Dual-module dynamic tandem cascade network system for predicting preoperative t stage of gastric cancer

Abstract

A double-module dynamic series cascade network system for predicting preoperative T stage of gastric cancer belongs to the technical field of medical artificial intelligence. The system adopts a deep learning architecture of double-module dynamic series connection. The first module realizes T1-T4 stage screening based on a hybrid model of parallel CNN and hierarchical Transformer. If it is judged as T1-T3 stage, the output result is output, and the second module is not entered. If it is judged as T4 stage, the second module is automatically triggered to perform T4 subtype differentiation task based on ResNet-152 submodel, and the output result is T4a or T4b. The system uses postoperative pathological results as the T stage gold standard, and shows high accuracy and universality in multicenter retrospective and prospective verification. The results show that the macro average AUC of the model in external verification reaches 0.964, the accuracy is 94.4%, and the T4 subtype recognition accuracy is highest, reaching 96.2%. The present application does not depend on labeled data, can significantly improve the accuracy and consistency of preoperative staging of gastric cancer, realize automatic and fine intelligent evaluation, has strong generalization ability and important clinical application value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical image analysis and artificial intelligence, specifically relating to a dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer. Background Technology

[0002] Gastric cancer is one of the malignant tumors with high incidence and mortality rates worldwide. Clinical staging of gastric cancer is crucial for developing treatment plans and predicting prognosis. Traditional gastric cancer staging primarily relies on imaging and endoscopy; however, these methods have limitations in accuracy and reliability. In recent years, with the development of computer vision and deep learning technologies, automated analysis based on image data has gradually become a research hotspot. Some studies have attempted to use CT images for gastric cancer staging, but most methods rely on labeled data, limiting their widespread clinical application.

[0003] Currently, clinical T-staging of gastric cancer primarily relies on imaging findings, especially CT scans. CT images provide information such as the tumor's morphology, size, and relationship to surrounding tissues; however, physicians' interpretation of CT images is often influenced by subjective factors, leading to low accuracy in staging results. Existing imaging staging methods typically require physician experience and judgment, and there are significant differences among physicians. Furthermore, existing automated staging models largely rely on labeled data for training, lacking sufficient generalization ability and struggling to adapt to imaging data from different hospitals and equipment.

[0004] Currently, the accuracy of preoperative T-staging for gastric cancer is generally low. Existing AI models are mostly limited to predicting only certain T-stages, such as early-to-late stage binary classification, or only T1-T4 stages, lacking differentiation between T4a and T4b, which often poses a challenge for clinicians. Due to the lack of effective auxiliary tools, doctors may misjudge the stage, thus affecting the patient's treatment plan and prognosis. Furthermore, existing models lack transparency in clinical application, making it difficult for doctors to understand the model's decision-making process, limiting their application in actual clinical practice. Although some studies have shown that AI assistance can improve the diagnostic performance of radiologists, the overall accuracy is still lower than that of the model itself, demonstrating the current limitations of technology in clinical decision support. Therefore, developing an end-to-end model based on label-free CT images can effectively improve the accuracy of gastric cancer T-staging and has significant clinical implications. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a dual-module dynamic cascaded network system (TRACE) for predicting preoperative T-staging of gastric cancer, thereby solving the technical problems of existing technologies where preoperative staging relies on manual interpretation, is highly subjective, has difficulty in finely distinguishing T4 subtypes, and lacks interpretability and consistency. This invention constructs a dual-channel CNN-Transformer fusion network and a dynamic cascaded structure to predict the main clinical T-staging of gastric cancer in an end-to-end manner, covering T1, T2, T3, T4a, and T4b stages. It is currently the most comprehensive gastric cancer T-staging prediction model, and it also considers accuracy, versatility, interpretability, and clinical application value.

[0006] The technical solution adopted in this invention is: a dual-module dynamic cascaded network system for predicting preoperative T stage of gastric cancer, comprising:

[0007] (1) The first stage multi-classification module is used to receive standardized venous phase CT image data and use a dual-channel CNN-Transformer hybrid architecture to extract features and fuse semantics of the input images to achieve preliminary prediction of gastric cancer T1-T4 stage;

[0008] The working process of the dual-channel CNN-Transformer hybrid architecture in the first-stage multi-classification module is as follows:

[0009] Step 1: Parallel ResNet-50 and DenseNet-121 backbone networks are used to extract global semantic features and local texture features, respectively.

[0010] Step 2: Map the two feature maps, global semantic features and local texture features, into a token sequence after performing 1×1 convolution and flattening operations respectively;

[0011] Step 3: Input the token sequence into the multi-layer Transformer module to model the spatial context relationship and obtain the global embedding vectors respectively. and ;

[0012] Step 4: After stitching the two together, a linear projection layer is used to obtain the fused feature. And add learnable CLS tags;

[0013] Step 5: Further integrate global semantic information through the lightweight Vision Transformer module, and finally achieve T1-T4 four-class classification prediction through the multilayer perceptron (MLP).

