Intelligent four-classification method for T stage of gastric cancer in progression stage based on low-tension water-filled stomach CT venous phase enhanced image

Abstract

The invention discloses a progression stage gastric cancer T stage intelligent four-classification method based on a low water filling stomach CT venous phase enhanced image, and relates to the technical field of medical imaging prediction. The method comprises the following steps: collecting an abdominal vein phase enhanced CT image of gastric cancer in a progression phase containing a hypotonic water filling stomach; the method comprises the following steps: preprocessing an abdominal vein phase enhanced CT image containing a low-tension water filling stomach to obtain sample data; a 3D-Conv-Unet network is utilized to automatically identify and segment a development stage gastric cancer primary lesion CT image area, and an image physician quickly confirms the image area to obtain an MASK image; and carrying out channel splicing on the vein phase enhanced CT full-layer abdominal image and the MASK image, processing by adopting a 3D convolution kernel to obtain an image sequence, and inputting the image sequence into a SmallFousNet network to obtain a progression phase gastric cancer T-stage classification result.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging prediction, and in particular to an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images. Background Art

[0002] Gastric cancer is a common malignant tumor with a relatively high incidence rate worldwide. Its homogeneous and accurate imaging staging is the basis for formulating precise treatment plans and is crucial to the patient's prognosis. In recent years, with the development of computer technology, automatic recognition and analysis technology based on medical images has gradually been applied in clinical practice, providing a new means for the diagnosis and staging of gastric cancer.

[0003] Among them, T staging is a technology for judging the depth of gastric cancer invasion of the gastric wall, which is crucial for formulating personalized treatment plans. Today, CT examination is the preferred examination technology for imaging evaluation before gastric cancer treatment, and the existing technology has many shortcomings in T staging before treatment of advanced gastric cancer, including the failure to use hypotonic water-filled gastric CT images, different levels of judgment skills of imaging physicians, limited application scope of models or poor accessibility to clinical applications, and there is no commercial artificial intelligence software that can be effectively applied to T staging of CT images before gastric cancer treatment. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images, so as to solve the shortcomings of the background technology.

[0005] The technical solution proposed in the present invention is implemented in the following steps: an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images, comprising:

[0006] Step 1: Perform an abdominal CT scan of the gastric cancer patient with hypotonic water-filled stomach to collect abdominal CT images of the entire stomach to prepare for the prediction of T staging results of advanced gastric cancer;

[0007] Step 2, preprocessing the abdominal CT image including the hypotonic water-filled stomach CT venous phase enhanced image to obtain sample data;

[0008] Step 3: Process the sample data based on the 3D-Conv-Unet network to automatically identify and segment the CT image area of ​​the primary lesion of advanced gastric cancer;

[0009] Step 4: The imaging physician confirms the range of the CT image area of ​​the primary lesion of advanced gastric cancer segmented by the 3D-Conv-Unet network, and corrects any errors to obtain a MASK image;

[0010] Step 5: Channel-joining of the venous phase enhanced CT full-layer abdominal image at the level where the MASK image is located and the MASK image is performed, and the 3D convolution kernel is used for processing to obtain an image sequence;

[0011] Step 6: Input the acquired image sequence into the SmallFocusNet network to obtain the T staging classification result of advanced gastric cancer.

[0012] Preferably, in step 1, the gastric cancer patient is fasting, hypotonic and takes warm water before CT scanning; the abdominal CT image includes hypotonic water-filled gastric CT venous phase enhanced image; the T stage of advanced gastric cancer includes T2, T3, T4a and T4b;

[0013] Abdominal CT scans include plain CT of the entire stomach, arterial phase enhanced CT, and venous phase enhanced CT; among them, venous phase enhanced CT is used to obtain venous phase enhanced images of hypotonic water-filled stomach CT.

[0014] Preferably, in step 2, the preprocessing of the abdominal CT image includes: exporting a 5 mm abdominal CT venous phase enhanced image including the entire stomach, converting its DICOMDIR format into NIFTI format, performing standardization processing, normalizing the pixel values ​​of the image to the range of [0,1], and obtaining sample data.

