Intelligent segmentation system for intestinal ulcer regions based on neural networks

Abstract

This invention relates to the field of image segmentation technology, specifically to an intelligent segmentation system for intestinal ulcer regions based on neural networks. Image blocks are segmented using grayscale values. Then, combining the color and distribution characteristics of the abnormal regions themselves, and based on the RGB values ​​and area of ​​the segmented regions, regions of interest (ROIs) exhibiting abnormalities are obtained. Since ROIs contain various anomalies and reflective features, the initial ulcer region is further determined based on the grayscale distribution of pixels within the ROI to filter out non-ulcer influences such as redness and swelling. Then, reflective effects are eliminated based on the direction, distance, and shape features of movement, resulting in a more accurate target ulcer region. The target ulcer region is labeled and input into the neural network for training, achieving intelligent segmentation of intestinal ulcer regions. This invention can eliminate other non-ulcer influences and improve the accuracy of ulcer region image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation, in particular to an intestinal ulcer region intelligent segmentation system based on a neural network. BACKGROUND

[0002] Intestinal ulcer refers to the local damage or ulceration of the intestinal wall, usually accompanied by inflammatory response. Intestinal ulcer region intelligent segmentation usually obtains high-resolution images of the intestinal tract through endoscopy, and uses machine learning technology to automatically identify and separate the region or structure of interest.

[0003] In the prior art, when implementing intestinal ulcer region intelligent segmentation, since there is a certain color difference between the ulcer region and the healthy mucosa region, the region growing algorithm is often used to obtain the part with color different from the healthy mucosa region as the ulcer region. However, in the actual labeling process, non-ulcer abnormalities and reflections will have an impact, causing the identification of the ulcer region to be mixed with other regions, polluting the data source, and further causing poor accuracy of intestinal ulcer region automatic identification and segmentation. SUMMARY

[0004] In order to solve the technical problem that in the related art, non-ulcer abnormalities and reflections will have an impact, causing the identification of the ulcer region to be mixed with other regions, polluting the data source, and further causing poor accuracy of intestinal ulcer region automatic identification and segmentation, the present application provides an intestinal ulcer region intelligent segmentation system based on a neural network, and the technical solution adopted is as follows:

[0005] The present application provides an intestinal ulcer region intelligent segmentation system based on a neural network, which comprises:

[0006] An acquisition module is configured to acquire an endoscopic image of an intestinal ulcer region, and perform grayscale processing to obtain an endoscopic grayscale image.

[0007] A screening module is configured to perform image segmentation on the endoscopic grayscale image based on the grayscale value, to obtain a segmented region, and to screen an interest region showing abnormalities according to the RGB value of the pixel points in the corresponding endoscopic image and the area of the segmented region.

[0008] A region analysis module is configured to determine an initial ulcer region according to the grayscale value distribution of the pixel points in the interest region, to determine the moving direction and distance of the same region in the adjacent frame endoscopic grayscale image according to the moving speed and interval time of the endoscope, to determine a matching region in combination with the moving direction, moving distance, and shape similarity of the initial ulcer regions in different frames, and to perform morphological analysis on the matching regions in different frames to analyze the reflection influence of normal mucosa, and to obtain a target ulcer region.

[0009] The segmentation module is configured to label the target ulcer area and input the labeled target ulcer area into the neural network for training to obtain a segmentation network, and perform intelligent segmentation on the intestinal ulcer area based on the segmentation network.

[0010] Further, the endoscopy grayscale image is segmented based on the grayscale value to obtain a segmentation region, including:

[0011] Based on a region growing algorithm, 50 seed points are randomly and uniformly set in the image, a growth rule is set as a grayscale value difference less than 15, a region growing process is performed, and a segmentation region is obtained.

[0012] Further, the RGB value is a luminance value in different channels of RGB, and the interest region showing an abnormality is screened according to the RGB value of a pixel point in different segmentation regions in the corresponding endoscopy image and the area of the segmentation region, including:

[0013] A luminance screening index of each segmentation region is determined according to a difference between the luminance value of different pixel points in the segmentation region in different channels of RGB and the overall luminance value.

[0014] A ratio of the area of the segmentation region to the area of the endoscopy grayscale image is calculated to obtain an area screening index.

[0015] A difference between the area screening index and the luminance screening index is calculated, and a normalization process is performed to obtain an interest screening coefficient of each segmentation region.

