A digital pathology-based tumor grading intelligent evaluation method and system

Abstract

The present application belongs to the technical field of brain pathological diagnosis, and particularly relates to a tumor grading intelligent evaluation method and system based on digital pathology. Through a tumor region and invasive edge division module, the present application adopts an encoder-decoder segmentation network based on deep learning to perform pixel-level classification on standardized whole slice images, marks tumor parenchymal regions, necrotic regions, vascular regions, normal brain parenchymal regions and residual neuron cell body regions, extracts the boundary contour line of the tumor parenchymal region, and outwardly expands the set pixel width to form a tumor invasive edge region containing an inner edge band and an outer edge band. This design accurately locks the analysis focus on the invasive front region containing the maximum amount of diagnostic information, eliminates irrelevant signal interference of the tumor core region and the normal brain parenchymal region, and effectively improves the signal-to-noise ratio of the grading evaluation.

Description

Technical Field

[0001] This invention belongs to the field of brain pathology diagnosis technology, specifically relating to a method and system for intelligent assessment of tumor grading based on digital pathology. Background Technology

[0002] Gliomas are the most common primary malignant tumors of the central nervous system, characterized by invasive growth and a high recurrence rate. Histopathologically, glioma cells often infiltrate surrounding normal brain tissue along structures such as white matter fiber bundles and vascular basement membranes, forming a tumor invasion periphery that differs significantly from the tumor core region in cell density, arrangement, and microenvironmental interaction. The morphological characteristics of this region directly reflect the tumor's invasiveness and are an important basis for judging the degree of tumor malignancy, predicting patient prognosis, and guiding the extent of surgical resection.

[0003] Currently, pathologists primarily assess the invasiveness of gliomas through manual observation of tissue sections under an optical microscope. They focus on morphological features such as the distribution density of tumor cells within the invasive border region, the positional relationship with blood vessels, and interactions with residual neurons. However, this manual assessment method is highly subjective and inconsistent, lacking a unified quantitative standard for interpreting the degree of invasion. The assessment results heavily rely on the pathologist's personal experience and subjective judgment, with low consistency between different doctors and even among the same doctor at different times, directly impacting the reliability of clinical decisions. Furthermore, the information utilization is insufficient, and brain-specific analytical dimensions are lacking. In particular, glioma cells exhibit directional migration along white matter fiber bundles, a typical brain-specific biological behavior distinguishing them from other solid tumors. However, existing computer-aided analysis methods fail to extract directional features related to white matter fiber orientation, resulting in the loss of crucial brain-specific diagnostic information and reducing the accuracy of brain tumor assessment.

[0004] To address the aforementioned issues, this application presents a method and system for intelligent assessment of tumor grading based on digital pathology. Summary of the Invention

[0005] To address the shortcomings of the prior art mentioned in the background section, this application proposes a digital pathology-based intelligent assessment method and system for tumor grading. This method extracts interpretable topological feature sets of invasion margins from multiple dimensions with clear pathological meanings, and effectively fuses these features to achieve an accurate, transparent, and brain-specific intelligent grading assessment method, thereby solving the problems in the background section.

[0006] Firstly, to achieve the above objectives, this application provides a digital pathology-based intelligent assessment method for tumor grading, comprising the following specific steps:

[0007] Step S1: Obtain a digital pathological whole-slice image of a glioma tissue section, and perform background correction and color standardization processing on the digital pathological whole-slice image to obtain a standardized whole-slice image.

[0008] Step S2: Based on the standardized whole slice image, a deep learning segmentation network is used to identify and segment the tumor core region, and the tumor invasion edge region is obtained by expanding outward by a preset width according to the boundary of the tumor core region.

[0009] Step S3: In the tumor invasion edge region, extract the invasion edge topological feature set reflecting the interaction between tumor cells and brain microenvironment. The invasion edge topological feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells.

[0010] Step S4: Input the invasion edge topological feature group into the pre-constructed multi-source feature fusion network based on the attention mechanism, and calculate and output the comprehensive invasion potential index through the multi-source feature fusion network;

[0011] Step S5: Compare the comprehensive invasive potential index with the preset grading threshold range, output the invasiveness grading result of the glioma, and generate a visualized pathology report marked with key invasive regions by combining the spatial distribution of various features in the invasive edge topological feature group.

[0012] Based on the above scheme, step S2 specifically includes the following steps:

[0013] Step S21: Use a deep learning-based encoder-decoder segmentation network to perform pixel-level classification on the standardized whole-slice image, marking the tumor parenchyma region, necrotic region, vascular region, normal brain parenchyma region, and residual neuronal cell body region.

[0014] Step S22: Extract the boundary contour of the tumor solid region, and extend the boundary contour outwards from the tumor by a set pixel width. Define the annular region formed after the extension as the tumor invasion edge region.

[0015] Step S23: Divide the tumor invasion edge region radially into an inner edge zone near the tumor core and an outer edge zone near normal brain tissue.

[0016] Based on the above scheme, the preferred method for calculating the tumor cell density gradient value along the invasion front in step S3 is as follows:

[0017] First, multiple sampling rays are generated in the region at the edge of the tumor invasion, along a direction perpendicular to the boundary of the tumor core;

[0018] Then, multiple sampling points are set at equal intervals from the inner edge to the outer edge on each sampling ray, and the ratio of the number of tumor cell nuclei to the sampling area at each sampling point is calculated to obtain the local tumor cell density.

