Tumor classification method and system based on tumor hyperspectral image

Abstract

The invention relates to the technical field of hyperspectral image processing, and particularly discloses a tumor classification method and system based on a tumor hyperspectral image, which can provide richer and more meticulous tumor tissue feature information including spectral features, morphological features, metabolic features and the like through a multiband reflection spectral image acquired by hyperspectral imaging equipment. The comprehensive information acquisition is helpful for doctors and researchers to comprehensively understand the characteristics of tumor tissues, and an important basis is provided for accurate diagnosis and personalized treatment. The invasiveness is one of the important indexes of the tumor malignancy degree, and knowing the invasiveness characteristics of the tumor tissue is beneficial to evaluating the growth and diffusion capabilities of the tumor tissue, so that important reference is provided for formulating a treatment scheme and predicting the prognosis of a patient. Meanwhile, the dynamic influence of spatial distribution on the tumor tissue metastasis capacity is quantified, and the metastasis potential of the tumor tissue can be evaluated more accurately. Doctors can be helped to formulate a treatment scheme better, the transfer process is intervened in time, and the treatment effect and the survival rate are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image processing, and relates to a tumor classification method and system based on tumor hyperspectral images. Background Art

[0002] The tumor classification method based on hyperspectral images is of irreplaceable importance and necessity in clinical diagnosis and treatment. First of all, hyperspectral imaging technology can capture spectral information of hundreds of continuous bands in the visible light to near-infrared range, and can reveal the chemical composition, metabolic state and microenvironment characteristics of tumor tissues at the molecular level, which cannot be achieved by traditional medical images. For example, the spectral reflectance difference between tumor tissues and normal tissues in specific bands can reach more than 30%, and this difference provides a quantitative basis for accurately distinguishing tumor subtypes and judging malignancy.

[0003] However, the current tumor classification technology based on hyperspectral images still has significant defects and drawbacks, seriously restricting its clinical transformation and application. The redundancy of high-dimensional spectral data and noise interference lead to low feature extraction efficiency. Hyperspectral images usually contain hundreds of bands, but only a small number of bands (such as 5-10) are related to tumor-specific metabolism. A large number of irrelevant bands not only increase the computational complexity, but also introduce interferences such as uneven illumination and equipment noise, resulting in a decline in the robustness of the classification model. For example, in areas with high tissue water content (such as the edema zone), the spectral attenuation of the 940nm band is easily affected by scattering noise, causing the model to misidentify the inflammatory reaction area as the tumor infiltration area, with a high error rate. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the present invention provides a tumor classification method and system based on tumor hyperspectral images to solve the above technical problems.

[0005] In order to achieve the above and other purposes, the technical solutions adopted by the present invention are as follows: On the one hand, the present invention provides a tumor classification method based on tumor hyperspectral images, and the method includes the following steps: Step S1: Obtain the multi-band reflectance spectral image of the tumor tissue sample through a hyperspectral imaging device, and perform noise elimination and normalization processing on it; Step S2: Quantify the invasive characteristics of the tumor tissue by spatial domain morphological analysis and spectral domain metabolic feature extraction; Step S3: Based on the spectral scattering attenuation rate of the tumor edge region in a fixed band, quantify the dynamic influence of the spatial distribution on the tumor tissue metastasis ability to obtain the metastasis ability characteristics of the tumor tissue; Step S4: Integrate the invasive and metastasis ability characteristics through a dynamic weight allocation network, and output the comprehensive malignancy grading of the tumor tissue.

[0006] Step S1 includes: Step S11: Use a hyperspectral imaging device to scan the tumor tissue at a preset spatial resolution to obtain a multi - band reflectance spectral image of the tumor tissue sample; Step S12: Calculate the spectral gradient value of adjacent bands for each pixel point in the spectral image. The specific calculation method is: subtract the reflectance of the previous band from the reflectance of the next band for each pixel point, and then divide by the band interval width to obtain the spectral gradient value of adjacent bands for each pixel point; set a gradient mutation threshold. If the spectral gradient values of three consecutive adjacent bands of the same pixel point in the spectral dimension all exceed the gradient mutation threshold, then determine the above - mentioned three adjacent bands as abnormal bands; Step S13: Perform the detection in Step S12 for all pixel points of the multi - band reflectance spectral image of the tumor tissue sample, and count the frequency of the appearance of abnormal bands. If a certain abnormal band is marked as abnormal for more than a set ratio of pixel points in the spectral image, then add the entire abnormal band to the mask blacklist; for the remaining band image, use a 7×7 pixel search window, calculate the similarity weight of the pixel point gray values within the window, and perform weighted averaging on each pixel point based on the weight matrix to eliminate speckle noise; Step S14: Collect reference data based on a standard whiteboard under the same imaging conditions, calculate the device gain coefficient matrix, and perform reflectance normalization on the hyperspectral image; subsequently, select the spectral curve of the central pixel point in the tumor region as a reference benchmark, calculate the cosine similarity of the spectra of other pixel points. If the similarity is lower than the experimentally verified error tolerance threshold, then use bicubic interpolation to regenerate the spectral data of the distorted region.

[0007] Step S14 includes the following sub - steps: Step S141: Under the same environmental conditions, use the same hyperspectral device to photograph the standard whiteboard to obtain its hyperspectral image data cube; Step S142: For each band, calculate the average reflectance of all pixel points in the standard whiteboard image, divide the calibrated reflectance value of the standard whiteboard by the average reflectance of the corresponding band to obtain the device gain coefficient matrix; Step S143: Multiply the reflectance value of each pixel point in the multi - band reflectance spectral image corresponding to the tumor tissue sample by the coefficient at the corresponding position in the gain coefficient matrix, and output the normalized reflectance data. The range of the normalized reflectance data is from 0 to 1, where 1 represents complete reflection; Step S144: Select a uniform area of 10×10 pixels in the core region of the multi-band reflectance spectral image corresponding to the tumor tissue sample, calculate its average spectrum as the reference benchmark. For each pixel, multiply its spectral curve with the reference spectrum point by point according to the wavelength band, sum the results, and then divide by the square root of the sum of the squares of their respective spectral values to obtain the similarity score. If the similarity score is lower than the experimentally verified error tolerance threshold, it is determined that the spectrum of this pixel is distorted. Step S145: With the distorted pixel as the center, take the spectral data with qualified similarity within the surrounding 5×5 pixel area, calculate the interpolation through a cubic polynomial weight function, and generate a new spectral curve.