[0014] (2) The second-stage subtype subdivision module is automatically triggered when the prediction result of the first stage is T4. Based on the deep residual network, the T4-stage sample is classified into T4a and T4b subtypes, and the final prediction result is output.

[0015] (3) Control and scheduling module, used to dynamically call the second stage model according to the output results of the first stage, so as to realize the adaptive cascading operation of the two modules;

[0016] (4) Results output module, used to output gastric cancer T stage prediction results, interpretability heatmap and dynamic decision curve.

[0017] Furthermore, the process of obtaining the standardized venous phase CT image data is as follows: acquire multi-center enhanced CT image data, and perform noise reduction, resampling and intensity normalization preprocessing on the CT images to obtain standardized venous phase CT image data.

[0018] Furthermore, the second-stage subtype segmentation module uses ResNet-152 as its backbone network structure; it achieves multi-scale feature extraction and high-level semantic modeling through a deep convolution stacked structure, thereby effectively capturing key imaging features such as serous membrane continuity disruption and invasion of adjacent organs.

[0019] The second-stage subtype subdivision module input includes the master CT slice of the T4 stage sample and its two adjacent slices above and below it, in order to introduce spatial context information.

[0020] Furthermore, for the ResNet-152 structure, the original fully connected layers were removed and the number of output channels was adjusted to 2 to adapt to binary classification tasks.

[0021] Furthermore, the control and scheduling module dynamically distributes traffic based on the prediction results of the first stage: when the prediction result is for period T1, T2, or T3, the result is output directly and the process is terminated; when the prediction result is for period T4, the second stage subtype subdivision module is activated.

[0022] Furthermore, the system incorporates a self-supervised pre-training mechanism, dynamic learning rate scheduling, image data augmentation, and early stopping strategy during model training.

[0023] Furthermore, the result output module includes:

[0024] a. Calculate the gradient weights of the convolutional layer feature maps with respect to the predicted class to generate a Grad-CAM heatmap;

[0025] b. By fusing radiomics scores and clinical parameters through a regression model, a dynamic decision curve is generated.

[0026] Furthermore, the Grad-CAM heatmap is used to visualize the region of interest in the model, helping to interpret the depth of gastric cancer invasion and its relationship with adjacent organs.

[0027] The beneficial effects of this invention are as follows: The TRACE system of this invention realizes the automation and refinement of gastric cancer T staging. From the original CT images to T4 subtype identification, no manual annotation and ROI segmentation are required, which greatly reduces human intervention and diagnostic bias.

[0028] The dual-channel CNN-Transformer architecture of this invention fully integrates local texture and global semantic information, significantly improving the model's recognition performance and generalization ability.

[0029] This invention achieves adaptive scheduling of the reasoning process through a dynamic concatenation mechanism, which not only ensures the accuracy of subcategories but also reduces the overall computational complexity.

[0030] The introduction of Grad-CAM visualization heatmaps and dynamic nodal plots in this invention gives the model good interpretability and clinical verifiability. Doctors can intuitively understand the model's focus areas and the contribution of each parameter to the prediction, thus enhancing clinical trust.

[0031] This invention supports the input of multi-center, multi-device CT images, has good transferability and real-time inference performance, and can be widely used in preoperative assessment and clinical decision support systems for gastric cancer, providing reliable technical support for precision medicine.

[0032] This invention provides an end-to-end prediction of the main clinical T stages of gastric cancer, covering T1, T2, T3, T4a, and T4b, making it the most comprehensive gastric cancer T-stage prediction model currently available. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall framework of a dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of Grad-CAM visualization of tumor features in enhanced CT images of T1-T4 stage gastric cancer according to an embodiment of the present invention. In this diagram, A is a Grad-CAM heatmap of representative cases of T1-T4 stage gastric cancer, showing the model's attention distribution on the CT images. B is a Grad-CAM heatmap of representative cases of T4 stage subtypes (T4a and T4b). Colors from red (high attention) to blue (low attention) represent the intensity of model attention.

[0035] Note: Grad-CAM is usually interpreted as gradient-weighted class activation mapping. GC in the diagram stands for Gastric Cancer, a common abbreviation.

[0036] Figure 3This is a comparative graph showing the performance of radiologists, the TRACE model, and doctors assisted by the TRACE model in a prospective cohort. In section A, the balanced accuracy of predictions by 12 radiologists using independent image reading versus the TRACE model-assisted method is shown; the dashed line represents the reference level of the TRACE model. Section B shows the misclassification trends of different types under the three models, including under-staging and over-staging. Section C compares the sensitivity of the three models at each T-stage (T1–T4b). Detailed Implementation

[0037] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0038] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0039] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0040] To achieve the above objectives, the first aspect of the present invention provides a dual-module dynamic cascaded network system for predicting preoperative T stage of gastric cancer, the system operation process including the following steps:

[0041] First, multicenter enhanced CT image data were acquired, and the CT images were preprocessed with noise reduction, resampling, and intensity normalization to obtain standardized venous phase CT images.

