Processing method and apparatus for performing m-stage classification prediction on basis of CT images

Abstract

A processing method and apparatus for performing M-stage classification prediction on the basis of CT images. The method comprises: on the basis of a U-Net model, constructing a first and a second semantic segmentation model; on the basis of a machine learning classification model, constructing a first prediction model; on the basis of a deep learning classification model, constructing a second prediction model; training the first and second semantic segmentation models and the first and second prediction models; after the training of the four types of models is completed, receiving a CT image and a prediction mode; if the prediction mode is a first mode, performing prediction on the basis of the CT image, the first and second semantic segmentation models and the first prediction model; if the prediction mode is a second mode, performing prediction on the basis of the CT image, the first and second semantic segmentation models and the second prediction model; and outputting an obtained prediction type. The method can improve the real-time performance and efficiency of prediction.

Description

A processing method and apparatus for M-stage classification prediction based on CT images.

[0001] This application claims priority to Chinese Patent Application No. 202410515492.1, filed on April 26, 2024, entitled "A Processing Method and Apparatus for M-Phase Classification Prediction Based on CT Images". Technical Field

[0002] This invention relates to the field of data processing technology, and in particular to a processing method and apparatus for M-phase classification prediction based on CT images. Background Technology

[0003] The M staging system is a gastric cancer staging system that classifies gastric tumors based on whether they have metastasized to distant organs (such as the liver), distant lymph nodes (such as para-aortic lymph nodes, retropancreatic lymph nodes, and root mesenteric lymph nodes), or the peritoneum. Common M staging types include M0 and M1, where M0 indicates no distant metastasis and M1 indicates distant metastasis. Currently, medical experts can manually identify M staging types based on computed tomography (CT) images. However, this manual identification method is clearly limited by the experts' time constraints, resulting in significant time-sensitivity issues, especially when processing large batches of CT images. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a processing method, apparatus, electronic device, and computer-readable storage medium for M-stage classification prediction based on CT images. This invention constructs two semantic segmentation models based on the model structure of the U-Net model. The first semantic segmentation model is used for semantic segmentation of semantic points in the gastric tumor lesion area on CT images, and the second semantic segmentation model is used for semantic segmentation of semantic points in the lesion area of ​​distal key tissues on CT images. A first prediction model capable of M-stage classification prediction based on input radiomics features is constructed based on a type of machine learning classification model, and a second prediction model capable of M-stage classification prediction based on input 3D CT images with lesion point markings is constructed based on a type of deep learning classification model. The above four models are trained separately. After the training of the above four models, different combinations of prediction models are selected for prediction based on the received prediction mode: when the prediction mode is the first mode, the first and second semantic segmentation models and the first prediction model are used to predict the M-stage of the received CT images; when the prediction mode is the second mode, the first and second semantic segmentation models and the second prediction model are used to predict the M-stage of the received CT images. This invention can improve the real-time performance and efficiency of prediction.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for M-phase classification prediction based on CT images, the method comprising:

[0006] A semantic segmentation model based on the U-Net model is constructed to segment the lesion area of ​​a gastric tumor on CT images, denoted as the first semantic segmentation model; a semantic segmentation model based on the U-Net model is constructed to segment the lesion area of ​​distal key tissues on the CT images, denoted as the second semantic segmentation model; a prediction model based on a machine learning classification model is constructed to predict the M-stage based on radiomics features, denoted as the first prediction model; and a prediction model based on a deep learning classification model is constructed to predict the M-stage based on a 3D CT image with lesion point markers, denoted as the second prediction model; wherein, the CT... The image is a three-dimensional abdominal CT scan with three-dimensional information of the stomach and the aforementioned distal key tissues; the distal key tissues include at least the liver, para-aortic lymph node region, retropancreatic lymph node region, root mesenteric lymph node region, and peritoneum; the machine learning classification model includes at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models; the deep learning classification model includes at least classification prediction models based on ResNet models and classification prediction models based on Transformer models;

[0007] The first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model are trained respectively.

[0008] After the four types of models are trained, a first CT image and a first prediction mode are received; the first prediction mode is identified; if the first prediction mode is a first mode, a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction mode is a second mode, a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and the obtained first prediction type is output as the prediction result corresponding to the first CT image; wherein, the first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first prediction mode includes a first mode and a second mode; the first prediction type includes type M0 and type M1.

[0009] Preferably, the step of training the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively specifically includes:

[0010] A first subject set is formed by recruiting multiple subjects who do not have gastric cancer, multiple subjects who have gastric cancer but have not metastasized to any of the aforementioned distant key tissues, and multiple subjects who have gastric cancer and have metastasized to one or more of the aforementioned distant key tissues. The first subject set includes multiple first subjects, each corresponding to a first M phase type. The first M phase type includes M0 type and M1 type. The first M phase type corresponding to the first subjects who do not have gastric cancer or who have gastric cancer but have not metastasized to any of the aforementioned distant key tissues is M0 type, and the first M phase type corresponding to the first subjects who have gastric cancer and have metastasized to one or more of the aforementioned distant key tissues is M1 type.

[0011] The most recent CT images of each of the first subjects are collected as corresponding first sample images; and each of the first sample images and the corresponding first M-phase type are combined to form a corresponding first raw data record; and all the obtained first raw data records are combined to form a corresponding first raw dataset; the first sample image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first sample image includes multiple first sample voxels; each first sample voxel corresponds to a first voxel pixel value.

[0012] Based on the first original dataset, four types of model training datasets are constructed and denoted as the first model dataset, the second model dataset, the third model dataset, and the fourth model dataset.

[0013] The first semantic segmentation model is trained based on the first model dataset;

[0014] The second semantic segmentation model is trained based on the second model dataset;

[0015] The first prediction model is trained based on the third model dataset;

[0016] The second prediction model is trained based on the fourth model dataset;

[0017] The training of the four types of models was confirmed to be complete after the first and second semantic segmentation models and the first and second prediction models had all completed their training.

[0018] Furthermore, the first model dataset includes multiple first model data records; the first model data records include a first training image and a first label semantic map; the three-dimensional size of the first label semantic map is consistent with the three-dimensional size of the first training image; the first label semantic map includes multiple first label voxels; each first label voxel corresponds to a binary first label semantic, with a value of 0 or 1; when the first label semantic is 0, it indicates that the corresponding first label voxel is a background voxel; when it is 1, it indicates that the corresponding first label voxel is a voxel of a gastric tumor lesion area.

[0019] The second model dataset includes multiple second model data records; the second model data records include second training images and second label semantic maps; the three-dimensional dimensions of the second label semantic map are consistent with the three-dimensional dimensions of the second training images; the second label semantic map includes multiple second label voxels; each second label voxel corresponds to a binary second label semantic, with a value of 0 or 1; when the second label semantic is 0, it indicates that the corresponding first label voxel is a background voxel, and when it is 1, it indicates that the corresponding second label voxel is a distant critical tissue lesion voxel;

[0020] The third model dataset includes multiple third model data records; each third model data record includes a first training radiomics feature set and a first label type vector; the first training radiomics feature set includes a first gastric tumor feature set and a first distal key tissue feature set; the first distal key tissue feature set includes multiple first distal key tissue feature sets, each of which corresponds to one distal key tissue; the first gastric tumor feature set and all first distal key tissue feature sets are composed of a set of first-order feature sets, shape feature sets, and texture feature sets of radiomics features; the first label type vector includes two first label type probabilities, each of which corresponds to an M-stage type; the M-stage type includes M0 type and M1 type;

[0021] The fourth model dataset includes multiple fourth model data records; each fourth model data record includes a first training feature map and a second label type vector; the first training feature map includes multiple first feature voxels; each first feature voxel corresponds to a second voxel feature vector; the second voxel feature vector includes at least the second voxel pixel value and the second voxel semantics; the second voxel semantics is a ternary semantic value, taking the value 0, 1, or 2; when the second voxel semantics is 0, it indicates that the corresponding first feature voxel is a background voxel; when it is 1, it indicates that the corresponding first feature voxel is a gastric tumor lesion voxel; when it is 2, it indicates that the corresponding first feature voxel is a distal critical tissue lesion voxel; the second label type vector includes two second label type probabilities, each second label type probability corresponding to one M-stage type.

