A system for realizing precise prognosis prediction of gastric cancer patients based on deep learning

Abstract

The present application relates to pathological section image recognition technical field, specifically to a kind of system for realizing accurate prognosis prediction of gastric cancer patient based on deep learning, to realize the accurate prognosis prediction of gastric cancer cell tumor patient on pathological section level.The implementation package information acquisition module of the system, image processing module, wavelet parallel Mamba model module, result prediction module, explainability module, greatly improve the efficiency of image feature extraction and the accuracy of model prediction.The core target of the present application is to solve the problems of insufficient accuracy, poor result stability, limited model generalization ability and lack of explainability in existing gastric cancer prognosis prediction methods.After strict verification and testing, the wavelet parallel Mamba model of the present application shows excellent performance in predicting the overall survival time of gastric cancer patients in multiple centers, providing reliable support for clinical decision-making and demonstrating its great potential in the application of precision medicine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pathological section image recognition, and particularly relates to a system for realizing precise prognosis prediction of gastric cancer patients based on deep learning. BACKGROUND

[0002] Gastric cancer (GC) remains a major public health challenge, despite improvements in mortality due to better healthcare and lifestyle changes, but it still constitutes a serious threat, and there is an urgent need to strengthen screening and prevention strategies to address this growing public health burden.

[0003] Gastric cancer has significant histological diversity, which is evident in hematoxylin and eosin (H&E) staining, where different cell morphologies and tissue arrangements are observed. Different subtypes of gastric cancer exhibit different structural features, which affect their diagnosis and treatment. For example, tubular adenocarcinoma is the most common subtype, which is usually characterized by irregularly dilated and branching small tubes filled with intraluminal mucus, showing clear glandular structures. In contrast, less cohesive carcinomas, including signet ring cell carcinoma, are characterized by cells that are either solitary or arranged in small aggregates without forming good glands, often containing abundant mucin, which displaces the cell nucleus, forming a typical signet ring appearance. Diffuse gastric cancer is characterized by strong invasion, with tumor cells dispersed throughout the tissue, lacking tissue structure, often accompanied by necrosis and mitotic activity. In addition, mucinous adenocarcinoma is defined by extracellular pools of mucin, which make up a large part of the tumor volume, further complicating its histological classification. The heterogeneity of cell morphology and tissue arrangement between different subtypes emphasizes the complexity of gastric cancer and highlights the need for accurate histopathological assessment to guide effective treatment strategies.

[0004] The prognosis of gastric cancer varies significantly between its different subtypes, each with different histological features that can serve as potential prognostic biomarkers. For example, the EB virus subtype is associated with the best prognosis, while the genome-stable subtype is associated with the worst prognosis. Other subtypes, such as microsatellite instability and chromosomal instability, present intermediate survival rates. Identifying specific morphological features observed in hematoxylin and eosin staining, such as cell arrangement and nuclear atypia, is crucial for predicting patient prognosis. Given the complexity of these histological features, artificial intelligence (AI) is a promising solution to improve diagnostic accuracy and prognosis prediction. By analyzing large datasets of histopathological images, artificial intelligence can potentially reveal subtle patterns that traditional methods may overlook, thereby improving patient stratification according to their expected survival and guiding personalized treatment approaches.

[0005] Despite researchers' increasing application of artificial intelligence techniques, such as convolutional neural networks, multi-instance learning, and visual transformers, to pathological image diagnosis, these methods still face several significant challenges. One major limitation is their ability to effectively analyze the complex features of the immune microenvironment, where the interactions and spatial arrangements between different cell types are crucial for understanding tumor biology. Furthermore, while some AI models excel at capturing long-range dependencies in images, they often require substantial computational resources, potentially hindering their practicality in clinical settings. The sparsity of labeled data further complicates the problem, as the availability of many high-quality labeled datasets is limited, making it difficult to train models that generalize well across a wide range of scenarios. Moreover, the interpretability and transparency of AI-driven decision-making remain critical concerns; clinicians need to understand how these models arrive at their conclusions in order to trust and integrate them into their diagnostic workflows. Addressing these shortcomings is essential to advancing the application of AI in pathology and ensuring its effectiveness in improving patient outcomes.

[0006] Combining multiple omics analyses, particularly through the integration of pathomics and immune infiltration analysis or spatial transcriptomics, provides a powerful approach for effectively analyzing the complex characteristics of the immune microenvironment in cancer. Using techniques such as spatial transcriptomics, researchers can gain deep insights into the cellular composition and spatial organization of immune cells within tumors. This multidimensional analysis can identify specific immune cell types, their interactions, and signaling pathways influencing tumor progression and treatment response. For example, spatial transcriptomics provides gene expression data while preserving two-dimensional cellular location information, enabling a comprehensive understanding of how different regions within the tumor microenvironment participate in the immune response. Furthermore, combining these methods can reveal key prognostic biomarkers by highlighting unique transcriptional features associated with various immune cell states and their spatial arrangements.

