Esophageal cancer IPCL typing identification method and system based on image classification

Abstract

The application relates to an esophageal cancer IPCL typing recognition method and system based on image classification, and the method comprises the following steps: acquiring an esophageal cancer image, pretreating the esophageal cancer image, extracting a blood vessel image according to color characteristic values of the esophageal cancer image, extracting a distortion characteristic value of the blood vessel image, which is used for evaluating the distortion degree of the blood vessel, extracting a diameter characteristic value of the blood vessel image, which is used for evaluating the diameter of the blood vessel, extracting a divergence characteristic value of the blood vessel image, which is used for evaluating the edge strength of the blood vessel, calculating the similarity of the blood vessel image and a preset example image based on an LW-KNN algorithm according to the extracted characteristic values, and obtaining the category of the blood vessel image based on the similarity result. The application can automatically classify esophageal cancer medical images, can not only divide A / B types, but also can further subdivide the B type, and improves the accuracy and classification efficiency of the classification result.

Description

Technical Field

[0001] This invention relates to the field of image classification technology, and in particular to an image classification-based method and system for esophageal cancer IPCL typing and identification. Background Technology

[0002] Esophageal cancer is a malignant tumor originating from the esophagus, primarily affecting the epithelial cells within the esophagus. Its incidence varies geographically worldwide. Intraepillary capillary loops (IPCLs) are microvascular networks perpendicular to the squamous epithelium and are crucial for assessing the depth of esophageal cancer invasion. Morphological changes in the IPCL structure are key indicators for endoscopically determining the extent of esophageal cancer invasion. As tumor invasion increases, the structure of the intraepillary capillary loops gradually deteriorates. When squamous cell carcinoma invades the mucosa proper, the microvascular loops exhibit an irregular shape with increased diameter. When the tumor invades the muscularis mucosa or superficially invades the submucosa, the abnormal microvascular loop structure gradually disappears, showing significant elongation and deformation. When the tumor deeply invades the submucosa, the IPCL structure is usually completely destroyed, with vessel diameters at least three times that of superficially invading vessels, accompanied by increased diameter and the appearance of anomalously shaped tumor vessels. Therefore, accurately identifying these changes in IPCL is crucial for the assessment of esophageal lesions. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an esophageal cancer IPCL typing identification method and system based on image classification, with the aim of automatically classifying esophageal cancer medical images to improve classification efficiency and accuracy.

[0004] The technical solution adopted in this invention is as follows:

[0005] This invention provides an image classification-based method for esophageal cancer IPCL subtyping and identification, comprising:

[0006] Acquire esophageal cancer images and preprocess them: extract blood vessel images based on the color feature values ​​of the esophageal cancer images;

[0007] The distortion feature values ​​of the blood vessel images are extracted and used to assess the degree of distortion of the blood vessels.

[0008] The diameter feature value of the blood vessel image is extracted and used to assess the diameter of the blood vessel;

[0009] Divergence features of blood vessel images are extracted and used to assess the edge strength of blood vessels.

[0010] Based on the extracted feature values, the similarity between the blood vessel image and the preset example image is calculated using the LW-KNN algorithm. Based on the similarity results, the category of the blood vessel image is determined.

[0011] The preset example image includes X A ,X B1 ,X B2 ,X B3 In the classification phase of the LW-KNN algorithm, there are four categories, first classified by X. A As an example image, the weights of the three feature values ​​are set in the similarity calculation process. The predicted category of the image Y to be classified is calculated, and it is determined whether the predicted category is similar to that of X. A If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For the target image Y to be classified, which belongs to type IPCLB, X is then... B1 ,X B2 ,X B3 Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate the predicted category of the image Y to be classified. Then, compare the predicted category with X in sequence. B1 ,X B2 ,X B3 The category attributes are compared, and the target image Y to be classified is further divided into IPCLB1, IPCLB2 or IPCLB3 types.

[0012] The further technical solution is as follows:

[0013] The process of calculating the predicted category of the image Y to be classified includes:

[0014] The similarity is calculated using a locally weighted method:

[0015]

[0016] Where X is the example image, Y is the target image to be classified, and w i (X,Y) represents the weights of the similarity metric corresponding to the i-th feature value, where X... i Y i Let ||X| represent the eigenvalues ​​corresponding to the i-th eigenvalues ​​of X and Y, respectively. i -Y i || Represents X i With Y i The Euclidean distance, where σ is the bandwidth parameter of the Gaussian kernel;

[0017] Considering the similarity metric and the distance between neighbors, a weight is assigned to each neighbor of the target image to be classified. The weight assignment formula is as follows:

[0018]

[0019] Where ||XY|| is the Euclidean distance between X and Y;

[0020] By introducing the weights of neighbors, the votes of each neighbor are summed in a weighted manner, and the image category is determined through weighted voting.