[0042] Then, the standardized image is input into a dual-channel deep learning architecture, where ResNet-50 is used to extract global semantic features and DenseNet-121 is used to capture local texture features. The two feature streams are convolved and mapped to a unified token sequence, which is then input into the Transformer module for context modeling. The two convolutional skeletons, ResNet-50 and DenseNet-121, are used in parallel to extract complementary representations from the same gastric CT image.

[0043]

[0044] Where I is the input stomach CT image, Deep feature maps extracted for the ResNet-50 network; This indicates that this is a real number tensor; This represents the number of channels in the ResNet50 feature map, i.e., the depth dimension. The height of the ResNet50 feature map; This represents the width of the ResNet50 feature map.

[0045] This is a deep feature map extracted by the DenseNet-121 network. This indicates that this is a real number tensor; This represents the number of channels in the DenseNet121 feature map, i.e., the depth dimension. The height of the DenseNet121 feature map; This represents the width of the DenseNet121 feature map.

[0046] ResNet-50 captures deep-level semantics through residual connections, while DenseNet-121 enhances sensitivity to local details through dense connections. To unify feature dimensions, both are convolutional and flattened into a sequence of symbols:

[0047]

[0048] The ResNet-50 feature embedding representation obtained after the above transformation is shown below. This is the feature embedding representation for DenseNet-121.

[0049] Conv 1×1 (⋅) indicates that a 1×1 convolution is performed on the extracted feature map to adjust the number of channels. Flatten(⋅): Compresses the convolved 3D feature map into a 1D vector. Permute(⋅): Rearranges the dimensions of the flattened vector.

[0050] Where N is the number of symbols and d is the embedding dimension. .

[0051] The Transformer layer captures long-range dependencies through a multi-head self-attention mechanism, outputting a global embedding vector. and Further input into a lightweight visual Transformer (ViT) is used for fusion to obtain a comprehensive representation that combines global and local features. This fused feature... The multilayer perceptron outputs preliminary prediction results for stages T1 to T4, thus achieving the first stage of coarse classification.

[0052] Each sequence is processed by a multi-layer transformer encoder (6 layers, 8 heads) to model long-range dependencies and global context information:

[0053]

[0054] TransformerEnc(⋅) is a Transformer encoder module containing 6 layers of encoders and 8 attention heads, used to capture long-range dependencies and global context information in sequences. , It is an enhanced sequence representation obtained after processing by the Transformer encoder.

[0055] Then, the local spatial information is aggregated into a compact global embedding vector by the labeled mean pooling:

[0056]

[0057] yes The i-th tag in the sequence, yes The i-th tag in the sequence.

[0058] Cross-modal fusion and hierarchical ViT encoding: Global embedding vector and They are connected and projected into a shared embedded space:

[0059]

[0060] exist After being appended with a learnable CLS token, the merged sequence is fed into a lightweight visual transformer (ViT) encoder with 6 layers and 12 attention points:

[0061]

[0062] The final CLS representation summarizes the global semantics fused from the CNN branch and the Transformer context.

[0063] The CLS token output is fed into a two-layer MLP classifier for four-way T-stage prediction:

[0064]

[0065] The MLP is a two-layer classifier with the following structure: Layer 1: A linear layer that maps the input 768-dimensional vector to 256 dimensions; Activation function: ReLU, introducing a non-linear transformation; Regularization: Dropout (dropout rate 0.3) to prevent overfitting. Layer 2: A linear layer that maps the 256-dimensional vector to 4 dimensions (corresponding to 4 T-stage categories).

[0066] Information Flow Summary: Image input features are extracted using ResNet-50 and DenseNet-121, and CNN feature maps are used for word segmentation (including flattening, permutation, and projection); Transformer Block #1: Context modeling in each CNN stream; Mean Pooling: Obtaining the global embedding vector. Feature connectivity and projection: joint embedding Transformer Block #2 (ViT): Hierarchical integration and attention refinement; MLP classifier: 4-level T-level output.

[0067] When the prediction result of the first stage is for period T1, T2 or T3, the result is output directly and the process is terminated.

[0068] When the first-stage prediction is T4, the system automatically triggers the second-stage fine-tuning model to further subdivide T4 cases into T4a and T4b subtypes. This stage is based on an improved ResNet-152 network structure, with input including the main slice and its adjacent slices. Spatial context information enhances the model's ability to perceive extraserosal infiltration and invasion features of adjacent organs, ultimately outputting a refined T4 subtype determination. The entire process is dynamically managed by a control and scheduling module, enabling the model to adaptively execute multi-stage inference, significantly improving inference efficiency while maintaining discrimination accuracy.

[0069] In some specific implementations, to address the "black box" problem of deep learning models, this invention introduces an interpretability analysis module during the inference phase.