[0019] Finally, a linear fit was performed on the local tumor cell density and distance on each sampling ray to obtain a gradient value representing the decrease in density as the distance increases. The maximum value of the gradient values ​​of all sampling rays was taken as the tumor cell density gradient value along the invasion front.

[0020] Based on the above scheme, the preferred method for calculating the anisotropy index of tumor cells along the direction of white matter fiber bundles in step S3 is as follows:

[0021] First, within the tumor invasion margin region, the image is divided into multiple local sub-regions. For each local sub-region, the main direction of local tissue texture is estimated using structural tensor analysis or histogram of orientation gradients, which serves as the direction of white matter fiber bundles in that local sub-region.

[0022] Then, within each local sub-region, the linear density of tumor cell nuclei in the first direction is calculated along the principal direction, and the linear density of tumor cell nuclei in the second direction is calculated along the direction perpendicular to the principal direction.

[0023] Finally, the difference between the linear density in the first direction and the linear density in the second direction is divided by the sum of the two to obtain the directional aggregation ratio of the local sub-region. The arithmetic mean of the directional aggregation ratios of all local sub-regions is then calculated to obtain the anisotropy index of tumor cells along the direction of white matter fiber bundles.

[0024] Based on the above scheme, the preferred method for calculating the aggregation degree of perivascular tumor cells in step S3 is as follows:

[0025] First, within the tumor invasion margin region, the contours of all vessel cross-sections are extracted based on the marked vascular regions;

[0026] Then, for each blood vessel section, an annular region formed by extending outward from the outline of the blood vessel section by a set distance is extracted, and the ratio of the total number of tumor cell nuclei in the annular region to the area of ​​the annular region is calculated as the aggregation index of the blood vessel section.

[0027] Finally, the arithmetic mean of the aggregation indices of all blood vessel sections is calculated to obtain the aggregation degree of perivascular tumor cells.

[0028] The contact ratio of residual neurons encapsulated by tumor cells in step S3 is calculated as follows:

[0029] First, within the tumor invasion edge region, the cell membrane contours of residual neuron cell bodies are extracted based on the marked residual neuron cell body regions.

[0030] Then, for each residual neuron cell body, a buffer distance is set outside its cell membrane outline. The total length of the arc segment on the cell membrane outline that is within the buffer distance by at least one tumor cell nucleus is calculated and divided by the total perimeter of the neuron cell body cell membrane outline to obtain the encapsulation ratio of a single neuron.

[0031] Finally, the arithmetic mean of the encapsulation ratios of all residual neurons was calculated to obtain the contact ratio of residual neurons encapsulated by tumor cells.

[0032] Preferably, based on the above scheme, the attention-based multi-source feature fusion network in step S4 includes:

[0033] The first feature encoding branch is used to receive the tumor cell density gradient value and its spatial distribution statistics along the invasion front, and output the first feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the density gradient on all sampling rays;

[0034] The second feature encoding branch is used to receive the anisotropy index of tumor cells along the direction of white matter fiber bundles and its spatial distribution statistics, and output the second feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the anisotropy index in all local sub-regions;

[0035] The third feature encoding branch is used to receive the aggregation degree of perivascular tumor cells and their spatial distribution statistics, and output the third feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the aggregation degree on all vascular sections.

[0036] The fourth feature encoding branch is used to receive the contact ratio of residual neurons wrapped by tumor cells and its spatial distribution statistics, and outputs the fourth feature vector, where the spatial distribution statistics include the maximum, minimum and standard deviation of the contact ratio on all residual neurons.

[0037] The attention fusion layer is used to calculate the attention weight coefficients of the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector, and to perform a weighted sum of the four feature vectors based on the attention weight coefficients to obtain the fused feature vector.

[0038] A fully connected mapping layer is used to map the fused feature vector to a fused feature scalar;

[0039] Nonlinear activation units are used to map the fused feature scalar to the 0-1 interval through a nonlinear activation function to obtain the comprehensive index of invasion potential.

[0040] In a preferred embodiment based on the above scheme, step S5 specifically includes the following steps:

[0041] Step S51: Set a first grading threshold and a second grading threshold, wherein the first grading threshold is less than the second grading threshold;

[0042] Step S52: When the comprehensive index of invasion potential is less than the first classification threshold, it is determined to be classified as low invasion level.

[0043] When the comprehensive index of invasion potential is greater than or equal to the first grading threshold and less than the second grading threshold, it is judged as medium invasiveness.

[0044] When the comprehensive index of invasion potential is greater than or equal to the second classification threshold, it is judged as a high invasion level.

[0045] Step S53: Generate a visualized pathology report. The visualized pathology report includes the invasiveness grading results and a thermal map of the key invasive regions formed by normalizing and weighting the spatial distribution maps of the tumor cell density gradient value along the invasion front, the tumor cell anisotropy index along the white matter fiber bundle direction, the aggregation degree of perivascular tumor cells, and the contact ratio of residual neurons wrapped by tumor cells in the tumor invasion edge region. The weights of the weighting are determined by the attention weight coefficients calculated by the attention fusion layer.