[0022] Furthermore, the construction of four types of model training datasets based on the first original dataset, denoted as the first model dataset, the second model dataset, the third model dataset, and the fourth model dataset, specifically includes:

[0023] Each of the first original data records in the first original dataset is taken as the corresponding current original data record; and the first sample image and the corresponding first M-period type of the current original data record are taken as the corresponding current sample image and current sample type.

[0024] The current sample image is used as the corresponding first training image and second training image; and the corresponding first and second label semantic maps are constructed based on the three-dimensional dimensions of the first and second training images; and the first and second label semantics of all first and second label voxels in the first and second label semantic maps are initialized to 0.

[0025] Based on manual annotation or other machine annotation methods, all first sample voxel points located in the gastric tumor lesion area on the current sample image are marked as Class I voxel points, and all first sample voxel points located in the lesion areas of all distal key tissues on the current sample image are marked as Class II voxel points.

[0026] When the total number of voxel points of the aforementioned type is zero, a set of all-zero first-order feature sets, shape feature sets, and texture feature sets is set to form a corresponding first gastric tumor feature set;

[0027] When the total number of voxel points of the first type is greater than zero, the first label semantic of the first label voxel point corresponding to each voxel point of the first type on the first label semantic map is reset to 1; and the three-dimensional structure composed of all voxel points of the first type on the current sample image is recorded as the corresponding first type structure; and the first-order feature set, shape feature set and texture feature set of the radiomics features of the first type structure are calculated to form a corresponding first gastric tumor feature set;

[0028] When the total number of the two types of voxel points is zero, a set of all-zero first-order feature sets, shape feature sets, and texture feature sets are set for each of the distal critical tissues to form a corresponding first distal critical tissue feature set; and all the obtained first distal critical tissue feature sets are combined to form a corresponding first distal critical tissue feature set.

[0029] When the total number of the two types of voxels is greater than zero, the second label semantics of the second label voxels corresponding to each of the two types of voxels on the second label semantic map are reset to 1; and based on a preset point clustering algorithm, the two types of voxels on the current sample image corresponding to each of the remote critical tissues are clustered to obtain the corresponding first point clusters; and each of the first point clusters is used as the corresponding current point cluster, and it is identified whether the current point cluster is empty; if the current point cluster is empty, a set of all-zero first-order feature sets is set for the remote critical tissue corresponding to the current point cluster. The shape feature set and the texture feature set form a corresponding first remote critical tissue feature set; if the current point group is not empty, the three-dimensional structure composed of the current point group on the current sample image is recorded as the corresponding second-class structure, and the first-order features, shape features and texture features of the image omics features of the second-class structure are calculated to obtain the corresponding first-order feature set, the shape feature set and the texture feature set form a corresponding first remote critical tissue feature set; and all the obtained first remote critical tissue feature sets form a corresponding first remote critical tissue feature set;

[0030] The first training image omics feature set is composed of the first gastric tumor feature set corresponding to the current sample image and the first distal key tissue feature set.

[0031] And take each of the first sample voxel points of the current sample image as the corresponding current sample voxel point; and take the first and second label voxel points on the first and second label semantic maps corresponding to the current sample image as the corresponding first and second matching voxel points; take the pixel value of the first voxel point of the current sample voxel point as a corresponding second voxel point pixel value; and set a corresponding second voxel point semantic value to 0 when the first and second label semantic values ​​corresponding to the first and second matching voxel points are both 0; and set the first label semantic value corresponding to the first matching voxel point to 0. When the semantics of the second voxel point are set to 1, a corresponding second voxel point semantics of 1 is set to 2; and when the semantics of the second label corresponding to the second matching voxel point are 1, a corresponding second voxel point semantics of 2 is set to 2; and a corresponding first feature voxel point is constructed based on the three-dimensional voxel coordinates of the current sample voxel point, denoted as the corresponding current feature voxel point; and a second voxel point feature vector corresponding to the current feature voxel point is formed by the pixel value of the second voxel point corresponding to the current sample voxel point and the semantics of the second voxel point; and a corresponding first training feature map is formed by all the first feature voxel points corresponding to the current sample image.

[0032] The first and second label type vectors are constructed according to the current sample type. When the current sample type is M0, the probability of the first and second label types corresponding to type M0 in the corresponding first and second label type vectors is 1, and the probability of the first and second label types corresponding to type M1 is 0. When the current sample type is M1, the probability of the first and second label types corresponding to type M1 in the corresponding first and second label type vectors is 0, and the probability of the first and second label types corresponding to type M1 is 1.

[0033] The first model data record is composed of the first training image and the first label semantic map corresponding to the current original data record; the second model data record is composed of the second training image and the second label semantic map corresponding to the current original data record; the third model data record is composed of the first training image omics feature set and the first label type vector corresponding to the current original data record; and the fourth model data record is composed of the first training feature map and the second label type vector corresponding to the current original data record.

[0034] The first model dataset is composed of all the first model data records corresponding to the first original dataset; the second model dataset is composed of all the second model data records corresponding to the first original dataset; the third model dataset is composed of all the third model data records corresponding to the first original dataset; and the fourth model dataset is composed of all the fourth model data records corresponding to the first original dataset.

[0035] Furthermore, training the first semantic segmentation model based on the first model dataset specifically includes:

[0036] Step 51: Take the first first model data record of the first model dataset as the corresponding current model data record;

[0037] Step 52: Input the first training image recorded in the current model data into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first predicted semantic map.

[0038] Step 53: Input the first predicted semantic map and the first label semantic map of the current model data record into a preset first model loss function to calculate the corresponding first loss value;

[0039] The first model loss function is implemented based on the L1 loss function or the L2 loss function;

[0040] Step 54: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current model data record is the last first model data record in the first model dataset; if yes, proceed to step 55; otherwise, take the next first model data record in the first model dataset as the new current model data record and return to step 52 to continue training; if the first loss value does not meet the first loss value range, optimize the first semantic segmentation model parameters in one round based on the preset first model parameter optimizer in the direction of minimizing the first model loss function, and return to step 52 to continue training when the parameter optimization ends.

[0041] The first model parameter optimizer includes at least an SDG optimizer;

[0042] Step 55: Solidify the model parameters of the first semantic segmentation model and confirm that the training of the first semantic segmentation model has ended.

[0043] Furthermore, training the second semantic segmentation model based on the second model dataset specifically includes:

[0044] Step 61: Take the first second model data record of the second model dataset as the corresponding current model data record;

[0045] Step 62: Input the second training image recorded in the current model data into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the remote key tissue to obtain the corresponding second predicted semantic map;

[0046] Step 63: Input the second predicted semantic map and the second label semantic map of the current model data record into a preset second model loss function to calculate the corresponding second loss value;

[0047] The second model loss function is implemented based on either the L1 loss function or the L2 loss function.

[0048] Step 64: Identify whether the second loss value meets the preset second loss value range; if the second loss value meets the second loss value range, identify whether the current model data record is the last second model data record in the second model dataset; if yes, proceed to step 65; otherwise, take the next second model data record in the second model dataset as the new current model data record and return to step 62 to continue training; if the second loss value does not meet the second loss value range, optimize the second semantic segmentation model in one round based on the preset second model parameter optimizer in the direction of minimizing the second model loss function, and return to step 62 to continue training when the parameter optimization ends.

[0049] The second model parameter optimizer includes at least an SDG optimizer;

[0050] Step 65: Solidify the model parameters of the second semantic segmentation model and confirm that the training of the second semantic segmentation model has ended.

[0051] Furthermore, training the first prediction model based on the third model dataset specifically includes:

[0052] Step 701: Perform a full combination of the hyperparameters of the first prediction model to obtain multiple corresponding first hyperparameter combinations; and perform a hyperparameter setting on the first prediction model based on each first hyperparameter combination to obtain the corresponding first prediction model.

