A virtual staining method for hyperspectral images of unstained sections

Abstract

The present invention relates to a method for virtual staining of hyperspectral images of unstained slices, so as to solve the technical problems of insufficient texture information representation and inaccurate virtual staining in the current virtual staining methods of unstained slices. The method comprises: 1. acquiring and storing multiple sets of hyperspectral raw data; 2. data preprocessing to acquire hyperspectral data; 3. the hyperspectral data of unstained slices constitutes an experimental data set, and k clustering centers are selected according to the experimental data set so that the color similarity between the preliminary pseudo-color image under the k value and the hyperspectral image of the stained slice is the highest; 4. acquiring the optimal preliminary pseudo-color image; 5. acquiring the standard color matching ratio; 6. synthesizing the final pseudo-color image of the unstained slice of the tissue sample to be tested with the single spectrum image of the microstructure to acquire the hyperspectral virtual staining image of the unstained slice of the tissue sample to be tested. The method provided by the present invention is conducive to extracting more accurate hyperspectral data structure texture information.

Description

Technical Field

[0001] The invention relates to a marking method for tissue sections, and in particular to a hyperspectral image virtual staining method for unstained sections. Background Art

[0002] Conventional staining methods include hematoxylin-eosin (HE) staining and immunohistochemistry (IHC) staining, such as Figure 1 The figure shows the schematic diagram of the existing HE staining microstructure. HE staining is the most basic staining method, which mainly uses the different coloring of acid and alkaline dyes in different areas of the cell (nucleus appears purple-blue, cytoplasm appears pink) to reflect the tissue structure. Figure 2 Shown is a schematic diagram of the existing IHC staining microstructure. IHC is often described as a "molecular pathology" method because it reduces reliance on morphological features and instead uses changes in chemical composition (the presence or overexpression of certain antibody-labeled molecules) as the primary means of distinguishing normal tissue from cancerous tissue. IHC has high sensitivity and specificity, but can only produce detection signals for the markers of the antibodies used, and cannot detect unknown cancer markers. In comparison, microscopic images of unstained sections usually do not have rich colors and lack detailed textures. Therefore, in current tissue pathology research, the staining process is still essential, and the microscopic tissue structure needs to be chemically stained to highlight important diagnostic information.

[0003] During the section staining process, due to differences in experimental instruments, differences in different chemical formulations, and the influence of human operation, the same tissue section may produce different staining effects. When performing automated image analysis, changes in the appearance of such sections do pose a great challenge to the establishment of efficient models. In addition, the staining process is usually time-consuming. If the type of cancer is difficult to determine, multiple and multiple types of staining are required, which is extremely unfavorable for rapid judgment of the cancer type and may affect the patient's optimal treatment time. Especially for frozen sections during surgery, surgical time is very precious, and the staining and diagnostic efficiency of frozen sections is even more critical.

[0004] At present, some research results have proved the feasibility of applying multispectral technology to the study of unstained samples. For example, Irshad et al. disclosed the use of multispectral imaging to automatically detect mitosis in breast cancer pathology images in Multi-channels statistical and morphological features based mitosis detection in breast cancer histopathology [J]. Conf Proc IEEE Eng Med Biol Soc, 2013, 2013 (2013): 6091-6094., and proved the correlation between spectral information and HE staining characteristics; Kopriva et al. disclosed the use of multispectral technology to segment unstained cancer microscopic images in Unsupervised decomposition of low-intensity low-dimensional multi-spectral fluorescent images for tumour demarcation [J]. Medical Image Analysis, 2009, 13 (3): 507-518., and converted the multi-channel images into spectral expressions through nonlinear mapping. This nonlinear mapping is also a form of virtual staining. However, multispectral technology has the problem of insufficient description of texture information and inaccurate virtual staining. Summary of the invention

[0005] The present invention aims to solve the technical problems of insufficient texture information representation and inaccurate virtual staining in current virtual staining methods for unstained sections, and proposes a hyperspectral image virtual staining method for unstained sections, which aims to use microscopic hyperspectral technology, adopt a spectral clustering method to spectrally label different tissue structures, and combine the significant morphological information of a single spectral segment to achieve hyperspectral virtual staining of unstained sections, thereby replacing or simulating the staining process without affecting the diagnostic efficiency and accuracy.

