Deep learning-based immune score data processing method and system

Abstract

This invention relates to the field of medical image processing technology, specifically to a method and system for processing immune scoring data based on deep learning. The method includes: acquiring an immunohistochemically stained whole-section image to be scored; dividing the whole-section image into non-overlapping image blocks; performing multi-level spectral separation and decomposition on each image block to construct a spatial density distribution map of immune cells in the whole-section; segmenting the tumor central region and invasion edge region on the spatial density distribution map according to tissue morphology gradients, and extracting the corresponding local density distribution. This invention, by inverting the effective chemotaxis coefficient and effective proliferation coefficient of the reaction-diffusion coupling system from the spatial density distribution map of immune cells, can directly quantify the degree of mechanistic deviation of immune cell chemotaxis and proliferation capacity in the tumor microenvironment, overcoming the limitations of traditional methods that rely solely on static cell density counting, and giving the scoring a clear biophysical meaning.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for processing immune scoring data based on deep learning. Background Technology

[0002] Immunoscoring is a method used to quantify the state of immune cell infiltration in the tumor microenvironment. Traditional immunoscoring schemes mainly rely on counting the density of immune cells in tissue sections after immunohistochemical staining, counting the number of positive cells in the central and invasive edge regions of the tumor, and then classifying them according to the distribution of the statistical values ​​in the population. This paradigm, which is based on static density counting, is essentially a summary of a snapshot of the spatial distribution of immune cells, and it is difficult to reflect the degree of abnormality of mechanistic parameters such as cell migration and proliferation in the tumor microenvironment.

[0003] In recent years, methods have emerged that use deep learning to directly map full-slice images to scoring results, reducing the subjective influence of manual region delineation and counting. However, these methods solidify the feature extraction and scoring mapping process into an opaque model structure, resulting in output scores that lack interpretable biophysical meaning, making it difficult to explain the microenvironmental mechanism changes corresponding to the scores, and limiting their generalization ability to data from different sources. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for processing immune scoring data, comprising: Acquire the immunohistochemical stained whole slide image to be scored, divide the whole slide image into non-overlapping image blocks; perform multi-layer spectral separation decomposition on each image block to construct the spatial density distribution map of immune cells in the whole slide. On the spatial density distribution map, the tumor central region and the invasion edge region are segmented according to the histological morphology gradient, and the corresponding local density distribution is extracted. Using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed, and the effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system are inverted by the variational adjoint method. The effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion are compared with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained from non-tumor reference samples to generate a parameter deviation vector; An initial score is obtained by performing an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of a non-tumor reference.

[0005] Preferably, the immunohistochemically stained whole-section image to be scored is acquired, and the whole-section image is divided into non-overlapping image blocks, including: Acquire multispectral immunohistochemical staining whole slide images, extract spectral data of multispectral images under each preset narrowband filter channel, and form a multichannel spectral image cube; The first linear spectral unmixing is performed on the multi-channel spectral image cube. The initial abundance map of each immune-related marker is decomposed using the pre-determined chromatin matrix. All initial abundance maps are superimposed by channel to generate an initial distribution characterization map of immune cells. Calculate the pixel-by-pixel local entropy of the initial distribution characterization map of immune cells, and generate a heterogeneity index map characterizing the degree of spatial heterogeneity of immune cells. Multi-scale Gaussian difference filtering is applied to the heterogeneity index map to extract heterogeneous gradient ridges. The heterogeneous gradient ridges are used as constraint boundaries, and the whole slice image is divided into non-overlapping tissue micro-regions using marker watershed segmentation. For each tissue microregion, a quadtree recursively split is performed according to the preset minimum analysis unit size until the standard deviation of the heterogeneity index in the split subregion is lower than the set threshold. All the subregions obtained by the final split are used as non-overlapping image blocks.

[0006] Preferably, multi-layer spectral separation and decomposition are performed on each image patch to construct a spatial density distribution map of immune cells in the whole slice, including: Obtain a pre-constructed hierarchical immunostaining spectral library; wherein the hierarchical immunostaining spectral library contains a first layer of coarse separation spectral matrix and at least one subsequent layer of fine separation spectral matrix, and the coarse separation spectral matrix is ​​used to separate the immunomarker signal from the tissue autofluorescence background, and the fine separation spectral matrix is ​​used to distinguish the specific spectra of different immune cell subsets; For each image block, the spectral data within the image block are subjected to supervised non-negative least squares unmixing using the first-layer coarse separation spectral basis matrix to obtain the overall abundance map of immunomarkers after removing the autofluorescence background. Using the overall abundance map of immunomarkers as a spatial prior, under the subsequent layer-by-layer separation of the spectral basis matrix, a spatially weighted non-negative matrix decomposition is performed. The decomposition coefficients are applied with a local smoothing regularization term based on the overall abundance map of immunomarkers, and the abundance maps of each preset immune cell subpopulation are separated layer by layer. The abundance maps of each subpopulation output in the last layer are used as the immune cell density distribution of the image patch. After traversing all image blocks and obtaining the immune cell density distribution of each image block, the immune cell density on both sides of the boundary of adjacent image blocks is directionally weighted and smoothed to construct a spatial density distribution map of immune cells in the whole slice.

[0007] Preferably, using heterogeneous gradient ridges as guiding boundaries, the immune cell density on both sides of the boundary of adjacent image blocks is directionally weighted and smoothly fused, including: For any two adjacent image patches, extract their shared boundary and calculate the boundary normal vector and tangent vector point by point along the shared boundary; At each boundary point of the shared boundary, the direction of the heterogeneous gradient ridge is obtained in a preset local region centered on that boundary point, and the angle between the direction of the heterogeneous gradient ridge and the boundary normal vector is calculated. When the included angle is less than a preset included angle threshold, a first fusion weight along the normal vector direction is assigned to the boundary point; otherwise, a second fusion weight is assigned. The first fusion weight is less than the second fusion weight. For each boundary point of the shared boundary, a third fusion weight is set along the tangent vector direction; wherein the third fusion weight is greater than or equal to the second fusion weight; A local anisotropic Gaussian smoothing kernel is constructed by using the fusion weights of each boundary point along the normal vector direction and the fusion weights along the tangent vector direction. Within a preset width range on both sides of the shared boundary, the immune cell density of adjacent image blocks is convolved and fused using a locally anisotropic Gaussian smoothing kernel to eliminate block effects and preserve the spatial heterogeneity boundary of immune cells defined by heterogeneous gradient ridges.

[0008] Preferably, the tumor central region and invasion edge region are segmented according to the histological morphology gradient on the spatial density distribution map, and the corresponding local density distribution is extracted, including: On the spatial density distribution map of immune cells in a whole slice, calculate the gradient vector field of immune cell density and generate a density gradient magnitude map; In the density gradient magnitude map, the local maximum point of the gradient magnitude is used as the seed point. The path of the gradient magnitude is tracked along the gradient direction and the gradient magnitude is monotonically decreasing. The tracking is terminated when the gradient magnitude on the path drops to a preset proportion of the gradient magnitude of the seed point. The termination points are connected to form the candidate boundary of the tumor center. The area enclosed by the candidate boundary is taken as the tumor center area. Starting from the candidate boundary of the tumor center, expand outward along the gradient direction, and calculate the immune cell density difference between each pixel and its corresponding pixel back to the candidate boundary along the gradient direction during the expansion process. Pixels whose density difference values ​​converge to a preset convergence threshold range are marked as invasion edge termination points. All invasion edge termination points are connected to form the outer boundary of the invasion edge. The annular region between the candidate boundary of the tumor center and the outer boundary of the invasion edge is taken as the invasion edge region. Using masks of the tumor center and invasion edge regions, the local density distribution of immune cells in the whole slice was extracted from the spatial density distribution map of immune cells.

[0009] Preferably, using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are assigned as the immune cell density field of the tumor center region and the immune cell density field of the invasion edge region, respectively, and the immune cell density field in each region is set as a quasi-steady-state field. Establish the coupling relationship between the immune cell density field and the chemokine concentration field; wherein, the temporal variation of the immune cell density field is given by the immune cell diffusion flux driven by the chemokine concentration gradient and the immune cell self-proliferation flux, and the temporal variation of the chemokine concentration field is given by the immune cell density-dependent chemokine secretion and production and the natural decay of chemokines. Substituting the quasi-steady-state field as the solution of the coupling relationship when the time variation approaches zero into the coupling relationship, we obtain a reaction-diffusion coupled system with local density distribution as the spatial constraint.