[0053] Step 702: Set parameter K to a positive integer greater than or equal to 5; and divide the third model dataset into K equal subsets, and denote each subset as the corresponding first subset D. i ; 1 ≤ index i ≤ K;

[0054] Step 703, take any one of the first subsets D i DB serves as the corresponding first validation dataset i And will exclude the current first subset D i The remaining K-1 data points of the first subset D i Merge into a corresponding first test dataset DA i ; and the obtained first test datasets DA i and the corresponding first verification dataset DB i Form a corresponding first test and verification set;

[0055] Step 704: Take the first prediction model as the corresponding current prediction model;

[0056] Step 705: Take the first first test verification set as the corresponding current test verification set;

[0057] Step 706, the first test dataset DA of the current test verification set. i and the first verification dataset DB i As the corresponding current test dataset and current validation dataset;

[0058] Step 707: Take the first third model data record of the current test dataset as the corresponding current model data record;

[0059] Step 708: Input the first training image omics feature set recorded in the current model data into the current prediction model to perform M-period classification prediction to obtain the corresponding first prediction vector;

[0060] Wherein, the first prediction vector includes two first type prediction probabilities, each of the first type prediction probabilities corresponding to one M-stage type;

[0061] Step 709: Substitute the first prediction vector and the first label type vector of the current model data record into the preset third model loss function; and perform a round of model parameter optimization on the model parameters of the current prediction model based on the preset third model parameter optimizer in the direction of minimizing the third model loss function; and at the end of this round of parameter optimization, identify whether the current model data record is the last third model data record of the current test dataset. If so, proceed to step 710; otherwise, extract the next third model data record of the current test dataset as the new current model data record and return to step 708 for training.

[0062] The third model loss function is based on the binary cross-entropy loss function; the third model parameter optimizer includes at least the SDG optimizer.

[0063] Step 710: Set three counters initially set to 0 and denoted as TP counter, FP counter and FN counter;

[0064] Step 711: Traverse all the third model data records in the current validation dataset; during traversal, take the currently traversed third model data record as the corresponding current validation record; take the first training radiomics feature set and the first label type vector of the current validation record as the corresponding current radiomics feature set and current label type vector; take the M stage type corresponding to the first label type probability with a probability value of 1 in the current label type vector as the corresponding current label type; input the current radiomics feature set into the current prediction model to perform M stage classification prediction to obtain the corresponding second prediction vector; take the M stage type corresponding to the second type prediction probability with the highest probability in the second prediction vector as the corresponding current prediction type; compare the current prediction type with the current label type; if the current prediction type and the current label type are both M1 type, then increment the TP counter by 1; if the current prediction type is M1 type and the current label type is M0 type, then increment the FP counter by 1; if the current prediction type is M0 type and the current label type is M1 type, then increment the FN counter by 1.

[0065] Wherein, the second prediction vector includes two second type prediction probabilities, each second type prediction probability corresponding to one M-stage type;

[0066] Step 712: After traversing all the third model data records in the current validation dataset, calculate the corresponding first precision and first recall based on the TP counter, the FP counter, and the FN counter; and calculate the corresponding first score F1 based on the first precision and the first recall. i ;

[0067] Step 713: Identify whether the current test verification set is the Kth first test verification set; if yes, proceed to step 714; if no, take the next first test verification set as the new current test verification set and return to step 706.

[0068] Step 714, based on the obtained K first scores F1 i The first error E is obtained by calculating the mean square error.

[0069] Step 715: Identify whether the current prediction model is the last first prediction model; if yes, proceed to step 716; if no, take the next first prediction model as the new current prediction model and return to step 705.

[0070] Step 716: Take the full model parameter set of the first prediction model corresponding to the first error E with the smallest error as the corresponding optimal full model parameter set.

[0071] Step 717: Based on the optimal full model parameter set, solidify the model parameters of the first prediction model and confirm that the model training of the first prediction model has ended.

[0072] Furthermore, training the second prediction model based on the fourth model dataset specifically includes:

[0073] Step 81: Take the first fourth model data record of the fourth model dataset as the corresponding current model data record;

[0074] Step 82: Input the first training feature map of the current model data record into the second prediction model to perform M-period classification prediction to obtain the corresponding third prediction vector;

[0075] The third prediction vector includes two third-type prediction probabilities, each of which corresponds to one of the M-stage types.

[0076] Step 83: Input the third prediction vector and the second label type vector of the current model data record into the preset fourth model loss function to calculate the corresponding third loss value;

[0077] The fourth model loss function is implemented based on the binary classification cross-entropy loss function;

[0078] Step 84: Identify whether the third loss value meets the preset third loss value range; if the third loss value meets the third loss value range, identify whether the current model data record is the last fourth model data record in the fourth model dataset; if yes, proceed to step 85; otherwise, take the next fourth model data record in the fourth model dataset as the new current model data record and return to step 82 to continue training; if the third loss value does not meet the third loss value range, perform a round of model parameter optimization on the second prediction model based on the preset fourth model parameter optimizer in the direction of minimizing the fourth model loss function, and return to step 82 to continue training when the parameter optimization ends.

[0079] The fourth model parameter optimizer includes at least the SDG optimizer;

[0080] Step 85: Solidify the model parameters of the second prediction model and confirm that the training of the second prediction model is complete.

[0081] Preferably, the step of predicting the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model specifically includes:

[0082] Step 91: Input the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first semantic map;

[0083] The first CT image includes multiple first voxel points; the three-dimensional dimensions of the first semantic map are consistent with the three-dimensional dimensions of the first CT image; the first semantic map includes multiple second voxel points, and the first and second voxel points correspond one-to-one; each second voxel point corresponds to a binary gastric tumor lesion area semantic, with a value of 0 or 1; when the gastric tumor lesion area semantic is 0, it indicates that the corresponding second voxel point is a background voxel point, and when it is 1, it indicates that the corresponding second voxel point is a gastric tumor lesion area voxel point;

[0084] Step 92: Count the number of second voxel points with semantic value 1 in the gastric tumor lesion area on the first semantic map to obtain the corresponding first number; and identify whether the first number exceeds the preset first number threshold. If it exceeds the threshold, proceed to step 93; if it does not exceed the threshold, set the corresponding first prediction type to type M0 and proceed to step 99.

[0085] Step 93: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding second semantic map;

[0086] The second semantic map has the same three-dimensional dimensions as the first CT image. The second semantic map includes multiple third voxel points, and the first and third voxel points correspond one-to-one. Each third voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding third voxel point is a background voxel point; when it is 1, it indicates that the corresponding third voxel point is a remote critical tissue lesion voxel point.

[0087] Step 94, and count the number of third voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding second number; and identify whether the second number exceeds the preset second number threshold. If it exceeds, proceed to step 95; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 99.

[0088] Step 95: The first voxel points on the first CT image that correspond to the second voxel points with semantic value 1 in each of the aforementioned distal key tissue lesion areas are recorded as the corresponding first lesion points; and the three-dimensional structure composed of all the first lesion points is recorded as the corresponding first structure; and the first-order features, shape features, and texture features of the radiomics features of the first structure are calculated to obtain the corresponding first-order feature set, shape feature set, and texture feature set, which together form a corresponding second gastric tumor feature set;

[0089] Step 96: Based on a preset point clustering algorithm, cluster the first voxel points on the first CT image corresponding to each of the distal critical tissues to obtain corresponding second point groups; and use each of the second point groups as the corresponding current point groups, and identify whether the current point group is empty; if the current point group is empty, set a set of all-zero first-order feature set, shape feature set, and texture feature set for the distal critical tissue corresponding to the current point group to form a corresponding second distal critical tissue feature set; if the current point group is not empty, record the three-dimensional structure on the first CT image composed of the current point groups as the corresponding second structure, and calculate the first-order feature set, shape feature set, and texture feature set of the radiomics features of the second structure to form a corresponding second distal critical tissue feature set; and use all the obtained second distal critical tissue feature sets to form a corresponding second distal critical tissue feature set;

[0090] Step 97: A corresponding first radiomics feature set is formed by combining the obtained second gastric tumor feature set and the second distal key tissue feature set; and the first radiomics feature set is input into the first prediction model to perform M-stage classification prediction to obtain the corresponding fourth prediction vector;

[0091] The fourth prediction vector includes two fourth-type prediction probabilities, each of which corresponds to one M-stage type.