[0007] Furthermore, wavelet transform and parallel Mamba modules represent cutting-edge models in the field of artificial intelligence, particularly in image processing and analysis. These models are designed to be user-friendly, enabling researchers and practitioners to easily operate them. They meet the necessary requirements for deployment on cloud platforms and scanners, facilitating seamless integration with existing workflows. Wavelet transform enhances the ability to capture both low- and high-frequency components of images, while Mamba modules optimize the processing of these features through advanced attention mechanisms. The combination of these two technologies not only improves image quality and training convergence but also supports the translation of scientific research into practical applications. This combination makes the analysis of complex datasets more effective, paving the way for advancements in multiple fields such as medical imaging and diagnostics.

[0008] This invention relates to a wavelet-parallel Mamba model module based on full-view digital pathological slices, aiming to achieve accurate prognostic assessment and model interpretability for gastric cancer patients. The model design considers the needs of modern medical image analysis, combining residual connectivity, a parallel Mamba module, wavelet transform, and feature fusion modules to achieve efficient and accurate analysis.

[0009] The wavelet-parallel Mamba model preprocesses pathological slides using wavelet transform, decomposing the image into sub-bands of different scales to extract a variety of features, from minute cellular structures to overall tissue layout. This method enables the model to simultaneously identify gastric cancer-related spread patterns and detect microscopic morphological changes, constructing a rich dataset and enhancing the comprehensiveness and accuracy of the analysis. By introducing parallel processing Mamba layers and a state-space model, the model significantly improves computational efficiency, making it highly suitable for clinical environments requiring rapid and reliable prognostic information. Furthermore, the model combines residual connections and feature fusion modules to effectively maintain data integrity and mitigate the impact of information degradation. The feature fusion module dynamically adjusts the weights of different features to adapt to research objectives and data characteristics, thereby improving the model's flexibility and accuracy in prognostic prediction. This design not only reduces reliance on manual feature engineering but also effectively handles noise and incomplete data, ensuring the model's robustness under various conditions. The ability to automatically generate key features significantly improves the model's detection sensitivity and specificity in gastric cancer outcome prediction, providing strong data support for the development of personalized treatment plans. In summary, the wavelet parallel Mamba model, through advanced technical means, not only optimizes the pathological image analysis process but also provides a more efficient and accurate tool for clinical applications, thus promoting the development of precision medicine. Summary of the Invention

[0010] To address the current problems of low accuracy, poor stability, weak generalization ability, and poor interpretability in prognostic prediction of gastric cancer patients, this study proposes a novel deep learning model architecture called the wavelet parallel Mamba model. This model combines wavelet transform, parallel Mamba layers, residual connections, and feature fusion modules, which can achieve a large receptive field without over-parameterization and effectively capture local and global information of images.

[0011] First, a wavelet-parallel Mamba model based on full-view digital pathological slides was designed for accurate prognostic prediction of gastric cancer patients. The system mainly consists of five modules: information acquisition module, image processing module, wavelet-parallel Mamba model module, result prediction module, and interpretability module.

[0012] The information acquisition module is responsible for performing high-resolution scanning of pathological slides from gastric cancer patients to generate full-view digital pathological slide images containing the tumor area.

[0013] The image processing module preprocesses the full-view image, including image segmentation, slicing, and staining correction, to generate standardized image patches for subsequent analysis.

[0014] The wavelet parallel Mamba model module employs wavelet transform, parallel Mamba layers, residual connections, and feature fusion techniques. Specifically, it includes: (1) Wavelet transform: including discrete wavelet transform, convolution processing, inverse discrete wavelet transform, and the generation of the final feature map. Multi-level wavelet transform is used to decompose the feature map and capture macroscopic and microscopic pathological features in the image. (2) Parallel Mamba layers: including input normalization, data reshaping, Mamba state space model, data splicing, re-normalization, linear projection, and data restoration to the original shape. The state space model enables efficient processing of global features, and a lightweight parallel visual processing layer is used to improve computational efficiency. (3) Residual connection and feature fusion module: including feature map fusion, global average pooling, attention map calculation, feature map weighting, channel dimensionality reduction, attention modulation, and final feature fusion. Combined with residual connections and feature fusion, it ensures the stability of information flow and achieves adaptive feature fusion.