[0046] Secondly, this application provides a tumor grading intelligent assessment system based on digital pathology, which specifically includes: an image acquisition and standardization module, a tumor region and invasion edge division module, an invasion edge topology feature extraction module, a grading assessment network module, and a grading result output and visualization report generation module.

[0047] The image acquisition and standardization module is used to acquire digital pathological whole-slice images of glioma tissue sections, and to perform background correction and color standardization processing on the digital pathological whole-slice images to obtain standardized whole-slice images.

[0048] The tumor region and invasion edge segmentation module is used to identify and segment the tumor core region based on a standardized whole slice image using a deep learning segmentation network, and to expand the tumor invasion edge region outward by a preset width according to the boundary of the tumor core region.

[0049] The invasion edge topology feature extraction module is used to extract an invasion edge topology feature set reflecting the interaction between tumor cells and the brain microenvironment within the tumor invasion edge region. The invasion edge topology feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells.

[0050] The hierarchical evaluation network module is used to store and run a pre-built multi-source feature fusion network based on the attention mechanism, receive the invasion edge topology feature group and calculate and output the comprehensive invasion potential index.

[0051] The grading result output and visualization report generation module is used to compare the comprehensive invasive potential index with the preset grading threshold range, output the invasiveness grading result of glioma, and generate a visualization pathology report marked with key invasive regions by combining the spatial distribution of various features in the invasive edge topological feature group.

[0052] Thirdly, this application provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0053] The processor executes the aforementioned intelligent assessment method for tumor grading based on digital pathology by calling the computer program stored in the memory.

[0054] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned intelligent assessment method for tumor grading based on digital pathology.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] First, this invention introduces a tumor cell anisotropy index along the direction of white matter fiber bundles into digital pathological grading assessment. This index estimates the principal direction of local tissue texture as the direction of white matter fiber bundles through structural tensor analysis or directional gradient histogram, and calculates the linear density of tumor cell nuclei along the principal direction and perpendicular to the principal direction, respectively. The normalized difference ratio is used to quantify the degree of directional aggregation of tumor cells along white matter fiber bundles. This feature directly corresponds to the core biological behavior of glioma's directional migration along white matter fiber bundles, which can greatly enhance the disease specificity of grading assessment.

[0057] Secondly, the topological feature set of the invasion edge extracted in this invention includes four features: the tumor cell density gradient value along the invasion front, the tumor cell anisotropy index along the direction of white matter fiber bundles, the aggregation degree of perivascular tumor cells, and the contact ratio of residual neurons wrapped by tumor cells. Each feature has a clear pathological meaning, and pathologists can directly understand the tumor-microenvironment interaction pattern reflected by each feature. At the same time, the attention weight coefficient output by the attention fusion layer in the grading assessment network module directly reflects the relative contribution of the four features to the grading decision. Combined with the grading result output and the heat map of the key invasion area generated by the visualization report generation module based on the weighted superposition of attention weight coefficients, pathologists can fully trace and review the reasoning basis of the system, which can significantly improve clinical acceptability and diagnostic transparency.

[0058] Third, this invention employs a tumor region and invasion edge segmentation module, utilizing a deep learning-based encoder-decoder segmentation network to perform pixel-level classification on standardized whole-slice images, marking the tumor parenchyma region, necrotic region, vascular region, normal brain parenchyma region, and residual neuronal cell body region. Based on this, the boundary contour of the tumor parenchyma region is extracted and extended outwards by a set pixel width to form a tumor invasion edge region containing inner and outer edge bands. This design precisely focuses the analysis on the invasion front region containing the maximum diagnostic information, eliminating irrelevant signal interference from the tumor core region and normal brain parenchyma region, effectively improving the signal-to-noise ratio of the grading assessment. Furthermore, through the first, second, third, and fourth feature encoding branches, four features and their spatial distribution statistics are mapped into feature vectors. Then, the attention fusion layer adaptively learns the attention weight coefficients of each feature vector and performs a weighted summation. This fusion mechanism fully utilizes the complementary information of multi-dimensional features to improve grading accuracy while retaining the advantages of each feature being independent, traceable, and interpretable. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a schematic diagram of the overall process of the intelligent assessment method for tumor grading based on digital pathology of the present invention.

[0061] Figure 2 This is a flowchart illustrating step S2 in this invention;

[0062] Figure 3 This is a system block diagram of a tumor grading intelligent assessment system based on digital pathology according to the present invention. Detailed Implementation

[0063] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0064] Example 1

[0065] To address the technical problems raised in the background art, this application provides a preferred embodiment: such as Figures 1-3As shown, a digital pathology-based intelligent assessment method for tumor grading includes the following specific steps:

[0066] Step S1: Obtain a digital pathological whole-slice image of a glioma tissue section, and perform background correction and color standardization processing on the digital pathological whole-slice image to obtain a standardized whole-slice image.

[0067] Step S2: Based on the standardized whole slice image, a deep learning segmentation network is used to identify and segment the tumor core region, and the tumor invasion edge region is obtained by expanding outward by a preset width according to the boundary of the tumor core region.

[0068] Step S3: In the tumor invasion edge region, extract the invasion edge topological feature set reflecting the interaction between tumor cells and brain microenvironment. The invasion edge topological feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells.

