Morphological characteristic recognition method, device and equipment and computer readable storage medium

Abstract

A morphological characteristic identification method, device and equipment and a computer readable storage medium. The method comprises: performing feature extraction on a depth marker segmented image and a gastroscope marker segmented image, obtaining first feature quantization values and second feature quantization values, and inputting the first feature quantization values and the second feature quantization values into a first machine learning classifier for classification to obtain a first morphological identification result; if the first morphological identification result is mixed type, performing feature extraction on the depth marker segmented image and the gastroscope marker segmented image, obtaining third feature quantization values and fourth feature quantization values, and inputting the third feature quantization values and the fourth feature quantization values into a second machine learning classifier for classification to obtain a second morphological identification result, wherein the second morphological identification result is any one of diffuse type and nested type. The present application fully considers the influence of feature quantization values of multiple different attributes of gastroscope images and depth images on the accuracy and intuitiveness of image processing, and can effectively improve the identification efficiency and accuracy of gastric mucosa concave-convex characteristics.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method, apparatus, device, and computer-readable storage medium for morphological characteristic recognition. Background Technology

[0002] Gastric cancer is the fifth most common cancer worldwide and the third leading cause of cancer death, causing numerous deaths annually. Early diagnosis and treatment of gastric cancer are crucial, and digestive endoscopy is the first-line method for detecting early-stage gastric cancer. Within the gastric cavity, there are non-gastric cancer lesions that need to be differentiated from early-stage gastric cancer lesions. Several experts and scholars have described key endoscopic features of these lesions, such as surface roughness, boundary characteristics, lesion color, and lesion morphology. Analyzing multiple endoscopic features will help construct a lesion diagnostic system based on various characteristics. Among these, lesion morphology is a key feature, but previous lesion morphology classifications, such as the Paris classification, are subjective descriptions based on clinical experience, resulting in significant heterogeneity in application. Currently, no method has been found that can quantify the surface roughness and morphological characteristics of lesions. Summary of the Invention

[0003] This application provides a method, apparatus, device, and computer-readable storage medium for morphological characteristic identification, which can solve the technical problem of high heterogeneity in the subjective description of lesion morphology based on clinical experience in the prior art.

[0004] In a first aspect, an embodiment of this application provides a morphological characteristic recognition method, the morphological characteristic recognition method comprising:

[0005] Acquire gastroscopy images;

[0006] Depth prediction is performed on the gastroscopy image to obtain a depth prediction image;

[0007] The gastroscopy image is segmented by markers to determine the marker regions, thereby obtaining a segmented gastroscopy image.

[0008] The depth prediction image is segmented according to the marker regions to obtain a depth marker segmentation image;

[0009] The depth marker segmentation image is subjected to feature extraction of multiple first preset attributes to obtain a first feature quantization value corresponding to each first preset attribute;

[0010] The gastroscopy marker segmentation image is subjected to feature extraction of multiple second preset attributes to obtain the second feature quantization value corresponding to each second preset attribute;

[0011] Multiple first feature quantization values ​​and multiple second feature quantization values ​​are input into a first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge, and mixed type;

[0012] If the first morphology recognition result is a mixed type, then multiple third preset attributes are extracted from the depth marker segmentation image to obtain the third feature quantization value corresponding to each third preset attribute, and multiple fourth preset attributes are extracted from the gastroscopy marker segmentation image to obtain the fourth feature quantization value corresponding to each fourth preset attribute.

[0013] Multiple third feature quantization values ​​and multiple fourth feature quantization values ​​are input into a second machine learning classifier for classification to obtain a second morphology recognition result, which is either diffuse or nested.

[0014] Secondly, embodiments of this application provide a morphological characteristic recognition device, the morphological characteristic recognition device comprising:

[0015] The acquisition module is used to acquire gastroscopy images;

[0016] The depth prediction module is used to perform depth prediction on the gastroscopy image to obtain a depth prediction image;

[0017] The first segmentation module is used to segment the gastroscopy image by markers, determine the marker regions, and then obtain a segmented gastroscopy image by markers.

[0018] The second segmentation module is used to segment the depth prediction image according to the marker region to obtain a depth marker segmentation image;

[0019] The feature extraction module is used to extract features of multiple first preset attributes from the depth marker segmentation image to obtain a first feature quantization value corresponding to each first preset attribute; and to extract features of multiple second preset attributes from the gastroscopy marker segmentation image to obtain a second feature quantization value corresponding to each second preset attribute.

[0020] The first classification module is used to input multiple first feature quantization values ​​and multiple second feature quantization values ​​into the first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge and mixed type;

[0021] The feature extraction module is also used to extract features of multiple third preset attributes from the depth marker segmentation image if the first morphology recognition result is a mixed type, and to obtain the third feature quantization value corresponding to each third preset attribute; and to extract features of multiple fourth preset attributes from the gastroscopy marker segmentation image, and to obtain the fourth feature quantization value corresponding to each fourth preset attribute.

[0022] The second classification module is used to input multiple third feature quantization values ​​and multiple fourth feature quantization values ​​into the second machine learning classifier for classification, and obtain the second morphology recognition result, wherein the second morphology recognition result is either diffuse or nested.

[0023] Thirdly, this application provides a morphological characteristic recognition device, which includes a processor, a memory, and a morphological characteristic recognition program stored in the memory and executable by the processor. When the morphological characteristic recognition program is executed by the processor, it implements the steps of the morphological characteristic recognition method as described above.

[0024] Fourthly, embodiments of this application provide a computer-readable storage medium storing a morphological characteristic recognition program, wherein when the morphological characteristic recognition program is executed by a processor, it implements the steps of the morphological characteristic recognition method as described above.

[0025] The beneficial effects of the technical solutions provided in this application include at least the following:

[0026] In this application, a gastroscopy image is acquired; depth prediction is performed on the gastroscopy image to obtain a depth prediction image; marker segmentation is performed on the gastroscopy image to determine marker regions, thereby obtaining a gastroscopy marker segmentation image; the depth prediction image is segmented according to the marker regions to obtain a depth marker segmentation image; features of multiple first preset attributes are extracted from the depth marker segmentation image to obtain a first feature quantization value corresponding to each first preset attribute; features of multiple second preset attributes are extracted from the gastroscopy marker segmentation image to obtain a second feature quantization value corresponding to each second preset attribute; the multiple first feature quantization values ​​and the multiple second feature quantization values ​​are input into a first machine learning algorithm. A classifier is used to classify the gastroscopy image to obtain a first morphology recognition result, which is any one of depression, bulge, or mixed type. If the first morphology recognition result is mixed type, features of multiple third preset attributes are extracted from the depth marker segmentation image to obtain a third feature quantization value corresponding to each third preset attribute. Features of multiple fourth preset attributes are extracted from the gastroscopy marker segmentation image to obtain a fourth feature quantization value corresponding to each fourth preset attribute. The multiple third feature quantization values ​​and multiple fourth feature quantization values ​​are input into a second machine learning classifier for classification to obtain a second morphology recognition result, which is any one of diffuse type or nested type. This application fully considers the impact of feature quantization values ​​of multiple different attributes of gastroscopy images and depth images on the accuracy and intuitiveness of image processing, and can effectively improve the recognition efficiency and accuracy of gastric mucosal concavity and convexity characteristics. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an embodiment of the morphological characteristic recognition method of this application;