[0015] The outcome prediction module uses negative partial log-likelihood loss to predict patient survival risk and optimize training results based on the fused feature set. The system outputs the C-index score of the predicted score through five-fold cross-validation and external test queue validation.

[0016] The interpretability module outputs attention scores and risk scores from the model. The system achieves model interpretability by performing attention heatmaps, t-SNE dimensionality reduction clustering, and receptive field calculations, as well as multi-omics analyses, including spatial transcriptomics, high-throughput transcriptomics, and whole-exome sequencing, to explore the biological processes related to the risk score. Finally, the output risk score and clinical indicators of gastric cancer patients, including stage, grade, age, and gender, are used to construct a nodal table for predicting patient survival, which is then deployed to a slide scanner and an online platform.

[0017] This application discloses a system for accurately predicting the prognosis of gastric cancer patients using a wavelet parallel Mamba model at the level of routine pathological slides. By combining clinical indicators and risk scores, a nomogram is created that is easy for clinicians to use and understand. It exhibits better consistency, sensitivity, specificity, and interpretability, while significantly reducing the economic burden on patients and contributing to the precision diagnosis and treatment of gastric cancer.

[0018] The system disclosed in this invention can accurately predict the prognosis of gastric cancer patients at the level of conventional pathological slides, and can be deployed on cloud platforms and scanners, reducing the economic burden on patients and helping to promote precision diagnosis and treatment of gastric cancer patients.

[0019] This invention employs three visualization methods to achieve model interpretability, including attention heatmap rendering, t-SNE dimensionality reduction clustering, and receptive field calculation, thereby improving the transparency of model decision-making.

[0020] This invention combines multi-omics data, including spatial transcriptomics, high-throughput transcriptomics, and whole-exome sequencing, overcoming the limitations of artificial intelligence in analyzing the complexity of the immune microenvironment.

[0021] This invention introduces an innovative negative partial log-likelihood survival loss (NLLSurvLoss) to accurately predict patient survival time and perform risk assessment.

[0022] The wavelet parallel Mamba model involved in this invention implements wavelet transform through the WTConv layer, providing an effective multi-frequency response and smoothly expanding as the receiving domain expands.

[0023] The feature fusion module in the wavelet parallel Mamba model involved in this invention solves the potential degradation effect caused by residual connections, ensuring efficient transmission of information flow.

[0024] The wavelet parallel Mamba model involved in this invention employs a PVMLayer for parallel processing of visual Mamba features. It achieves high-performance feature extraction with low computational complexity while maintaining the total number of channels. Attached Figure Description

[0025] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0026] Figure 1 Overall flowchart of prognostic prediction for gastric cancer patients based on wavelet parallel Mamba model;

[0027] Figure 2 Algorithm 1: Algorithm flow for digital pathology slide preprocessing and label information construction;

[0028] Figure 3 Algorithm 2: Development, training, and validation of wavelet parallel Mamba models;

[0029] Figure 4 Diagram of the wavelet parallel Mamba model mechanism;

[0030] Figure 5 Attention heatmap of wavelet parallel Mamba model;

[0031] Figure 6 t-SNE dimensionality reduction clustering graph of wavelet parallel Mamba model;

[0032] Figure 7 Receptive field computation diagram of wavelet parallel Mamba model;

[0033] Figure 8 Nodal table of model risk scores combined with clinical indicators. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0035] The following is a system for accurate prognostic prediction of gastric cancer patients based on deep learning, which is involved in this invention. Figure 1 It includes the following sub-modules:

[0036] S1, Information Collection Module;

[0037] S2, Image Processing Module;

[0038] S3, Wavelet Parallel Mamba Model Module;

[0039] S4, Result Prediction Module;

[0040] S5, Interpretability Module.

[0041] The information collection module includes the following inclusion criteria for gastric cancer patients: the patient's age is greater than 18 years at the time of diagnosis; the gastric cancer diagnosis is confirmed by a qualified pathologist according to the latest World Health Organization classification of digestive system tumors; the patient does not have serious cardiovascular disease or mental illness; the patient has complete follow-up data, and their pathological tissue sections and paraffin blocks are well preserved; the patient's survival status includes death, survival, or censoring, where the starting point for calculating overall survival is the date of pathological diagnosis, and the endpoint is the date of the patient's death or the date of the last follow-up visit.

[0042] The information collection module includes the following information: basic clinical information of gastric cancer patients, covering age, gender, pathological grade, lesion location, medical history, treatment history, family history, and biomarker status (including Ki-67, PD-L1, HER-2, CerB-2, CDX2, P53, etc.); patient follow-up data, including the date of diagnosis, disease progression and time, survival status (death or survival), date of death or last follow-up date; full-field digital pathological slides of pathological tissue after fixation, embedding and staining, image data acquired by a high-magnification digital scanner, and image blocks extracted from them for subsequent analysis.

