CN117036762A - Multi-mode data clustering method - Google Patents

Abstract

The invention discloses a multi-mode data clustering method, which belongs to the technical field of data processing and comprises the steps of obtaining a sample data set; extracting edge characteristics of the image data; extracting differential features of transcriptome data, the differential features including mRNA features and miRNA features; calculating a correlation coefficient matrix of each sample data; a soft threshold value is adopted to carry out nonlinear mapping on the correlation coefficient matrix; calculating the connectivity of each sample and the rest samples; calculating discretized connectivity and corresponding probability to obtain a distance matrix between samples; pre-clustering the sample data by a K-means++ clustering algorithm; converting the inter-sample distance matrix into an inter-sample similarity matrix; constructing a kernel matrix according to the pre-clustering information; iterating the similarity matrix among the samples according to the kernel matrix; integrating the similarity matrix among samples in the mRNA data view, the microRNA data view and the Image data view to obtain a similarity fusion matrix among samples; and clustering samples by a spectral clustering algorithm.

Description

A multimodal data clustering method

Technical field

The invention belongs to the field of data processing technology, and specifically relates to a multi-modal data clustering method.

Background technique

In clinical diagnosis, differences in response to clinical treatment among different patients with the same tumor are often caused by tumor heterogeneity. Many studies have proven the existence of tumor heterogeneity. This heterogeneity may be attributed to Mutations during tumor cell proliferation and differentiation. Tumor heterogeneity will eventually translate into phenotypic differences, which not only refer to differences in the response of different patients with the same tumor to the same drug treatment experiment, but also reflect differences in various biomarkers in the patient's tumor microenvironment.

On the one hand, transcriptomic data is a very critical biomarker for different tumor subtypes. In 2015, RikkeKarlin Jepsen et al. used microRNA expression data in colorectal cancer samples and found that microRNA-92a, microRNA-375, and microRNA-424 were differentially expressed in different colorectal cancer tumor subtypes. Transcriptome data reflects the expression of genes in cells and can provide a large amount of gene expression information, including the expression levels of genes under different conditions, and can reveal differences in cell functions, metabolic pathways, signaling pathways, etc. Transcriptome data usually have high-dimensional feature vectors, which allows more gene expression changes to be considered in cluster analysis, helping to discover small differences. However, due to the high-dimensional nature of transcriptome data, clustering algorithms may cause increased computational complexity when processing large-scale data.

On the other hand, histopathological images play an important role in the early identification and diagnosis of cancer. The use of analysis of pathological images to participate in cancer diagnosis has been applied and developed for many years; Kowal et al. compared and tested the method used for cell nucleus segmentation. Different algorithms are used to determine whether the patient's tumor is benign by analyzing a data set of cancer images, with an accuracy of more than 96%. Histopathology images visually display the morphology and structure of tissue cells, which can help doctors or pathologists quickly observe the characteristics of samples and discover potential abnormalities or pathological changes. However, the interpretation and clustering of histopathological images often requires subjective judgment by professional pathologists, which may be affected by individual differences and subjective experience. Moreover, obtaining high-quality histopathological images requires tissue sectioning, staining and other processing, which is costly and time-consuming.

In summary, transcriptome data and histopathology images each have their own advantages and disadvantages in clustering cancer samples.

Contents of the invention

In order to solve the technical problems in the prior art that the transcriptome data has high data dimensions, the clustering algorithm has high computational complexity, histopathological images are susceptible to the influence of subjective factors, and the clustering accuracy is poor, the present invention provides a multi-modal Data clustering methods.

first

The present invention provides a multi-modal data clustering method, including:

S101: Obtain a sample data set. The sample data set includes multiple sample data, and each sample data includes image data and transcriptome data;

S102: Filter the image data through a bilateral filter;

S103: Introduce the Sobel operator to calculate the gradient information of pixels in the filtered image data. The gradient information includes gradient intensity and gradient direction;

S104: When there is multiple gradient information in the filtered image data, retain the maximum value pixels and suppress the non-maximum value pixels;