[0092] Step 98: Take the M-stage type corresponding to the fourth type prediction probability with the highest probability in the fourth prediction vector as the corresponding first prediction type;

[0093] Step 99: Output the obtained first prediction type as the prediction result for this time.

[0094] Preferably, the step of predicting the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model specifically includes:

[0095] Step 101: Input the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding third semantic map;

[0096] The first CT image includes multiple first voxel points; each first voxel point corresponds to a first voxel point pixel value; the three-dimensional size of the third semantic map is consistent with the three-dimensional size of the first CT image; the third semantic map includes multiple fourth voxel points, with a one-to-one correspondence between the first and fourth voxel points; each fourth voxel point corresponds to a binary gastric tumor lesion area semantic, with a value of 0 or 1; when the gastric tumor lesion area semantic is 0, it indicates that the corresponding fourth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fourth voxel point is a gastric tumor lesion area voxel point.

[0097] Step 102: Count the number of fourth voxel points with semantic value 1 in the gastric tumor lesion area on the third semantic map to obtain the corresponding third number; and identify whether the third number exceeds the preset first number threshold. If it exceeds the threshold, proceed to step 103; if it does not exceed the threshold, set the corresponding first prediction type to type M0 and proceed to step 108.

[0098] Step 103: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding fourth semantic map;

[0099] The fourth semantic map has the same three-dimensional dimensions as the first CT image. The fourth semantic map includes multiple fifth voxel points, with a one-to-one correspondence between the first and fifth voxel points. Each fifth voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding fifth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fifth voxel point is a remote critical tissue lesion voxel point.

[0100] Step 104: Count the number of fifth voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding fourth number; and identify whether the fourth number exceeds the preset second number threshold. If it exceeds, proceed to step 105; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 108.

[0101] Step 105: Each first voxel point in the first CT image is taken as the corresponding current voxel point; the fourth and fifth voxel points on the third and fourth semantic maps corresponding to the current voxel points are recorded as the corresponding third and fourth matching voxel points; the pixel value of the first voxel point of the current voxel point is taken as a corresponding feature voxel point pixel value; when the semantics of the gastric tumor lesion area and the semantics of the distal critical tissue lesion area corresponding to the third and fourth matching voxel points are both 0, a corresponding feature voxel point semantics is set to 0; and when the semantics of the gastric tumor lesion area corresponding to the third matching voxel point is... When the semantics of the feature voxel point are set to 1, a corresponding feature voxel point is set to 2; and when the semantics of the distal critical tissue lesion area corresponding to the fourth matching voxel point are 1, a corresponding feature voxel point is set to 2; and a corresponding second feature voxel point is constructed based on the three-dimensional voxel coordinates of the current voxel point, denoted as the corresponding current feature voxel point; and a feature voxel point feature vector corresponding to the current feature voxel point is formed by the pixel value of the feature voxel point corresponding to the current voxel point and the semantics of the feature voxel point; and a corresponding first feature map is formed by all the second feature voxel points corresponding to the first CT image.

[0102] Step 106: Input the first feature map into the second prediction model to perform M-period classification prediction to obtain the corresponding fifth prediction vector;

[0103] The fifth prediction vector includes two fifth-type prediction probabilities, each of which corresponds to one of the M-stage types.

[0104] Step 107: Take the M-stage type corresponding to the fifth type prediction probability with the highest probability in the fifth prediction vector as the corresponding first prediction type;

[0105] Step 108: Output the obtained first prediction type as the prediction result for this time.

[0106] A second aspect of the present invention provides an apparatus for implementing the processing method for M-phase classification prediction based on CT images as described in the first aspect above, the apparatus comprising: a model building module, a model training module, and a model application module;

[0107] The model building module is used to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of gastric tumor lesions on CT images, denoted as the corresponding first semantic segmentation model; and to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of lesion areas of distal key tissues on the CT images, denoted as the corresponding second semantic segmentation model; and to construct a prediction model based on a machine learning classification model for M-stage classification prediction based on radiomics features, denoted as the corresponding first prediction model; and to construct a prediction model based on a deep learning classification model for M-stage classification prediction based on 3D CT images with lesion point markings, denoted as the corresponding second prediction model; wherein... The CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the distal key tissues include at least the liver, the para-aortic lymph node region, the retropancreatic lymph node region, the root mesenteric lymph node region, and the peritoneum; the types of machine learning classification models include at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models; the types of deep learning classification models include at least classification prediction models based on ResNet models and classification prediction models based on Transformer models;

[0108] The model training module is used to train the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively;

[0109] The model application module is used to receive a first CT image and a first prediction mode after the training of four types of models; and to identify the first prediction mode; if the first prediction mode is a first mode, then a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction mode is a second mode, then a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and the obtained first prediction type is output as the prediction result corresponding to the first CT image; wherein, the first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first prediction mode includes a first mode and a second mode; the first prediction type includes M0 type and M1 type.

[0110] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0111] The processor is used to couple with the memory, read and execute instructions in the memory to implement the steps of the method described in the first aspect above;

[0112] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0113] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a computer, cause the computer to perform the instructions described in the first aspect.

[0114] This invention provides a processing method, apparatus, electronic device, and computer-readable storage medium for M-stage classification prediction based on CT images. As described above, this invention constructs two semantic segmentation models based on the U-Net model structure. The first semantic segmentation model is used for semantic segmentation of semantic points in the gastric tumor lesion area on CT images, and the second semantic segmentation model is used for semantic segmentation of semantic points in the lesion area of ​​distal key tissues on CT images. A first prediction model capable of M-stage classification prediction based on input radiomics features is constructed based on a type of machine learning classification model, and a second prediction model capable of M-stage classification prediction based on input 3D CT images with lesion point markings is constructed based on a type of deep learning classification model. These four models are trained separately. After training, different combinations of prediction models are selected based on the received prediction mode: when the prediction mode is the first mode, the first and second semantic segmentation models and the first prediction model are used to predict the M-stage of the received CT images; when the prediction mode is the second mode, the first and second semantic segmentation models and the second prediction model are used to predict the M-stage of the received CT images. The embodiments of the present invention improve both the real-time performance and the efficiency of prediction. Attached Figure Description

[0115] Figure 1 is a schematic diagram of a processing method for M-stage classification prediction based on CT images provided in Embodiment 1 of the present invention;

[0116] Figure 2 is a module structure diagram of a processing device for M-phase classification prediction based on CT images provided in Embodiment 2 of the present invention;

[0117] Figure 3 is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0118] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0119] Embodiment 1 of the present invention provides a processing method for M-stage classification prediction based on CT images. Figure 1 is a schematic diagram of the processing method for M-stage classification prediction based on CT images provided in Embodiment 1 of the present invention. The method mainly includes the following steps:

[0120] Step 1: Based on the U-Net model, construct a semantic segmentation model for semantic point segmentation of gastric tumor lesions on CT images, denoted as the first semantic segmentation model; and based on the U-Net model, construct a semantic segmentation model for semantic point segmentation of lesions in distal key tissues on CT images, denoted as the second semantic segmentation model; and based on a type of machine learning classification model, construct a prediction model for M-stage classification based on radiomics features, denoted as the first prediction model; and based on a type of deep learning classification model, construct a prediction model for M-stage classification based on 3D CT images with lesion point markings, denoted as the second prediction model.

[0121] Here, the CT image in this embodiment of the invention is a three-dimensional abdominal CT image with three-dimensional information of the stomach and distal key tissues; wherein, the distal key tissues include at least the liver, the para-aortic lymph node region, the retropancreatic lymph node region, the mesenteric root lymph node region, and the peritoneum. The machine learning classification models in this embodiment of the invention include at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models. The deep learning classification models in this embodiment of the invention include at least classification prediction models based on ResNet models and classification prediction models based on Transformer models.