[0006] The technical solution of the present invention is:

[0007] A method for virtual staining of a hyperspectral image of an unstained slice is special in that it comprises the following steps:

[0008] S1. Using microscopic hyperspectral imaging technology to collect multiple sets of hyperspectral raw data of tissue samples, and storing the hyperspectral raw data; each set of hyperspectral raw data includes unstained slice hyperspectral raw data and stained slice hyperspectral raw data of the same tissue sample;

[0009] S2, using a smoothing denoising method to perform image-dimensional data preprocessing on the hyperspectral raw data to obtain hyperspectral data, the hyperspectral data including unstained slice hyperspectral data and stained slice hyperspectral data, and obtaining a stained slice hyperspectral image according to the stained slice hyperspectral data;

[0010] S3, the unstained slice hyperspectral data constitutes the experimental data set, and k cluster centers are selected according to the experimental data set so that the color similarity between the preliminary pseudo-color image under the k value and the stained slice hyperspectral image is the highest;

[0011] S4, according to the k value selected in step S3, the spectral characteristics of each pixel point of the unstained slice hyperspectral data are divided by the spectral clustering method, and each pixel point is marked according to the division result to obtain the optimal preliminary pseudo-color image after the spectral marking of the unstained slice hyperspectral data;

[0012] S5. According to color similarity, the optimal preliminary pseudo-color image is compared with the stained slice hyperspectral image to obtain a standard color ratio, and the color ratio of each type of tissue in the unstained slice is adjusted according to the standard color ratio to obtain a final pseudo-color image;

[0013] S6, collecting the raw hyperspectral data of the unstained slice of the tissue sample to be tested, and preprocessing it in a smoothing and denoising manner to obtain the hyperspectral data of the unstained slice of the tissue sample to be tested;

[0014] According to the k value selected in step S3, the hyperspectral data of the unstained slice of the tissue sample to be tested is processed by the spectral clustering method, and the color ratio of each type of tissue in the unstained slice of the tissue sample to be tested is adjusted according to the standard color ratio in step S5 to obtain a final pseudo-color image of the unstained slice of the tissue sample to be tested;

[0015] The structural texture information of the hyperspectral data of the unstained slice of the tissue sample to be tested is extracted by a twice-stacked principal component transformation method to obtain a single-spectral image of the microstructure. The single-spectral image of the microstructure is synthesized with the final pseudo-color image of the unstained slice of the tissue sample to be tested to obtain a hyperspectral virtual staining image of the unstained slice of the tissue sample to be tested.

[0016] Furthermore, in step S3, k cluster centers are selected so that the color similarity between the preliminary pseudo-color image and the stained slice hyperspectral image under the k value is the highest. The specific process is as follows:

[0017] S3.1. Randomly select a cluster center from the experimental data set as the initial cluster center c 1 , calculate each coordinate sample point x i The shortest distance D(x i ), then calculate x i The probability of being selected as the next cluster center is Pi , the calculation formula is

[0018]

[0019] Among them, i is the i-th sample point, c i The i-th is the cluster center;

[0020] Select the sample point with the highest probability as the next cluster center;

[0021] S3.2, repeat step S3.1 until k cluster centers are selected, where k is an integer and 2≤k≤8, to obtain a preliminary pseudo-color image;

[0022] S3.3, taking different k values, repeating steps S3.1-S3.2 to obtain preliminary pseudo-color images under various k values, comparing the color similarity of each preliminary pseudo-color image with the stained slice hyperspectral image, and selecting the k value corresponding to the preliminary pseudo-color image with the highest similarity.

[0023] Furthermore, in step S1, collecting tissue samples specifically includes adjusting the focusing mechanism of the microscope hyperspectral imaging device so that the tissue sample is in the observation position to collect hyperspectral raw data, and then keeping the magnification and the light source intensity unchanged, moving the slice or replacing other slices to collect hyperspectral raw data.

[0024] Furthermore, in step S1, the unstained slice hyperspectral raw data and the stained slice hyperspectral raw data of the same tissue sample are obtained as follows: two adjacent slices are cut from the same tissue sample, one is stained and sealed according to the normal process as the stained slice, and the data collected is the stained slice hyperspectral raw data; the other is directly sealed as the unstained slice, and the data collected is the unstained slice hyperspectral raw data.

[0025] Further, in step S6, the twice stacked principal component transformation method includes a first transformation and a second transformation;

[0026] The first transformation separates the noise based on the noise covariance matrix of different spectral bands in the unstained slice hyperspectral data, so as to make the transformed noise data have only the minimum variance and no correlation between the spectral bands;

[0027] The second transformation is a standard principal component transformation of the noise data to remove the noise data and obtain a characteristic image of the unstained slice hyperspectral data, which is used to extract the structural texture information of the hyperspectral data in the characteristic image of the unstained slice hyperspectral data.