[0028] Figure 2 This is a schematic diagram showing the comparison between a gastroscopy image and a depth prediction image in one embodiment of the morphological characteristic recognition method of this application;

[0029] Figure 3 This is a schematic diagram showing the comparison between a gastroscopy image and a segmented image of gastroscopy markers in one embodiment of the morphological characteristic recognition method of this application;

[0030] Figure 4 This is a schematic diagram showing the three classification results in one embodiment of the morphological characteristic recognition method of this application;

[0031] Figure 5 This is a schematic diagram showing the corresponding two classification results in one embodiment of the morphological characteristic recognition method of this application;

[0032] Figure 6 This is a schematic diagram of the functional modules of an embodiment of the morphological characteristic recognition device of this application;

[0033] Figure 7 This is a schematic diagram of the hardware structure of the morphological characteristic recognition device involved in the embodiments of this application. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0035] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0036] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0037] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0038] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0040] In a first aspect, embodiments of this application provide a method for identifying morphological characteristics.

[0041] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the morphological characteristic recognition method of this application. Figure 1 As shown, the morphological characteristic recognition method includes:

[0042] Step S10: Obtain gastroscopy images;

[0043] In this embodiment, the gastroscopy image can be a white light image, a stained image, or a stained magnified image. This gastroscopy image is an electronic gastroscopy image of the stomach taken through an electronic endoscope. Specifically, the gastroscopy image can be acquired using a gastroscope or obtained from an image library pre-stored in the memory of a computer device.

[0044] Step S20: Perform depth prediction on the gastroscopy image to obtain a depth prediction image;

[0045] In this embodiment, the depth prediction image can be used to calculate the actual distance from the image to the lens during gastroscopy. (Refer to...) Figure 2 , Figure 2 This is a schematic diagram showing the comparison between a gastroscopy image and a depth prediction image in one embodiment of the morphological characteristic recognition method of this application. Figure 2 The original image in the image is the gastroscopy image.

[0046] Specifically, gastroscopy images and depth images are used as sample images to pre-train a depth prediction model. For example, an image depth prediction model such as DenseDepth is selected. In one specific implementation, the gastroscopy image is used as the input of the trained segmentation model, and the output of the depth prediction model is the depth prediction image.

[0047] Step S30: Perform marker segmentation on the gastroscopy image, determine the marker region, and then obtain a gastroscopy marker segmentation image;

[0048] In this embodiment, the gastroscopy marker segmentation image refers to the gastroscopy marker segmentation image obtained by segmenting the gastroscopy image according to the region where the marker is located, thus containing the marker region. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram showing the comparison between a gastroscopy image and a segmented image of gastroscopy markers in one embodiment of the morphological characteristic recognition method of this application. Figure 3 The central image is the gastroscopy image, and the marker segmentation image is the gastroscopy marker segmentation image.

[0049] Specifically, gastroscopy images and marker segmentation images are used as sample images. A marker segmentation model is pre-trained, for example, image segmentation network models such as Unet++ or Mask-RCNN are selected. In one specific implementation, the gastroscopy image is used as the input to the trained segmentation model, and the output of the segmentation model is the marker segmentation image. Understandably, in this embodiment, by obtaining the gastroscopy marker segmentation image, multiple preset attribute features can be quantized on the gastroscopy marker segmentation image in subsequent steps, thereby improving the accuracy of quantization.

[0050] Step S40: Segment the depth prediction image according to the marker region to obtain a depth marker segmentation image;

[0051] In this embodiment, the depth marker segmentation image refers to the depth marker segmentation image obtained by segmenting the depth prediction image according to the region where the marker is located. Understandably, in this embodiment, obtaining the depth marker segmentation image allows for subsequent quantization of multiple preset attribute features, thereby improving the accuracy of the quantization.

[0052] Step S50: Extract features of multiple first preset attributes from the depth marker segmentation image to obtain the first feature quantization value corresponding to each first preset attribute;

[0053] In this embodiment, the first preset attribute refers to multiple attributes of the depth marker segmentation image, such as the distance attribute, depth attribute, and area ratio attribute of the depth marker segmentation image, and the first feature quantization value refers to the quantization value corresponding to the feature of each first preset attribute.

[0054] Specifically, feature extraction methods are used to extract features from the depth marker segmentation image to obtain a first feature quantization value. The feature extraction method can be a combination of manual feature extraction methods and image feature analysis algorithms such as pixel neighborhood mean calculation and maximum pixel value extraction to calculate the first feature quantization value. Alternatively, it can be a deep learning feature extraction method, such as convolutional neural network (CNN) or UNet++. The specific method can be selected based on the features of the first preset attribute, and there are no restrictions here.

[0055] In this embodiment, by extracting features from the depth marker segmentation image and obtaining the corresponding first feature quantization value, the quantization calculation of the features of each first preset attribute of the depth marker segmentation 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 multiple first feature quantization values, thereby improving the accuracy of image concavity and convexity characteristics recognition.

[0056] Step S60: Extract features of multiple second preset attributes from the segmented image of gastroscopy markers to obtain the second feature quantization value corresponding to each second preset attribute;

[0057] In this embodiment, the second preset attribute refers to multiple attributes of the gastroscopy marker segmentation image, such as the color gradient attribute, surrounding liquid texture attribute, and reflected light spot attribute of the gastroscopy marker segmentation image, and the second feature quantization value refers to the quantization value corresponding to the feature of each second preset attribute.

[0058] Specifically, feature extraction methods are used to extract features from the segmented image of gastroscopy markers to obtain a second feature quantization value. The feature extraction method can be a combination of manual feature extraction methods and image feature analysis algorithms such as pixel neighborhood mean calculation and maximum pixel value extraction to calculate the second feature quantization value. Alternatively, it can be a deep learning feature extraction method, such as convolutional neural network (CNN) or UNet++. The specific method can be selected according to the features of the second preset attribute, and there are no restrictions here.

[0059] In this embodiment, by extracting features from the segmented image of the gastroscopy markers, the corresponding second feature quantization value is obtained, thereby realizing the quantization calculation of the features of each second preset attribute of the segmented image of the gastroscopy markers. This makes the feature quantization value more comprehensive and rich, so that subsequent accurate and intuitive image analysis and recognition can be performed based on these multiple second feature quantization values, thus improving the accuracy of image concavity and convexity characteristics recognition.