[0008] Step S2 includes: Step S21: Use the Canny edge detection algorithm to process the grayscale information in the spatial domain of the hyperspectral image. Set the standard deviation of Gaussian filtering σ = 1.5 to smooth the noise, and extract the initial edge contour based on the high-low threshold ratio of 1:3. Perform a morphological closing operation on the initial edge to fill the gaps and generate a continuous tumor-normal tissue boundary mask. Subsequently, calculate the boundary fractal dimension based on the box-counting method: divide the mask image into grids of different sizes, count the number of grids covering the boundary, and fit the slope in the double-logarithmic coordinate system as the fractal dimension value. Further extract the curvature sequence along the boundary, calculate the curvature value of each point through the second derivative method, discretize it into 10 intervals and count the probability distribution, and finally calculate the Shannon entropy as the curvature entropy. Step S22: At the 580 nm wavelength band, use the Otsu adaptive threshold segmentation method to extract the blood vessel area from the spatial domain image, and calculate the proportion of the blood vessel area as the blood vessel density index. Synchronously detect the edema area at the 940 nm wavelength band, generate a binary mask by setting a threshold with a reflectance lower than 30% of the average value of the whole image, and perform morphological erosion with a 5×5 circular structuring element to eliminate the interference of scattered point noise. Thus, calculate the metabolic activity score, and the calculation formula is to use the reflectance at the 940 nm wavelength band as the numerator of the fraction and the reflectance at the lipid wavelength band in the hyperspectral image as the denominator of the fraction, and thus calculate the metabolic activity score. Step S23: Quantitatively analyze the degree of collagen degradation in the tumor area through Raman spectroscopy technology. The specific implementation process is as follows: First, detect the characteristic peak intensity of the tumor tissue at the 1550 cm⁻¹ wavelength band in the Raman spectrum, use the numerical integration method to calculate the spectral intensity integral value of this wavelength band, and compare it with the corresponding wavelength band intensity integral reference value of the normal tissue adjacent to the cancer of the same patient. Calculate the collagen intensity ratio between the tumor area and the normal tissue, and subtract this ratio from 1 to obtain the degradation index DI.

[0009] The metastatic ability characteristics of the tumor tissue are specifically the attenuation coefficients of each pixel at each wavelength band. The specific calculation logic is as follows: For each pixel point (x, y), taking the reflection intensity I(λ) of the standard whiteboard in the wavelength band λ as a reference, comparing it with the actually measured reflection intensity I(x, y, λ) of the tumor region, and combining with the optical path length d, the attenuation coefficient of each pixel point in each wavelength band is obtained through the formula u(x, y, λ) = -(1 / d)·ln(I(x, y, λ) / I(λ)).

[0010] Step S4 includes: Step S41: Perform Z-score normalization on the invasive features to eliminate the dimensional differences between different features. At the same time, perform Min-Max normalization on the metastasis ability features and map them to the interval [0, 1] to adapt to the model input requirements; Step S42: Design a dynamic weight allocation network. The input layer includes static features and dynamic features. Model the temporal attenuation coefficient through a gated recurrent unit to generate dynamic feature weights. At the same time, determine the static feature weights based on ridge regression optimization; Step S43: Input the weighted and fused features into a weighted summation model to obtain the initial malignancy score of the tumor tissue. Compare the initial malignancy score of the tumor tissue with the initial malignancy score intervals corresponding to each preset malignancy level; thereby obtaining the malignancy level of the tumor tissue.

[0011] Step S42 includes: Step S421: Perform Z-score normalization on the invasive features and perform piecewise normalization on the metastasis ability features to eliminate dimensional differences; encode the invasive features as static features and the metastasis ability features as dynamic features; at the same time, encode the two types of feature parameters as static feature vectors and dynamic feature tensors respectively as the network input; Step S422: Use a gated recurrent unit to model the dynamic feature tensor. The network structure includes two hidden layers, the time step = 7, control the temporal dependence relationship through the update gate and the reset gate, and output the hidden state h(t) at each time step; finally, input the hidden state sequence into the fully connected layer to generate dynamic feature weights; Step S423: Based on the static feature vector and the clinical metastasis label, construct a ridge regression model. The objective function is to minimize the sum of the prediction error and the L2 norm of the weight parameters, and solve to obtain the static weight vector.

[0012] A tumor classification system based on tumor hyperspectral images, including: Hyperspectral image processing module: Obtain the multi-band reflection spectral image of the tumor tissue sample through a hyperspectral imaging device, and perform noise elimination and normalization processing on it; Image invasion feature extraction module: Quantify the invasive features of the tumor tissue through spatial domain morphological analysis and spectral domain metabolic feature extraction; Image metastasis feature extraction module: Based on the spectral scattering attenuation rate of the tumor edge region in a fixed band, quantify the dynamic impact of the spatial distribution on the metastasis ability of tumor tissues, and obtain the metastasis ability characteristics of tumor tissues; Tumor grade determination module: Integrate the invasion and metastasis ability characteristics through a dynamic weight allocation network, and output the comprehensive malignancy grading of tumor tissues.

[0013] On the other hand, the present invention provides a tumor classification system based on tumor hyperspectral images, including a hyperspectral image processing module, an image invasion feature extraction module, an image metastasis feature extraction module, and a tumor grade determination module, wherein each module is connected by wired and / or wireless connection methods to achieve data transmission between modules; Hyperspectral image processing module: Obtain the multi-band reflection spectral image of the tumor tissue sample through a hyperspectral imaging device, and perform noise elimination and normalization processing on it; Image invasion feature extraction module: Quantify and obtain the invasion characteristics of tumor tissues through spatial domain morphological analysis and spectral domain metabolic feature extraction; Image metastasis feature extraction module: Based on the spectral scattering attenuation rate of the tumor edge region in a fixed band, quantify the dynamic impact of the spatial distribution on the metastasis ability of tumor tissues, and obtain the metastasis ability characteristics of tumor tissues; Tumor grade determination module: Integrate the invasion and metastasis ability characteristics through a dynamic weight allocation network, and output the comprehensive malignancy grading of tumor tissues.

[0014] As described above, the tumor classification method and system based on tumor hyperspectral images provided by the present invention have at least the following beneficial effects: The tumor classification method and system based on tumor hyperspectral images provided by the present invention can achieve all-round and multi-dimensional information acquisition and analysis of tumor tissues by using a hyperspectral imaging device to obtain the multi-band reflection spectral image of the tumor tissue sample and performing noise elimination and normalization processing on it, combined with spatial domain morphological analysis and spectral domain metabolic feature extraction. The multi-band reflection spectral image obtained through the hyperspectral imaging device can provide richer and more detailed tumor tissue feature information, including spectral features, morphological features, and metabolic features, etc. This comprehensive information acquisition helps doctors and researchers fully understand the characteristics of tumor tissues and provides an important basis for accurate diagnosis and personalized treatment.