[0010] Preferably, the effective chemotaxis coefficient and effective multiplication coefficient in a reaction-diffusion coupled system are inverted using the variational adjoint method, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are taken as the observation target field, and the immune cell density field output by the reaction-diffusion coupling system under quasi-steady-state conditions is taken as the prediction field. A loss functional is constructed, where the loss functional is the integral of the point-by-point deviation between the prediction field and the observation target field over the entire spatial domain. Construct the adjoint system corresponding to the reaction-diffusion coupling system, take the variation of the loss functional with respect to the immune cell density field as the driving source term of the adjoint system, solve the adjoint system in reverse time, and obtain the spatiotemporal adjoint variable field; In the forward equation of the reaction-diffusion coupled system, the partial derivatives of the chemotaxis flux term with respect to the effective chemotaxis coefficient and the partial derivatives of the proliferation flux term with respect to the effective proliferation coefficient are extracted. The adjoint variable field is then spatially integrated with the partial derivative information to obtain the gradients of the loss functional with respect to the effective chemotaxis coefficient and the effective proliferation coefficient. The gradient-driven gradient descent optimization process iteratively updates the effective chemotaxis coefficient and the effective multiplication coefficient. When the value of the loss functional drops below the preset convergence threshold, the iteration stops, and the current effective chemotaxis coefficient and the effective multiplication coefficient are output as the inversion result.

[0011] Preferably, the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion are differiated from the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples to generate a parameter deviation vector, including: A set of non-tumor reference samples with the same tissue type as the immunohistochemically stained whole-slice images to be scored was obtained. For the whole-slice images of each sample in the non-tumor reference sample set, the same image block division, multi-layer spectral separation and decomposition, spatial density distribution map construction, region segmentation, local density extraction, reaction-diffusion coupling system construction and variational adjoint inversion were used as for the samples to be scored to obtain the effective chemotaxis coefficient and effective proliferation coefficient of each sample. Robust statistical estimations are performed on the chemotaxis coefficient set and the proliferation coefficient set respectively, and the statistical center value of the chemotaxis coefficient set is used as the benchmark chemotaxis coefficient, and the statistical center value of the proliferation coefficient set is used as the benchmark proliferation coefficient. The effective chemotaxis coefficient of the sample to be scored obtained by subtracting the baseline chemotaxis coefficient is used to obtain the chemotaxis coefficient deviation; the effective proliferation coefficient of the sample to be scored is subtracted from the baseline proliferation coefficient to obtain the proliferation coefficient deviation. The deviations of the chemotaxis coefficient and the deviations of the proliferation coefficient are arranged into a two-dimensional vector, which serves as the parameter deviation vector.

[0012] Preferably, an initial score is obtained by performing an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score, including: Obtain a predetermined immune score projection vector; wherein, the immune score projection vector is a two-dimensional vector, the two components of which correspond to the chemotaxis deviation dimension and the proliferation deviation dimension, respectively. The two-dimensional vector is obtained by fitting the clinical immune score to the chemotaxis coefficient deviation and the proliferation coefficient deviation of the reference population. The initial score is obtained by performing an inner product operation between the parameter deviation vector and the immune score projection vector. Obtain the initial score set of the preset reference population, determine the quantile of the initial score in the initial score set, map the quantile to the preset percentage scoring scale, and obtain the final immune score.

[0013] A deep learning-based immune scoring data processing system, applicable to the aforementioned deep learning-based immune scoring data processing methods, including: The image segmentation processing module is used to acquire the immunohistochemical stained whole slide image to be scored, divide the whole slide image into non-overlapping image blocks, perform multi-layer spectral separation decomposition on each image block, and construct a spatial density distribution map of immune cells in the whole slide. The region segmentation and extraction module is used to segment the tumor central region and the invasion edge region on the spatial density distribution map according to the histological morphology gradient, and extract the corresponding local density distribution. The mechanism parameter inversion module is used to construct a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration, using local density distribution as a quasi-steady-state observation constraint. It then uses the variational adjoint method to invert the effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system. The parameter deviation generation module is used to differiate the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples to generate a parameter deviation vector; The immune score output module is used to perform an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector to obtain an initial score. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of a non-tumor reference.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention, by inverting the effective chemotaxis coefficient and effective proliferation coefficient of the reaction-diffusion coupling system from the spatial density distribution map of immune cells, can directly quantify the degree of mechanistic deviation of the chemotaxis and proliferation capacity of immune cells in the tumor microenvironment. This overcomes the limitations of traditional methods that rely solely on static cell density counting, giving the scoring a clear biophysical meaning. Furthermore, by differentiating the inverted mechanism parameters with the baseline parameters obtained from non-tumor reference samples under the same processing procedure, it can eliminate systematic biases introduced by differences in processing procedures, achieving individualized quantitative assessment of mechanisms. Moreover, the scoring process, from image patch division, spectral separation and decomposition, region segmentation to parameter inversion and projection scoring, is based on a defined physical model and mathematical operations at each step, avoiding the black-box defects of deep learning end-to-end scoring methods. The scoring results are traceable and the process is verifiable. This invention employs an adaptive image block segmentation and directional weighted fusion strategy driven by the heterogeneity of immune cell distribution. This strategy can eliminate block effects while preserving the spatial heterogeneity boundaries of immune cells, providing accurate spatial constraints for subsequent response-diffusion inversion and improving the reliability of scoring results. Furthermore, by adaptively segmenting the tumor center region and invasion edge region based on the decay law of density gradient amplitude and the convergence condition of density difference, it can objectively reflect the tissue gradient characteristics of the spatial distribution of immune cells, avoiding the subjectivity of manually delineated regions. Moreover, by establishing benchmark parameters through non-tumor reference samples and percentile standardization of the reference population, individual scores can be mapped to a unified scoring scale, making scores between different samples comparable. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention; Figure 2This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0016] In the diagram: 1. Image block processing module; 2. Region segmentation and extraction module; 3. Mechanism parameter inversion module; 4. Parameter deviation generation module; 5. Immune score output module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, please refer to Figure 1 This invention provides a technical solution: a deep learning-based method for processing immune scoring data, comprising: S1. Obtain the immunohistochemical stained whole slide image to be scored, and divide the whole slide image into non-overlapping image blocks; perform multi-layer spectral separation decomposition on each image block to construct the spatial density distribution map of immune cells in the whole slide. S2. On the spatial density distribution map, the tumor central region and the invasion edge region are segmented according to the histological morphology gradient, and the corresponding local density distribution is extracted. S3. Using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed. The effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system are inverted using the variational adjoint method. S4. Difference the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained from non-tumor reference samples to generate a parameter deviation vector; S5. The parameter deviation vector is multiplied by the predetermined immune score projection vector to obtain the initial score. The initial score is then standardized based on the percentile of the preset reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of the non-tumor reference.

[0019] It should be noted that, firstly, the immunohistochemically stained whole-section image to be scored is acquired. This image is a tissue section scan image acquired through a multispectral imaging system, which can simultaneously record the spatial distribution information of multiple immunomarkers. The whole-section image is divided into non-overlapping image blocks. The purpose is to decompose the large-scale, high-resolution whole-section image into local units that are easy to calculate block by block, laying the foundation for subsequent quantitative analysis. Multi-level spectral separation decomposition is performed on each image block. This refers to using spectral unmixing technology to separate the abundance information of different immune cell subsets from the multispectral signal layer by layer. This process first separates the tissue autofluorescence background from the immunomarker signal, and then finely distinguishes the characteristic spectra of different immune cell subsets, thereby obtaining the local density of various immune cells in each image block. After all image blocks have completed multi-level spectral separation decomposition, the local immune cell densities of each image block are stitched and fused to construct an immune cell spatial density distribution map covering the entire tissue section. Next, the tumor center region and the invasion edge region are segmented on the constructed spatial density distribution map according to the histological morphology gradient, and the corresponding local density distribution is extracted. The tumor center and the invasion edge are two key analysis regions defined in the immune scoring standard. This invention does not rely on manual delineation, but directly calculates the gradient field of density change on the immune cell density distribution map. The boundary of the tumor center region is automatically identified by the amplitude change law of the density gradient. Then, the outer boundary of the invasion edge is determined by tracking the change trend of density difference value outward along the gradient direction, thereby realizing adaptive segmentation of the two regions. After the segmentation is completed, the immune cell density values ​​of each pixel in the tumor center region and the invasion edge region are extracted to form local density distribution data. Then, using the extracted local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed. This coupling system characterizes the interaction of two core dynamic processes in the tumor microenvironment: first, immune cells migrate in a directional manner driven by the chemokine concentration gradient; second, immune cells proliferate in local areas. Under the quasi-steady-state assumption, the density distribution of immune cells tends to a time-steady state. At this point, the actual measured local density distribution is substituted into the coupling system as the observation value of this steady state, so that the system equation contains two unknown parameters to be solved, namely the effective chemotaxis coefficient and the effective proliferation coefficient. The effective chemotaxis coefficient reflects the strength of the immune cell migration ability under the action of chemokines, and the effective proliferation coefficient reflects the strength of the immune cell self-proliferation ability in the tumor local area. Subsequently, the effective chemotaxis coefficient and effective multiplication coefficient in the coupled system are inverted using the variational adjoint method. Specifically, the deviation between the predicted density field output by the coupled system under given parameters and the measured local density distribution is constructed as a loss functional. The adjoint system corresponding to the coupled system is introduced. By solving the coupled system in the forward direction and the adjoint system in the reverse direction, the gradient information of the loss functional with respect to the undetermined parameters is calculated. Then, the parameters are iteratively updated along the gradient descent direction until the deviation between the predicted density field and the measured density distribution converges. The parameter values ​​at this time are output as the inversion result. The advantage of the variational adjoint method is that it can efficiently handle the sensitivity calculation of spatially distributed observation data to model parameters and avoid the computational burden of large-scale parameter space search. After obtaining the effective chemotaxis coefficient and effective proliferation coefficient of the sample to be scored, they are differiated from the baseline chemotaxis coefficient and baseline proliferation coefficient obtained from non-tumor reference samples to generate a parameter deviation vector. Non-tumor reference samples refer to samples with the same tissue type but without tumorous changes. These samples are processed and analyzed using the same image processing and analysis procedures as the sample to be scored to obtain the chemotaxis coefficient and proliferation coefficient of each reference sample. The statistical center value of these coefficients is determined as the benchmark through robust statistical methods. The two coefficients of the sample to be scored are subtracted from the corresponding benchmarks to obtain the deviation of the chemotaxis coefficient and the deviation of the proliferation coefficient, which are arranged into a two-dimensional vector, namely the parameter deviation vector. The two dimensions of this vector represent the degree of deviation of the sample to be scored from normal tissue in terms of immune cell chemotaxis ability and proliferation ability, respectively. Finally, the initial score is obtained by performing an inner product operation on the parameter deviation vector using the pre-determined immune score projection vector. The immune score projection vector is a two-dimensional weight vector pre-fitted using a reference population sample. Its two components correspond to the chemotaxis deviation dimension and the proliferation deviation dimension, respectively, reflecting the relative importance of the deviations in these two dimensions in the comprehensive immune status assessment. The inner product operation compresses the deviation information of the two dimensions of the parameter deviation vector into a scalar, namely the initial score. The initial score is then standardized based on the percentile of the pre-determined reference population, which maps the quantile position of the initial score in the initial score set of the reference population to a pre-determined scoring scale, thereby obtaining the final immune score. The final immune score characterizes the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored relative to the non-tumor reference. The higher the value, the more significant the deviation.

