A plain scan CT image analysis method and system for gastric cancer screening

Abstract

This invention relates to the field of intelligent medical image analysis technology, and discloses a method and system for analyzing plain CT images for gastric cancer screening. The method uses a deep learning model enhanced with anatomical prior knowledge to locate the gastric region and obtain an image of that region. It constructs a multi-scale three-dimensional feature extraction network targeting the imaging manifestations of gastric cancer, extracting multi-scale three-dimensional feature maps of mucosal enhancement and gastric wall thickening. Based on the continuity of the gastric wall structure, it constructs a voxel-structure dynamic heterogeneous map, and aggregates node information through a graph neural network to obtain an aggregated feature vector. The aggregated feature vector is input into a multilayer perceptron classifier, which outputs a three-classification result: normal, benign lesion, and malignant lesion. This invention solves the problems of inaccurate localization, insufficient feature extraction, and missing structural correlations in plain CT gastric cancer screening through anatomical prior constraints, multi-scale feature fusion, and graph structure modeling, achieving high-precision screening.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, specifically to a method and system for analyzing plain CT images used in gastric cancer screening. Background Technology

[0002] In the field of abdominal disease examination, plain CT scans are widely used due to their non-invasiveness, affordability, and high availability. However, plain CT scans lack contrast enhancement information. When used for gastric cancer screening, the key imaging features of gastric cancer, such as mucosal enhancement and gastric wall thickening, have extremely low contrast with normal tissues, making it difficult to effectively extract relevant features using traditional methods.

[0003] Existing deep learning-based gastric cancer screening methods lack precise localization of the gastric region, leading to a large amount of irrelevant background information interfering with classification decisions. At the same time, existing feature extraction networks mostly use single-scale convolution operations, which cannot simultaneously capture the macroscopic morphological changes of gastric wall thickening and the local texture details of mucosal enhancement. In addition, the gastric wall has a well-defined multi-layered anatomical structure with strong structural continuity between adjacent voxels. Existing methods ignore this structural prior information and treat voxels as independent feature points, failing to fully utilize the spatial correlation of the gastric wall structure, resulting in limited classification accuracy, especially poor performance on minority class samples. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a method and system for analyzing plain CT images for gastric cancer screening, thereby solving the problems of inaccurate localization, insufficient feature extraction, and lack of structural correlation in plain CT gastric cancer screening.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for analyzing plain CT images for gastric cancer screening, the method comprising: Step S1: Based on the deep learning model enhanced by anatomical prior knowledge, locate the stomach region in the plain CT image and obtain the stomach region image; Step S2: Construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations, extract multi-scale three-dimensional feature maps from gastric region images, including gastric cancer imaging manifestations including mucosal enhancement features and gastric wall thickening features; Step S3: Based on the continuity of the stomach wall structure, the voxel features in the multi-scale three-dimensional feature map are associated with the stomach wall structure features to construct a voxel-structure dynamic heterogeneous map. The node information of the dynamic heterogeneous map is aggregated through a graph neural network to obtain the aggregated feature vector of the stomach region. Step S4: Input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

[0006] Preferably, in one possible implementation of the first aspect, the deep learning model based on anatomical prior knowledge enhancement includes a pre-trained gastric anatomy segmentation network and a gastric anatomy atlas prior constraint module. The gastric anatomy segmentation network adopts an encoder-decoder architecture. The encoder extracts multi-level features of the gastric region through three-dimensional convolutional layers and residual connections, while the decoder restores spatial resolution through deconvolutional layers and skip connections. The gastric anatomy atlas prior constraint module aligns the spatial probability distribution of the gastric anatomy atlas with the feature map output by the segmentation network. It enhances the feature response of the gastric wall boundary and mucosal surface through the atlas-guided attention mechanism. The gastric anatomy atlas includes anatomical structural information of the gastric antrum, gastric body, gastric fundus, and each layer of the gastric wall.

[0007] Preferably, in one possible implementation of the first aspect, the gastric anatomical atlas prior constraint module performing the atlas alignment operation includes: Register the gastric anatomical atlas to the space of plain CT images to generate a three-dimensional probability mask of the gastric wall structure; The similarity weights between the probability mask and the segmentation feature map are calculated using a graph attention gating unit, and the features of the stomach wall region in the feature map are selectively enhanced. The atlas attention gating unit uses learnable convolutional kernels to achieve local feature fusion and introduces an atlas prior loss function to constrain the spatial consistency between the segmentation network output and the anatomical atlas. The graph prior loss function is constructed based on the Dice similarity coefficient and boundary distance loss.

