Mass spectrometry imaging spatial super-resolution reconstruction method and system based on label propagation network

Abstract

The application discloses a mass spectrum imaging space super-resolution reconstruction method and system based on a label propagation network, and the method comprises the following steps: preprocessing MSI data to obtain a low-dimensional representation of the MSI data; registering the low-dimensional representation with an H&E staining image; performing spatial segmentation on the H&E staining image to obtain a plurality of image blocks, and constructing a label propagation network for each image block; taking the MSI data as the label of a matching node in the network, propagating along the network, and updating the label of the network node; fusing the node label with a statistical-based feature to construct a regression model to calculate the feature of each superpixel, and realizing the spatial super-resolution reconstruction of the MSI. The application fuses the structural information of the H&E image in the MSI, realizes the global multi-modal registration through local affine transformation registration, replaces the H&E staining microscopic image with other modal high-resolution medical images, can realize the mass spectrum imaging super-resolution reconstruction driven by other modalities, and improves the spatial resolution of the MSI.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a mass spectrometry imaging spatial super-resolution reconstruction method and system based on a label propagation network. Background Art

[0002] Spatial metabolomics is a field of omics research focused on the detection and interpretation of metabolites, lipids, drugs, and other small molecules within the spatial context of cells, tissues, and organs. Spatial metabolomics is a rapidly emerging field that addresses numerous biomedical questions, including the tumor molecular microenvironment, early developmental regulation, and metabolic dysregulation during infection and inflammation. This growing research demand has driven the rapid development of mass spectrometry imaging (MSI). MSI is a high-throughput imaging technique used for the identification, quantification, and distribution of proteins, lipids, and chemical metabolites in single cells and complex multicellular tissues. Mass spectrometry imaging plays a key role in the field of metabolomics because it ensures high sensitivity and requires minimal sample preparation, thereby minimizing analyte loss. Furthermore, mass spectrometry imaging can be used to analyze cells in situ, which not only reduces the perturbations encountered in traditional metabolomics but also allows for the correlation of molecular information with spatial context, such as cell morphology or cell-cell interactions. However, the high spatial resolution (~10μm) and point-by-point acquisition strategy of mass spectrometry imaging make MSI experiments last too long, which can easily lead to adverse effects such as sample deterioration. Super-resolution is an effective way to solve this problem. Currently, there are two major methods to improve the spatial resolution of MSI. One is based on hardware technology. For example, Yu Jiancheng et al. proposed a super-resolution biomolecular mass spectrometry imaging device and its working method. Based on the mass spectrometry imaging technology of coded aperture and mass spectrometer, they changed the inherent framework of the existing lattice optical scanning and achieved the goal of super-resolution imaging; Gu Zhongze et al. proposed a high spatial resolution mass spectrometry imaging device, which makes it easier to obtain mass spectrometry imaging with higher spatial resolution; Zhao Weiqian et al. proposed a high spatial resolution laser dual-axis differential confocal spectroscopy-mass spectrometry microscopy method and device, which provides a new method for obtaining high spatial resolution MSI. The second approach is based on algorithmic techniques. For example, Professor Dong Jiyang and his colleagues constructed a super-resolution reconstruction model based on deep neural networks, using H&E-stained images as reference images to achieve MSI super-resolution reconstruction through automatic registration and self-supervised learning. Zhang Xiaozhe and his colleagues also proposed a deep learning-based super-resolution reconstruction method for mass spectrometry images, which overcomes the image resolution limitations of hardware-based methods and reduces experimental costs. Compared to hardware-based approaches to improve MSI spatial resolution, algorithmic techniques are not constrained by physical conditions or economic costs, and have gradually become a practical and efficient technical means for MSI research.

[0003] Because MSI data contains relatively little spatial information, it is often possible to improve the spatial resolution of MSI images from the perspective of image reconstruction by learning from the fine cellular structures of other high-resolution medical images. Hematoxylin-eosin (H&E) stain has become a standard stain for human histological examinations. This staining technique can reveal fine cellular details and, to a certain extent, infer ultrastructural features. Hematoxylin dye is alkaline and can stain basophilic structures such as RNA, nuclei, and ribosomes in the cytoplasm into blue-purple; eosin dye is acidic and can stain eosinophilic structures in tissues into pink, such as intracellular and intercellular proteins, including most tissues in the cytoplasm. Well-stained sections can observe a large number of intracellular details under a light microscope. H&E-stained images are characterized by rich tissue structure and high spatial resolution.