[0020] In an optional embodiment, acquiring an immunohistochemically stained whole-section image to be scored, and dividing the whole-section image into non-overlapping image blocks, includes: Acquire multispectral immunohistochemical staining whole slide images, extract spectral data of multispectral images under each preset narrowband filter channel, and form a multichannel spectral image cube; It should be noted that multispectral immunohistochemical staining whole-section images are obtained by scanning tissue sections stained with multiple immunomarkers simultaneously using a multispectral imaging system. Unlike ordinary color images, this image records the light intensity values ​​of a set of continuous or discrete wavelength channels at each spatial pixel position. The preset narrowband filter channel refers to a specific wavelength range selected during imaging. Each channel corresponds to the absorption peak position of the staining characteristic spectrum of one or a class of immunomarkers. After extracting spectral data from each filter channel, these data are stacked layer by layer according to the channel order to form a three-dimensional data array. Two dimensions are the rows and columns of the image, and the third dimension is the spectral channel. This three-dimensional data array is the multichannel spectral image cube. The first linear spectral unmixing is performed on the multi-channel spectral image cube. The initial abundance map of each immune-related marker is decomposed using the pre-determined chromatin matrix. All initial abundance maps are superimposed by channel to generate an initial distribution characterization map of immune cells. It should be noted that the first linear spectral unmixing refers to the preliminary decomposition of the spectral image cube using a linear spectral mixing model. Its core principle is that the spectral signal of each pixel in the image can be obtained by linearly superimposing the spectra of several pure staining components according to their abundance ratios. The pre-determined chromatin matrix is ​​a standard set of spectra obtained by performing spectral measurements on individual stained samples before the experiment. Each column in this matrix corresponds to the characteristic spectral curve of an immunomarker staining. By linearly unmixing the multi-channel spectral image cube with the chromatin matrix, the abundance distribution of each immunomarker across the entire slice can be obtained. Each marker corresponds to an initial abundance map. The higher the pixel value in the map, the stronger the expression of the marker at that spatial location. By stacking all the initial abundance maps layer by layer according to the channel direction, the overall spatial distribution of all immune markers can be comprehensively reflected. The resulting comprehensive image is the initial distribution characterization map of immune cells. For example, after using spectral unmixing to separate the abundance distribution of each immune marker from the original multispectral image, each abundance map corresponds to the distribution area of ​​different immune cell subsets in the tissue. By stacking them, the approximate outline of the enriched and sparse areas of immune cells in the whole tissue can be visually presented. Calculate the pixel-by-pixel local entropy of the initial distribution characterization map of immune cells, and generate a heterogeneity index map characterizing the degree of spatial heterogeneity of immune cells. It should be noted that local entropy is an indicator that measures the complexity of signals within a local region. In this step, for each pixel on the initial distribution characterization map of immune cells, a local window of a preset size is taken around it, and the degree of disorder in the distribution of pixel values ​​within the window is calculated, i.e., the local entropy value. If the abundance distribution of immune markers within the local window is uniform, the local entropy is high, reflecting that the distribution of immune cells in this region is relatively disordered and lacks obvious aggregation patterns. If the abundance of immune markers within the local window is relatively clustered or sparse, the local entropy is low, reflecting that the distribution of immune cells in this region is relatively uniform. The local entropy value of each pixel is assigned to the pixel position to generate a heterogeneity index map with the same size as the original map, where high-value pixels correspond to regions with disordered immune cell distribution, and low-value pixels correspond to regions with homogeneous immune cell distribution. Multi-scale Gaussian difference filtering is applied to the heterogeneity index map to extract heterogeneous gradient ridges. The heterogeneous gradient ridges are used as constraint boundaries, and the whole slice image is divided into non-overlapping tissue micro-regions using marker watershed segmentation. It should be noted that multi-scale Gaussian difference filtering refers to smoothing the heterogeneity index map and performing difference operations using multiple Gaussian kernels of different scales to enhance the edges of heterogeneity changes at different scales in the image. In this process, kernels with smaller scale parameters can capture local details of heterogeneity changes, while kernels with larger scale parameters can capture the transition trend of large-scale heterogeneity distribution. The locations with the most significant and continuous gradient changes are extracted from the multi-scale difference filtering results. These locations are connected to form heterogeneous gradient ridges. The heterogeneous gradient ridges mark the boundary locations where the heterogeneity of immune cell distribution in the tissue changes significantly. The extracted heterogeneous gradient ridges are used as the constraint boundaries of the watershed segmentation algorithm. The watershed algorithm treats these ridges as insurmountable watersheds, dividing the entire image region into multiple closed tissue micro-regions. The heterogeneity of immune cell distribution within each micro-region is relatively consistent, while adjacent micro-regions are naturally separated by heterogeneous gradient ridges. For each tissue micro-region, a quadtree recursively split is performed according to the preset minimum analysis unit size until the standard deviation of the heterogeneity index in the split sub-region is lower than the set threshold. All the sub-regions obtained by the final split are used as non-overlapping image patches. It should be noted that quadtree recursive splitting is an adaptive spatial partitioning method. For each tissue microregion, the dispersion of the heterogeneity index values ​​of all pixels within it is first determined. If the standard deviation is higher than a set threshold, it indicates that there is still significant heterogeneity difference within the microregion, and splitting needs to continue. If the standard deviation is lower than the set threshold, it indicates that the heterogeneity within the microregion is sufficiently uniform, and splitting can be stopped. During splitting, a region is divided into four subregions of the same size along both the horizontal and vertical directions. Then, the calculation and judgment of the heterogeneity index standard deviation are repeated for each subregion. The preset minimum analysis unit size refers to the minimum side length that the subregion is allowed to reach during the splitting process. When the subregion size will be smaller than this size after further splitting, splitting will not continue even if the standard deviation has not yet converged, to ensure that each image block has enough pixels to support the statistical stability of subsequent multilayer spectral separation and decomposition. After the above recursive splitting, all the remaining subregions are the non-overlapping image blocks.

