A Modeling Method for Pathological Slice Recognition Based on the Simplicity Complex Neural Network (SMNN)

Abstract

This invention discloses a pathological slide recognition modeling method based on a Simple Complex Neural Network (SNN) model. The method includes: acquiring medical images and inputting them into a pre-trained UNI pathological model as the backbone network to obtain feature vectors; dynamically selecting K Simple Complex Neural Network (SNN) models using the MoE sparse routing mechanism to model high-dimensional manifold structures of specific categories; using fully connected layers to reduce the dimensionality of the feature vectors in the SNN models, inputting the reduced data into a Simple Complex Neural Network (SCM) layer, and evaluating its proximity to manifolds of various categories; weighting the outputs of the activated K SCM models based on combined weights learned through sparse routing, and converting the summed results into category probabilities using SoftMax for output, thus obtaining the recognition result. By modeling high-dimensional manifolds using SCM neural network kernels, this method addresses the shortcomings of existing techniques in complex feature modeling, rejection capability, reliability, cross-disease generalization ability, and high-dimensional distance calculation in pathological slide analysis.

Description

Technical Field

[0001] This invention relates to the field of pathological slide image analysis technology, and in particular to a pathological slide recognition and modeling method based on the Simple ComplexNN model. Background Technology

[0002] With the continuous improvement of digitalization in the healthcare field, AI-based pathology slide recognition technology is gradually becoming an important tool in clinical diagnosis. Pathology slides are the gold standard for diagnosing cancer and many other diseases, and traditional analysis methods rely entirely on pathologists' experience in judging tissue samples. However, this process is usually time-consuming, and the results can vary due to the pathologist's subjective experience, presenting certain limitations. In recent years, the rapid development of artificial intelligence technology has made automated analysis of pathology slides possible. By modeling pathology slide images using deep learning methods, it is possible to efficiently process large-scale data while improving diagnostic accuracy and further reducing the workload of pathologists.

[0003] Artificial intelligence (AI) technology based on pathological slides has demonstrated broad application value in cancer diagnosis. For example, pathological image analysis has become an important tool for decision support in the diagnosis of various cancer types, including breast cancer, lung cancer, and esophageal adenocarcinoma. In practical applications, pathological slide analysis not only helps detect cancerous areas but also supports precision medicine, such as the development of personalized treatment plans for patients. Therefore, AI analysis technology based on pathological slides has significant research value in clinical practice. However, the high dimensionality, high resolution, and complex structure of pathological slide images present numerous challenges to their automated analysis. These technical difficulties include how to efficiently model the complex features of pathological slides, how to handle the rejection of samples from unseen categories, and how to enhance the reliability of the model to meet the needs of clinical environments. Therefore, researching downstream task models that can solve these problems is of great significance for promoting the application of pathological slide analysis technology in real-world scenarios.

[0004] In recent years, various technical solutions have been proposed for the automated analysis of pathological slide images. Most of these technologies are based on deep learning methods, combining the image characteristics of pathological slides to construct a complete framework from feature extraction to downstream task applications. For example, in feature extraction, large-scale models based on self-supervised learning have made some progress in pathological image analysis. The UNI model, a representative of basic pathology models, can extract general features of cancer slides by pre-training on large-scale unlabeled data, improving the feature representation capabilities for downstream tasks. Furthermore, the CONCH model, through contrastive learning, has achieved performance improvements in multiple tasks such as classification, segmentation, and image annotation; the Prov-GigaPath model and the Virchow model have also performed well in full-slide pathological image analysis and pan-cancer detection tasks, respectively. The success of these basic models demonstrates that pre-training strategies based on large-scale pathological data can provide strong support for multi-task analysis of pathological slides.