[0008] Preferably, in one possible implementation of the first aspect, the multi-scale three-dimensional feature extraction network includes coarse-scale feature extraction branches and fine-scale feature extraction branches arranged in parallel. The coarse-scale feature extraction branch uses a three-dimensional dilated convolutional layer to expand the receptive field and capture macroscopic morphological changes in gastric wall thickening features; The fine-scale feature extraction branch uses a three-dimensional densely connected convolutional layer to extract local texture details of mucosal enhancement features; The output feature maps of the two branches are channel-stitched and spatially aligned by the feature fusion module. The feature fusion module uses an adaptive weight allocation mechanism to dynamically adjust the contribution of features at different scales.

[0009] Preferably, in one possible implementation of the first aspect, the adaptive weight allocation mechanism of the feature fusion module is implemented through a channel attention module and a spatial attention module; The channel attention module calculates the global statistics for each feature channel and generates a channel weight vector, while the spatial attention module calculates the importance of each spatial location in the feature map and generates a spatial weight matrix. The channel weight vector and the spatial weight matrix are multiplied element by element and then weighted and fused with the original feature map to obtain the fused multi-scale three-dimensional feature map. Both the channel attention module and the spatial attention module use fully connected layers and convolutional layers to achieve nonlinear transformations.

[0010] Preferably, in one possible implementation of the first aspect, the process of constructing the voxel-structure dynamic heterogeneity map includes: Voxel node generation and structural edge connection; A voxel node consists of the position coordinates and feature vector of each voxel in a multi-scale 3D feature map; The structural edges are established according to the continuity rules of the gastric wall structure. The continuity rules of the gastric wall structure define the connection relationship between adjacent voxels based on the anatomical layers of the gastric wall, including intramural connections, intramuscular connections, and translayer connections. Each structural edge is assigned a connection strength weight, which is calculated using a joint function of voxel feature similarity and spatial distance.

[0011] Preferably, in one possible implementation of the first aspect, the graph neural network uses a multi-layer graph convolutional network to aggregate node information of a dynamic heterogeneous graph; Each layer of the graph convolutional network performs node feature updates and edge weight updates; Node feature updates are achieved by aggregating the features of neighboring nodes and edge weights, and information flow is controlled by a gated loop unit. Edge weight updates recalculate connection strength weights based on the updated node features, forming a dynamic graph structure; After iterative execution of the multi-layer graph convolutional network, the aggregated feature vector of the stomach region is obtained by performing global average pooling on all node features.

[0012] Preferably, in one possible implementation of the first aspect, the multilayer perceptron classifier includes an input layer, multiple hidden layers, and an output layer; The input layer receives aggregated feature vectors, the hidden layers are processed with batch normalization and dropout regularization, and each hidden layer is followed by a non-linear activation function; The output layer uses the Softmax function to generate the probability distribution of gastric cancer screening classification results, which include three categories: normal, benign lesions, and malignant lesions. The number of neurons in the hidden layers decreases layer by layer, and the output dimension of the last hidden layer matches the number of classification categories.

[0013] Preferably, in one possible implementation of the first aspect, the training process of the multilayer perceptron classifier employs a focus loss function and label smoothing regularization; The focus loss function adjusts the weights of difficult and easy samples to enhance the learning ability for minority class samples; Label smoothing regularization converts the one-hot encoding of the true label into a smooth distribution, preventing the model from overfitting; During training, an adaptive moment estimation algorithm is used to optimize network parameters, and the learning rate is dynamically reduced based on the performance of the validation set.

[0014] In a second aspect, the present invention provides a plain CT image analysis system for gastric cancer screening, the system being used to implement a plain CT image analysis method for gastric cancer screening as described in the first aspect, comprising: The image preprocessing module, based on a deep learning model enhanced with anatomical prior knowledge, locates the stomach region in plain CT images and obtains stomach region images; The multi-scale feature extraction module is used to construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations. It extracts multi-scale three-dimensional feature maps from gastric region images. Gastric cancer imaging manifestations include mucosal enhancement features and gastric wall thickening features. The graph neural network aggregation module is used to associate voxel features in the multi-scale three-dimensional feature map with gastric wall structural features based on the continuity of the gastric wall structure, construct a voxel-structure dynamic heterogeneous map, and obtain the aggregated feature vector of the gastric region by aggregating the node information of the dynamic heterogeneous map through the graph neural network. The classification decision module is used to input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

[0015] The beneficial effects of this invention are as follows: by introducing a deep learning model enhanced by anatomical prior knowledge, combined with a gastric anatomical atlas prior constraint module, and utilizing the atlas-guided attention mechanism to enhance the feature response of the gastric wall boundary and mucosal surface, the accuracy of gastric region localization is improved.