[0004] After tissue sections are prepared for MSI, they are placed in a mass spectrometer and analyzed using a predefined position array. Data is recorded as a continuous list of mass spectra, each associated with a positional pixel. Typically, a mass spectrometric image of a single tissue section comprises tens of thousands of pixels, each described by a mass spectrum, resulting in a very large MSI dataset. Higher spatial resolution means more small pixels, which increases data acquisition time and data load. Furthermore, smaller pixels can excite less sample, resulting in weaker ion signals. From an analytical chemistry perspective, higher spatial resolution means fewer ions are detected above the limit of quantification, resulting in more noise in the dataset. Due to the high dimensionality and complexity of MSI datasets, data dimensionality reduction is necessary to facilitate subsequent analysis. MSI data dimensionality reduction is essentially a feature extraction process, which is effective for global spatial features but may overlook local spatial information. Furthermore, in mass spectrometry imaging, Goodwin et al. have demonstrated that factors such as the storage time of tissue sections at room temperature after sectioning and multiple freeze-thaw cycles significantly impact MSI datasets for peptides, proteins, and drugs. Therefore, various factors can introduce systemic noise, causing errors in both the measured MSI data and the reduced-dimensionality data relative to the true data. These two data characteristics result in MSI data having lower spatial resolution than H&E-stained images. Therefore, researchers can improve the spatial resolution of MSI by incorporating the texture structure of H&E-stained images as external reference information. This not only facilitates medical observation but also helps analyze the pathological information contained in MSI, which has important implications for clinical diagnosis and disease treatment. Summary of the Invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the existing technology and propose a mass spectrometry imaging spatial super-resolution reconstruction method and system based on a label propagation network, which can improve the spatial resolution of MSI.

[0006] The technical solutions of the present invention are as follows.

[0007] On the one hand, a mass spectrometry imaging spatial super-resolution reconstruction method based on a label propagation network comprises:

[0008] In the MSI data preprocessing step, the raw mass spectrometry imaging MSI data is preprocessed to obtain a three-dimensional MSI data matrix X; the spectral data of the data matrix δ is reduced in dimension to obtain a low-dimensional representation E of the MSI data;

[0009] In the image registration step, the low-dimensional representation E of the MSI data is initially aligned with the H&E-stained microscopic image H using rigid body transformation and image cropping to obtain a coarsely registered MSI image. The spatial resolution of E is upsampled to the resolution of the H&E-stained microscopic image, and the upsampled MSI data is precisely registered with the H&E-stained image H using a local affine transformation to obtain the registered MSI upsampled data matrix R.

[0010] The label propagation network construction step is to spatially segment the H&E stained image H to obtain multiple image blocks; a label propagation network G is constructed for each image block, where the nodes of the network are the image pixels of the H&E stained image H, and the connecting edges of the network represent the neighborhood structure similarity of two corresponding pixels;

[0011] In the label generation and propagation step, in each label propagation network G, the MSI upsampled data matrix R is used as the label of the corresponding node in the label propagation network G, and the known label is propagated along the connection edge in the network G to the node with the unknown label, and the node state of the network G is iteratively updated as the MSI feature of each node;

[0012] In the MSI super-resolution reconstruction step, the statistical features of each pixel of the H&E stained image H are extracted and fused with the MSI features of the corresponding nodes of the label propagation network to form a feature vector. A regression relationship is established between the MSI data and the feature vector to reconstruct the MSI image S with high spatial resolution.

[0013] Preferably, the MSI data preprocessing step specifically includes:

[0014] The mass spectrometry signal of each pixel point is subjected to baseline correction, peak alignment, normalization and peak extraction to obtain an MSI ion image;

[0015] The MSI ion image is subjected to unsupervised compression and dimensionality reduction to obtain the MSI dimensionality reduction data E of the tissue section to be analyzed.

[0016] Preferably, the image registration step specifically includes:

[0017] The low-dimensional representation E of the MSI data is upsampled to obtain an MSI data matrix with the same spatial resolution as the H&E stained microscopic image H;

[0018] Extract the feature points of the H&E stained image H and the upsampled MSI data matrix and pair them. Any three adjacent feature points constitute an image patch.