[0005] However, these foundational large-scale models typically focus only on feature extraction, and classification techniques for downstream tasks still rely on traditional methods. For example, classic classification methods such as fully connected networks (FC), K-nearest neighbors (KNN), and ProtoNet are often used to generate classification results from features extracted from the foundational models. Furthermore, innovative models proposed in recent years, such as the ROAM model with a pyramidal Transformer structure, have further improved the ability to identify cancer types such as gliomas; Kolmogorov-Arnold Networks (KANs) have demonstrated higher interpretability and performance in small-scale tasks by improving the activation function of multilayer perceptrons. However, these technical solutions still face many limitations when dealing with the practical needs of pathological slide images. The following technical problems exist: 1. Existing models have limited ability to represent complex image features in high-dimensional spaces. Pathological slide images are usually distributed in high-dimensional Euclidean space, with complex manifold structures and significant local distribution differences. Most existing downstream task models use simple feature mapping strategies, which cannot capture the refined feature expressions of samples. In addition, when modeling high-dimensional data distributions, these methods struggle to accurately describe the manifold subspace where sample points reside, resulting in limited feature representation capabilities. 2. Insufficient ability to reject samples of unknown categories. Pathological slide analysis systems may encounter new categories of samples not seen in the training set in clinical settings. Current technologies often fail to wisely reject these samples, tending instead to misclassify them into known categories. This mechanism not only increases the risk of model misdiagnosis but also seriously affects the model's reliability in practical applications. 3. The reliability and interpretability of the model still need improvement. Existing deep learning models generally lack transparency; their decision-making processes are often considered "black boxes," making it difficult to provide doctors with intuitive and reliable support. In particular, while classic neural network models such as multilayer perceptrons have high accuracy, the geometric meaning of their outputs is unclear, making it difficult to gain the complete trust of pathologists. 4. Poor generalization ability across diseases. Current technologies are mostly optimized for specific types of cancer, but often perform poorly on other cancer types or out-of-distribution samples, limiting their wide applicability in real-world clinical scenarios. 5. Calculating the distance from sample points to manifolds in high-dimensional space still presents technical challenges. Although some studies have attempted to model data manifolds using geometric tools (such as simplicities), existing methods mostly remain at the stage of primitive geometric representation and fail to fully integrate modern deep learning techniques for efficient model training. Furthermore, calculating the distance from sample points to the simplicity still relies on iterative algorithms (such as gradient descent), whose efficiency and convergence are poor in high-dimensional scenarios.

[0006] In summary, existing technologies still have shortcomings in areas such as complex feature modeling, rejection capability, reliability, cross-disease generalization ability, and high-dimensional distance calculation in pathological slide analysis. These problems limit the further promotion of pathological slide analysis technology in clinical practice. Summary of the Invention

[0007] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, the purpose of this invention is to propose a pathological slide recognition modeling method based on a Simple Complex NN model. This method models high-dimensional manifolds using simple complex neuron kernels and proposes a deterministic algorithm to solve the high-dimensional distance calculation problem. This addresses the shortcomings of existing technologies in complex feature modeling, rejection capability, reliability, cross-disease generalization ability, and high-dimensional distance calculation in pathological slide analysis.

[0008] To achieve the above objectives, this invention proposes a pathological slide recognition modeling method based on a Simple Complex NN model, comprising: Medical images are acquired and input into a pre-trained UNI pathology model, which serves as the backbone network, to obtain feature vectors. The feature vectors dynamically select K Simple Complex NN models through the MoE sparse routing mechanism to model high-dimensional manifold structures of specific categories; The Simplicial Complex NN model reduces the dimensionality of feature vectors through fully connected layers, and then inputs the reduced data into the Simple Complex Kernel neuron layer to evaluate its similarity to each category of manifold. The combined weights based on sparse route learning are used to weight and summarize the outputs of the activated K Simple Complex NN models. The summation result is then converted into class probabilities using SoftMax and output to obtain the recognition result.

[0009] According to some embodiments of the present invention, a method for constructing a Simple Complex NN model includes: Constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model; A Gaussian activation function for the simple complex neuron nucleus is constructed to quantify the relationship between the input sample points and the simple complex.

[0010] According to some embodiments of the present invention, constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model includes: The simplex is defined as a convex geometry consisting of simplex complexes. An affine independent point is defined, and its mathematical expression is as follows: ; in, It is a d-dimensional simplex; Let be the vertex of the i-th simplex, representing the key feature point of the categorical data; The convex combination coefficients of the vertices of the i-th simplex are used to determine the position of the sample point in the simplex. A simplex complex is composed of multiple simplexes, forming a unified structure by sharing vertices or edges. Its dimension is determined by the highest-dimensional simplex in the set. Generate and optimize simple complex shapes.

[0011] According to some embodiments of the present invention, the generation and optimization of simple complexes includes: Initialization of simplicial complex vertices: From the training data of each category, select d+1 representative key points as the vertices of the initial simplicial complex; the vertex selection is based on the geometric characteristics of the data distribution to ensure that these points can cover the main regions of the category. Optimization process: During training, the position of the vertices of the simplex complex is dynamically adjusted using the backpropagation algorithm to make the simplex complex more closely resemble the true manifold of the class data; the dilatation product is set; and the value of the dilatation product radius r is optimized.