[0028] Beneficial effects of the present invention:

[0029] 1. The present invention proposes a hyperspectral virtual staining method, which relies on microscopic hyperspectral technology to achieve an effect similar to traditional HE staining, can assist in pathological judgment, and at the same time brings new ideas to the research of label-free pathological diagnosis.

[0030] 2. The present invention selects the best number of cluster centers, uses the best number of cluster centers to process the hyperspectral raw data using the spectral clustering method, obtains the best preliminary pseudo-color image, and improves the accuracy of hyperspectral virtual coloring.

[0031] 3. In the two-layered principal component transformation, the present invention separates the signal from the noise, concentrates the important information in a limited feature set, and aggregates the spectral features into a class feature vector. Some weak signals will be enhanced in the denoising process, which is conducive to extracting more accurate hyperspectral data structure texture information.

[0032] 4. The method provided by the present invention can speed up the processing time of tissue samples, effectively reduce costs, and at the same time prevent the adverse effects of the staining process on the samples, thereby increasing the objectivity of pathological judgment.

[0033] 5. The method provided by the present invention is not only applicable to unstained sections, but also to frozen sections during surgery, which can improve the staining and diagnostic efficiency of frozen sections. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the existing HE-stained microstructure;

[0035] Figure 2 It is a schematic diagram of the existing IHC staining microstructure;

[0036] Figure 3 is a black and white image of a 10× hyperspectral pseudo-color image of an unstained gastric tissue section in an embodiment of the present invention;

[0037] Figure 4 is a black and white image of a 4× hyperspectral pseudo-color image of an unstained gastric tissue section in an embodiment of the present invention;

[0038] Figure 5 is a black and white image of a 20× hyperspectral pseudo-color image of an unstained gastric tissue section in an embodiment of the present invention;

[0039] Figure 6 Schematic diagram of comparison between a black and white image (A) of a hyperspectral pseudo-color image of a stained section of a first gastric cancer sample and a black and white image (B) of a hyperspectral virtual stained image of an unstained section in an embodiment of the present invention;

[0040] Figure 7Schematic diagram of comparison between a black and white image (C) of a hyperspectral pseudo-color image of a stained section of a second gastric cancer sample and a black and white image (D) of a hyperspectral virtual stained image of an unstained section in an embodiment of the present invention;

[0041] Figure 8 Schematic diagram of comparison between a black and white image (A) of a hyperspectral pseudo-color image of a stained section of a third gastric cancer sample and a black and white image (B) of a hyperspectral virtual stained image of an unstained section in an embodiment of the present invention;

[0042] Fig. 9 Schematic diagram of comparison between a black and white image (C) of a hyperspectral pseudo-color image of a stained section of a fourth gastric cancer sample and a black and white image (D) of a hyperspectral virtual stained image of an unstained section in an embodiment of the present invention;

[0043] Fig.10 Schematic diagram of the comparison of the final black-and-white image of the pseudo-color image (A), the microscopic tissue single-spectrum image (B) and the black-and-white image of the hyperspectral virtual staining image (C) of the first sample of greater omental metastatic adenocarcinoma in an embodiment of the present invention;

[0044] Fig.11 Schematic diagram of comparison of a black-and-white image (D) of a final pseudo-color image, a single-spectrum image of a microscopic tissue (E), and a black-and-white image (F) of a hyperspectral virtual staining image of a second sample of greater omental metastatic adenocarcinoma in an embodiment of the present invention;

[0045] Fig.12 Schematic diagram of comparison between a black and white image (B) of a hyperspectral virtual stained image and a black and white image (A) of a HE stained image in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] This embodiment provides a method for virtual staining of a hyperspectral image of an unstained slice, the method comprising the following steps:

[0047] S1. Multiple groups of hyperspectral raw data of gastric tissue samples are collected by using microscopic hyperspectral imaging technology, and the hyperspectral raw data are stored; each group of hyperspectral raw data includes unstained slice hyperspectral raw data and stained slice hyperspectral raw data of the same gastric tissue sample.

[0048] Specifically, the gastric tissue sample is collected by adjusting the focusing mechanism of the microscope hyperspectral imaging device to place the gastric tissue sample in the observation position for hyperspectral raw data collection, and then the magnification and light source intensity are kept unchanged, the slice is moved or other slices are replaced, and the hyperspectral raw data is collected; the collected data is saved as a 596×716×256 data cube; the unstained slice hyperspectral raw data and the stained slice hyperspectral raw data of the same gastric tissue sample are obtained as follows: two adjacent slices are cut from the same gastric tissue sample, one is stained and sealed according to the normal process as the stained slice, and the collected data is the stained slice hyperspectral raw data; the other is directly sealed as the unstained slice, and the collected data is the unstained slice hyperspectral raw data.