[0122] Step 2: Train the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively;

[0123] Specifically, this includes: Step 21, recruiting multiple subjects who do not have gastric cancer, multiple subjects who have gastric cancer but have not metastasized to any distant key tissues, and multiple subjects who have gastric cancer and have metastasized to one or more types of distant key tissues to form the first subject set;

[0124] The first subject set includes multiple first subjects, each corresponding to a first M phase type; the first M phase type includes M0 type and M1 type; the first M phase type corresponding to the first subject who does not have gastric cancer or has gastric cancer but has not metastasized to any distant key tissues is M0 type, and the first M phase type corresponding to the first subject who has gastric cancer and has metastasized to one or more types of distant key tissues is M1 type;

[0125] Step 22: Collect the most recent CT images of each first subject as the corresponding first sample images; and combine each first sample image and the corresponding first M-phase type to form a corresponding first raw data record; and combine all the obtained first raw data records to form the corresponding first raw dataset;

[0126] The first sample image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and distal key tissues; the first sample image includes multiple first sample voxels; each first sample voxel corresponds to a first voxel pixel value;

[0127] The first raw dataset obtained in this embodiment of the invention includes multiple first raw data records; the first raw data records include first sample images and corresponding first M-period types; the first M-period types include M0 type and M1 type;

[0128] Step 23: Construct four types of model training datasets based on the first original dataset, denoted as the first model dataset, the second model dataset, the third model dataset, and the fourth model dataset.

[0129] Specifically, it includes: step 2301, taking each of the first original data records of the first original dataset as the corresponding current original data record; and taking the first sample image and the corresponding first M-period type of the current original data record as the corresponding current sample image and current sample type;

[0130] Step 2302: The current sample image is used as the corresponding first training image and second training image; and the corresponding first and second label semantic maps are constructed based on the three-dimensional dimensions of the first and second training images; and the first and second label semantics of all first and second label voxels in the first and second label semantic maps are initialized to 0.

[0131] Step 2303: Based on manual annotation or other machine annotation methods, mark all first sample voxel points located in the gastric tumor lesion area on the current sample image as Class I voxel points, and mark all first sample voxel points located in the lesion areas of all distant key tissues on the current sample image as Class II voxel points.

[0132] Step 2304, and when the total number of voxel points of a certain type is zero, set a set of all zero first-order feature set, shape feature set and texture feature set to form a corresponding first gastric tumor feature set;

[0133] Step 2305: When the total number of voxel points of a class is greater than zero, reset the first label semantic of the first label voxel points corresponding to each class of voxel points on the first label semantic map to 1; and record the three-dimensional structure composed of all class of voxel points on the current sample image as the corresponding class of structure; and calculate the first-order feature set, shape feature set and texture feature set of the radiomics features of the class of structure to form a corresponding first gastric tumor feature set;

[0134] Step 2306: When the total number of Class II voxel points is zero, set a set of all-zero first-order feature sets, shape feature sets, and texture feature sets for each remote critical tissue to form a corresponding first remote critical tissue feature set; and form a corresponding first remote critical tissue feature set by all the obtained first remote critical tissue feature sets.

[0135] Step 2307: When the total number of Class II voxel points is greater than zero, reset the second label semantics of the second label voxel points corresponding to each Class II voxel point on the second label semantic map to 1; and based on the preset point clustering algorithm, cluster the Class II voxel points corresponding to each remote critical tissue on the current sample image to obtain the corresponding first point group; and take each first point group as the corresponding current point group, and identify whether the current point group is empty; if the current point group is empty, set a set of all zero first-order feature set, shape feature set and texture feature set for the remote critical tissue corresponding to the current point group to form a corresponding first remote critical tissue feature set; if the current point group is not empty, record the three-dimensional structure composed of the current point group on the current sample image as the corresponding Class II structure, and calculate the first-order feature set, shape feature set and texture feature set of the image omics features of the Class II structure to form a corresponding first remote critical tissue feature set; and form a corresponding first remote critical tissue feature set by all the obtained first remote critical tissue feature sets;

[0136] Here, the calculation methods for the first-order features, shape features, and texture features of radiomics features mentioned in the embodiments of the present invention are three publicly disclosed radiomics feature calculation methods. The calculation details can be obtained by consulting the publicly available technical documents, and will not be elaborated further here. The point clustering algorithm in the embodiments of the present invention is a conventional clustering algorithm, such as the K-means clustering algorithm, the DBSCAN clustering algorithm, etc., and the algorithm can be configured according to the actual application requirements.

[0137] Step 2308, and a corresponding first training image omics feature set is formed by the first gastric tumor feature set and the first distal key tissue feature set corresponding to the current sample image;

[0138] Step 2309: Each first sample voxel point in the current sample image is taken as the corresponding current sample voxel point; the first and second label voxels on the first and second label semantic maps corresponding to the current sample voxel point are recorded as the corresponding first and second matching voxels; the pixel value of the first voxel point of the current sample voxel point is taken as the pixel value of the corresponding second voxel point; when the first and second label semantics corresponding to the first and second matching voxels are both 0, a corresponding second voxel point semantic is set to 0; when the first label semantic of the first matching voxel point is 1, a corresponding second voxel point semantic is set to 1; when the second label semantic of the second matching voxel point is 1, a corresponding second voxel point semantic is set to 2; a corresponding first feature voxel point is constructed based on the three-dimensional voxel coordinates of the current sample voxel point and recorded as the corresponding current feature voxel point; a second voxel point feature vector corresponding to the current feature voxel point is composed of the pixel value of the second voxel point and the semantics of the second voxel point corresponding to the current sample voxel point; and a corresponding first training feature map is composed of all the first feature voxels corresponding to the current sample image.

[0139] Here, the first training feature map in this embodiment of the invention is essentially a three-dimensional CT image with lesion point markings;

[0140] Step 2310, and construct the corresponding first and second label type vectors according to the current sample type;

[0141] Specifically, when the current sample type is M0, the probability of the first and second label types corresponding to type M0 in the corresponding first and second label type vectors is 1 and the probability of the first and second label types corresponding to type M1 is 0; when the current sample type is M1, the probability of the first and second label types corresponding to type M1 in the corresponding first and second label type vectors is 0 and the probability of the first and second label types corresponding to type M1 is 1.

[0142] Step 2311: A first model data record is formed by combining the first training image and the first label semantic map corresponding to the current original data record; a second model data record is formed by combining the second training image and the second label semantic map corresponding to the current original data record; a third model data record is formed by combining the first training image omics feature set and the first label type vector corresponding to the current original data record; and a fourth model data record is formed by combining the first training feature map and the second label type vector corresponding to the current original data record.

[0143] Step 2312, and the first model dataset is composed of all the first model data records corresponding to the first original dataset; the second model dataset is composed of all the second model data records corresponding to the first original dataset; the third model dataset is composed of all the third model data records corresponding to the first original dataset; and the fourth model dataset is composed of all the fourth model data records corresponding to the first original dataset.

[0144] Here, the first model dataset obtained in this embodiment of the invention includes multiple first model data records; the first model data records include a first training image and a first label semantic map; the three-dimensional size of the first label semantic map is consistent with the three-dimensional size of the first training image; the first label semantic map includes multiple first label voxels; each first label voxel corresponds to a binary first label semantic, which takes a value of 0 or 1; when the first label semantic is 0, it indicates that the corresponding first label voxel is a background voxel, and when it is 1, it indicates that the corresponding first label voxel is a gastric tumor lesion area voxel;

[0145] The second model dataset obtained in this embodiment of the invention includes multiple second model data records; the second model data records include a second training image and a second label semantic map; the three-dimensional size of the second label semantic map is consistent with the three-dimensional size of the second training image; the second label semantic map includes multiple second label voxels; each second label voxel corresponds to a binary second label semantic, with a value of 0 or 1; when the second label semantic is 0, it indicates that the corresponding first label voxel is a background voxel; when it is 1, it indicates that the corresponding second label voxel is a distant key tissue lesion voxel.