[0012] According to some embodiments of the present invention, constructing a Gaussian activation function for a simple complex neuron nucleus includes: ; in, Use the Gaussian activation function; Input sample points; It is a simple complex corresponding to the category; For sample points The minimum distance to the simplex complex is defined as: ,in, is the Euclidean distance between the sample point and the vertex of the simplex complex; The dilatation radius is used to control the activation range of the Gaussian kernel.

[0013] According to some embodiments of the present invention, setting the expansion product includes: The definition of the dilatation product is as follows: ; in, The dilatational product is the geometric region extended from a simple complex manifold. The radius of the dilatation product controls the size of the dilatation region; It is an n-dimensional space.

[0014] According to some embodiments of the present invention, the feature vector dynamically selects K Simple Complex NN models through the MoE sparse routing mechanism, including: Construct an expert pool consisting of several Simple Complex NN models, each focusing on a specific manifold or class modeling task; Generate query vector: Input data According to online searches Mapped to query vector The query vector is used for similarity matching with the key vectors in the expert pool. Expert retrieval and sparse selection: Each Simple Complex NN model is associated with a key vector during construction. , representing its position in the feature space; for the query vector Calculate its dot product with the key vectors in the expert pool: N is the number of experts in the expert pool; sparse routing operators are used. , choice and The relevant K Simple Complex NN models: ;in, T represents the set of indices of the K activated Simple Complex NN models; T is the transpose of the matrix.

[0015] According to some embodiments of the present invention, a weighted summation of the outputs of the activated K Simple Complex NN models is performed based on the combined weights learned from sparse route learning, including: For the K activated Simple Complex NN models, calculate the corresponding routing scores. Routing scores are normalized using an activation function. ; The K activated Simple Complex NN models on the input data Perform calculations and output. ; Output via route score Weighted summation: .

[0016] According to some embodiments of the present invention, it further includes: The overall accuracy of the output model constructed from K Simple Complex NN models is calculated. If the overall accuracy is greater than a preset accuracy threshold, the output model is considered qualified; otherwise, the output model is considered unqualified. The overall accuracy of the output model constructed from K Simple Complex NN models is calculated, including: The output is The evaluation section includes two parts: the output score of the classification task. Output score of the rejection task ,Keep In the form of, This indicates the number of samples for the classification task; the simple complex only applies to samples greater than a threshold. Meaningful; Classification accuracy The calculation is as follows: ; ; ; in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is greater than 0, that is, whether there is a case that exceeds the threshold; cls_true represents the number of correctly classified samples in the classification task, based on... Comparison calculation with label; For those who refuse to recognize Rejection accuracy The calculation is as follows: ; ; ; in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is equal to 0; ff_true represents the number of correct samples in the rejection task, through... judge; Overall accuracy : ; in: It is the total number of samples, i.e. .

[0017] This invention proposes a pathological slide recognition modeling method based on the Simple Complex NN model, which has the following beneficial effects: 1. To address the issue of insufficient feature representation capabilities in complex images, this paper proposes a model structure based on simple complex neuron kernels, focusing on the accurate modeling of data manifolds in high-dimensional space, specifically targeting the high-dimensional complexity of pathological slide images. By fitting the subspace distribution of the data manifold using the simple complex geometry, the model can effectively capture fine-grained features of sample points and accurately characterize the complex features of pathological slide images. Furthermore, the model design pays special attention to maintaining efficient feature representation capabilities in scenarios with significant local distribution differences, thereby improving performance in pathological slide classification tasks.

[0018] 2. To address the issue of rejection capability for samples with no known class, a geometrically meaningful rejection mechanism is designed by estimating the probability of a sample's class affiliation based on simple complex distance. This effectively addresses the problem of samples with no known class in the training set. For new class samples that may appear in pathological slide analysis, the model can proactively choose to reject classification, thereby enhancing the model's reliability in practical applications.

[0019] 3. To address the problem of efficiently calculating the distance from sample points to manifolds in high-dimensional space, a deterministic method, distinct from traditional gradient descent algorithms, is proposed. This method specifically addresses the efficient calculation of the distance from sample points to simplicities in high-dimensional spaces. It can efficiently solve for the minimum distance from sample points to simplicities, avoiding the slow convergence problem of gradient descent. This deterministic algorithm has significant advantages in computational efficiency and accuracy, providing technical support for the widespread application of the model in high-dimensional scenarios.