[0049] The test samples of this example include intraoperative frozen sections of gastric cancer tissue (from PLA 301 Hospital) and unstained sections of gastric cancer tissue (from Guangzhou Panyu District First Hospital), collected as shown in Figures 3 to 5 , 4×, 10× and 20× hyperspectral pseudo-color images of unstained gastric tissue sections. It can be clearly seen from the figures that compared with stained sections, unstained sections lack rich color information and structural texture information. For unstained sections, it is impossible to obtain sufficient morphological characteristics of cells and tissues, so it is difficult to make a direct diagnosis.

[0050] S2. Preprocess the image dimension of the hyperspectral raw data using a smoothing denoising method to obtain hyperspectral data. The hyperspectral data includes unstained slice hyperspectral data and stained slice hyperspectral data. A stained slice hyperspectral image is obtained based on the stained slice hyperspectral data.

[0051] S3, the unstained slice hyperspectral data constitutes the experimental data set, and k cluster centers are selected so that the color similarity between the preliminary pseudo-color image and the stained slice hyperspectral image under the k value is the highest; specifically,

[0052] S3.1. Randomly select a cluster center from the experimental data set as the initial cluster center c 1 ; Calculate each coordinate sample point x in the experimental data set i The shortest distance D(x i ), then calculate x i The probability of being selected as the next cluster center is P i , the calculation formula is

[0053]

[0054] Among them, i is the i-th sample point, c i is the i-th cluster center;

[0055] Select the sample point with the highest probability as the next cluster center.

[0056] S3.2. Repeat step S3.1 until k clustering centers are selected, where k is an integer and 2 ≤ k ≤ 8, to obtain a preliminary pseudo-color image.

[0057] S3.3. Set different values of k, repeat steps S3.1 - S3.2, obtain the preliminary pseudo-color images under each value of k, compare the color similarity between each preliminary pseudo-color image and the stained section hyperspectral image, and select the value of k corresponding to the highest similarity.

[0058] The spectral clustering adopted measures the similarity between samples based on the Euclidean distance and clusters each sample according to the similarity. During this process, the number of clustering centers has a great influence on the overall result. Once the selection is not good, it may be impossible to obtain an effective similarity result. Therefore, the clustering center k is set to 2 - 8 in turn, and the best k value is determined by comparing the experimental effects. The idea of selecting the clustering center is that if n initial clustering centers have been selected (0 < n < k), then when selecting the (n + 1)-th clustering center, the points farther away from the current n clustering centers have a higher probability of being selected as the (n + 1)-th clustering center. In this embodiment, after comparing the color similarity, comparing each preliminary pseudo-color image with the stained section hyperspectral image, setting the value of k to 6 is more in line with the actual situation.

[0059] S4. According to the value of k selected in step S3, use spectral clustering to divide the spectral features of each pixel point of the unstained section hyperspectral data, mark each pixel point according to the division result, and obtain the optimal preliminary pseudo-color image after spectral marking of the unstained section hyperspectral data.

[0060] S5. According to the color similarity, compare the optimal preliminary pseudo-color image with the stained section hyperspectral image, obtain the standard color ratio, and adjust the color ratio of various tissues in the section according to the standard color ratio to obtain the final pseudo-color image.

[0061] This embodiment shows the processing results of poorly differentiated gastric cancer tissues. See Figure 6 , A mainly shows the characteristics of the stained section cells in the first group of gastric cancer samples, and B uses the spectral clustering method to mark the unstained sections in the first group of gastric cancer samples to obtain a hyperspectral virtual staining image; see Figure 7 , C mainly shows the characteristics of the stained section cell regions in the second group of gastric cancer samples, and D uses the spectral clustering method to mark the unstained sections in the second group of gastric cancer samples to obtain a hyperspectral virtual staining image, clearly showing the regional characteristics of the fibrous tissue in the gastric cancer samples; see Figure 8, A mainly shows the characteristics of the stained slice cells in the third group of gastric cancer samples, and B uses the spectral clustering method to mark the unstained slices in the third group of gastric cancer samples to obtain a hyperspectral virtual staining image; see Fig. 9 , C mainly shows the regional characteristics of cells in the stained sections of the fourth group of gastric cancer samples. In D, the spectral clustering method is used to mark the unstained sections of the fourth group of gastric cancer samples to obtain a hyperspectral virtual staining image, which clearly shows the regional characteristics of the fibrous tissue in the gastric cancer samples.