[0070] Grad-CAM heatmaps are generated by calculating the gradient weights of each feature map in the convolutional layer with respect to the predicted category, thereby visualizing the key regions that the model focuses on. The heatmaps visually demonstrate the extent and infiltration boundaries of lesions that the model focuses on during the decision-making process, corroborating the doctor's interpretation and significantly improving the model's transparency and understandability.

[0071] This invention also uses a regression model to screen key variables from high-dimensional deep features extracted from fully connected layers, calculates radiomics scores, and combines them with clinical parameters (such as carcinoembryonic antigen CEA, differentiation degree, etc.) to construct a dynamic nomogram, achieving an organic integration of radiological features and clinical indicators. This nomogram can quantitatively demonstrate the contribution of multidimensional factors to staging results, providing physicians with personalized risk prediction and decision support.

[0072] Figure 1 This is a schematic diagram of the overall architecture of the TRACE model; the first stage demonstrates the preliminary classification model. Gastric cancer CT images are processed by a two-branch convolutional neural network (CNN) and a Transformer structure. ResNet-50 is used to extract local features, and DenseNet-121 is used to capture structural features. After compression by convolutional layers, these features are input into a multi-layer Transformer encoder and fused into a unified representation, thus completing the preliminary staging of T1-T4 stages. The second stage demonstrates the refined model. For cases predicted as T4 in the first stage, a ResNet-152 backbone network is used in conjunction with information from adjacent slices to further distinguish between T4a and T4b subtypes. TRACE aims to simulate the stepwise diagnostic process of clinical staging, achieving automation and precision in gastric cancer staging. CNN stands for Convolutional Neural Network. TRACE is an abbreviation for Two-component Radiology-guided Autonomous CascadeEngine for Gastric T-stagings.

[0073] This system is used to predict the clinical T stage of gastric cancer (including T4a and T4b), improve the accuracy of preoperative T staging assessment for gastric cancer by radiologists and clinicians, and solve the current clinical problem of low accuracy in preoperative T staging.

[0074] Example 1: Intelligent identification of gastric cancer T-staging:

[0075] In a preferred embodiment of the present invention, the first-stage intelligent gastric cancer T-staging recognition module adopts a dual-channel CNN-Transformer hybrid network architecture to perform T1, T2, T3, and T4 four-class classification prediction on the input standardized CT images. Its core idea is to achieve joint modeling of local texture features and global spatial semantic information through parallel convolutional neural networks and hierarchical Transformer modules, thereby realizing multi-scale perception of features at different infiltration depths.

[0076] Specifically, venous phase images are first read from a multicenter prospective CT dataset after normalization and preprocessing. A sample index is then constructed using a custom image dataset to ensure a one-to-one correspondence between images and labels. Each image is then resized, normalized, and tensorized before being input into the model. The model includes two independent feature extraction channels, based on ResNet-50 and DenseNet-121 backbone networks, respectively. This dual-channel design aims to leverage ResNet to extract high-level global semantic information while utilizing DenseNet to capture local texture and subtle structural differences, thereby achieving complementary fusion of global and local features.

[0077] In terms of structural design, each CNN backbone removes the fully connected layers at the ends, retaining only the feature maps output by the convolutional layers. After 1×1 convolutional dimensionality reduction and Flattening operations, the two-dimensional feature maps are mapped to a one-dimensional token sequence, and the dimensional order is adjusted through a custom Permute layer to adapt to the Transformer input format. Subsequently, the sequence feature input is a hierarchical Transformer module consisting of six stacked Transformer Encoders, which captures long-range dependencies and spatial context relationships within the image through a multi-head self-attention mechanism. The feature sequence encoded by the Transformer undergoes global average pooling to obtain embedding vectors, and preliminary class discrimination is achieved through linear layers. This module can enhance the overall discriminative power and robustness of the network while maintaining sensitivity to local features.

[0078] During the model training phase, this invention uses the cross-entropy loss function as the optimization objective and employs the Adam optimizer for parameter updates, with a learning rate set to 1×10⁻⁶. -4 The training data is divided into training and validation sets in an 8:2 ratio. After each training cycle, the accuracy is calculated on the validation set to dynamically evaluate model performance and save the best-performing weight parameters. The entire training process is performed in a GPU environment to accelerate model convergence and ensure computational efficiency for high-dimensional feature extraction.

[0079] During the model fusion phase, the separately trained ResNet-50 and DenseNet-121 networks independently output feature files (i.e., features_resnet50.csv and features_densenet121.csv). The system then precisely aligns and concatenates the two sets of features according to the sample IDs to obtain a joint feature matrix with expanded dimensions. This matrix is ​​input into a custom lightweight visual Transformer (ViT) fusion module for cross-channel feature modeling. The ViT module achieves global information interaction in a six-layer Transformer encoding structure by adding learnable CLS tags, outputting representative embedding vectors, which are then passed through a two-layer multilayer perceptron (MLP) to achieve T1 to T4 four-class classification predictions.