[0016] The segmentation region with the interest screening coefficient less than or equal to a preset background threshold value is combined as the interest region.

[0017] Further, a luminance screening index of each segmentation region is determined according to a difference between the luminance value of different pixel points in the segmentation region in different channels of RGB and the overall luminance value, including:

[0018] The luminance mean value of all pixel points in the endoscopy image in different channels of RGB is calculated.

[0019] The absolute value of the difference between the RGB value of the different pixel points in the segmentation region and the luminance mean value in the corresponding channel is taken as a channel luminance coefficient in the corresponding channel, wherein the channel luminance coefficient includes an R channel luminance coefficient, a G channel luminance coefficient, and a B channel luminance coefficient.

[0020] The mean value of the R channel luminance coefficient, the G channel luminance coefficient, and the B channel luminance coefficient of all pixel points in the segmentation region is taken as the luminance screening index.

[0021] Further, the initial ulcer area is determined according to the grayscale value distribution of the pixel points in the interest region, including:

[0022] performing Otsu threshold segmentation on the interest region to obtain two to-be-analyzed regions;

[0023] taking the to-be-analyzed region with a large mean gray value as an initial ulcer region.

[0024] Further, for the moving speed and interval time of the endoscope, the moving direction and moving distance of the same region in the adjacent frame endoscopic gray images are determined, including:

[0025] taking the opposite direction of the moving direction of the endoscope as the moving direction of the adjacent frame endoscopic gray image at the corresponding moment;

[0026] calculating the product value of the moving speed and the interval time of the adjacent frame endoscopic gray image acquisition as the moving distance.

[0027] Further, the moving direction, the moving distance, and the shape similarity of the initial ulcer regions in different frames are combined to determine the matching region, including:

[0028] taking the initial ulcer regions with matching moving direction and moving distance in adjacent two frames as reference regions;

[0029] determining the morphological center points of the reference regions in adjacent two frames, and superimposing the morphological center points to determine the overlapping part of the two reference regions as an overlapping region;

[0030] performing matching analysis according to the area comparison of the overlapping region and the two reference regions to determine the matching region.

[0031] Further, the matching region is determined according to the area comparison of the overlapping region and the two reference regions, including:

[0032] calculating the mean value of the areas of the two reference regions as a reference mean value;

[0033] calculating the ratio of the area of the overlapping region to the reference mean value as a matching degree;

[0034] taking the two reference regions with a matching degree greater than a preset matching threshold as the matching region.

[0035] Further, morphological analysis is performed on the matching regions in different frames to analyze the reflection of normal mucosa and obtain a target ulcer region, including:

[0036] taking the moving direction of the endoscope as the sixth direction in the 8-neighbor chain code to determine the chain code value sequence of the matching regions in different frames;

[0037] performing dynamic time warping processing on the chain code value sequence of the matching regions in adjacent frames to obtain a DTW value, and taking the inverse of the DTW value as a sequence similarity after normalization processing;

[0038] The matching region with the sequence similarity greater than the preset similarity threshold is taken as the target ulcer region.

[0039] Further, the neural network is a U-NET neural network.

[0040] The present application has the following advantages:

[0041] In the embodiment of the present application, the image block is segmented by the gray value, and then the interest region showing the abnormality is obtained based on the RGB value and area of the segmented region in combination with the color and distribution characteristics of the abnormal region. Since the interest region contains various polyps, redness and reflection conditions, the initial ulcer region is determined based on the gray value distribution of the pixel points in the interest region to screen out the non-ulcer influence such as polyps and redness regions. Then, the reflection influence is eliminated based on the moving direction, moving distance and shape characteristics to obtain a more accurate target ulcer region. The target ulcer region is labeled and input into the neural network for training to realize intelligent segmentation of the intestinal ulcer region. The present application can effectively eliminate a series of influences such as non-ulcer influence and reflection to obtain more accurate and reliable ulcer labeling, extract a more complete ulcer region, establish a more accurate intelligent segmentation system, and realize more accurate automatic recognition and image segmentation of the intestinal ulcer region. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art and their advantages, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0043] Figure 1 A neural network-based intestinal ulcer region intelligent segmentation system structure diagram provided by an embodiment of the present application;

[0044] Figure 2 An endoscope moving schematic diagram provided by an embodiment of the present application;

[0045] Figure 3 A reflection change schematic diagram under different frames provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive objectives, the specific implementation, structure, features and effects of a neural network-based intestinal ulcer region intelligent segmentation system according to the present application are described in detail below in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0048] The specific scheme of the neural network-based intestinal ulcer region intelligent segmentation system provided by the present application is described below in combination with the drawings.