[0021] In an optional embodiment, multi-layer spectral separation decomposition is performed on each image patch to construct a spatial density distribution map of immune cells in the whole slice, including: Obtain a pre-constructed hierarchical immunostaining spectral library; wherein the hierarchical immunostaining spectral library contains a first layer of coarse separation spectral matrix and at least one subsequent layer of fine separation spectral matrix, and the coarse separation spectral matrix is ​​used to separate the immunomarker signal from the tissue autofluorescence background, and the fine separation spectral matrix is ​​used to distinguish the specific spectra of different immune cell subsets; It should be noted that the hierarchical immunostaining spectral library is a spectral reference database pre-established through standardized measurements. Its construction involves separately acquiring spectra of known immunomarker-stained samples and tissue autofluorescence samples, extracting their respective characteristic spectral curves, and then cataloging and storing them hierarchically. The first layer, the coarse separation spectral basis matrix, contains two main categories of spectral basis vectors: one is the joint characteristic spectrum of all used immunomarkers, and the other is the typical spectrum of tissue autofluorescence components. Subsequent layers, the fine separation spectral basis matrix, pre-define a corresponding characteristic spectrum for each immune cell subpopulation, such as the specific spectral absorption patterns exhibited by single-positive and double-positive cell subpopulations under their combined labeling. The purpose of this hierarchical structure is to first separate useful immune signals from interfering tissue autofluorescence background at a coarser granularity, and then perform precise subpopulation-level decomposition of the background-disrupted immune signals at a finer granularity. For each image block, the spectral data within the image block are subjected to supervised non-negative least squares unmixing using the first-layer coarse separation spectral basis matrix to obtain the overall abundance map of immunomarkers after removing the autofluorescence background. It should be noted that the purpose of this step is to decompose the spectral signal of each pixel within each image block into the contribution components of immunomarkers and the contribution components of tissue autofluorescence, and to remove the autofluorescence component. Supervised non-negative least squares unmixing refers to finding a set of abundance coefficients that minimize the sum of squares of the differences between the reconstructed spectrum and the measured spectrum, under the premise that the abundance values ​​of each component satisfy the non-negative constraint, using the first-layer coarse separation spectral basis matrix as known supervisory information. The component of the abundance coefficient that corresponds to the joint characteristic spectrum of immunomarkers is spatially recombined to form the overall abundance map of immunomarkers, which records the combined intensity of the staining signals of all immunomarkers at each pixel location. Using the overall abundance map of immunomarkers as a spatial prior, under the subsequent layer-by-layer separation of the spectral basis matrix, a spatially weighted non-negative matrix decomposition is performed. The decomposition coefficients are applied with a local smoothing regularization term based on the overall abundance map of immunomarkers, and the abundance maps of each preset immune cell subpopulation are separated layer by layer. The abundance maps of each subpopulation output in the last layer are used as the immune cell density distribution of the image patch. It should be noted that the overall abundance map of immunomarkers, as a spatial prior, refers to using the overall abundance information obtained in the aforementioned steps as a constraint in the subsequent fine separation process. This directly suppresses the fine separation coefficients of regions with weak signals on the overall abundance map, preventing background noise from being incorrectly amplified during the fine separation stage. Spatially weighted nonnegative matrix decomposition decomposes the spectral data matrix of the image patch under the fine separation spectral basis matrix, while applying a local smoothing regularization term to the decomposition coefficients. This regularization term ensures that adjacent pixels with similar abundance values ​​on the overall abundance map of immunomarkers tend to obtain similar subpopulation abundance coefficients, thus guaranteeing that the separated subpopulation abundance maps are spatially continuous without generating random noise scattering. Layer-by-layer separation refers to first separating the abundance map of the first immune cell subpopulation, then continuing to separate the second subpopulation from the remaining signal, and so on, until all preset subpopulations are separated. The final layer output subpopulation abundance maps directly reflect the density estimates of various immune cells at each pixel location; these maps together constitute the immune cell density distribution of the image patch. After traversing all image blocks and obtaining the immune cell density distribution of each image block, the immune cell density on both sides of the boundary of adjacent image blocks is directionally weighted and smoothed to construct a spatial density distribution map of immune cells in the whole slice. It should be noted that since each image patch undergoes independent multi-level spectral separation and decomposition, adjacent image patches may exhibit density discontinuities on both sides of the shared boundary after independent processing, i.e., block effect. These boundary regions require smoothing and fusion processing. This step utilizes the heterogeneous gradient ridges extracted during the image patch segmentation stage as fusion guide boundaries. These heterogeneous gradient ridges identify locations where substantial changes occur in the heterogeneity of immune cell distribution within the tissue. Over-smoothing should be avoided at these locations to prevent erasing the true spatial heterogeneity boundaries of immune cells. Directional weighted smoothing fusion refers to smoothing the shared boundary between two adjacent image patches. Different smoothing intensities with different weights are applied along the direction perpendicular to the boundary and along the direction parallel to the boundary: when the boundary point coincides with a heterogeneous gradient ridge, the smoothing weight is reduced along the direction perpendicular to the boundary to preserve the density difference on both sides of the boundary, and the smoothing weight is increased along the direction parallel to the boundary to ensure the density continuity of each point on the boundary; when no heterogeneous gradient ridge passes through the boundary point, normal smoothing weights are used in both directions to eliminate stitching marks; after directional weighted smoothing fusion processing, the immune cell density distribution of each image patch is stitched together patch by patch into a continuous and complete whole-slice immune cell spatial density distribution map.

[0022] In an optional embodiment, using heterogeneous gradient ridges as guiding boundaries, directional weighted smoothing fusion of immune cell densities on both sides of adjacent image block boundaries is performed, including: For any two adjacent image patches, extract their shared boundary and calculate the boundary normal vector and tangent vector point by point along the shared boundary; It should be noted that the shared boundary refers to the boundary line between two adjacent image patches that are spatially adjacent. During image processing, the results on both sides of this boundary belong to different image patches and are calculated independently. The boundary normal vector is the direction vector perpendicular to the boundary direction at each point on the boundary, pointing from the current image patch across the boundary to the adjacent image patch. The tangent vector is the direction vector along the boundary extension direction at each point on the boundary, perpendicular to the normal vector. The purpose of calculating the boundary normal vector and tangent vector is to provide a pair of mutually orthogonal reference directions for subsequent directional fusion, so that the fusion strength can be independently controlled along the cross-boundary direction and along the boundary tangent direction. For example, when two image patches are adjacent left and right, the shared boundary is a vertical line segment. Take a point on the boundary, whose normal vector points horizontally and whose tangent vector points vertically. This set of directional information is updated point by point at different boundary points as the boundary direction changes. At each boundary point of the shared boundary, the direction of the heterogeneous gradient ridge is obtained in a preset local region centered on that boundary point, and the angle between the direction of the heterogeneous gradient ridge and the boundary normal vector is calculated. It should be noted that the preset local region is a fixed-size neighborhood window centered on the current boundary point. Within this window, it searches for whether a heterogeneous gradient ridge line crosses the boundary. If it does, its local extension direction at that point is extracted. The angle between the ridge line direction and the boundary normal vector is calculated. The purpose is to determine whether the direction of the tissue structure where the heterogeneity of immune cell distribution changes substantially at the boundary point is consistent with the cross-boundary direction. The smaller the angle, the closer the ridge line is to being perpendicular to the boundary, which means that the shared boundary exactly crosses a real boundary of mutation in the spatial distribution of immune cells. When the included angle is less than a preset included angle threshold, a first fusion weight along the normal vector direction is assigned to the boundary point; otherwise, a second fusion weight is assigned. The first fusion weight is less than the second fusion weight. It should be noted that the preset angle threshold is a criterion used to distinguish whether the density difference at the current boundary point originates from changes in the distribution of real immune cells or from the stitching effect caused by independent processing of image patches. Usually, a small angle value can be selected. When the angle between the ridge direction and the normal vector direction is less than the threshold, it means that the density change at this location is highly consistent with the heterogeneous boundary of the tissue and should be protected. Therefore, a lower first fusion weight is assigned in the direction perpendicular to the boundary to reduce the smoothing intensity and avoid smoothing out the real density difference. When the angle is greater than or equal to the threshold, it means that there is no significant histological boundary of immune cell distribution at this location. The density difference on both sides of the boundary is more likely to come from the stitching effect caused by block processing. Therefore, a higher second fusion weight is given in this direction to fully smooth the transition. For each boundary point of the shared boundary, a third fusion weight is set along the tangent vector direction; wherein the third fusion weight is greater than or equal to the second fusion weight; It should be noted that the tangent vector direction is parallel to the direction of the boundary itself. Smoothing along this direction will not weaken the protective effect of density difference across the boundary. On the contrary, it helps to maintain the spatial continuity of immune cell density between points on the boundary. Setting the third fusion weight to be greater than or equal to the second fusion weight ensures that there is a sufficiently strong smoothing effect along the direction of the boundary. This can make the fused boundary area present a smooth and continuous transition in both visual and analytical terms, avoiding jagged density jumps at the boundary. A local anisotropic Gaussian smoothing kernel is constructed by using the fusion weights of each boundary point along the normal vector direction and the fusion weights along the tangent vector direction. It should be noted that the locally anisotropic Gaussian smoothing kernel is a directional smoothing filter, shaped like a two-dimensional Gaussian surface. However, the broadening degree of this surface differs in two orthogonal directions: the broadening in the normal vector direction is controlled by the fusion weight of the normal vector direction at that point, with smaller weights resulting in narrower broadening and weaker smoothing; the broadening in the tangent vector direction is controlled by a third fusion weight, with larger weights resulting in wider broadening and stronger smoothing. By mapping the fusion weights of the normal and tangent vector directions at different boundary points to the shape parameters of the anisotropic Gaussian smoothing kernel, a customized smoothing kernel adapted to the local heterogeneous ridge conditions can be generated at each boundary point. Within a preset width range on both sides of the shared boundary, the immune cell density of adjacent image blocks is convolved and fused using a locally anisotropic Gaussian smoothing kernel to eliminate block effects and preserve the spatial heterogeneity boundary of immune cells defined by the heterogeneous gradient ridge. It should be noted that the preset width range refers to a strip-shaped region extending a certain number of pixels from the shared boundary to both sides. Convolutional fusion is performed only within this strip-shaped region to avoid unnecessary modifications to the already stable density estimation results within image blocks far from the boundary. Specifically, convolutional fusion replaces the immune cell density value of each pixel within this strip-shaped region with a weighted average of the local anisotropic Gaussian smoothing kernel corresponding to that pixel's location and the original density values ​​in its neighborhood. Since the shape of the anisotropic Gaussian smoothing kernel adaptively changes with the boundary point position, density differences perpendicular to the boundary direction are preserved in boundary segments with heterogeneous gradient ridges; in boundary segments without ridges, smoothing in both directions is fully expanded to eliminate stitching marks. For example, in two... When the shared boundary of adjacent image blocks just crosses the outer boundary of the invasion edge, the heterogeneous gradient ridge corresponding to the outer boundary is approximately perpendicular to the boundary. The normal vector direction of each point on this boundary segment is assigned a low fusion weight, the smoothing kernel is narrower and shorter in the normal vector direction, the density jump across the boundary is preserved, and a clear boundary is maintained between the high-density area of ​​the invasion edge and the low-density matrix area outside. However, when the other shared boundary segment of the adjacent block is located in the tumor center region and is not crossed by any heterogeneous gradient ridge, the fusion weight remains high in both directions, the smoothing kernel is fully expanded in both directions, the inter-block seams are effectively eliminated, and the entire slice of immune cell spatial density distribution map is achieved while eliminating the block effect and perfectly preserving the spatial heterogeneous structure of the tissue's own immune cells.

