Method for scoring pathological microenvironment of neuroblastoma based on spatial omics prior and double teacher distillation

Abstract

The application provides a scoring method for scoring the pathological microenvironment of neuroblastoma based on spatial omics prior and double-teacher distillation, comprising the following steps: first step: multi-sample single-cell transcriptome integration and cell deconstruction based on iNMF algorithm; second step: construction and cell deconstruction of multi-modal spatial omics dataset; third step: spatial niche definition based on sliding window sampling; fourth step: high-precision preprocessing and quality control of pathological images; fifth step: double-teacher feature library precalculation and expert head preheating; sixth step: cross-architecture knowledge distillation and 5-fold robust model training; seventh step: whole film integration reasoning and clinical prognosis evaluation. The method simultaneously fuses the "expert head" preheating through the double-teacher model distillation, and "injects" the global view into the local extraction capability of ResNet50, so that the lightweight model has the "overall situation" to understand the complex spatial niche, and the prognosis risk assessment accuracy is higher.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of medical artificial intelligence and bioinformatics, and in particular to a scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation. Specifically, it is a method and system that uses deep learning to quantitatively predict spatial niche (NICHE) scores defined by spatial transcriptomics from conventional pathological (HE) images for prognostic assessment. Background Technology

[0002] Spatial transcriptomics (ST) currently has the following limitations: although ST technology can provide high-dimensional molecular spatial information, it is expensive and has extremely high sample requirements, making it impossible to widely adopt in clinical practice.

[0003] The shortcomings of existing pathological evaluation are that clinical pathological diagnosis (such as neuroblastoma classification) relies on the doctor's visual assessment of the proportion of stroma (such as Schwann cells and stromal cells), which has problems such as strong subjectivity, difficulty in quantification, and insensitivity to weak molecular signals.

[0004] Current computational pathology faces the following challenges: training molecular-level regression models directly on HE images often encounters technical difficulties such as small data volume (on the scale of patients), high annotation noise, and non-convergence of cross-architecture model training. Summary of the Invention

[0005] In view of the above, it is necessary to provide a scoring method and system for neuroblastoma pathological microenvironment scoring based on spatial omics priors and dual-teacher distillation. This method and system adopt dual-teacher model distillation while integrating "expert head" warm-up and "injecting" the global perspective into the local extraction capability of ResNet50, so that the lightweight model has a "big picture" of understanding complex spatial niches (NICHE), thereby making the prognostic risk assessment more accurate.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A scoring method for the pathological microenvironment of neuroblastoma based on spatial omics priors and dual-teacher distillation includes the following steps:

[0008] Step 1: Multi-sample single-cell transcriptome integration and cell deconstruction based on the iNMF algorithm:

[0009] We collected neuroblastoma tissue samples and constructed a single-cell dataset. We then used the iNMF algorithm to perform deep integration of the single-cell dataset, achieving cross-sample heterogeneity integration. We performed cluster analysis in the integrated feature space and defined a precise cell type dictionary based on marker genes.

[0010] Step 2: Construction of multimodal spatial omics dataset and cell deconstruction:

[0011] Visium HD high-resolution spatial transcriptome sequencing was performed on the matched NB samples from the neuroblastoma tissue samples collected in the first step to obtain spatial expression data and HE staining for pathological sections. The RCTD algorithm was used to deconvolve the spatial spots of Visium HD with the cell type dictionary obtained in the first step as reference data to obtain a spatial coordinate table of cell types.

[0012] Step 3: Spatial niche definition based on sliding window sampling:

[0013] A sliding window sampling strategy was used on the spatial coordinate table of cell types obtained in the second step to obtain the cell proportion matrix of each virtual spot. For the cell proportion matrix of each virtual spot, a non-negative matrix factorization algorithm was performed to decompose it into a NICHE contribution matrix H. Based on the NICHE contribution matrix H, k types of tumor niche NICHE and k-dimensional continuous NICHE scores corresponding to each coordinate point on each HE slice were obtained. Based on the tumor niche and the ID of the virtual spot, the spatial ecotype matrix of the virtual spot was obtained. At the same time, based on the tumor niche and cell type, the cell type weight matrix of the tumor microenvironment ecotype was obtained.

[0014] Step 4: High-precision preprocessing and quality control of pathological images:

[0015] Using the coordinates in the cell proportion matrix of the virtual spots obtained in step 3, the whole-slice HE image is cut by sliding window to generate virtual spot image patches corresponding to the spatial coordinates of the physical size of the virtual spots; the scores of the k tumor microenvironment ecotypes in the virtual spot spatial ecotype matrix are mapped to the virtual spot image patches to obtain the ecotype score of each virtual spot patch.

[0016] Step 5: Pre-calculation of the dual-teacher feature database and warm-up of expert heads:

[0017] 5.1 A dual-teacher model consisting of Teacher A and Teacher B is used to extract features from the virtual blob image patches obtained in step four, and a feature library of virtual blob image patches is pre-calculated and generated; among them, Teacher A is used to extract pathological features, and Teacher B is used to extract geometric structural features.

[0018] 5.2. Perform morphological predictability detection analysis to achieve expert head preheating. The specific steps are as follows: completely fix the parameters of the dual-teacher model, and map the image patches in the feature library generated in step 5.1 into a high-dimensional feature vector space; train a multilayer perceptron regressor in this high-dimensional feature vector space, and use 5-fold cross-validation to score and train the k types of NICHE obtained in step 3; calculate the theoretical prediction upper limit of each NICHE, and based on the "consensus report", select several indicators with significant morphological signals, and then use them for expert head training to obtain the preheated "expert head";

[0019] 5.3 The preheated "expert head" uses a prefix stripping algorithm to solve the key name mismatch problem in independent and ensemble training of the multilayer perceptron, and then uses a three-in-one collaborative training method. Here, the three-in-one collaborative training refers to the application of task constraints, semantic constraints, and structural constraints, and the use of a hierarchical learning rate strategy to explore the backbone network and perform life-saving fine-tuning of the regression head.

[0020] Step 6: Cross-architecture knowledge distillation and 5-fold robust model training:

[0021] 6.1 Student Model Architecture Design and Weight Implantation: A lightweight convolutional neural network was selected as the backbone network of the student model, and two independent linear projections were constructed in parallel after the global average pooling layer of the backbone network to map the features to the pathological semantic space and the geometric structure space; the optimal MLP regression head weights obtained by training in step 5 were loaded into the output of the student model.