[0043] The image processing module preprocesses digitized full-view slice images of gastric cancer patients, mainly including: segmentation: removing useless background from the image; slicing: cutting the image into multiple non-overlapping 256×256 pixel image blocks; feature extraction: extracting the feature vector of each image block using pre-trained pathology domain large language models (UNI, CONCH, RESNET).

[0044] Patients were followed up by telephone. Patient survival status was categorized as deceased, alive, or censored. Overall survival was calculated from the date of the patient's pathological diagnosis to the date of death or the last follow-up visit. Follow-up information for all gastric cancer patients in the cohort was obtained via telephone. Gastric cancer patient pathological slides were digitized using a digital pathology slide scanner at 40x magnification. Image patches were extracted from the digital pathology slides to capture complex structural information. (For details, please refer to the corresponding algorithm flowchart.) Figure 2 Design and implement a wavelet parallel Mamba model that fuses local and global information. Local information is captured through a global-to-local spatial aggregation module, while the parallel processing Mamba layer is used to capture global information for feature fusion. (For the wavelet parallel Mamba model development and training algorithm flowchart, please refer to...) Figure 3 ).

[0045] The model designed in this invention is a wavelet-parallel Mamba model focused on digital pathological slide analysis, aiming to solve the prognostic assessment problem of gastric cancer patients. The model design integrates modern artificial intelligence technologies such as wavelet transform, parallel processing Mamba layers, and feature fusion modules, and possesses the following core components (…). Figure 4 ):

[0046] Wavelet Transform: To overcome the problem of excessive parameters caused by increasing the convolution kernel size to simulate the global receiving domain, this invention employs a wavelet transform method. This method can achieve a very large receiving domain without over-parameterization. For a k×k receiving domain, the number of trainable parameters in the algorithm increases only logarithmically with k. The proposed layer, called WTConv, can replace existing architectures, generate effective multi-frequency responses, and gracefully scale with the expansion of the receiving domain.

[0047] Parallel Mamba Layer: This invention introduces a lightweight visual Mamba UNet, in which the parallel Mamba layer can process visual Mamba features in parallel, achieving faster computation speed while keeping the total number of channels constant.

[0048] Residual connections: To enhance the stability and performance of deep network training, residual connections are added to the model. This design allows signals to be directly passed from the previous layer to the next while preserving high-level abstract features;

[0049] Feature fusion: Considering that residual connections may lead to the degradation effect of deep networks, this invention introduces a feature fusion module, which can adaptively combine local and global features to solve the problem of unnatural visual perception.

[0050] The parallel processing Mamba layer, residual connection, wavelet transform, and feature fusion module of the image training module are implemented as follows:

[0051] 1. Wavelet Transform: This includes Discrete Wavelet Transform, Convolution Processing, Inverse Discrete Wavelet Transform, and Generation of the Final Feature Map.

[0052] (1) Discrete wavelet transform:

[0053] (1)

[0054] Given a pair of wavelet filters ( , ) and input data tensor ,in It refers to the batch size. It is the number of channels. and These are the height and width of the image. Here, Indicates the first The low-frequency subband of the layer, This is the filter index, which is convolved with the input data by inversion and translation. This process is applied to all subbands, producing... , , .

[0055] (2) Convolution processing:

[0056] (2)

[0057] Apply a convolution operation to each subband after the wavelet transform. For each subband , , , Perform the above operations. Among them... These are the weights of the convolutional layer. This represents a two-dimensional convolution operation.

[0058] (3) Inverse Discrete Wavelet Transform:

[0059] (3)

[0060] (4)

[0061] Given a pair of wavelet filters ( , These convolved subbands are combined, and an inverse wavelet transform is applied to reconstruct the data. This process is repeated for all subbands to reconstruct the data. , , .

[0062] (4) Generation of the final feature map:

[0063] (4)

[0064] Finally, the inversely transformed subbands are combined back into the original image space and added to the output of the original convolutional layer to generate the final feature map. It is the output of the model. These are the weights of the base convolutional layer. It is the number of layers in the wavelet transform.

[0065] 2. Parallel Mamba Layer: Includes input normalization, data reshaping, Mamba state-space model, concatenating the processed data, re-normalization, linear projection, and data restoration to its original shape.

[0066] (1) Input normalization:

[0067] (5)

[0068] Given input data tensor ,in It refers to the batch size. It is the number of channels. and These are the height and width of the image. and They represent The mean and standard deviation.