[0021] Class(Y) = argmac c ∑ Y∈NearestNeighbors(Y) Weight(X,Y)·I(Class(X)=c)

[0022] Where Class(Y) is the predicted class of sample Y, NearNeightbor(Y) represents the nearest neighbor of the target image Y to be classified, and I(·) is an indicator function, I(Class(X)=c) is 1 when the class of the example image X is c, and 0 otherwise.

[0023] The step of extracting vascular images based on the color feature values ​​of esophageal cancer images includes:

[0024] Calculate the equalization color feature value Q of the blood vessel image eq (x,y):

[0025]

[0026] Among them, Q (x,y) Q represents the original color feature value of the pixel at coordinates (x, y) in the image, representing the original color information; min(Q) and max(Q) are the minimum and maximum color feature values ​​among all pixels in the image, respectively.

[0027] Threshold segmentation based on equalized color feature values:

[0028]

[0029] Among them, S (x,y) It represents the pixel at coordinates (x, y) in the segmented image, and T is the threshold used to divide the target region and the background region, which is calculated through histogram analysis.

[0030] The original color feature value Q (x,y) This represents the grayscale value of the image at coordinates (x, y).

[0031] The extraction of the distortion feature values ​​includes:

[0032] Calculate the curvature K of any point i on the blood vessel curve in a blood vessel image. i Obtain the average curvature of the entire blood vessel as the tortuosity feature value K:

[0033]

[0034] Where, subscript i is the number of the i-th pixel on the blood vessel edge curve, and N is the total number of pixels on the blood vessel edge curve;

[0035] The torsion rate K at any point i on the blood vessel edge curve i for

[0036]

[0037] Where K1 and K2 are the curves on the pipe edge that intersect with K i The curvature at two points that are close to each other.

[0038] The extraction of the diameter feature value includes:

[0039] Based on the vascular image, obtain the edge curve of the vascular vessel and the central axis of the vascular vessel;

[0040] Define search area Where t is the width of the blood vessel image, (x,y) is the coordinate of the pixel on the central axis of the blood vessel, and θ is the angle between the curvature normal of the central axis of the blood vessel at the pixel (x,y) and the y-axis.

[0041] The search area includes the edge of the blood vessel. The radius at (x,y) is obtained by calculating the distance from the edge curve within the search area to the nearest pixel point.

[0042] Calculate the radius at each pixel along the central axis of the blood vessel to obtain the diameter feature value R:

[0043]

[0044] Where N is the number of all pixels on the central axis of the blood vessel, and P i It is the radius at the i-th pixel.

[0045] The extraction of the divergence feature value includes:

[0046] The edge strength is calculated using the Sobe operator: G x G y These are two convolutional kernels of the Sobel operator, used to detect edges in the horizontal and vertical directions of the image, respectively; the angle θ = arctan2(G) is calculated for each pixel in each image. y G x );

[0047] Calculate the divergence of the current pixel i: j is the number of the pixels surrounding pixel i, and the total number of these pixels is k;

[0048] Calculate the divergence eigenvalue:

[0049]

[0050] Among them, D m It is the arithmetic mean of the divergence of all pixels in the image.

[0051] This invention also provides an image classification-based IPCL subtyping and identification system for esophageal cancer, comprising:

[0052] The image segmentation module is used to acquire esophageal cancer images and preprocess them: extracting blood vessel images based on the color feature values ​​of the esophageal cancer images;

[0053] The feature extraction module extracts the distortion feature value of the blood vessel image, which is used to assess the degree of distortion of the blood vessel; extracts the diameter feature value of the blood vessel image, which is used to assess the diameter of the blood vessel; and extracts the divergence feature value of the blood vessel image, which is used to assess the edge strength of the blood vessel.

[0054] The calculation module calculates the similarity between the extracted feature values ​​and the preset example images based on the LW-KNN algorithm, and determines the category of the blood vessel image based on the similarity results.