[0022] 6.2 Stratified Ranking and 5-Fold Cross-Validation Based on Patient Heterogeneity: The cumulative value of k NICHE scores for each patient is calculated to obtain a comprehensive index reflecting the complexity of the tumor microenvironment. All patients are then ranked in descending order based on this comprehensive index. Patients are cyclically assigned to 5 Folds according to the ranking order. During the assignment process, group constraints are used to strictly lock the patient ID to ensure that all patches from the same source can only exist in the same set.

[0023] 6.3 Weak signal index directional rescue mechanism is adopted to enhance the weak signals or sparsely distributed indices found in the third step of NMF decomposition;

[0024] 6.4, obtain the parameter model of the optimal student model trained on each fold of data;

[0025] Step 7: Integrated Reasoning and Clinical Prognostic Assessment of the Entire Film:

[0026] 7.1 The image to be predicted is subjected to normalization preprocessing. The specific steps of preprocessing include physical scale alignment, high-overlap sliding window sampling, and coordinate locking naming method.

[0027] 7.2 A multi-model ensemble voting inference mechanism consisting of a 5-fold ensemble learning strategy is used for prediction: the parameter models of the 5 best student models obtained in step 6 are loaded in parallel, and each image patch after preprocessing in step 7.1 is used as input to obtain the 5 model prediction values ​​for each image patch for each NICHE index, and the arithmetic mean of the 5 model prediction values ​​is used as the final NICHE score Sn of the image patch; in addition, the standard deviation of the 5 model prediction values ​​is calculated simultaneously, and the magnitude of the standard deviation value is used as the confidence index of the system prediction.

[0028] 7.3 Spatial Heatmap Reconstruction and Clinical Indicator Conversion: Using the center point coordinate information in the image block file name contained in the coordinate locking naming method during preprocessing in step 7.1, the NICHE score Sn calculated in step 7.2 is mapped back to a two-dimensional matrix corresponding to the proportion of the original image, resulting in a reconstructed score matrix. A Gaussian filtering algorithm is applied to the reconstructed score matrix to generate a visually continuous spatial distribution heatmap. By setting a score threshold for each indicator, the area of ​​high-scoring regions exceeding the score threshold in the spatial distribution heatmap is calculated to obtain the area proportion of different NICHE indicators in the spatial distribution heatmap. Based on the area proportion, the spatial overlap, exclusion, or inclusion relationships of different NICHE indicators are analyzed to define quantitative clinical indicators reflecting the complexity of the tumor microenvironment.

[0029] 7.4 Based on the quantitative clinical indicators obtained in step 7.3, the evaluation results are derived.

[0030] Preferably, in the first step, the single-cell dataset is obtained by performing single-cell transcriptome sequencing on collected neuroblastoma tissue samples.

[0031] Preferably, in the first step, when integrating across sample heterogeneity, the iNMF algorithm decomposes the single-cell dataset into shared factors and sample-specific factors.

[0032] Preferably, in the first step, the cell type dictionary includes transcriptome matrix data for each single cell, i.e., cell ID * gene name, and the type name corresponding to each cell ID, i.e., cell type metadata.

[0033] Preferably, in the second step, the spatial coordinate table of cell types includes the x-axis, y-axis and the corresponding cell types, which is an annotation of the same cell types on the spatial transcriptome data as in the first step.

[0034] Preferably, in the third step, the cell proportion matrix of the virtual spots is obtained by calculating the average percentage of various cell types in the window set by the sliding window sampling strategy.

[0035] Preferably, in the third step, when performing a nonnegative matrix factorization algorithm on the cell proportion matrix of the virtual spots, the value of k is obtained by the following method: determining the search interval for the value of k, performing traversal tests within the search interval, and using evaluation indicators for comprehensive evaluation to obtain the optimal value of k; the above evaluation indicators include the co-table correlation coefficient, the sum of squared residuals and reconstruction error, and the dispersion coefficient.

[0036] Preferably, in the fourth step, the resulting virtual speckle image patch is filtered after the sliding window is cut.

[0037] Preferably, in step 5.1, Teacher A model is a pathological semantic model, which uses the UNI basic model to extract pathological features; Teacher B model is a spatial structure model, which uses the DINOv2 model to extract geometric structure features.

[0038] Preferably, in step 6.3, the weak signal indicator targeted rescue mechanism is a dual enhancement strategy of "sampling-loss", which includes weighted random sampling and multi-task collaborative loss function optimization;

[0039] The specific steps of weighted random sampling are as follows: before the start of each training iteration Epoch, scan the label vectors of all Pats in the training set; then, define a "rare sample detection function"; if the score of the weak signal or sparse distribution index of a Pat exceeds a preset threshold, increase the sampling weight of the sample to 10 times that of ordinary samples, otherwise retain the original sample.

[0040] The specific steps for optimizing the multi-task collaborative loss function are as follows: First, joint gradient descent is performed, that is, a synchronous optimization approach is adopted, in which L is calculated simultaneously in each mini-batch. task L uni With L dino The numerical value is obtained, and the total loss function L is calculated using the chain rule. total The gradient of the student model parameters; where L total =α⋅L task +β1⋅L uni +β2⋅L dino In the formula, α is the task loss weight, β1 and β2 are both distillation weights, and L task It is the weighted task loss, L uni It is a semantic consistency constraint, L dino It is a structural consistency constraint;

[0041] Secondly, a hierarchical dynamic step size adjustment is implemented. Specifically, the AdamW optimizer is used in conjunction with a cosine annealing learning rate scheduling algorithm. In the early stages of training, a large distillation weight forces the student model to align features. In the later stages of training, as the learning rate decreases, the model enters a fine-tuning phase to ensure L... taskThe indicators have reached their optimal levels.

[0042] Next, gradient clipping technology is used in L task When weak signal indicators generate abnormal gradients due to oversampling, the global gradient magnitude is truncated to ensure the numerical stability of multi-task collaborative training.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. Most models rely on "manual annotation" by pathologists, which is limited by human visual resolution and subjective experience, and cannot reflect underlying molecular expression. This invention, however, employs Visium HD (2μm level) spatial transcriptomics technology as the "source of truth," combined with RCTD deconvolution, to directly advance label precision to the cell-scale level. Furthermore, it innovatively proposes a sliding window sampling strategy with a 500μm window and a 100μm stride. This not only simulates the physical capture range of spatial transcriptomics but also solves the fault tolerance problem of physical coordinate alignment between HE images and omics blobules through "high overlap sampling," constructing an extremely robust "morphomic" correspondence matrix.

[0045] 2. This invention achieves cross-architecture distillation from Transformer (ViT) to Convolution (CNN) in the field of neuroblastoma. This approach "injects" the global perspective of ViT into the local extraction capabilities of ResNet50, enabling the lightweight model to have a "big picture" understanding complex spatial niches (NICHE).