[0049] Please refer to Figure 1 which shows a neural network-based intestinal ulcer region intelligent segmentation system structure diagram provided by an embodiment of the present application, which includes an acquisition module 101, a screening module 102, a region analysis module 103 and a segmentation module 104.

[0050] The acquisition module 101 is used to acquire an endoscopic image of the intestinal ulcer region, and a grayscale processing is performed to obtain an endoscopic grayscale image.

[0051] Intestinal ulcer refers to the local damage or ulceration of the inner wall of the intestine, usually accompanied by inflammatory reaction. Intestinal ulcer region intelligent segmentation usually obtains high-resolution images of the intestine through endoscopy, and uses machine learning technology to automatically identify and separate the region or structure of interest.

[0052] The intestinal ulcer region usually appears as an irregular shape, a blurred and uneven shape in the intestinal image obtained through endoscopy, and usually presents a brightness difference different from the surrounding healthy tissue. The color of the ulcer area is usually darker than the surrounding normal tissue, and may present a darker reddish-brown color, accompanied by varying degrees of bleeding or exudation. The surface of the ulcer in the image may present uneven or granular, and the surrounding may appear signs of inflammatory reaction, such as hyperemia or edema. By analyzing these features, combined with neural network technology, more accurate automatic identification and segmentation of intestinal ulcer region can be achieved.

[0053] In the embodiment of the present application, the endoscope device is used to scan at a stable speed, and continuous frames of high-resolution images of the intestine can be obtained according to fixed interval time to obtain the endoscopic image.

[0054] It should be noted that the endoscopic image in the embodiment of the present application is an image after image preprocessing, each frame of image is subjected to denoising processing using a filtering technology (such as Gaussian filtering), random noise in the image is removed, the brightness and contrast of different images are unified, the image is subjected to normalization processing, the numerical value range is unified, and the neural network processing is facilitated.

[0055] After the endoscopic image is determined, the endoscopic image is subjected to image graying processing, which may be, for example, mean graying processing, so as to obtain an endoscopic grayscale image.

[0056] The screening module 102 is configured to perform image segmentation on the endoscopic grayscale image based on the grayscale value, to obtain a segmentation region, and to screen an interest region showing an abnormality according to the RGB value of the pixel point in the corresponding endoscopic image in different segmentation regions and the area of the segmentation region.

[0057] When the intestinal ulcer region is intelligently segmented, there are interference structures such as polyps in the intestinal tract, which may be misjudged as ulcer regions. In order to obtain a more accurate intelligent segmentation result of the ulcer region, the suspected regions are extracted from the image, and the special features of the ulcer regions in multiple frames of images are combined to analyze the feature performance of each suspected region, so as to screen an interest region showing an abnormality, and then a more accurate intelligent segmentation system is established.

[0058] The abnormal regions in the intestinal tract such as polyps and ulcers have obvious image feature differences with the normal tissue of the intestinal tract in the image, and the differences mainly manifest in color differences. The ulcer base often shows the feature of hyperemia, and is relatively brighter in color compared with the normal intestinal tract, and the internal color is more chaotic due to the mixing of other necrotic tissues and secretions. The normal mucosa region (background) is relatively darker in color, and the connected domain area of the normal mucosa region is much larger than that of other regions, so the normal mucosa region and the abnormal region can be divided by the gray difference and the connected domain size, and the abnormal region is taken as the interest region.

[0059] The normal mucosa region and the abnormal region have obvious differences in gray scale. First, the endoscopic grayscale image is subjected to initial gray feature segmentation, the endoscopic grayscale image is subjected to image segmentation based on the grayscale value, and the segmentation region is obtained, including: 50 seed points are randomly and uniformly set in the image based on a region growing algorithm, the growth rule is set to be that the grayscale value difference is less than 15, the region growing processing is performed, and the segmentation region is obtained.