[0060] Step S70: Input multiple first feature quantization values ​​and multiple second feature quantization values ​​into the first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge and mixed type;

[0061] In this embodiment, the first machine learning classifier can be implemented by learning a machine learning algorithm model with classification capabilities from samples. The machine learning algorithm model can be one or more of the following: neural networks (e.g., convolutional neural networks, backpropagation neural networks, etc.), logistic regression models, support vector machines, decision trees, random forests, perceptrons, and other machine learning models. As part of training such a machine learning model, the training inputs are the quantized values ​​of each first feature and each second feature, such as distance attributes, depth attributes, area ratio attributes, color gradient attributes, surrounding liquid texture attributes, and reflected light spot attributes. Through training, a classifier is established that establishes the correspondence between the first feature value set, the second feature value set, and the concave / convex characteristics of the gastroscopy image to be identified. This enables the preset classifier to determine whether the classification result of the gastroscopy image to be identified is concave, convex, or a mixed result.

[0062] In this embodiment, the classifier is a three-classifier, meaning it yields three classification results: a concave result, a convex result, or a mixed result. (Refer to...) Figure 4 , Figure 4 This is a schematic diagram showing the three classification results in one embodiment of the morphological characteristic recognition method of this application.

[0063] Understandably, this embodiment fully considers the impact of the feature quantization values ​​of multiple different attributes of depth marker segmentation images and the feature quantization values ​​of multiple different attributes of gastroscopy marker segmentation images on the accuracy and intuitiveness of image processing. By extracting features with richer information and quantifying and comprehensively processing features of multiple different attributes, the rationality of feature value quantization is improved. Compared with the traditional method of only considering single feature information and single statistical comparison, the efficiency of recognizing concave and convex characteristics of gastroscopy images is greatly improved.

[0064] Step S80: If the first morphology recognition result is a mixed type, then perform feature extraction of multiple third preset attributes on the depth marker segmentation image to obtain the third feature quantization value corresponding to each third preset attribute, and perform feature extraction of multiple fourth preset attributes on the gastroscopy marker segmentation image to obtain the fourth feature quantization value corresponding to each fourth preset attribute.

[0065] In this embodiment, the third preset attribute refers to multiple attributes of the depth marker segmentation image, such as the brightness distribution attribute and zero-point crossover attribute of the depth marker segmentation image, and the third feature quantization value refers to the quantization value corresponding to the feature of each third preset attribute.

[0066] Specifically, feature extraction methods are used to extract features from the depth marker segmentation image to obtain a third feature quantization value. The feature extraction method can be a combination of manual feature extraction methods and image feature analysis algorithms such as pixel neighborhood mean calculation and maximum pixel value extraction to calculate the third feature quantization value. Alternatively, it can be a deep learning feature extraction method, such as convolutional neural network (CNN) or UNet++. The specific method can be selected based on the features of the third preset attribute, and there are no restrictions here.

[0067] In this embodiment, by extracting features from the depth marker segmentation image, the corresponding third feature quantization value is obtained, thereby realizing the quantization calculation of the features of each third preset attribute of the depth marker segmentation image. This makes the feature quantization value more comprehensive and rich, so that subsequent accurate and intuitive image analysis and recognition can be performed based on these multiple third feature quantization values, thus improving the accuracy of image concavity and convexity characteristics recognition.

[0068] The fourth preset attribute refers to multiple attributes of the gastroscopy marker segmentation image, such as the surface regularity attribute and surface roughness attribute of the gastroscopy marker segmentation image. The fourth feature quantization value refers to the quantization value corresponding to the feature of each fourth preset attribute.

[0069] Specifically, feature extraction methods are used to extract features from the segmented images of gastroscopy markers to obtain the fourth feature quantization value. The feature extraction method can be a combination of manual feature extraction methods and image feature analysis algorithms such as pixel neighborhood mean calculation and maximum pixel value extraction to calculate the fourth feature quantization value. Alternatively, it can be a deep learning feature extraction method, such as convolutional neural network (CNN) or UNet++. The specific method can be selected according to the features of the fourth preset attribute, and there are no restrictions here.

[0070] In this embodiment, by extracting features from the segmented image of the gastroscopy markers, the corresponding fourth feature quantization value is obtained, thereby realizing the quantization calculation of the features of each fourth preset attribute of the segmented image of the gastroscopy markers. This makes the feature quantization value more comprehensive and rich, so that subsequent accurate and intuitive image analysis and recognition can be performed based on these multiple fourth feature quantization values, thus improving the accuracy of image concavity and convexity characteristics recognition.

[0071] Step S90: Input multiple third feature quantization values ​​and multiple fourth feature quantization values ​​into the second machine learning classifier for classification to obtain the second morphology recognition result, wherein the second morphology recognition result is either diffuse or nested.

[0072] In this embodiment, the second machine learning classifier can be implemented by learning a machine learning algorithm model with classification capabilities through sample learning. The machine learning algorithm model can be one or more of the following: neural networks (e.g., convolutional neural networks, backpropagation neural networks, etc.), logistic regression models, support vector machines, decision trees, random forests, perceptrons, and other machine learning models. As part of training such a machine learning model, the training inputs are the quantized values ​​of each third feature and each fourth feature, such as brightness distribution attributes, zero-point crossover attributes, surface regularity attributes, surface roughness attributes, etc. Through training, a classifier is established that corresponds to the set of third feature values, the set of fourth feature values, and the mixed classification of the gastroscopy image to be identified. This enables the preset classifier to determine whether the classification result of the gastroscopy image to be identified is diffuse or nested.

[0073] In this embodiment, the classifier is a binary classifier, meaning it yields two classification results: either a diffuse or nested classification. (Refer to...) Figure 5 , Figure 5 This is a schematic diagram showing the corresponding classification results of two types in one embodiment of the morphological characteristic recognition method of this application.

[0074] Understandably, this embodiment fully considers the impact of the feature quantization values ​​of multiple different attributes of the hybrid depth marker segmentation image and the feature quantization values ​​of multiple different attributes of the hybrid gastroscopy marker segmentation image on the accuracy and intuitiveness of image processing. By extracting features with richer information and quantifying and comprehensively processing features of multiple different attributes, the rationality of feature value quantification is improved. Compared with the traditional method that only considers single feature information and single statistical comparison, the efficiency of hybrid classification recognition of gastroscopy images is greatly improved.