[0019] The H&E stained image H and the corresponding image patch of the upsampled MSI data matrix are registered one by one by affine transformation to obtain the registered MSI upsampled data matrix R.

[0020] Preferably, the label propagation network construction step specifically includes:

[0021] Perform spatial segmentation on the H&E stained image H to obtain multiple image blocks. For each image block, construct a label propagation network G, where the nodes of the network correspond to the pixels of the H&E stained image H.

[0022] The H&E staining value of the local area of ​​the pixel is used as the feature vector, the structural similarity between two pixels is calculated, and it is used as the connection edge weight between the corresponding nodes to obtain the label propagation network G.

[0023] Preferably, the steps of generating and propagating the label specifically include:

[0024] The MSI upsampled data matrix R is used as the label of the corresponding node in the label propagation network G, and the label is propagated along the connection edge in the network G to update the status of other nodes;

[0025] Combining the propagation results of each label propagation network, the MSI characteristics of each node are obtained.

[0026] Preferably, the MSI super-resolution reconstruction step specifically includes:

[0027] The statistical features of each pixel in the H&E stained image H are obtained using multiple feature extraction methods, and the statistical features are fused with the MSI features of the corresponding nodes in the label propagation network to form the fused features of each pixel.

[0028] A regression model is constructed between the fusion features of the pixels and their MSI features to predict the MSI data of the upsampled pixels and obtain the super-resolution reconstructed MSI image S.

[0029] On the other hand, a mass spectrometry imaging spatial super-resolution reconstruction system based on a label propagation network includes:

[0030] The MSI data preprocessing module is used to preprocess the original mass spectrometry imaging MSI data to obtain a three-dimensional MSI data matrix X; the spectral data of the data matrix X is reduced in dimension to obtain a low-dimensional representation E of the MSI data;

[0031] The image registration module is used to preliminarily align the low-dimensional representation E of the MSI data with the H&E-stained microscopic image H using rigid body transformation and image cropping to obtain a coarsely registered MSI image. The spatial resolution of E is upsampled to the resolution of the H&E-stained microscopic image, and the upsampled MSI data is precisely registered with the H&E-stained image H through local affine transformation to obtain the registered MSI upsampled data matrix R.

[0032] The label propagation network construction module is used to spatially segment the H&E stained image H to obtain multiple image blocks. A label propagation network G is constructed for each image block. The nodes of the network are the pixels of the H&E stained image H, and the connecting edges of the network represent the neighborhood structure similarity between two corresponding pixels.

[0033] The label generation and propagation module is used to use the MSI upsampled data matrix R as the label of the corresponding node in each label propagation network G, and propagate the known label along the connection edge in the network G to the node with unknown label, and iteratively update the node state of the network G as the MSI feature of each node;

[0034] The MSI super-resolution reconstruction module is used to extract the statistical features of each pixel in the H&E staining image H and fuse the statistical features with the MSI features of the corresponding nodes in the label propagation network to form a feature vector. A regression relationship is established between the MSI data and the feature vector to reconstruct the MSI image S with high spatial resolution.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention increases the linear and nonlinear information of the H&E image pixels themselves and the surrounding pixels, and at the same time uses the local texture structure information of the high-resolution modality H&E stained image to guide the propagation of the MSI pixel intensity feature, so that the MSI pixel intensity feature is integrated into the local texture structure of the H&E stained image, improving the image perception quality of low-resolution modalities such as MSI, and avoiding the defect of the upsampling technology that its own information is insufficient to achieve super-resolution reconstruction;

[0037] (2) The present invention achieves registration through a patch-based registration method, which alleviates the problem that the linear registration method cannot cope with slice distortion and is closer to the deformation restoration scene in the real world;

[0038] (3) Replacing H&E stained images with high-resolution medical images of other modalities can achieve super-resolution reconstruction of low-resolution modality images driven by other modalities and improve the spatial resolution of MSI. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 Flowchart of a method for spatial super-resolution reconstruction of mass spectrometry imaging based on a label propagation network according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of a method for spatial super-resolution reconstruction of mass spectrometry imaging based on a label propagation network according to an embodiment of the present invention;

[0042] Figure 3 The embodiment of the present invention is to reduce the dimension of low-resolution MSI data E h×v×3 The super-resolution reconstruction image after implementing the present invention and the comparison with the results of the common bicubic interpolation upsampling method and Plas method;

[0043] Figure 4 This is a structural diagram of a mass spectrometry imaging spatial super-resolution reconstruction system based on a label propagation network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is apparent that the embodiments described are only some, not all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0045] See also Figure 1 and Figure 2 As shown, the present invention provides a mass spectrometry imaging spatial super-resolution reconstruction method based on a label propagation network, which includes the following steps.

