Colorectal cancer image segmentation method based on nuclear matroid feature guidance

Abstract

The present invention relates to a colorectal cancer image segmentation method based on nuclear matroid feature guidance, which belongs to the field of medical image processing. First, the input features are obtained by preprocessing the colorectal cancer image, and then a quantized linear transformation is performed to form an initial query matrix and an initial value matrix; secondly, the initial query matrix is ​​subjected to nuclear granulation calculation to form information granules based on nuclear relations; then, a matroid is generated based on the nuclear information granules to obtain nuclear matroid information granules, and the information granules are screened using the basis of the nuclear matroid; then, a feature-guided approximate relationship matrix is ​​obtained based on the closed set in the screened nuclear matroid, and the initial value matrix is ​​subjected to a relationship-guided enhancement operation using the guided relationship matrix to form an enhanced feature matrix; finally, a colorectal cancer segmentation image is generated through a multi-layer stacked convolutional network. This method improves the redundancy capability of uncertain features and improves the robustness of the colorectal cancer image segmentation processing as a whole.

Description

Technical Field

[0001] The invention relates to a colorectal cancer image segmentation method based on nuclear matroid feature guidance, and belongs to the field of medical image processing. Background Art

[0002] The early symptoms of colorectal cancer patients are not obvious, and the early lesions are difficult to judge manually. By segmenting the medical images of the colorectum and highlighting the appearance contours of key parts, favorable conditions can be provided for further disease analysis and diagnosis. At present, the commonly used colorectal cancer image segmentation methods are "threshold segmentation, joint optimization, machine learning" and so on. For example, Sun Xixia et al. disclosed the patent "An HT-29 image segmentation method based on an improved teaching and learning optimization algorithm (application number: 201910258297.4)", Liu Qiao et al. disclosed the patent "A colon cancer image segmentation method based on deep learning (application number: 202311323597.9)", and Shi Kaibing et al. disclosed the patent "A colorectal cancer image segmentation method, system and electronic device (application number: 202410033369.6)". The existing methods ignore the colorectal cancer pathology knowledge contained in the correlation between different feature points of the image, and the accuracy of image segmentation needs to be improved.

[0003] In recent years, feature relationship guided methods represented by attention mechanisms have effectively solved the above problems. Such methods mainly guide and strengthen key features by calculating the dot product relationship of feature bits and constructing a relationship weight matrix, thereby effectively utilizing the information contained in the feature relationship and ultimately achieving effective image segmentation. For example, Wang Li et al. published a patent entitled "A method, device and medium for attention image segmentation (application number: 202110217268.0)", Zhang Shujun et al. published a paper entitled "SAU-Net: A medical image segmentation method based on U-Net and self-attention mechanism [J]. Journal of Electronics, 2022(010):050", and Chen Fang et al. published a paper entitled "Bone ultrasound image segmentation network based on sequence attention and local phase guidance [J]. Acta Automatica Sinica, 2024, 50(5):970-979". However, due to the complex illumination changes and unclear lesion appearance in colorectal cancer images, there are a large number of uncertain features in such medical images. Direct guidance by multiplying feature relationships can easily lead to feature attention guidance deviations, which greatly reduces the robustness of image segmentation processing. Therefore, it is urgent to construct a more reasonable feature guidance method to break through the current technical bottleneck. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a colorectal cancer image segmentation method based on nuclear matroid feature guidance. When there are a large number of uncertain features in the colorectal cancer image, such as complex illumination changes and unclear lesion appearance contours, the colorectal cancer image features are kernelized and calculated, and the information granule structure is further optimized using the basis of the nuclear matroid. The image features are guided and enhanced multiple times by solving the closed set in the nuclear matroid. Finally, a colon cancer region segmentation image is output through a multi-layer superimposed separable variable convolutional network. This method improves the redundancy capability of uncertain features and improves the robustness of colorectal cancer image segmentation processing as a whole.

[0005] The technical solution of the present invention is: a colorectal cancer image segmentation method based on nuclear matroid feature guidance, and the specific steps are as follows:

[0006] Step 1: Preprocess the colorectal cancer image to obtain input features;

[0007] Step 2: Perform two quantized linear transformations on the input features to obtain the initial query matrix and the initial value matrix;

[0008] Step 3: Perform kernel granulation calculation on the initial query matrix to form information granules based on kernel relations;

[0009] Step 4: Generate a matroid on the information granule to obtain the core matroid information granule, use the basis of the core matroid to screen the core matroid information granule, and generate the core matroid feature matrix;

[0010] Step 5: Solve the closed set in the characteristic matrix of the core pseudo-array to obtain the characteristic guided approximate relationship matrix;

[0011] Step 6: Use the feature-guided approximate relationship matrix to perform a relationship-guided enhancement operation on the initial value matrix to form an enhanced feature matrix;

[0012] Step 7: Repeat Step 2-Step 6. If the number of repetitions reaches the set number of enhancement transformation rounds, proceed to Step 8.