[0075] In this embodiment, a gastroscopy image is acquired; depth prediction is performed on the gastroscopy image to obtain a depth prediction image; marker segmentation is performed on the gastroscopy image to determine marker regions, thereby obtaining a gastroscopy marker segmentation image; the depth prediction image is segmented according to the marker regions to obtain a depth marker segmentation image; features of multiple first preset attributes are extracted from the depth marker segmentation image to obtain a first feature quantization value corresponding to each first preset attribute; features of multiple second preset attributes are extracted from the gastroscopy marker segmentation image to obtain a second feature quantization value corresponding to each second preset attribute; the multiple first feature quantization values ​​and the multiple second feature quantization values ​​are input into a first machine learning algorithm. A classifier is used to classify the gastroscopy image to obtain a first morphology recognition result, which is any one of depression, bulge, or mixed type. If the first morphology recognition result is mixed type, features of multiple third preset attributes are extracted from the depth marker segmentation image to obtain a third feature quantization value corresponding to each third preset attribute. Features of multiple fourth preset attributes are extracted from the gastroscopy marker segmentation image to obtain a fourth feature quantization value corresponding to each fourth preset attribute. The multiple third feature quantization values ​​and multiple fourth feature quantization values ​​are input into a second machine learning classifier for classification to obtain a second morphology recognition result, which is any one of diffuse type or nested type. This embodiment fully considers the impact of feature quantization values ​​of multiple different attributes of gastroscopy images and depth images on the accuracy and intuitiveness of image processing, and can effectively improve the recognition efficiency and accuracy of gastric mucosal convexity characteristics.

[0076] Further, in one embodiment, the plurality of first preset attributes include a distance attribute, a depth attribute, and an area proportion attribute, and the corresponding first feature quantization values ​​are the distance feature quantization value, the depth feature quantization value, and the area proportion feature quantization value, respectively. Step S50 includes:

[0077] Step S501: The pixel values ​​of each pixel in the marker region of the depth marker segmentation image are adjusted based on the adjustment value and then summed. The ratio of the summation result to the area of ​​the marker region and 255 is used as the distance feature quantization value.

[0078] In this embodiment, the distance feature quantization value label1 is calculated using the following formula:

[0079]

[0080] Where img(p, q) is the pixel value at (p, q) in the marker region of the depth marker segmentation image, (w1, h1) is the coordinate of the lower right corner of the minimum horizontal bounding rectangle of the marker in the depth marker segmentation image, (w0, h0) is the coordinate of the upper left corner of the minimum horizontal bounding rectangle of the marker in the depth marker segmentation image, S is the area of ​​the marker region in the depth marker segmentation image, and δ is the adjustment value.

[0081] Furthermore, in one embodiment, the morphological characteristic recognition method further includes:

[0082] M reference gastroscopy images are acquired, each reference gastroscopy image being an image taken of the same stomach as the previous gastroscopy image. For each reference gastroscopy image, depth prediction is performed to obtain a reference depth prediction image, and non-marker regions in the reference depth prediction image are identified. Reference pixel values ​​are obtained based on the area of ​​the non-marker regions in the M reference depth prediction images and the pixel values ​​of the pixels within the non-marker regions. Adjustment values ​​are obtained based on the reference pixel values, the area of ​​the non-marker regions in the depth prediction images, and the pixel values ​​of the pixels within the non-marker regions.

[0083] In this embodiment, the reference pixel value μ is obtained based on the following formula:

[0084]

[0085] Among them, SUM' k S' is the sum of the pixel values ​​of all pixels within the non-landmark region in the Kth reference depth prediction image. k Let be the area of ​​the non-marker region in the Kth reference depth prediction image.

[0086] After obtaining the reference pixel value μ, the adjustment value is obtained according to the following formula:

[0087]

[0088] Where img(i,j) is the pixel value at (i,j) in the non-marker region of the depth marker segmentation image, (W,H) are the width and height of the depth marker segmentation image, and S' is the area of ​​the non-marker region in the depth prediction image. Based on the above formula, the adjustment value δ can be obtained.

[0089] Step S502: The depth of the marker region in the depth marker segmentation image relative to the background mucosa is used as the depth feature quantization value.

[0090] In this embodiment, the ratio of the sum of the pixel values ​​within the marker region in the depth marker segmentation image after adjustment relative to the adjustment value to the area of ​​the marker region, compared to the change in the reference pixel value of the non-marker region, is used as the depth feature quantization value. The depth feature quantization value label2 can be calculated using the following formula:

[0091]

[0092] Where img(p, q) is the pixel value at (p, q) in the marker region of the depth marker segmentation image, (w1, h1) is the coordinate of the lower right corner of the minimum horizontal bounding rectangle of the marker in the depth marker segmentation image, w0, h0) is the coordinate of the upper left corner of the minimum horizontal bounding rectangle of the marker in the depth marker segmentation image, S is the area of ​​the marker region in the depth marker segmentation image, δ is the adjustment value, and μ is the reference pixel value.

[0093] Step S503: Count the number of target pixels in the marker region of the depth marker segmentation image whose pixel value exceeds the pixel threshold, and use the ratio of the number of target pixels to the area of ​​the marker region as the area proportion feature quantization value.

[0094] In this embodiment, the pixel value of each pixel within the marker region in the depth marker segmentation image is compared with a pixel threshold, and the number of target pixels Num whose pixel value exceeds the pixel threshold is counted. Then, the area proportion feature quantization value label3 is obtained using the following formula:

[0095]

[0096] Where S is the area of ​​the marker region in the depth marker segmentation image.

[0097] Further, in one embodiment, the plurality of second preset attributes include a color gradient attribute, a surrounding liquid texture attribute, and a reflected light spot attribute, and the corresponding second feature quantization values ​​are the color gradient feature quantization value, the surrounding liquid texture feature quantization value, and the reflected light spot feature quantization value, respectively. Step S60 includes:

[0098] Step S601: Construct multiple circles with the centroid of the marker in the gastroscopy marker segmentation image as the center, and use the ratio of the average pixel value inside the circle to the change in the circle radius as the color gradient feature quantization value.

[0099] In this embodiment, the gradual color change of the marker from its center to its edge can reflect the image's convexity and concavity characteristics to a certain extent.

[0100] Specifically, based on the connected components, the centroids of the markers in the segmented image of the gastroscopy markers are obtained. X circles are constructed with these centroids as centers, with a minimum radius of r0 and a maximum radius of r1. The gradient of the circle radius variation can then be obtained. The color gradient quantization value label4 is:

[0101]

[0102] in, Where img(i,j) is the pixel value of the pixel at (i,j) within the annular interval corresponding to r0+(x-1)·Δr~r0+x·Δr; Wherein, ing(i,j) is the pixel value of the pixel at (i,j) inside the circle corresponding to r0; m and n are the width and height of the segmented image of the gastroscopy marker, respectively; and x is a variable.

[0103] Step S602: Using the centroid of the marker in the gastroscopy marker segmentation image as the center, enlarge the radius of the maximum outer circle of the marker according to a preset ratio, and multiply the ratio of the variance of the liquid texture feature vector to the liquid texture feature vector in the enlarged circular area by the circle enlargement ratio to obtain the quantization value of the surrounding liquid texture feature.

[0104] In this embodiment, if the marker is concave, the surrounding liquid extends from the surrounding background mucosa to the marker area; if the marker is convex, the surrounding liquid surrounds the marker in a circular shape from the base of the marker. The texture of the liquid near the marker can reflect the concavity and convexity characteristics of the image to a certain extent.