[0015] Secondly, through spatial domain morphological analysis and spectral domain metabolic feature extraction, the quantification and analysis of the invasion characteristics of tumor tissues can be realized. Invasiveness is one of the important indicators of tumor malignancy. Understanding the invasion characteristics of tumor tissues helps to evaluate its growth and diffusion ability, and provides an important reference for formulating treatment plans and predicting the prognosis of patients.

[0016] In addition, quantifying the dynamic impact of the spatial distribution on the metastatic ability of tumor tissue based on the spectral scattering attenuation rate in the tumor marginal region can more accurately evaluate the metastatic potential of tumor tissue. Tumor metastasis is one of the important challenges in tumor treatment. By analyzing and evaluating the characteristics of metastatic ability, it can help doctors better formulate treatment plans, intervene in the metastatic process in a timely manner, improve the treatment effect and survival rate. At the same time, it helps to improve the accuracy of tumor diagnosis and the level of personalized treatment, providing better treatment effects and quality of life for patients, and having important technical advantages and application values. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 Schematic diagram of the connection of each step of the method of the present invention.

[0019] Figure 2 Schematic diagram of the connection of each module of the system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The above content is only an example and explanation of the concept of the present invention. Those skilled in the art of this technology can make various modifications or supplements to the described specific embodiments or use similar methods for substitution, as long as they do not deviate from the concept of the invention or exceed the scope defined by this claims, they should belong to the protection scope of the present invention.

[0021] Embodiment 1 Please refer to Figure 1 As shown, a tumor classification method based on tumor hyperspectral images, the method includes the following steps: Step S1: Obtain the multi-band reflection spectral image of the tumor tissue sample through a hyperspectral imaging device, and perform noise elimination and normalization processing on it; Step S2: Quantify the invasive characteristics of the tumor tissue through spatial domain morphological analysis and spectral domain metabolic feature extraction, where the invasive characteristics of the tumor tissue include fractal dimension, curvature entropy, vascular density index, metabolic activity score, and degradation index; It should be added that the fractal dimension quantifies the geometric complexity and irregularity of the tumor-normal tissue boundary; the boundary fractal feature analysis based on the box-counting method. The curvature entropy characterizes the heterogeneity of the boundary curvature distribution and reflects the disorder of the local invasion dynamics at the tumor edge. The discrete degree of the curvature distribution probability is calculated through Shannon entropy. The vascular density index reflects the angiogenesis activity inside the tumor and is positively correlated with the metastatic ability. The metabolic activity score evaluates the glycolytic activity of the tumor and is related to the intensity of the Warburg effect (aerobic glycolysis). It is a normalized score of the reflectance ratio between the 940 nm (water molecule) and 1450 nm (lipid) bands. The degradation index quantifies the degree of damage to collagen in the extracellular matrix (ECM) and reflects the tumor microenvironment remodeling ability.

[0022] Step S3: Based on the spectral scattering attenuation rate of the tumor edge region in a fixed band, quantify the dynamic impact of the spatial distribution on the metastatic ability of the tumor tissue to obtain the metastatic ability characteristics of the tumor tissue; Step S4: Integrate the invasiveness and metastatic ability characteristics through a dynamic weight assignment network and output the comprehensive malignancy grading of the tumor tissue.

[0023] Step S1 includes: Step S11: Use a hyperspectral imaging device to scan the tumor tissue at a preset spatial resolution to obtain a multi-band reflectance spectral image of the tumor tissue sample; Use a hyperspectral imaging device to scan the tumor tissue at a spatial resolution of 10 μm / pixel, with a spectral coverage range of 400 - 2500 nm, including 300 continuous bands; set the scanning parameters: exposure time 50 ms, light source intensity 1000 Lux, objective magnification 20×, to ensure clear imaging of the microstructures on the tumor surface (such as blood vessel branches, cell gaps); collect the data and store it as a three-dimensional data cube H_i(X, Y, λ), where X / Y are the spatial dimension coordinates, λ is the spectral band index, and i is the number index of each pixel point; the three-dimensional data includes, but is not limited to, the wavelength position and its light intensity information corresponding to each pixel point in the spectrometer.

[0024] Step S12: Calculate the spectral gradient value of adjacent bands for each pixel point in the spectral image. The specific calculation method is: subtract the reflectance of the previous band from the reflectance of the next band for each pixel point, and then divide by the band interval width to obtain the spectral gradient value of adjacent bands calculated for each pixel point; set the gradient mutation threshold of ±20%. If the spectral gradient values of three consecutive adjacent bands in the spectral dimension of the same pixel point all exceed the gradient mutation threshold, then the above three adjacent bands are determined to be abnormal bands; The three consecutive adjacent bands in this step refer to three adjacent bands of the same pixel point in the spectral dimension (λ axis), rather than the pixel points in the spatial dimension. The specific logic is: For a single pixel (e.g., the pixel at coordinates x = 100, y = 200), on its spectral reflectance curve, calculate the gradient between adjacent bands in the order of wavelengths (λ1, λ2, …, λ300). If the reflectance of this pixel has a gradient change exceeding ±20% in each of the three consecutive band intervals of λ2 - λ1, λ3 - λ2, and λ4 - λ3, then λ2, λ3, and λ4 are determined to be abnormal bands (possibly caused by device thermal noise or sudden changes in illumination). Perform the above detection on all pixels in the entire image, and count the frequency of occurrence of abnormal bands. If a certain band (e.g., λ50) is marked as abnormal for more than 5% of the pixels in the entire image, then the entire band is added to the mask blacklist.

[0025] Example illustration: Suppose the reflectance gradients of a certain pixel in bands λ100 - λ102 are +25%, -22%, and +28% respectively (all exceeding the ±20% threshold), then: The abnormal bands are λ100, λ101, and λ102; If λ100 of multiple pixels is marked and the frequency > 5%, then the entire λ100 band is finally masked.

[0026] Step S13: Perform the detection in Step S12 on all pixels of the multi - band reflectance spectral image of the tumor tissue sample, and count the frequency of occurrence of abnormal bands. If a certain abnormal band is marked as abnormal for pixels exceeding a set ratio in the spectral image, then the entire abnormal band is added to the mask blacklist; for the remaining band images, use a 7×7 pixel search window to calculate the similarity weight of the pixel gray values within the window, and perform weighted averaging on each pixel based on the weight matrix to eliminate speckle noise. Step S14: Collect reference data by using a standard whiteboard under the same imaging conditions, calculate the device gain coefficient matrix, and perform reflectance normalization on the hyperspectral image; subsequently, select the spectral curve of the central pixel of the tumor region as the reference benchmark, calculate the cosine similarity of the spectra of other pixels. If the similarity is lower than the experimentally verified error tolerance threshold, then use bicubic interpolation to regenerate the spectral data of the distorted region.