[0146] The third model dataset obtained in this embodiment of the invention includes multiple third model data records; the third model data records include a first training radiomics feature set and a first label type vector; the first training radiomics feature set includes a first gastric tumor feature set and a first distal key tissue feature set; the first distal key tissue feature set includes multiple first distal key tissue feature sets, each first distal key tissue feature set corresponding to a distal key tissue; the first gastric tumor feature set and all first distal key tissue feature sets are composed of a set of first-order feature sets, shape feature sets, and texture feature sets of radiomics features; the first label type vector includes two first label type probabilities, each first label type probability corresponding to an M-staging type; the M-staging type includes M0 type and M1 type;

[0147] The fourth model dataset obtained in this embodiment of the invention includes multiple fourth model data records; each fourth model data record includes a first training feature map and a second label type vector; the first training feature map includes multiple first feature voxels; each first feature voxel corresponds to a second voxel feature vector; the second voxel feature vector includes at least the second voxel pixel value and the second voxel semantics; the second voxel semantics is a ternary semantic value, taking the value 0, 1, or 2; when the second voxel semantics is 0, it indicates that the corresponding first feature voxel is a background voxel; when it is 1, it indicates that the corresponding first feature voxel is a gastric tumor lesion voxel; when it is 2, it indicates that the corresponding first feature voxel is a distal critical tissue lesion voxel; the second label type vector includes two second label type probabilities, each second label type probability corresponding to an M-stage type.

[0148] Step 24: Train the first semantic segmentation model based on the first model dataset;

[0149] Specifically, this includes: step 241, taking the first first model data record of the first model dataset as the corresponding current model data record;

[0150] Step 242: Input the first training image recorded in the current model data into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first predicted semantic map;

[0151] Step 243: Input the first predicted semantic map and the first label semantic map of the current model data record into the preset first model loss function to calculate the corresponding first loss value;

[0152] The first model loss function is implemented based on either the L1 loss function or the L2 loss function.

[0153] Step 244: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current model data record is the last first model data record in the first model dataset; if so, proceed to step 245; otherwise, take the next first model data record in the first model dataset as the new current model data record and return to step 242 to continue training; if the first loss value does not meet the first loss value range, optimize the first semantic segmentation model parameters in one round based on the preset first model parameter optimizer in the direction of minimizing the first model loss function, and return to step 242 to continue training when the parameter optimization ends.

[0154] The first model parameter optimizer includes at least the SDG optimizer; the first loss range is a pre-set loss range.

[0155] Step 245: Solidify the model parameters of the first semantic segmentation model and confirm that the training of the first semantic segmentation model has ended;

[0156] Step 25: Train the second semantic segmentation model based on the second model dataset;

[0157] Specifically, this includes: Step 251, taking the first second model data record of the second model dataset as the corresponding current model data record;

[0158] Step 252: Input the second training image recorded in the current model data into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the remote key tissue to obtain the corresponding second predicted semantic map;

[0159] Step 253: Input the second predicted semantic map and the second label semantic map of the current model data record into the preset second model loss function to calculate the corresponding second loss value;

[0160] The second model loss function is implemented based on either the L1 loss function or the L2 loss function.

[0161] Step 254: Identify whether the second loss value meets the preset range of the second loss value; if the second loss value meets the range of the second loss value, identify whether the current model data record is the last second model data record in the second model dataset; if so, proceed to step 255; otherwise, take the next second model data record in the second model dataset as the new current model data record and return to step 252 to continue training; if the second loss value does not meet the range of the second loss value, optimize the second semantic segmentation model parameters in one round based on the preset second model parameter optimizer in the direction of minimizing the second model loss function, and return to step 252 to continue training when the parameter optimization ends.

[0162] The second model parameter optimizer includes at least the SDG optimizer; the second loss range is a pre-set loss range.

[0163] Step 255: Solidify the model parameters of the second semantic segmentation model and confirm that the training of the second semantic segmentation model has ended;

[0164] Step 26: Train the first prediction model based on the third model dataset;

[0165] Specifically, it includes: step 2601, performing a full combination of the hyperparameters of the first prediction model to obtain multiple corresponding first hyperparameter combinations; and setting the hyperparameters of the first prediction model once based on each first hyperparameter combination to obtain the corresponding first prediction model;

[0166] Step 2602: Set parameter K to a positive integer greater than or equal to 5; and divide the third model dataset into K equal subsets, denoting each subset as the corresponding first subset D. i ;

[0167] Where 1 ≤ index i ≤ K;

[0168] Step 2603, take any first subset D i DB serves as the corresponding first validation dataset i And will exclude the current first subset D i The remaining K-1 first subsets D i Merge into a corresponding first test dataset DA i ; and from the obtained first test datasets DA i and the corresponding first validation dataset DB i Form a corresponding first test and verification set;

[0169] Step 2604: Take the first prediction model as the corresponding current prediction model;

[0170] Step 2605: Take the first test validation set as the corresponding current test validation set;

[0171] Step 2606: The first test dataset DA of the current test validation set... i and the first validation dataset DB i As the corresponding current test dataset and current validation dataset;

[0172] Step 2607: Take the first third model data record of the current test dataset as the corresponding current model data record;

[0173] Step 2608: Input the first training image omics feature set of the current model data record into the current prediction model to perform M-period classification prediction to obtain the corresponding first prediction vector;

[0174] The first prediction vector includes two first-type prediction probabilities, each of which corresponds to an M-stage type.

[0175] Step 2609: Substitute the first prediction vector and the first label type vector of the current model data record into the preset third model loss function; and perform a round of model parameter optimization on the model parameters of the current prediction model based on the preset third model parameter optimizer in the direction of minimizing the third model loss function; and at the end of this round of parameter optimization, identify whether the current model data record is the last third model data record of the current test dataset. If so, proceed to step 2610; otherwise, extract the next third model data record of the current test dataset as the new current model data record and return to step 2608 for training.

[0176] The third model loss function is based on the binary cross-entropy loss function; the third model parameter optimizer includes at least the SDG optimizer.

[0177] Step 2610: Set three counters initially set to 0 and denoted as TP counter, FP counter and FN counter;

[0178] Step 2611: Traverse all third model data records in the current validation dataset; during traversal, use the currently traversed third model data record as the corresponding current validation record; use the first training radiomics feature set and the first label type vector of the current validation record as the corresponding current radiomics feature set and the current label type vector; use the M stage type corresponding to the first label type probability with a probability value of 1 in the current label type vector as the corresponding current label type; input the current radiomics feature set into the current prediction model to perform M stage classification prediction to obtain the corresponding second prediction vector; use the M stage type corresponding to the second type prediction probability with the highest probability in the second prediction vector as the corresponding current prediction type; compare the current prediction type with the current label type; if both the current prediction type and the current label type are M1 type, increment the TP counter by 1; if the current prediction type is M1 type and the current label type is M0 type, increment the FP counter by 1; if the current prediction type is M0 type and the current label type is M1 type, increment the FN counter by 1.

[0179] The second prediction vector includes two second-type prediction probabilities, each of which corresponds to an M-stage type.

[0180] Step 2612: After traversing all third-model data records in the current validation dataset, calculate the first precision and first recall based on the TP counter, FP counter, and FN counter; and calculate the first F1 score based on the first precision and first recall. i ;

[0181] here,

[0182] First accuracy

[0183] First Recall Rate

[0184] First score

[0185] Step 2613: Identify whether the current test validation set is the Kth first test validation set; if yes, proceed to step 2614; if no, take the next first test validation set as the new current test validation set and return to step 2606.

[0186] Step 2614, based on the obtained K first scores F1 i The first error E is obtained by calculating the mean square error.

[0187] here,

[0188] Step 2615: Identify whether the current prediction model is the last first-to-last prediction model; if yes, proceed to step 2616; if no, use the next first-to-last prediction model as the new current prediction model and return to step 2605.

[0189] Step 2616: Take the full model parameter set of the first prediction model corresponding to the first error E with the smallest error as the corresponding optimal full model parameter set;

[0190] Step 2617: Based on the optimal full model parameter set, solidify the model parameters of the first prediction model and confirm that the model training of the first prediction model is completed;

[0191] Step 27: Train the second prediction model based on the fourth model dataset;

[0192] Specifically, this includes: Step 271, taking the first fourth model data record of the fourth model dataset as the corresponding current model data record;

[0193] Step 272: Input the first training feature map of the current model data record into the second prediction model to perform M-period classification prediction to obtain the corresponding third prediction vector;

[0194] The third prediction vector includes two third-type prediction probabilities, each of which corresponds to an M-stage type.