[0080] Finally, this stage outputs the T-stage category (T1, T2, T3, or T4) for each patient, while saving the deep feature vector for use in the second stage's T4a / T4b subtype subdivision model. The classification results of this stage constitute the first-level output of the TRACE system, realizing an end-to-end automated process from raw CT images to preliminary prediction of gastric cancer T-stage, providing structured input for subsequent refined staging and interpretability analysis.

[0081] Example 2: Sample of T4 stage:

[0082] In another preferred embodiment of the present invention, for patient samples identified as T4 stage in the first stage, a T4 subtype subdivision module is further constructed to achieve automatic differentiation between T4a and T4b. This module is also based on a deep convolutional neural network, using ResNet-152 as the backbone network to achieve fine-grained identification of the depth of invasion of the gastric wall serosal layer and the characteristics of involvement of surrounding structures.

[0083] In the data preparation phase, the system first filters out sample images predicted as category T4 based on the output of the first phase, and combines this with manually labeled data to form the T4a / T4b binary classification dataset. All sample images undergo uniform size adjustment, tensor quantization, and normalization preprocessing to ensure scale consistency and distribution stability of the model input. Subsequently, the data is grouped according to the data labeling file T4ab.csv, with the training set and test set distinguished by the group field, achieving independent data validation. A custom data loader CTDataset is used to synchronously read image paths, labels, and metadata, ensuring a strict correspondence between the network input and the ground truth labels.

[0084] In terms of model architecture design, the T4 subtype discrimination module supports two backbone network structures: ResNet-152. It also loads ImageNet pre-trained weights to enhance the model's feature generalization ability. For the ResNet-152 structure, the original fully connected layers are removed and the number of output channels is adjusted to 2 to adapt to binary classification tasks. This module achieves multi-scale feature extraction and high-level semantic modeling through a deep convolutional stacked structure, thereby effectively capturing key imaging features such as serous membrane continuity disruption and invasion of adjacent organs.

[0085] During model training, the Cross-Entropy Loss function is used as the objective function, and the Adam optimizer is used for gradient updates. The initial learning rate is set to 0.001, and the number of training epochs is 200. In each training epoch, the system performs batch training on the training set (batch size is 32) and calculates the training loss in real time to monitor convergence. After each training epoch, the model performs inference on the validation set and outputs the non-cancer and cancer probability values ​​for each image, using a 0.5 threshold to determine the class label. Simultaneously, the system automatically generates prediction result files for the training and test sets (named train_predictions_xxx.csv and test_predictions_xxx.csv, respectively), containing sample paths, ground truth labels, predicted labels, and class probabilities, facilitating subsequent performance evaluation and result traceability.

[0086] Both training and inference in this module are performed in a GPU environment, significantly improving the computational efficiency of the model on high-resolution CT data. After training, the system saves the final weight file in .pth format for clinical validation and subsequent model integration. The final T4a / T4b classification results can be linked with the first-stage T1-T4 output results to form a two-level progressive T-staging prediction system, thereby achieving fully automated grading and identification of gastric cancer T-staging.

[0087] Example 3: Model Evaluation:

[0088] For model evaluation, the system uses Scikit-learn to calculate accuracy, recall, specificity, precision, F1-score, and AUC, and generates a classification report. The results show that the proposed two-stage deep learning system can accurately distinguish between T4a and T4b subtypes, significantly improving the fine-grained accuracy of staging prediction and its clinical decision support value while maintaining model interpretability and clinical usability.

[0089] In embodiments of this invention, the TRACE system comprises two sub-models: a T-stage multi-class sub-model (first stage) and a T4a / T4b binary sub-model (second stage). To systematically evaluate the performance of these two sub-models, this invention conducted comprehensive technical validation on a development cohort (DC), two independent external retrospective validation cohorts (EVC1, EVC2), and a prospective validation cohort (PVC). The performance evaluation of the sub-models primarily focuses on two dimensions: discriminative power and calibration performance. Discriminative power evaluation employs a one-vs-rest strategy to calculate the area under the receiver operating characteristic (AUC) curve for each class, and performs a macro-AUC calculation on the AUCs of each class to reflect the model's overall discriminative ability under class imbalance conditions. Simultaneously, conventional metrics such as classification accuracy, sensitivity, specificity, precision, and F1 score are calculated, and scatter plots or probability distribution heatmaps of the classes are plotted to visually display the model's output distribution and error patterns across different confidence intervals. All discriminant performance metrics were obtained using k-fold cross-validation (k is usually 5) within the development cohort to obtain the mean and 95% confidence interval, and were independently evaluated in external and prospective cohorts to test generalization ability. Statistical significance was tested using the DeLong test to compare AUC differences between different models or cohorts, and the McNemar test or chi-square test was used to compare classification metrics. The significance threshold was set at p < 0.05.