[0046] 3. This invention uses an "expert head warm-up" step to explore the "predictability upper limit" of 16 indicators in morphology before formal training. This strategy allows for the scientific elimination of noisy indicators based on consensus reports and targeted "rescue" of weak signal indicators such as N11. This engineering logic of "first finding the upper limit and then fine-tuning" greatly improves the convergence efficiency of the model with a small sample size (57 cases).

[0047] 4. This invention employs stratified sorting, meaning that, addressing the extremely uneven distribution of NICHE scores, stratified sampling is used after sorting patients by their total scores, ensuring that each fold in the 5-fold validation has complete representativeness from "extremely high scores" to "typically low scores." Secondly, the invention utilizes a weak signal rescue strategy, combining Weighted Random Sampler and a weighted task loss function. This combination successfully overcomes the common problem of high-score sample sparsity in medical images, improving the predictive efficiency of weak signal indicators such as N11 by more than 250% (from...). Figure 3 (This can be derived from...)

[0048] 5. In existing technologies, models can only replicate doctors' existing classifications (such as Gleason or Shimada). This invention, through an AI model, successfully reconstructs the expensive NICHE_12 (Schwann cell niche) distribution on inexpensive, ordinary HE slides. Spatial transcriptome sequencing is very expensive, costing around 30,000 to 50,000 RMB per sample. Experiments have shown that this digital score is significantly correlated with patient survival (p=0.0024).

[0049] 6. The hardware required for this invention is also very inexpensive, requiring only a Tesla T4 16G graphics card to complete training and prediction; moreover, this invention compresses the "wisdom" of UNI and DINOv2 into a lightweight ResNet50 backbone network through cross-architecture (Transformer to CNN) knowledge distillation. Attached Figure Description

[0050] Figure 1 It is a table containing the final NICHE score Sn and standard deviation and variance obtained through step 7.2 of this invention.

[0051] Figure 2 It's a survival curve.

[0052] Figure 3 The table summarizes the signal performance before and after signal rescue. In the table, Mean_R2 is the R2 value before rescue and Std_R2 is the R2 value after rescue. As can be seen from the table, before signal rescue, the R2 of N11 was only 0.098, but after signal rescue, it reached 0.250.

[0053] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0054] In a preferred embodiment of the present invention, the scoring method for the pathological microenvironment of neuroblastoma based on spatial omics priors and dual-teacher distillation includes the following steps:

[0055] Step 1: Multi-sample single-cell transcriptome integration and cell deconstruction based on the iNMF algorithm (Integrated Nonnegative Matrix Factorization Algorithm):

[0056] Neuroblastoma tissue samples were collected, and single-cell transcriptome sequencing was performed on the collected neuroblastoma tissue samples to construct a single-cell dataset. The iNMF algorithm was used to deeply integrate the single-cell dataset to achieve cross-sample heterogeneity integration. During integration, the iNMF algorithm decomposed the single-cell gene expression matrix into shared factors and sample-specific factors. Cluster analysis was performed in the integrated feature space, and a precise cell type dictionary was defined based on the marker genes. The cell type dictionary includes the transcriptome matrix data of each single cell, i.e., cell ID*gene name, and the type name corresponding to each cell ID, i.e., cell type metadata. For example, cell types are defined as adrenergic cells (ADRN), mesenchymal cells (MES), endothelial cells, T cells, B cells, myeloid cells, and Schwann cells.

[0057] In this step, the present invention uses the iNMF algorithm for integration, which can effectively eliminate batch effects caused by different patients and different sequencing batches, while preserving the universal immune and tumor biological characteristics across samples.

[0058] Compared to conventional integration methods, the iNMF algorithm ensures that the cell types annotated subsequently (such as specific tumor subpopulations and Schwann cell precursors) are highly consistent across samples, thus providing a highly pure reference map for spatial deconvolution.

[0059] In this embodiment, 107 neuroblastoma (NB) tissue samples were collected and 10x Genomics single-cell transcriptome sequencing was performed. The iNMF algorithm was used to deeply integrate the 107 samples, decomposing them into shared factors and dataset-specific factors. Specifically, the single-cell gene expression matrix in the single-cell dataset was decomposed.

[0060] Step 2: Construction of multimodal spatial omics dataset and cell deconstruction:

[0061] Visium HD high-resolution spatial transcriptome sequencing was performed on the matched NB samples from the neuroblastoma tissue samples collected in the first step. Sequencing yielded spatial expression data and HE staining for pathological sections. The RCTD algorithm was used, with the cell type dictionary obtained in the first step as reference data, that is, with the transcriptome matrix data of single cells and cell type metadata as reference data, to deconvolve the spatial spectra of Visium HD to obtain a spatial coordinate table of cell types. The spatial coordinate table of cell types includes the x-axis, y-axis and the corresponding cell type, which is an annotation of the same cell types as in the first step on the spatial transcriptome data.

[0062] This step enables precise annotation of various cell types at every coordinate point in space.

[0063] The Visium HD used above is a spatial transcriptome sequencing technology from 10xgenomics. Sequencing yields spatial expression data, which is the expression value of genes at each coordinate. This step also produces HE staining of pathological sections, which corresponds one-to-one with the spatial transcriptome.

[0064] Spatial expression data is the expression of genes at each coordinate, and HE staining corresponds one-to-one with the spatial transcriptome.

[0065] RCTD stands for Robust Cell Type Decomposition.

[0066] In this embodiment, Visium HD high-resolution spatial transcriptome sequencing was performed on 57 matched NB samples out of the 107 cases mentioned above to obtain the same cell type annotations on the spatial transcriptome data as in the first step, and to obtain a spatial coordinate table of cell types, i.e., x-axis, y-axis and corresponding cell types.

[0067] Step 3: Spatial niche definition based on sliding window sampling:

[0068] A sliding window sampling strategy was used to obtain the cell proportion matrix for each virtual spot by applying the spatial coordinate table of cell types obtained in the second step. For the cell proportion matrix of each virtual spot, a non-negative matrix factorization algorithm (hereinafter referred to as NMF) was performed to decompose it into a NICHE contribution matrix H. Based on the NICHE contribution matrix H, k types of tumor niche NICHE and k-dimensional continuous NICHE scores corresponding to each coordinate point on each HE slice were obtained. Based on the tumor niche and the ID of the virtual spot, the spatial ecotype matrix of the virtual spot was obtained. At the same time, the cell type weight matrix of the tumor microenvironment ecotype was obtained based on the tumor niche and cell type.