[0015] Preferably, in step 3, the method for automatically identifying and segmenting the CT image area of ​​the primary lesion of advanced gastric cancer is as follows:

[0016] Input the sample data as the first input image to the 3D-Conv-Unet network;

[0017] The encoder is used to extract higher-level features from the input image through downsampling operations to obtain high-order features;

[0018] The decoder is used to gradually restore the lesions in the predicted image through upsampling operations. In this process, it is spliced ​​multiple times with the high-order features output by the encoder until the size of the feature map is restored to its original size, and the CT image area of ​​the primary lesion of advanced gastric cancer is obtained.

[0019] Preferably, in step 5, by counting all lesion areas, eight lesion areas are retained for each case. For cases with insufficient number, the last image is copied until the conditions are met. For cases with more than eight images, the middle image part is preferentially selected to ensure that the number and dimension of images input into the 3D convolutional network are consistent.

[0020] Preferably, in step 5, it also includes: by formula I concat =concat(I CT ,M) Channel stitching of the venous phase enhanced CT full-layer abdominal image and the MASK image; where Iconcat Represents the stitched image; I CT represents the CT full-layer abdominal image; M represents the MASK image; concat represents the channel concatenation operation;

[0021] By formula X = Conv3D (I concat ) The spliced ​​images are processed with a 3D convolution kernel to obtain an image sequence; wherein X is the image sequence; Conv3D represents a three-dimensional convolution operation.

[0022] Preferably, in step 6, the method for obtaining the T staging classification result of advanced gastric cancer is as follows:

[0023] The spliced ​​image sequence is used as the second input image and input into the SmallFocusNet network; wherein the SmallFocusNet network includes the primary feature extraction layer ResBlock1, the deeper residual blocks ResBlock2 and ResBlock3, and the SEBlock combined with the channel attention mechanism;

[0024] ResBlock1: Perform preliminary feature extraction on the second input image through the formula F1=ReLU(Conv(X))+X; where F1 is the output feature map of ResBlock1; X is the image sequence and also the second input image; ReLU is the activation function; Conv is the convolution operation;

[0025] ResBlock2 and ResBlock3: Deep feature learning is performed through the formula F2=ReLU(Conv(F1))+F1 and the formula F3=ReLU(Conv(F2))+F2; where F2 is the output feature map of ResBlock2 and F3 is the output feature map of ResBlock3.

[0026] Preferably, in step 6, SEBlock combined with the channel attention mechanism: by formula Z C =GlobalAvgPool(F3) performs global average pooling on the output feature map F3 of ResBlock3 to obtain a 1×1×C feature vector; where Z C is the feature vector after global average pooling; C is the number of channels; GlobalAvgPool is the global average pooling operation;

[0027] Through the first fully connected layer, the features Z of C channels are C Compress to C / r channels, r is the reduction rate, and use the ReLU activation function for nonlinear transformation;

[0028] The calculation formula is as follows:

[0029] e=δ(g(Z C ,w1));

[0030] Among them, e is the intermediate feature vector after the first fully connected layer and the ReLU activation function; g(Z C ,w1) is the output of the first fully connected layer; w1 is the weight matrix of the first fully connected layer; δ represents the ReLU activation function;

[0031] By passing e through the second fully connected layer and performing Sigmoid activation through the formula s=σ(w2·e), we get the gated unit vector s; where s is the gated unit vector; w2 is the weight matrix of the first fully connected layer; σ represents the Sigmoid activation function;

[0032] By formula F SE =s·F3 multiplies the learned weights by the output feature map of ResBlock3; where F SE is the weighted feature map.

[0033] Preferably, in step 6, the T stage classification result is outputted by: inputting the weighted feature map into the classifier, and converting the output of the fully connected layer into a probability distribution through a softmax function;

[0034] The calculation formula is as follows:

[0035] Y m =softmax(F SE );

[0036] Where Y m represents the probability distribution of predicting the T stage of advanced gastric cancer; m = 1, 2, 3, 4, corresponding to the four categories of T2, T3, T4a and T4b respectively;

[0037] The predicted result with the highest probability is taken as the final T staging classification result of advanced gastric cancer.

[0038] The beneficial effects achieved by the present invention are as follows:

[0039] The present invention uses a 3D-Conv-Unet network to intelligently identify the hypotonic water-filled stomach CT venous phase enhanced images of advanced gastric cancer, and marks the primary tumor on the image. The CT venous phase images obtained by venous phase enhanced CT are conducive to more clearly displaying the location, size and morphology of the lesion, thereby improving the accuracy of intelligent classification of gastric cancer T staging.