[0027] Step S14 includes the following sub - steps: Step S141: Under the same environmental conditions, use the same hyperspectral device to photograph the standard whiteboard to obtain its hyperspectral image data cube; Step S142: For each band, calculate the average reflectance of all pixels in the standard whiteboard image, divide the calibrated reflectance value of the standard whiteboard by the average reflectance of the corresponding band to obtain the device gain coefficient matrix; Step S143: Multiply the reflectance value of each pixel in the multi-band reflectance spectrum image corresponding to the tumor tissue sample by the coefficient at the corresponding position in the gain coefficient matrix, and output the normalized reflectance data. The range of the normalized reflectance data is from 0 to 1, where 1 represents complete reflection. Step S144: Select a 10×10 pixel uniform area in the core region of the multi-band reflectance spectrum image corresponding to the tumor tissue sample, calculate its average spectrum as the reference benchmark. For each pixel, multiply its spectral curve and the reference spectrum point by point by band and then sum them, and then divide by the square root of the sum of the squares of their respective spectral values to obtain the similarity score. If the similarity score is lower than the experimentally verified error tolerance threshold, it is determined that the spectral distortion of this pixel point occurs. Step S145: Take the spectral data with qualified similarity within the surrounding 5×5 pixel area centered on the distorted pixel point, calculate the interpolation through the cubic polynomial weight function, and generate a new spectral curve.

[0028] In the embodiment of the present invention, first, data of a standard whiteboard and a tumor tissue are collected by a hyperspectral imaging device in a standard experimental environment. The standard whiteboard is a polytetrafluoroethylene board with a calibrated reflectance value of 99%, and its surface is polished at the nanometer level to ensure reflection uniformity. The spectral coverage range of the hyperspectral device is 400 - 2500 nm, including 300 continuous bands. The spatial resolution is set to 10 μm / pixel, which can clearly capture the morphological details of tumor microvessels and cell gaps. The ambient light intensity is constantly 1000 Lux, the exposure time is 50 ms, and the objective magnification is 20× to avoid imaging blur caused by light fluctuations or defocusing. After the data collection of the standard whiteboard is completed, the system automatically generates a three-dimensional data cube W(x, y, λ), where the reflectance data of each pixel is stored as a 16-bit floating-point number. Using the full-image statistical analysis method, for each spectral band, calculate the sum of the reflectance values of all pixel points in the standard whiteboard image, and then divide by the total number of pixel points to obtain the average reflectance of this band. Divide the calibrated reflectance value (99%) of the standard whiteboard by the average reflectance of the corresponding band to generate a gain coefficient matrix related to the image spatial position and band. This matrix is used to compensate for the reflectance measurement deviation caused by light source attenuation and sensor sensitivity differences. For example, if the average reflectance of the standard whiteboard at a certain band λ = 580 nm is 0.85, then the gain coefficient of this band is 0.99 / 0.85 ≈ 1.1647, which is used to compensate for the non-linear response of the sensor.

[0029] For each pixel point in the original reflectance data of the tumor tissue, multiply the reflectance value at the corresponding band and position by the coefficient value at the same position in the gain coefficient matrix, and output the normalized reflectance. The range of the normalized reflectance is limited to 0 to 1, where 1 represents complete reflection (consistent with the standard whiteboard), and 0 represents no reflection (such as an absorbent material). This step effectively eliminates the interference of device hardware differences (such as light source attenuation and sensor dark current) on the spectral data.

[0030] Further screen the reference spectra of the tumor core region: In the three-dimensional data cube of the tumor tissue, select the homogeneous region (10×10 pixels) marked by the pathologist. Step 1: Multiply the reference spectrum (the average spectrum of the tumor core region) and the spectral data of the target pixel point by point in the order of bands, and accumulate the product results of all bands; Step 2: Calculate the sum of the squares of the reflectances of the reference spectrum and the target pixel spectrum respectively, and then take the square root of each sum of squares; Divide the accumulated result of Step 1 by the product of the two square root results of Step 2 to obtain the similarity score. If the score is lower than the dynamic error threshold (set to 0.95 through experimental verification), it is determined that the pixel has spectral distortion due to sample tilt or local bleeding. For the distorted pixel point (x, y), intercept a 5×5 pixel neighborhood centered on this point, and screen the pixel points with a similarity score ≥0.95 in this area as valid data samples; Assign weights to each sample point, and the weight value is calculated by a cubic polynomial function. The function form is: the combination of the cube, square, first power, and constant term of the distance from the sample point to the center; Use the least squares method to fit the spectral data of the sample points to solve the optimal coefficient combination of the cubic polynomial function; According to the weight function and the sample spectral data, calculate the reconstructed spectral value of the distorted pixel point. Re-import the reconstructed spectral value of the distorted pixel point into the three-dimensional data cube to obtain a new spectral curve.

[0031] Step S2 includes: Step S21: Process the spatial domain gray information of the hyperspectral image using the Canny edge detection algorithm, set the Gaussian filter standard deviation σ = 1.5 to smooth the noise, and extract the initial edge contour based on the high-low threshold ratio of 1:3; Perform a morphological closing operation on the initial edge to fill the gaps and generate a continuous tumor-normal tissue boundary mask; Subsequently, calculate the boundary fractal dimension based on the box counting method: Divide the mask image into grids of different sizes, count the number of grids covering the boundary, and fit the slope in the double logarithmic coordinate as the fractal dimension value; Further extract the curvature sequence along the boundary, calculate the curvature value of each point by the second derivative method, discretize it into 10 intervals and count the probability distribution, and finally calculate the Shannon entropy as the curvature entropy; Step S22: In the 580 nm band, the Otsu adaptive threshold segmentation method is used to extract the vascular region from the spatial domain image, and the proportion of the vascular area is calculated as the vascular density index; simultaneously, in the 940 nm band, the edema region is detected, a binary mask is generated by setting a threshold with a reflectivity lower than 30% of the mean value of the whole image, and morphological erosion is performed using a 5×5 circular structural element to eliminate the interference of scattered point noise; thus, the metabolic activity score is calculated, and the calculation formula is to use the reflectivity in the 940 nm band as the numerator of the fraction and the reflectivity in the lipid band of the hyperspectral image as the denominator of the fraction, and the metabolic activity score is calculated accordingly; Step S23: The quantitative analysis of the collagen degradation degree in the tumor region is carried out by Raman spectroscopy technology, and the specific implementation process is as follows: First, the characteristic peak intensity of the tumor tissue is detected in the 1550 cm⁻¹ band of the Raman spectrum, the spectral intensity integral value of this band is calculated by the numerical integration method, and it is compared with the corresponding band intensity integral reference value of the normal tissue adjacent to the cancer of the same patient. By calculating the collagen intensity ratio between the tumor region and the normal tissue, and subtracting this ratio from 1 to obtain the degradation index DI.