[0062] S6, collecting the raw hyperspectral data of the unstained slice of the tissue sample to be tested, and obtaining the hyperspectral data of the unstained slice of the tissue sample to be tested after preprocessing by smoothing and denoising; processing the hyperspectral data of the unstained slice of the tissue sample to be tested by spectral clustering method according to the k value selected in step S3, and performing color matching on the unstained slice of the tissue sample to be tested according to the standard color matching in step S5 to obtain the final pseudo-color image of the unstained slice of the tissue sample to be tested; extracting the structural texture information of the hyperspectral data of the unstained slice of the tissue sample to be tested by using the principal component transformation method of two layers, and obtaining the single spectral image of the microstructure; synthesizing the single spectral image of the microstructure with the final pseudo-color image of the unstained slice of the tissue sample to be tested, and obtaining the hyperspectral virtual staining image of the unstained slice of the tissue sample to be tested.

[0063] It is understandable that the final pseudo-color image reflects the image information of the unstained slice, but when the data is preprocessed in the image dimension, the local structural texture information is lost. In order to make up for the lack of detailed texture information, it is necessary to use a twice-stacked principal component transformation method to add structural texture information to the final pseudo-color image.

[0064] See also Fig.10 , A is the processing result of spectral clustering, which simulates the staining information for the unstained slice to obtain the final pseudo-color image; B is the single-spectrum image of the microscopic tissue extracted after two stacked principal component transformations, which contains relatively rich tissue texture information; C is the hyperspectral virtual staining image formed by fusing the final pseudo-color image with the single-spectrum image of the microscopic tissue. Because of the addition of detailed information, the cell structure is clearer, and it can be seen that the characteristics of the red blood cells are more obvious.

[0065] See also Fig.11 , D is the processing result of spectral clustering, which simulates the staining information for the unstained slice to obtain the final pseudo-color image; E is the single-spectrum image of the microstructure extracted after two layers of principal component transformation, which contains relatively rich tissue texture information; F is the hyperspectral virtual staining image formed by fusing the final pseudo-color image with the single-spectrum image of the microstructure, which makes the texture characteristics of the fiber tissue in F more prominent, effectively improving the effect of virtual staining.

[0066] Specifically, the twice-stacked principal component transformation method includes a first transformation and a second transformation; the first transformation separates the noise based on the noise covariance matrix of different spectral bands in the unstained slice hyperspectral data, so as to make the transformed noise data have only the minimum variance and no correlation between the spectral bands; the second transformation is a standard principal component transformation of the noise data to remove the noise data and obtain the characteristic image of the unstained slice hyperspectral data, which is used to extract the structural texture information of the hyperspectral data in the characteristic image of the unstained slice hyperspectral data.

[0067] After the two-layered principal component transformation, different spectral segments in the feature domain will have different physical meanings. For the unstained slice image, in this embodiment, the transformed spectral segment 1 often represents the brightness background of the entire spectral segment, which is brighter than other spectral segments in the image; the spectral information is mainly concentrated in the 2nd to 6th spectral segments, the image is relatively clear, and noise gradually appears; usually after the 10th spectral segment, random noise will appear, covering the spectral information and spatial information of the target object. In general, in the two-layered principal component transformation, through the separation of signal and noise, important information is concentrated in a limited feature set, and the spectral features converge to the class feature vector. Some weak signals will be enhanced in the denoising process. This method is used to extract the single spectral segment image of the microscopic tissue, and synthesize it with the final pseudo-color image of the unstained slice of the tissue sample to be tested to achieve overall hyperspectral virtual staining.

[0068] To further illustrate the problem, the hyperspectral virtual dyeing image of this embodiment is compared with the image dyed by HE method, see Fig.12 It can be seen that the combination of the final pseudo-color image shown in B and the single-spectrum image of the microstructure can show finer fiber structure details, thereby effectively improving the resolution of the image.