S105: Perform denoising on the sample data after non-maximum suppression to obtain edge features of the image data;

S106: Extract differential features of transcriptome data. Differential features include mRNA features and miRNA features;

S107: Calculate the correlation coefficient matrix of each sample data based on the edge features, mRNA features and miRNA features of the sample data;

S108: Use soft thresholds to perform nonlinear mapping on the correlation coefficient matrix;

S109: Calculate the connectivity between each sample and the remaining samples;

S110: Use the Histogram algorithm to discretize the connectivity, calculate the discretized connectivity and the corresponding probability, and obtain the distance matrix between samples;

S111: Use the K-means++ clustering algorithm to pre-cluster the sample data in the mRNA data view, microRNA data view and Image data view to obtain pre-clustering information;

S112: Convert the distance matrix between samples into a similarity matrix between samples in the mRNA data view, microRNA data view and Image data view;

S113: Based on the pre-clustering information, construct the kernel matrix under the mRNA data view, microRNA data view and Image data view;

S114: According to the kernel matrix, iterate the similarity matrix between samples in the mRNA data view, microRNA data view and Image data view;

S115: Combine the similarity matrices between samples under the mRNA data view, microRNA data view and Image data view to obtain the similarity fusion matrix between samples;

S116: Use the spectral clustering algorithm to cluster the samples based on the similarity fusion matrix between samples.

Compared with the prior art, the present invention at least has the following beneficial technical effects:

In the present invention, the transcriptome data and histopathology images are integrated, the edge features of the histopathology images and the mRNA features and miRNA features of the transcriptome data are extracted, and the edge features of the histopathology images and the mRNA features of the transcriptome data are extracted. Perform multi-modal fusion with miRNA features to obtain a similarity fusion matrix between samples, and then perform automatic clustering based on the similarity fusion matrix between samples. For transcriptome data, we only need to focus on the mRNA features and miRNA features, which reduces the data dimension, reduces the complexity of the clustering algorithm, and reduces the influence of subjective factors in the disease assessment process, and improves clustering through multi-modal analysis. Accuracy of assessment.

Description of the drawings

The following will describe the preferred embodiments in a clear and easy-to-understand manner with reference to the accompanying drawings, and further explain the above-mentioned characteristics, technical features, advantages and implementation methods of the present invention.

Figure 1 is a schematic flow chart of a multi-modal data clustering method provided by the present invention.

Detailed ways

In order to explain the embodiments of the present invention or technical solutions in the prior art more clearly, the specific implementation modes of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without exerting creative efforts, other drawings can also be obtained based on these drawings, and obtain Other embodiments.

In order to keep the drawings concise, only the parts related to the invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to make the drawings concise and easy to understand, in some drawings, only one of the components with the same structure or function is schematically illustrated or labeled. In this article, "a" not only means "only one", but can also mean "more than one".

It will be further understood that the term "and/or" as used in the specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items. .

In this article, it should be noted that, unless otherwise clearly stated and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense. For example, it can be a fixed connection or a detachable connection, or Connected in one piece. The connection can be mechanical or electrical. It can be directly connected, or it can be indirectly connected through an intermediary, or it can be an internal connection between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood on a case-by-case basis.

In addition, in the description of the present invention, the terms "first", "second", etc. are only used to differentiate the description and cannot be understood as indicating or implying relative importance.

Example 1

In one embodiment, refer to FIG. 1 of the description, which shows a schematic flow chart of the multi-modal data clustering method provided by the present invention.

The invention provides a multi-modal data clustering method, including:

S101: Obtain sample data set.

The sample data set includes multiple sample data, and each sample data includes image data and transcriptome data.

Specifically, the sample data may be gastrointestinal cancer sample data.

S102: Filter the image data through a bilateral filter.

In a possible implementation, S102 specifically includes:

S1021: Convert sample data into a pixel matrix.

