H&e staining microscopic image driven mass spectrometry imaging super-resolution reconstruction method

Abstract

The application relates to an H&E staining microscopic image-driven mass spectrometry imaging super-resolution reconstruction method, which relates to image processing. The MSI original data is subjected to pretreatment such as spectral peak alignment, spectral peak extraction, normalization, baseline correction and compression dimension reduction, so as to obtain ion image data and dimension reduction data of the MSI; a super-resolution reconstruction model is constructed based on a deep neural network; the high-resolution H&E microscopic image and the MSI dimension reduction data are taken as inputs, the model is automatically registered and self-supervised learning is performed, so that the super-resolution reconstruction of the MSI dimension reduction data is realized; and the super-resolution reconstruction of the MSI ion image is realized by using transfer learning. The super-resolution image of the MSI dimension reduction data is used for iterative registration of the H&E image, so that the correction error of the H&E image caused by factors such as small information quantity of a single ion image, low signal-to-noise ratio and low spatial resolution is avoided; and the H&E staining microscopic image is replaced by a high-resolution medical image of other modalities, so that the super-resolution reconstruction of mass spectrometry imaging driven by other modalities can be realized.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a mass spectrometry imaging super-resolution reconstruction method driven by H&E staining microscopic images. Background Art

[0002] Mass spectrometry imaging (MSI) is a label-free, high-throughput analytical technique that enables qualitative, quantitative, and localized analysis of molecular information in biological samples. In recent years, MSI has become an important analytical tool in research such as drug development, tumor heterogeneity analysis, and disease-related biomarker screening. However, due to its low spatial resolution and data acquisition efficiency, achieving single-cell resolution is difficult, limiting the in-depth application of MSI technology in medical research.

[0003] Taking MALDI-MSI as an example, due to limitations in the laser scanning spot, mass spectrometer sensitivity, platform step size, and data acquisition rate, the spatial resolution of the resulting images is low, typically ranging from 50 to 200 μm. Mammalian cells, with diameters of approximately 1 to 20 μm, are difficult to study at the single-cell level using MALDI-MSI. However, given that single-cell studies can more reliably reveal the molecular mechanisms of disease, identify distinct molecular phenotypes, and facilitate the development of targeted drugs, improving the spatial resolution of MSI technology is highly significant.

[0004] Currently, there are two main approaches to improving the spatial resolution of MSI technology. The first is hardware-based technology, but due to physical constraints, the room for improvement is relatively limited. The second is super-resolution methods that integrate high-resolution imaging technologies from other modalities. For example, Raf Van de Plas et al. achieved super-resolution reconstruction of hematoxylin and eosin (H&E)-stained microscopic images by constructing a multivariate linear regression model between MSI ion images and MSI ion images. However, there are some limitations. First, this method uses MSI ion images to register H&E-stained images, but MSI ion images have low information content, low signal-to-noise ratio, and low spatial resolution, resulting in large registration errors for H&E-stained images. Second, this method super-resolves MSI images by constructing a linear regression model between H&E staining image information and MSI ion image information. However, this linear model has poor generalization ability, resulting in unsatisfactory MSI super-resolution results. Other super-resolution methods also have some problems. For example, their results rely on pre-processed image registration, ignoring the global and texture features of the H&E modality. Alternatively, methods based solely on external reference information may produce results with data distributions that differ from those of the MSI modality, making the results unreliable. Therefore, a more efficient super-resolution technique is urgently needed to achieve comprehensive characterization of MSI data.

[0005] Hematoxylin-eosin (H&E) stained microscopic images are a relatively mature histological imaging technology and a commonly used method for clinical pathology examination. Compared with MSI data, H&E stained microscopic images have rich histological anatomical structure information and higher spatial resolution. If the internal information of MSI and the external information of H&E stained microscopic images can be combined to improve the spatial resolution of MSI, it will not only be beneficial for medical observation, but also help to analyze the pathological information contained in MSI, which is of great significance for clinical diagnosis and disease treatment. Therefore, the development of MSI super-resolution technology driven by H&E stained microscopic images has important scientific significance and practical value. Summary of the Invention

[0006] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and provide a H&E staining microscopic image-driven mass spectrometry imaging super-resolution reconstruction method that uses deep learning to establish an end-to-end deep neural network to achieve better super-resolution reconstruction performance and improve the spatial resolution capability of MSI data.