[0046] S1, MSI data preprocessing steps.

[0047] Specifically, the original MSI spectrum data is preprocessed by baseline correction, peak alignment, normalization, and peak extraction to obtain the MSI ion image of the tissue sample; the spectrum data of the MSI ion image is compressed and dimensionally reduced to obtain the low-resolution MSI dimension-reduced data E h×v×3, where h and v are the number of pixels in the horizontal and vertical directions respectively.

[0048] In this embodiment, the MSI data preprocessing steps are specifically as follows.

[0049] S11, baseline correction was performed using a spline approximation of the baseline at the 10% quantile of ion intensity.

[0050] S12, peak alignment: Align the ion peaks of each pixel with the reference peak to correct the mass-to-charge ratio drift of each ion peak. The peak with the highest correlation with other ion peaks is selected as the reference peak.

[0051] S13, normalizing the spectrum data at each pixel point.

[0052] S14, spectral peak extraction: retain the spectral peaks that meet the following two conditions to achieve spectral peak extraction: first, the intensity of the spectral peak is more than twice the noise intensity, where the noise intensity is determined by the median absolute deviation method; second, within the 10ppm window, the spectral peak is at a local maximum, and finally the MSI ion image data is obtained.

[0053] S15, using uniform manifold approximation and projection (UMAP) to reduce the dimension of the MSI ion image data to obtain the MSI reduced dimension data E h×v×3 .

[0054] S2, image registration step.

[0055] Specifically, the H&E stained microscopic image H corresponding to the MSI data is obtained through rigid body transformation and image cropping. H×V×3 ,; for MSI dimension reduction data E h×v×3 Upsample to obtain data E with the same resolution as the H&E stained image H×V×3 ; Extract H&E staining images H H×V×3 and upsampled MSI data E H×V×3 The feature points are paired and affine transformation is performed to obtain the registered MSI data matrix R H×V×3 Among them, H and V are the number of pixels in the horizontal and vertical directions respectively.

[0056] In this embodiment, the image registration steps are specifically as follows.

[0057] S21, through rigid body transformation and image shearing, obtain high-resolution H&E staining image H that matches the MSI dimensionality reduction data H×V×3 , and reduce the dimension of data E by upsampling MSI h×v×3 , obtain the MSI data matrix E with the same resolution as the H&E staining image H×V×3 ;

[0058] S22, H&E staining image H H×V×3 With MSI upsampled data E H×V×3 Perform feature point extraction and pairing to find multiple pairs of feature points.

[0059] S23, based on multiple pairs of feature points, the H&E stained image H H×V×3 With MSI upsampled data E H×V×3 Divide into HE-MSI corresponding patch pairs, so as to perform affine transformation registration between patches.

[0060] For example, every three adjacent feature points form a Patch, so that H H×V×3 and E H×V×3 Divide into a series of Patch pairs; use affine transformation to transform E H×V×3 Patches are registered one by one to H H×V×3 superior.

[0061] S24, combining the results of each pair of patch registration to obtain the registered MSI data matrix R H×V×3 .

[0062] S3, label propagation network construction step.

[0063] Specifically, H&E staining images H H×V×3 Perform spatial segmentation, build a label propagation network G for each image block, and transform the H&E stained image H H×V×3 Pixels are used as nodes in the H&E stained image H H×V×3 In

[15] , a local area centered on each pixel is used to represent the central pixel. The distance between two pixels is calculated based on this local area. The structural similarity between the two pixels is calculated based on this distance, and the structural similarity is used as the connecting edge between the two corresponding nodes.

[0064] The specific steps for constructing the label propagation network are as follows.

[0065] S31, H&E stained images H H×V×3 In

[15] , a local area centered on each pixel represents the central pixel, the distance between two pixels is calculated based on this local area, and the structural similarity between the two pixels is calculated based on this distance.

[0066] d ij =||SW(x i )-SW(x j )|| 2

[0067] Among them, SW(x i ) represents x i The distance is calculated using a 3×3 window with centered at .