S1022: Nonlinearly fuse the current pixel with pixels within a neighborhood with a radius of 1 pixel:

Among them, g(i,j) represents the pixel value after nonlinear fusion at the current pixel point (i,j), and S(i,j) represents a radius of 1 pixel around the current pixel point (i,j). A collection of pixels within the neighborhood of pixels, (k,l) represents the pixel coordinates around the current pixel (i,j), f(k,l) represents the grayscale at the pixel (k,l) Degree value, w(i,j,k,l) represents the weight parameter between the current pixel point (i,j) and the pixel point (k,l).

Among them, the weight parameter w(i,j,k,l) between the current pixel point (i,j) and the pixel point (k,l) is calculated as:

w(i,j,k,l)=d(i,j,k,l)·r(i,j,k,l)

Among them, d(i,j,k,l) represents the spatial domain weight between the current pixel point (i,j) and the pixel point (k,l), and r(i,j,k,l) represents the current pixel point. The pixel domain weight between (i, j) and pixel point (k, l), σ _d represents the spatial domain standard deviation, and σ _r represents the pixel domain standard deviation.

In the present invention, the bilateral filter is a nonlinear filter that takes into account the gray value and spatial position of the pixels during the filtering process, which can preserve the edge characteristics of the image and make the image look clearer and sharper. Bilateral filters perform nonlinear fusion of the gray values of surrounding pixels, which means that during the filtering process, the fusion relationship between pixels is not a simple weighted average, but is adjusted based on the similarity between pixels. . This better preserves the details and texture information of the image.

S103: Introduce the Sobel operator to calculate the gradient information of pixels in the filtered image data. The gradient information includes gradient intensity and gradient direction.

Among them, the Sobel operator is a commonly used image processing operator used to detect edges in images. It is a discrete difference operator that finds the edge of the image by calculating the gradient of the horizontal and vertical directions of the image pixels.

Among them, gradient information is an important feature in image processing and computer vision, used to describe the degree of change and edge information of the image. It can help us understand the changes in pixel values in the image, so as to locate and identify features such as edges, textures, shapes, etc. in the image.

In a possible implementation, S103 specifically includes:

S1031: Introduce the Sobel operator to calculate the horizontal feature matrix S _x and vertical feature matrix S _y of the filtered sample data:

S1032: Calculate the horizontal gradient G _x and the vertical gradient G _y according to the horizontal feature matrix S _x and the vertical feature matrix S _y :

G _x =S _x I

G _y =S _y I

Among them, I represents the gray value matrix of the sample data after filtering.

S1033: Calculate the gradient intensity G and gradient direction θ of the pixel:

In the present invention, by introducing the Sobel operator and calculating the horizontal and vertical feature matrices of the filtered image data, the gradient information of the image in the horizontal and vertical directions can be obtained. Then, the gradient intensity and gradient direction of the pixel are calculated based on the gradient features, thereby achieving edge detection and feature extraction of the image. This gradient information is very important for subsequent image analysis and processing steps, and can provide richer and more useful information for image data.

S104: When there is multiple gradient information in the filtered image data, the maximum value pixels are retained and the non-maximum value pixels are suppressed.

In a possible implementation, S104 specifically includes:

S1041: Select four preset edge angles, respectively 0,

S1042: When the gradient intensity of the pixel is greater than the gradient of the preset edge angle, the current pixel is determined to be a maximum value pixel and retained. Otherwise, the current pixel is determined to be a non-maximum value pixel and suppressed.

In the present invention, the processing method of retaining the maximum value pixels and suppressing the non-maximum value pixels helps to improve the effect of image edge detection, enhance image features, and reduce the impact of noise, thereby obtaining a more accurate and clear image. Edge information provides a better foundation for subsequent image processing and analysis tasks.

S105: Perform denoising processing on the sample data after non-maximum value suppression to obtain edge features of the image data.

In a possible implementation, S105 specifically includes:

S1051: Set high threshold and low threshold.

S1052: When the edge pixel gradient value of a pixel is greater than the high threshold, the pixel is determined to be a strong edge pixel.