[0060] The region growing algorithm is an algorithm known to those skilled in the art, the region growing processing is realized by randomly and uniformly setting 50 seed points in the image, and the growth rule is set to be that the grayscale value difference is less than 15, and no further limitation and elaboration is made.

[0061] Wherein, the RGB value in the endoscopic image is the brightness value of red, green and blue color light, and the normal mucosa and abnormal area can be effectively distinguished by the RGB value.

[0062] Further, in some embodiments of the present application, the interest region showing abnormality is screened according to the RGB value of the pixel points in different segmentation regions in the corresponding endoscopic image and the area of the segmentation region, comprising: determining the brightness screening index of each segmentation region according to the difference between the brightness value of different pixel points in the segmentation region under different channels of RGB and the overall brightness value; calculating the ratio of the area of the segmentation region to the area of the endoscopic grayscale image to obtain the area screening index; calculating the difference between the area screening index and the brightness screening index, and normalizing to obtain the interest screening coefficient of each segmentation region; and combining the segmentation region with the interest screening coefficient less than or equal to the preset background threshold as the interest region.

[0063] Wherein, the brightness screening coefficient is the parameter information for screening the segmentation region, since the normal mucosa accounts for the majority, the overall brightness can effectively represent the normal condition, and the brightness screening index of each segmentation region can be determined by the difference between the brightness value of different pixel points in the segmentation region under different channels of RGB and the overall brightness value, so that the abnormal analysis can be realized.

[0064] Further, in some embodiments of the present application, the brightness screening index of each segmentation region is determined according to the difference between the brightness value of different pixel points in the segmentation region under different channels of RGB and the overall brightness value, comprising: calculating the average brightness value of all pixel points in the endoscopic image under different channels of RGB; taking the absolute value of the difference between the RGB value of different pixel points in the segmentation region and the average brightness value under the corresponding channel as the channel brightness coefficient under the corresponding channel, wherein the channel brightness coefficient comprises the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient; and taking the average of the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient of all pixel points in the segmentation region as the brightness screening index.

[0065] Wherein, the absolute value of the difference between the R channel value of the pixel point and the average value of the R channel value of all pixel points is taken as the R channel brightness coefficient; the absolute value of the difference between the G channel value of the pixel point and the average value of the G channel value of all pixel points is taken as the G channel brightness coefficient; and the absolute value of the difference between the B channel value of the pixel point and the average value of the B channel value of all pixel points is taken as the B channel brightness coefficient.

[0066] The greater the values of the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient, the greater the difference between the corresponding channel and the normal mucosa region, that is, the more likely the corresponding pixel point is an abnormal pixel point, and therefore, the average of the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient of all pixel points in the segmentation region is taken as the brightness screening index.

[0067] Since the texture inside the normal mucosa region changes relatively gently, the area of the normal mucosa region is relatively large when region growing is performed, and the texture inside the abnormal region is more chaotic due to the inclusion of polyps, ulcers, and a large amount of necrotic tissue and secretions, and the area of the abnormal region is relatively small. Based on the area characteristics, effective screening can be achieved. The ratio of the area of the segmented region to the area of the endoscopic grayscale image is calculated to obtain an area screening index.

[0068] Since the larger the value of the brightness screening index is, the more abnormal the region is under all color channels, and the larger the area screening index is, the more consistent the region is with the area characteristics of the normal mucosa region, based on the characteristics, the difference between the area screening index and the brightness screening index is calculated, and normalization processing is performed to obtain an interest screening coefficient of each segmented region. It can be known that the larger the value of the interest screening coefficient is, the more likely the corresponding segmented region is the normal mucosa region.

[0069] The preset background threshold is a threshold value of the interest screening coefficient, and the normal mucosa is regarded as the background, so that the background region and the interest region are effectively distinguished based on the interest screening coefficient. The interest region is the region under the influence of the abnormality. The preset background threshold can be specifically, for example, 0.5, that is, the segmented region combination with an interest screening coefficient less than or equal to 0.5 is combined as the interest region, and for the same reason, the segmented region combination with an interest screening coefficient greater than 0.5 is combined as the background region of the normal mucosa.

[0070] The region analysis module 103 is configured to determine an initial ulcer region according to the grayscale value distribution of the pixel points in the interest region, determine the moving direction and the moving distance of the same region in the adjacent frame endoscopic grayscale images according to the moving speed and the interval time of the endoscope, determine a matching region in combination with the moving direction, the moving distance, and the shape similarity of the initial ulcer regions in different frames, and perform morphological analysis on the matching regions in different frames to analyze the reflection of the normal mucosa and obtain a target ulcer region.