[0069] (2) Data Reshaping:

[0070] (6)

[0071] (7)

[0072] Continue to normalize Remodeling ,in It is the total number of pixels in the image, and through Process it to convert To match the format required by the model. (3) Mamba state-space model:

[0073] (8)

[0074] (9)

[0075] The reshaped Divided into four parts Each part is passed The model is processed. It is a state-space model operation. This is the scaling parameter for skipping connections.

[0076] (4) Data after splicing:

[0077] (10)

[0078] Will The four parts of the processed model are then pieced back together.

[0079] (5) Normalize again:

[0080] (11)

[0081] spliced Perform normalization again. and They represent The mean and standard deviation.

[0082] (6) Linear projection:

[0083] (12)

[0084] Normalized data is projected through a linear layer. and These are the weights and biases of the linear layer, respectively.

[0085] (7) Data restoration to its original shape:

[0086] (13)

[0087] Finally, the data is restored to its original image shape.

[0088] 3. Residual connection

[0089] (14)

[0090] In practice, It represents the result of a series of operations such as convolution and non-linear activation functions.

[0091] 4. Feature Fusion Module: Includes feature map fusion, global average pooling, attention map computation, feature map weighting, channel dimensionality reduction, attention modulation, and final feature fusion.

[0092] (1) Feature map fusion:

[0093] (15)

[0094] Given two feature maps and ,in It refers to the batch size. , It is the number of channels in the feature map. , , These are the depth, height, and width of the feature map. These two feature maps are concatenated along the channel direction to generate a new feature map. .

[0095] (2) Global average pooling:

[0096] (16)

[0097] Next, adaptive global average pooling is applied to the concatenated feature maps. This is to extract global context information. Here It is a single vector for each batch.

[0098] (3) Attention map calculation:

[0099] (17)

[0100] Use a convolutional layer and Activation function, computes an attention map This diagram illustrates the importance of each channel. .

[0101] (4) Feature map weighting:

[0102] (18)

[0103] attention map Feature map of splicing Multiplication is used to weight the feature maps. Here... , This indicates element-wise multiplication.

[0104] (5) Channel dimensionality reduction:

[0105] (19)

[0106] The number of channels in the feature map is reduced by another convolutional layer. Here... .

[0107] (6) Attention Modulation:

[0108] (20)

[0109] (twenty one)

[0110] Additionally, through two different convolutional layers and The activation function computes two additional attention maps, which weight the original feature maps respectively. Here , .

[0111] (7) Final feature fusion:

[0112] (twenty two)

[0113] Finally, the feature map after dimensionality reduction of the channel is multiplied by the two weighted feature maps to obtain the final feature fusion result. It is the fused feature map, which is the output of the module.

[0114] 5. The hyperparameter settings for model training are as follows: During model training, this experiment selects the Adam optimizer, sets the learning rate to 0.002, the batch size to 1, and the number of training epochs to 50.

[0115] 6. Model Output: After model training, the fused feature information for each image patch can be output. This feature information contains rich image and tissue information. Then, through a classifier, the predicted hazard value is obtained. Hazards are calculated, and the sigmoid function is used to map the hazard values ​​to the (0, 1) interval. Finally, the survival function is calculated using the cumulative product. In the survival analysis, we use the negative log-likelihood as the loss function based on the Cox proportional hazards model. It measures the difference between the model-predicted hazard value and the actual observed survival time. The specific form of the loss function is as follows:

[0116] (1) Cumulative product of survival functions:

[0117] (twenty three)

[0118] here Indicates the first The probability of survival at the end of a time interval.

[0119] (2) Filling the survival function:

[0120] (twenty four)

[0121] here It is the survival function after filling, taking into account all patients. They survived for the entire period of time.

[0122] (3) Loss of untruncated samples:

[0123] (25)

[0124] here It is the sum of the losses of all untruncated samples.

[0125] (4) Loss due to truncation:

[0126] (26)

[0127] here It is the sum of the losses of all truncated samples.

[0128] After the above training, the wavelet parallel Mamba model involved in this invention achieved accurate prediction of the overall survival of gastric cancer patients in the TCGA-STAD internal validation and ZN-STAD external validation cohorts (Table 1).

[0129] Table 1. Wavelet parallel Mamba model accurately predicted overall survival for gastric cancer patients in the TCGA-STAD internal validation and ZN-STAD external validation cohorts.