[0013] Step 8: Generate colorectal cancer segmentation images through a multi-layer stacked convolutional network.

[0014] Specifically, the separable dilated convolution can adaptively expand the feature receptive field according to spatial changes and improve the accuracy of the output image.

[0015] The Step 1 is specifically as follows:

[0016] The preprocessing includes stretching and resizing the image data;

[0017] The preprocessed image data is input into a multi-layer convolutional neural network to extract features and obtain input features.

[0018] Specifically, the main feature information in the image data is captured through preprocessing to obtain high-quality input data and provide meaningful representation for subsequent tasks.

[0019] The Step 2 is specifically as follows:

[0020] quantizing the input features;

[0021] Extract features from the quantized input features through separable dilated convolution;

[0022] The extracted features are transformed twice linearly to construct an initial query matrix and an initial value matrix respectively.

[0023] Specifically, the linear layer maps the input data from the original feature space to a new feature space, making the important features of the data obvious and suppressing irrelevant features.

[0024] The Step 3 is specifically as follows:

[0025] Based on the Gaussian kernel function and setting the Gaussian kernel granulation hyperparameter, a kernel granulation operation is performed on the initial query matrix.

[0026] Specifically, the obtained granular features contain richer feature attributes.

[0027] The Step 4 is specifically as follows:

[0028] Convert the kernelized matrix into a row echelon matrix;

[0029] If the kernel-granulated matrix is ​​a zero matrix or a full-rank matrix, the column vector is directly returned as the basis of the kernel pseudo-matrix; otherwise, Gaussian elimination is performed on the kernel-granulated matrix to obtain an upper triangular matrix;

[0030] Selecting the column where the pivot element is located from the upper triangular matrix;

[0031] The column where the pivot element is located is extracted from the kernel-granulated matrix as the basis of the kernel pseudo-matrix.

[0032] Specifically, since the prokaryotic matroid information granules contain a large amount of redundant information, information granules with stronger representation capabilities can be obtained through basis screening of the nuclear matroid.

[0033] The closed set in the characteristic matrix of the core matroid in Step 5 is specifically solved as follows:

[0034] Solving the adjacency matrix with zero diagonals in the kernel matroid characteristic matrix;

[0035] Solving the diagonal matrix and the Boolean column matrix in the kernel matroid characteristic matrix;

[0036] Multiplying the kernel matroid feature matrix with the diagonal matrix and the Boolean column matrix to obtain an approximate relationship matrix;

[0037] The closed set of the characteristic matrix of the kernel matroid is obtained by minimizing the adjacency matrix and the approximate relationship matrix element by element.

[0038] Specifically, a feature space with lower approximation capability is obtained through closed set operations within the kernel pseudo-matrix, which effectively handles the uncertainty characteristics in the information granules.

[0039] The enhanced operations in Step 6 are specifically as follows:

[0040] After multiplying the feature-guided approximate relationship matrix with the initial value matrix, an addition operation is performed with the initial value matrix to obtain an enhanced feature matrix.

[0041] Specifically, the feature matrix will enhance the key feature areas based on the initial value matrix to make the target features more obvious.

[0042] The specific number of rounds of enhancement transformation setting in Step 7 is:

[0043] Through multiple experiments on colorectal cancer images, it was demonstrated that the visualization effect of the enhanced feature matrix was best when the number of cycles was set to 8-12 times.

[0044] Specifically, multiple cycles are performed to continuously superimpose and enhance the key feature areas of the generated feature matrix, thereby improving the representation capability of the feature matrix.