[0105] Specifically, based on the connected components, the centroid and maximum circumcircle of the markers in the segmented image of the gastroscopy markers are obtained. The radius of the maximum circumcircle is then magnified by a factor of ε, using the centroid as the center. Within this magnified circle, the liquid is segmented. If no liquid is segmented, the texture quantization value label5 around the liquid is set to 0. If the liquid is segmented, resulting in a segmented image around the liquid, the LBP feature extraction class from the skimage toolkit in Phyton is used to extract texture features from the segmented image, resulting in a one-dimensional feature vector list. w Then the quantization value of the surrounding liquid texture Among them, std(list w (a list of one-dimensional feature vectors) w Standard deviation, mean(list) w (a list of one-dimensional feature vectors) w The average value.

[0106] Step S603: The ratio of the area of ​​the reflected light spot on the surface of the gastroscopy marker segmentation image to the area of ​​the marker is multiplied by the ratio of the distance between the centroid of the reflected light spot and the centroid of the marker to the diagonal of the marker's smallest circumscribed rectangle, and the reflected light spot feature quantization value is used.

[0107] In this embodiment, when the gastric mucosa is in a raised or diffuse form, it is easy to reflect light and produce light spots.

[0108] Specifically, the reflected light spot is segmented in the segmented image of the gastroscopic marker. If no light spot is segmented, the quantization value of the reflected light spot, label6, is set to 0. If the reflected light spot is segmented, a segmented image of the reflected light spot is obtained. Based on the connected components, the centroid O and area S of the marker region in the segmented image of the gastroscopic marker are obtained. O The centroid G of the reflected light spot region and the area S of the reflected light spot region. G Then the quantization value of the reflected light spot Where d OG W is the distance between the centroid of the reflected light spot area and the centroid of the marker area. O H O The width and height are the minimum bounding rectangle of the marker area.

[0109] Further, in one embodiment, the plurality of third preset attributes include a brightness distribution attribute and a zero-point crossover attribute, and the corresponding third feature quantization values ​​are the brightness distribution feature quantization value and the zero-point crossover feature quantization value, respectively. The step of extracting features from the depth marker segmentation image using the plurality of third preset attributes to obtain the third feature quantization value corresponding to each third preset attribute includes:

[0110] The product of the distance between the equivalent centroids of pixels above the threshold and pixels below the threshold in the depth marker segmentation image and the ratio of the ratio of the diagonal of the minimum bounding rectangle of the depth marker segmentation image to the threshold relative to 255 is used as the quantization value of the brightness distribution feature.

[0111] In this embodiment, the brightness distribution in the depth marker segmentation image represents the degree of aggregation of depressions or ridges. The more dispersed the brightness, the more it tends to be diffuse, and vice versa.

[0112] Specifically, a threshold ω is set, and the equivalent centroid of pixels in the depth marker segmentation image that are above the threshold ω is calculated to be O. > (x > ,y > ), calculate the equivalent centroid of pixels below the threshold ω in the depth marker segmentation image as O. ≤ (x ≤ ,y ≤ Then the quantized value of the brightness distribution Among them WS and H S Find the width and height of the minimum bounding rectangle for segmenting the image for depth markers.

[0113] The ratio of the sum of the number of 0 and 1 alternations in each row to the sum of the number of 0 and 1 alternations in each column after binarization of the image segmented by depth markers is used as the zero-point crossover feature quantization value.

[0114] In this embodiment, the depth marker segmentation image is binarized to obtain its corresponding mask image containing only 0s and 1s. The number of times C in each row of the mask image changes from 1 to 0 is counted. rj , where j≤H s H s The height of the depth marker segmentation image is represented by C; the number of times each column of the mask image changes from 1 to 0 is counted. ci , where i≤W s W s The width of the image segmented by the depth marker is represented by the brightness distribution quantization value.

[0115] Further, in one embodiment, the plurality of fourth preset attributes include surface regularity attributes and surface roughness attributes, and the corresponding fourth feature quantization values ​​are surface regularity feature quantization values ​​and surface roughness feature quantization values, respectively. The step of extracting features from the segmented image of gastroscopic markers using the plurality of fourth preset attributes to obtain the fourth feature quantization value corresponding to each fourth preset attribute includes:

[0116] The two-dimensional image entropy of the segmented image based on gastroscopy markers is used as the quantization value of surface regularity features;

[0117] In this embodiment, the more disordered the surface texture of the markers in the gastroscopy image, the more diffuse the gastric mucosa morphology tends to be.

[0118] Specifically, if represented by surface image entropy, then the surface regularity quantization value... Among them, P i This represents the probability of gray level i occurring. The one-dimensional entropy of an image can represent the clustering characteristics of its gray-level distribution, but it cannot reflect the spatial characteristics of the gray-level distribution. To characterize this spatial characteristic, a feature quantity reflecting the spatial characteristics of the gray-level distribution is introduced on top of the one-dimensional entropy to form the two-dimensional entropy of the image. The mean gray level of the image's neighborhood is chosen as the spatial feature quantity of the gray-level distribution, forming a feature pair with the pixel gray level, denoted as (i, j), where i represents the gray level value of the pixel, and j represents the mean gray level of the neighborhood. Where f(i,j) is the frequency of occurrence of the feature pair (i,j), and N is the size of the segmented image of the gastroscopy marker.

[0119] The sum of the absolute values ​​of the differences between the pixel values ​​of each column in the row containing the centroid of the gastroscopic marker in the segmented image and the average pixel value of the segmented image is used as the quantization value of the surface roughness feature.

[0120] In this embodiment, the rougher the surface of the marker in the gastroscopy image, the more diffuse the gastric mucosa morphology tends to be, and the higher the surface roughness quantification value... Where W0 is a baseline parallel to the segmentation line of the gastroscopy landmark image, 0 <W0<W W W0 is the x-coordinate of the centroid of the segmented region in the gastroscopy marker segmentation image, and P... mean Average pixel values ​​of segmented images of gastroscopy markers. abs is the absolute value function, W W H W These represent the width and height of the segmented image of the gastroscopy markers, respectively.

[0121] Further, in one embodiment, the first machine learning classifier includes a first feature fitting sub-network and a first classification sub-network, and step S70 includes:

[0122] Multiple first feature quantization values ​​and multiple second feature quantization values ​​are input into a first machine learning classifier; a first feature fitting sub-network performs a weighted summation of the multiple first feature quantization values ​​and multiple second feature quantization values ​​according to the weighting coefficients corresponding to the multiple first feature quantization values ​​and multiple second feature quantization values ​​to obtain a first fused feature value; a first classification sub-network obtains a first morphology recognition result based on the first fused feature value;

[0123] In this embodiment, the first feature fitting subnetwork is used to fit each first feature quantization value and each second feature quantization value to obtain the corresponding weights for each first feature quantization value and each second feature quantization value. Continuing with the example of the first feature quantization values ​​label1~label3 and the second feature quantization values ​​label4~label6 in the above embodiment, the corresponding weights λ1~λ6 are determined using decision trees, random forests, etc., and then the fused feature values ​​are... Based on θ1, the first classification sub-network is used for analysis to obtain the first morphology recognition result, which is any one of depression, bulge, and mixed type.