[0023] In an optional embodiment, the tumor central region and invasion edge region are segmented on the spatial density distribution map according to the histological morphology gradient, and the corresponding local density distribution is extracted, including: On the spatial density distribution map of immune cells in a whole slice, calculate the gradient vector field of immune cell density and generate a density gradient magnitude map; It should be noted that the gradient vector field of immune cell density describes the direction and rate of change of immune cell density in space within the entire slice. At each pixel position on the density distribution map, the first-order change of density value is calculated along the horizontal and vertical directions of the image, respectively. The changes in these two directions are combined into a two-dimensional vector. The direction of this vector represents the direction of the fastest increase in density, and the length of this vector represents the intensity of the density change. The gradient vector length at each pixel is extracted and assigned to that pixel to form a new image, which is the density gradient magnitude map. The brighter the position in the image, the more intense the change in immune cell density at that position. In the density gradient magnitude map, the local maximum point of the gradient magnitude is used as the seed point. The path of the gradient magnitude is tracked along the gradient direction and the gradient magnitude is monotonically decreasing. The tracking is terminated when the gradient magnitude on the path drops to a preset proportion of the gradient magnitude of the seed point. The termination points are connected to form the candidate boundary of the tumor center. The area enclosed by the candidate boundary is taken as the tumor center area. It should be noted that local maxima of gradient magnitude refer to those pixel locations where the gradient magnitude reaches its maximum within a certain local neighborhood and is greater than that of the surrounding pixels. These points are usually located at the steepest point where the density of immune cells transitions from a homogeneous high-density region to a transitional region. Using these maxima points as starting seed points, the algorithm moves outward step by step along the gradient direction at each seed point. During the movement, the gradient magnitude should show a monotonically decreasing trend, indicating that it is moving from the position of the most drastic density change to a region of gentle density change. When the gradient magnitude of a point on the tracking path decreases to a preset proportion of the gradient magnitude of the starting seed point of the path, such as decreasing to half, the tracking is terminated. This termination point is considered to be a position where the density change has become gentle and no longer has obvious transitional characteristics. All termination points from adjacent seed points are connected sequentially according to their spatial proximity to form a closed curve, which is the candidate boundary of the tumor center. The internal region enclosed by this boundary is characterized by high density and low gradient and is identified as the tumor center region. Starting from the candidate boundary of the tumor center, expand outward along the gradient direction, and calculate the immune cell density difference between each pixel and its corresponding pixel back to the candidate boundary along the gradient direction during the expansion process. It should be noted that the purpose of this step is to explore outwards from the candidate boundary at the center of the tumor, using the magnitude of density decrease to determine the termination position of immune cell infiltration. Starting from each point on the candidate boundary, the movement proceeds outwards along the gradient direction at that point. For each external pixel passed during the movement, its corresponding starting pixel on the candidate boundary can be found by tracing back along the opposite direction of the gradient. The density difference value is calculated by subtracting the density value of the corresponding boundary starting pixel from the immune cell density value of the external pixel. This difference value reflects the cumulative decrease in immune cell density from the boundary to that location. Since immune cells in the tumor microenvironment typically decrease gradually from the high-density area at the center to the periphery, the density difference value gradually increases in the invasion edge area until it stabilizes after reaching the surrounding non-tumor tissue. Pixels whose density difference values ​​converge to a preset convergence threshold range are marked as invasion edge termination points. All invasion edge termination points are connected to form the outer boundary of the invasion edge. The annular region between the candidate boundary of the tumor center and the outer boundary of the invasion edge is taken as the invasion edge region. It should be noted that density difference convergence refers to the phenomenon that as the outward expansion distance increases, the change in density difference between adjacent pixels becomes smaller and smaller, and the difference value tends to reach a relatively stable plateau level. The preset convergence threshold range is a very small range of variation. When the difference between the density difference of a pixel and the density difference of the previous pixel falls within this threshold range, and subsequent pixels continuing outward also continuously meet the convergence condition, this pixel is marked as the invasion edge termination point, indicating that the immune cell density has no longer been significantly affected by the tumor microenvironment and has entered the background level from this point. Connecting the invasion edge termination points marked in all directions according to their spatial positions forms an outer boundary surrounding the tumor center region. The annular region between this outer boundary and the candidate boundary of the tumor center is the invasion edge region. Using masks of the tumor center region and the invasion edge region, the local density distribution of the tumor center and the local density distribution of the invasion edge were extracted from the spatial density distribution map of immune cells in the whole slice. It should be noted that the mask for the tumor center region is a binary image of the same size as the density distribution map of the whole slice, with a value of 1 for pixels belonging to the tumor center region and a value of 0 for all other pixels. Similarly, the mask for the invasion edge region has a value of 1 for pixels belonging to the ring-shaped invasion edge region and a value of 0 for all other pixels. By mapping these two masks to the spatial density distribution map of immune cells in the whole slice pixel by pixel and retaining the density values ​​of the pixels with a mask value of 1, the local density distribution of the tumor center and the local density distribution of the invasion edge are obtained.

[0024] In an optional embodiment, using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are assigned as the immune cell density field of the tumor center region and the immune cell density field of the invasion edge region, respectively, and the immune cell density field in each region is set as a quasi-steady-state field. It should be noted that the local density distribution at the tumor center and the local density distribution at the invasion edge are the spatial distribution data of immune cell density in two regions obtained by segmenting and extracting the spatial density distribution map of immune cells from the whole slice in the aforementioned steps. Directly assigning this measured data to the immune cell density field of the corresponding region means that the spatial distribution value of the density field is based on the actual tissue slice observation results, rather than using assumed values ​​or empirical values ​​from literature. Setting the immune cell density field in each region as a quasi-steady-state field means that at the time of tumor tissue sampling, the spatiotemporal distribution of immune cells has reached a dynamic equilibrium state, and the density field no longer changes significantly with time. The rationality of this quasi-steady-state assumption is that the tissue slice records a static snapshot of a certain moment in the process of tumor development. At this time section, the sum of the net inflow and net proliferation of immune cells is approximately 0, and the density distribution shows a stable spatial pattern. Establish the coupling relationship between the immune cell density field and the chemokine concentration field; wherein, the temporal variation of the immune cell density field is given by the immune cell diffusion flux driven by the chemokine concentration gradient and the immune cell self-proliferation flux, and the temporal variation of the chemokine concentration field is given by the immune cell density-dependent chemokine secretion and production and the natural decay of chemokines. It should be noted that the coupling relationship between the immune cell density field and the chemokine concentration field describes the interactive dynamics of two core elements in the tumor microenvironment. On the one hand, the distribution changes of immune cells in tissues are driven by two processes: first, the diffusion process driven by the chemokine concentration gradient, where immune cells migrate directionally along the direction of increasing chemokine concentration, and the migration flux is proportional to the spatial rate of change of chemokine concentration and the effective chemokine coefficient; second, the proliferation process of immune cells themselves, where the increase in local immune cell density is proportional to the current density and the effective proliferation coefficient. On the other hand, the dynamic changes in the chemokine concentration field are also determined by two processes: first, the immune cell density field... Chemokines are produced in two ways: firstly, through cell density-dependent secretion, where they are continuously released at a certain rate by immune cells themselves or by other stromal cells in the tumor microenvironment where immune cells are present; and secondly, through natural decay of chemokines, where they are degraded or inactivated in tissues with a certain half-life. For example, in a typical solid tumor, the presence of a large number of immune cells in the central region of the tumor leads to vigorous secretion and high concentration of chemokines in that region, forming a concentration gradient that gradually decreases from the center outwards. This gradient further guides peripheral immune cells to migrate towards the center, while the density of local immune cells also increases due to proliferation. The chemokine concentration field and the immune cell density field form a positive feedback-driven coupling dynamic. Substituting the quasi-steady-state field as the solution of the coupling relationship when the time variation approaches zero into the coupling relationship, we obtain a reaction-diffusion coupled system with local density distribution as spatial constraint. It should be noted that under quasi-steady-state conditions, both the immune cell density field and the chemokine concentration field no longer change with time; that is, all time-related partial derivatives in the coupling equation are zero. Substituting the aforementioned quasi-steady-state field, which has been assigned the value of the measured local density distribution, into the equation after setting each time term to zero, all spatial distribution terms in the equation are fixed by the known local density distribution values. The remaining undetermined parts are the two undetermined parameters: the effective chemokine coefficient and the effective proliferation coefficient. The resulting reaction-diffusion coupling system is no longer a kinetic equation for forward simulation, but a set of constraint equations with the measured density of real tissue as spatial constraints and the undetermined parameters as unknowns. For example, in a whole-slice analysis of non-small cell lung cancer, after substituting the measured high value of immune cell density in the tumor center region and the measured density gradient value in the invasive edge region as quasi-steady-state fields into the coupling system, the gradient relationship between the spatial distribution of chemokines and the density field within the system is completely determined by the measured data. The only unknowns to be solved in the equation are the effective chemokine coefficient and the effective proliferation coefficient.