[0069] Step S4: Input the invasion edge topological feature group into the pre-constructed multi-source feature fusion network based on the attention mechanism, and calculate and output the comprehensive invasion potential index through the multi-source feature fusion network;

[0070] Spatial distribution statistics of invasion edge topological feature groups:

[0071] Before performing step S4, the spatial distribution statistics of the features extracted in step S3 need to be performed to fully construct the input information of the network.

[0072] Specifically, this step involves: calculating the maximum, minimum, and standard deviation of the tumor cell density gradient along the invasion front on all sampling rays; calculating the maximum, minimum, and standard deviation of the tumor cell anisotropy index along the white matter fiber bundle direction on all local sub-regions; calculating the maximum, minimum, and standard deviation of the perivascular tumor cell aggregation degree on all vascular cross sections; and calculating the maximum, minimum, and standard deviation of the contact ratio of residual neurons wrapped by tumor cells on all residual neurons.

[0073] These spatial distribution statistics, together with the feature values ​​calculated in step S3, constitute the complete input vector of the attention-based multi-source feature fusion network.

[0074] Step S5: Compare the comprehensive invasive potential index with the preset grading threshold range, output the invasiveness grading result of the glioma, and generate a visualized pathology report marked with key invasive regions by combining the spatial distribution of various features in the invasive edge topological feature group.

[0075] The advantages of this embodiment compared to the prior art are as follows: The method provided by the present invention covers the entire automated process from image acquisition and standardization module to acquire standardized whole slice images, tumor region and invasion edge segmentation module to segment tumor invasion edge regions, invasion edge topological feature extraction module to extract invasion edge topological feature groups, grading assessment network module to calculate the comprehensive index of invasion potential, to grading result output and visualization report generation module to output grading results and visualization pathology reports. It can be seamlessly embedded into the existing digital pathology scanning and diagnosis workflow without increasing the additional operational burden on pathologists, and has good prospects for clinical application transformation and promotion value.

[0076] Furthermore:

[0077] In an optional embodiment, step S2 specifically includes the following steps:

[0078] Step S21: Use a deep learning-based encoder-decoder segmentation network to perform pixel-level classification on the standardized whole-slice image, marking the tumor parenchyma region, necrotic region, vascular region, normal brain parenchyma region, and residual neuronal cell body region.

[0079] Step S22: Extract the boundary contour of the tumor solid region, and extend the boundary contour outwards from the tumor by a set pixel width. Define the annular region formed after the extension as the tumor invasion edge region.

[0080] Step S23: Divide the tumor invasion edge region radially into an inner edge zone near the tumor core and an outer edge zone near normal brain tissue.

[0081] It should be noted that the deep learning-based encoder-decoder segmentation network used in step S21 can be a known fully convolutional network architecture such as U-Net, DeepLab series, or FCN. When training this deep learning segmentation network, a large number of whole-slice images of glioma pathology can be collected, and pathologists can perform pixel-level annotations on regions such as tumor parenchyma, necrosis, blood vessels, normal brain parenchyma, and residual neuronal cell bodies in the images to construct a training dataset. Subsequently, conventional supervised learning methods such as cross-entropy loss function and stochastic gradient descent optimizer are used to train the network until it reaches the preset segmentation accuracy on the validation set. After training, the network can be used to automatically classify the standardized whole-slice images at the pixel level.

[0082] In an optional embodiment, the tumor cell density gradient value along the invasion front in step S3 is calculated as follows:

[0083] First, multiple sampling rays are generated in the region at the edge of the tumor invasion, along a direction perpendicular to the boundary of the tumor core;

[0084] Then, multiple sampling points are set at equal intervals from the inner edge to the outer edge on each sampling ray, and the ratio of the number of tumor cell nuclei to the sampling area at each sampling point is calculated to obtain the local tumor cell density.

[0085] Finally, a linear fit was performed on the local tumor cell density and distance on each sampling ray to obtain a gradient value representing the decrease in density as the distance increases. The maximum value of the gradient values ​​of all sampling rays was taken as the tumor cell density gradient value along the invasion front.

[0086] In an optional embodiment, the anisotropy index of tumor cells along the direction of white matter fiber bundles in step S3 is calculated as follows:

[0087] First, within the tumor invasion margin region, the image is divided into multiple local sub-regions. For each local sub-region, the main direction of local tissue texture is estimated using structural tensor analysis or histogram of orientation gradients, which serves as the direction of white matter fiber bundles in that local sub-region.

[0088] Then, within each local sub-region, the linear density of tumor cell nuclei in the first direction is calculated along the principal direction, and the linear density of tumor cell nuclei in the second direction is calculated along the direction perpendicular to the principal direction.

[0089] Finally, the difference between the linear density in the first direction and the linear density in the second direction is divided by the sum of the two to obtain the directional aggregation ratio of the local sub-region. The arithmetic mean of the directional aggregation ratios of all local sub-regions is then calculated to obtain the anisotropy index of tumor cells along the direction of white matter fiber bundles.

[0090] In an optional embodiment, the aggregation degree of perivascular tumor cells in step S3 is calculated as follows:

[0091] First, within the tumor invasion margin region, the contours of all vessel cross-sections are extracted based on the marked vascular regions;

[0092] Then, for each blood vessel section, an annular region formed by extending outward from the outline of the blood vessel section by a set distance is extracted, and the ratio of the total number of tumor cell nuclei in the annular region to the area of ​​the annular region is calculated as the aggregation index of the blood vessel section.