[0195] Step 273: Input the third prediction vector and the second label type vector of the current model data record into the preset fourth model loss function to calculate the corresponding third loss value;

[0196] The loss function of the fourth model is based on the binary classification cross-entropy loss function;

[0197] Step 274: Identify whether the third loss value meets the preset range of the third loss value; if the third loss value meets the range of the third loss value, identify whether the current model data record is the last fourth model data record in the fourth model dataset; if so, proceed to step 275; otherwise, take the next fourth model data record in the fourth model dataset as the new current model data record and return to step 272 to continue training; if the third loss value does not meet the range of the third loss value, optimize the second prediction model parameters in one round based on the preset fourth model parameter optimizer in the direction of minimizing the fourth model loss function, and return to step 272 to continue training when the parameter optimization ends.

[0198] The fourth model parameter optimizer includes at least the SDG optimizer; the third loss range is a pre-set loss range.

[0199] Step 275: Solidify the model parameters of the second prediction model and confirm that the training of the second prediction model is complete.

[0200] Step 28, and confirm the completion of training for the four types of models after the first and second semantic segmentation models and the first and second prediction models have all completed model training.

[0201] Step 3: After the four types of models have been trained, the first CT image and the first prediction pattern are received; the first prediction pattern is identified; if the first prediction pattern is the first pattern, the corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction pattern is the second pattern, the corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and the obtained first prediction type is output as the prediction result corresponding to the first CT image.

[0202] The first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and distal key tissues; the first CT image includes multiple first voxel points; each first voxel point corresponds to a first voxel point pixel value; the first prediction mode includes a first mode and a second mode; the first prediction type includes M0 type and M1 type.

[0203] Specifically, this includes: Step 31, after the training of the four types of models is completed, receiving the first CT image and the first prediction pattern;

[0204] Step 32, and identify the first prediction pattern;

[0205] Step 33: If the first prediction mode is the first mode, then the corresponding first prediction type is obtained by predicting based on the first CT image, the first semantic segmentation model, the second semantic segmentation model and the first prediction model.

[0206] Specifically, it includes: Step 331, inputting the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first semantic map;

[0207] The three-dimensional dimensions of the first semantic map are consistent with those of the first CT image. The first semantic map includes multiple second voxel points, and the first and second voxel points correspond one-to-one. Each second voxel point corresponds to a binary semantic area of ​​a gastric tumor lesion, with a value of 0 or 1. When the semantic area of ​​a gastric tumor lesion is 0, it indicates that the corresponding second voxel point is a background voxel point; when it is 1, it indicates that the corresponding second voxel point is a gastric tumor lesion voxel point.

[0208] Step 332: Count the number of second voxel points with semantic value 1 in the gastric tumor lesion area on the first semantic map to obtain the corresponding first number; and identify whether the first number exceeds the preset first number threshold. If it exceeds, proceed to step 333; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 339.

[0209] Here, the first quantity threshold is a pre-set integer value;

[0210] Step 333: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding second semantic map;

[0211] The second semantic map has the same three-dimensional dimensions as the first CT image. The second semantic map includes multiple third voxel points, with a one-to-one correspondence between the first and third voxel points. Each third voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding third voxel point is a background voxel point; when it is 1, it indicates that the corresponding third voxel point is a remote critical tissue lesion voxel point.

[0212] Step 334: Count the number of third voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding second number; and identify whether the second number exceeds the preset second number threshold. If it exceeds, proceed to step 335; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 339.

[0213] Here, the second quantity threshold is a pre-set integer value;

[0214] Step 335: Record the first voxel points on the first CT image that correspond to the second voxel points with semantic value of 1 in each distal key tissue lesion area as the corresponding first lesion points; record the three-dimensional structure composed of all the first lesion points as the corresponding first structure; and calculate the first-order features, shape features and texture features of the radiomics features of the first structure to obtain the corresponding first-order feature set, shape feature set and texture feature set to form a corresponding second gastric tumor feature set;

[0215] Step 336: Based on a preset point clustering algorithm, cluster the first voxel points on the first CT image corresponding to each distal critical tissue to obtain the corresponding second point clusters; and use each second point cluster as the corresponding current point cluster, and identify whether the current point cluster is empty; if the current point cluster is empty, set a set of all-zero first-order feature set, shape feature set, and texture feature set for the distal critical tissue corresponding to the current point cluster to form a corresponding second distal critical tissue feature set; if the current point cluster is not empty, record the three-dimensional structure on the first CT image composed of the current point clusters as the corresponding second structure, and calculate the first-order feature set, shape feature set, and texture feature set of the radiomics features of the second structure to form a corresponding second distal critical tissue feature set; and use all the obtained second distal critical tissue feature sets to form a corresponding second distal critical tissue feature set;

[0216] Step 337: The obtained second gastric tumor feature set and second distal key tissue feature set are combined to form a corresponding first radiomics feature set; and the first radiomics feature set is input into the first prediction model to perform M-phase classification prediction to obtain the corresponding fourth prediction vector;

[0217] The fourth prediction vector includes two fourth-type prediction probabilities, each of which corresponds to an M-stage type.

[0218] Step 338: Take the M-stage type corresponding to the fourth type prediction probability with the highest probability in the fourth prediction vector as the corresponding first prediction type;

[0219] Step 339: Output the first prediction type obtained as the prediction result for this time;

[0220] Step 34: If the first prediction mode is the second mode, then the corresponding first prediction type is obtained by predicting based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model.

[0221] Specifically, this includes: Step 341, inputting the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding third semantic map;

[0222] The three-dimensional dimensions of the third semantic map are consistent with those of the first CT image. The third semantic map includes multiple fourth voxel points, with a one-to-one correspondence between the first and fourth voxel points. Each fourth voxel point corresponds to a binary semantic representation of a gastric tumor lesion area, with a value of 0 or 1. When the semantic representation of the gastric tumor lesion area is 0, it indicates that the corresponding fourth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fourth voxel point is a gastric tumor lesion area voxel point.

[0223] Step 342: Count the number of fourth voxel points with semantic value 1 in the gastric tumor lesion area on the third semantic map to obtain the corresponding third number; and identify whether the third number exceeds the preset first number threshold. If it exceeds, proceed to step 343; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 348.

[0224] Step 343: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding fourth semantic map;

[0225] The three-dimensional dimensions of the fourth semantic map are consistent with those of the first CT image. The fourth semantic map includes multiple fifth voxel points, with a one-to-one correspondence between the first and fifth voxel points. Each fifth voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding fifth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fifth voxel point is a remote critical tissue lesion voxel point.

[0226] Step 344: Count the number of fifth voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding fourth number; and identify whether the fourth number exceeds the preset second number threshold. If it exceeds, proceed to step 345; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 348.

[0227] Step 345: Each first voxel point in the first CT image is taken as the corresponding current voxel point; the fourth and fifth voxel points on the third and fourth semantic maps corresponding to the current voxel points are recorded as the corresponding third and fourth matching voxel points; the pixel value of the first voxel point of the current voxel point is taken as a corresponding feature voxel point pixel value; when the semantics of the gastric tumor lesion area and the semantics of the distal key tissue lesion area corresponding to the third and fourth matching voxel points are both 0, a corresponding feature voxel point semantics is set to 0; and when the semantics of the gastric tumor lesion area corresponding to the third matching voxel point is... When the semantics of a feature voxel point are set to 1, a corresponding feature voxel point is set to 2 when the semantics of the distal critical tissue lesion area corresponding to the fourth matching voxel point are 1. A corresponding second feature voxel point is constructed based on the three-dimensional voxel coordinates of the current voxel point and denoted as the corresponding current feature voxel point. A feature vector of the feature voxel point corresponding to the current feature voxel point is formed by the pixel value of the feature voxel point corresponding to the current voxel point and the semantics of the feature voxel point. A corresponding first feature map is formed by all the second feature voxel points corresponding to the first CT image.

[0228] Here, the first feature map in this embodiment of the invention is essentially a three-dimensional CT image with lesion point markings;

[0229] Step 346: Input the first feature map into the second prediction model to perform M-period classification prediction to obtain the corresponding fifth prediction vector;

[0230] The fifth prediction vector includes two fifth-type prediction probabilities, each of which corresponds to an M-stage type.