[0055] The preset example image includes X A ,X B1 ,X B2 ,X B3 In the classification phase of the LW-KNN algorithm, there are four categories, first classified by X. A As an example image, the weights of the three feature values ​​are set in the similarity calculation process. The predicted category of the image Y to be classified is calculated, and it is determined whether the predicted category is similar to that of X. A If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For the target image Y to be classified, which belongs to type IPCLB, X is then... B1 ,X B2 ,X B3 Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate the predicted category of the image Y to be classified. Then, compare the predicted category with X in sequence. B1 ,X B2 ,X B3 The category attributes are compared, and the target image Y to be classified is further divided into IPCLB1, IPCLB2 or IPCLB3 types.

[0056] The beneficial effects of this invention are as follows:

[0057] This invention achieves accurate classification of esophageal cancer images through image processing and feature parameter extraction. The locally weighted K-nearest neighbor (LW-KNN) algorithm provides a more comprehensive similarity measure and neighbor weight allocation. By introducing weights into the similarity measure and considering the weights of features, the similarity between images can be more comprehensively characterized. When performing neighbor voting, this algorithm not only considers the distance between neighbors but also fully considers the influence weight of each neighbor on the target sample, thereby improving the robustness and accuracy of the model.

[0058] By adjusting the weights corresponding to various feature values, this invention can not only distinguish between types A and B, but also further subdivide type B, resulting in more refined classification and identification results.

[0059] The typing results obtained in this invention can be used as a dataset for training the IPCL typing model. Based on more refined classification results, the IPCL typing model can be better trained and optimized, improving the diagnostic accuracy of esophageal cancer. It not only has broad application prospects in clinical practice but also provides a powerful tool and method for related research.

[0060] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the logic flow of the method in an embodiment of the present invention. Detailed Implementation

[0062] The specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0063] See Figure 1 This embodiment of an image classification-based IPCL subtyping and identification method for esophageal cancer includes:

[0064] Acquire esophageal cancer images and preprocess them: extract blood vessel images based on the color feature values ​​of the esophageal cancer images;

[0065] The distortion feature values ​​of the blood vessel images are extracted and used to assess the degree of distortion of the blood vessels.

[0066] The diameter feature value of the blood vessel image is extracted and used to assess the diameter of the blood vessel;

[0067] Divergence features of blood vessel images are extracted and used to assess the edge strength of blood vessels.

[0068] Based on the extracted feature values, the similarity between the blood vessel image and the preset example image is calculated using the LW-KNN algorithm. Based on the similarity results, the category of the blood vessel image is determined.

[0069] The preset example image includes X A ,X B1 ,X B2 ,X B3 In the classification phase of the LW-KNN algorithm, there are four categories, first classified by X. A As an example image, the weights of the three feature values ​​are set in the similarity calculation process. The predicted category of the image Y to be classified is calculated, and it is determined whether the predicted category is similar to that of X. A If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For the target image Y to be classified, which belongs to type IPCLB, X is then... B1 ,X B2 ,X B3 Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate the predicted category of the image Y to be classified. Then, compare the predicted category with X in sequence. B1 ,X B2 ,X B3 The category attributes are compared, and the target image Y to be classified is further divided into IPCLB1, IPCLB2 or IPCLB3 types.

[0070] Specifically, esophageal cancer images can be images acquired using NBI endoscopy.

[0071] Specifically, the calculation process for predicting the category of the image Y to be classified includes:

[0072] The similarity is calculated using a locally weighted method:

[0073]

[0074] Where X is the example image, Y is the target image to be classified, and w i (X,Y) represents the weights of the similarity metric corresponding to the i-th feature value, where X... i Y i Let ||X| represent the eigenvalues ​​corresponding to the i-th eigenvalues ​​of X and Y, respectively. i -Y i || Represents X i With Y i The Euclidean distance, where σ is the bandwidth parameter of the Gaussian kernel;

[0075] Considering the similarity metric and the distance between neighbors, a weight is assigned to each neighbor of the target image to be classified. The weight assignment formula is as follows:

[0076]

[0077] Where ||XY|| is the Euclidean distance between X and Y;

[0078] By introducing the weights of neighbors, the votes of each neighbor are summed in a weighted manner, and the image category is determined through weighted voting.

[0079] Class(Y) = argmac c ∑ Y∈NearestNeighbors(Y) Weight(X,Y)·I(Class(X)=c)

[0080] Where Class(Y) is the predicted class of sample Y, NearNeightbor(Y) represents the nearest neighbor of the target image Y to be classified, and I(·) is an indicator function, I(Class(X)=c) is 1 when the class of the example image X is c, and 0 otherwise.