[0093] Finally, the arithmetic mean of the aggregation indices of all blood vessel sections is calculated to obtain the aggregation degree of perivascular tumor cells.

[0094] The contact ratio of residual neurons encapsulated by tumor cells in step S3 is calculated as follows:

[0095] First, within the tumor invasion edge region, the cell membrane contours of residual neuron cell bodies are extracted based on the marked residual neuron cell body regions.

[0096] Then, for each residual neuron cell body, a buffer distance is set outside its cell membrane outline. The total length of the arc segment on the cell membrane outline that is within the buffer distance by at least one tumor cell nucleus is calculated and divided by the total perimeter of the neuron cell body cell membrane outline to obtain the encapsulation ratio of a single neuron.

[0097] Finally, the arithmetic mean of the encapsulation ratios of all residual neurons was calculated to obtain the contact ratio of residual neurons encapsulated by tumor cells.

[0098] It should be noted that: in the calculation of the tumor cell density gradient value along the invasion front, the maximum value of all sampled ray gradient values ​​is taken instead of the average value. This can effectively capture the leading edge region of tumor cells invading along the local structure in a finger-like protrusion manner, avoiding the dilution of strong invasion signals by averaging. In the calculation of the aggregation degree of tumor cells around blood vessels, the area of ​​the annular region is used instead of the blood vessel perimeter as the normalization factor, eliminating the systematic bias caused by the difference in the size of the blood vessel cross-section. In the calculation of the contact ratio of residual neurons wrapped by tumor cells, a buffer distance is set outside the cell membrane contour line of the residual neuron cell body for judgment, which improves the robustness of the recognition of intercellular interactions. The above optimizations make the quantitative results of various features more objective and stable in reflecting the true pathological state.

[0099] Furthermore:

[0100] In an optional embodiment, the attention-based multi-source feature fusion network in step S4 includes:

[0101] The first feature encoding branch is used to receive the tumor cell density gradient value and its spatial distribution statistics along the invasion front, and output the first feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the density gradient on all sampling rays;

[0102] The second feature encoding branch is used to receive the anisotropy index of tumor cells along the direction of white matter fiber bundles and its spatial distribution statistics, and output the second feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the anisotropy index in all local sub-regions;

[0103] The third feature encoding branch is used to receive the aggregation degree of perivascular tumor cells and their spatial distribution statistics, and output the third feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the aggregation degree on all vascular sections.

[0104] The fourth feature encoding branch is used to receive the contact ratio of residual neurons wrapped by tumor cells and its spatial distribution statistics, and outputs the fourth feature vector, where the spatial distribution statistics include the maximum, minimum and standard deviation of the contact ratio on all residual neurons.

[0105] The attention fusion layer is used to calculate the attention weight coefficients of the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector, and to perform a weighted sum of the four feature vectors based on the attention weight coefficients to obtain the fused feature vector.

[0106] A fully connected mapping layer is used to map the fused feature vector to a fused feature scalar;

[0107] Nonlinear activation units are used to map the fused feature scalar to the 0-1 interval through a nonlinear activation function to obtain the comprehensive index of invasion potential.

[0108] It should be noted that the first feature encoding branch, the second feature encoding branch, the third feature encoding branch, and the fourth feature encoding branch are all composed of at least two fully connected network layers, and each fully connected network layer is followed by a non-linear activation function and a dropout regularization layer;

[0109] The nonlinear activation unit uses the Sigmoid function to map the fused feature scalar to the interval between 0 and 1 to obtain the comprehensive index of invasion potential.

[0110] In an optional embodiment, step S5 specifically includes the following steps:

[0111] Step S51: Set a first grading threshold and a second grading threshold, wherein the first grading threshold is less than the second grading threshold;

[0112] Step S52: When the comprehensive index of invasion potential is less than the first classification threshold, it is determined to be classified as low invasion level.

[0113] When the comprehensive index of invasion potential is greater than or equal to the first grading threshold and less than the second grading threshold, it is judged as medium invasiveness.

[0114] When the comprehensive index of invasion potential is greater than or equal to the second classification threshold, it is judged as a high invasion level.

[0115] Step S53: Generate a visualized pathology report. The visualized pathology report includes the invasiveness grading results and a thermal map of the key invasive regions formed by normalizing and weighting the spatial distribution maps of the tumor cell density gradient value along the invasion front, the tumor cell anisotropy index along the white matter fiber bundle direction, the aggregation degree of perivascular tumor cells, and the contact ratio of residual neurons wrapped by tumor cells in the tumor invasion edge region. The weights of the weighting are determined by the attention weight coefficients calculated by the attention fusion layer.

[0116] To enable those skilled in the art to implement the present invention, the following provides an exemplary description of the construction and training process of the attention-based multi-source feature fusion network.

[0117] Training sample construction:

[0118] First, a large number of digital pathological whole-slice images of gliomas, each independently graded for invasiveness by at least two senior pathologists, were collected. The consensus-based grading (low-invasive, intermediate-invasive, high-invasive) was used as the gold standard label. For each whole-slice image, following steps S1 to S3, the topological features of the invasive edge and the corresponding spatial distribution statistics were extracted to form a training sample. The input to this training sample consisted of: the tumor cell density gradient along the invasive front, the tumor cell anisotropy index along the white matter fiber bundles, the aggregation degree of perivascular tumor cells, the contact ratio of residual neurons encased in tumor cells, and the maximum, minimum, and standard deviation of each of the above four features. The expected output of this training sample was the target value of the comprehensive invasive potential index corresponding to the gold standard grading.