[0032] The calculation logic of the fractal dimension (FD) in the embodiment of the present invention is: The tumor boundary mask image is divided into grids of different sizes (the side length ranges from 2 pixels to 64 pixels), and the number of grids covering the boundary is counted for each grid size. The logarithms of the reciprocal of the grid size and the corresponding number of grids are taken respectively, and the linear relationship between the two is fitted, and the absolute value of the slope of the fitted straight line is used as the fractal dimension (FD). If FD is greater than 1.7, it is determined as an invasive growth pattern.

[0033] The calculation logic of the curvature entropy (CE) is: The local curvature value is calculated point by point along the tumor boundary, the curvature range [-1, 1] is equally divided into 10 intervals, and the occurrence probability of the curvature value in each interval is counted. Based on the probability distribution, the Shannon entropy (that is, the negative value of the sum of the products of the probabilities of all intervals and their logarithms) is calculated as the curvature entropy (CE). If CE is greater than 3.5, it is determined as a highly invasive boundary.

[0034] The calculation logic of the vascular density index (VDI) is: In the 580 nm band image, the vascular region is extracted by the adaptive threshold segmentation method, and the proportion of the number of vascular pixels to the total number of pixels in the tumor region is counted as the vascular density index (VDI). If VDI is greater than 0.15, it is determined as high angiogenesis activity.

[0035] The calculation logic of the metabolic activity score (MAS) is: The reflectivity values in the 940 nm (water molecule absorption peak) and 1450 nm (lipid absorption peak) bands are extracted, and the ratio of the two is normalized. Subtract the reference mean value of the normal tissue from this ratio, and then divide it by its standard deviation to obtain the metabolic activity score (MAS). If MAS is greater than 0.9, it is determined as an abnormal increase in glycolytic activity.

[0036] The calculation logic of the degradation index (DI) is as follows: The integrated intensity values of the tumor region and normal tissue in the Raman spectrum at the 1550 cm⁻¹ band (characteristic peak of collagen) are compared. The integrated value of the tumor region is divided by the integrated value of the normal tissue. After obtaining the ratio, subtract this ratio from 1 to obtain the degradation index (DI). If DI is greater than 0.6, it is determined that the extracellular matrix is severely damaged.

[0037] In the embodiments of the present invention, first, a hyperspectral imaging device is used to collect multi-band images of tumor tissues. The device uses a high-sensitivity CCD sensor (resolution 2048×2048 pixels) and a tunable filter (spectral range 400 - 1700 nm) to ensure that the spectral resolution at key bands such as 580 nm, 940 nm, and 1450 nm reaches 5 nm. During the acquisition process, the exposure time of the device is set to 20 ms, the aperture is f / 2.8, and the gain parameter is dynamically adjusted according to the tissue surface reflectivity to balance the image signal-to-noise ratio and detail retention. The original hyperspectral data is transmitted to the processing terminal through optical fibers and stored in 16-bit RAW format to avoid compression loss; when preprocessing the spatial domain gray information of the hyperspectral image, the Canny edge detection algorithm is used, and the Gaussian filter kernel size is set to 5×5 (standard deviation σ = 1.5) to effectively suppress noise and retain the microstructural features of the tumor-normal tissue boundary. The high and low thresholds are set in a 1:3 ratio (high threshold 45, low threshold 15), and false edges are eliminated through non-maximum suppression to generate an initial contour. To further improve the boundary continuity, morphological closing operations are performed on the initial edge: using a 3×3 rectangular structuring element, first dilating and then eroding to fill gaps and smooth irregular protrusions. The masked image after the closing operation is input into the fractal dimension (FD) calculation module, and the boundary complexity is quantified based on the box-counting method: The mask is divided into grids with side lengths ranging from 2^1 to 2^6 pixels, and the number of grids N(k) covering the boundary is counted. The linear regression curve of log(N(k)) and log(1 / k) is fitted, and the absolute value of its slope is the FD value. Experimental verification shows that when FD > 1.7, the tumor exhibits an invasive growth pattern (such as a crab-claw-shaped edge). Further, a curvature sequence is extracted along the boundary, and the local curvature value is calculated using the second derivative method. Specifically, the curvature is calculated through the coordinate difference of a three-point sliding window (the previous point, the current point, and the next point). After discretizing it into 10 equal-width intervals, the probability distribution is statistically analyzed, and finally the Shannon entropy is calculated. When CE > 3.5, it is determined to be highly invasive.

[0038] In the 580 nm band (absorption peak of oxyhemoglobin), the Otsu adaptive threshold segmentation method is used to extract the vascular region: the segmentation threshold is automatically determined by maximizing the inter-class variance, and after generating a binary mask, the proportion of the vascular area is calculated (VDI = number of vascular pixels / total number of pixels in the tumor region). VDI > 0.15 indicates abnormal angiogenesis activity. Simultaneously, the edema region is detected in the 940 nm band (absorption peak of water molecules): the pixel reflectance is compared with the mean value of the whole image, and the region where the reflectance is lower than 30% of the mean value is determined as the edema area, and morphological erosion is performed using a 5×5 circular structuring element to eliminate isolated noise points. The metabolic activity score is obtained by calculating the reflectance ratio (R940 / R1450) of the 940 nm and 1450 nm (lipid characteristic peak) bands, and is normalized by comparing with the reference value of normal tissue (μ ± 2σ). MAS > 0.9 indicates an abnormally elevated glycolytic activity.

[0039] A Raman spectrometer (laser wavelength 785 nm, spectral resolution 2 cm⁻¹) is used to scan the tumor region, and spectral signals are collected in the 1550 cm⁻¹ band (characteristic vibration peak of collagen). After baseline correction (polynomial fitting) and smoothing processing (Savitzky-Golay filter), numerical integration is performed on the 1550 cm⁻¹ peak (integration range 1520 - 1580 cm⁻¹) to calculate the intensity integral value I_tumor of the tumor region. The same-peak integral value I_normal of the normal tissue adjacent to the cancer is used as a control, and the degradation index (DI) is defined as DI = 1 - I_tumor / I_normal. When DI > 0.6, it indicates severe collagen degradation, which is strongly positively correlated with tumor invasiveness.

[0040] The metastatic ability characteristics of tumor tissue are specifically the attenuation coefficients of each pixel point in each band, and the specific calculation logic is as follows: For each pixel point (x, y), taking the reflection intensity I(λ) of the standard white board in the band λ as a reference, the actually measured reflection intensity I(x, y, λ) of the tumor region is compared with it. Combining the optical path length d, the attenuation coefficient of each pixel point in each band is obtained through the formula u(x, y, λ) = -(1 / d)·ln(I(x, y, λ) / I(λ)).