[0081] The similarity calculation principle based on the LW-KNN algorithm in this embodiment is as follows:

[0082] The distortion feature value K, diameter feature value R, and divergence feature value D for each image. ▽ Form a 1*3 feature matrix [K,R,D ▽ The feature values ​​determine the image's position in the feature space, and their similarity is determined by measuring the distance between two positions in the feature space. The choice of feature values ​​and the calculation of weights both affect the final similarity measure. This is mainly reflected in two aspects: distance calculation, where feature values ​​directly influence distance calculation. Differences between different feature values ​​will be amplified or diminished in the distance calculation, thus affecting the similarity measure result; and weight calculation, where the weights w introduced in the similarity measure... i (X,Y) can be adjusted according to the importance of the eigenvalues.

[0083] Specifically, in this embodiment, X A As an example image, the weights of three feature values ​​in the similarity calculation process are set, including: the weight corresponding to the twist rate feature value is 1, and the weights corresponding to the diameter feature value and the divergence feature value are both 0. That is, the IPCLA class and the IPCLB class are selected by twist rate feature value.

[0084] Specifically, in this embodiment, X B1 ,X B2 ,X B3 Replace each image with the example image and reset the weights of the three feature values ​​in the similarity calculation process, including:

[0085] The weight corresponding to the twist feature value is 0.7, and the weights corresponding to the diameter feature value and the divergence feature value are both 1, thus distinguishing the B3 class;

[0086] The weights corresponding to the twist feature value are 0.7, the weights corresponding to the diameter feature value are 0.5, and the weights corresponding to the divergence feature value are 0.7, thus distinguishing the B2 class.

[0087] The weights for the distortion feature values ​​are 1, the diameter feature values ​​are 0.5, and the divergence feature values ​​are 0.3, thus distinguishing the B1 class.

[0088] Specifically, extracting the vascular image based on the color feature values ​​of the esophageal cancer image includes:

[0089] Calculate the equalization color feature value Q of the blood vessel image eq (x,y):

[0090]

[0091] Among them, Q (x,y) Q represents the original color feature value of the pixel at coordinates (x, y) in the image, representing the original color information; min(Q) and max(Q) are the minimum and maximum color feature values ​​among all pixels in the image, respectively.

[0092] Threshold segmentation based on equalized color feature values:

[0093]

[0094] Among them, S (x,y) It represents the pixel at coordinates (x, y) in the segmented image, and T is the threshold used to divide the target region and the background region, which is calculated through histogram analysis.

[0095] The original color feature value Q (x,y) Here is the gray value of the image at coordinates (x, y):

[0096] Q (x,y) =0.299·R(x,y)+0.587·G(x,y)+0.144·B(x,y)

[0097] Where R(x,y), G(x,y), and B(x,y) are the red channel value, green channel value, and blue channel value at (x,y), respectively.

[0098] Specifically, the extraction of the distortion feature values ​​includes:

[0099] Calculate the curvature K of any point i on the blood vessel curve in a blood vessel image. i Obtain the average curvature of the entire blood vessel as the tortuosity feature value K:

[0100]

[0101] Where, subscript i is the number of the i-th pixel on the blood vessel edge curve, and N is the total number of pixels on the blood vessel edge curve;

[0102] The torsion rate K at any point i on the blood vessel edge curve i for

[0103]

[0104] Where K1 and K2 are the curves on the pipe edge that intersect with K i The curvature at two points that are close to each other.

[0105] Specifically, the extraction of the diameter feature value includes:

[0106] Based on the vascular image, obtain the edge curve of the vascular vessel and the central axis of the vascular vessel;

[0107] Define search area Where t is the width of the blood vessel image, (x,y) is the coordinate of the pixel on the central axis of the blood vessel, and θ is the angle between the curvature normal of the central axis of the blood vessel at the pixel (x,y) and the y-axis.

[0108] The search area includes the edge of the blood vessel. The radius at (x,y) is obtained by calculating the distance from the edge curve within the search area to the nearest pixel point.

[0109] Calculate the radius at each pixel along the central axis of the blood vessel to obtain the diameter feature value R:

[0110]

[0111] Where N is the number of all pixels on the central axis of the blood vessel, and P i It is the radius at the i-th pixel.

[0112] Specifically, the extraction of the divergence feature value includes:

[0113] The edge strength is calculated using the Sobe operator: G x G y These are two convolutional kernels of the Sobel operator, used to detect edges in the horizontal and vertical directions of the image, respectively; the angle θ = arctan2(G) is calculated for each pixel in each image. y G x );

[0114] Calculate the divergence of the current pixel i: j is the number of the pixels surrounding pixel i, and the total number of these pixels is k;

[0115] Calculate the divergence eigenvalue:

[0116]

[0117] Among them, D m It is the arithmetic mean of the divergence of all pixels in the image.