[0119] Network training and loss function:

[0120] The multi-source feature fusion network is trained using an end-to-end supervised learning approach. Since the invasive classification task is essentially an ordered multi-classification problem, the cross-entropy loss function commonly used in classification tasks or the mean squared error loss function for ordered regression problems is preferably used during training.

[0121] In one specific embodiment, we map the hierarchical labels (low, medium, high) to target values ​​of the comprehensive invasion potential index (e.g., 0.1, 0.5, 0.9), and use mean squared error as the loss function to measure the difference between the network's predicted values ​​and the target values. All weight parameters in the network are iteratively updated using the backpropagation algorithm and the Adam optimizer. During training, strategies such as early stopping, Dropout regularization, and learning rate decay can be employed to prevent overfitting and improve the model's generalization ability. After sufficient training on a large-scale dataset, the attention fusion layer in the network can adaptively learn the differentiated contribution weights of the four topological features to the hierarchical decision, thereby enabling the model to converge.

[0122] Network verification and threshold setting:

[0123] After network training is complete, a separate validation set is used to set the first and second grading thresholds in step S5. Specifically, by adjusting the thresholds, the grading accuracy, sensitivity, and specificity on the validation set are brought to a clinically acceptable level. The network and parameters after training and threshold setting are saved.

[0124] Example 2

[0125] Based on the same inventive concept as in Embodiment 1, such as Figure 3As shown, this embodiment provides a tumor grading intelligent assessment system based on digital pathology, which specifically includes: an image acquisition and standardization module, a tumor region and invasion edge division module, an invasion edge topology feature extraction module, a grading assessment network module, and a grading result output and visualization report generation module.

[0126] The image acquisition and standardization module is used to acquire digital pathological whole-slice images of glioma tissue sections, and to perform background correction and color standardization processing on the digital pathological whole-slice images to obtain standardized whole-slice images.

[0127] The tumor region and invasion edge segmentation module is used to identify and segment the tumor core region based on a standardized whole slice image using a deep learning segmentation network, and to expand the tumor invasion edge region outward by a preset width according to the boundary of the tumor core region.

[0128] The invasion edge topology feature extraction module is used to extract the invasion edge topology feature set reflecting the interaction between tumor cells and brain microenvironment within the tumor invasion edge region. The invasion edge topology feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells.

[0129] The hierarchical evaluation network module is used to store and run a pre-built attention-based multi-source feature fusion network, receive invasion edge topological feature groups and calculate and output a comprehensive invasion potential index;

[0130] The grading result output and visualization report generation module compares the comprehensive invasive potential index with the preset grading threshold range, outputs the invasiveness grading result of glioma, and generates a visualization pathology report marked with key invasive areas by combining the spatial distribution of various features in the invasive edge topological feature group.

[0131] The parameters and steps for implementing the corresponding functions of each unit module in the intelligent assessment system for tumor grading based on digital pathology of the present invention described above can be referred to the parameters and steps in the embodiments of the intelligent assessment method for tumor grading based on digital pathology described above, and will not be repeated here.

[0132] Specific implementation method: First, the digital pathological whole slide image of the glioma tissue section is obtained through the image acquisition and standardization module. The background of the image is corrected and the staining distribution is unified by the color standardization method to obtain a standardized whole slide image.

[0133] Then, the standardized whole-slice image is processed by the tumor region and invasion edge segmentation module as follows: a deep learning-based encoder-decoder segmentation network is used for pixel-level classification to mark the tumor parenchyma region, necrotic region, vascular region, normal brain parenchyma region and residual neuronal cell body region. The boundary contour line of the tumor parenchyma region is extracted, and the boundary is used as a reference to extend a preset pixel width outward from the tumor. The resulting ring-shaped region is taken as the tumor invasion edge region. The ring-shaped region is divided into an inner edge zone near the tumor core and an outer edge zone near the normal brain tissue along the radial direction.

[0134] Secondly, the invasion edge topological feature extraction module extracts the invasion edge topological feature set reflecting the interaction between tumor cells and brain microenvironment in the tumor invasion edge region. Then, the maximum value, minimum value and standard deviation of the above four features in all computing units are calculated as spatial distribution statistics.

[0135] Subsequently, the four feature values ​​and their spatial distribution statistics are input into a pre-constructed multi-source feature fusion network based on the attention mechanism through a hierarchical evaluation network module. Each feature is mapped into a feature vector through four independent feature encoding branches. The attention fusion layer calculates the attention weight coefficients and performs weighted summation to obtain the fused feature vector. Then, the fully connected mapping layer and nonlinear activation unit output the comprehensive index of invasion potential in the range of zero to one.

[0136] Finally, the comprehensive index of invasive potential is compared with the preset first and second categorization thresholds through the categorization result output and visualization report generation module. If it is less than the first categorization threshold, a low invasiveness categorization is output; if it is between the two, a medium invasiveness categorization is output; and if it is greater than or equal to the second categorization threshold, a high invasiveness categorization is output. At the same time, based on the spatial distribution map of the above four features, the attention weight coefficients output by the attention fusion layer are normalized and weighted to generate a heat map of the key invasive areas. The invasiveness categorization results and the heat map are then integrated to form a visualized pathology report.