[0040] Channel splicing of the MASK image and the full-level CT abdominal image at the venous phase of the MASK layer is conducive to forming multimodal feature information; and the 3D convolution processing method adopted can better capture spatial information and improve the depth and breadth of feature extraction;

[0041] The present invention uses the SmallFocusNet network to intelligently classify the T staging of advanced gastric cancer, including T2, T3, T4a and T4b, and achieves a high-accuracy prediction of the T staging of advanced gastric cancer. It has important guiding significance for clinical diagnosis and helps doctors to formulate more accurate treatment plans. The application of this technology is expected to promote the development of AI homogeneous evaluation before gastric cancer treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below in conjunction with the accompanying drawings.

[0043] Figure 1 The present invention is a flowchart of an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0045] See also Figure 1 The present invention provides an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images, comprising the following steps:

[0046] Step 1: Perform an abdominal CT scan of the gastric cancer patient with hypotonic water-filled stomach to collect abdominal CT images of the entire stomach to prepare for the prediction of T staging results of advanced gastric cancer;

[0047] Among them, gastric cancer patients are prepared to fast, hypotonic and drink warm water before CT scanning; abdominal CT images include hypotonic water-filled gastric CT venous phase enhanced images; advanced gastric cancer T staging is divided according to the depth of tumor infiltration into the gastric wall, including T2, T3, T4a and T4b;

[0048] T2 means the tumor has infiltrated the muscular layer of the stomach wall;

[0049] T3 means that the tumor has penetrated the gastric wall muscle layer and reached the subserosa layer, but has not yet invaded the serosa;

[0050] T4a means that the tumor has penetrated all layers of the gastric wall, that is, it has not only penetrated the entire muscle layer and subserosa, but also invaded or penetrated the serosa;

[0051] T4b means the tumor has penetrated the stomach wall and invaded adjacent structures or organs outside the stomach wall;

[0052] Furthermore, the abdominal CT scan includes plain CT of the whole stomach, arterial phase enhanced CT, and venous phase enhanced CT; wherein, venous phase enhanced CT is used to obtain venous phase enhanced images of hypotonic water-filled stomach CT;

[0053] It should be noted that the patient needs to fast for more than 4 hours to fill the stomach with hypotonic water, and the patient is given an intramuscular injection of 20 mg of racemic anisodamine hydrochloride injection about 30 minutes before the CT scan. Immediately after the injection, 800-1200 ml of warm water is taken orally; 200 ml is drunk immediately before the examination.

[0054] In the implementation steps of the present invention, plain scan CT may not be able to clearly display the details of the lesion or the vascular structure due to the lack of iodine contrast agent; the CT venous phase images obtained by venous phase enhanced CT can increase the local contrast and help to discover features that are not easily observed by plain scan.

[0055] Step 2, preprocessing the abdominal CT image including the hypotonic water-filled stomach CT venous phase enhanced image to obtain sample data;

[0056] Furthermore, the preprocessing of the abdominal CT image includes: exporting a 5 mm abdominal CT venous phase enhanced image including the entire stomach, converting the DICOMDIR format of the image into the NIFTI format, performing standardization processing, normalizing the pixel values ​​of the image to the range of [0,1], and obtaining sample data;

[0057] It should be noted that NIFTI is a file format used to store and transmit medical imaging data. It contains important data about the image, such as image size, resolution, layer thickness, layer spacing, pixel spacing, etc. It is more suitable for data storage and analysis in the field of imaging.

[0058] Step 3: Process the sample data based on the 3D-Conv-Unet network to automatically identify and segment the CT image area of ​​the primary lesion of advanced gastric cancer;

[0059] Among them, the 3D-Conv-Unet network is a convolutional neural network improved based on the U-Net architecture. By introducing 3D convolution operations, it can make full use of the spatial information of three-dimensional images and improve the accuracy and robustness of automatic segmentation.

[0060] Furthermore, in the 3D-Conv-Unet network, each layer of the encoder and decoder is composed of two convolution blocks; wherein the activation function is the ReLU function, and the convolution kernel is 3×3×3;

[0061] Input the sample data as the first input image to the 3D-Conv-Unet network;

[0062] The encoder is used to extract higher-level features from the input image through downsampling operations to obtain high-order features;

[0063] The decoder is used to gradually restore the lesions in the predicted image through upsampling operations. In this process, the high-order features output by the encoder are spliced ​​multiple times until the size of the feature map is restored to its original size, and the CT image area of ​​the primary lesion of advanced gastric cancer is obtained.