[0124] The second machine learning classifier includes a second feature fitting subnetwork and a second classification subnetwork. Step S90 includes:

[0125] Multiple third feature quantization values ​​and multiple fourth feature quantization values ​​are input into the second machine learning classifier; the second feature fitting sub-network performs a weighted summation of the multiple third feature quantization values ​​and multiple fourth feature quantization values ​​according to the weighting coefficients corresponding to the multiple third feature quantization values ​​and multiple fourth feature quantization values ​​to obtain the second fusion feature value; the second classification sub-network obtains the second morphological recognition result based on the second fusion feature value.

[0126] In this embodiment, the second feature fitting sub-network is used to fit each third feature quantization value and each fourth feature quantization value to obtain the corresponding weights for fitting each third feature quantization value and each fourth feature quantization value. The third feature quantization value is then used as described in the previous embodiment. and Fourth feature quantization value and For example, using decision trees, random forests, etc., the corresponding weights are determined as λ. h1 , λ h2 , λ h3 , λ h4 Then the fused eigenvalues Based on θ2, a second classification sub-network is used for analysis to obtain the second morphology recognition result, which is either diffuse or nested.

[0127] In this embodiment, by fusing and calculating the quantization values ​​of each third feature and each fourth feature, the information features of the gastroscopy image are made richer and the quantization is more accurate, thereby improving the recognition accuracy and efficiency of the concave and convex characteristics of the gastroscopy image.

[0128] Secondly, embodiments of this application also provide a morphological characteristic recognition device.

[0129] In one embodiment, reference is made to Figure 6 , Figure 6 This is a schematic diagram of the functional modules of an embodiment of the morphological characteristic recognition device of this application. Figure 6 As shown, the morphological characteristic recognition device includes:

[0130] Acquisition module 10 is used to acquire gastroscopy images;

[0131] Depth prediction module 20 is used to perform depth prediction on the gastroscopy image to obtain a depth prediction image;

[0132] The first segmentation module 30 is used to perform marker segmentation on the gastroscopy image, determine the marker region, and then obtain a gastroscopy marker segmentation image.

[0133] The second segmentation module 40 is used to segment the depth prediction image according to the marker region to obtain a depth marker segmentation image;

[0134] The feature extraction module 50 is used to extract features of multiple first preset attributes from the depth marker segmentation image to obtain a first feature quantization value corresponding to each first preset attribute; and to extract features of multiple second preset attributes from the gastroscopy marker segmentation image to obtain a second feature quantization value corresponding to each second preset attribute.

[0135] The first classification module 60 is used to input multiple first feature quantization values ​​and multiple second feature quantization values ​​into the first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge and mixed type;

[0136] The feature extraction module 50 is also used to extract features of multiple third preset attributes from the depth marker segmentation image if the first morphology recognition result is a mixed type, and to obtain the third feature quantization value corresponding to each third preset attribute; and to extract features of multiple fourth preset attributes from the gastroscopy marker segmentation image, and to obtain the fourth feature quantization value corresponding to each fourth preset attribute.

[0137] The second classification module 70 is used to input multiple third feature quantization values ​​and multiple fourth feature quantization values ​​into the second machine learning classifier for classification to obtain the second morphology recognition result, wherein the second morphology recognition result is either diffuse or nested.

[0138] Further, in one embodiment, the plurality of first preset attributes include a distance attribute, a depth attribute, and an area percentage attribute, and the corresponding first feature quantization values ​​are the distance feature quantization value, the depth feature quantization value, and the area percentage feature quantization value, respectively. The feature extraction module 50 is used for:

[0139] The pixel values ​​of each pixel within the marker region in the depth marker segmentation image are adjusted based on the adjustment value and then summed. The ratio of the summation result to the area of ​​the marker region and 255 is used as the distance feature quantization value.

[0140] The depth of the marker region in the image relative to the background mucosa is used as the depth feature quantization value;

[0141] The number of target pixels with pixel values ​​exceeding a pixel threshold within the marker region in the depth marker segmentation image is counted, and the ratio of the number of target pixels to the area of ​​the marker region is used as the quantized value of the area proportion feature.

[0142] Furthermore, in one embodiment, the morphological characteristic recognition device further includes an adjustment value determination module, used for:

[0143] Acquire M reference gastroscopy images, wherein the reference gastroscopy images are images taken of the same stomach as the gastroscopy images;

[0144] For each reference gastroscopy image, depth prediction is performed on the reference gastroscopy image to obtain a reference depth prediction image, and non-marker regions in the reference depth prediction image are identified.

[0145] The reference pixel value is obtained by predicting the area of ​​non-marker regions in M ​​reference depth images and the pixel value of the pixels in the non-marker regions.

[0146] The adjustment value is obtained based on the reference pixel value, the area of ​​the non-marker region in the depth prediction image, and the pixel value of the pixel in the non-marker region.

[0147] Further, in one embodiment, the plurality of second preset attributes include a color gradient attribute, a surrounding liquid texture attribute, and a reflected light spot attribute, and the corresponding second feature quantization values ​​are the color gradient feature quantization value, the surrounding liquid texture feature quantization value, and the reflected light spot feature quantization value, respectively. The feature extraction module 50 is used for:

[0148] Multiple circles are constructed with the centroid of the marker in the gastroscopy marker segmentation image as the center, and the ratio of the average pixel value within the circle to the change in the circle radius is used as the quantization value of the color gradient feature.

[0149] Using the centroid of the marker in the segmented image of the gastroscopy marker as the center, the radius of the largest outer circle of the marker is enlarged by a preset ratio. The ratio of the variance of the liquid texture feature vector to the liquid texture feature vector in the enlarged circular area is multiplied by the circle enlargement ratio to obtain the quantization value of the surrounding liquid texture feature.

[0150] The ratio of the area of ​​the reflected light spot on the surface of the gastroscopic marker to the area of ​​the marker in the segmented image is multiplied by the ratio of the distance between the centroid of the reflected light spot and the centroid of the marker relative to the diagonal of the smallest circumscribed rectangle of the marker, and the reflected light spot feature quantification value is used.

[0151] Further, in one embodiment, the plurality of third preset attributes include a brightness distribution attribute and a zero-point crossover attribute, and the corresponding third feature quantization values ​​are the brightness distribution feature quantization value and the zero-point crossover feature quantization value, respectively. The feature extraction module 50 is used for:

[0152] The product of the distance between the equivalent centroids of pixels above the threshold and pixels below the threshold in the depth marker segmentation image and the ratio of the ratio of the diagonal of the minimum bounding rectangle of the depth marker segmentation image to the threshold relative to 255 is used as the quantization value of the brightness distribution feature.