[0137] Example 3

[0138] Based on the same inventive concept as Embodiment 1, this embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0139] The processor executes the aforementioned intelligent assessment method for tumor grading based on digital pathology by calling the computer program stored in the memory.

[0140] It should be noted that all computer programs for a digital pathology-based intelligent assessment method for tumor grading are implemented using the C programming language.

[0141] Example 4

[0142] Based on the same inventive concept as in Embodiment 1, this embodiment proposes a computer-readable storage medium having an erasable and rewritable computer program stored thereon.

[0143] When a computer program runs on a computer device, it causes the computer device to perform the aforementioned intelligent assessment method for tumor grading based on digital pathology.

[0144] For example, computer-readable storage media can be read-only memory, random access memory, read-only optical disc, magnetic tape, floppy disk, and optical data storage devices.

[0145] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the embodiments for IoT devices and media are relatively simple in description because they are fundamentally similar to the method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0146] The systems, media, and methods provided in the embodiments of the present invention are in one-to-one correspondence. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0152] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0153] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0154] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A digital pathology-based intelligent assessment method for tumor grading, characterized in that, Includes the following steps: Step S1: Obtain a digital pathological whole-slice image of a glioma tissue section, and perform background correction and color standardization processing on the digital pathological whole-slice image to obtain a standardized whole-slice image. Step S2: Based on the standardized whole slice image, a deep learning segmentation network is used to identify and segment the tumor core region, and the tumor invasion edge region is obtained by expanding outward by a preset width according to the boundary of the tumor core region. Step S3: In the tumor invasion edge region, extract the invasion edge topological feature set reflecting the interaction between tumor cells and brain microenvironment. The invasion edge topological feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells. Step S4: Input the invasion edge topological feature group into the pre-constructed multi-source feature fusion network based on the attention mechanism, and calculate and output the comprehensive invasion potential index through the multi-source feature fusion network; Step S5: Compare the comprehensive invasive potential index with the preset grading threshold range, output the invasiveness grading result of the glioma, and generate a visualized pathology report marked with key invasive regions by combining the spatial distribution of various features in the invasive edge topological feature group.

2. The intelligent assessment method for tumor grading based on digital pathology according to claim 1, characterized in that: Step S2 specifically includes the following steps: Step S21: Use a deep learning-based encoder-decoder segmentation network to perform pixel-level classification on the standardized whole-slice image, marking the tumor parenchyma region, necrotic region, vascular region, normal brain parenchyma region, and residual neuronal cell body region. Step S22: Extract the boundary contour of the tumor solid region, and extend the boundary contour outwards from the tumor by a set pixel width. Define the annular region formed after the extension as the tumor invasion edge region. Step S23: Divide the tumor invasion edge region radially into an inner edge zone near the tumor core and an outer edge zone near normal brain tissue.

3. The intelligent assessment method for tumor grading based on digital pathology according to claim 2, characterized in that: The calculation method for the tumor cell density gradient value along the invasion front in step S3 is as follows: First, multiple sampling rays are generated in the region at the edge of the tumor invasion, along a direction perpendicular to the boundary of the tumor core; Then, multiple sampling points are set at equal intervals from the inner edge to the outer edge on each sampling ray, and the ratio of the number of tumor cell nuclei to the sampling area at each sampling point is calculated to obtain the local tumor cell density. Finally, a linear fit was performed on the local tumor cell density and distance on each sampling ray to obtain a gradient value representing the decrease in density as the distance increases. The maximum value of the gradient values ​​of all sampling rays was taken as the tumor cell density gradient value along the invasion front.

4. The intelligent assessment method for tumor grading based on digital pathology according to claim 3, characterized in that: The anisotropy index of tumor cells along the direction of white matter fiber bundles in step S3 is calculated as follows: First, within the tumor invasion margin region, the image is divided into multiple local sub-regions. For each local sub-region, the main direction of local tissue texture is estimated using structural tensor analysis or histogram of orientation gradients, which serves as the direction of white matter fiber bundles in that local sub-region. Then, within each local sub-region, the linear density of tumor cell nuclei in the first direction is calculated along the principal direction, and the linear density of tumor cell nuclei in the second direction is calculated along the direction perpendicular to the principal direction. Finally, the difference between the linear density in the first direction and the linear density in the second direction is divided by the sum of the two to obtain the directional aggregation ratio of the local sub-region. The arithmetic mean of the directional aggregation ratios of all local sub-regions is then calculated to obtain the anisotropy index of tumor cells along the direction of white matter fiber bundles.