Claims

1. A method for virtual staining of hyperspectral images of unstained sections, characterized in that: The following steps are involved: S1. Using microscopic hyperspectral imaging technology to collect multiple sets of hyperspectral raw data of tissue samples, and storing the hyperspectral raw data; each set of hyperspectral raw data includes unstained slice hyperspectral raw data and stained slice hyperspectral raw data of the same tissue sample; S2, using a smoothing denoising method to perform image-dimensional data preprocessing on the hyperspectral raw data to obtain hyperspectral data, the hyperspectral data including unstained slice hyperspectral data and stained slice hyperspectral data, and obtaining a stained slice hyperspectral image according to the stained slice hyperspectral data; S3, the unstained slice hyperspectral data constitutes the experimental data set, and k cluster centers are selected according to the experimental data set so that the color similarity between the preliminary pseudo-color image under the k value and the stained slice hyperspectral image is the highest; S4, according to the k value selected in step S3, the spectral characteristics of each pixel point of the unstained slice hyperspectral data are divided by the spectral clustering method, and each pixel point is marked according to the division result to obtain the optimal preliminary pseudo-color image after the spectral marking of the unstained slice hyperspectral data; S5. According to color similarity, the optimal preliminary pseudo-color image is compared with the stained slice hyperspectral image to obtain a standard color ratio, and the color ratio of each type of tissue in the unstained slice is adjusted according to the standard color ratio to obtain a final pseudo-color image; S6, collecting the raw hyperspectral data of the unstained slice of the tissue sample to be tested, and preprocessing it in a smoothing and denoising manner to obtain the hyperspectral data of the unstained slice of the tissue sample to be tested; According to the k value selected in step S3, the hyperspectral data of the unstained slice of the tissue sample to be tested is processed by the spectral clustering method, and the color ratio of each type of tissue in the unstained slice of the tissue sample to be tested is adjusted according to the standard color ratio in step S5 to obtain a final pseudo-color image of the unstained slice of the tissue sample to be tested; The structural texture information of the hyperspectral data of the unstained slice of the tissue sample to be tested is extracted by a twice-stacked principal component transformation method to obtain a single-spectral image of the microstructure. The single-spectral image of the microstructure is synthesized with the final pseudo-color image of the unstained slice of the tissue sample to be tested to obtain a hyperspectral virtual staining image of the unstained slice of the tissue sample to be tested.

2. The method for virtual staining of hyperspectral images of unstained sections according to claim 1, characterized in that: In step S3, k cluster centers are selected so that the color similarity between the preliminary pseudo-color image and the stained slice hyperspectral image under the k value is the highest. The specific process is as follows: S3.

1. Randomly select a cluster center from the experimental data set as the initial cluster center c1; calculate each coordinate sample point x i The shortest distance D(x i ), then calculate x i The probability of being selected as the next cluster center is P i , the calculation formula is Among them, i is the i-th sample point, c i is the i-th cluster center; Select the sample point with the highest probability as the next cluster center; S3.2, repeat step S3.1 until k cluster centers are selected, where k is an integer and 2≤k≤8, to obtain a preliminary pseudo-color image; S3.3, taking different k values, repeating steps S3.1-S3.2 to obtain preliminary pseudo-color images under various k values, comparing the color similarity of each preliminary pseudo-color image with the stained slice hyperspectral image, and selecting the k value corresponding to the preliminary pseudo-color image with the highest similarity.

3. The method for virtual staining of hyperspectral images of unstained sections according to claim 2, characterized in that: In step S1, collecting tissue samples specifically involves adjusting the focusing mechanism of the microscope hyperspectral imaging device so that the tissue sample is in the observation position to collect hyperspectral raw data, and then keeping the magnification and light source intensity unchanged, moving the slice or replacing other slices to collect hyperspectral raw data.

4. The method for virtual staining of hyperspectral images of unstained sections according to claim 3, characterized in that: In step S1, the method for obtaining the hyperspectral raw data of the unstained section and the hyperspectral raw data of the stained section of the same tissue sample is as follows: two adjacent sections are cut from the same tissue sample, one is stained and sealed according to the normal process as the stained section, and the data collected is the hyperspectral raw data of the stained section; the other is directly sealed as the unstained section, and the data collected is the hyperspectral raw data of the unstained section.

5. The method for virtual staining of hyperspectral images of unstained sections according to any one of claims 1 to 4, characterized in that: In step S6, the twice stacked principal component transformation method includes a first transformation and a second transformation; The first transformation separates the noise based on the noise covariance matrix of different spectral bands in the unstained slice hyperspectral data, so as to make the transformed noise data have only the minimum variance and no correlation between the spectral bands; The second transformation is a standard principal component transformation of the noise data to remove the noise data and obtain a characteristic image of the unstained slice hyperspectral data, which is used to extract the structural texture information of the hyperspectral data in the characteristic image of the unstained slice hyperspectral data.

Images