[0153] The ratio of the sum of the number of 0 and 1 alternations in each row to the sum of the number of 0 and 1 alternations in each column after binarization of the image segmented by depth markers is used as the zero-point crossover feature quantization value.

[0154] Further, in one embodiment, the plurality of fourth preset attributes include surface regularity attributes and surface roughness attributes, and the corresponding fourth feature quantization values ​​are surface regularity feature quantization values ​​and surface roughness feature quantization values, respectively. The feature extraction module 50 is used for:

[0155] The two-dimensional image entropy of the segmented image based on gastroscopy markers is used as the quantization value of surface regularity features;

[0156] The sum of the absolute values ​​of the differences between the pixel values ​​of each column in the row containing the centroid of the gastroscopic marker in the segmented image and the average pixel value of the segmented image is used as the quantization value of the surface roughness feature.

[0157] Furthermore, in one embodiment, the first machine learning classifier includes a first feature fitting subnetwork and a first classification subnetwork, and the first classification module 60 is used for:

[0158] Input multiple quantized values ​​of the first feature and multiple quantized values ​​of the second feature into the first machine learning classifier;

[0159] The first feature fitting sub-network performs a weighted summation of multiple first feature quantization values ​​and multiple second feature quantization values ​​according to the weighting coefficients corresponding to multiple first feature quantization values ​​and multiple second feature quantization values, to obtain the first fusion feature value;

[0160] The first classification sub-network obtains the first morphological recognition result based on the first fusion feature value;

[0161] The second machine learning classifier includes a second feature fitting subnetwork and a second classification subnetwork. The second classification module 70 is used for:

[0162] Multiple quantized values ​​of the third feature and multiple quantized values ​​of the fourth feature are input into the second machine learning classifier;

[0163] The second feature fitting sub-network performs a weighted summation of multiple third feature quantization values ​​and multiple fourth feature quantization values ​​according to the weighting coefficients corresponding to the multiple third feature quantization values ​​and multiple fourth feature quantization values, to obtain the second fusion feature value;

[0164] The second classification sub-network obtains the second morphological recognition result based on the second fusion feature value.

[0165] The functions of each module in the above-mentioned morphological characteristic recognition device correspond to the steps in the above-mentioned morphological characteristic recognition method embodiment, and their functions and implementation processes will not be described in detail here.

[0166] Thirdly, embodiments of this application provide a morphological characteristic recognition device, which can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0167] Reference Figure 7 , Figure 7 This is a schematic diagram of the hardware structure of the morphological characteristic recognition device involved in the embodiments of this application. In the embodiments of this application, the morphological characteristic recognition device may include a processor, a memory, a communication interface, and a communication bus.

[0168] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0169] Communication interfaces include input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting internal components of the morphology recognition device, as well as interfaces used for interconnecting the morphology recognition device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0170] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0171] The processor can be a general-purpose processor, which can call the morphological characteristic recognition program stored in the memory and execute the morphological characteristic recognition method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the morphological characteristic recognition program is called can be referred to in the various embodiments of the morphological characteristic recognition method of this application, and will not be repeated here.

[0172] Those skilled in the art will understand that Figure 7 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0173] Fourthly, embodiments of this application also provide a readable storage medium.

[0174] The present application has a morphological characteristic recognition program stored on a readable storage medium, wherein when the morphological characteristic recognition program is executed by a processor, it implements the steps of the morphological characteristic recognition method as described above.

[0175] The method implemented when the morphological characteristic recognition program is executed can be referred to in various embodiments of the morphological characteristic recognition method of this application, and will not be repeated here.

[0176] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0177] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0178] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for identifying morphological characteristics, characterized in that, The morphological characteristic recognition method includes: Acquire gastroscopy images; Depth prediction is performed on the gastroscopy image to obtain a depth prediction image; The gastroscopy image is segmented by markers to determine the marker regions, thereby obtaining a segmented gastroscopy image. The depth prediction image is segmented according to the marker regions to obtain a depth marker segmentation image; The depth marker segmentation image is subjected to feature extraction of multiple first preset attributes to obtain a first feature quantization value corresponding to each first preset attribute. The multiple first preset attributes include distance attribute, depth attribute and area proportion attribute, and the corresponding first feature quantization values ​​are distance feature quantization value, depth feature quantization value and area proportion feature quantization value, respectively. The gastroscopy marker segmentation image is subjected to feature extraction of multiple second preset attributes to obtain the second feature quantization value corresponding to each second preset attribute. The multiple second preset attributes include color gradient attribute, surrounding liquid texture attribute and reflected light spot attribute, and the corresponding second feature quantization values ​​are color gradient feature quantization value, surrounding liquid texture feature quantization value and reflected light spot feature quantization value, respectively. Multiple first feature quantization values ​​and multiple second feature quantization values ​​are input into a first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge, and mixed type; If the first morphology recognition result is a hybrid type, then multiple third preset attributes are extracted from the depth marker segmentation image to obtain the third feature quantization value corresponding to each third preset attribute. Similarly, multiple fourth preset attributes are extracted from the gastroscopy marker segmentation image to obtain the fourth feature quantization value corresponding to each fourth preset attribute. The multiple third preset attributes include a brightness distribution attribute and a zero-point crossover attribute, with the corresponding third feature quantization values ​​being the brightness distribution feature quantization value and the zero-point crossover feature quantization value, respectively. The multiple fourth preset attributes include a surface regularity attribute and a surface roughness attribute, with the corresponding fourth feature quantization values ​​being the surface regularity feature quantization value and the surface roughness feature quantization value, respectively. Multiple third feature quantization values ​​and multiple fourth feature quantization values ​​are input into a second machine learning classifier for classification to obtain a second morphology recognition result, which is either diffuse or nested.

2. The morphological characteristic recognition method as described in claim 1, characterized in that, The steps of extracting features from a depth marker segmentation image for multiple first preset attributes and obtaining a first feature quantization value corresponding to each first preset attribute include: The pixel values ​​of each pixel within the marker region in the depth marker segmentation image are adjusted based on the adjustment value and then summed. The ratio of the summation result to the area of ​​the marker region and 255 is used as the distance feature quantization value. The depth of the marker region in the image relative to the background mucosa is used as the depth feature quantization value; The number of target pixels with pixel values ​​exceeding a pixel threshold within the marker region in the depth marker segmentation image is counted, and the ratio of the number of target pixels to the area of ​​the marker region is used as the quantized value of the area proportion feature.

3. The morphological characteristic recognition method as described in claim 2, characterized in that, The morphological characteristic recognition method further includes: Acquire M reference gastroscopy images, wherein the reference gastroscopy images are images taken of the same stomach as the gastroscopy images; For each reference gastroscopy image, depth prediction is performed on the reference gastroscopy image to obtain a reference depth prediction image, and non-marker regions in the reference depth prediction image are identified. The reference pixel value is obtained by predicting the area of ​​non-marker regions in M ​​reference depth images and the pixel value of the pixels in the non-marker regions. The adjustment value is obtained based on the reference pixel value, the area of ​​the non-marker region in the depth prediction image, and the pixel value of the pixel in the non-marker region.