5. The intelligent assessment method for tumor grading based on digital pathology according to claim 4, characterized in that: The calculation method for the perivascular tumor cell aggregation degree in step S3 is as follows: First, within the tumor invasion margin region, the contours of all vessel cross-sections are extracted based on the marked vascular regions; Then, for each blood vessel section, an annular region formed by extending outward from the outline of the blood vessel section by a set distance is extracted, and the ratio of the total number of tumor cell nuclei in the annular region to the area of ​​the annular region is calculated as the aggregation index of the blood vessel section. Finally, the arithmetic mean of the aggregation indices of all blood vessel sections is calculated to obtain the aggregation degree of perivascular tumor cells. The contact ratio of residual neurons encapsulated by tumor cells in step S3 is calculated as follows: First, within the tumor invasion edge region, the cell membrane contours of residual neuron cell bodies are extracted based on the marked residual neuron cell body regions. Then, for each residual neuron cell body, a buffer distance is set outside its cell membrane outline. The total length of the arc segment on the cell membrane outline that is within the buffer distance by at least one tumor cell nucleus is calculated and divided by the total perimeter of the neuron cell body cell membrane outline to obtain the encapsulation ratio of a single neuron. Finally, the arithmetic mean of the encapsulation ratios of all residual neurons was calculated to obtain the contact ratio of residual neurons encapsulated by tumor cells.

6. The intelligent assessment method for tumor grading based on digital pathology according to claim 5, characterized in that: The attention-based multi-source feature fusion network in step S4 includes: The first feature encoding branch is used to receive the tumor cell density gradient value and its spatial distribution statistics along the invasion front, and output the first feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the density gradient on all sampling rays; The second feature encoding branch is used to receive the anisotropy index of tumor cells along the direction of white matter fiber bundles and its spatial distribution statistics, and output the second feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the anisotropy index in all local sub-regions; The third feature encoding branch is used to receive the aggregation degree of perivascular tumor cells and their spatial distribution statistics, and output the third feature vector, wherein the spatial distribution statistics include the maximum value, minimum value and standard deviation of the aggregation degree on all vascular sections. The fourth feature encoding branch is used to receive the contact ratio of residual neurons wrapped by tumor cells and its spatial distribution statistics, and outputs the fourth feature vector, where the spatial distribution statistics include the maximum, minimum and standard deviation of the contact ratio on all residual neurons. The attention fusion layer is used to calculate the attention weight coefficients of the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector, and to perform a weighted sum of the four feature vectors based on the attention weight coefficients to obtain the fused feature vector. A fully connected mapping layer is used to map the fused feature vector to a fused feature scalar; Nonlinear activation units are used to map the fused feature scalar to the 0-1 interval through a nonlinear activation function to obtain the comprehensive index of invasion potential.

7. The intelligent assessment method for tumor grading based on digital pathology according to claim 6, characterized in that: Step S5 specifically includes the following steps: Step S51: Set a first grading threshold and a second grading threshold, wherein the first grading threshold is less than the second grading threshold; Step S52: When the comprehensive index of invasion potential is less than the first classification threshold, it is determined to be classified as low invasion level. When the comprehensive index of invasion potential is greater than or equal to the first grading threshold and less than the second grading threshold, it is judged as medium invasiveness. When the comprehensive index of invasion potential is greater than or equal to the second classification threshold, it is judged as a high invasion level. Step S53: Generate a visualized pathology report. The visualized pathology report includes the invasiveness grading results and a thermal map of the key invasive regions formed by normalizing and weighting the spatial distribution maps of the tumor cell density gradient value along the invasion front, the tumor cell anisotropy index along the white matter fiber bundle direction, the aggregation degree of perivascular tumor cells, and the contact ratio of residual neurons wrapped by tumor cells in the tumor invasion edge region. The weights of the weighting are determined by the attention weight coefficients calculated by the attention fusion layer.

8. A digital pathology-based intelligent tumor grading assessment system, used to execute the digital pathology-based intelligent tumor grading assessment method as described in any one of claims 1-7, characterized in that, Specifically, it includes: an image acquisition and standardization module, a tumor region and invasion edge segmentation module, an invasion edge topology feature extraction module, a grading assessment network module, and a grading result output and visualization report generation module; The image acquisition and standardization module is used to acquire digital pathological whole-slice images of glioma tissue sections, and to perform background correction and color standardization processing on the digital pathological whole-slice images to obtain standardized whole-slice images. The tumor region and invasion edge segmentation module is used to identify and segment the tumor core region based on a standardized whole slice image using a deep learning segmentation network, and to expand outward by a preset width according to the boundary of the tumor core region to obtain the tumor invasion edge region. The invasion edge topology feature extraction module is used to extract an invasion edge topology feature set reflecting the interaction between tumor cells and the brain microenvironment within the tumor invasion edge region. The invasion edge topology feature set includes: tumor cell density gradient value along the invasion front, tumor cell anisotropy index along the direction of white matter fiber bundles, perivascular tumor cell aggregation degree, and contact ratio of residual neurons wrapped by tumor cells. The hierarchical evaluation network module is used to store and run a pre-built multi-source feature fusion network based on the attention mechanism, receive the invasion edge topology feature group and calculate and output the comprehensive invasion potential index. The grading result output and visualization report generation module is used to compare the comprehensive invasive potential index with the preset grading threshold range, output the invasiveness grading result of glioma, and generate a visualization pathology report marked with key invasive regions by combining the spatial distribution of various features in the invasive edge topological feature group.

9. An electronic device, comprising: A processor and a memory, wherein the memory stores a computer program that can be called by the processor, characterized in that: the processor executes a digital pathology-based intelligent assessment method for tumor grading as described in any one of claims 1-7 by calling the computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform a digital pathology-based intelligent assessment method for tumor grading as described in any one of claims 1-7.

Images