[0069] In this step, the cell proportion matrix of the virtual spot is obtained by calculating the average percentage of each type of cell in the window set by the sliding window sampling strategy. For example, in this embodiment, the window size is set to 500 micrometers × 500 micrometers, and the sliding step size is 100 micrometers; each window is defined as a virtual spot, and then the average percentage of each type of cell in the window is calculated to obtain the cell proportion matrix of each virtual spot.

[0070] When performing a nonnegative matrix factorization algorithm on the cell proportion matrix of virtual spots, the value of k is obtained by the following method: determining the search interval for the value of k, performing traversal tests within the search interval, and using evaluation indicators for comprehensive evaluation to obtain the optimal value of k; the above evaluation indicators include the co-table correlation coefficient, the sum of squared residuals and reconstruction error, and the dispersion coefficient.

[0071] For example, in this implementation, the search interval is [1, 50]. Each value in the search interval is comprehensively evaluated. Among the evaluation metrics used, the cophenetic correlation coefficient is used to evaluate the stability of the clustering results. The inflection point before the coefficient decreases significantly is selected. The higher the value, the more stable and reliable the clustering structure or latent variable relationship of the model is. The residual sum of squares (RSS) and reconstruction error are used to evaluate the degree to which the decomposition results restore the original cell distribution information. The smaller the value, the less information loss and the stronger the model's ability to fit the data. The dispersion coefficient ensures that there is significant feature discrimination between different NICHEs. Generally, a higher dispersion coefficient is selected.

[0072] Through traversal testing of 1-50 dimensions, the experimental results show that when k=16, the model achieves the best balance between reconstruction Fidelity and cluster stability (i.e., the mathematical "golden inflection point"). Therefore, the value of k in this implementation is determined to be 16.

[0073] This step yields 16 tumor niche NICHEs representing specific cell spatial interaction patterns and a 16-dimensional continuous NICHE score for each coordinate point on each HE slice, serving as the "gold standard" label for deep learning.

[0074] In this invention, when performing a nonnegative matrix factorization algorithm on the cell proportion matrix of the virtual spots in the third step, the invention does not subjectively set the number of NICHEs, but adopts a data-driven iterative search algorithm. In this way, the resulting k tumor niche NICHEs and NICHE scores can comprehensively cover the complex cellular composition of neuroblastoma, including the tumor core area, the stromal infiltration area, the immune accumulation area, and the specific Schwann cell development microenvironment, while effectively avoiding overfitting caused by excessive dimensionality (i.e., avoiding the decomposition of random noise into independent NICHEs).

[0075] In addition, the virtual spot spatial ecotype matrix obtained in the third step is a matrix with column vectors representing tumor microenvironment ecotype IDs (NICHE IDs) and horizontal vectors representing virtual spot IDs (pseudo ids); the cell type weight matrix of the tumor microenvironment ecotype is a matrix with column vectors representing tumor microenvironment ecotype IDs (NICHE IDs) and horizontal vectors representing cell types. This weight matrix shows the cell type weight value of each ecotype. The higher the weight, the higher the importance of this cell in the corresponding ecotype, and the more it dominates the formation of the corresponding ecotype.

[0076] Step 4: High-precision preprocessing and quality control of pathological images:

[0077] Using the coordinates in the cell proportion matrix of the virtual spots obtained in step 3, the whole-slice HE image (that is, the pathological graphic corresponding to the spatial transcriptome detection sample) is cut by sliding window to generate virtual spot image patches corresponding to the spatial coordinates of the physical size of the virtual spots; the scores of k tumor microenvironment ecotypes in the virtual spot spatial ecotype matrix are mapped to the virtual spot image patches to obtain the ecotype score of each virtual spot patch.

[0078] Furthermore, the resulting virtual speckle image patches are filtered after sliding window segmentation. Specifically, an Ultra-clean QC method is used, comprising four modules: HSV saturation filtering to remove low-saturation backgrounds with S-channel values ​​below 20; Otsu adaptive binarization to automatically segment tissues from the background based on brightness distribution; Laplacian variance verification to calculate image sharpness and remove out-of-focus or textureless areas with variance below 50, ensuring the model learns effective features; and color normalization using the Macenko algorithm, employing a standard HE slice as a reference template to unify the hue of all image patches.

[0079] Step 5: Pre-calculation of the dual-teacher feature database and warm-up of expert heads:

[0080] 5.1 A dual-teacher model consisting of Teacher A and Teacher B is used to extract features from the virtual speckle image patches obtained in step four, and a feature library of virtual speckle image patches is pre-calculated and generated; among them, Teacher A extracts pathological features and Teacher B extracts geometric structural features.

[0081] Specifically, the teacher A model in this implementation is a pathological semantic model, which uses the UNI basic model (ViT-L architecture) to extract 1024-dimensional pathological features and stores them as an h5 file "dino_features.h5".

[0082] Teacher B model is a spatial structure model that uses the DINOv2 model (ViT architecture with registers) to extract 1024-dimensional geometric structure features and store them as an h5 file uni_features.h5.

[0083] In this embodiment, this step generates a feature library of 190,000 virtual spot patches, which are stored offline in HDF5 format.

[0084] 5.2. Perform morphological predictability detection analysis to achieve expert head preheating. The specific steps are as follows: completely fix the parameters of the dual-teacher model, and map the image patches in the feature library generated in step 5.1 into a high-dimensional feature vector space; train a multilayer perceptron regressor in this high-dimensional feature vector space, and use 5-fold cross-validation to score and train the k types of NICHE obtained in step 3; calculate the theoretical prediction upper limit of each NICHE, and based on the "consensus report", select several indicators with significant morphological signals, and then use them for expert head training to obtain the preheated "expert head".

[0085] In other words, for each of the k NICHEs, calculate its own score and the predicted value obtained from the model, and based on this, calculate the theoretical prediction upper limit R for each NICHE. 2 Score, R 2 The score measures how well a predictive model fits the real data; the closer the score is to 1, the better the prediction. Based on the "consensus report," indicators with significant morphological signals are selected, decisively eliminating purely molecularly driven or morphologically invisible noise indicators. The "consensus report" is a performance list of K NICHE indicators summarized after this 5-fold validation. Its core content includes: the average R² of each indicator (representing the upper limit of morphological prediction), the MAE / correlation coefficient of each indicator, and the prediction stability (standard deviation). It's called "consensus" because it represents the degree of agreement between "image features" and "spatial truth." If a NICHE has a high R², it means that the image features and this indicator have reached a "consensus"—that is, the signal is indeed morphologically visible.