[0025] In an optional embodiment, the effective chemotaxis coefficient and effective multiplication coefficient in the reaction-diffusion coupled system are inverted using the variational adjoint method, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are taken as the observation target field, and the immune cell density field output by the reaction-diffusion coupling system under quasi-steady-state conditions is taken as the prediction field. A loss functional is constructed, where the loss functional is the integral of the point-by-point deviation between the prediction field and the observation target field over the entire spatial domain. It should be noted that the observed target field refers to the measured spatial distribution of immune cell density in the tumor center and invasion edge regions obtained through the aforementioned image segmentation and density extraction steps; the predicted field refers to the spatial distribution of immune cell density output by the reaction-diffusion coupling system under quasi-steady-state conditions, given a set of effective chemotaxis coefficients and effective proliferation coefficients; the loss functional is a scalar index that measures the overall deviation between the predicted field and the observed target field. It is constructed by traversing every spatial pixel position in both regions, calculating the difference between the predicted density value and the observed density value at that position, and accumulating the deviations of all positions over the entire spatial region to obtain the total deviation; the smaller the value of the loss functional, the closer the predicted density field output by the coupling system is to the measured density field under the current parameters. Construct the adjoint system corresponding to the reaction-diffusion coupling system, take the variation of the loss functional with respect to the immune cell density field as the driving source term of the adjoint system, solve the adjoint system in reverse time, and obtain the spatiotemporal adjoint variable field; It should be noted that the adjoint system is an auxiliary system that is mathematically dual to the original reaction-diffusion coupled system. Its construction is based on the joint variational derivation of the original physical field equations and the loss functional. The driving source term refers to the excitation source that propels the operation of the adjoint system. In this step, the sensitivity information of the loss functional to the immune cell density field at each spatial location is used as the driving input of the adjoint system. This sensitivity information reflects the contribution weight of the deviation between the predicted density field and the observed target field at a certain point to the overall loss. The inverse time solution means that the solution direction of the adjoint system is opposite to the time progression direction of the forward coupled system, working backward from the terminal moment of the quasi-steady state to the initial moment. After the solution is completed, the values ​​of the adjoint variables at all spatial locations and all inverse time steps are obtained, i.e., the spatiotemporal adjoint variable field. The adjoint variable field can be understood as a sensitivity distribution field, which records how a small perturbation at any spatial location on the immune cell density field at any moment will affect the final loss functional value. In the forward equation of the reaction-diffusion coupled system, the partial derivatives of the chemotaxis flux term with respect to the effective chemotaxis coefficient and the partial derivatives of the proliferation flux term with respect to the effective proliferation coefficient are extracted. The adjoint variable field is then spatially integrated with the partial derivative information to obtain the gradients of the loss functional with respect to the effective chemotaxis coefficient and the effective proliferation coefficient. It should be noted that the chemotactic flux term in the forward equation describes the effect of the spatial displacement of immune cells driven by the chemokine gradient on density changes. The magnitude of this flux term is linearly dependent on the effective chemotactic coefficient. Extracting its partial derivative with respect to the effective chemotactic coefficient yields the response pattern of this flux term to changes in the chemotactic coefficient. The proliferation flux term describes the effect of local proliferation of immune cells on density changes. Its magnitude is also linearly dependent on the effective proliferation coefficient. Extracting its partial derivative with respect to the effective proliferation coefficient yields the response pattern of this flux term to changes in the proliferation coefficient. The accompanying variable field obtained from the above solution is then multiplied point-by-point with these two sets of partial derivative information at all spatial locations and all time steps, and then summed. This yields two scalar values, corresponding to the gradient of the loss functional with respect to the effective chemotaxis coefficient and the effective multiplication coefficient, respectively, under the current parameter values. The gradient indicates the direction and step size in which the parameters should be adjusted to decrease the loss functional. Compared to gradient estimation methods that trial and error the parameters one by one, such as finite difference methods, the adjoint method only needs to solve the coupled system once in the forward direction and the adjoint system once in the reverse direction to obtain the accurate gradients of both parameters simultaneously, reducing computational costs. For example, if the gradient value with respect to the effective chemotaxis coefficient obtained after one adjoint solution is positive and large, it indicates that the current chemotaxis coefficient is too high, resulting in an excessively strong chemotaxis-driven migration effect in the predicted density field. The chemotaxis coefficient should be reduced in the opposite direction of the gradient to reduce the loss functional. The gradient-driven gradient descent optimization process iteratively updates the effective chemotaxis coefficient and the effective multiplication coefficient. When the value of the loss functional drops below the preset convergence threshold, the iteration stops and the current effective chemotaxis coefficient and effective multiplication coefficient are output as the inversion result. It should be noted that the gradient descent optimization process refers to an iterative strategy that uses the gradient information calculated at each step to successively correct the parameter values ​​in the opposite direction of the gradient. In each iteration, the current chemotaxis coefficient and proliferation coefficient are first used to solve the coupled system in the forward direction to obtain the predicted density field and calculate the loss functional value. Then, the adjoint system is constructed and solved in the reverse direction to obtain the adjoint variable field. Next, the gradient is calculated using the spatial integration of the adjoint variable field and the partial derivative information. Finally, the parameter values ​​are updated along the gradient descent direction with a preset step size. As the iteration progresses, the loss functional value gradually decreases, indicating that the predicted density field is getting closer and closer to the observed target field. The preset convergence threshold is the threshold value for determining whether the iteration should terminate. When the loss functional value drops below this threshold, it indicates that the deviation between the predicted density and the measured density is small enough that continuing the iteration will have negligible improvement on the parameters, and the iteration terminates at this time. The effective chemotaxis coefficient and effective proliferation coefficient obtained in the final iteration are output as the inversion result.