[0064] In some preferred embodiments, each layer of the encoder is composed of two convolution blocks, the first convolution block performs a convolution operation on the first input image, and the second convolution block performs a convolution operation on the output of the first convolution block;

[0065] The calculation formula is as follows:

[0066] O i =ReLU(Conv(ReLU(Conv(I i ,W i1 )),W i2 ));

[0067] In the formula, I i is the first input image before convolution of the i-th layer; O i is the output image after convolution of the i-th layer; i represents the number of layers, i is a positive integer; W i1 and W i2 is the convolution kernel weight of the first and second convolution blocks of the i-th convolution layer; ReLU is the activation function; Conv is the convolution operation;

[0068] In the downsampling operation, the size of the convolutional layer output image is reduced through the maximum pooling operation;

[0069] The calculation formula is as follows:

[0070] S i =Pooling(O i );

[0071] In the formula, S i is the feature map of the down-sampling output of the i-th layer; Pooling represents the pooling operation; O i is the output image of the convolutional layer of the i-th layer;

[0072] In some preferred embodiments, the decoder restores the downsampled feature map to its original size by upsampling, and concatenates it with the feature map of the corresponding level in the encoder;

[0073] The calculation formula is as follows:

[0074] U j =Upsample(S j );

[0075] F j =Concat(U j ,O j );

[0076] Among them, U j is the upsampled output of the jth layer, S j is the feature map of the corresponding level in the encoder; F j is the feature map output after concatenation of the jth layer; O j The output image of the j-th convolutional layer; Upsample represents the upsampling operation, and Concat represents the concatenation operation of the feature map;

[0077] After multiple layers of decoding, the feature map size is restored to its original size, and the CT image area of ​​the primary lesion of advanced gastric cancer is obtained;

[0078] In the implementation steps of the present invention, 3D-Conv-Unet is a deep learning network based on three-dimensional convolution, which combines the advantages of the U-Net architecture and is suitable for medical image segmentation tasks; the network extracts features through an encoder and a decoder restores spatial information to achieve high-precision image segmentation.

[0079] Step 4: The imaging physician confirms the range of the CT image area of ​​the primary lesion of advanced gastric cancer segmented by the 3D-Conv-Unet network, and corrects any errors to obtain a MASK image;

[0080] Step 5: Channel-joining of the venous phase enhanced CT full-layer abdominal image at the level where the MASK image is located and the MASK image is performed, and the 3D convolution kernel is used for processing to obtain an image sequence;

[0081] Among them, by counting all the lesion areas, eight lesion areas are retained for each case. For cases with insufficient number of images, the last image is copied until the conditions are met. For cases with more than eight images, the middle image part is selected first to ensure that the number and dimension of images input into the 3D convolutional network are consistent;

[0082] Furthermore, through formula I concat =concat(I CT ,M) Channel stitching of the venous phase enhanced CT full-layer abdominal image and the MASK image; where I concat Represents the stitched image; I CT represents the CT full-layer abdominal image; M represents the MASK image; concat represents the channel concatenation operation;

[0083] By formula X = Conv3D (I concat) The spliced ​​images are processed using a 3D convolution kernel to obtain an image sequence; wherein X is the image sequence; Conv3D represents a three-dimensional convolution operation;

[0084] In the implementation steps of the present invention, by channel splicing the venous phase enhanced CT image and the mask image, two different types of information can be fused together. The venous phase enhanced CT image provides rich anatomical structure and tissue density information, while the mask image highlights the lesion area. This combination enables the model to simultaneously learn the texture features of normal tissues and the spatial location and morphological features of lesions, thereby more comprehensively understanding the data.

[0085] Step 6: Input the acquired image sequence into the SmallFocusNet network to obtain the T staging classification result of advanced gastric cancer;

[0086] Among them, the SmallFocusNet network includes the primary feature extraction layer (ResBlock1), deeper residual blocks (ResBlock2 and ResBlock3), and SEBlock combined with the channel attention mechanism;

[0087] Furthermore, the spliced ​​image sequence is used as the second input image and input into the SmallFocusNet network;

[0088] ResBlock1: Perform preliminary feature extraction on the second input image through the formula F1=ReLU(Conv(X))+X; where F1 is the output feature map of ResBlock1; X is the image sequence and also the second input image; ReLU is the activation function; Conv is the convolution operation;