[0086] In this embodiment, 190,000 HE image patches are mapped into a high-dimensional feature vector space. The 16 NICHE metrics derived from NMF decomposition are used for scoring training. Five-fold cross-validation is performed to select consensus metrics with Mean_R2 > 0.15 (such as NICHE_12, NICHE_11, etc.). After removing morphologically invisible noise metrics, 10 trainable tumor ecotypes (such as N12 R) are obtained. 2 =0.74, specifically N12 as shown in Table 1, which represents the Schwann cell-dominated tumor ecotype. This is then used for expert head training. This step yields the pre-warmed "expert head" parameter file "best_selected_head.pth".

[0087] Table 1. N12 is the Schwann cell-dominant tumor ecotype.

[0088] NICHE Mean_R2 Std_R2 N12 0.7421079993247985 0.11850942870527614 N14 0.4523025870323181 0.2489007560456852 N4 0.3948558926582336 0.2432610275259537 N8 0.3617557883262634 0.1450135036963012 N13 0.3080237627029419 0.16919547645893934 N10 0.30699973106384276 0.5868698114076905 N7 0.2801594972610474 0.321096585245813 N6 0.27250994443893434 0.30115690386262267 N11 0.25006287097930907 0.11583270524061991 N9 0.22138769626617433 0.20392827520452142 N1 0.07504327297210693 0.22108212007577374 N15 -0.319373786 0.7645285859901199 N3 -0.596569443 0.9692482918177858 N2 -0.771775162 1.9014633833517136 N5 -1.37641201 1.7385677261567036 N16 -6.731486392 9.86470124825275

[0089] Table 1 shows the mean and standard deviation of R2 for each tumor ecotype, with the mean R2 for N12 reaching 0.74.

[0090] 5.3 The preheated "expert head" uses a prefix stripping algorithm to solve the key name mismatch problem in independent and ensemble training of the multilayer perceptron, and then uses a three-in-one collaborative training method. Here, the three-in-one collaborative training refers to the application of task constraints, semantic constraints, and structural constraints, and the use of a hierarchical learning rate strategy to explore the backbone network and perform life-saving fine-tuning of the regression head.

[0091] In this step, the "expert head" warmed up in step 5.2 is trained independently, and its parameter keys have a specific prefix (such as net.0.weight). However, when integrated into the student model, the key names have changed to regressor.0.weight. At this point, a prefix stripping algorithm must be executed to dynamically remove redundant characters from the parameter dictionary, thereby achieving accurate parameter mapping. Therefore, step 5.3 uses the prefix stripping algorithm to process the "expert head," which can solve the key mismatch problem in multilayer perceptron training between independent and ensemble training, ensuring lossless transfer of preheated knowledge.

[0092] Furthermore, to protect the "expert head" knowledge from being destroyed by the large gradients generated by the unaligned backbone network in step 5.2, this invention implements a hierarchical learning rate strategy, specifically employing a three-pronged collaborative training approach. This includes: task constraints, using weighted mean squared error (WMSE) to ensure the student model inherits the regression accuracy of the experts; semantic constraints (UNI-driven), forcing the intermediate features of the student model to align to the pathological semantic space; structural constraints (DINOv2-driven), using geometric priors to correct the local receptive field of the convolutional kernels; and a hierarchical learning rate strategy, using a 3×10^6 learning rate for the backbone network. −5 The learning rate was explored, and a more subtle 1×10⁻⁶ was applied to the implanted Regressor (regression head, i.e., "expert head"). −5 The learning rate is then fine-tuned to ensure its survival.

[0093] Step 6: Cross-architecture knowledge distillation and 5-fold robust model training:

[0094] 6.1 Student Model Architecture Design and Weight Implantation: A lightweight convolutional neural network was selected as the backbone network of the student model, and two independent linear projections were constructed in parallel after the global average pooling layer of the backbone network to map the features to the pathological semantic space and the geometric structure space; the optimal MLP regression head weights obtained by training in step 5 were loaded into the output of the student model.

[0095] In this step, two independent linear projections are constructed in parallel: the UNI projection head and the DINOv2 projection head. Both projection heads contain linear layers and layer normalization (LayerNorm). The UNI projection head maps features to the pathological semantic space, while the DINOv2 projection head maps features to the geometric structure space. The reason for setting up these two linear projections in parallel is that the feature dimensions of the student model (CNN) and the teacher model (ViT) are inconsistent (2048 dimensions vs. 1024 dimensions).

[0096] In this embodiment, the backbone network of the student model is preferably a lightweight convolutional neural network ResNet50, which can ensure inference efficiency on clinical edge computing devices. In the fifth step, the trained expert head is implanted into the student model. During the loading process, the key name automatic rematching algorithm (Key Stripping) is executed to remove the prefix redundancy when the expert head is trained independently, so as to achieve lossless transfer of cross-model parameters.

[0097] 6.2 Stratified ranking and 5-fold cross-validation based on patient heterogeneity: The cumulative value of k NICHE scores for each patient is calculated to obtain a comprehensive index reflecting the complexity of the tumor microenvironment, and all patients are ranked in descending order based on this comprehensive index; patients are cyclically assigned to 5 Folds according to the ranking order; during the assignment process, group constraints are used to strictly lock the patient ID to ensure that all patches from the same source can only exist in the same set.

[0098] In other words, in this invention, in response to the high individual differences exhibited by the samples, instead of using conventional random splitting, a systematic hierarchical sampling based on outcome intensity is performed. That is, a sorting mechanism is used to perform descending sorting, hierarchical distribution, and group constraints to ensure that image patches from the same source exist in only one set.

[0099] In this implementation, for the 57 samples, the cumulative value of each patient's 16 NICHE scores was first calculated as a comprehensive indicator reflecting the complexity of the tumor microenvironment, and the 57 patients were sorted in descending order. Next, stratified distribution was performed, cyclically assigning patients to 5 folds (i.e., 0, 1, 2, 3, 4, 0, 1...) according to the sorting order, ensuring that each fold included representative samples with high NICHE scores. Then, group constraints were applied, strictly locking patient IDs during the splitting process to ensure that all 190,000 patches from the same source existed only in the same set (training, validation, or testing), completely preventing data leakage.

[0100] 6.3 Weak signal index directional rescue mechanism is adopted to enhance the weak signals or sparsely distributed indices found in the third step of NMF decomposition.

[0101] Among them, the weak signal indicator targeted rescue mechanism is a dual enhancement strategy of "sampling-loss", which includes weighted random sampling and multi-task collaborative loss function optimization.

