic gastritis, W ₀ Binarizing the atrophic gastritis segmentation map according to a certain set threshold to obtain a maximum variance row, wherein W is greater than 0 ₀ ＜W _W ；

According to the formula:

obtaining a villus sample characteristic quantization value label ₆ Wherein N is _r Is the number of villus-like regions, s _ri Is the area of the villus-like region, w _ri ，h _ri Is the width and height of the minimum circumscribing rectangle of the villus-like region, n _ri Is the corner point of the fluff-like area and is broken at the corner point to obtain the number of fluff sections;

according to the formula:

obtaining a bias white characteristic quantized value label ₇ Wherein n is _b Is the maximum number of class pixel points, (r) _bi ,g _bi ,b _bi ) Pixel values of each pixel point of the category having the largest number of colors;

according to the formula:

obtaining a bright blue ridge characteristic quantized value label ₈ Wherein W is _L ，H _L Is the width and height of the bright blue ridge segmented image, img _L Is to segment the image for bright blue ridges, (x) _RSL ,y _RSL ) Is the centroid coordinate of the bright blue ridge segmented image, S _RSL Is the bright blue ridge dividing image area, (x) _RS ,y _RS ) Is the centroid coordinates of the segmentation image of the dyeing and enlarging atrophic gastritis, S _RS Is to dye and enlarge the area of the atrophic gastritis area,

is the standard bright blue ridge three channel average pixel value, PL is the bright blue ridge split image average pixel value.

6. The method for predicting the risk of progression of gastrointestinal epithelial metaplasia according to claim 2, wherein the feature extraction of the third tag attribute is performed on the intestinal segmented image, and the third feature quantification value corresponding to the third tag attribute is obtained, comprising the steps of:

according to the formula:

obtaining a position characteristic quantized value label ₉ Wherein S is _C Is the intestinal divided image area, [ (x) _FDX ,y _FDX ),(x _FDD ,y _FDD ),(x _FJ ,y _FJ ),(x _FTX ,y _FTX ),(x _FTD ,y _FTD )]Is the centroid coordinate of the stomach risk part segmentation image [ S ] _FDX ,S _FDD ,S _FJ ,S _FTX ,S _FTD ]Is the area of the risk site, list _d ＝[d _FDX ,d _FDD ,d _FJ ,d _FTX ,d _FTD ]The distance between the intestinal segmented image and each risk part;

according to the formula:

obtaining morphological feature quantized value label ₁₀ Wherein W is _C ，H _C The width and height of the smallest circumscribed rectangle of the intestinal segmented image are respectively, (x) _C ,y _C ) Is centroid coordinate of intestinal segmented image, r _C Is the minimum circumcircle radius of the intestinal segmented image, n _f Is the number of non-enteric segmented regions within the minimum circumscribing circle, s _fi Area of non-enteric segmented region, (x) _fi ,y _fi ) Is the smallest circumscribing circular center of the intestinal segmented image.

7. The method for predicting the risk of progression of gastrointestinal epithelial metaplasia according to claim 2, wherein the third feature quantification value is input into a trained machine learning classifier for classification, and a gastroscopic image intestinal metaplasia risk level classification result is obtained, wherein the classifier comprises a feature fitting sub-network and a classification sub-network;

the obtained gastroscope image enteric risk level classification result comprises the following steps:

fitting the characteristic quantized values of the tag attributes by adopting the characteristic fitting sub-network to obtain a judgment coefficient;

and based on the judging coefficient, analyzing by adopting the classifying sub-network to obtain the identification result.

8. A gastrointestinal epithelial metaplasia progression risk prediction device, configured to:

performing gastroscopic image intestinal grade classification by acquiring characteristic quantification values of tag attributes of the marker segmentation image, the atrophic gastritis segmentation image and the intestinal segmentation image;

The first label attribute of the marker segmentation image comprises a blood vessel attribute, a fold attribute, a color attribute and a diffuse attribute;

the second label attribute of the atrophic gastritis segmented image comprises a roughness attribute, a villus attribute, a off-white attribute and a bright blue ridge attribute;

the third tag attribute of the enteric segmented image comprises a position attribute and a morphology attribute.

9. A gastrointestinal epithelial metaplasia progression risk prediction apparatus according to claim 8, comprising:

the acquisition module is used for acquiring gastroscope images;

the segmentation module is used for carrying out marker segmentation on the gastroscope image to obtain a marker segmentation image;

the feature extraction module is used for extracting features of the first tag attributes from the marker segmentation image and obtaining first feature quantized values corresponding to the first tag attributes;

the feature extraction module is further used for extracting features of second tag attributes from the atrophic gastritis segmented image, obtaining second feature quantized values corresponding to the second tag attributes, extracting features of third tag attributes from the enteric segmented image, and obtaining third feature quantized values corresponding to the third tag attributes;

The generation module is used for inputting the first characteristic quantized value into a trained machine learning classifier to classify, obtaining a gastroscope image atrophic gastritis classification result, and inputting the second characteristic quantized value into the trained machine learning classifier to classify, obtaining a gastroscope image enteron classification result; and inputting the third characteristic quantization value into a trained machine learning classifier for classification to obtain a gastroscope image intestinal risk level classification result.

10. The gastrointestinal epithelial metaplasia progression risk prediction apparatus according to claim 8, wherein said feature extraction module is configured to:

feature extraction of a first tag attribute is performed on the marker segmentation image, and a first feature quantization value corresponding to the first tag attribute is obtained, and the method comprises the following steps:

acquiring a blood vessel segmentation image in the marker segmentation image;

according to the formula:

obtaining a blood vessel characteristic quantification value label ₁ Wherein n is _v Is the number of blood vessels in the blood vessel segmentation image, s _v The blood vessel area in the blood vessel segmentation image is that of the marker segmentation image;

according to the formula:

obtaining a fold characteristic quantization value label ₂ Wherein s is _z Is the fold area, n, in the marker region in the marker-segmented image _z Is the number of folds in the marker region in the marker-segmented image, S is the marker-segmented image area, S _f Is the fold area, n, in the non-marker region in the marker-segmented image _f Is the number of folds in the non-marker region in the marker-segmented image;

according to the formula:

obtaining a color characteristic quantization value label ₃ Wherein r is _c ，g _c ，b _c Is the three-channel average color value, S is the area of the marker region in the marker-partitioned image, n _c The average number is calculated for the rest list after the colors close to black are removed;

according to the formula:

obtaining diffuse characteristic quantized value label ₄ Wherein M is the number of images taken of the entire stomach wall, w _Qi ，h _Qi Is the width and height of each image, S _i The area from each image to the marker is divided by the marker and is based on the connected domain, w _bi ，h _bi Is the width and height of the marker, (x) _bi ,y _bi ) Is the centroid coordinates of the marker.

Images