[0020] 4. A dedicated classification framework for downstream tasks of basic pathology models is implemented. Addressing the lack of such a framework in existing basic pathology models (such as UNI and CONCH), a downstream classification framework combining simple complex modeling and a hybrid expert (MoE) routing strategy is designed. This framework allows the model to dynamically learn the combined weights of different simple complexes, adapting to diverse distributions of pathological slide features, thereby significantly improving the accuracy and efficiency of classification tasks and constructing a robust downstream task classification system for pathological slide analysis.

[0021] 5. Enhance the reliability and transparency of the model: By utilizing the geometric modeling characteristics of simple complex neuronal nuclei, the model's decision-making process is given clear geometric meaning, significantly improving the interpretability of the model's output. Through the combination of geometric and topological theories, the model can provide transparent decision-making basis while ensuring high performance, enhancing the reliability of pathological slide analysis technology in practical applications.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] 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: Figure 1 This is a flowchart of a pathological slide recognition and modeling method based on a Simple Complex NN model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a pathological slide recognition modeling method based on a Simple Complex NN model according to an embodiment of the present invention. Detailed Implementation

[0025] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] like Figure 1 As shown, this embodiment of the invention proposes a pathological slide recognition modeling method based on the Simple Complex NN model, including steps S1-S4: S1. Acquire medical images and input them into the UNI pre-trained pathological model, which serves as the backbone network, to obtain feature vectors; S2. The feature vectors dynamically select K Simple Complex NN models through the MoE sparse routing mechanism to model high-dimensional manifold structures of specific categories. The S3 Simple Complex NN model reduces the dimensionality of feature vectors through fully connected layers, and then inputs the reduced data into the Simple Complex Kernel neuron layer to evaluate its proximity to each category of manifold. S4. The combined weights based on sparse route learning are used to weight and summarize the outputs of the activated K Simple Complex NN models. The summation result is then converted into class probabilities using SoftMax and output to obtain the recognition result.

[0027] The working principle of the above technical solution is as follows: K Simple Complex NN models are also K expert networks, each composed of a specialized Simple Complex NN neural network. The UNI model (Universal Pathology Model) is a large-scale pathology model based on self-supervised learning, designed to extract universal features from pathological images as feature inputs for downstream tasks, improving performance in tasks such as classification and segmentation.

[0028] The beneficial effects of the above technical solution are as follows: This invention proposes a pathological slide recognition modeling method based on the Simple Complex NN model for pathological slide classification tasks. All parameters in the model, including the vertices of the simple complex kernel group, the radius of the dilatation product, and the linear weights, are optimized through backpropagation to achieve accurate fitting and classification prediction of high-dimensional manifold distributions. The Simple Complex NN model not only effectively captures the complex manifold distribution of pathological slide data but also significantly improves computational efficiency and classification performance. By modeling high-dimensional manifolds using simple complex neuron kernels, a deterministic algorithm is proposed to solve the high-dimensional distance calculation problem, addressing the shortcomings of existing technologies in complex feature modeling, rejection capability, reliability, cross-disease generalization ability, and high-dimensional distance calculation in pathological slide analysis.

[0029] like Figure 2As shown, Extract Feature means extracting features, i.e., acquiring medical images and inputting them into the UNI pre-trained pathology model as the backbone network to obtain feature vectors. Query means querying, i.e., dynamically selecting K Simple Complex NN models from the feature vectors through the MoE sparse routing mechanism. use top K indices to retrieval experts means using the Top K indexes to retrieve experts. The Top K index can be understood as a ranking mechanism used to quickly retrieve the K most relevant or important experts from a large amount of expert information. The Top K index is a data structure or algorithm used to quickly find the top K elements in a large amount of data. compute slimilarity means calculating similarity. ReduceDimension means dimensionality reduction, i.e., the Simple Complex NN model performs dimensionality reduction processing on the feature vectors through fully connected layers. calculate distance means calculating distance, i.e., inputting the dimensionality-reduced data into the Simple Complex Kernel neuron layer to evaluate its proximity to each category of manifold. weighted sum means weighted summation, i.e., weighted summation of the outputs of the activated K Simple Complex NN models based on the combined weights learned by sparse routing. The Softmax function is an activation function widely used in machine learning, especially in the field of deep learning. It converts the aggregated results into class probabilities and outputs them to obtain the recognition results.

[0030] According to some embodiments of the present invention, a method for constructing a Simple Complex NN model includes: Constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model; A Gaussian activation function for the simple complex neuron nucleus is constructed to quantify the relationship between the input sample points and the simple complex.