[0102] The specific steps of weighted random sampling are as follows: Before the start of each training iteration epoch, the label vectors of all patches in the training set are scanned; then, a "rare sample detection function" is defined. If the score of a patch's weak signal or sparsely distributed indicator exceeds a preset threshold, the sampling weight of that sample is increased to 10 times that of ordinary samples; otherwise, the original sample is retained. In other words, for weighted random sampling, weights are evaluated first, and then rare samples are weighted based on triggering logic. This forces the model to encounter these weak signal features with a high probability in each training mini-batch, solving the problem of "prediction degradation" or "prediction of all zeros" caused by the model's excessive tendency for training samples to score 0.

[0103] The specific steps for optimizing the multi-task collaborative loss function are as follows: First, joint gradient descent is performed, that is, a synchronous optimization approach is adopted, in which L is calculated simultaneously in each mini-batch. task L uni With L dino The numerical value is obtained, and the total loss function L is calculated using the chain rule. total The gradient of the student model parameters; where L total =α⋅L task +β1⋅L uni +β2⋅L dino In the formula, α is the task loss weight (e.g., 5.0 in this embodiment), β1 and β2 are distillation weights (e.g., 1.5 and 1.0 in this embodiment), L task It is the weighted task loss, L uni It is a semantic consistency constraint, L dino It is a structural consistency constraint.

[0104] Secondly, hierarchical dynamic step size adjustment is employed. Specifically, the AdamW optimizer is used in conjunction with the Cosine Annealing Scheduler. In the early stages of training, larger distillation weights force the student model to align features. Later in training, as the learning rate decreases, the model enters a fine-tuning phase, ensuring that L... task The indicators have reached their optimal levels.

[0105] Next, gradient clipping technology is used in L task When weak signal indicators generate abnormal gradients due to oversampling, the global gradient magnitude is truncated (e.g., the threshold is set to 1.0) to ensure the numerical stability of multi-task collaborative training.

[0106] In this invention, the total loss function L total It is a three-in-one composite structure, in which the weighted task loss L taskThe mean squared error (MSE) is weighted by terms, and higher weight coefficients (e.g., 2.0-3.0) are assigned to the regression error of weak signals (such as NICHE_11 and NICHE_1) to compensate for the insufficient signal strength.

[0107] Feature consistency constraint (L) uni ,L dino ): By calculating the distance between the projected features of the student model and the pre-stored features of the teacher model in the HDF5 (Hierarchical Data Format) library using MSE, the feature space of the student model ResNet50 is aligned with the larger model; specifically, semantic consistency constraint L uni This guides student models to mimic the high-level semantic space of the UNI pathological model, ensuring that the models can recognize deep pathological features such as malignant cell atypia and immune infiltration; structural consistency constraints L dino This guides students to model the geometric feature space of the DINOv2 general vision model, ensuring that the model can recognize underlying structural priors such as organizational boundaries and spatial textures.

[0108] The main steps in obtaining consistency constraints are offline feature extraction, online spatial alignment, and distance calculation. Through these three steps, the professional features pre-stored in the teacher model are transformed into distance constraints in the loss function, which forces the student model to internalize and align the pathological semantics and structural logic of the base model by "narrowing the gap" during training.

[0109] In the offline feature extraction process, before training, all image patches are input into the UNI and DINOv2 models with frozen parameters to extract the 1024-dimensional feature vectors corresponding to each image patch, and then serialized and stored using HDF5 to build an offline "expert feature library".

[0110] Online spatial alignment is a key step. Since the feature dimensions (2048 dimensions) output by the student model do not match those of the teacher model (1024 dimensions), this invention designs two independent linear projection layers after the student model.

[0111] During training, the Euclidean distance between the projected features generated by the student model and the teacher features read from the feature library is calculated in real time using mean squared error (MSE).

[0112] 6.4 The parameter model of the optimal student model trained on each fold of data is obtained. Additionally, the R-values ​​for each fold under different tumor ecotypes are also obtained. 2 The mean values ​​are shown in Table 2.

[0113] Table 2 R of each fold in different tumor ecotypes 2 mean

[0114] NICHE_12_R2 NICHE_14_R2 NICHE_4_R2 NICHE_8_R2 NICHE_13_R2 NICHE_10_R2 NICHE_7_R2 NICHE_6_R2 NICHE_11_R2 NICHE_9_R2 FOLD_1 0.96 0.52 0.15 0.36 0.23 0.49 0.06 0.35 0.32 0.32 FOLD_2 0.73 0.43 0.54 0.09 -0.19 -0.25 -0.12 0.34 -0.18 0.07 FOLD_3 0.61 0.37 -0.36 0.30 0.38 0.32 0.64 0.11 -0.01 0.13 FOLD_4 0.62 0.65 0.65 0.64 0.29 0.39 0.15 0.11 0.29 0.07 FOLD_5 0.78 -0.39 0.43 -0.16 0.31 0.63 0.50 0.37 0.32 0.01

[0115] Through R 2 The mean can measure the accuracy of the niche score prediction.

[0116] Step 7: Integrated Reasoning and Clinical Prognostic Assessment of the Entire Film:

[0117] 7.1 The image to be predicted is subjected to normalization preprocessing. The specific steps of preprocessing include physical scale alignment, high-overlap sliding window sampling, and coordinate-locked naming method.

[0118] Among them, physical scale alignment is performed on the original SVS format slices to be inferred. Based on the pixel physical size (e.g., 0.5µm / pixel) in its metadata, a bicubic interpolation algorithm is used to uniformly scale them to the same physical resolution as the training set (e.g., a scaling factor of 0.383), generating a high-resolution PNG benchmark image.

[0119] High-overlap tiling performs tiling cuts on the reference map; for example, in this embodiment, in order to ensure the continuity and spatial smoothness of the generated heatmap, the cut size is set to 383 pixels (corresponding to the original scale of 1000µm), and the sliding step size is set to 77 pixels (corresponding to the original scale of 200µm).

[0120] The coordinate-locked naming convention automatically names each generated image patch in the format of sample name_center point X coordinate_center point Y coordinate.png. This naming rule establishes a unique mapping between each predicted score and its physical location in biological tissue.

[0121] 7.2 A multi-model ensemble voting inference mechanism consisting of a 5-fold ensemble learning strategy is used for prediction: The parameter models of the 5 best student models obtained in step 6 are loaded in parallel, and each image patch after preprocessing in step 7.1 is used as input to obtain the 5 model prediction values ​​for each image patch for each NICHE index, and the arithmetic mean of the 5 model prediction values ​​is used as the final NICHE score Sn of the image patch; In addition, the standard deviation of the 5 model prediction values ​​is calculated simultaneously, and the magnitude of the standard deviation value is used as the confidence index of the system prediction.