[0090] Calibration performance evaluation aims to quantify the consistency between model-predicted probabilities and actual observed incidence rates. Quantitative analysis is performed using probability calibration curves and average Brier scores. The probability calibration curve is plotted by calculating the average predicted probability and true positive rate within each bin (e.g., 10 equal parts) based on the predicted probability, visually indicating the direction and magnitude of deviation from the model's output probability. The average Brier score serves as a joint measure of overall calibration and accuracy, used to compare model stability across different cohorts. To further improve probabilistic interpretability, post-calibration methods such as isometric or Isotonic / Platt scaling can be applied to the output probabilities in the development set, and the calibration effect can be reproduced on external cohorts to avoid overfitting. All calibration analysis graphs, values, and confidence intervals are recorded and archived with the model version.

[0091] Regarding the attention region analysis of the sub-models, this invention employs interpretable tools such as Gradient Weighted Class Activation Mapping (Grad-CAM) to visualize the internal activations of the two sub-models. The specific process is as follows: during sample prediction, the activation map of the last convolutional layer and the gradient of the corresponding predicted class are extracted. A heatmap is generated through weighted summation and ReLU operation. This heatmap is then upsampled and superimposed back onto the original CT slices for visualization. Figure 2 As shown: A is the Grad-CAM heatmap of representative cases of gastric cancer in stages T1-T4, showing the attention distribution of the model on CT images.

[0092] B shows the Grad-CAM heatmap of representative cases of T4 stage subtypes (T4a and T4b). Colors range from red (high attention) to blue (low attention) to indicate the intensity of model attention. Grad-CAM is a gradient-weighted class activation map. GC in the figure is an abbreviation for Gastric Cancer. To quantitatively assess the clinical rationality of the model's regions of interest, this invention performs pixel-by-pixel or region-by-region overlap statistics between the Grad-CAM-generated heatmap and the regions of interest (ROIs) annotated by radiologists on the same case, calculating indicators such as IoU (intersection-over-union) and Dice coefficient, and using the Kappa coefficient to evaluate the consistency between the model's regions of interest and expert annotations. This type of visualization and consistency evaluation not only demonstrates that the features learned by the model are consistent with clinical interpretation but also serves as an important basis for model iteration and error analysis.

[0093] The comprehensive performance evaluation of TRACE focuses on the end-to-end prediction process, covering the overall performance from input CT to final T staging (including T4a / T4b). Regarding discriminant performance, in addition to the aforementioned macro-average AUC and conventional classification metrics, category-probability heatmaps and confusion matrices are generated to identify easily confused staging categories and analyze the reasons for misclassification (e.g., misclassification as adhesion to adjacent organs or artifact effects). In terms of calibration, cross-cohort comparisons are conducted, reporting Brier scores and calibration curves for different cohorts. Case retrospectives and statistical descriptions of probability biases observed in PVCs are performed to assess the model's applicability and robustness in real-world clinical workflows. To measure the model's clinical auxiliary value, decision curve analysis (DCA) can be further performed if necessary to compare the net benefit differences between TRACE and radiologists or existing clinical tools (e.g., EUS staging), and to quantify the reduction in overdiagnosis or underdiagnosis probability that can be achieved using TRACE.

[0094] In implementing this invention, all code, model weights, feature tables, and evaluation results used in experimental analyses are version-managed and properly archived. The generated intermediate files include model prediction probability files, Grad-CAM heatmaps, calibration charts, confusion matrices, and SHAP visualizations, for reproducibility and auditability. For software implementation, PyTorch or TensorFlow deep learning frameworks are preferred for building and training neural networks. scikit-learn is used for traditional machine learning model and metric calculations. Statistical tests are implemented using statsmodels or R, and the SHAP library is used for attribution explanation. For hardware, a GPU server with CUDA acceleration is preferred to ensure efficient training and batch inference. Overall, the integrated evaluation system described in this invention not only quantifies the discrimination and calibration capabilities of sub-models but also provides rich interpretable evidence through visualization and statistical analysis, thereby effectively supporting the promotion and application of TRACE in multi-center real-world clinical scenarios.

[0095] Example 4: Predictive performance for periods T1-T4:

[0096] This invention has been approved by the ethics review committee and granted an exemption from informed consent (Approval No.: 2024PS1840K). The study included 3,874 patients from three medical centers between January 2019 and June 2025: Liaoning Cancer Hospital (Hospital A), Shengjing Hospital of China Medical University (Hospital B), and Zhejiang Cancer Hospital (Hospital C). Patient recruitment criteria and sample size estimation, including the inclusion criteria, are as follows:

[0097] (1) Histologically confirmed gastric adenocarcinoma (AJCC 8th edition, stage T1-T4); (2) Abdominal CT scan completed within 30 days prior to surgery, with complete imaging data (original DICOM file and corresponding clinical record); (3) Image quality meets diagnostic criteria (slice thickness ≤3 mm, no motion artifacts, sufficient contrast enhancement); (4) No history of gastric surgery, radiotherapy or chemotherapy.