[0031] The working principle and beneficial effects of the above technical solution: From a generalized geometric perspective, a neuron can be understood as the following basic operations: ; in, This represents the geometric mapping of the neuron to the input data. For bias terms, This is a non-linear activation function. Neurons of this type typically rely on simple geometric interfaces (such as hyperplanes or hyperspheres) to partition the data. However, this classification method cannot effectively represent the distribution patterns of complex data, especially in high-dimensional spaces where the distribution of data points often forms low-dimensional manifold structures.

[0032] To address this issue, a simple complex neuron kernel is proposed. Based on the simple complex theory in geometry, it constructs a simple complex manifold to fit the distribution of categorical data and combines this with a Gaussian activation function to calculate the distance between the input sample and the manifold, thus achieving manifold fitting in classification tasks. The classification approach of fitting the manifold and its distribution gives each parameter a clear meaning, giving the model rejection capability and strong interpretability. It provides a scientific estimate of the reliability of the output results and avoids blind predictions based on data points it doesn't understand. Therefore, it is more suitable for downstream tasks such as medical imaging, where the reliability of results is strictly required.

[0033] The simplex is the core structure of the simplex neuron nucleus, used to represent the geometric manifold of categorical data. In high-dimensional space... In this context, a simplex complex is a set of simplexes.

[0034] The Gaussian activation function quantifies how close a sample point is to a simplex; the closer the point, the higher the activation value; the farther the point, the faster the activation value decays.

[0035] According to some embodiments of the present invention, constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model includes: The simplex is defined as a convex geometry consisting of simplex complexes. An affine independent point is defined, and its mathematical expression is as follows: ; in, It is a d-dimensional simplex; Let be the vertex of the i-th simplex, representing the key feature point of the categorical data; The convex combination coefficients of the vertices of the i-th simplex are used to determine the position of the sample point in the simplex. A simplex complex is composed of multiple simplexes, forming a unified structure by sharing vertices or edges. Its dimension is determined by the highest-dimensional simplex in the set. Generate and optimize simple complex shapes.

[0036] The working principle and beneficial effects of the above technical solution are as follows: The definition of a simplex ensures its topological consistency: if a simplex belongs to a simplex complex, then all its faces also belong to that simplex complex; if two simplexes intersect in a simplex complex, then their intersection is the common face in the simplex complex. Through this structure, the simplex can represent the geometric manifold distribution of categorical data with a small number of feature points, capturing the complex shape of categorical data in high-dimensional space. This definition allows the simplex to dynamically adjust vertex positions to fit the actual distribution of categorical data. The manifold structure of specific categorical data can be approximated by optimizing vertex positions, thus supporting subsequent manifold modeling. This method can adapt to the actual distribution of categorical data, providing geometric support for the entire model, while making the learning of manifold structures more efficient and robust.

[0037] According to some embodiments of the present invention, the generation and optimization of simple complexes includes: Initialization of simplicial complex vertices: From the training data of each category, select d+1 representative key points as the vertices of the initial simplicial complex; the vertex selection is based on the geometric characteristics of the data distribution to ensure that these points can cover the main regions of the category. Optimization process: During training, the position of the vertices of the simplex complex is dynamically adjusted using the backpropagation algorithm to make the simplex complex more closely resemble the true manifold of the class data; the dilatation product is set; and the value of the dilatation product radius r is optimized.

[0038] The working principle and beneficial effects of the above technical solution are as follows: The generation and optimization of the simplicial complex are key steps in the simplicial complex neuron nucleus. During training, the position of the simplicial complex vertices is dynamically adjusted using the backpropagation algorithm, making the simplicial complex more closely resemble the true manifold of the categorical data. Simultaneously, the value of the dilatational product radius *r* is optimized to ensure the adaptability of the Gaussian activation function, thereby improving the accuracy of the classification boundary. Each category of data uses an independent simplicial complex representation; during training, the simplicial complex corresponding to each category is optimized to ensure the discriminative power between different category manifolds.

[0039] According to some embodiments of the present invention, constructing a Gaussian activation function for a simple complex neuron nucleus includes: ; in, Use the Gaussian activation function; Input sample points; It is a simple complex corresponding to the category; For sample points The minimum distance to the simplex complex is defined as: ,in, is the Euclidean distance between the sample point and the vertex of the simplex complex; The dilatation radius is used to control the activation range of the Gaussian kernel.

[0040] In one embodiment, the simple complex neuron nucleus completes the classification task through the following process: Feature dimensionality reduction: Input features are reduced in dimensionality through a fully connected layer and mapped to the feature space of a simple complex kernel in order to calculate the distance to the manifold.