S1053: When the edge pixel gradient value of a pixel is between the low threshold and the high threshold, determine the pixel to be a weak edge pixel.

S1054: When the edge pixel gradient value of a pixel is less than the low threshold, suppress the pixel. Among them, the high threshold TH and the low threshold TL are determined as follows:

TH＝0.5max(H*)

TL＝0.1max(H*)

Among them, H* represents the pixel matrix after non-maximum suppression, and max(H*) represents the maximum value in the pixel matrix after non-maximum suppression.

Among them, denoising the sample data after non-maximum suppression and extracting edge features of the image data can help us extract important visual information from the image for various image processing and computer vision tasks, thereby improving Algorithm performance and accuracy.

S106: Extract differential features of transcriptome data. Differential features include mRNA features and miRNA features.

Among them, mRNA (messenger RNA) is a type of RNA molecule that participates in the gene expression process within cells and transcribes the genetic information in DNA into the amino acid sequence of proteins. mRNA characteristics usually refer to the characteristics of analyzing and describing the expression levels of mRNA in transcriptomic studies.

Among them, miRNA (microRNA) is a type of short non-coding RNA molecules that participates in the regulation of gene expression in cells and inhibits the transcription or translation of target genes by binding to mRNA targets. miRNA signature usually refers to the characteristics of the expression and function of miRNA in miRNAomics research.

S107: Calculate the correlation coefficient matrix of each sample data based on the edge features, mRNA features and miRNA features of the sample data.

Specifically, the Pearson correlation coefficient can be calculated to construct the Pearson correlation coefficient matrix.

S108: Use soft thresholds to perform nonlinear mapping on the correlation coefficient matrix.

In a possible implementation, S108 is specifically:

The correlation coefficient matrix is nonlinearly mapped through the following formula:

a _ij =|a _ij | ^β

Sm _×m ＝[ _aij ] _m×m

Among them, S represents the correlation coefficient matrix, a _ij represents the correlation coefficient between the i-th sample and the j-th sample, β represents the soft threshold, and m represents the number of samples.

Among them, the value range of the soft threshold is β∈[2,20].

In the present invention, soft threshold mapping can enhance the similarity between related samples. In the correlation coefficient matrix, the correlation coefficient values between related samples are higher. After soft threshold mapping, the correlation coefficients between related samples will be further enhanced, making them more closely clustered together to form a clearer category. or cluster. Using soft thresholds for nonlinear mapping of the correlation coefficient matrix can help improve the accuracy and stability of the clustering algorithm and better discover the intrinsic relationships between samples and the potential structure of the data.

S109: Calculate the connectivity between each sample and the remaining samples.

Among them, connectivity measures the degree of correlation between samples. By calculating the connectivity between a sample and other samples, we can know how similar each sample is to other samples, thereby understanding which samples in the data have a strong connection with each other and which samples are relatively independent or uncorrelated.

In a possible implementation, S109 is specifically:

Calculate the connectivity between the current sample and the remaining samples through the following formula:

Among them, k _i represents the connectivity between the i-th sample and the remaining samples.

In the present invention, calculating the connectivity between a sample and other samples helps to understand the degree of correlation between samples, provides important information and basis for the clustering algorithm, thereby improving the accuracy and interpretability of the clustering results.

S110: Use the Histogram algorithm to discretize the connectivity, calculate the discretized connectivity and the corresponding probability, and obtain the distance matrix between samples.

Among them, the Histogram algorithm is a method for data discretization. It maps continuous data into discrete intervals, thereby quantizing the data into different values.

In a possible implementation, S110 is specifically:

Calculate the probability under discretized connectivity through the following formula:

log ₁₀ (p(k _i ))∝(1/log ₁₀ (k _i ))

Among them, p(k _i ) represents the probability under connectivity k _i .

In the present invention, the Histogram algorithm is used to discretize the connectivity and calculate the probability, which can convert the original continuous connectivity data into a discretized feature representation, providing effective information for the construction of the distance matrix between samples, thereby providing subsequent Provides helpful support and guidance on data processing and clustering tasks.