[0068] S32, H&E stained image H is transformed by K-means H×V×3 Divide so that H H×V×3 Segmented into 6 categories (Class1~Class6), the segmentation result F={F k ,(1≤k≤6)}.

[0069] S33, whole H&E stained image H H×V×3 Each pixel is a node X = [X L ;X U ], X L ={x1, x2, ...x l} is a labeled node, X U ={x l+1 , x l+2 ,…x n} is the node of the label to be predicted, based on the segmentation result F={F k ,(1≤k≤6)}, all nodes are divided into 6 corresponding categories.

[0070] S34, according to the segmentation result F={F k ,(1≤k≤6)}, in each category, the structural similarity W between two pixels kij As the connecting edge between the corresponding two nodes, the similarity matrix W of the category is constructed k , W={W k ,(1≤k≤6)}.

[0071] W kij =exp{-d ij / σ 2}

[0072] Here, σ is the control parameter of the propagation length.

[0073] S35, calculate the probability propagation matrix T of each category by normalization k , thus constructing a label propagation network G for each segmentation category (Class). Where T={T k ,(1≤k≤6)}, which is expressed as follows:

[0074]

[0075] Among them, u, v represent the horizontal axis coordinate and vertical axis coordinate of a matrix element respectively, and col represents the matrix W k The total number of columns.

[0076] S4, label generation and propagation steps.

[0077] Specifically, within each label propagation network G, the value of the MSI data matrix E is used as the known label of the corresponding node in G. The label is propagated along the connection edge in the network G to update the status of each node. Combining the propagation results of each label propagation network, the MSI feature of each superpixel is obtained.

[0078] The steps for label generation and propagation are as follows.

[0079] S41, according to the MSI data matrix R after registration H×V×3 As the label of the corresponding node, the MSI original data X L ={x1, x2, ...x l} is labeled Y L ={y i· ,(1≤i≤l)}, upsampling node X U ={x l+1 , x l+2 ,…x n} is labeled Y U ={y i. ,(l+1≤i≤n)},Y U Represents the label to be predicted.

[0080] S42, based on the segmentation result F={F k ,(1≤k≤6)}, divide the labels into corresponding categories, that is, classify the node X and the label matrix Y into 6 categories, (X,Y)={(X k ,Y k ), (1≤k≤6)}, establish a separate label propagation network in each category, propagate the known labels to the nodes with unknown labels until the matrix is ​​converged

[0081] While: {Y k t+1 =T k ×Y k t}

[0082] S43, combining the propagation results of each label propagation network to obtain the H&E staining image H H×V×3 The corresponding MSI features of each pixel.

[0083] S5, MSI super-resolution reconstruction step.

[0084] Specifically, extract the H&E staining image H H×V×3 The statistical features of each pixel in the image are combined with the propagation results of the label network to perform feature fusion; regression modeling is used to obtain the mapping relationship between the fusion features and the MSI data, so as to predict and reconstruct the high-resolution MSI image S H×V×3 .

[0085] The MSI super-resolution reconstruction steps are as follows.

[0086] S51, using principal component analysis (PCA), non-negative matrix factorization (NMF), independent component analysis (ICA), mean, variance, Gaussian filtering, Sobel operator (Sobel) method to extract H&E staining image H H×V×3 The statistical characteristics of each pixel in .

[0087] S52, the MSI features obtained by the label network and the pixel statistical features obtained in S51 are combined into a complete feature group for feature fusion.

[0088] S53, establish feature groups and MSI dimension reduction data E through support vector regression (SVR) h×v×3 The nonlinear mapping relationship between them is then used to predict the MSI super-resolution reconstructed image S. H×V×3 .

[0089] The following example uses super-resolution reconstruction of MSI data and H&E stained images of a mouse brain to illustrate how the generated image retains the texture information of the H&E stained image while also preserving the MSI information, improving image quality and increasing MSI resolution. Figure 3 As shown, the MSI dimension reduction data E h×v×3 The super-resolution reconstruction images after implementing the present invention and the comparison with the results of the ordinary bicubic interpolation upsampling method and the Plas method are shown. The experimental results show that compared with the results of the ordinary upsampling method and the Plas method, the present invention can learn the texture information in the H&E stained image while retaining the MSI information itself, avoiding the defect of the upsampling technology that the inherent information is insufficient to achieve super-resolution reconstruction. At the same time, the registration is achieved through the patch-based registration method, which alleviates the problem that the linear registration method cannot cope with slice distortion. Therefore, the present invention can achieve better super-resolution reconstruction performance.