[0071] The background region and the interest region are separated by color and connected domain size difference. Since the ulcer region in the interest region is similar to the polyp and the red and swollen region, specific analysis of the interest region is still required to locate the ulcer region.

[0072] The initial ulcer region is determined according to the grayscale value distribution of the pixel points in the interest region, including: performing Otsu threshold segmentation on the interest region to obtain two to-be-analyzed regions; and taking the to-be-analyzed region with a larger average grayscale value as the initial ulcer region.

[0073] The obvious feature of the ulcer area in the image is that there will be obvious necrotic tissue or exudate such as pus around the ulcer area, and the exudate is distributed in the form of spots around or inside the ulcer area, and the color of the exudate is usually white, yellow or gray-green, which shows a high gray value in the gray image.

[0074] In the formula, the Otsu threshold method is a segmentation algorithm known to those skilled in the art, and the image can be divided into two types by the Otsu threshold, one is the polyp and the redness area with generally dark red color and low gray value, and the other is the ulcer area with high gray value. Therefore, the initial ulcer area is obtained based on the Otsu threshold.

[0075] When the intestinal endoscopic image is collected, the light source on the endoscope may cause the reflection of the mucosal tissue, resulting in the appearance of light spots on some areas. Since the light spots appear white in the image, they may also be identified as the initial ulcer area. That is, the endoscope performs illumination in the body, and the normal mucosa reflects the light, thereby producing an abnormal gray area. In the initial ulcer area identification process, the light reflection of the normal mucosa is not effectively distinguished, and therefore, the influence of the light reflection of the normal mucosa needs to be analyzed.

[0076] The initial ulcer area corresponding to the reflection point is caused by the light source on the endoscope, and as the endoscope moves, the reflection point caused by the endoscope also changes with the movement of the endoscope. Therefore, by analyzing the consistency of the initial ulcer area in multiple images, it is determined whether the initial ulcer area corresponds to the real existing adherent area or the reflection caused by the light source of the endoscope.

[0077] Firstly, the image change under the adjacent frames is determined according to the movement of the endoscope. According to the moving speed and the interval time of the endoscope, the moving direction and the moving distance of the same area in the adjacent endoscopic gray images are determined, including: taking the opposite direction of the moving direction of the endoscope as the moving direction of the adjacent endoscopic gray images at the corresponding moment; and taking the product value of the moving speed and the interval time of the adjacent endoscopic gray image collection as the moving distance.

[0078] In the embodiment of the present application, referring to Figure 2 , Figure 2 The endoscope movement schematic diagram provided by one embodiment of the present application is shown in the adjacent two frames of the f frame and the f+1 frame, the moving direction of the interest area is the opposite direction of the moving direction of the endoscope, and the endoscope moves at a constant speed in the adjacent two frames, so as to take the product value of the moving speed and the interval time of the adjacent endoscopic gray image collection as the moving distance. Thus, the moving direction and the moving distance of the interest area corresponding to the endoscopic gray image are determined.

[0079] Further, in some embodiments of the present application, the matching area is determined in combination with the moving direction, the moving distance, and the shape similarity of the initial ulcer areas in different frames, including: matching the initial ulcer areas in the two adjacent frames in the moving direction and the moving distance as the reference areas; determining the morphological center points of the reference areas in the two adjacent frames, and matching the morphological center points to determine the overlapping part of the two reference areas as the overlapping area; and performing matching analysis according to the area comparison of the overlapping area and the two reference areas to determine the matching area.

[0080] In the embodiments of the present application, the morphological center points of the initial ulcer areas can be determined first, and the initial ulcer areas with the most similar distance in the moving direction of the morphological center points in the two adjacent frames are matched as the reference areas, that is, the reference areas correspond to the two initial ulcer areas in the two adjacent frames.

[0081] The overlapping area is the part of the two reference areas that overlaps itself, and the matching area can be analyzed according to the area ratio of the overlapping area in the two reference areas.