[0007] The present invention preprocesses mass spectrometry imaging (MSI) raw data through peak alignment, peak extraction, normalization, baseline correction, and compression and dimensionality reduction to obtain MSI ion image data and dimensionality reduction data; constructs a super-resolution reconstruction model based on a deep neural network; uses high-resolution H&E microscopic images and MSI dimensionality reduction data as input, and allows the model to perform automatic registration and self-supervised learning to achieve super-resolution reconstruction of MSI dimensionality reduction data; and utilizes transfer learning to achieve super-resolution reconstruction of MSI ion images.

[0008] The present invention comprises the steps of:

[0009] 1) MSI data acquisition and preprocessing: A mass spectrometer is used to acquire mass spectrometry data of the tissue section to be analyzed on a two-dimensional dot matrix. After preprocessing the obtained mass spectrometry data, an MSI ion image X of the tissue section to be analyzed is obtained;

[0010] 2) H&E staining microscopic image analysis: Take an adjacent section of the tissue section to be analyzed, wash it, stain it with hematoxylin, stain it with eosin, dehydrate it, make it transparent, dry it, mount it, and examine it under a microscope to obtain a high-resolution H&E staining microscopic image Y;

[0011] 3) MSI data compression: Perform unsupervised compression and dimensionality reduction on the spectrum data of the MSI ion image X to obtain the MSI dimensionality reduction data E of the tissue section to be analyzed;

[0012] 4) Construction of super-resolution reconstruction model: Constructing an end-to-end super-resolution reconstruction model based on deep neural network Designing the model's loss function And the corresponding model optimizer to initialize the parameters;

[0013] 5) Training of super-resolution reconstruction model for dimensionality reduction data: Super-resolution reconstruction model Self-supervised learning is used to obtain the super-resolution reconstructed image S of the dimensionality-reduced data E of the tissue section MSI to be analyzed;

[0014] 6) Transfer learning of ion image super-resolution model: Transfer super-resolution reconstruction model Migrate to Ionic Image X i Super-resolution model In

[15] , transfer learning is used to achieve super-resolution reconstruction of MSI ion images.

[0015] In step 1), the specific steps of MSI data collection and preprocessing include:

[0016] ① Cryosection the sample to be analyzed. After thawing, the tissue sections are placed on cold indium tin oxide (ITO)-coated microscope slides and dried in a vacuum desiccator.

[0017] ② The dried tissue slice surface is coated with a matrix, and the mass spectrometer is used to collect the mass spectrometry signal of each pixel point of the tissue slice along the two-dimensional dot matrix direction;

[0018] ③ Perform peak alignment, peak extraction, normalization and baseline correction on the mass spectrometry signals of each pixel to obtain the MSI ion image data X = {X i ,i=1,2,…,I}.

[0019] In step 2), the specific steps of the H&E staining microscopic image analysis include:

[0020] ① Select an adjacent tissue section to be analyzed for MSI, wash the section, and stain the nucleus and cytoplasm with hematoxylin (H) and eosin (E), respectively;

[0021] ② Dehydrate, clear, air dry and seal the H&E stained tissue sections;

[0022] ③ The processed H&E stained sections are examined under a microscope, and high-resolution H&E staining image data Y is collected.

[0023] In step 5), the specific steps of training the dimensionality reduction data super-resolution model are as follows: using the MSI dimensionality reduction data E and the high-resolution H&E stained microscopic image Y as training samples, and letting the super-resolution reconstruction model Perform self-supervised learning and iteratively update model parameters until the loss function After converging to the optimal value, a super-resolution reconstructed image S of the dimension-reduced data E of the tissue slice MSI to be analyzed is obtained.

[0024] In step 6), the specific method of transfer learning of the ion image super-resolution model is: Migrate to the i-th ion image X i Super-resolution model In the middle; keep the parameters of the registration module unchanged; use the ion image X i and high-resolution H&E image Y training Other parameters are used to obtain the super-resolution ion image R i , realizing MSI super-resolution reconstruction of tissue sections R={R i ,i=1,2,…,I}.