[0041] Activation value calculation: The activation value is calculated for each class simplex using the Gaussian activation function to measure the proximity of the sample point to the class manifold.

[0042] Category probability normalization: The activation values ​​of all categories are normalized using the SoftMax function to generate a category probability distribution. ; in, This represents the probability that sample x belongs to category c; The final classification result is output based on the normalized probability distribution.

[0043] According to some embodiments of the present invention, setting the expansion product includes: The definition of the dilatation product is as follows: ; in, The dilatational product is the geometric region extended from a simple complex manifold. The radius of the dilatation product controls the size of the dilatation region; It is an n-dimensional space.

[0044] The dilatational product is a form of r-thickening in a metric space. By defining the dilatational product, it's possible to effectively quantify the neighborhood of a simplicial complex manifold and assign higher weights to points closer to that manifold in classification tasks. This geometric expansion property of the dilatational product allows it to more flexibly adapt to the manifold distribution of complex data, while enhancing the robustness of the simplicial complex kernel in classification tasks.

[0045] According to some embodiments of the present invention, the feature vector dynamically selects K Simple Complex NN models through the MoE sparse routing mechanism, including: Construct an expert pool consisting of several Simple Complex NN models, each focusing on a specific manifold or class modeling task; Generate query vector: Input data According to online searches Mapped to query vector The query vector is used for similarity matching with the key vectors in the expert pool. Expert retrieval and sparse selection: Each Simple Complex NN model is associated with a key vector during construction. , representing its position in the feature space; for the query vector Calculate its dot product with the key vectors in the expert pool: N is the number of experts in the expert pool; sparse routing operators are used. , choice and The relevant K Simple Complex NN models: ;in, T represents the set of indices of the K activated Simple Complex NN models; T is the transpose of the matrix.

[0046] The working principle and beneficial effects of the above technical solution: The MoE (Mixture of Experts) routing mechanism significantly improves training and inference efficiency by dynamically selecting some experts from the expert pool. The structural characteristics of the expert pool include: broad feature space coverage: multiple Simple Complex NN models cover various categories or feature distributions; sparse activation mechanism: not all experts participate in the computation simultaneously, but relevant experts are dynamically selected based on the input data.

[0047] The core of MoE routing is to retrieve the most relevant Simple Complex NN from the expert pool through query and sparse routing mechanisms, i.e., to generate query vectors and perform expert retrieval and sparse selection.

[0048] Characteristics of the routing mechanism: Sparse activation: Only K Simple Complex NN models are activated for computation in the expert pool, reducing computational complexity and resource consumption. Dynamic adaptation: The most relevant Simple Complex NN model is dynamically matched through a query mechanism, ensuring that the input data can be modeled most effectively. Through the MoE routing mechanism, efficient scheduling of the expert pool is achieved, and sparse routing ensures the efficiency and flexibility of the inference process, ultimately outputting an aggregated result from K Simple Complex NN models of different dimensions.

[0049] During the training of a Simple Complex Neural Network (SMNN) model, it is necessary to determine the model's key parameters to effectively fit the distribution manifold of the data and complete the classification task. The training process mainly revolves around the following three types of parameters: Vertices of a d-dimensional simplex: The vertices p_i ∈ R^n of each simplex are used to construct the geometry of the d-dimensional simplex manifold.

[0050] The radius of the dilatation product r>0: This parameter determines the size of the buffer zone around the manifold, which measures how close the input data points are to the manifold.

[0051] Linear weights w: used to assign different weights to the classification boundaries of the manifold, ensuring that the manifold can adapt to the complexity of the data distribution.

[0052] To ensure the effectiveness of manifold construction, the goal of training is to project data points of a specified class onto the manifold of that class to the greatest extent possible, covering the distribution of the data classes while minimizing error. Specifically: In the initial training phase, the d+1 vertices of the d-dimensional simplex are randomly initialized without specifying a class.

[0053] The distribution of vertices is optimized by minimizing the cross-entropy loss function, so that it gradually covers the manifold of a specific category.

[0054] During training, the parameters of the model are optimized so that the vertices constituting the simplex can fit the class manifold well and complete the manifold recovery.

[0055] This training process can be roughly understood as fitting a manifold. By optimizing the shape of the simple complex and the radius of the dilatation product, the data distribution of the specified category is covered to the greatest extent, and accurate classification is ultimately achieved.