[0082] According to the area comparison of the overlapping area and the two reference areas, the matching analysis is performed to determine the matching area, including: calculating the area average of the two reference areas as the reference average; calculating the ratio of the area of the overlapping area to the reference average as the matching degree; and taking the two reference areas with the matching degree greater than the preset matching threshold as the matching area.

[0083] The preset matching threshold is a threshold value of the matching degree, and the preset matching threshold can be 0.5, that is, the two reference areas with the matching degree greater than 0.5 are taken as the matching area.

[0084] Referring to Figure 3 , Figure 3 The present application provides a schematic diagram of reflection changes in different frames; since the movement of the endoscope in different frames will cause the shape of the reflection area itself to change, and the actual ulcer area will not change much, therefore, the reflection is detected and identified based on this feature.

[0085] The identification of the matching area can only represent the area feature, and cannot represent the shape feature, therefore, further analysis of the shape feature is required.

[0086] Further, in some embodiments of the present application, the matching areas of different frames are subjected to morphological analysis to analyze the reflection of normal mucosa and obtain the target ulcer area, including: taking the moving direction of the endoscope as the sixth direction in the 8-neighbor chain code, determining the chain code value sequence of the matching areas of different frames; performing dynamic time warping processing on the chain code value sequence of the matching areas of adjacent frames to obtain a DTW value, and taking the reciprocal of the DTW value and normalizing it as the sequence similarity; taking the matching area with a sequence similarity greater than a preset similarity threshold as the target ulcer area.

[0087] In the embodiments of the present application, the eight chain codes are combined for morphological analysis to determine the eight chain code directions, and then the chain code value sequence of the matching areas of different frames is obtained, and the elements in the chain code value sequence are the chain code directions of the next pixel point relative to the previous pixel point in a fixed direction (such as clockwise or counterclockwise), so the chain code value sequence represents the morphological features.

[0088] The dynamic time warping algorithm can be used to perform dynamic time warping processing on the chain code value sequence of the matching areas of adjacent frames to obtain a DTW value, and the greater the DTW value, the greater the difference between the two chain code value sequences, and the smaller the similarity, so the reciprocal of the DTW value is normalized as the sequence similarity.

[0089] The preset similarity threshold is a threshold value of the sequence similarity, and in the embodiments of the present application, the preset similarity threshold can be specifically, for example, 0.8, that is, the matching area with a sequence similarity greater than 0.8 is taken as the target ulcer area. In this way, the reflection is eliminated, and a more accurate and reliable target ulcer area is determined.

[0090] The segmentation module 104 is configured to label the target ulcer area and input it into a neural network for training to obtain a segmentation network and perform intelligent segmentation of the intestinal ulcer area based on the segmentation network.

[0091] After the target ulcer area is obtained, it can be labeled, and then a neural network is trained based on the characteristics of the target ulcer area to realize intelligent segmentation of the intestinal ulcer area based on the neural network. The neural network is a U-NET neural network, and specifically, the target ulcer area is labeled; the labeled image is taken as a training set for training the neural network to train the U-NET neural network; the trained model is verified and optimized; thus, the intelligent segmentation of the intestinal ulcer area based on the neural network is realized.

[0092] In the embodiment of the present application, the image block is segmented by the gray value, then the interest region showing the abnormality is obtained based on the RGB value and area of the segmented region combined with the color and distribution characteristics of the abnormal region. Since the interest region contains various abnormal conditions and reflection conditions, the initial ulcer region is determined based on the gray value distribution of the pixel points in the interest region to screen out non-ulcer effects such as red swelling, then the reflection effect is eliminated based on the moving direction, moving distance and shape characteristics to obtain a more accurate target ulcer region. The target ulcer region is labeled and input into the neural network for training to realize intelligent segmentation of the intestinal ulcer region. The present application can effectively eliminate a series of influences such as non-ulcer abnormalities and reflections, obtain more accurate and reliable ulcer labels, extract more complete ulcer regions, establish a more accurate intelligent segmentation system, and realize more accurate automatic recognition and segmentation of the intestinal ulcer region.