S111: Use the K-means++ clustering algorithm to pre-cluster the sample data in the mRNA data view, microRNA data view and Image data view to obtain pre-clustering information.

Optionally, the value range of the number of clustering categories K of the K-means++ clustering algorithm is K=[2,10].

In the present invention, the sample data is pre-clustered under different data views through the K-means++ clustering algorithm, which can convert high-dimensional data into low-dimensional clustering labels, discover structures and patterns in the data, and provide Subsequent clustering, clustering, and data fusion provide useful information and foundation.

S112: Convert the distance matrix between samples into a similarity matrix between samples in the mRNA data view, microRNA data view and Image data view.

In a possible implementation, S112 is specifically:

S1121: Convert the distance between samples into the similarity between samples through the following formula:

Among them, w _ij represents the similarity between the i-th sample and the j-th sample, d _ij represents the distance between the i-th sample and the j-th sample, and ε _ij represents the distance between the i-th sample and the j-th sample. _{The adaptive parameters between _ _} the average of the distances.

S1122: Construct the similarity matrix P between samples according to the following formula:

Among them, P _ij represents the element value of the i-th row and j-th column in the similarity matrix between samples, and w _ik represents the similarity between the i-th sample and the k-th sample.

In the present invention, converting the distance matrix between samples into the similarity matrix between samples helps to better describe the similarity and correlation between samples, flexibly adjust the similarity calculation, and provide information for subsequent data analysis and fusion. More meaningful input and foundation.

S113: Based on the pre-clustering information, construct the kernel matrix under the mRNA data view, microRNA data view and Image data view.

In a possible implementation, S113 is specifically:

Construct the kernel matrix S through the following formula:

Among them, S _ij represents the element value of the i-th row and j-th column in the kernel matrix, and C _i represents the i-th sample category.

In the present invention, in multi-modal data clustering, each data view provides different types of feature information, and the construction of the kernel matrix can synthesize these feature information to more comprehensively describe the association between samples. sex. By constructing a kernel matrix, sample similarity information under different data views can be integrated, improving clustering accuracy and stability, and reducing computational complexity, thereby playing an important role in multi-modal data clustering.

S114: According to the kernel matrix, iterate the similarity matrix between samples in the mRNA data view, microRNA data view and Image data view.

In a possible implementation, S114 is specifically:

Iterate the similarity matrix between samples through the following formula:

Among them, P ^v represents the similarity matrix between samples under the v-th view, P ^k represents the similarity matrix between samples under the k-th view, S ^v represents the kernel matrix under the v-th view, (·) ^T represents the matrix Transpose.

In the present invention, by iteratively updating the similarity matrix between samples, the sample similarity information under different views can be fused, thereby obtaining a more comprehensive sample similarity matrix. This can make full use of the information provided by different views and improve the performance of the clustering algorithm. During the iteration process, the similarity matrix between samples continuously approaches the kernel matrix, which is equivalent to transferring sample similarity information under different views to other views. Doing so can help make up for the lack of information between different views and improve the accuracy of sample clustering.

S115: Combine the similarity matrices between samples under the mRNA data view, microRNA data view and Image data view to obtain the similarity fusion matrix between samples.

In a possible implementation, S115 is specifically:

Through the following formula, the similarity matrix between samples in the mRNA data view, microRNA data view and Image data view is combined to obtain the similarity fusion matrix between samples:

Among them, P represents the similarity fusion matrix between samples, and P ^v represents the similarity matrix between samples under the vth view.

In the present invention, different views provide respective feature information. By synthesizing the sample similarity matrices under different views, the information from different views can be fused together to obtain more comprehensive and richer sample similarity information. Doing so can improve the clustering algorithm's ability to understand and judge the correlation between samples, thereby improving clustering performance. Sample similarity matrices under different views may reflect different aspects of similarity relationships. By combining the similarity matrices of these views, different aspects of similarity information can be combined to more fully describe the similarity between samples. Doing so helps overcome the possible limitations and biases of a single view and improves the stability and robustness of the clustering algorithm.