[0118] This embodiment also provides an esophageal cancer IPCL typing and identification system based on image classification, including:

[0119] The image segmentation module is used to acquire esophageal cancer images and preprocess them: extracting blood vessel images based on the color feature values ​​of the esophageal cancer images;

[0120] The feature extraction module extracts the distortion feature value of the blood vessel image, which is used to assess the degree of distortion of the blood vessel; extracts the diameter feature value of the blood vessel image, which is used to assess the diameter of the blood vessel; and extracts the divergence feature value of the blood vessel image, which is used to assess the edge strength of the blood vessel.

[0121] The calculation module calculates the similarity between the extracted feature values ​​and the preset example images based on the LW-KNN algorithm, and determines the category of the blood vessel image based on the similarity results.

[0122] The preset example image includes X A ,X B1 ,X B2 ,X B3 In the classification phase of the LW-KNN algorithm, there are four categories, first classified by X. A As an example image, the weights of the three feature values ​​are set in the similarity calculation process. The predicted category of the image Y to be classified is calculated, and it is determined whether the predicted category is similar to that of X. A If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For the target image Y to be classified, which belongs to type IPCLB, X is then... B1 ,X B2 ,X B3 Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate the predicted category of the image Y to be classified. Then, compare the predicted category with X in sequence. B1 ,X B2 ,X B3 The category attributes are compared, and the target image Y to be classified is further divided into IPCLB1, IPCLB2 or IPCLB3 types.

[0123] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying IPCL (Intracytoplasmic Chronic Collateralization) subtypes of esophageal cancer based on image classification, characterized in that, include: Acquire esophageal cancer images and preprocess them: extract blood vessel images based on the color feature values ​​of the esophageal cancer images; The distortion feature values ​​of the blood vessel images are extracted and used to assess the degree of distortion of the blood vessels. The diameter feature value of the blood vessel image is extracted and used to assess the diameter of the blood vessel; Divergence features of blood vessel images are extracted and used to assess the edge strength of blood vessels. Based on the extracted feature values, the similarity between the blood vessel image and the preset example image is calculated using the LW-KNN algorithm. Based on the similarity results, the category of the blood vessel image is determined. The preset example images include In the classification phase of the LW-KNN algorithm, the four categories are first... As an example image, the weights of the three feature values ​​are set in the similarity calculation process to obtain the image to be classified. The predicted category, determine whether the predicted category is consistent with... If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For target images to be classified that belong to type IPCLB... Then Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate to obtain the image to be classified. The predicted categories are then sequentially compared with... The category attributes are compared to further classify the target image. It is divided into IPCLB1 type, IPCLB2 type or IPCLB3 type; The extraction of the distortion feature values ​​includes: Calculate any point on the blood vessel curve in a blood vessel image i curvature Obtain the average curvature of the entire blood vessel as the tortuosity feature value K: Among them, subscript i The first on the curve of the blood vessel edge i The number of each pixel This represents the total number of pixels on the edge curve of the blood vessel. any point on the curve of the blood vessel edge i Twist rate at for in, For the curve of the pipe edge and The curvature at two close points; The extraction of the diameter feature value includes: Based on the vascular image, obtain the edge curve of the vascular vessel and the central axis of the vascular vessel; Define search area ,in, The width of the blood vessel image. These are the coordinates of the pixels on the central axis of the blood vessel. The central axis of the blood vessel at the pixel point The angle between the curvature normal and the y-axis at that point; The search area includes the edge of blood vessels, and the edge curve within the search area is calculated to a point. Get the nearest pixel The radius at that location; Calculate the radius at each pixel along the central axis of the blood vessel to obtain the diameter feature value. : in, It is the number of all pixels along the central axis of the blood vessel. It is in the The radius of each pixel; The extraction of the divergence feature value includes: The edge strength is calculated using the Sobe operator: , , These are two convolutional kernels of the Sobel operator, used to detect edges in the horizontal and vertical directions of the image, respectively; calculating the angle of each pixel in each image. ; Calculate the current pixel i Divergence: , j is the number of the pixels surrounding pixel i, and the total number of them is k; Calculate the divergence eigenvalue: in, It is the arithmetic mean of the divergence of all pixels in the image.