[0130] TCGA-STAD ZN-STAD C-Index (%) 72.48±0.068 70.63±0.077

[0131] 7. Model Interpretability Implementation: To achieve model interpretability, the wavelet parallel Mamba model involved in this invention outputs an attention score after training. Using the output attention score, we will improve model interpretability through three visualization methods: attention heatmap, t-SNE dimensionality reduction clustering, and receptive field calculation.

[0132] (1) Attention heatmap: The attention heatmap is used to visualize the contribution of different regions of the input image to the model output; for the preprocessed image patch, its attention score is expressed as:

[0133] (27)

[0134] in It is a location Attention score This is the calculated feature response. Finally, the attention heatmap is obtained by combining the attention score with the original image:

[0135] (28)

[0136] in, It is the generated heatmap. The original image is in position Pixel values ​​(see attention heatmap of wavelet parallel Mamba model) Figure 5 ).

[0137] (2) t-SNE dimensionality reduction clustering: t-SNE (t-distributed random neighborhood embedding) is a technique for dimensionality reduction of high-dimensional data. Its goal is to embed data points in a high-dimensional space into a low-dimensional space while maintaining the relative distance between data points. First, prepare the feature matrix extracted from the pathological image patch:

[0138] (29)

[0139] in, The number of image patches, This represents the original feature dimension. Then, t-SNE dimensionality reduction is performed:

[0140] (30)

[0141] in, Indicates at a given point In this case, point The probability of being chosen as a neighbor; Is and point The standard deviation of the relevant Gaussian distribution. In low-dimensional space, t-SNE defines similarity using the following formula:

[0142] (31)

[0143] here, Represents the midpoint in a low-dimensional space and points The similarity. For each image patch Perform K-Means clustering and evaluation:

[0144] (32)

[0145] (33)

[0146] in, express The average distance to other points in the same cluster. express The average distance to the nearest other cluster point is the global evaluation value for all image patches. The mean of the clusters is maximized to select the optimal number of clusters. Then visualize it:

[0147] (34)

[0148] in, It is the generated clustering graph. Each image patch is assigned a cluster label. The original image is in position Pixel values ​​(see t-SNE dimensionality reduction clustering diagram of wavelet parallel Mamba model) Figure 6 ).

[0149] (3) Receptive Field Calculation: The receptive field is the size of the input image region corresponding to one pixel on each feature map in the neural network. First, the model's output is obtained through forward propagation. :

[0150] (35)

[0151] in For the multi-layer transformation of the model, High-dimensional feature tensors extracted from pathological image patches. For batch size, For the number of channels, Image height, The image width is then used for center point activation calculation to obtain the features of the center point. :

[0152] (36)

[0153] in For activation function, This represents the integer division calculation process. Then, the gradient is calculated via backpropagation:

[0154] (37)

[0155] in For feature dimension, The gradient is the center point with respect to the input. Gradient processing and contribution aggregation are then performed:

[0156] (38)

[0157] (39)

[0158] in For the activation function (see the receptive field computation graph of the wavelet parallel Mamba model), see [link / reference]. Figure 7 ).

[0159] 8. Exploration of Prognostic-Related Biological Processes: To explore the prognostic-related biological processes predicted by the model, the wavelet parallel Mamba model involved in this invention outputs an attention score and a risk score after training. Using the output risk score and interpretability results, multi-omics analysis is performed in conjunction with spatial transcriptomics, high-throughput transcriptomics, and whole-exome sequencing.

[0160] (1) Combination of heat map and spatial transcriptome: For the poor prognosis related regions and good prognosis related regions identified by heat map, the characteristic genes of the corresponding regions are obtained by using spatial transcriptome data, thereby identifying the biological processes that are beneficial and harmful to the prognosis of gastric cancer patients.

[0161] (2) Combination of risk score and high-throughput transcriptomics: using Cibersort, EPIC, xCELL, ESTIMATE immune infiltration algorithms to identify immune cell types associated with risk score;

[0162] Enrichment analysis was performed on gene modules co-expressed with risk scores to identify biological processes associated with survival outcomes in gastric cancer patients. Significance was calculated using hypergeometric distribution based on the enrichment analysis of co-expressed gene modules.

[0163] (40)

[0164] in, It is the total number of genes in the background gene set. It is the total number of genes in the target gene set. It is the number of significant genes in the target gene set. It is the number of genes that appear in both sets simultaneously.

[0165] (3) Combination of risk grouping and whole-exome sequencing: Gastric cancer patients were divided into high-risk and low-risk groups based on risk scores to identify differences in mutation landscape between the two groups. Statistical significance analysis was performed using the chi-square test or Fisher's exact test to compare the differences in mutation frequencies between the high-risk and low-risk groups.

[0166] (41)

[0167] in, It is the observed frequency. It is the expected frequency.