[0098] Exclusion criteria include: (1) unclear tumor boundaries on CT images; (2) receiving preoperative antitumor treatment, such as chemotherapy or radiotherapy; and (3) missing clinical or imaging data.

[0099] According to the research objectives, the samples were divided into five independent cohorts: 1,652 patients from Hospital A were used as the training set (DC) for constructing the T1-T4 multi-class classification model; another 567 T4 patients from Hospital A were selected to form the training set for the T4a / T4b binary classification model (T4 subtype supplementary training cohort, TSC-T4), which had no sample overlap with the DC cohort. 850 patients from Hospital B constituted external validation set 1 (EVC1); an additional 822 patients who visited Hospital B between March and June 2025 were included in the prospective validation set (PVC) to evaluate the model's performance in a real-time clinical setting. 550 patients from Hospital C constituted external validation set 2 (EVC2) to further validate the model's generalization ability across different geographic populations. Notably, independent T4 cases from the DC cohort were also used as an independent training cohort for the T4a / T4b binary classification model, thereby enhancing the independence and robustness of the sub-model training process. T-staging for all patients was determined according to the American Joint Committee on Cancer (AJCC) TNM staging system, 8th edition. Postoperative pathological T-staging was used as the gold standard in model training and multi-stage retrospective validation to ensure the accuracy and reliability of the assessment results. In prospective studies, the primary endpoint was defined as final postoperative pathological validation (for patients who underwent final surgery). Secondary endpoints included concordance with EUS, concordance with multidisciplinary team (MDT) assessments, and predictive stability during short-term follow-up.

[0100] This model demonstrates excellent classification performance in multicenter datasets. As shown in Table 1, the overall accuracy of the model in the Development Cohort (DC) is 97.0%, with high levels of sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for each T stage. The identification of T1 and T4 stages is particularly accurate. In the external validation cohort, the model's accuracy in EVC1 and EVC2 is 94.2% and 90.2%, respectively, with corresponding AUC values ​​of 0.957 and 0.939, demonstrating good generalization ability and robustness. In the prospective validation cohort (PVC), the model's accuracy is 85.9%, with an AUC value of 0.947, further supporting its potential application in real-world clinical scenarios. In summary, the model exhibits excellent discriminative power, calibration, and cross-center robustness in T1-T4 stage classification.

[0101] Table 1. Predictive performance of TRACE for periods T1-T4

[0102]

[0103] Acc, Accuracy; AUC, Area Under the Curve; 95% CI, 95% Confidence Interval; PPV, Positive Predictive Value; NPV, Negative Predictive Value. All performance metrics were calculated on independent validation sets. The macro average is the unweighted average of the categories, and confidence intervals were calculated using the bootstrap method. "-" indicates that the data was not applicable.

[0104] The model achieved an overall accuracy of 97.2% on the training set (TSC-T4), with sensitivity and specificity of 91.7% and 89.7%, respectively. In the independent validation cohort, the model also demonstrated good discriminative ability: accuracies of 95.0%, 86.7%, and 96.2% in DC-T4, EVC1-T4, and EVC2-T4, respectively, showing strong generalization performance. Furthermore, in the prospective validation cohort PVC-T4, the model achieved an accuracy of 84.9%, further validating its generalizability and operational stability in real-world clinical settings. Overall, the T4 subtype fine classification model demonstrated excellent performance in terms of discriminative accuracy, cross-center generalization ability, and clinical applicability, as shown in Table 2.

[0105] Table 2. Predictive performance of TRACE for T4 subtypes

[0106]

[0107] Acc, Accuracy; AUC, Area Under the Curve; 95% CI, 95% Confidence Interval; PPV, Positive Predictive Value; NPV, Negative Predictive Value. All performance metrics were calculated on independent validation sets, and confidence intervals were calculated using the DeLong method.

[0108] Example 5: TRACE assists radiologists in identifying T-staging:

[0109] To further validate the application value of the TRACE model in real-world clinical settings, this study conducted a prospective image interpretation study simulating actual clinical workflows. The prospective dataset included 822 gastric cancer patients scheduled for preoperative contrast-enhanced CT scans. All cases were independently reviewed by 12 gastroenterologists with 8 to 15 years of clinical experience from three tertiary hospitals, and clinical T-staging (cT1-cT4b) was assessed according to standard imaging criteria. After the initial assessment, a "memory clearing period" of at least one month was implemented to avoid recall bias. The same group of physicians then conducted a second assessment of the same cases, aided by predictions from the TRACE model. For cases where the AI ​​and physician judgments differed during the second assessment, the TRACE model results were for reference only, and the final prediction was based on the physician's opinion. This prospective study did not involve radiologists in clinical decision-making.