[0089] ResBlock2 and ResBlock3: Deep feature learning is performed using the formula F2 = ReLU(Conv(F1)) + F1 and the formula F3 = ReLU(Conv(F2)) + F2; where F2 is the output feature map of ResBlock2 and F3 is the output feature map of ResBlock3;

[0090] SEBlock combined with channel attention mechanism:

[0091] By formula Z C =GlobalAvgPool(F3) performs global average pooling on the output feature map F3 of ResBlock3 to obtain a 1×1×C feature vector; where Z C is the feature vector after global average pooling; C is the number of channels; GlobalAvgPool is the global average pooling operation;

[0092] It should be noted that this step is to compress each feature map into a real number, thereby obtaining the global receptive field of each channel and compressing the information of the spatial dimension into the channel descriptor;

[0093] Through the first fully connected layer, the features Z of C channels are C Compress to C / r channels and use ReLU activation function for nonlinear transformation, where r is the reduction rate, usually set to 16;

[0094] The calculation formula is as follows:

[0095] e=δ(g(Z C ,w1));

[0096] Among them, e is the intermediate feature vector after the first fully connected layer and the ReLU activation function; g(Z C ,w1) is the output of the first fully connected layer; w1 is the weight matrix of the first fully connected layer; δ represents the ReLU activation function;

[0097] It should be noted that this step is to reduce the number of parameters and increase the nonlinear expression ability of the network;

[0098] Pass e through the second fully connected layer and perform Sigmoid activation through the formula s=σ(w2·e) to obtain the gated unit vector s; where s is the gated unit vector; w2 is the weight matrix of the first fully connected layer; σ represents the Sigmoid activation function, which limits the output between 0 and 1;

[0099] It should be noted that this step generates the importance weight of each channel for subsequent weighted operations.

[0100] By formula F SE =s·F3 multiplies the learned weights by the output feature map of ResBlock3; where F SE is the weighted feature map;

[0101] It should be noted that this step realizes the recalibration of features, thereby enhancing important features and suppressing unimportant features;

[0102] T-stage classification result output: The weighted feature map is input into the classifier, and the output of the fully connected layer is converted into a probability distribution through the softmax function;

[0103] The calculation formula is as follows:

[0104] Y m =softmax(F SE );

[0105] Where Y mrepresents the probability distribution of predicting the T stage of advanced gastric cancer; m = 1, 2, 3, 4, corresponding to the four categories of T2, T3, T4a and T4b respectively;

[0106] The predicted result with the highest probability is taken as the final T staging classification result of advanced gastric cancer.

[0107] In the embodiments provided by the present invention, it should be understood that the disclosed system can be implemented in other ways. For example, the above-described embodiments of the invention are only illustrative, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation.

[0108] The modules described as separate accessories may or may not be physically separated, and the accessories shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0109] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0110] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images, characterized in that: include: Step 1: Perform an abdominal CT scan of the gastric cancer patient with hypotonic water-filled stomach to collect abdominal CT images of the entire stomach to prepare for the prediction of T staging results of advanced gastric cancer; Step 2, preprocessing the abdominal CT image including the hypotonic water-filled stomach CT venous phase enhanced image to obtain sample data; Step 3: Process the sample data based on the 3D-Conv-Unet network to automatically identify and segment the CT image area of ​​the primary lesion of advanced gastric cancer; Step 4: The imaging physician confirms the range of the CT image area of ​​the primary lesion of advanced gastric cancer segmented by the 3D-Conv-Unet network, and corrects any errors to obtain a MASK image; Step 5: Channel-joining of the venous phase enhanced CT full-layer abdominal image at the level where the MASK image is located and the MASK image is performed, and the 3D convolution kernel is used for processing to obtain an image sequence; Step 6: Input the acquired image sequence into the SmallFocusNet network to obtain the T staging classification result of advanced gastric cancer.

2. According to claim 1, an intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images, characterized in that: In step 1, gastric cancer patients are prepared with fasting, hypotonicity, and warm water before CT scanning; abdominal CT images include hypotonic water-filled gastric CT venous phase enhanced images; advanced gastric cancer T stages include T2, T3, T4a, and T4b; Abdominal CT scans include plain CT of the entire stomach, arterial phase enhanced CT, and venous phase enhanced CT; among them, venous phase enhanced CT is used to obtain venous phase enhanced images of hypotonic water-filled stomach CT.

3. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 2, characterized in that: In step 2, the preprocessing of the abdominal CT image includes: exporting a 5 mm abdominal CT venous phase enhanced image including the entire stomach, converting its DICOMDIR format into NIFTI format, performing standardization processing, normalizing the pixel values ​​of the image to the range of [0,1], and obtaining sample data.

4. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 3, characterized in that: In step 3, the method for automatically identifying and segmenting the CT image area of ​​the primary lesion of advanced gastric cancer is as follows: Input the sample data as the first input image to the 3D-Conv-Unet network; The encoder is used to extract higher-level features from the input image through downsampling operations to obtain high-order features; The decoder is used to gradually restore the lesions in the predicted image through upsampling operations. In this process, it is spliced ​​multiple times with the high-order features output by the encoder until the size of the feature map is restored to its original size, and the CT image area of ​​the primary lesion of advanced gastric cancer is obtained.

5. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 4, characterized in that: In step 5, By counting all the lesion areas, eight lesion areas are retained for each case. For cases with insufficient number of lesion areas, the last one is copied until the conditions are met. For cases with more than eight images, the middle image part is selected first to ensure that the number and dimension of images input into the 3D convolutional network are consistent.

6. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 5, characterized in that: In step 5, it also includes: using formula I concat =concat(I CT ,M) Channel stitching of the venous phase enhanced CT full-layer abdominal image and the MASK image; where I concat Represents the stitched image; I CT represents the CT full-layer abdominal image; M represents the MASK image; concat represents the channel concatenation operation; By formula X = Conv3D (I concat ) The spliced ​​images are processed with a 3D convolution kernel to obtain an image sequence; wherein X is the image sequence; Conv3D represents a three-dimensional convolution operation.

7. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 6, characterized in that: In step 6, the method for obtaining the T staging classification result of advanced gastric cancer is as follows: The spliced ​​image sequence is used as the second input image and input into the SmallFocusNet network; wherein the SmallFocusNet network includes the primary feature extraction layer ResBlock1, the deeper residual blocks ResBlock2 and ResBlock3, and the SEBlock combined with the channel attention mechanism; ResBlock1: Perform preliminary feature extraction on the second input image through the formula F1=ReLU(Conv(X))+X; where F1 is the output feature map of ResBlock1; X is the image sequence and also the second input image; ReLU is the activation function; Conv is the convolution operation; ResBlock2 and ResBlock3: Deep feature learning is performed through the formula F2=ReLU(Conv(F1))+F1 and the formula F3=ReLU(Conv(F2))+F2; where F2 is the output feature map of ResBlock2 and F3 is the output feature map of ResBlock3.

8. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 7, characterized in that: SEBlock combined with channel attention mechanism in step 6: By formula Z C =GlobalAvgPool(F3) performs global average pooling on the output feature map F3 of ResBlock3 to obtain a 1×1×C feature vector; where Z C is the feature vector after global average pooling; C is the number of channels; GlobalAvgPool is the global average pooling operation; Through the first fully connected layer, the features Z of C channels are C Compress to C / r channels, r is the reduction rate, and use the ReLU activation function for nonlinear transformation; The calculation formula is as follows: e=δ(g(Z C ,w1)); Among them, e is the intermediate feature vector after the first fully connected layer and the ReLU activation function; g(Z C ,w1) is the output of the first fully connected layer; w1 is the weight matrix of the first fully connected layer; δ represents the ReLU activation function; By passing e through the second fully connected layer and performing Sigmoid activation through the formula s=σ(w2·e), we get the gated unit vector s; where s is the gated unit vector; w2 is the weight matrix of the first fully connected layer; σ represents the Sigmoid activation function; By formula F SE =s·F3 multiplies the learned weights by the output feature map of ResBlock3; where F SE is the weighted feature map.

9. The intelligent four-classification method for T staging of advanced gastric cancer based on hypotonic water-filled gastric CT venous phase enhanced images according to claim 8, characterized in that: In step 6, the T-stage classification results are output: the weighted feature map is input into the classifier, and the output of the fully connected layer is converted into a probability distribution through the softmax function; The calculation formula is as follows: Y m =softmax(F SE ); Where Y m represents the probability distribution of predicting the T stage of advanced gastric cancer; m = 1, 2, 3, 4, corresponding to the four categories of T2, T3, T4a and T4b respectively; The predicted result with the highest probability is taken as the final T staging classification result of advanced gastric cancer.

Images