[0045] The beneficial effects of the present invention are as follows: the relationship-guided reinforcement in the present invention is based on the nuclear granulation structure, forming a redundant capability for uncertain features; further, the nuclear information granule structure is optimized by the basis of the nuclear matroid, thereby improving the characterization capability of the information granule. Thus, the instability problem of the existing feature relationship-guided method is solved, the robustness of the colorectal cancer image segmentation processing is improved as a whole, and favorable conditions are created for subsequent medical analysis and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall flow chart of the present invention;

[0047] Figure 2 It is a structural diagram of the core matroid relationship guidance model of the present invention;

[0048] Figure 3 Schematic diagram of the convolution operator principle used in the present invention;

[0049] Figure 4 It is a schematic diagram before and after the colorectal cancer image segmentation processing of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0051] Example 1: Figure 1 As shown in FIG. 1 , a colorectal cancer image segmentation method based on nuclear matroid feature guidance is shown in FIG. 1 . The specific steps are as follows:

[0052] Step 1: Preprocess the colorectal cancer image to obtain input features (represents a set of d vectors of length n);

[0053] Step 2: Input features Perform two quantized linear transformations to obtain an initial query matrix and an initial value matrix.

[0054] After the input feature matrix is ​​quantized and convolved, a quantized feature is formed. and . It is a separable dilated convolution. The specific execution process is divided into two steps: separable convolution and dilated convolution. Figure 2 shown.

[0055] Quantify the features and Perform linear transformation: First, normalize the quantized feature matrix, then pass the residual connection, and finally pass the parameterized rectified linear activation function to output the quantized linear feature matrix. and , as the initial query matrix and initial value matrix respectively.

[0056] The data is expressed in the following form:

[0057] , (1)

[0058] (2)

[0059] (3)

[0060] Quantize the input feature matrix: In formula (1), It is the input feature to be quantized. In the present invention, 1-bit quantization, 2-bit quantization, and 4-bit quantization can be selected. and For the convolution kernel quantization operation, you can also choose 1-bit quantization, 2-bit quantization, and 4-bit quantization; and The convolution kernel weights can be learned and obtained through training. is the quantization error reduction factor, which is set to 0.2-0.5 in the present invention. represents the weight threshold, which is set to 0.3-0.5 in the present invention; in formula (2), is a normalization operation, is the residual, is a parameterized rectified linear activation function; Formula (3) is a 1-bit quantization example, where is a sign function;

[0061] Step 3: Initial query matrix Perform nuclear granulation calculations to form information particles based on nuclear relations.

[0062] Initial query matrix (representing a set of d vectors of length n), There are d features in total, and the key matrix is is the transpose of the initial query matrix, that is . Calculated by Gaussian kernel function and The relationship characteristics of nucleation and granulation :

[0063] (4)

[0064] In the formula, It is a feature and The Euclidean distance between Represents a mathematical constant ( .71828) as the base, It is a hyperparameter, and 0.05-0.35 is selected in the present invention.

[0065] Using nucleation-granulation relationships Guided feature vector Each token in constructs an information granule, encapsulating all intrinsic feature information in the feature vector (representing a set of n vectors of length n).

[0066] Step 4: In the information particle Generate a matroid on it, get the nuclear matroid information granules, use the basis of the nuclear matroid to screen the nuclear matroid information granules, and generate the nuclear matroid feature matrix.

[0067] Information particles Convert it to a row echelon matrix, and then determine whether the matrix is ​​a zero matrix or a full rank matrix. If it is, directly return the column vector as the pseudo-matrix basis; otherwise, perform Gaussian elimination on the matrix to solve the upper triangular matrix, select the column where the main element is located from the upper triangular matrix, and remove the column where the main element is located from The obtained pseudo-array basis is the generated core pseudo-array feature matrix , so as to realize the screening of nuclear information particles, the formula is as follows:

[0068] (5)

[0069] in, represents the independent set in the kernelization matrix, Represents a maximum-finding function.

[0070] Step 5: Solve the closed set in the characteristic matrix of the kernel pseudo-array and obtain the characteristic guided approximate relationship matrix.

[0071] The relationship formula between the closed set and the lower approximation is as follows:

[0072] (6)

[0073] (7)

[0074] (8)

[0075] If and only if ;

[0076] In formula (6), represents the set of x features under the nucleation relation in Step 3. Find the closure operator of the union Formula (7) gives the approximate definition of the classical rough set. represents the characteristic lower approximation space that satisfies the kernel granulation (including all elements that definitely belong to the target set X). Formula (8) gives the closed set form of the kernel pseudo-matrix if and only if , X belongs to the subset of lower approximation and has the approximation ability of rough set.

[0077] From the kernel matroid feature matrix Get the lower approximate relationship guidance matrix ( ), firstly, the kernel matroid feature matrix Solve for the adjacency matrix G with zero diagonal; then solve The diagonal matrix and Boolean column matrices , There are n features in total. Then the kernel relationship approximation matrix and the lower approximate relation guidance matrix The formula is as follows:

[0078] (9)

[0079] (10)

[0080] in, Represents the matrix element-by-element minimization function.