[0090] In summary, the present invention has developed a method for spatial super-resolution reconstruction of mass spectrometry imaging based on a label propagation network, which uses affine transformation between patches for registration, alleviating the problem that the linear registration method cannot cope with the distortion of experimental slices. The present invention not only increases the linear information and nonlinear information of the H&E pixel itself and the surrounding pixels, but also uses the local texture structure information of the high-resolution modality H&E staining image to guide the propagation of MSI pixel intensity features, so that the MSI pixel intensity features are integrated into the local texture structure of the H&E staining image, improving the image perception quality of the MSI low-resolution modality, and avoiding the defect that the self-information in the upsampling technology is insufficient to achieve super-resolution reconstruction. In addition, the present invention replaces the H&E staining image with a high-resolution medical image of other modalities, and can perform super-resolution reconstruction of mass spectrometry imaging and information of other modalities, further expanding the application of MSI technology in biomedical research.

[0091] See also Figure 4 As shown, this embodiment also discloses a mass spectrometry imaging spatial super-resolution reconstruction system based on a label propagation network, comprising:

[0092] The MSI data preprocessing module 401 is used to preprocess the original mass spectrometry imaging MSI data to obtain a three-dimensional MSI data matrix X; reduce the dimensionality of the spectrum data of the data matrix X to obtain a low-dimensional representation E of the MSI data;

[0093] Image registration module 402 is configured to preliminarily align the low-dimensional representation E of the MSI data with the H&E stained microscopic image H using rigid body transformation and image cropping to obtain a coarsely registered MSI image; upsample the spatial resolution of E to the resolution of the H&E stained microscopic image, and accurately register the upsampled MSI data with the H&E stained image H using a local affine transformation to obtain a registered MSI upsampled data matrix R;

[0094] The label propagation network construction module 403 is used to spatially segment the H&E stained image H to obtain multiple image blocks; a label propagation network G is constructed for each image block, where the nodes of the network are the image pixels of the H&E stained image H, and the connecting edges of the network represent the neighborhood structure similarity between two corresponding pixels;

[0095] The label generation and propagation module 404 is used to use the MSI upsampled data matrix R as the label of the corresponding node in each label propagation network G, propagate the known label along the connection edge in the network G to the node with the unknown label, and iteratively update the node state of the network G as the MSI feature of each node;

[0096] The MSI super-resolution reconstruction module 405 is used to extract the statistical features of each pixel of the H&E stained image H and fuse the statistical features with the MSI features of the corresponding nodes of the label propagation network to form a feature vector; establish a regression relationship between the MSI data and the feature vector, and reconstruct a high spatial resolution MSI image S.

[0097] Specific Implementation of a Mass Spectrum Imaging Spatial Super-resolution Reconstruction System Based on a Label Propagation Network The same mass spectrometry imaging spatial super-resolution reconstruction method based on a label propagation network is not repeated in this embodiment.