[0093] It should be noted that the above-mentioned embodiment sequence of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0094] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

Claims

1. A neural network-based intelligent segmentation system for intestinal ulcer regions, characterized by, The method comprises the following steps: An acquisition module is configured to acquire an endoscopic image of an intestinal ulcer region, and to obtain an endoscopic grayscale image through grayscale processing; A screening module is configured to perform image segmentation on the endoscopic grayscale image based on a grayscale value, to obtain a segmented region, and to screen an interest region that exhibits an abnormality according to RGB values of pixel points in different segmented regions in a corresponding endoscopic image and areas of the segmented regions; A region analysis module is configured to determine an initial ulcer region according to a grayscale value distribution of the pixel points in the interest region, to determine a moving direction and a moving distance of a same region in adjacent frame endoscopic grayscale images according to a moving speed and an interval time of an endoscope, to determine a matching region in combination with the moving direction, the moving distance, and a shape similarity of the initial ulcer regions in different frames, and to perform morphological analysis on the matching regions in different frames to analyze a reflection influence of normal mucosa and to obtain a target ulcer region; the region analysis module is specifically configured to: take a reverse direction of a moving direction of the endoscope as a moving direction of the adjacent frame endoscopic grayscale images at a corresponding moment; calculate a product value of the moving speed and an interval time at which the adjacent frame endoscopic grayscale images are collected as the moving distance; take an initial ulcer region that matches the moving direction and the moving distance in adjacent two frames as a reference region; determine morphological center points of the reference regions in the adjacent two frames, and superimpose the morphological center points to determine an overlapping part of the two reference regions as an overlapping region; calculate an area average value of the two reference regions as a reference average value; calculate a ratio of an area of the overlapping region to the reference average value as a matching degree; take the two reference regions with the matching degree greater than a preset matching threshold as the matching region; take the moving direction of the endoscope as a sixth direction in an 8-neighbor chain code to determine a chain code value sequence of the matching regions in different frames; perform dynamic time warping processing on the chain code value sequences of the matching regions in adjacent frames to obtain a DTW value, and normalize a reciprocal of the DTW value as a sequence similarity; take the matching region with the sequence similarity greater than a preset similarity threshold as the target ulcer region; A segmentation module is configured to label the target ulcer region, input the target ulcer region into a neural network for training, obtain a segmentation network, and perform intelligent segmentation of the intestinal ulcer region based on the segmentation network.

2. The neural network-based intelligent intestinal ulcer region segmentation system of claim 1, wherein, The image segmentation on the endoscopic grayscale image based on the grayscale value comprises: Based on a region growing algorithm, 50 seed points are randomly and uniformly set in the image, a growth rule is set as a grayscale value difference less than 15, and region growing processing is performed to obtain the segmented region.

3. The neural network-based intelligent intestinal ulcer region segmentation system of claim 1, wherein, The RGB values are luminance values in different channels of RGB, and the interest region that exhibits the abnormality is screened according to the RGB values of the pixel points in different segmented regions in a corresponding endoscopic image and areas of the segmented regions, which comprises: A luminance screening index of each segmented region is determined according to a difference between the luminance values of different pixel points in the segmented region in different channels of RGB and a whole luminance value; An area screening index is obtained by calculating a ratio of the area of the segmented region to the area of the endoscopic grayscale image; A difference value between the area screening index and the luminance screening index is calculated, and a normalization processing is performed to obtain an interest screening coefficient of each segmented region. Combine the segmented regions with the interest screening coefficient less than or equal to the preset background threshold value as the interest region.

4. The neural network-based intelligent intestinal ulcer region segmentation system of claim 3, wherein, Determine a brightness screening index of each segmented region according to the difference between the brightness value of each pixel in the segmented region under different channels of RGB and the overall brightness value, including: Calculate the brightness mean value of all pixel points in the endoscopic image under different channels of RGB; Take the absolute value of the difference between the RGB value of each pixel in the segmented region and the brightness mean value under the corresponding channel as the channel brightness coefficient under the corresponding channel, wherein the channel brightness coefficient includes the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient; Take the mean value of the R channel brightness coefficient, the G channel brightness coefficient and the B channel brightness coefficient of all pixel points in the segmented region as the brightness screening index.

5. The neural network-based intelligent intestinal ulcer region segmentation system of claim 1, wherein, Determine the initial ulcer region according to the gray value distribution of the pixel points in the interest region, including: Perform Otsu threshold segmentation on the interest region to obtain two to-be-analyzed regions; Take the to-be-analyzed region with a larger gray value mean value as the initial ulcer region.

6. The neural network-based intelligent intestinal ulcer region segmentation system of claim 1, wherein, The neural network is a U-NET neural network.

Images