[0016] By constructing parallel coarse-scale and fine-scale 3D feature extraction branches, the macroscopic morphology of gastric wall thickening and the local texture details of mucosal enhancement are captured respectively. Adaptive fusion is achieved by using a channel-space dual attention mechanism, so that multi-scale features can be fully utilized.

[0017] By constructing a voxel-structure dynamic heterogeneous map and using a graph neural network for aggregation, the continuity of the gastric wall structure was fully modeled, enabling a deep correlation between voxel features and anatomical structural features, thereby improving the ability to characterize the imaging manifestations of gastric cancer.

[0018] The classifier employs focus loss and label smoothing regularization to enhance its learning ability for minority class samples and prevent overfitting. The overall scheme achieves high-precision gastric cancer screening classification under plain CT conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This application provides a flowchart of a plain CT image analysis method for gastric cancer screening. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: As Figure 1 As shown, the present invention provides a method for analyzing plain CT images for gastric cancer screening, comprising: Step S1: Based on the deep learning model enhanced by anatomical prior knowledge, locate the stomach region in the plain CT image to obtain the stomach region image.

[0023] In this embodiment, the deep learning model enhanced by anatomical prior knowledge consists of a pre-trained gastric anatomy segmentation network and a gastric anatomy atlas prior constraint module. The gastric anatomy segmentation network adopts an encoder-decoder architecture. The encoder part contains five sequentially connected 3D convolutional downsampling stages. Each stage consists of two consecutive 3D convolutional layers with a kernel size of 3×3×3, a batch normalization layer, and a ReLU activation function, forming a base block. Each stage uses two such base blocks. Residual connections are introduced within each base block. At the end of each stage, a 3D max pooling layer with a stride of 2 and a pooling kernel size of 2×2×2 is used for downsampling, halving the spatial size of the feature map and doubling the number of channels. The initial convolutional layer maps a single-channel CT image to 64 feature channels.

[0024] The decoder also comprises five upsampling stages, symmetrical to the encoder stages. Each stage begins with a 3D deconvolutional layer with a kernel size of 2×2×2 and a stride of 2, doubling the spatial resolution of the input feature map and halving the number of channels. Subsequently, this upsampled feature map is concatenated with the feature map from the corresponding layer of the encoder via skip connections along the channel dimension. The concatenated feature map is then processed by two 3D convolutional blocks with the same structure as the encoder's basic blocks. After the output of the final decoding stage, a 1×1×1 3D convolutional layer is connected to reduce the number of channels to 1, and a 3D probability segmentation map P with the same size as the input image is output through a sigmoid activation function.

[0025] The gastric anatomy atlas prior constraint module integrates spatial prior knowledge of standard gastric anatomy atlases into the training process of the segmentation network. During model training, for each input training plain CT image, a registration algorithm based on normalized mutual information is first used to register the standard atlas to the coordinate space of the CT image, generating a corresponding 3D probability mask M. Atlas attention gating units are embedded in each skip connection path of the segmentation network decoder. For the decoder in the... Feature map of layer and encoder feature fusion and feature maps from the corresponding level of the encoder. First, apply the mask Downsampling to For the same space dimensions, The operation of the graph attention gating unit is as follows: First, for... and Each feature is transformed using a separate 1×1×1 3D convolutional layer. Then, the two transformed feature tensors are concatenated along the channel dimension. Next, another 1×1×1 3D convolutional layer is applied to the concatenated features to compress the number of channels to 1, and a spatial attention weight map is generated using a sigmoid activation function. The weight map is multiplied element-wise with the original encoder features to obtain the features enhanced by the map prior modulation. Finally, With decoder features Channel splicing is performed and the data is fed into the subsequent decoder convolutional block.

[0026] The model is trained and optimized using a composite loss function. The total loss function is... Due to segmentation loss and graph prior constraint loss Weighted summation yields:

[0027] in, The weighting factor is set to 0.5 in this embodiment.

[0028] Segmentation loss A linear combination of binary cross-entropy loss and Dice loss is used:

[0029] in, The probability graph for network prediction is the first The probability value of a voxel. Label the corresponding voxel values ​​(0 or 1) in the mask for the actual stomach region. The total number of voxels.

[0030] Prior Constraints on Graphical Constraints Loss It is used to enhance the spatial consistency between network segmentation results and anatomical atlases, and consists of two parts: Regional Dice Loss Calculate the Dice loss between the predicted region and the map mask region A (after binarization):

[0031] Boundary distance loss : Calculate the predicted boundary voxel set voxel set at the boundary of the map mask The average symmetric surface distance between them is calculated. First, boundary voxels are extracted using morphological gradient operations, and then the average surface distance is calculated. :

[0032] in, Voxel representation to set The Euclidean distance of the nearest voxel in the middle. Voxel representation to set The Euclidean distance to the nearest voxel. Then the boundary distance loss is... .