[0098] The principles and operation of the present invention are illustrated through the above-mentioned specific implementation examples. These examples are intended to provide readers with a clear understanding framework to grasp the core ideas and key operational points of the present invention. However, it should be made clear that these examples are not intended to limit the scope of application of the present invention. For professionals in this field, various forms of improvements and innovations based on the core ideas of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for spatial super-resolution reconstruction of mass spectrometry imaging based on a label propagation network, characterized in that: include: MSI data preprocessing step: preprocessing the original mass spectrometry imaging MSI data to obtain a three-dimensional MSI data matrix X; Reduce the dimensionality of the spectral data of the data matrix X to obtain the low-dimensional representation E of the MSI data; In the image registration step, the low-dimensional representation E of the MSI data is initially aligned with the H&E-stained microscopic image H using rigid body transformation and image cropping to obtain a coarsely registered MSI image. The spatial resolution of E is upsampled to the resolution of the H&E-stained microscopic image, and the upsampled MSI data is precisely registered with the H&E-stained image H using a local affine transformation to obtain the registered MSI upsampled data matrix R. The label propagation network construction step is to spatially segment the H&E stained image H to obtain multiple image blocks; a label propagation network G is constructed for each image block, where the nodes of the network are the image pixels of the H&E stained image H, and the connecting edges of the network represent the neighborhood structure similarity of two corresponding pixels; In the label generation and propagation step, in each label propagation network G, the MSI upsampled data matrix E is used as the label of the corresponding node in the label propagation network G, and the known label is propagated along the connection edge in the network G to the node with the unknown label. The node state of the network G is iteratively updated as the MSI feature of each node; In the MSI super-resolution reconstruction step, the statistical features of each pixel of the H&E stained image H are extracted and fused with the MSI features of the corresponding nodes of the label propagation network to form a feature vector. A regression relationship is established between the MSI data and the feature vector to reconstruct the high spatial resolution MSI image S. The image registration step specifically includes: The low-dimensional representation E of the MSI data is upsampled to obtain an MSI data matrix with the same spatial resolution as the H&E stained microscopic image H; Extract the feature points of the H&E stained image H and the upsampled MSI data matrix and pair them. Any three adjacent feature points constitute an image patch. Perform affine transformation registration on the H&E stained image H and the corresponding image patch of the upsampled MSI data matrix one by one to obtain the registered MSI upsampled data matrix R; The label propagation network construction step specifically includes: Perform spatial segmentation on the H&E stained image H to obtain multiple image blocks. For each image block, construct a label propagation network G, where the nodes of the network correspond to the pixels of the H&E stained image H. The H&E staining value of the local area of ​​the pixel is used as the feature vector, the structural similarity between two pixels is calculated, and it is used as the connection edge weight between the corresponding nodes to obtain the label propagation network G; The steps of generating and propagating the label specifically include: The MSI upsampled data matrix R is used as the label of the corresponding node in the label propagation network G, and the label is propagated along the connection edge in the network G to update the status of other nodes; Combining the propagation results of each label propagation network, the MSI characteristics of each node are obtained; The MSI super-resolution reconstruction step specifically includes: The statistical features of each pixel in the H&E stained image H are obtained using multiple feature extraction methods, and the statistical features are fused with the MSI features of the corresponding nodes in the label propagation network to form the fused features of each pixel. A regression model is constructed between the fusion features of the pixels and their MSI features to predict the MSI data of the upsampled pixels and obtain the super-resolution reconstructed MSI image S.

2. The method for spatial super-resolution reconstruction of mass spectrometry imaging based on label propagation network according to claim 1, characterized in that: The MSI data preprocessing step specifically includes: The mass spectrometry signal of each pixel point is subjected to baseline correction, peak alignment, normalization and peak extraction to obtain an MSI ion image; The MSI ion image is subjected to unsupervised compression and dimensionality reduction to obtain the MSI dimensionality reduction data E of the tissue section to be analyzed.

3. A mass spectrometry imaging spatial super-resolution reconstruction system based on a label propagation network, characterized in that: Based on the method according to claim 1 or 2, comprising: The MSI data preprocessing module is used to preprocess the original mass spectrometry imaging MSI data to obtain a three-dimensional MSI data matrix X; the spectral data of the data matrix X is reduced in dimension to obtain a low-dimensional representation E of the MSI data; The image registration module is used to preliminarily align the low-dimensional representation E of the MSI data with the H&E-stained microscopic image H using rigid body transformation and image cropping to obtain a coarsely registered MSI image. The spatial resolution of E is upsampled to the resolution of the H&E-stained microscopic image, and the upsampled MSI data is precisely registered with the H&E-stained image H through local affine transformation to obtain the registered MSI upsampled data matrix R. The label propagation network construction module is used to spatially segment the H&E stained image H to obtain multiple image blocks. A label propagation network G is constructed for each image block. The nodes of the network are the pixels of the H&E stained image H, and the connecting edges of the network represent the neighborhood structure similarity between two corresponding pixels. The label generation and propagation module is used to use the MSI upsampled data matrix R as the label of the corresponding node in each label propagation network G, and propagate the known label along the connection edge in the network G to the node with unknown label, and iteratively update the node state of the network G as the MSI feature of each node; The MSI super-resolution reconstruction module is used to extract the statistical features of each pixel in the H&E staining image H and fuse the statistical features with the MSI features of the corresponding nodes in the label propagation network to form a feature vector. A regression relationship is established between the MSI data and the feature vector to reconstruct the MSI image S with high spatial resolution.

Images