S116: Use the spectral clustering algorithm to cluster the samples based on the similarity fusion matrix between samples.

Among them, the spectral clustering algorithm is an unsupervised clustering algorithm based on graph theory and spectral theory. It represents the sample data in the form of a graph and uses the feature vectors of the graph to divide the data samples and classify similar samples into the same category.

Compared with the prior art, the present invention at least has the following beneficial technical effects:

In the present invention, the transcriptome data and histopathology images are integrated, the edge features of the histopathology images and the mRNA features and miRNA features of the transcriptome data are extracted, and the edge features of the histopathology images and the mRNA features of the transcriptome data are extracted. Perform multi-modal fusion with miRNA features to obtain a similarity fusion matrix between samples, and then perform automatic clustering based on the similarity fusion matrix between samples. For transcriptome data, we only need to focus on the mRNA features and miRNA features, which reduces the data dimension, reduces the complexity of the clustering algorithm, and reduces the influence of subjective factors in the disease assessment process, and improves clustering through multi-modal analysis. Assessment accuracy.

The technical features of the above embodiments can be combined in any way. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, all possible combinations should be used. It is considered to be within the scope of this manual.

The above embodiments only express several embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the scope of protection of the patent of the present invention should be determined by the appended claims.

Claims (10)

1. A multimodal data clustering method, characterized by: S101: Obtain a sample data set, the sample data set includes multiple sample data, each of the sample data includes image data and transcriptome data; S102: Filter the image data through a bilateral filter; S103: Introduce the Sobel operator to calculate the gradient information of pixels in the filtered image data, where the gradient information includes gradient intensity and gradient direction; S104: When there is multiple gradient information in the filtered image data, retain the maximum value pixels and suppress the non-maximum value pixels; S105: Perform denoising processing on the sample data after non-maximum value suppression to obtain edge features of the image data; S106: Extract differential features of the transcriptome data, where the differential features include mRNA features and miRNA features; S107: Calculate the correlation coefficient matrix of each sample data according to the edge features, mRNA features and miRNA features of the sample data; S108: Use soft thresholds to perform nonlinear mapping on the correlation coefficient matrix; S109: Calculate the connectivity between each sample and the remaining samples; S110: Use the Histogram algorithm to discretize the connectivity, calculate the discretized connectivity and the corresponding probability, and obtain the distance matrix between samples; S111: Use the K-means++ clustering algorithm to pre-cluster the sample data in the mRNA data view, microRNA data view and Image data view to obtain pre-clustering information; S112: Convert the inter-sample distance matrix into an inter-sample similarity matrix in the mRNA data view, microRNA data view and Image data view; S113: Based on the pre-clustering information, construct the kernel matrix under the mRNA data view, microRNA data view and Image data view; S114: According to the kernel matrix, iterate the similarity matrix between samples under the mRNA data view, microRNA data view and Image data view; S115: Combine the similarity matrices between samples under the mRNA data view, microRNA data view and Image data view to obtain the similarity fusion matrix between samples; S116: Use the spectral clustering algorithm to cluster the samples according to the similarity fusion matrix between samples. 2. The multi-modal data clustering method according to claim 1, characterized in that said S102 specifically includes: S1021: Convert the sample data into a pixel matrix; S1022: Nonlinearly fuse the current pixel with pixels within a neighborhood with a radius of 1 pixel: Among them, g(i,j) represents the pixel value after nonlinear fusion at the current pixel point (i,j), and S(i,j) represents a radius of 1 pixel around the current pixel point (i,j). A collection of pixels within the neighborhood of pixels, (k,l) represents the pixel coordinates around the current pixel (i,j), f(k,l) represents the grayscale at the pixel (k,l) Degree value, w(i,j,k,l) represents the weight parameter between the current pixel point (i,j) and the pixel point (k,l); Among them, the weight parameter w(i,j,k,l) between the current pixel point (i,j) and the pixel point (k,l) is calculated as: w(i,j,k,l)=d(i,j,k,l)·r(i,j,k,l) Among them, d(i,j,k,l) represents the spatial domain weight between the current pixel point (i,j) and the pixel point (k,l), and r(i,j,k,l) represents the current pixel point. The pixel domain weight between (i, j) and pixel point (k, l), σ _d represents the spatial domain standard deviation, and σ _r represents the pixel domain standard deviation. 