[0033] Therefore, the prior loss of the map is:

[0034] in As a balance factor, it is set to 1 in this embodiment.

[0035] The model was trained using the Adam optimizer, with an initial learning rate set to... The batch size was set to 2. The training data consisted of 2000 plain CT images with ternary gastric region annotations. Performance was monitored using a validation set during training; the learning rate was multiplied by 0.1 if the validation set loss did not decrease for 10 consecutive epochs. The total number of training epochs was 300. After model training, for new plain CT images, the probability map P output by the network was binarized with a threshold of 0.5, and the largest 3D connected region was extracted to obtain the finally localized gastric region image.

[0036] Step S2: Construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations, extract multi-scale three-dimensional feature maps from gastric region images, including mucosal enhancement features and gastric wall thickening features.

[0037] In this embodiment, a multi-scale 3D feature extraction network is used to extract 3D images of the stomach region. As input, where , , These represent the height, width, and depth of the image, respectively. The network employs a parallel dual-branch architecture consisting of a coarse-scale feature extraction branch and a fine-scale feature extraction branch. The two branches process the input image independently, and their outputs are integrated by a feature fusion module to ultimately generate a fused multi-scale 3D feature map. ,in , ', ' is the spatial size after downsampling. This represents the number of feature channels.

[0038] The coarse-scale feature extraction branch aims to capture macroscopic contextual information characterizing the overall morphological abnormalities and thickening features of the gastric wall. This branch mainly consists of four sequentially connected 3D convolutional modules, denoted as... Each module The specific operation is as follows: for the input feature map First, a convolution kernel with a size of 3×3×3 and a dilatation rate of 100% is applied. A three-dimensional dilated convolutional layer, wherein the void ratio sequence is set to The output of this convolutional layer is then connected to a batch normalization layer and a ReLU activation function, i.e. Next, a second 3D standard convolutional layer with a kernel size of 3×3×3 is applied, followed by batch normalization and the ReLU activation function, to obtain... Finally, downsampling is performed using a 3D max-pooling layer with a kernel size of 2×2×2 and a stride of 2 to output the features of this module. This design allows the convolution kernel to effectively increase the receptive field while keeping the number of parameters constant. By gradually increasing the size, it integrates information from distant voxels, and formally expresses its effective receptive field size and... Positive correlation. The input to this branch. First, the feature map is mapped to a 32-channel feature map through an initial 3D convolutional layer. After being processed step-by-step by the four modules described above, the final output feature map is obtained. .

[0039] The fine-scale feature extraction branch is dedicated to extracting local texture details that characterize mucosal surface reinforcement, roughness, and fine-structure heterogeneity. At the core of this branch are three cascaded 3D densely connected blocks, each containing… A dense layer. Let the first... The input of each densely connected block is , its first The operation for a dense layer is as follows: ,in This is for dimensionality reduction convolution. Next, through Layer transformation, output feature map The final output of this dense layer is This means concatenating the outputs of all previous layers along the channel dimension. A transition layer is placed after each densely connected block, containing... Convolution and Average pooling is used to compress channels and control dimensionality. The growth rate (number of new channels per layer) for this branch is set to... After processing, the final output feature map is obtained. Its spatial dimensions and Consistent.

[0040] The feature fusion module is responsible for... and Adaptive weighted fusion is performed. First, channel splicing is performed: The calculation process for the channel attention submodule is as follows: For... Global spatial average pooling is used to obtain the statistical vector. Channel weights are then generated using two fully connected layers and a sigmoid function. ,in , For ReLU function, The sigmoid function is used. The spatial attention submodule calculation process is as follows: compression along the channel dimension. Obtain spatial summary Then, the spatial weight matrix is ​​generated by the Sigmoid function. Adaptive fusion computation is as follows: ,in This indicates that the channel weight is multiplied element-wise with the spatial weight after broadcasting. This represents element-wise multiplication. Finally, the channels are combined using a 1×1×1 convolution. Number of output channels .

[0041] The parameters of this multi-scale 3D feature extraction network are initialized using the Kaiming normal distribution, i.e., the weights... ,in The layer input connection number. Training uses an adaptive moment estimation optimizer, with the parameter update rule as follows: ,in and For the first and second moment estimates after bias correction, the initial learning rate is... hyperparameters , , The batch size is set to 4. The training objective is supervised by the main loss function of the subsequent classification task, and an auxiliary feature difference loss can be introduced. For example, a loss term based on the Hilbert-Schmidt Independence Criterion can be used to maximize the independence of features in the two branches: The total loss is ,in The weights are adjustable during training to balance the network's performance. The network takes the stomach region image output from previous steps as input and its output... This will be passed to the subsequent graph neural network aggregation module.