[0025] Compared with the prior art, the advantages of the present invention are:

[0026] 1. Use the super-resolution image of MSI dimension reduction data to iteratively register the H&E image to avoid the influence of factors such as the small amount of information, low signal-to-noise ratio and low spatial resolution of the ion image on the registration accuracy of the H&E image;

[0027] 2. Super-resolution reconstruction of single ion images is achieved through transfer learning, making the MSI super-resolution reconstruction results more robust and reliable;

[0028] 3. This invention is not only applicable to 3D MSI images but also enables super-resolution reconstruction of MSI ion images through transfer learning. By replacing H&E-stained microscopic images with high-resolution medical images of other modalities, super-resolution reconstruction of mass spectrometry images and information from other modalities can be achieved, further expanding the application of MSI technology in biomedical research. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a system framework diagram of the present invention.

[0030] Figure 2 To reduce the dimension of MSI data E h×v×3 Super-resolution reconstruction image after implementing the present invention and comparison with the Raf Van de Plas method.

[0031] Figure 3 For the ion image (X i ) h×v Super-resolution reconstruction image after implementing the present invention. DETAILED DESCRIPTION

[0032] In order to make the advantages of the technical solutions of the present invention more clear, the present invention will be further described in the following embodiments with reference to the accompanying drawings.

[0033] The embodiment of the present invention includes the following steps:

[0034] S1: The tissue sample to be analyzed is frozen and sliced. After thawing, it is placed on an indium tin oxide (ITO)-coated microscope slide and dried in a vacuum desiccator for pretreatment. The dried slice surface is coated with a matrix and data is collected using a MALDI-TOF imager. After matrix-assisted laser desorption ionization time-of-flight mass spectrometry (MALDI-TOF) analysis, raw MSI data is obtained. The raw MSI spectrum data are preprocessed by peak alignment, peak extraction, normalization, and baseline correction. At the 10% quantile of the ion intensity, they are baseline-corrected using a spline approximation baseline to obtain the MSI ion image X of the tissue section to be analyzed. h×v×I ={(X i ) h×v , i=1,2,…,I}, where h and v are the number of horizontal and vertical pixels respectively, I is the number of ions, X i is the i-th ion image.

[0035] S2: Take an adjacent slice of the tissue to be analyzed and wash it. Stain the cell nucleus and cytoplasm of the slice with hematoxylin and eosin respectively. Then dehydrate, make it transparent, dry it and seal it. Finally, examine it under a microscope to obtain a high-resolution H&E staining microscopic image Y. H×V×3 , where H and V are the number of pixels in the horizontal and vertical directions respectively.

[0036] S3: Using the Uniform Manifold Approximation and Projection (UMAP) method, the acquired MSI ion image X h×v×I The spectrum data is compressed and reduced to obtain the MSI dimension reduction data E h×v×3 .

[0037] S4: Building a super-resolution model based on deep neural networks Designing an appropriate loss function Use an appropriate optimizer.

[0038] S5: Dimensionality reduction data E using MSI h×v×3 and high-resolution H&E staining microscopic images Y H×V×3 As training samples, that is, super-resolution models The input of the super-resolution model Perform self-supervised learning and optimize model parameters until the loss function Converge to the optimal value and obtain the super-resolution image S of the MSI dimensionality reduction data H×V×3 .

[0039] S6: Super-resolution model Migrate to Ionic Image X i (i=1, 2, ..., I) super-resolution model In the example, keeping the parameters of the registration module unchanged, the ion image Xi and high-resolution H&E staining images Y H×V×3 Training the model Other parameters are used to obtain super-resolution ion images (R i ) H×V , realize the super-resolution reconstruction of MSI R={R i , i=1,2,…,I}.

[0040] A specific embodiment is given below, and the specific steps are:

[0041] 1. Sample preparation, MSI data acquisition and preprocessing

[0042] (1) Tissue samples were cryosectioned (12 μm thick), thawed, mounted on ITO-coated microscope slides, and dried in a vacuum desiccator (~2 h).

[0043] (2) The dried slices were coated with a matrix and MALDI-TOF MSI analysis was performed on a Bruker Autoflex Speed ​​TOF mass spectrometer in positive linear mode using FlexControl 3.3 software. Each point was acquired 100 times at a repetition rate of 1 kHz, with a resolution of 100 μm and a mass-to-charge ratio range of m / z 4500 to m / z 14400.

[0044] (3) Peak alignment: Align the ion peaks of each pixel in the original MSI data 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.

[0045] (4) Spectral peak extraction: Spectral peaks that meet the following two conditions are retained 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, the spectral peak is at a local maximum within a 10 ppm window.