[0168] 9. Model Integration with Clinical Indicators and Deployment: To achieve model deployment and clinical translation, the wavelet parallel Mamba model involved in this invention outputs an attention score after training. Using the output risk score, we construct a nomogram for predicting patient survival by combining it with clinical indicators of gastric cancer patients (including stage, grade, age, and gender). This nomogram is then deployed to slide scanners and online platforms for use by clinicians and patients (see the constructed nomogram for details). Figure 8 ).

[0169] The wavelet parallel Mamba model constructed in this invention provides an important tool for predicting the overall survival of gastric cancer patients. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of this invention.

[0170] 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. The above-described embodiments of this invention do not constitute a limitation on the scope of protection of this invention.

Claims

1. A system for accurate prognostic prediction of gastric cancer patients based on deep learning, characterized in that, The system implementation includes the following modular steps: S1, Information Collection Module; S2, Image Processing Module; S3, Wavelet Parallel Mamba Model Module: This model integrates local and global information and optimizes features through a feature fusion module. The model adopts a parallel structure to simultaneously capture microscopic and macroscopic pathological features. The specific model structure includes the following sub-modules: wavelet transform sub-module, dual parallel Mamba sub-module, residual connection sub-module, and feature fusion sub-module. S4. Result prediction module: After the model training is completed, the output image patch fusion features are used to generate predicted risk values ​​using a classifier. The predicted risk values ​​are mapped to [0,1] using the sigmoid function, and the survival function is calculated using the cumulative product. The negative log-likelihood loss is also introduced. S5, Interpretability Module.

2. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, The module S1, namely the information collection module, includes the following: setting inclusion criteria for gastric cancer patients, collecting basic clinical information of gastric cancer patients, and digitizing pathological tissue sections.

3. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, The module S2, namely the image processing module, includes: segmenting, slicing and staining the digitized image to generate standardized image blocks, and using pre-trained pathology domain large language models UNI, CONCH and RESNET to extract the feature vector of each image block for subsequent analysis.

4. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, The module S3, namely the wavelet parallel Mamba model module, includes the following four steps in its wavelet transform submodule: Step 1, Discrete Wavelet Transform: （1） Given a wavelet filter pair , and input data tensors ,in It refers to the batch size. It is the number of channels. and These are the height and width of the image; here, Indicates the first The low-frequency subband of the layer, It is the index of the filter. The filter convolves with the input data by inversion and translation. This process is applied to all subbands, producing... , , ; Step 2, Convolution Processing: （2） Apply a convolution operation to each subband after wavelet transform, for each subband , , , Perform the above operations, where These are the weights of the convolutional layer. Represents a two-dimensional convolution operation; Step 3, Inverse Discrete Wavelet Transform: （3） （4） Given a wavelet filter pair , These convolutional subbands are combined, and inverse wavelet transform is applied to reconstruct the data. This process is repeated to reconstruct all the subbands, resulting in... , , ; Step 4: Generation of the final feature map: （5） Finally, the inversely transformed subbands are combined back into the original image space and added to the output of the original convolutional layer to generate the final feature map. It is the output of the model. These are the weights of the base convolutional layer. It is the number of layers in the wavelet transform.

5. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, The module S3, namely the wavelet parallel Mamba model module, includes the following steps in its dual parallel Mamba submodule: Step 1, Input Normalization: （6） Given input data tensor ,in It refers to the batch size. It is the number of channels. and These are the height and width of the image. and They represent The mean and standard deviation; Step 2, Data Reshaping: （7） （8） Continue to normalize Remodeling ,in It is the total number of pixels in the image, and through Process it to convert To match the format required by the model; Step 3, Mamba state-space model: （9） （10） The reshaped Divided into four parts Each part is passed The model is processed, where It is a state-space model operation. This is the scaling parameter for skipping connections; Step 4: Data after splicing and processing: （11） Will The four parts of the processed model are then pieced back together. Step 5: Normalize again: （12） spliced Perform further normalization. and They represent The mean and standard deviation; Step 6, Linear Projection: （13） The normalized data is projected through a linear layer, where and These are the weights and biases of the linear layer, respectively. Step 7: Restore the data to its original shape: （14） Finally, the data is restored to the original image shape, where For batch size, For the number of channels, Image height, This represents the image width.

6. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, The operation of module S3, namely the wavelet parallel Mamba model module, and its feature fusion submodule consists of the following 7 steps: Step 1, Feature Map Fusion: （15） Given two feature maps and ,in It refers to the batch size. , It is the number of channels in the feature map. , , It consists of the depth, height, and width of the feature maps. These two feature maps are then concatenated along the channel direction to generate a new feature map. ; Step 2, Global Average Pooling: （16） Next, adaptive global average pooling is applied to the concatenated feature maps. To extract global context information, here It is a single vector for each batch; Step 3, Attention Map Calculation: （17） Use a convolutional layer and Activation function, computes an attention map The diagram illustrates the importance of each channel; where ; Step 4, Feature Map Weighting: （18） attention map Feature map of splicing Multiplication, weighting the feature maps, here , This represents element-wise multiplication; Step 5, Channel Dimensionality Reduction: （19） The number of channels in the feature map is reduced by another convolutional layer; here ; Step 6, Attention Modulation: （20） （21） Additionally, through two different convolutional layers and The activation function computes two additional attention maps, which weight the original feature maps respectively. , ; Step 7, Final Feature Fusion: （22） Finally, the feature map after dimensionality reduction of the channel is multiplied by the two weighted feature maps to obtain the final feature fusion result. It is the fused feature map, which is the output of the module.

7. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, Module S5, the interpretability module, enhances model interpretability through three visualization methods: attention heatmap, t-SNE dimensionality reduction clustering, and receptive field. Attention heatmap: An attention heatmap is used to visualize the contribution of different regions of the input image to the model output; for a preprocessed image patch, its attention score is expressed as: （23） in It is a location Attention score It is the calculated feature response; finally, the attention heatmap is obtained by combining the attention score with the original image: （24） in, It is the generated heatmap. The original image is in position Pixel values; t-SNE dimensionality reduction clustering: First, prepare the feature matrix extracted from the pathological image patch: （25） in, The number of image patches, The original feature dimension is used; then t-SNE dimensionality reduction is performed: （26） in, Indicates at a given point In this case, point The probability of being chosen as a neighbor; Is and point The standard deviation of the relevant Gaussian distribution; in low-dimensional space, t-SNE uses the following formula to define similarity: （27） here, Represents the midpoint in a low-dimensional space and points The similarity; for each image patch Perform K-means clustering and evaluation: （28） （29） in, express The average distance to other points in the same cluster. express The average distance to the nearest other cluster point is the global evaluation value for all image patches. The mean of the clusters is maximized to select the optimal number of clusters. Then visualize it: （30） in, It is the generated clustering graph. Each image patch is assigned a cluster label. The original image is in position Pixel values; Receptive field calculation: The receptive field is the size of the input image region corresponding to one pixel on each feature map in the neural network; first, the model undergoes forward propagation to obtain the model's output. : （31） in For the multi-layer transformation of the model, High-dimensional feature tensors extracted from pathological image patches. For batch size, For the number of channels, Image height, The image width is used; then, center point activation calculation is performed to obtain the features of the center point. : （32） in For activation function, This represents the integer division calculation process, followed by backpropagation gradient calculation: （33） in For feature dimension, The gradient of the center point with respect to the input is then used for gradient processing and contribution aggregation. （34） （35） in This is the activation function.

8. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, Module S5, the interpretability module, further includes multi-omics analysis using the risk score output by the model in conjunction with spatial transcriptomics, high-throughput transcriptomics, and whole exome sequencing to further enhance the interpretability of the model from a biological perspective. Specifically, it includes the following four aspects: The combination of risk score and spatial transcriptomics: the higher the risk score, the worse the prognosis of the corresponding patient. The corresponding risk score is visualized in the form of a heat map. For the poor prognosis-related regions and good prognosis-related regions identified by the heat map, the characteristic expression genes of the corresponding regions are obtained by using spatial transcriptomics data, thereby identifying the biological processes that are beneficial and harmful to the prognosis of gastric cancer patients, and thus realizing the interpretability of the model from the gene expression level. The combination of risk scoring and high-throughput transcriptomics: an algorithm that uses four gene transcriptome data (Cibersort, EPIC, xCELL, ESTIMATE) to infer immune cell infiltration, identifies immune cell types associated with risk scores, and thus achieves model interpretability at the level of immune cell infiltration. The combination of risk grouping and whole-exome sequencing: gastric cancer patients are divided into high-risk and low-risk groups based on risk scores, and differences in mutation landscape between the two groups are identified, thereby achieving the interpretability of the model at the gene mutation level. The differences in gene mutations between different risk groups are statistically significant using at least one of the chi-square test and Fisher's exact test.

9. The system for accurate prognostic prediction of gastric cancer patients based on deep learning according to claim 1, characterized in that, We constructed a nodal table to predict patient survival using the risk score output by the model and clinical indicators of gastric cancer patients, and deployed it to slide scanners and online platforms for use by clinicians and patients.

Images