[0056] According to some embodiments of the present invention, a weighted summation of the outputs of the activated K Simple Complex NN models is performed based on the combined weights learned from sparse route learning, including: For the K activated Simple Complex NN models, calculate the corresponding routing scores. Routing scores are normalized using an activation function. ; The K activated Simple Complex NN models on the input data Perform calculations and output. ; Output via route score Weighted summation: .

[0057] According to some embodiments of the present invention, it further includes: The overall accuracy of the output model constructed from K Simple Complex NN models is calculated. If the overall accuracy is greater than a preset accuracy threshold, the output model is considered qualified; otherwise, the output model is considered unqualified. The overall accuracy of the output model constructed from K Simple Complex NN models is calculated, including: The output is The evaluation section includes two parts: the output score of the classification task. Output score of the rejection task ,Keep In the form of, This indicates the number of samples for the classification task; the simple complex only applies to samples greater than a threshold. Meaningful, and the surfaces can be distinguished; Classification accuracy The calculation is as follows: ; ; ; in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is greater than 0, that is, whether there is a case that exceeds the threshold; cls_true represents the number of correctly classified samples in the classification task, based on... Comparison calculation with label; For those who refuse to recognize Rejection accuracy The calculation is as follows: ; ; ; in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is equal to 0; ff_true represents the number of correct samples in the rejection task, through... judge; Overall accuracy : ; in: It is the total number of samples, i.e. .

[0058] The working principle and beneficial effects of the above technical solution are: to accurately verify whether the output model constructed from K Simple Complex NN models is qualified, thereby improving the accuracy of the model output.

[0059] This invention, by combining topological geometry theory and machine learning techniques, effectively addresses the shortcomings of traditional pathological image analysis methods in areas such as complex feature representation, rejection capability, few-shot learning, and computational efficiency, and has the following significant beneficial effects: 1. Enhanced representation of complex features: Simple Complex NN captures the geometric relationships of class feature points and their manifold distribution in high-dimensional space through simple complex structures, enabling accurate modeling of complex morphological features in pathological sections. It has demonstrated outstanding performance in classification and typing tasks of multi-organ pathological images such as gastric cancer and colorectal cancer.

[0060] 2. Reliable rejection capability: By leveraging the category manifold buffer region defined by the dilatation product, it can effectively determine whether unseen samples belong to the training category. This rejection mechanism is of great value in the detection of rare diseases (such as lymphoma, lung adenocarcinoma, uterine sarcoma, etc.), significantly reducing the misdiagnosis rate and providing more efficient diagnostic support for pathologists.

[0061] 3. The significant advantages of few-shot learning: The Simple Complex NN's ability to fit low-dimensional manifolds in high-dimensional space through its simple complex neuron nuclei gives it excellent learning capabilities even with limited labeled data. The model demonstrates significantly better efficiency and robustness than traditional methods in few-shot scenarios.

[0062] 4. A geometrically meaningful classification mechanism: classification decisions are quantified by the geometric distance between sample points and simple complex manifolds, resulting in results with clear geometric meaning. The transparency and interpretability of the model output support pathologists' understanding and trust, particularly excelling in hierarchical clustering and similarity analysis.

[0063] 5. Its broad applicability and excellent generalization ability demonstrate outstanding performance not only in common pathological tasks such as gastric and colorectal cancer, but also in accurately rejecting and classifying unseen categories (such as rare diseases). Its adaptability and generalization performance indicate that this model can serve as a general-purpose tool for multi-task pathological image analysis.

[0064] 6. Optimization of computational efficiency: By introducing the MoE sparse routing mechanism, the most relevant simple complex network experts are dynamically selected from the expert pool, which significantly reduces computational overhead and improves the efficiency and flexibility of multi-class tasks, making it suitable for large-scale clinical data analysis.

[0065] In summary, by innovatively designing simple complex neuronal nuclei and applying topological geometry theory to pathological image recognition, we have successfully achieved accurate expression of complex features, rare disease rejection, and small-sample learning capabilities. This provides an efficient and reliable solution for the field of pathological image analysis, with significant clinical value and broad application prospects.

[0066] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for pathological slide recognition and modeling based on a Simple Complex NN model, characterized in that, include: Medical images are acquired and input into a pre-trained UNI pathology model, which serves as the backbone network, to obtain feature vectors. The feature vectors dynamically select K Simple Complex NN models through the MoE sparse routing mechanism to model high-dimensional manifold structures of specific classes; The Simplicial Complex NN model reduces the dimensionality of feature vectors through fully connected layers, and then inputs the reduced data into the Simple Complex Kernel neuron layer to evaluate its similarity to each category of manifold. The combined weights based on sparse route learning are used to weight and summarize the outputs of the activated K Simple Complex NN models. The summation result is then converted into class probabilities using SoftMax and output to obtain the recognition result.

2. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 1, characterized in that, Methods for constructing Simplicial Complex NN models include: Constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model; A Gaussian activation function for the simple complex neuron nucleus is constructed to quantify the relationship between the input sample points and the simple complex.

3. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 2, comprising constructing the simple complex of the simple complex neuron nucleus in the Simple Complex NN model, including: The simplex is defined as a convex geometry consisting of simplex complexes. An affine independent point is defined, and its mathematical expression is as follows: ； in, It is a d-dimensional simplex; Let be the vertex of the i-th simplex, representing the key feature point of the categorical data; The convex combination coefficients of the vertices of the i-th simplex are used to determine the position of the sample point in the simplex. A simplex complex is composed of multiple simplexes, forming a unified structure by sharing vertices or edges. Its dimension is determined by the highest-dimensional simplex in the set. Generate and optimize simple complex shapes.

4. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 3, comprising generating and optimizing simple complex shapes, including: Initialization of simplicial complex vertices: From the training data of each category, select d+1 representative key points as the vertices of the initial simplicial complex; the vertex selection is based on the geometric characteristics of the data distribution to ensure that these points can cover the main regions of the category. Optimization process: During training, the position of the vertices of the simplex complex is dynamically adjusted using the backpropagation algorithm to make the simplex complex more closely resemble the true manifold of the class data; the dilatation product is set; and the value of the dilatation product radius r is optimized.

5. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 3, comprising constructing a Gaussian activation function for the simple complex neuron nucleus, including: ； in, Use the Gaussian activation function; Input sample points; It is a simple complex corresponding to the category; For sample points The minimum distance to the simplex complex is defined as: ,in, is the Euclidean distance between the sample point and the vertex of the simplex complex; The dilatation radius is used to control the activation range of the Gaussian kernel.

6. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 5, wherein the dilatation product is defined, including: The definition of the dilatation product is as follows: ； in, The dilatational product is the geometric region extended from a simple complex manifold. The radius of the dilatation product controls the size of the dilatation region; It is an n-dimensional space.

7. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 6, wherein the feature vector dynamically selects K Simple Complex NN models through the MoE sparse routing mechanism, including: Construct an expert pool consisting of several Simple Complex NN models, each focusing on a modeling task for a specific manifold or class; Generate query vector: Input data According to online searches Mapped to query vector The query vector is used for similarity matching with the key vectors in the expert pool. Expert retrieval and sparse selection: Each Simple Complex NN model is associated with a key vector during construction. , representing its position in the feature space; for the query vector Calculate its dot product with the key vectors in the expert pool: N is the number of experts in the expert pool; Use sparse routing operators , choice and The relevant K Simple Complex NN models: ;in, T represents the set of indices of the K activated Simple Complex NN models; T is the transpose of the matrix.

8. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 7, wherein the outputs of the activated K Simple Complex NN models are weighted and summarized based on the combined weights of sparse route learning, including: For the K activated Simple Complex NN models, calculate the corresponding routing scores. ; The routing scores are normalized using an activation function: ; The K activated Simple Complex NN models on the input data Perform calculations and output. ; Output via route score Weighted summation: .

9. The pathological slide recognition modeling method based on the Simple Complex NN model as described in claim 1, further comprising: The overall accuracy of the output model constructed from K Simple Complex NN models is calculated. If the overall accuracy is greater than a preset accuracy threshold, the output model is considered qualified; otherwise, the output model is considered unqualified. The overall accuracy of the output model constructed from K Simple Complex NN models is calculated, including: The output is The evaluation section includes two parts: the output score of the classification task. Output score of the rejection task ,Keep In the form of, This indicates the number of samples for the classification task; the simplex complex only applies to samples greater than a threshold. Meaningful; Classification accuracy The calculation is as follows: ； ； ； in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is greater than 0, that is, whether there is a case that exceeds the threshold; cls_true represents the number of correctly classified samples in the classification task, based on... Comparison calculation with label; For those who refuse to recognize Rejection accuracy The calculation is as follows: ； ； ； in: It is a Boolean condition used to determine... Does the value exceed the threshold? And whether the sum on the last dimension is equal to 0; ff_true represents the number of correct samples in the rejection task, through... judge; Overall accuracy : ； in: It is the total number of samples, i.e. .

Images