[0046] (5) Normalize the spectral data at each pixel.

[0047] (6) At the 10% quantile of ion intensity, they are baseline corrected using a spline approximation baseline to obtain the MSI ion image data X h×v×I .

[0048] 2. H&E staining microscopic image data acquisition

[0049] (1) Clean the adjacent sections of the MSI analysis sections.

[0050] (2) The cell nucleus and cytoplasm were stained with hematoxylin (H) and eosin (E), and the tissue sections after H&E staining were dehydrated, transparent, air-dried and sealed. The H&E stained sections were examined under a microscope to complete the data collection. H×V×3 The images were acquired with a Mirax scanner at a resolution of 5 μm.

[0051] 3. Use Uniform Manifold Approximation and Projection (UMAP) to calculate the MSI ion image data X obtained in step 1. h×v×I Perform dimensionality reduction to obtain MSI dimensionality reduction data E h×v×3 .

[0052] 4. Build an end-to-end super-resolution reconstruction model based on deep neural networks Designing an appropriate loss function Select an appropriate optimizer and initialize the parameters:

[0053] (1) Constructing a registration network based on STN to eliminate the Y H×V×3 With MSI dimensionality reduction data E h×v×3 Position offset between.

[0054] The registration network consists of three parts: local network, grid generator, and sampler. The local network extracts the H&E stained microscopic image Y H×V×3 The grid generator then uses θ to transform the input H&E stained microscopic image Y H×V×3 The points in the image are sampled and transformed, and the interpolation method is used to generate the registered H&E staining image. Among them, the local network consists of two convolutional neural network (CNN) modules, one flatten layer and one fully connected module. Each CNN module consists of a 2D CNN layer with a convolution kernel of 3×3, a pooling layer with a pooling kernel of 2×2, and a ReLU activation layer; the flatten layer is used for the transition between CNN and the fully connected layer; the fully connected module consists of a fully connected layer with 32 nodes, a ReLU activation layer and a fully connected layer with 6 nodes.

[0055] (2) Construct a mapping network based on the registered H&E stained images Mapping to generate MSI prediction graph E F H×V×3 The conversion network consists of four CNN modules. The first three CNN modules consist of a 2D CNN layer with a 3×3 convolution kernel, a batch normalization (BN) layer, and a ReLU activation layer. The last CNN module consists of a 2D CNN layer with a 3×3 convolution kernel and a BN layer.

[0056] (3) Construct a fusion network to use H&E stained images and MSI dimensionality reduction data E h×v×3 Based on the spatial features of H×V×3 The fusion network consists of an encoder and a decoder. The encoder consists of four downsampling CNN modules, each of which consists of a 2D CNN layer with a 3×3 convolution kernel, a 2×2 downsampling layer, a batch normalization layer, and a ReLU activation layer. The decoder consists of four upsampling CNN modules. The first three CNN modules consist of a 2D CNN layer with a 3×3 convolution kernel, a 2×2 upsampling layer, a batch normalization layer, and a ReLU activation layer. The last CNN module consists of a 2D CNN layer with a 3×3 convolution kernel, a 2×2 upsampling layer, a batch normalization layer, and a ReLU activation layer. The input of each CNN module consists of the output of the previous CNN module and the output of the CNN module corresponding to the resolution in the encoder.

[0057] (4) Construct a filtering network to eliminate H&E staining images The redundant features in the MSI are used to generate the super-resolution image S of the MSI dimensionality reduction data. H×V×3 The filtering network consists of a 4-layer CNN module. The first three CNN modules consist of a 2D CNN layer with a 3×3 convolution kernel and a ReLU activation layer. The last CNN module consists of a 2D CNN layer with a 3×3 convolution kernel and a BN layer.

[0058] (5) Select the adaptive moment estimation (Adam) optimizer with parameters β1 and β2 set to 0.5 and 0.99, and the learning rate set to 10 -4 , the network parameters are initialized to normal distribution N(0,0.02).

[0059] (6) Loss function of super-resolution reconstruction model The design is as follows:

[0060]

[0061] Among them, ω1, ω2, ω3ω4, ω5 are regularization parameters, and the five Loss functions are:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] Among them, ||·||1 is the L1 norm, d(·) is the downsampling operator with Lanczos kernel, ||·||2 is the L2 norm, and k is the window size. Loss function Used to reduce registration error; So that the fused image Z H×V×3 H&E stained images The characteristic distribution of Used to reduce the error between adjacent pixels; So that the fused image Z H×V×3 With the dimension-reduced data E h×v×3 The spatial structure of Used to filter out H&E stained images Unique features in .