2. The esophageal cancer IPCL typing and identification method based on image classification according to claim 1, characterized in that, Images to be classified The calculation process for the predicted category includes: The similarity is calculated using a locally weighted method: in, For example images, For the target image to be classified, For the corresponding number i The weights of the similarity measures corresponding to the feature values. , They are respectively , The i The eigenvalues ​​corresponding to the eigenvalues, represent and European distance, It is the bandwidth parameter of the Gaussian kernel; Considering the similarity metric and the distance between neighbors, a weight is assigned to each neighbor of the target image to be classified. The weight assignment formula is as follows: in, yes , The Euclidean distance between them; By introducing the weights of neighbors, the votes of each neighbor are summed in a weighted manner, and the image category is determined through weighted voting. in, It is a sample The prediction category Represents the target image to be classified The nearest neighbor, For indicator functions, In the example image The value is 1 when the category is c, and 0 otherwise.

3. The esophageal cancer IPCL typing and identification method based on image classification according to claim 1, characterized in that, The step of extracting vascular images based on the color feature values ​​of esophageal cancer images includes: Calculate the equalization color feature values ​​of blood vessel images : in, In the image at coordinates The original color feature value of the pixel at that location represents the original color information; , These are the minimum and maximum color feature values ​​among all pixels in the image, respectively. Threshold segmentation based on equalized color feature values: in, The segmented image in coordinates The pixel at the location is T, which is the threshold used to divide the target region and the background region, and is calculated through histogram analysis.

4. The esophageal cancer IPCL typing and identification method based on image classification according to claim 3, characterized in that, The original color feature value For the image in coordinates The grayscale value at that location.

5. An image classification-based IPCL subtyping and identification system for esophageal cancer, characterized in that, include: The image segmentation module is used to acquire esophageal cancer images and preprocess them: extracting blood vessel images based on the color feature values ​​of the esophageal cancer images; The feature extraction module extracts the distortion feature value of the blood vessel image, which is used to assess the degree of distortion of the blood vessel; extracts the diameter feature value of the blood vessel image, which is used to assess the diameter of the blood vessel; and extracts the divergence feature value of the blood vessel image, which is used to assess the edge strength of the blood vessel. The calculation module calculates the similarity between the extracted feature values ​​and the preset example images based on the LW-KNN algorithm, and determines the category of the blood vessel image based on the similarity results. The preset example images include In the classification phase of the LW-KNN algorithm, the four categories are first... As an example image, the weights of the three feature values ​​are set in the similarity calculation process to obtain the image to be classified. The predicted category, determine whether the predicted category is consistent with... If the category attributes are the same, it is determined to be of type IPCLA; otherwise, it is determined to be of type IPCLB. For target images to be classified that belong to type IPCLB... Then Replace each image with the example image, reset the weights of the three feature values ​​in the similarity calculation process, and recalculate to obtain the image to be classified. The predicted categories are then sequentially compared with... The category attributes are compared to further classify the target image. It is divided into IPCLB1 type, IPCLB2 type or IPCLB3 type; The extraction of the distortion feature values ​​includes: Calculate any point on the blood vessel curve in a blood vessel image i curvature Obtain the average curvature of the entire blood vessel as the tortuosity feature value K: Among them, subscript i The first on the curve of the blood vessel edge i The number of each pixel This represents the total number of pixels on the edge curve of the blood vessel. any point on the curve of the blood vessel edge i Twist rate at for in, For the curve of the pipe edge and The curvature at two close points; The extraction of the diameter feature value includes: Based on the vascular image, obtain the edge curve of the vascular vessel and the central axis of the vascular vessel; Define search area ,in, The width of the blood vessel image. These are the coordinates of the pixels on the central axis of the blood vessel. The central axis of the blood vessel at the pixel point The angle between the curvature normal and the y-axis at that point; The search area includes the edge of blood vessels, and the edge curve within the search area is calculated to a point. Get the nearest pixel The radius at that location; Calculate the radius at each pixel along the central axis of the blood vessel to obtain the diameter feature value. : in, It is the number of all pixels along the central axis of the blood vessel. It is in the The radius of each pixel; The extraction of the divergence feature value includes: The edge strength is calculated using the Sobe operator: , , These are two convolutional kernels of the Sobel operator, used to detect edges in the horizontal and vertical directions of the image, respectively; calculating the angle of each pixel in each image. ; Calculate the current pixel i Divergence: , j is the number of the pixels surrounding pixel i, and the total number of them is k; Calculate the divergence eigenvalue: in, It is the arithmetic mean of the divergence of all pixels in the image.

Images