[0081] Step 6: Use the feature-guided approximate relationship matrix to perform a relationship-guided enhancement operation on the initial value matrix to form an enhanced feature matrix;

[0082] The feature enhancement mapping formula is expressed as follows:

[0083] (11)

[0084] in, is the feature enhancement matrix;

[0085] In order to effectively preserve the original data information, the feature enhancement matrix is ​​superimposed on the initial value matrix:

[0086] (12)

[0087] in, is the final enhanced matrix.

[0088] Step 7: Repeat Step 2-Step 6. If the number of repetitions reaches the set number of enhancement transformation rounds, proceed to Step 8.

[0089] The final enhanced matrix The initial query matrix and the initial value matrix are reconstructed as input features. The present invention sets the cycle to 8-12 times.

[0090] Step 8: Generate colorectal cancer segmentation images through a multi-layer stacked convolutional network.

[0091] like Figure 3 As shown, the adaptive spatial transformation of the deformable convolution and the dilated convolution are used to obtain a larger spatial receptive field, thereby outputting an effective feature-enhanced image. The present invention sets three layers of convolution stacking, with 128 convolution kernels in the first layer, 256 convolution kernels in the second layer, and 256 convolution kernels in the third layer.

[0092] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A colorectal cancer image segmentation method based on nuclear matroid feature guidance, characterized by: Step 1: Preprocess the colorectal cancer image to obtain input features; Step 2: Perform two quantized linear transformations on the input features to obtain the initial query matrix and the initial value matrix; Step 3: Perform kernel granulation calculation on the initial query matrix to form information granules based on kernel relations; Step 4: Generate a matroid on the information granule to obtain the core matroid information granule, use the basis of the core matroid to screen the core matroid information granule, and generate the core matroid feature matrix; Step 5: Solve the closed set in the characteristic matrix of the core pseudo-array to obtain the characteristic guided approximate relationship matrix; Step 6: Use the feature-guided approximate relationship matrix to perform a relationship-guided enhancement operation on the initial value matrix to form an enhanced feature matrix; Step 7: Repeat Step 2-Step 6. If the number of repetitions reaches the set number of enhancement transformation rounds, proceed to Step 8. Step 8: Generate colorectal cancer segmentation images through a multi-layer stacked convolutional network; The Step 3 is specifically as follows: Based on a Gaussian kernel function and setting Gaussian kernel granulation hyperparameters, a kernel granulation operation is performed on the initial query matrix; The Step 4 is specifically as follows: Convert the kernelized matrix into a row echelon matrix; If the kernel-granulated matrix is ​​a zero matrix or a full-rank matrix, the column vector is directly returned as the basis of the kernel pseudo-matrix; otherwise, Gaussian elimination is performed on the kernel-granulated matrix to obtain an upper triangular matrix; Selecting the column where the pivot element is located from the upper triangular matrix; Extracting the column where the pivot element is located from the kernel-granulated matrix as the basis of the kernel pseudo-matrix; The closed set in the characteristic matrix of the core matroid in Step 5 is specifically solved as follows: Solving the adjacency matrix with zero diagonals in the kernel matroid characteristic matrix; Solving the diagonal matrix and the Boolean column matrix in the kernel matroid characteristic matrix; Multiplying the kernel matroid feature matrix with the diagonal matrix and the Boolean column matrix to obtain an approximate relationship matrix; The closed set of the characteristic matrix of the kernel matroid is obtained by minimizing the adjacency matrix and the approximate relationship matrix element by element. The enhanced operations in Step 6 are specifically as follows: After multiplying the feature-guided approximate relationship matrix with the initial value matrix, an addition operation is performed with the initial value matrix to obtain an enhanced feature matrix.

2. The method for colorectal cancer image segmentation based on nuclear matroid feature guidance according to claim 1, characterized in that: The Step 1 is specifically as follows: The preprocessing includes stretching and resizing the image data; The preprocessed image data is input into a multi-layer convolutional neural network to extract features and obtain input features.

3. The method for colorectal cancer image segmentation based on nuclear matroid feature guidance according to claim 1, characterized in that: The Step 2 is specifically as follows: quantizing the input features; Extract features from the quantized input features through separable dilated convolution; The extracted features are transformed twice linearly to construct an initial query matrix and an initial value matrix respectively.

Images