[0042] Step S3: Based on the continuity of the stomach wall structure, the voxel features in the multi-scale three-dimensional feature map are associated with the stomach wall structure features to construct a voxel-structure dynamic heterogeneous map. The node information of the dynamic heterogeneous map is then aggregated through a graph neural network to obtain the aggregated feature vector of the stomach region.

[0043] In this embodiment, the input is a multi-scale 3D feature map output from a multi-scale 3D feature extraction network. ,in The output is an aggregated feature vector of the stomach region. .

[0044] First, construct a voxel-structure dynamic heterogeneity graph. The set of nodes in the graph From feature map Each voxel in the text is composed of [a specific unit]. For a spatial location coordinate [a specific value], The voxel, and its corresponding node Includes three-dimensional spatial coordinate attributes With feature vector attributes The feature vector is directly taken from In the corresponding position 3D features. The total number of nodes in the graph is . .

[0045] set of edges The connection is established based on the continuity rules of the stomach wall structure. These rules define the connections between nodes based on the anatomical layers of the stomach wall. Specifically, connections are established within the 26-neighborhood of each voxel node. The connection type is determined according to the anatomical layer to which the node belongs, including intramucosaural connections, intramuscular connections, and cross-layer connections. Anatomical layer information is obtained by binarizing the probability segmentation map output by the pre-trained stomach anatomy segmentation network with a threshold of 0.5. For each node… Its neighboring nodes Establish a directed edge .

[0046] Each edge Includes a learnable connection strength weight The initial value of this weight is calculated by a joint function that comprehensively considers voxel feature similarity and spatial distance. Feature similarity By calculating node feature vectors and The cosine similarity between them is obtained as follows:

[0047] Spatial proximity By calculating node coordinates and The negative exponent of the Euclidean distance between them is obtained as follows:

[0048] Among them, distance scale parameter The initial value of the connection strength weight is calculated as follows: and for each node Normalize all outgoing edge weights so that ,in For nodes The set of neighboring nodes.

[0049] Subsequently, a four-layer graph convolutional network is used to aggregate node information in the constructed voxel-structure dynamic heterogeneous graph. Let the first layer be... The input node feature matrix of the layer graph convolutional network is The edge weight matrix is ,in For the first The feature dimension of the layer, initially It is formed by stacking the original feature vectors of all nodes. The elements are The feature dimensions of each layer are set to... , , , , .

[0050] Each layer of the graph convolutional network performs two operations: node feature update and edge weight update. In the node feature update stage, an aggregation method based on attention mechanisms and gated recurrent units is employed. For nodes... First, calculate its relationship with neighboring nodes. raw attention scores :

[0051] in The weight matrix is ​​a learnable matrix. For learnable attention vectors, This represents a vector concatenation operation; LeakyReLU has a negative slope coefficient of 0.2. It utilizes the current edge weights. The score is modulated, and then the nodes are... The scores of all neighbors (including itself) are softmax normalized to obtain the normalized attention coefficients. :

[0052] Next, the aggregated information is calculated. :

[0053] Subsequently, a gated loop unit is used to control the aggregated information. Current node state Update. Calculate the update gate. Reset door and candidate hidden state :

[0054]

[0055]

[0056] in For the sigmoid function, This represents element-wise multiplication. , , , , , These are learnable parameters. Ultimately, the node... In the Features of the layer The update is as follows:

[0057] During the edge weight update phase, based on the updated node features and Then, recalculate the connection strength weights. First, for each pair of connected nodes... Calculate an original association score :

[0058] in The learnable edge feature transformation matrix. For learnable parameter vectors, For each source node The set of original scores corresponding to all its outgoing edges Perform softmax normalization to obtain the updated dynamic edge weights. :

[0059] This operation enables the graph structure to dynamically adjust the connection strength based on the learned node features, forming a dynamic heterogeneous graph.

[0060] After iterative execution through a four-layer graph convolutional network, the final feature representations of all nodes are obtained. Finally, regarding All node features are subjected to global average pooling along the node dimension to obtain an aggregated feature vector representing the entire stomach region. :

[0061] This 256-dimensional vector It will be used as input to the classifier.