[0068] (7) Super-resolution model Training, reduce the MSI dimension data E h×v×3 and high-resolution H&E staining microscopic images Y H×V×3 As input to the model. First, only the loss function is applied The network was trained by iterating 1,000 times under the rule of applying all loss functions; secondly, the network was trained by iterating 25,000 times under the rule of applying all loss functions; finally, the super-resolution image S of the MSI dimensionality reduction data was obtained. H×V×3 , with a 16-fold increase in spatial resolution. The models were trained in the PyTorch environment using an NVIDIA GTX2080Ti GPU.

[0069] (8) Model Migrate to Ionic Image X i Super-resolution model and keep the parameters of the registration module unchanged; use the ion image X i and high-resolution H&E images Y H×V×3 Training the model Other parameters are used to obtain super-resolution ion images (R i ) H×V , realize the super-resolution reconstruction of MSI R={R i , i=1,2,…,I}.

[0070] Figure 1 The system framework diagram of the present invention is given. H×V×3 H&E stained microscopic images. is the H&E stained image after registration, E F H×V×3 H&E stained image based on registration The MSI prediction map generated by mapping, Eh×v×3 For MSI dimensionality reduction data, E u H×V×3 For E h×v×3 The upsampling result, Z H×V×3 For E u H×V×3 and The fusion result, S H×V×3 It's E h×v×3 The super-resolution reconstruction results, (X i ) h×v is the i-th ion image, For X i The upsampling result, (R i ) H×V For X i R is the super-resolution reconstruction result of the MSI ion image X.

[0071] According to an embodiment of the present invention, super-resolution reconstruction is performed on MSI data and H&E-stained microscopic images of a mouse brain. The generated images have the advantage of high resolution while having the macromolecular information of the MSI data. Figure 2 To reduce the dimension of MSI data E h×v×3 The super-resolution reconstruction image after implementing the present invention and the comparison with the Raf Van de Plas method show that compared with the Raf Van de Plas method, the present invention can accurately correct the H&E image through the registration network. At the same time, the present invention realizes nonlinearity through the neural network, and the super-resolution reconstruction result is consistent with the dimensionality reduction data E h×v×3 The spatial structure of the image tends to be consistent, and the unique features in the H&E staining image are filtered out. Therefore, the present invention can achieve better super-resolution reconstruction performance. Figure 3 For the ion image (X i ) h×v The super-resolution reconstruction images after implementing the present invention show that the present invention is not only applicable to three-dimensional MSI images, but can also perform super-resolution reconstruction of MSI ion images through transfer learning.

[0072] In summary, the present invention develops a super-resolution reconstruction method for mass spectrometry imaging driven by H&E staining microscopic images, uses deep learning to establish an end-to-end deep neural network, and improves the spatial information of MSI data; uses super-resolution images of MSI dimensionality reduction data to iteratively align H&E images, avoiding the difficulty of accurately correcting H&E images due to factors such as the small amount of information, low signal-to-noise ratio, and low spatial resolution of ion images, which causes errors in MSI super-resolution reconstruction. In addition, the present invention achieves super-resolution reconstruction of single ion images through transfer learning, making the MSI super-resolution reconstruction results more robust and reliable, and replaces H&E staining microscopic images with high-resolution medical images of other modalities, so that mass spectrometry imaging and information of other modalities can be super-resolved and reconstructed, further expanding the application of MSI technology in biomedical research.