[0110] Figure 3 In the diagram, A represents the balanced accuracy of predictions made by 12 radiologists using independent image reading and the TRACE model-assisted method. The dashed line represents the reference level of the TRACE model. B shows the misclassification trends of different types under the three models, including under-staging and over-staging. C compares the sensitivity of the three models at each T stage (T1-T4b). The TRACE model showed the highest detection rate across all stages, significantly outperforming independent image reading. TRACE is a cascaded network of a dual-module dynamic cascade architecture for gastric cancer T staging. Figure 3As shown, with AI assistance, the average sensitivity of radiologists significantly improved, and the overall accuracy jumped from 41.2% to 79.6%. Furthermore, the misclassification rate decreased significantly: the proportion of understaging decreased from 33.5% (275 cases) to 11.2% (92 cases), and the proportion of overstaging decreased from 25.3% (208 cases) to 9.2% (76 cases). This indicates that the TRACE model can significantly reduce systematic bias in clinical interpretation and greatly improve the consistency and accuracy of staging judgment. Although the diagnostic performance of both junior and senior radiologists improved significantly with AI assistance (both differences were statistically significant, P<0.05), their overall performance was still lower than the discrimination ability of the TRACE model itself. Specifically, with AI assistance, the overall accuracy of radiologists reached 79.6%, with a weighted Kappa coefficient of 0.735, while the accuracy of the TRACE model was 82.4%, with a weighted Kappa of 0.816. This indicates that although AI can significantly improve the consistency and accuracy of physicians' diagnoses, the model itself still has advantages in terms of discriminative efficacy and stability.

[0111] The above embodiments are only used to illustrate the present invention. All equivalent transformations and improvements made on the basis of the technical solutions of the present invention are within the protection scope of the present invention.

Claims

1. A dual-module dynamic tandem cascade network system for predicting preoperative T staging of gastric cancer, characterized by, include: (1) The first stage multi-classification module is used to receive standardized venous phase CT image data and use a dual-channel CNN-Transformer hybrid architecture to extract features and fuse semantics of the input images to achieve preliminary prediction of gastric cancer T1-T4 stage; The working process of the dual-channel CNN-Transformer hybrid architecture in the first-stage multi-classification module is as follows: Step 1: Parallel ResNet-50 and DenseNet-121 backbone networks are used to extract global semantic features and local texture features, respectively. Step 2: Map the two feature maps, global semantic features and local texture features, into a token sequence after performing 1×1 convolution and flattening operations respectively; Step 3: Input the token sequence into the multi-layer Transformer module to model the spatial context relationship and obtain the global embedding vectors respectively. and ; Step 4: After stitching the two together, a linear projection layer is used to obtain the fused feature. And add learnable CLS tags; Step 5: Further integrate global semantic information through the lightweight Vision Transformer module, and finally achieve T1-T4 four-class classification prediction through the multilayer perceptron (MLP). (2) The second-stage subtype subdivision module is automatically triggered when the prediction result of the first stage is T4. Based on the deep residual network, the T4-stage sample is classified into T4a and T4b subtypes, and the final prediction result is output. (3) Control and scheduling module, used to dynamically call the second stage model according to the output results of the first stage, so as to realize the adaptive cascading operation of the two modules; (4) Results output module, used to output gastric cancer T stage prediction results, interpretability heatmap and dynamic decision curve.

2. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 1, characterized in that: The process of obtaining the standardized venous phase CT image data is as follows: acquire multi-center enhanced CT image data, and perform noise reduction, resampling and intensity normalization preprocessing on the CT images to obtain standardized venous phase CT image data.

3. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 1, characterized in that: The second-stage subtype segmentation module uses ResNet-152 as its backbone network structure. It achieves multi-scale feature extraction and high-level semantic modeling through a deep convolution stacked structure, thereby effectively capturing key radiological signs of serous membrane discontinuity disruption and invasion of adjacent organs. The second-stage subtype subdivision module input includes the master CT slice of the T4 stage sample and its two adjacent slices above and below it, in order to introduce spatial context information.

4. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 3, characterized in that: For the ResNet-152 architecture, the original fully connected layers were removed and the number of output channels was adjusted to 2 to adapt to binary classification tasks.

5. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 1, characterized in that: The control and scheduling module dynamically distributes traffic based on the prediction results of the first stage: when the prediction result is for period T1, T2 or T3, the result is output directly and the process is terminated; when the prediction result is for period T4, the second stage subtype subdivision module is activated.

6. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 1, characterized in that: The system incorporates a self-supervised pre-training mechanism, dynamic learning rate scheduling, image data augmentation, and early stopping strategy during model training.

7. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 1, characterized in that: The result output module includes: a. Calculate the gradient weights of the convolutional layer feature maps with respect to the predicted class to generate a Grad-CAM heatmap; b. By fusing radiomics scores and clinical parameters through a regression model, a dynamic decision curve is generated.

8. The dual-module dynamic cascaded network system for predicting preoperative T staging of gastric cancer according to claim 7, characterized in that: The Grad-CAM heatmap is used to visualize the region of interest in the model and to help interpret the depth of gastric cancer invasion and its relationship with adjacent organs.

Images