3. The multi-modal data clustering method according to claim 1, characterized in that said S103 specifically includes: S1031: Introduce the Sobel operator to calculate the horizontal feature matrix S _x and vertical feature matrix S _y of the filtered sample data: S1032: Calculate the horizontal gradient G _x and the vertical gradient G _y according to the horizontal feature matrix S _x and the vertical feature matrix _Sy : G _x =S _x I G _y =S _y I Among them, I represents the gray value matrix of the filtered sample data; S1033: Calculate the gradient intensity G and gradient direction θ of the pixel: 4. The multi-modal data clustering method according to claim 1, characterized in that said S105 specifically includes: S1051: Set high threshold and low threshold; S1052: When the edge pixel gradient value of a pixel is greater than the high threshold, determine the pixel to be a strong edge pixel; S1053: When the edge pixel gradient value of a pixel is between the low threshold and the high threshold, determine the pixel to be a weak edge pixel; S1054: When the edge pixel gradient value of a pixel is less than the low threshold, suppress the pixel. Wherein, the determination method of the high threshold TH and the low threshold TL is: TH＝0.5max(H*) TL＝0.1max(H*) Among them, H* represents the pixel matrix after non-maximum suppression, and max(H*) represents the maximum value in the pixel matrix after non-maximum suppression. 5. The multimodal data clustering method according to claim 1, characterized in that, the S108 is specifically: The correlation coefficient matrix is nonlinearly mapped through the following formula: a _ij =|a _ij | ^β Sm _×m ＝[ _aij ] _m×m Among them, S represents the correlation coefficient matrix, a _ij represents the correlation coefficient between the i-th sample and the j-th sample, β represents the soft threshold, and m represents the number of samples. 6. The multi-modal data clustering method according to claim 1, characterized in that, the S109 is specifically: Calculate the connectivity between the current sample and the remaining samples through the following formula: Among them, k _i represents the connectivity between the i-th sample and the remaining samples. 7. The multi-modal data clustering method according to claim 1, characterized in that said S112 specifically includes: S1121: Convert the distance between samples into similarity between samples through the following formula: Among them, w _ij represents the similarity between the i-th sample and the j-th sample, d _ij represents the distance between the i-th sample and the j-th sample, and ε _ij represents the distance between the i-th sample and the j-th sample. _{The adaptive parameters between _ _} the average of the distance; S1122: Construct the similarity matrix P between samples according to the following formula: Among them, P _ij represents the element value of the i-th row and j-th column in the similarity matrix between samples, and w _ik represents the similarity between the i-th sample and the k-th sample. 8. The multi-modal data clustering method according to claim 1, characterized in that, the S113 is specifically: Construct the kernel matrix S through the following formula: Among them, S _ij represents the element value of the i-th row and j-th column in the kernel matrix, and C _i represents the category of the i-th sample. 9. The multi-modal data clustering method according to claim 1, characterized in that, the S114 is specifically: Iterate the similarity matrix between samples through the following formula: Among them, P ^v represents the similarity matrix between samples under the v-th view, P ^k represents the similarity matrix between samples under the k-th view, S ^v represents the kernel matrix under the v-th view, (·) ^T represents the matrix Transpose. 10. The multi-modal data clustering method according to claim 1, characterized in that, the S115 is specifically: Through the following formula, the similarity matrix between samples in the mRNA data view, microRNA data view and Image data view is combined to obtain the similarity fusion matrix between samples: Among them, P represents the similarity fusion matrix between samples, and P ^v represents the similarity matrix between samples under the vth view.