[0026] In an optional embodiment, the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion are differiated from the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples to generate a parameter deviation vector, including: A set of non-tumor reference samples with the same tissue type as the immunohistochemically stained whole-slice images to be scored was obtained. For the whole-slice images of each sample in the non-tumor reference sample set, the same image block division, multi-layer spectral separation and decomposition, spatial density distribution map construction, region segmentation, local density extraction, reaction-diffusion coupling system construction and variational adjoint inversion were used as for the samples to be scored to obtain the effective chemotaxis coefficient and effective proliferation coefficient of each sample. It should be noted that the non-tumor reference sample set refers to a group of control samples with the same tissue origin but confirmed by pathology to be free of tumorous changes. For example, normal tissue samples taken from the same organ or normal tissue samples next to benign lesions. These samples are processed using the exact same procedure as the samples to be scored. This means that every step from image patch segmentation to variational inversion maintains consistent processing parameters and algorithm configurations. The purpose is to ensure that the effective chemotaxis coefficient and effective proliferation coefficient output by the reference samples are in the same physical dimension and quantization framework as the samples to be scored, eliminating systematic bias introduced by differences in processing procedures. Each sample will generate a pair of effective chemotaxis coefficients and effective proliferation coefficients after processing. The coefficients of all samples are collected to form the chemotaxis coefficient set and the proliferation coefficient set, respectively. Robust statistical estimations are performed on the chemotaxis coefficient set and the proliferation coefficient set respectively, and the statistical center value of the chemotaxis coefficient set is used as the benchmark chemotaxis coefficient, and the statistical center value of the proliferation coefficient set is used as the benchmark proliferation coefficient. It should be noted that robust statistical estimation refers to the use of statistical methods that are insensitive to extreme outliers in the data distribution when estimating the central tendency of a set of data. This ensures that the estimation results can reflect the typical level of most normal samples without being affected by a few samples with particularly large or small deviations. Robust estimators such as the median or truncated mean can be used. The central values ​​obtained by robust statistical estimation of the chemotaxis coefficient set and the proliferation coefficient set are defined as the benchmark chemotaxis coefficient and the benchmark proliferation coefficient, respectively. These two benchmark values ​​represent the reference level of the chemotactic and proliferative capacity of immune cells under normal conditions for this tissue type. The effective chemotaxis coefficient of the sample to be scored obtained by subtracting the baseline chemotaxis coefficient is used to obtain the chemotaxis coefficient deviation; the effective proliferation coefficient of the sample to be scored is subtracted from the baseline proliferation coefficient to obtain the proliferation coefficient deviation. It should be noted that the subtraction operation directly calculates the difference between the individual mechanism parameters of the sample to be scored and the baseline parameters of the normal sample of the same tissue type. The sign and magnitude of the difference reflect the direction and magnitude of the deviation of the individual tumor microenvironment from the normal tissue in terms of immune cell migration ability and local proliferation ability. A positive deviation of the chemotaxis coefficient indicates that the chemotactic migration ability of immune cells in the tumor tissue is stronger than that in the normal tissue, while a negative value indicates that it is weaker than that in the normal tissue; the same applies to the deviation of the proliferation coefficient. The deviations of the chemotaxis coefficient and the deviations of the proliferation coefficient are arranged into a two-dimensional vector, which serves as the parameter deviation vector. It should be noted that organizing the two deviations into an ordered two-dimensional structure, with each dimension representing the degree of deviation of a type of mechanism parameter, makes the parameter deviation vector a geometric object with a clear direction and length in the two-dimensional mechanism parameter deviation space. The length of the vector reflects the combined magnitude of the deviations of the two mechanism parameters, and the direction of the vector reflects the relative contribution ratio of chemotaxis deviation and proliferation deviation.

[0027] In an optional embodiment, an initial score is obtained by performing an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score, including: Obtain a predetermined immune score projection vector; wherein, the immune score projection vector is a two-dimensional vector, the two components of which correspond to the chemotaxis deviation dimension and the proliferation deviation dimension, respectively. The two-dimensional vector is obtained by fitting the clinical immune score to the chemotaxis coefficient deviation and the proliferation coefficient deviation of the reference population. It should be noted that the immune score projection vector is a fixed two-dimensional weighted vector pre-determined through fitting with a large sample of reference population data. Its function is to define the projection direction from the two-dimensional parameter deviation space to the one-dimensional scoring axis. The construction method of this projection vector is as follows: a set of reference population samples with known clinical immune scores are collected, and the aforementioned processing procedures are performed on each sample to obtain its chemotaxis coefficient deviation and proliferation coefficient deviation. Then, the two-dimensional deviation is used as the independent variable and the known clinical immune score is used as the target dependent variable for fitting. A set of weight coefficients is found that makes the fitted score best match the known clinical score. This set of weight coefficients is arranged in the order of chemotaxis deviation dimension and proliferation deviation dimension, which is the immune score projection vector. The ratio of the two components in this projection vector reflects the relative importance of chemotaxis capacity deviation and proliferation capacity deviation in the comprehensive immune status assessment. The initial score is obtained by performing an inner product operation between the parameter deviation vector and the immune score projection vector. It should be noted that the inner product operation refers to multiplying the two components of the parameter deviation vector with the corresponding components of the immune score projection vector, and then adding the two products to obtain a scalar value. The geometric meaning of this operation is to project the parameter deviation vector onto the direction defined by the immune score projection vector, and the length of the projection is the initial score. The initial score integrates information from two dimensions: chemotaxis deviation and proliferation deviation. The greater the deviation and the more consistent the deviation direction is with the projection direction, the higher the initial score. Obtain the initial score set of the preset reference population, determine the quantile of the initial score in the initial score set, map the quantile to the preset percentage score scale, and obtain the final immune score; It should be noted that the initial score set of the preset reference population refers to the score data pool formed after calculating the initial score of each individual sample in the above reference population according to the same process; determining the quantile of the initial score of the sample to be scored in this set means calculating the percentile position of the initial score in the initial score of the reference population in ascending order. This operation transforms the absolute value of the individual score into its relative position in the population, eliminating the numerical scale differences that may be introduced by different batches, different populations or different experimental conditions; mapping the quantile to the preset percentage score scale directly converts the quantile into a percentage score that is easy to understand and compare. Finally, the immune score characterizes the relative level of the deviation of the individual's tumor microenvironment immune cell chemotaxis and proliferation capacity from the non-tumor reference in the same type of sample population.

[0028] Example 2, please refer to Figure 2 This invention provides a technical solution: a deep learning-based immune scoring data processing system, applicable to the aforementioned deep learning-based immune scoring data processing method, comprising: Image segmentation processing module 1 is used to acquire the immunohistochemical staining whole slide image to be scored, divide the whole slide image into non-overlapping image blocks, perform multi-layer spectral separation decomposition on each image block, and construct a spatial density distribution map of immune cells in the whole slide. Region segmentation and extraction module 2 is used to segment the tumor central region and invasion edge region on the spatial density distribution map according to the histological morphology gradient and extract the corresponding local density distribution. Mechanism parameter inversion module 3 is used to construct a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration using local density distribution as a quasi-steady-state observation constraint, and to invert the effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system through the variational adjoint method. The parameter deviation generation module 4 is used to generate a parameter deviation vector by differencing the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples. The immune score output module 5 is used to perform an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector to obtain an initial score. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of a non-tumor reference.

[0029] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A deep learning-based method for processing immune scoring data, characterized in that, include: Obtain the immunohistochemical stained whole slide image to be scored, and divide the whole slide image into non-overlapping image blocks; Multi-level spectral separation and decomposition were performed on each image patch to construct a spatial density distribution map of immune cells in the whole slice. On the spatial density distribution map, the tumor central region and the invasion edge region are segmented according to the histological morphology gradient, and the corresponding local density distribution is extracted. Using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed, and the effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system are inverted by the variational adjoint method. The effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion are compared with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained from non-tumor reference samples to generate a parameter deviation vector; An initial score is obtained by performing an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of a non-tumor reference.

2. The method for processing immune scoring data based on deep learning according to claim 1, characterized in that, Obtain the immunohistochemically stained whole-section image to be scored, and divide the whole-section image into non-overlapping image blocks, including: Acquire multispectral immunohistochemical stained whole-section images, extract spectral data of the multispectral images under each preset narrowband filter channel, and form a multichannel spectral image cube; The first linear spectral unmixing is performed on the multi-channel spectral image cube. The initial abundance map of each immune-related marker is decomposed using the pre-determined chromatin matrix. All initial abundance maps are superimposed by channel to generate an initial distribution characterization map of immune cells. Calculate the pixel-by-pixel local entropy of the initial distribution characterization map of immune cells, and generate a heterogeneity index map characterizing the degree of spatial heterogeneity of immune cells. Multi-scale Gaussian difference filtering is applied to the heterogeneity index map to extract heterogeneous gradient ridges. The heterogeneous gradient ridges are used as constraint boundaries, and the whole slice image is divided into non-overlapping tissue micro-regions using marker watershed segmentation. For each tissue microregion, a quadtree recursively split is performed according to the preset minimum analysis unit size until the standard deviation of the heterogeneity index in the split subregion is lower than the set threshold. All the subregions obtained by the final split are used as non-overlapping image blocks.

3. The method for processing immune scoring data based on deep learning according to claim 2, characterized in that, Multi-level spectral separation and decomposition were performed on each image patch to construct a spatial density distribution map of immune cells in the whole slice, including: Obtain a pre-constructed hierarchical immunostaining spectral library; wherein the hierarchical immunostaining spectral library contains a first layer of coarse separation spectral matrix and at least one subsequent layer of fine separation spectral matrix, and the coarse separation spectral matrix is ​​used to separate the immunomarker signal from the tissue autofluorescence background, and the fine separation spectral matrix is ​​used to distinguish the specific spectra of different immune cell subsets; For each image block, the spectral data within the image block are subjected to supervised non-negative least squares unmixing using the first-layer coarse separation spectral basis matrix to obtain the overall abundance map of immunomarkers after removing the autofluorescence background. Using the overall abundance map of immunomarkers as a spatial prior, under the subsequent layer-by-layer separation of the spectral basis matrix, a spatially weighted non-negative matrix decomposition is performed. The decomposition coefficients are applied with a local smoothing regularization term based on the overall abundance map of immunomarkers, and the abundance maps of each preset immune cell subpopulation are separated layer by layer. The abundance maps of each subpopulation output in the last layer are used as the immune cell density distribution of the image patch. After traversing all image blocks and obtaining the immune cell density distribution of each image block, the immune cell density on both sides of the boundary of adjacent image blocks is directionally weighted and smoothed to construct a spatial density distribution map of immune cells in the whole slice.