[0062] The training of this neural network module is performed jointly with the entire end-to-end system. The network parameters are optimized using an adaptive moment estimation algorithm, with an initial learning rate of... First-order moment estimation of attenuation rate Second-order moment estimation of attenuation rate During training, the focus loss function is used for classification, and label smoothing regularization is applied to the true labels. For categories containing... A classification task involving multiple categories, using the one-hot vectors of the true labels. Transform into a smooth distribution :

[0063] Where the smoothing parameter Let the model's predicted probability distribution for the samples be... Its true category index is The corresponding predicted probability is Focus loss The calculation is as follows:

[0064] Among them weighting factors , , Focus parameters This loss function supervises the entire end-to-end network during training, with a batch size of 8. The learning rate is scheduled using a decay strategy based on validation set performance; if the validation set loss does not decrease for five consecutive training epochs, the learning rate is multiplied by a factor of 0.5, and the total number of training epochs is 150.

[0065] Step S4: Input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

[0066] In this embodiment, the multilayer perceptron classifier uses the aggregated feature vector obtained in the previous step. The classifier uses a four-layer feedforward neural network structure as input. This includes an input layer, two hidden layers, and an output layer. The input layer directly receives a 256-dimensional aggregated feature vector. The first hidden layer contains 128 neurons and performs a linear transformation on the input. The weight matrix Bias vector After linear transformation, batch normalization is applied to calculate the mean of the batch data. With variance Normalize the output ,in Then scaling and offsetting are performed. Scaling parameters With offset parameter These are learnable parameters. The batch-normalized output is then connected to a modified linear unit activation function. Subsequently, dropout regularization is applied, and probabilistic regularization is used during forward propagation. Randomly The neuron output is set to zero to prevent overfitting.

[0067] Output of the first hidden layer This serves as the input to the second hidden layer. The second hidden layer contains 64 neurons and performs a linear transformation. ,in , Similarly, batch normalization, ReLU activation function, and dropout regularization with a probability of 0.5 are applied sequentially to obtain the output. The number of neurons in the hidden layers decreases from 128 to 64, and the output dimension of the last hidden layer matches the final number of classification categories.

[0068] The output layer receives the output of the second hidden layer. Perform linear transformation ,in , The Softmax function is then applied to convert the linear output into a probability distribution. The calculation formula is: ,in This probability distribution corresponds to three categories: normal, benign lesions, and malignant lesions. The final gastric cancer screening classification result is the category with the highest probability value. .

[0069] The training process of this multilayer perceptron classifier is jointly optimized end-to-end with the aforementioned graph neural network module. Training employs a target function combining a focus loss function and label smoothing regularization. For a single training sample, let its true label vector after label smoothing be... The model predicts the probability distribution as follows: Focus loss The calculation formula is: Among them, the focusing parameters Set to 2.0 to reduce the loss contribution of easily classified samples. Weighting factor For balancing categories, this embodiment is set as follows: , , This assigns higher weights to a minority of classes. Label smoothing regularization performs one-hot encoding of the true labels during training. Smoothing is performed using the following formula: The number of categories Smoothing parameters .

[0070] The network parameters are optimized using an adaptive moment estimation algorithm. This algorithm maintains the parameters... First moment estimation and second-order moment estimation The update rules are as follows: , .in For the current iteration Time loss function with respect to parameters gradient, decay rate , .right and Perform deviation correction: , The parameter update formula is as follows: Initial learning rate Set as The learning rate scheduling employs a dynamic decay strategy based on validation set performance. If the validation set loss does not decrease for five consecutive training epochs, the current learning rate is multiplied by a decay factor of 0.5. The model is trained using a batch size of 8, with a total of 150 training epochs. After training, the network parameters are fixed. For a new plain CT image, the aggregated feature vector extracted is used to output the gastric cancer screening classification result through this forward propagation process.

[0071] Example 2: This invention provides a method for analyzing plain CT images for gastric cancer screening, and also includes an implementation method based on a self-supervised pre-training and progressive feature fusion network.

[0072] Specifically, the multi-scale 3D feature extraction network adopts a coarse-to-fine serial progressive feature pyramid structure, which gradually integrates deep semantic features and shallow detail features through a top-down path, and uses deformable convolution to adaptively adjust the receptive field to align with the anatomical structure of the stomach wall.

[0073] The graph neural network aggregation module uses a multi-head graph attention mechanism to replace the standard graph convolution operation, and enhances the aggregation capability of heterogeneous node information by modeling the relationship between different subspaces.

[0074] In addition, before the overall training of the model, a large number of unlabeled plain CT images are used to perform self-supervised pre-training on the feature extraction network through a contrastive learning paradigm, so as to learn more discriminative gastric image representations and improve the generalization performance and robustness of the final classification model on small-scale labeled datasets.