Claims

1. A mass spectrometry imaging super-resolution reconstruction method driven by H&E staining microscopic images, characterized in that The steps include: 1) MSI data acquisition and preprocessing: A mass spectrometer is used to acquire mass spectrometry data of the tissue section to be analyzed on a two-dimensional dot matrix. After preprocessing the obtained mass spectrometry data, an MSI ion image X of the tissue section to be analyzed is obtained; 2) H&E staining microscopic image analysis: Take an adjacent section of the tissue section to be analyzed, wash it, stain it with hematoxylin, stain it with eosin, dehydrate it, make it transparent, dry it, mount it, and examine it under a microscope to obtain a high-resolution H&E staining microscopic image Y; 3) MSI data compression: Perform unsupervised compression and dimensionality reduction on the spectrum data of the MSI ion image X to obtain the MSI dimensionality reduction data E of the tissue section to be analyzed; 4) Construction of super-resolution reconstruction model: Constructing an end-to-end super-resolution reconstruction model based on deep neural network Designing the model's loss function And the corresponding model optimizer to initialize the parameters; (1) Constructing a registration network based on STN to eliminate the Y H×V×3 With MSI dimensionality reduction data E h×v×3 Position offset between (2) Construct a mapping network based on the registered H&E stained images Mapping to generate MSI prediction graph E F H×V×3 ; (3) Construct a fusion network to use H&E stained images and MSI dimensionality reduction data E h×v×3 Based on the spatial features of H×V×3 ,The fusion network consists of an encoder and a decoder; (4) Construct a filtering network to eliminate H&E staining images The redundant features in the MSI are used to generate the super-resolution image S of the MSI dimensionality reduction data. H×V×3 ,The filtering network consists of 4 layers of CNN modules; (5) Selecting an adaptive moment estimation optimizer; (6) Loss function of super-resolution reconstruction model The design is as follows: Among them, ω1, ω2, ω3, ω4, ω5 are regularization parameters, Used to reduce registration error; So that the fused image Z H×V×3 H&E stained images The characteristic distribution of Used to reduce the error between adjacent pixels; So that the fused image Z H×V×3 With the dimension-reduced data E h×v×3 The spatial structure of Used to filter out H&E stained images Unique features in 5) Training of super-resolution reconstruction model for dimensionality reduction data: Super-resolution reconstruction model Self-supervised learning is used to obtain the super-resolution reconstructed image S of the dimensionality-reduced data E of the tissue section MSI to be analyzed; 6) Transfer learning of ion image super-resolution model: Transfer super-resolution reconstruction model Migrate to Ionic Image X i Super-resolution model In [15], transfer learning is used to achieve super-resolution reconstruction of MSI ion images.

2. The H&E staining microscopic image driven mass spectrometry imaging super-resolution reconstruction method according to claim 1, characterized in that In step 1), the specific steps of MSI data collection and preprocessing include: ① Cryosection the sample to be analyzed. After thawing, place the tissue sections on cold indium tin oxide (ITO)-coated microscope slides and dry them in a vacuum desiccator. ② The dried tissue slice surface is coated with a matrix, and the mass spectrometer is used to collect the mass spectrometry signal of each pixel point of the tissue slice along the two-dimensional dot matrix direction; ③ Perform peak alignment, peak extraction, normalization and baseline correction on the mass spectrometry signal of each pixel point to obtain the MSI ion image data X={X i ,i=1,2,…,I}.

3. The H&E staining microscopic image driven mass spectrometry imaging super-resolution reconstruction method according to claim 1, characterized in that In step 2), the specific steps of the H&E staining microscopic image analysis include: ① Select an adjacent tissue section to be analyzed for MSI, wash the section, and stain the nucleus and cytoplasm with hematoxylin H and eosin E, respectively; ② Dehydrate, clear, air dry and seal the H&E stained tissue sections; ③ The processed H&E stained sections are examined under a microscope, and high-resolution H&E staining image data Y is collected.

4. The H&E staining microscopic image driven mass spectrometry imaging super-resolution reconstruction method according to claim 1, characterized in that In step 5), the specific steps of training the super-resolution reconstruction model of the dimensionality reduction data are as follows: using the MSI dimensionality reduction data E and the high-resolution H&E stained microscopic image Y as training samples, and letting the super-resolution reconstruction model Perform self-supervised learning and iteratively update model parameters until the loss function After converging to the optimal value, a super-resolution reconstructed image S of the dimension-reduced data E of the tissue slice MSI to be analyzed is obtained.

5. The H&E staining microscopic image driven mass spectrometry imaging super-resolution reconstruction method according to claim 1, characterized in that In step 6), the specific method of transfer learning of the ion image super-resolution model is: Migrate to the i-th ion image X i Super-resolution model In the middle; keep the parameters of the registration module unchanged; use the ion image X i and high-resolution H&E image Y training Other parameters are used to obtain the super-resolution ion image R i , realizing MSI super-resolution reconstruction of tissue sections R={R i ,i=1,2,…,I}.

Images