4. The method for processing immune scoring data based on deep learning according to claim 3, characterized in that, Using heterogeneous gradient ridges as guiding boundaries, directional weighted smoothing fusion of immune cell densities on both sides of adjacent image patch boundaries is performed, including: For any two adjacent image patches, extract their shared boundary and calculate the boundary normal vector and tangent vector point by point along the shared boundary; At each boundary point of the shared boundary, the direction of the heterogeneous gradient ridge is obtained in a preset local region centered on that boundary point, and the angle between the direction of the heterogeneous gradient ridge and the boundary normal vector is calculated. When the included angle is less than a preset included angle threshold, a first fusion weight along the normal vector direction is assigned to the boundary point; otherwise, a second fusion weight is assigned. The first fusion weight is less than the second fusion weight. For each boundary point of the shared boundary, a third fusion weight is set along the tangent vector direction; wherein the third fusion weight is greater than or equal to the second fusion weight; A local anisotropic Gaussian smoothing kernel is constructed by using the fusion weights of each boundary point along the normal vector direction and the fusion weights along the tangent vector direction. Within a preset width range on both sides of the shared boundary, the immune cell density of adjacent image blocks is convolved and fused using a locally anisotropic Gaussian smoothing kernel to eliminate block effects and preserve the spatial heterogeneity boundary of immune cells defined by heterogeneous gradient ridges.

5. The deep learning-based immune scoring data processing method according to claim 4, characterized in that, On the spatial density distribution map, the tumor central region and the invasion edge region are segmented according to the histological morphology gradient, and the corresponding local density distribution is extracted, including: On the spatial density distribution map of immune cells in a whole slice, calculate the gradient vector field of immune cell density and generate a density gradient magnitude map; In the density gradient magnitude map, the local maximum point of the gradient magnitude is used as the seed point. The path of the gradient magnitude is tracked along the gradient direction and the gradient magnitude is monotonically decreasing. The tracking is terminated when the gradient magnitude on the path drops to a preset proportion of the gradient magnitude of the seed point. The termination points are connected to form the candidate boundary of the tumor center. The area enclosed by the candidate boundary is taken as the tumor center area. Starting from the candidate boundary of the tumor center, expand outward along the gradient direction, and calculate the immune cell density difference between each pixel and its corresponding pixel back to the candidate boundary along the gradient direction during the expansion process. Pixels whose density difference values ​​converge to a preset convergence threshold range are marked as invasion edge termination points. All invasion edge termination points are connected to form the outer boundary of the invasion edge. The annular region between the candidate boundary of the tumor center and the outer boundary of the invasion edge is taken as the invasion edge region. Using masks of the tumor center and invasion edge regions, the local density distribution of immune cells in the whole slice was extracted from the spatial density distribution map of immune cells.

6. The method for processing immune scoring data based on deep learning according to claim 5, characterized in that, Using local density distribution as a quasi-steady-state observation constraint, a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration is constructed, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are assigned as the immune cell density field of the tumor center region and the immune cell density field of the invasion edge region, respectively, and the immune cell density field in each region is set as a quasi-steady-state field. Establish the coupling relationship between the immune cell density field and the chemokine concentration field; wherein, the temporal variation of the immune cell density field is given by the immune cell diffusion flux driven by the chemokine concentration gradient and the immune cell self-proliferation flux, and the temporal variation of the chemokine concentration field is given by the immune cell density-dependent chemokine secretion and production and the natural decay of chemokines. Substituting the quasi-steady-state field as the solution of the coupling relationship when the time variation approaches zero into the coupling relationship, we obtain a reaction-diffusion coupled system with local density distribution as the spatial constraint.

7. The method for processing immune scoring data based on deep learning according to claim 6, characterized in that, The effective chemotaxis coefficient and effective multiplication coefficient in a reaction-diffusion coupled system are inverted using the variational adjoint method, including: The local density distribution at the tumor center and the local density distribution at the invasion edge are taken as the observation target field, and the immune cell density field output by the reaction-diffusion coupling system under quasi-steady-state conditions is taken as the prediction field. A loss functional is constructed, where the loss functional is the integral of the point-by-point deviation between the prediction field and the observation target field over the entire spatial domain. Construct the adjoint system corresponding to the reaction-diffusion coupling system, take the variation of the loss functional with respect to the immune cell density field as the driving source term of the adjoint system, solve the adjoint system in reverse time, and obtain the spatiotemporal adjoint variable field; In the forward equation of the reaction-diffusion coupled system, the partial derivatives of the chemotaxis flux term with respect to the effective chemotaxis coefficient and the partial derivatives of the proliferation flux term with respect to the effective proliferation coefficient are extracted. The adjoint variable field is then spatially integrated with the partial derivative information to obtain the gradients of the loss functional with respect to the effective chemotaxis coefficient and the effective proliferation coefficient. The gradient-driven gradient descent optimization process iteratively updates the effective chemotaxis coefficient and the effective multiplication coefficient. When the value of the loss functional drops below the preset convergence threshold, the iteration stops, and the current effective chemotaxis coefficient and the effective multiplication coefficient are output as the inversion result.

8. The method for processing immune scoring data based on deep learning according to claim 7, characterized in that, The effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion are differiated from the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples to generate a parameter deviation vector, including: A set of non-tumor reference samples with the same tissue type as the immunohistochemically stained whole-slice images to be scored was obtained. For the whole-slice images of each sample in the non-tumor reference sample set, the same image block division, multi-layer spectral separation and decomposition, spatial density distribution map construction, region segmentation, local density extraction, reaction-diffusion coupling system construction and variational adjoint inversion were used as for the samples to be scored to obtain the effective chemotaxis coefficient and effective proliferation coefficient of each sample. Robust statistical estimations are performed on the chemotaxis coefficient set and the proliferation coefficient set respectively, and the statistical center value of the chemotaxis coefficient set is used as the benchmark chemotaxis coefficient, and the statistical center value of the proliferation coefficient set is used as the benchmark proliferation coefficient. The effective chemotaxis coefficient of the sample to be scored obtained by subtracting the baseline chemotaxis coefficient is used to obtain the chemotaxis coefficient deviation; the effective proliferation coefficient of the sample to be scored is subtracted from the baseline proliferation coefficient to obtain the proliferation coefficient deviation. The deviations of the chemotaxis coefficient and the deviations of the proliferation coefficient are arranged into a two-dimensional vector, which serves as the parameter deviation vector.

9. The method for processing immune scoring data based on deep learning according to claim 8, characterized in that, An initial score is obtained by performing an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector. This initial score is then standardized to percentiles based on a pre-defined reference population to obtain the final immune score, which includes: Obtain a predetermined immune score projection vector; wherein, the immune score projection vector is a two-dimensional vector, the two components of which correspond to the chemotaxis deviation dimension and the proliferation deviation dimension, respectively. The two-dimensional vector is obtained by fitting the clinical immune score to the chemotaxis coefficient deviation and the proliferation coefficient deviation of the reference population. The initial score is obtained by performing an inner product operation between the parameter deviation vector and the immune score projection vector. Obtain the initial score set of the preset reference population, determine the quantile of the initial score in the initial score set, map the quantile to the preset percentage scoring scale, and obtain the final immune score.

10. A deep learning-based immune scoring data processing system, applicable to the deep learning-based immune scoring data processing method according to any one of claims 1-9, characterized in that, include: The image segmentation processing module is used to acquire the immunohistochemical staining whole slide image to be scored and divide the whole slide image into non-overlapping image blocks. Multi-level spectral separation and decomposition were performed on each image patch to construct a spatial density distribution map of immune cells in the whole slice. The region segmentation and extraction module is used to segment the tumor central region and the invasion edge region on the spatial density distribution map according to the histological morphology gradient, and extract the corresponding local density distribution. The mechanism parameter inversion module is used to construct a reaction-diffusion coupling system describing the interaction between immune cell density and chemokine concentration, using local density distribution as a quasi-steady-state observation constraint. It then uses the variational adjoint method to invert the effective chemokine coefficient and effective proliferation coefficient in the reaction-diffusion coupling system. The parameter deviation generation module is used to differiate the effective chemotaxis coefficient and effective proliferation coefficient obtained by inversion with the baseline chemotaxis coefficient and baseline proliferation coefficient obtained statistically from non-tumor reference samples to generate a parameter deviation vector; The immune score output module is used to perform an inner product operation on the parameter deviation vector using a pre-determined immune score projection vector to obtain an initial score. The initial score is then standardized based on percentiles of a pre-defined reference population to obtain the final immune score. The immune score is used to characterize the degree of deviation of the tumor microenvironment immune cell chemotaxis and proliferation capacity of the sample to be scored from that of a non-tumor reference.

Images