[0075] Example 3: Experimental results of a plain CT image analysis method for gastric cancer screening provided by this invention. This example uses 1000 plain CT image data for model validation, including 240 pathologically confirmed positive cases of gastric cancer, 160 cases of benign lesions, and 600 normal cases. All images were tested after standardized preprocessing.

[0076] On the test set, the method of this invention achieved an overall accuracy of 84.2% in the three-category classification task of normal, benign, and malignant lesions, with a macro-average AUC of 0.928. Specifically, the sensitivity for malignant lesion detection reached 89.5%, and the specificity reached 91.2%; the sensitivity for benign lesion identification was 78.3%. For the early gastric cancer (T1 stage) subgroup (86 cases), the detection sensitivity reached 79.1%, indicating that this method has a high ability to identify low-contrast early lesions.

[0077] Ablation experiments further validated the roles of each core module: removing the anatomical prior knowledge enhancement module reduced the Dice coefficient for gastric region localization from 0.873 to 0.836; removing the voxel-structure dynamic heterogeneous graph neural network decreased the recall rate for malignant lesion classification by 4.3 percentage points; and removing focal loss and label smoothing regularization increased the error rate for benign / malignant lesion classification by 3.1 percentage points. These results demonstrate that the anatomical prior constraints, graph structure modeling, and optimization strategies proposed in this invention all play a substantial role in improving the performance of plain CT gastric cancer screening, validating the effectiveness and reliability of this method in clinical settings.

[0078] Example 4: This invention provides a plain CT image analysis system for gastric cancer screening, comprising: The image preprocessing module uses a deep learning model enhanced with prior anatomical knowledge to locate the stomach region in a plain CT image and obtain an image of the stomach region.

[0079] The multi-scale feature extraction module is used to construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations. It extracts multi-scale three-dimensional feature maps from gastric region images. Gastric cancer imaging manifestations include mucosal enhancement features and gastric wall thickening features.

[0080] The graph neural network aggregation module is used to associate voxel features in the multi-scale three-dimensional feature map with gastric wall structural features based on the continuity of the gastric wall structure, construct a voxel-structure dynamic heterogeneous map, and obtain the aggregated feature vector of the gastric region by aggregating the node information of the dynamic heterogeneous map through the graph neural network.

[0081] The classification decision module is used to input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

[0082] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for analyzing plain CT images for gastric cancer screening, characterized in that, The method includes: Step S1: Based on the deep learning model enhanced by anatomical prior knowledge, locate the stomach region in the plain CT image and obtain the stomach region image; Step S2: Construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations, extract multi-scale three-dimensional feature maps from gastric region images, including gastric cancer imaging manifestations including mucosal enhancement features and gastric wall thickening features; Step S3: Based on the continuity of the stomach wall structure, the voxel features in the multi-scale three-dimensional feature map are associated with the stomach wall structure features to construct a voxel-structure dynamic heterogeneous map. The node information of the dynamic heterogeneous map is aggregated through a graph neural network to obtain the aggregated feature vector of the stomach region. Step S4: Input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

2. The method for analyzing plain CT images for gastric cancer screening as described in claim 1, characterized in that, The deep learning model based on anatomical prior knowledge enhancement includes a pre-trained gastric anatomy segmentation network and a gastric anatomy atlas prior constraint module. The gastric anatomy segmentation network adopts an encoder-decoder architecture. The encoder extracts multi-level features of the gastric region through three-dimensional convolutional layers and residual connections, while the decoder restores spatial resolution through deconvolutional layers and skip connections. The gastric anatomy atlas prior constraint module aligns the spatial probability distribution of the gastric anatomy atlas with the feature map output by the segmentation network. It enhances the feature response of the gastric wall boundary and mucosal surface through the atlas-guided attention mechanism. The gastric anatomy atlas includes anatomical structural information of the gastric antrum, gastric body, gastric fundus, and each layer of the gastric wall.

3. The plain CT image analysis method for gastric cancer screening as described in claim 2, characterized in that, The prior constraint module for the gastric anatomy atlas performs atlas alignment operations including: Register the gastric anatomical atlas to the space of plain CT images to generate a three-dimensional probability mask of the gastric wall structure; The similarity weights between the probability mask and the segmentation feature map are calculated using a graph attention gating unit, and the features of the stomach wall region in the feature map are selectively enhanced. The atlas attention gating unit uses learnable convolutional kernels to achieve local feature fusion and introduces an atlas prior loss function to constrain the spatial consistency between the segmentation network output and the anatomical atlas. The graph prior loss function is constructed based on the Dice similarity coefficient and boundary distance loss.