[0122] That is, for this invention, the 5-fold ensemble learning strategy can eliminate the prediction bias of a single model under specific data partitioning. In this strategy, the final NICHE score Sn of the image patch is calculated through a mean fusion algorithm, i.e.,

[0123]

[0124] In the formula, Let m be the predicted value of the m-th model for the n-th NICHE index.

[0125] Meanwhile, uncertainty is quantified by calculating confidence index, where a lower standard deviation indicates a higher degree of consensus among models and stronger clinical reference value.

[0126] Through this step, you can obtain the following: Figure 1 The table shown, Figure 1 Only the first three niche scores and a few of the first image patches are shown.

[0127] 7.3 Spatial Heatmap Reconstruction and Clinical Indicator Conversion: Using the center point coordinate information in the image block file name contained in the coordinate locking naming method during preprocessing in step 7.1, the NICHE score Sn calculated in step 7.2 is mapped back to a two-dimensional matrix corresponding to the proportion of the original image, resulting in the reconstructed score matrix. A Gaussian filtering algorithm (e.g., σ=1.5) is applied to the reconstructed score matrix to generate a visually continuous spatial distribution heatmap. By setting a score threshold for each indicator, the area of ​​high-scoring regions exceeding the score threshold in the spatial distribution heatmap is calculated to obtain the area proportion of different NICHE indicators in the spatial distribution heatmap. Based on the area proportion, the spatial overlap, exclusion, or inclusion relationships of different NICHE indicators are analyzed to define quantitative clinical indicators reflecting the complexity of the tumor microenvironment.

[0128] 7.4 Based on the quantitative clinical indicators obtained in step 7.3, the evaluation results are derived.

[0129] Based on the aforementioned mean predictive scores, prognostic analysis can be performed in conjunction with clinical information. This clinical information includes follow-up time and data on the occurrence of symptoms (relapse, death, metastasis, and other adverse prognostic events). The specific prognostic analysis steps are as follows:

[0130] Based on the area ratio of different NICHE indicators obtained from the model, patients are automatically divided into high-risk and low-risk groups. An adaptive search algorithm based on the maximum rank statistic is used for this division. During execution, this algorithm automatically traverses all possible values ​​of the DSSI score within the [20th, 80th] percentile interval in the patient cohort as candidate thresholds. For each candidate threshold, the statistical significance of the difference in overall survival between the high-risk and low-risk groups is calculated using the Log-rank test. The value that maximizes the Log-rank χ² statistic (i.e., minimizes the P-value) is selected as the final classification threshold (in this embodiment, the automatically determined optimal cutoff point is 0.43854). Then, a survival analysis curve (KM curve) is generated to differentiate patient prognoses.

[0131] That is, the adaptive search algorithm partitioning method provided by this invention can automatically find the cut point of "maximizing survival difference", ensuring that the stratification criteria are driven by objective clinical data rather than subjectively set by researchers, thus achieving "automatic partitioning" without human intervention.

[0132] Taking this implementation method as an example, when performing prognostic correlation analysis:

[0133] Generate a full-slice spatial distribution map based on the prediction results, and calculate the area percentage of key indicators (such as NICHE_12).

[0134] Combining clinical data, the Kaplan-Meier survival curve was used to demonstrate the significant correlation between this digital score and patient prognosis (experimental confirmation p = 0.0024).

[0135] In summary, based on the scoring method of this invention, combined with clinical data such as the patient's follow-up survival status and the recorded follow-up time, prognostic risk assessment can be achieved.

[0136] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A scoring method for the pathological microenvironment of neuroblastoma based on spatial omics priors and dual-teacher distillation, characterized by: Includes the following steps, Step 1: Multi-sample single-cell transcriptome integration and cell deconstruction based on the iNMF algorithm: We collected neuroblastoma tissue samples and constructed a single-cell dataset. We then used the iNMF algorithm to perform deep integration of the single-cell dataset, achieving cross-sample heterogeneity integration. We performed cluster analysis in the integrated feature space and defined a precise cell type dictionary based on marker genes. Step 2: Construction of multimodal spatial omics dataset and cell deconstruction: Visium HD high-resolution spatial transcriptome sequencing was performed on the matched NB samples from the neuroblastoma tissue samples collected in the first step to obtain spatial expression data and HE staining for pathological sections. Using the RCTD algorithm, with the cell type dictionary obtained in the first step as reference data, the spatial spectra of Visium HD are deconvolved to obtain the spatial coordinate table of cell types; Step 3: Spatial niche definition based on sliding window sampling: A sliding window sampling strategy was used on the spatial coordinate table of cell types obtained in the second step to obtain the cell proportion matrix of each virtual spot. For the cell proportion matrix of each virtual spot, a non-negative matrix factorization algorithm was performed to decompose it into a NICHE contribution matrix H. Based on the NICHE contribution matrix H, k types of tumor niche NICHE and k-dimensional continuous NICHE scores corresponding to each coordinate point on each HE slice were obtained. Based on the tumor niche and the ID of the virtual spot, the spatial ecotype matrix of the virtual spot was obtained. At the same time, based on the tumor niche and cell type, the cell type weight matrix of the tumor microenvironment ecotype was obtained. Step 4: High-precision preprocessing and quality control of pathological images: Using the coordinates in the cell proportion matrix of the virtual spots obtained in step 3, the whole-slice HE image is cut by sliding window to generate virtual spot image patches corresponding to the spatial coordinates of the physical size of the virtual spots; the scores of the k tumor microenvironment ecotypes in the virtual spot spatial ecotype matrix are mapped to the virtual spot image patches to obtain the ecotype score of each virtual spot patch. Step 5: Pre-calculation of the dual-teacher feature database and warm-up of expert heads: 5.1 A dual-teacher model consisting of Teacher A and Teacher B is used to extract features from the virtual blob image patches obtained in step four, and a feature library of virtual blob image patches is pre-calculated and generated; among them, Teacher A is used to extract pathological features, and Teacher B is used to extract geometric structural features. 5.