[0041] In the embodiments of the present invention, a hyperspectral imaging system is used to perform multi-band optical property quantification analysis on tumor tissues. The hyperspectral imaging device used is equipped with a quantum well infrared detector (QWIP), with a spectral coverage range of 400 - 2500 nm, a spectral resolution of 3 nm, and an adjustable spatial resolution (up to 1024×1024 pixels), ensuring accurate capture of the spatial heterogeneity characteristics of the tumor microenvironment in the visible to near-infrared bands. The device is built-in with a standard whiteboard calibration module, whose reflectivity is stably maintained at 99% ± 0.5% in the range of 400 - 2500 nm, for providing the reference reflection intensity I(λ) of each band. The optical path length d is accurately measured by a laser rangefinder (error < 1μm) and dynamically adjusted according to the tissue section thickness (typical value d = 10μm). For each pixel point (x,y) and the target band λ (such as 580 nm, 940 nm, 1450 nm), the system synchronously acquires the reflection intensity I(λ) of the standard whiteboard and the reflection intensity I(x,y,λ) of the tumor region. The data is stored with a 16-bit depth and the detector noise is eliminated through dark current correction.

[0042] For each pixel point (x,y), the measured reflection intensity I(x,y,λ) is compared with the standard whiteboard reference value I(λ), the intensity attenuation ratio is calculated through natural logarithm operation, and then combined with the reciprocal of the optical path length d to obtain the attenuation coefficient μ(x,y,λ) per unit path. Specifically, the system automatically executes the following calculation process: For each pixel point, calculate I(x,y,λ) / I(λ). If the ratio approaches 0 (i.e., extremely weak reflection), trigger the overflow protection mechanism (the threshold is set as I(x,y,λ) < 0.01I(λ)); Take the natural logarithm ln(I(x,y,λ) / I(λ)) of the intensity ratio, and the negative sign indicates the attenuation direction; Divide the above result by the optical path length d (unit: mm) to obtain μ(x,y,λ)=-ln(I(x,y,λ) / I(λ)) / d, whose unit is mm⁻¹.

[0043] Step S4 includes: Step S41: Perform Z-score normalization on the invasive features to eliminate the dimensional differences between different features. At the same time, perform Min-Max normalization on the metastasis ability features and map them to the [0,1] interval to adapt to the model input requirements; Step S42: Design a dynamic weight allocation network. The input layer includes static features and dynamic features. Model the temporal attenuation coefficient through a gated recurrent unit to generate dynamic feature weights, and at the same time optimize and determine the static feature weights based on ridge regression; Step S43: Input the weighted and fused features into the weighted summation model to obtain the initial malignancy score of the tumor tissue, and compare the initial malignancy score of the tumor tissue with the initial malignancy score intervals corresponding to the preset malignancy grades; thereby obtaining the malignancy grade of the tumor tissue; wherein the weighted summation model is specifically expressed as: Take the fractal dimension, curvature entropy, blood vessel density index, metabolic activity score, and degradation index in the invasiveness features as the numerators of each fraction, take the standard fractal dimension threshold, standard curvature entropy threshold, standard blood vessel density index threshold, standard metabolic activity score threshold, and standard degradation index threshold corresponding to the benign grade of the tumor tissue as the denominators of the corresponding fractions, and multiply each fraction by the corresponding static feature weight; Take the attenuation coefficients of each pixel point of the metastasis ability feature in each band as the numerators of each fraction, take the reference attenuation coefficients of each pixel point of the tumor tissue corresponding to the benign grade in each band as the denominators of the corresponding fractions, and multiply each fraction by the corresponding dynamic feature weight; then perform summation and mean calculation for each band and each pixel. Overall, add the weighted summation result of the invasiveness features and the weighted summation result of the metastasis ability features, and divide by the number 2 to obtain the initial malignancy evaluation coefficient of the tumor tissue; multiply the initial malignancy evaluation coefficient of the tumor tissue by the reference score corresponding to the preset unit initial malignancy evaluation coefficient to finally obtain the initial malignancy score of the tumor tissue.

[0044] Step S42 includes: Step S421: Perform Z-score standardization on the invasiveness features and piecewise standardization on the metastasis ability features to eliminate the dimensional difference; encode the invasiveness features as static features and the metastasis ability features as dynamic features; at the same time, encode the two types of feature parameters as static feature vectors and dynamic feature tensors respectively as the network input. Step S422: Use a gated recurrent unit to model the dynamic feature tensor. The network structure includes two hidden layers (number of neurons = 64), time step = 7 (corresponding to one week of continuous monitoring data), and control the temporal dependence relationship through the update gate and reset gate, and output the hidden state h(t) at each time step; finally, input the hidden state sequence into the fully connected layer to generate the dynamic feature weight. Step S423: Based on the static feature vector and the clinical metastasis label, construct a ridge regression model. The objective function is to minimize the sum of the prediction error and the L2 norm of the weight parameters, and solve to obtain the static weight vector.

[0045] In the embodiment of the present invention, the clinical metastasis label is a binary variable (0 = no metastasis, 1 = metastasis exists) comprehensively determined according to the imaging and pathological results of the patient's postoperative follow-up. The specific parameters include but are not limited to: Metastatic site: such as liver metastasis, lung metastasis, bone metastasis, etc., recorded by the lesion location in the imaging report; Metastasis time: the interval from surgical resection to the first detection of metastatic lesions (e.g., <6 months is short-term metastasis, >6 months is long-term metastasis); Metastatic burden: the total number and diameter of target lesions evaluated according to the RECIST 1.1 standard.

[0046] The specific steps for constructing the ridge regression model in the embodiments of the present invention are as follows: Step 1: Data preprocessing Perform Z-score standardization on static features: subtract the mean from each feature and then divide by the standard deviation; binarize the labels: mark the presence of metastasis as 1, otherwise 0; divide the training set and test set in a 7:3 ratio to avoid data leakage; Adopt 5-fold cross-validation on the training set to test the range of the regularization coefficient λ (e.g., 0.1, 0.5, 1.0, 5.0); select the λ value that minimizes the mean squared error (MSE) of the validation set (e.g., λ = 0.5).

[0047] Construct the ridge regression objective function: minimize the weighted combination of the prediction error (the sum of the squares of the difference between the true label and the predicted value) and the L2 norm of the weight parameters (the sum of the squares of the weights), that is, the total loss = Σ(prediction error)² + λ × Σ(weights)²; directly calculate the weights through the closed-form solution (analytical solution) of matrix operations.

[0048] In the embodiments of the present invention, the L2 norm is the sum of the squares of the weight parameters (e.g., if the weight vector (W = [w_1, w_2, w_3]), then the L2 norm = (w_1^2 + w_2^2 + w_3^2)), and its role is: by penalizing large weight values, prevent the model from over-relying on a few features (overfitting).