4. The morphological characteristic recognition method as described in claim 1, characterized in that, The steps for extracting features from a segmented image of gastroscopic markers for multiple second preset attributes, and obtaining the second feature quantization value corresponding to each second preset attribute, include: Multiple circles are constructed with the centroid of the marker in the gastroscopy marker segmentation image as the center, and the ratio of the average pixel value within the circle to the change in the circle radius is used as the quantization value of the color gradient feature. Using the centroid of the marker in the segmented image of the gastroscopy marker as the center, the radius of the largest outer circle of the marker is enlarged by a preset ratio. The ratio of the variance of the liquid texture feature vector to the liquid texture feature vector in the enlarged circular area is multiplied by the circle enlargement ratio to obtain the quantization value of the surrounding liquid texture feature. The ratio of the area of ​​the reflected light spot on the surface of the gastroscopic marker to the area of ​​the marker in the segmented image is multiplied by the ratio of the distance between the centroid of the reflected light spot and the centroid of the marker relative to the diagonal of the smallest circumscribed rectangle of the marker, and the reflected light spot feature quantification value is used.

5. The morphological characteristic recognition method as described in claim 1, characterized in that, The steps for extracting features from a depth marker segmentation image using multiple third preset attributes to obtain the third feature quantization value corresponding to each third preset attribute include: The product of the distance between the equivalent centroids of pixels above the threshold and pixels below the threshold in the depth marker segmentation image and the ratio of the ratio of the diagonal of the minimum bounding rectangle of the depth marker segmentation image to the threshold relative to 255 is used as the quantization value of the brightness distribution feature. The ratio of the sum of the number of 0 and 1 alternations in each row to the sum of the number of 0 and 1 alternations in each column after binarization of the image segmented by depth markers is used as the zero-point crossover feature quantization value.

6. The morphological characteristic recognition method as described in claim 1, characterized in that, The steps for extracting features from multiple fourth preset attributes of the segmented image of gastroscopic markers and obtaining the fourth feature quantization value corresponding to each fourth preset attribute include: The two-dimensional image entropy of the segmented image based on gastroscopy markers is used as the quantization value of surface regularity features; The sum of the absolute values ​​of the differences between the pixel values ​​of each column in the row containing the centroid of the gastroscopic marker in the segmented image and the average pixel value of the segmented image is used as the quantization value of the surface roughness feature.

7. The morphological characteristic recognition method according to any one of claims 1 to 6, characterized in that, The first machine learning classifier includes a first feature fitting subnetwork and a first classification subnetwork. The step of inputting multiple first feature quantization values ​​and multiple second feature quantization values ​​into the first machine learning classifier for classification to obtain a first morphology recognition result includes: Input multiple quantized values ​​of the first feature and multiple quantized values ​​of the second feature into the first machine learning classifier; The first feature fitting sub-network performs a weighted summation of multiple first feature quantization values ​​and multiple second feature quantization values ​​according to the weighting coefficients corresponding to multiple first feature quantization values ​​and multiple second feature quantization values, to obtain the first fusion feature value; The first classification sub-network obtains the first morphological recognition result based on the first fusion feature value; The second machine learning classifier includes a second feature fitting subnetwork and a second classification subnetwork. The step of inputting multiple third feature quantization values ​​and multiple fourth feature quantization values ​​into the second machine learning classifier for classification to obtain the second morphology recognition result includes: Multiple quantized values ​​of the third feature and multiple quantized values ​​of the fourth feature are input into the second machine learning classifier; The second feature fitting sub-network performs a weighted summation of multiple third feature quantization values ​​and multiple fourth feature quantization values ​​according to the weighting coefficients corresponding to the multiple third feature quantization values ​​and multiple fourth feature quantization values, to obtain the second fusion feature value; The second classification sub-network obtains the second morphological recognition result based on the second fusion feature value.

8. A morphological characteristic recognition device, characterized in that, The morphological characteristic recognition device includes: The acquisition module is used to acquire gastroscopy images; The depth prediction module is used to perform depth prediction on the gastroscopy image to obtain a depth prediction image; The first segmentation module is used to segment the gastroscopy image by markers, determine the marker regions, and then obtain a segmented gastroscopy image by markers. The second segmentation module is used to segment the depth prediction image according to the marker region to obtain a depth marker segmentation image; The feature extraction module is used to extract features from the depth marker segmentation image using multiple first preset attributes, obtaining a first feature quantization value corresponding to each first preset attribute; and to extract features from the gastroscopy marker segmentation image using multiple second preset attributes, obtaining a second feature quantization value corresponding to each second preset attribute; wherein, the multiple first preset attributes include a distance attribute, a depth attribute, and an area proportion attribute, and the corresponding first feature quantization values ​​are distance feature quantization value, depth feature quantization value, and area proportion feature quantization value, respectively; the multiple second preset attributes include a color gradient attribute, a surrounding liquid texture attribute, and a reflected light spot attribute, and the corresponding second feature quantization values ​​are color gradient feature quantization value, surrounding liquid texture feature quantization value, and reflected light spot feature quantization value, respectively; The first classification module is used to input multiple first feature quantization values ​​and multiple second feature quantization values ​​into the first machine learning classifier for classification to obtain a first morphology recognition result, wherein the first morphology recognition result is any one of depression, bulge and mixed type; The feature extraction module is further configured to, if the first morphology recognition result is a hybrid type, extract features of multiple third preset attributes from the depth marker segmentation image to obtain a third feature quantization value corresponding to each third preset attribute, and extract features of multiple fourth preset attributes from the gastroscopy marker segmentation image to obtain a fourth feature quantization value corresponding to each fourth preset attribute; wherein, the multiple third preset attributes include a brightness distribution attribute and a zero-point crossover attribute, and the corresponding third feature quantization values ​​are the brightness distribution feature quantization value and the zero-point crossover feature quantization value, respectively; the multiple fourth preset attributes include a surface regularity attribute and a surface roughness attribute, and the corresponding fourth feature quantization values ​​are the surface regularity feature quantization value and the surface roughness feature quantization value, respectively; The second classification module is used to input multiple third feature quantization values ​​and multiple fourth feature quantization values ​​into the second machine learning classifier for classification, and obtain the second morphology recognition result, wherein the second morphology recognition result is either diffuse or nested.

9. A morphological characteristic recognition device, characterized in that, The morphological characteristic recognition device includes a processor, a memory, and a morphological characteristic recognition program stored in the memory and executable by the processor, wherein when the morphological characteristic recognition program is executed by the processor, it implements the steps of the morphological characteristic recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a morphological characteristic recognition program, wherein when the morphological characteristic recognition program is executed by a processor, it implements the steps of the morphological characteristic recognition method as described in any one of claims 1 to 7.

Images