4. The method for analyzing plain CT images for gastric cancer screening as described in claim 1, characterized in that, The multi-scale 3D feature extraction network includes parallel coarse-scale feature extraction branches and fine-scale feature extraction branches; The coarse-scale feature extraction branch uses a three-dimensional dilated convolutional layer to expand the receptive field and capture macroscopic morphological changes in gastric wall thickening features; The fine-scale feature extraction branch uses a three-dimensional densely connected convolutional layer to extract local texture details of mucosal enhancement features; The output feature maps of the two branches are channel-stitched and spatially aligned by the feature fusion module. The feature fusion module uses an adaptive weight allocation mechanism to dynamically adjust the contribution of features at different scales.

5. The plain CT image analysis method for gastric cancer screening as described in claim 4, characterized in that, The adaptive weight allocation mechanism of the feature fusion module is implemented through the channel attention module and the spatial attention module; The channel attention module calculates the global statistics for each feature channel and generates a channel weight vector, while the spatial attention module calculates the importance of each spatial location in the feature map and generates a spatial weight matrix. The channel weight vector and the spatial weight matrix are multiplied element by element and then weighted and fused with the original feature map to obtain the fused multi-scale three-dimensional feature map. Both the channel attention module and the spatial attention module use fully connected layers and convolutional layers to achieve nonlinear transformations.

6. The plain CT image analysis method for gastric cancer screening as described in claim 1, characterized in that, The process of constructing the voxel-structure dynamic heterogeneity diagram includes: Voxel node generation and structural edge connection; A voxel node consists of the position coordinates and feature vector of each voxel in a multi-scale 3D feature map; The structural edges are established according to the continuity rules of the gastric wall structure. The continuity rules of the gastric wall structure define the connection relationship between adjacent voxels based on the anatomical layers of the gastric wall, including intramural connections, intramuscular connections, and translayer connections. Each structural edge is assigned a connection strength weight, which is calculated using a joint function of voxel feature similarity and spatial distance.

7. The method for analyzing plain CT images for gastric cancer screening as described in claim 6, characterized in that, The graph neural network uses a multi-layer graph convolutional network to aggregate node information of a dynamic heterogeneous graph; Each layer of the graph convolutional network performs node feature updates and edge weight updates; Node feature updates are achieved by aggregating the features of neighboring nodes and edge weights, and information flow is controlled by a gated loop unit. Edge weight updates recalculate connection strength weights based on the updated node features, forming a dynamic graph structure; After iterative execution of the multi-layer graph convolutional network, the aggregated feature vector of the stomach region is obtained by performing global average pooling on all node features.

8. The method for analyzing plain CT images for gastric cancer screening as described in claim 1, characterized in that, The multilayer perceptron classifier includes an input layer, multiple hidden layers, and an output layer; The input layer receives aggregated feature vectors, the hidden layers are processed with batch normalization and dropout regularization, and each hidden layer is followed by a non-linear activation function; The output layer uses the Softmax function to generate the probability distribution of gastric cancer screening classification results, which include three categories: normal, benign lesions, and malignant lesions. The number of neurons in the hidden layers decreases layer by layer, and the output dimension of the last hidden layer matches the number of classification categories.

9. The method for analyzing plain CT images for gastric cancer screening as described in claim 8, characterized in that, The training process of the multilayer perceptron classifier employs a focus loss function and label smoothing regularization. The focus loss function adjusts the weights of difficult and easy samples to enhance the learning ability for minority class samples; Label smoothing regularization converts the one-hot encoding of the true label into a smooth distribution, preventing the model from overfitting; During training, an adaptive moment estimation algorithm is used to optimize network parameters, and the learning rate is dynamically reduced based on the performance of the validation set.

10. A plain CT image analysis system for gastric cancer screening, characterized in that, The system is used to implement a plain CT image analysis method for gastric cancer screening as described in any one of claims 1 to 9, comprising: The image preprocessing module, based on a deep learning model enhanced with anatomical prior knowledge, locates the stomach region in plain CT images and obtains stomach region images; The multi-scale feature extraction module is used to construct a multi-scale three-dimensional feature extraction network for gastric cancer imaging manifestations. It extracts multi-scale three-dimensional feature maps from gastric region images. Gastric cancer imaging manifestations include mucosal enhancement features and gastric wall thickening features. The graph neural network aggregation module is used to associate voxel features in the multi-scale three-dimensional feature map with gastric wall structural features based on the continuity of the gastric wall structure, construct a voxel-structure dynamic heterogeneous map, and obtain the aggregated feature vector of the gastric region by aggregating the node information of the dynamic heterogeneous map through the graph neural network. The classification decision module is used to input the aggregated feature vector into the multilayer perceptron classifier and output the gastric cancer screening classification result corresponding to the plain CT image.

Images