[0049] The operation process for solving the static weight vector in the embodiments of the present invention includes: Step 1: Construct the feature matrix and label vector: Static feature matrix X: dimension n×5 (n = number of samples); Metastasis label vector y: dimension n×1. Step 2: Matrix operations Calculate X^T·X: a 5×5 covariance matrix; Add the regularization term λ: obtain X^T·X + λ; Find the inverse matrix (X^T·X + λ)^{-1}; Calculate X^Ty: a 5×1 vector; Final weight = (X^T·X + λ)^{-1}X^Ty.

[0050] In the embodiments of the present invention, Z-score normalization is performed on the invasive features: First, calculate the mean and standard deviation of each feature in all samples, subtract the mean of each feature value from it and then divide by the standard deviation, so as to eliminate the dimensional differences between different features. Piecewise normalization is performed on the metastasis ability features: Taking every 24 hours as a time window, calculate the mean and standard deviation of the attenuation coefficients of each band within the window respectively, and convert the original value into the normalized value within the window to retain the dynamic time series characteristics. The normalized invasive features are encoded into a static feature vector (with a dimension of 5, corresponding to each invasive feature parameter respectively), and the time series attenuation coefficient is encoded into a dynamic feature tensor (with a dimension of the time step multiplied by 5, and the time step is set to 7 days to cover the complete metabolic cycle). The static feature vector and the dynamic feature tensor are input into the network through independent data channels to ensure modality specificity.

[0051] The gated recurrent unit (GRU) is used to perform time series modeling on the dynamic feature tensor. The network contains two hidden layers (with 64 neurons in each layer), and the time step is fixed at 7 days. The retention ratio of historical information is controlled by the update gate, and the influence of the current input on the state update is adjusted by the reset gate, so as to capture the short-term fluctuations and long-term trends of the time series attenuation coefficient. The dynamic feature vector (including the normalized attenuation coefficients of three bands) is input at each time step, and the corresponding hidden state is output and passed to the next time step; finally, the hidden state sequence of all time steps is input into the fully connected layer (with the activation function being Softmax) to generate the dynamic weight vector.

[0052] Based on the static feature vector and the clinical metastasis label (binary classification: 0 indicates no metastasis, 1 indicates the presence of metastasis), a ridge regression model is constructed. The objective function is defined as minimizing the sum of squared prediction errors and the L2 regularization term of the weight parameters simultaneously: The prediction error is the sum of squared differences between the true label and the model prediction value, and the L2 regularization term is the sum of squares of all weight values multiplied by the regularization coefficient (selected as 0.5 through cross-validation). The weight vector is directly solved through a closed-form solution. The specific steps include: multiplying the transposed standardized static feature matrix by itself, adding the product of the regularization coefficient and the identity matrix, taking the inverse of the resulting matrix, and then multiplying by the product of the transposed feature matrix and the label vector.

[0053] Example 2 Please refer to Figure 2 As shown, the tumor classification system based on tumor hyperspectral images includes a hyperspectral image processing module, an image invasion feature extraction module, an image metastasis feature extraction module, and a tumor grade determination module, where each module is connected by wired and / or wireless connection methods to realize data transmission between each module; Hyperspectral image processing module: Obtain the multi-band reflection spectral image of the tumor tissue sample through a hyperspectral imaging device, and perform noise elimination and normalization processing on it; Image invasion feature extraction module: Through spatial domain morphological analysis and spectral domain metabolic feature extraction, the invasiveness features of tumor tissues are quantitatively obtained; Image metastasis feature extraction module: Based on the spectral scattering attenuation rate of the tumor edge region in a fixed band, the dynamic influence of the spatial distribution on the metastasis ability of tumor tissues is quantified to obtain the metastasis ability features of tumor tissues; Tumor grade determination module: By integrating the invasiveness and metastasis ability features through a dynamic weight allocation network, the comprehensive malignancy grading of tumor tissues is output.

[0054] It should be understood that in various embodiments of the present application, the magnitudes of the sequence numbers of the above processes do not mean the order of execution, and the order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0055] It should be understood that determining B based on A does not mean determining B only based on A, and B can also be determined based on A and / or other information.

[0056] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0057] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A tumor classification method based on tumor hyperspectral images, characterized in that: include: Step S1: obtaining a multi-band reflectance spectrum image of a tumor tissue sample by a hyperspectral imaging device, and performing noise elimination and standardization processing on the image; Step S2: quantify the invasive characteristics of tumor tissue by spatial domain morphological analysis and spectral domain metabolic feature extraction; Step S3: quantify the dynamic effect of spatial distribution on the metastatic ability of tumor tissue based on the spectral scattering attenuation rate of the tumor edge area in a fixed band, and obtain the metastatic ability characteristics of tumor tissue; Step S4: Integrate the invasiveness and metastasis ability characteristics through a dynamic weight allocation network to output a comprehensive malignancy grade of the tumor tissue.

2. The tumor classification method based on tumor hyperspectral images according to claim 1, characterized in that: Step S1 includes: Step S11, using a hyperspectral imaging device to scan the tumor tissue at a preset spatial resolution, and scanning to obtain a multi-band reflectance spectrum image of the tumor tissue sample; Step S12, calculating the spectral gradient value of the adjacent band for each pixel point in the spectral image, the specific calculation method is: subtracting the reflectance of the previous band from the reflectance of the next band of each pixel point, and dividing by the band interval width, to obtain the spectral gradient value of the adjacent band calculated for each pixel point; setting a gradient mutation threshold, if the spectral gradient values ​​of three consecutive adjacent bands of the same pixel point in the spectral dimension all exceed the gradient mutation threshold, then the above three adjacent bands are determined to be abnormal bands; Step S13, performing step S12 detection on all pixel points of the multi-band reflectance spectrum image of the tumor tissue sample, counting the frequency of abnormal bands, if the pixel points of a certain abnormal band in the spectrum image exceed the set ratio, the abnormal band is marked as abnormal, and the entire abnormal band is added to the mask blacklist; for the remaining band images, a 7×7 pixel search window is used to calculate the similarity weight of the grayscale value of the pixel points in the window, and a weighted average is performed on each pixel point based on the weight matrix to eliminate speckle noise; Step S14: Collect reference data under the same imaging conditions based on a standard whiteboard, calculate the device gain coefficient matrix, and normalize the reflectance of the hyperspectral image; then select the spectral curve of the central pixel of the tumor area as a reference benchmark, calculate the cosine similarity of the spectra of other pixel points, and if the similarity is lower than the error tolerance threshold verified by the experiment, use the bicubic interpolation method to regenerate the spectral data of the distorted area.