2. Perform morphological predictability detection analysis to achieve expert head preheating. The specific steps are as follows: completely fix the parameters of the dual-teacher model, and map the image patches in the feature library generated in step 5.1 into a high-dimensional feature vector space; train a multilayer perceptron regressor in this high-dimensional feature vector space, and use 5-fold cross-validation to score and train the k types of NICHE obtained in step 3; calculate the theoretical prediction upper limit of each NICHE, and based on the "consensus report", select several indicators with significant morphological signals, and then use them for expert head training to obtain the preheated "expert head"; 5.3 The preheated "expert head" uses a prefix stripping algorithm to solve the key name mismatch problem in independent and ensemble training of the multilayer perceptron, and then uses a three-in-one collaborative training method. Here, the three-in-one collaborative training refers to the application of task constraints, semantic constraints, and structural constraints, and the use of a hierarchical learning rate strategy to explore the backbone network and perform life-saving fine-tuning of the regression head. Step 6: Cross-architecture knowledge distillation and 5-fold robust model training: 6.1 Student Model Architecture Design and Weight Implantation: A lightweight convolutional neural network was selected as the backbone network of the student model, and two independent linear projections were constructed in parallel after the global average pooling layer of the backbone network to map the features to the pathological semantic space and the geometric structure space; the optimal MLP regression head weights obtained by training in step 5 were loaded into the output of the student model. 6.2 Stratified Ranking and 5-Fold Cross-Validation Based on Patient Heterogeneity: The cumulative value of k NICHE scores for each patient is calculated to obtain a comprehensive index reflecting the complexity of the tumor microenvironment. All patients are then ranked in descending order based on this comprehensive index. Patients are cyclically assigned to 5 Folds according to the ranking order. During the assignment process, group constraints are used to strictly lock the patient ID to ensure that all patches from the same source can only exist in the same set. 6.3 Weak signal index directional rescue mechanism is adopted to enhance the weak signals or sparsely distributed indices found in the third step of NMF decomposition; 6.4, obtain the parameter model of the optimal student model trained on each fold of data; Step 7: Integrated reasoning and clinical prognostic assessment of the entire film: 7.1 The image to be predicted is subjected to normalization preprocessing. The specific steps of preprocessing include physical scale alignment, high-overlap sliding window sampling, and coordinate locking naming method. 7.2 A multi-model ensemble voting inference mechanism consisting of a 5-fold ensemble learning strategy is used for prediction: the parameter models of the 5 best student models obtained in step 6 are loaded in parallel, and each image patch after preprocessing in step 7.1 is used as input to obtain the 5 model prediction values ​​for each image patch for each NICHE index, and the arithmetic mean of the 5 model prediction values ​​is used as the final NICHE score Sn of the image patch; in addition, the standard deviation of the 5 model prediction values ​​is calculated simultaneously, and the magnitude of the standard deviation value is used as the confidence index of the system prediction. 7.3 Spatial Heatmap Reconstruction and Clinical Indicator Conversion: Using the center point coordinate information in the image block file name contained in the coordinate locking naming method during preprocessing in step 7.1, the NICHE score Sn calculated in step 7.2 is mapped back to a two-dimensional matrix corresponding to the proportion of the original image, resulting in a reconstructed score matrix. A Gaussian filtering algorithm is applied to the reconstructed score matrix to generate a visually continuous spatial distribution heatmap. By setting a score threshold for each indicator, the area of ​​high-scoring regions exceeding the score threshold in the spatial distribution heatmap is calculated to obtain the area proportion of different NICHE indicators in the spatial distribution heatmap. Based on the area proportion, the spatial overlap, exclusion, or inclusion relationships of different NICHE indicators are analyzed to define quantitative clinical indicators reflecting the complexity of the tumor microenvironment. 7.4 Based on the quantitative clinical indicators obtained in step 7.3, the evaluation results are derived.

2. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the first step, the single-cell dataset was obtained by performing single-cell transcriptome sequencing on collected neuroblastoma tissue samples.

3. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the first step, when integrating heterogeneous data across samples, the iNMF algorithm decomposes the single-cell dataset into shared factors and sample-specific factors.

4. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the first step, the cell type dictionary includes transcriptome matrix data for each single cell, i.e., cell ID * gene name, and the type name corresponding to each cell ID, i.e., cell type metadata.

5. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the second step, the spatial coordinate table of cell types includes the x-axis, y-axis, and corresponding cell types. It is an annotation of the same cell types as in the first step on the spatial transcriptome data.

6. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the third step, the cell proportion matrix of the virtual spots is obtained by calculating the average percentage of various types of cells in the window set by the sliding window sampling strategy.

7. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the third step, when performing the nonnegative matrix factorization algorithm on the cell proportion matrix of the virtual spots, the value of k is obtained by the following method: determining the search interval for the value of k, performing traversal tests within the search interval, and using evaluation indicators for comprehensive evaluation to obtain the optimal value of k; the above evaluation indicators include the co-table correlation coefficient, the sum of squared residuals and reconstruction error, and the dispersion coefficient.

8. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In the fourth step, the resulting virtual speckle image patch is filtered after sliding window cutting.

9. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In step 5.1, Teacher A model is a pathological semantic model, which uses the UNI basic model to extract pathological features; Teacher B model is a spatial structure model that uses the DINOv2 model to extract geometric structure features.

10. The scoring method for neuroblastoma pathological microenvironment based on spatial omics priors and dual-teacher distillation as described in claim 1, characterized in that: In step 6.3, the weak signal indicator targeted rescue mechanism is a dual enhancement strategy of "sampling-loss", which includes weighted random sampling and multi-task collaborative loss function optimization; The specific steps of weighted random sampling are as follows: before the start of each training iteration Epoch, scan the label vectors of all Pats in the training set; then, define a "rare sample detection function"; if the score of the weak signal or sparse distribution index of a Pat exceeds a preset threshold, increase the sampling weight of the sample to 10 times that of ordinary samples, otherwise retain the original sample. The specific steps for optimizing the multi-task collaborative loss function are as follows: First, joint gradient descent is performed, that is, a synchronous optimization approach is adopted, in which L is calculated simultaneously in each mini-batch. task L uni With L dino The numerical value is obtained, and the total loss function L is calculated using the chain rule. total The gradient of the student model parameters; where L total =α⋅L task +β1⋅L uni +β2⋅L dino In the formula, α is the task loss weight, β1 and β2 are both distillation weights, and L task It is the weighted task loss, L uni It is a semantic consistency constraint, L dino It is a structural consistency constraint; Secondly, a hierarchical dynamic step size adjustment is implemented. Specifically, the AdamW optimizer is used in conjunction with a cosine annealing learning rate scheduling algorithm. In the early stages of training, a large distillation weight forces the student model to align features. In the later stages of training, as the learning rate decreases, the model enters a fine-tuning phase to ensure L... task The indicators have reached their optimal levels. Next, gradient clipping technology is used in L task When weak signal indicators generate abnormal gradients due to oversampling, the global gradient magnitude is truncated to ensure the numerical stability of multi-task collaborative training.

Images