3. The tumor classification method based on tumor hyperspectral images according to claim 2 is characterized in that: Step S14 includes the following sub-steps: Step S141: Under the same environmental conditions, use the same hyperspectral device to photograph a standard whiteboard to obtain its hyperspectral image data cube; Step S142: for each band, calculate the average reflectivity of all pixels in the standard whiteboard image, divide the calibrated reflectivity value of the standard whiteboard by the average reflectivity of the corresponding band, and obtain the device gain coefficient matrix; Step S143, multiplying the reflectance value of each pixel point in the multi-band reflectance spectrum image corresponding to the tumor tissue sample by the coefficient of the corresponding position in the gain coefficient matrix, and outputting normalized reflectance data, wherein the normalized reflectance data ranges from 0 to 1; Step S144, selecting a uniform area of ​​10×10 pixels in the core area of ​​the multi-band reflectance spectrum image corresponding to the tumor tissue sample, calculating its average spectrum as a reference benchmark, and for each pixel, multiplying its spectrum curve with the reference spectrum point by point according to the band, and then summing the sum, and then dividing it by the square root of the sum of the squares of the spectral values ​​of the two to obtain a similarity score; if the similarity score is lower than the error tolerance threshold verified by the experiment, it is determined that the spectrum of the pixel is distorted; Step S145 , taking the distorted pixel as the center, taking the spectral data with similarity meeting the standard in the surrounding 5×5 pixel area, and calculating the interpolation through the cubic polynomial weight function to generate a new spectral curve.

4. The tumor classification method based on tumor hyperspectral images according to claim 1, characterized in that: Step S2 includes: Step S21, using the Canny edge detection algorithm to process the spatial domain grayscale information of the hyperspectral image, setting the Gaussian filter standard deviation σ=1.5 to smooth the noise, and extracting the initial edge contour based on the high and low threshold ratio of 1:3; performing morphological closing operations on the initial edges, filling the gaps and generating a continuous tumor-normal tissue boundary mask; then calculating the boundary fractal dimension based on the box counting method: dividing the mask image into grids of different sizes, counting the number of grids covering the boundary, and fitting the slope under the double logarithmic coordinates as the fractal dimension value; further extracting the curvature sequence along the boundary, calculating the curvature value of each point by the quadratic differential method, discretizing it into 10 intervals and statistically distributing the probability, and finally calculating the Shannon entropy as the curvature entropy; Step S22: In the 580nm band, the Otsu adaptive threshold segmentation method is used to extract the vascular area from the spatial domain image, and the vascular area ratio is calculated as the vascular density index; the edema area is detected in the 940nm band simultaneously, a binary mask is generated by setting a threshold whose reflectance is 30% lower than the average value of the whole image, and a 5×5 circular structure element is used for morphological corrosion to eliminate the interference of scattered noise; the metabolic activity score is calculated in this way, and the calculation formula is that the reflectance of the 940nm band is used as the numerator of the fraction and the reflectance of the lipid band in the hyperspectral image is used as the denominator of the fraction, thereby calculating the metabolic activity score; Step S23: Quantitatively analyze the degree of collagen degradation in the tumor area by Raman spectroscopy. The specific implementation process is as follows: first, detect the characteristic peak intensity of the tumor tissue in the 1550cm⁻¹ band of the Raman spectrum, calculate the spectral intensity integral value of the band by a numerical integration method, and compare it with the corresponding band intensity integral reference value of the normal tissue adjacent to the cancer of the same patient, calculate the collagen intensity ratio of the tumor area and the normal tissue, and subtract the ratio from 1 to obtain the degradation index.

5. The tumor classification method based on tumor hyperspectral images according to claim 1, characterized in that: Step S3 includes: The metastatic ability characteristics of tumor tissue are specifically the attenuation coefficients of each pixel in each band. The specific calculation logic is: For each pixel point (x, y), the reflection intensity I(λ) of the standard white board in band λ is used as a benchmark, and the actual measured reflection intensity I(x, y, λ) of the tumor area is compared with it. Combined with the optical path length d, the attenuation coefficient of each pixel point in each band is obtained by the formula u(x, y, λ) = -(1 / d)·ln(I(x, y, λ) / I(λ)).

6. The tumor classification method based on tumor hyperspectral images according to claim 1, characterized in that: Step S4 includes: Step S41: Z-score normalization is performed on the invasiveness feature to eliminate the dimensional differences between different features, and Min-Max normalization is performed on the metastasis ability feature to map it to the [0,1] interval to adapt to the model input requirements; Step S42: Design a dynamic weight allocation network, where the input layer includes static features and dynamic features, model the temporal attenuation coefficient through a gated recurrent unit, generate dynamic feature weights, and determine static feature weights based on ridge regression optimization; Step S43: Input the weighted fused features into the weighted sum model to obtain the initial malignancy score of the tumor tissue, and compare the initial malignancy score of the tumor tissue with the preset initial malignancy score intervals corresponding to each malignancy level; thereby obtaining the malignancy level of the tumor tissue.

7. The tumor classification method based on tumor hyperspectral images according to claim 6, characterized in that: Step S42 includes: Step S421, performing Z-score standardization on the invasiveness feature, and performing segmented standardization on the metastasis ability feature to eliminate dimensional differences; encoding the invasiveness feature as a static feature, and encoding the metastasis ability feature as a dynamic feature; and encoding the two types of feature parameters as a static feature vector and a dynamic feature tensor, respectively, as network input; Step S422: The dynamic feature tensor is modeled using a gated recurrent unit. The network structure includes two hidden layers, and the time step length is 7. The timing dependency is controlled by updating the gate and resetting the gate, and the hidden state h(t) of each time step is output. Finally, the hidden state sequence is input into the fully connected layer to generate the dynamic feature weight. Step S423: Based on the static feature vector and the clinical metastasis label, a ridge regression model is constructed, and the objective function is to minimize the sum of the prediction error and the L2 norm of the weight parameter, and the static weight vector is obtained by solving.

8. A tumor classification system based on tumor hyperspectral images, characterized in that: The method for classifying tumors based on tumor hyperspectral images according to any one of claims 1 to 7 is implemented, comprising: Hyperspectral image processing module: obtains multi-band reflectance spectral images of tumor tissue samples through hyperspectral imaging equipment, and performs noise elimination and standardization processing on them; Image invasion feature extraction module: Through spatial domain morphological analysis and spectral domain metabolic feature extraction, the invasive characteristics of tumor tissue are quantified; Image transfer feature extraction module: Based on the spectral scattering attenuation rate of the tumor edge area in a fixed band, the dynamic impact of spatial distribution on the metastatic ability of tumor tissue is quantified to obtain the metastatic ability characteristics of tumor tissue; Tumor grade determination module: Integrates the invasiveness and metastasis characteristics through a dynamic weight allocation network to output a comprehensive malignancy grade of tumor tissue.

Images