[0231] Step 347: Take the M-stage type corresponding to the fifth type prediction probability with the highest probability in the fifth prediction vector as the corresponding first prediction type;

[0232] Step 348: Output the first prediction type as the prediction result for this time;

[0233] Step 35, and output the obtained first prediction type as the prediction result corresponding to the first CT image.

[0234] Figure 2 is a block diagram of a processing device for M-phase classification prediction based on CT images provided in Embodiment 2 of the present invention. This device can be a terminal device or server that implements the aforementioned method embodiments, or it can be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiments. For example, the device can be a device or chip system of the aforementioned terminal device or server. As shown in Figure 2, the device includes: a model building module 201, a model training module 202, and a model application module 203.

[0235] The model building module 201 is used to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of lesion areas of gastric tumors on CT images, denoted as the corresponding first semantic segmentation model; and to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of lesion areas of distal key tissues on CT images, denoted as the corresponding second semantic segmentation model; and to construct a prediction model based on a type of machine learning classification model for M-stage classification prediction based on radiomics features, denoted as the corresponding first prediction model; and to construct a prediction model based on a type of deep learning classification model for M-stage classification prediction based on 3D CT images with lesion point markers, denoted as the corresponding second prediction model. The model includes a CT image of the abdomen containing three-dimensional information of the stomach and distal key tissues. The distal key tissues include at least the liver, para-aortic lymph node region, retropancreatic lymph node region, mesenteric root lymph node region, and peritoneum. The machine learning classification model includes at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models. The deep learning classification model includes at least classification prediction models based on ResNet models and classification prediction models based on Transformer models.

[0236] The model training module 202 is used to train the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively.

[0237] The model application module 203 is used to receive a first CT image and a first prediction mode after the training of four types of models; and to identify the first prediction mode; if the first prediction mode is a first mode, then predict the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction mode is a second mode, then predict the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and output the obtained first prediction type as the prediction result corresponding to the first CT image; wherein, the first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and distal key tissues; the first prediction mode includes a first mode and a second mode; the first prediction type includes M0 type and M1 type.

[0238] The present invention provides a processing device for M-stage classification prediction based on CT images, which can execute the method steps in the above method embodiments. Its implementation principle and technical effect are similar, and will not be repeated here.

[0239] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing elements; they can be fully implemented in hardware; or some modules can be implemented by processing elements calling software, while others are implemented in hardware. For example, the model building module can be a separate processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0240] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-on-a-Chip (SOC).

[0241] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the foregoing method embodiments are generated. The computer described above can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The aforementioned computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the aforementioned computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, Bluetooth, microwave, etc.) means. The aforementioned computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0242] Figure 3 is a schematic diagram of an electronic device provided in Embodiment 3 of the present invention. This electronic device can be a terminal device or server implementing the methods of the aforementioned embodiments, or it can be a terminal device or server connected to the aforementioned terminal device or server implementing the methods of the aforementioned embodiments. As shown in Figure 3, the electronic device may include: a processor 301 (e.g., CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transmission and reception operations of the transceiver 303. The memory 302 may store various instructions for completing various processing functions and implementing the processing steps described in the methods of the aforementioned embodiments. Preferably, the electronic device involved in the embodiments of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The communication port 306 is used for communication between the electronic device and other peripherals.

[0243] The system bus 305 mentioned in Figure 3 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write libraries, and read-only libraries). Memory may include Random Access Memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0244] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0245] It should be noted that the embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when run on a computer, cause the computer to perform the methods and processes provided in the above embodiments.

[0246] This invention also provides a chip for executing instructions, which is used to perform the processing steps described in the foregoing method embodiments.

[0247] This invention provides a processing method, apparatus, electronic device, and computer-readable storage medium for M-stage classification prediction based on CT images. As described above, this invention constructs two semantic segmentation models based on the U-Net model structure. The first semantic segmentation model is used for semantic segmentation of semantic points in the gastric tumor lesion area on CT images, and the second semantic segmentation model is used for semantic segmentation of semantic points in the lesion area of ​​distal key tissues on CT images. A first prediction model capable of M-stage classification prediction based on input radiomics features is constructed based on a type of machine learning classification model, and a second prediction model capable of M-stage classification prediction based on input 3D CT images with lesion point markings is constructed based on a type of deep learning classification model. These four models are trained separately. After training, different combinations of prediction models are selected based on the received prediction mode: when the prediction mode is the first mode, the first and second semantic segmentation models and the first prediction model are used to predict the M-stage of the received CT images; when the prediction mode is the second mode, the first and second semantic segmentation models and the second prediction model are used to predict the M-stage of the received CT images. The embodiments of the present invention improve both the real-time performance and the efficiency of prediction.

[0248] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0249] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0250] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A processing method for M-stage classification prediction based on CT images, characterized in that, The method includes: A semantic segmentation model based on the U-Net model is constructed to segment the lesion area of ​​a gastric tumor on CT images, denoted as the first semantic segmentation model; a semantic segmentation model based on the U-Net model is constructed to segment the lesion area of ​​distal key tissues on the CT images, denoted as the second semantic segmentation model; a prediction model based on a machine learning classification model is constructed to predict the M-stage based on radiomics features, denoted as the first prediction model; and a prediction model based on a deep learning classification model is constructed to predict the M-stage based on a 3D CT image with lesion point markers, denoted as the second prediction model; wherein, the CT... The image is a three-dimensional abdominal CT scan with three-dimensional information of the stomach and the aforementioned distal key tissues; the distal key tissues include at least the liver, para-aortic lymph node region, retropancreatic lymph node region, root mesenteric lymph node region, and peritoneum; the machine learning classification model includes at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models; the deep learning classification model includes at least classification prediction models based on ResNet models and classification prediction models based on Transformer models; The first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model are trained respectively. After the four types of models are trained, a first CT image and a first prediction mode are received; the first prediction mode is identified; if the first prediction mode is a first mode, a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction mode is a second mode, a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and the obtained first prediction type is output as the prediction result corresponding to the first CT image; wherein, the first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first prediction mode includes a first mode and a second mode; the first prediction type includes type M0 and type M1.

2. The processing method for M-stage classification prediction based on CT images according to claim 1, characterized in that, The training of the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively includes: A first subject set is formed by recruiting multiple subjects who do not have gastric cancer, multiple subjects who have gastric cancer but have not metastasized to any of the aforementioned distant key tissues, and multiple subjects who have gastric cancer and have metastasized to one or more of the aforementioned distant key tissues. The first subject set includes multiple first subjects, each corresponding to a first M phase type. The first M phase type includes M0 type and M1 type. The first M phase type corresponding to the first subjects who do not have gastric cancer or who have gastric cancer but have not metastasized to any of the aforementioned distant key tissues is M0 type, and the first M phase type corresponding to the first subjects who have gastric cancer and have metastasized to one or more of the aforementioned distant key tissues is M1 type. The most recent CT images of each of the first subjects are collected as corresponding first sample images; and each of the first sample images and the corresponding first M-phase type are combined to form a corresponding first raw data record; and all the obtained first raw data records are combined to form a corresponding first raw dataset; the first sample image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first sample image includes multiple first sample voxels; each first sample voxel corresponds to a first voxel pixel value. Based on the first original dataset, four types of model training datasets are constructed and denoted as the first model dataset, the second model dataset, the third model dataset, and the fourth model dataset. The first semantic segmentation model is trained based on the first model dataset; The second semantic segmentation model is trained based on the second model dataset; The first prediction model is trained based on the third model dataset; The second prediction model is trained based on the fourth model dataset; The training of the four types of models was confirmed to be complete after the first and second semantic segmentation models and the first and second prediction models had all completed their training.

3. The processing method for M-stage classification prediction based on CT images according to claim 2, characterized in that, The first model dataset includes multiple first model data records; the first model data records include a first training image and a first label semantic map; the three-dimensional size of the first label semantic map is consistent with the three-dimensional size of the first training image; the first label semantic map includes multiple first label voxels; each first label voxel corresponds to a binary first label semantic, with a value of 0 or 1; when the first label semantic is 0, it indicates that the corresponding first label voxel is a background voxel; when it is 1, it indicates that the corresponding first label voxel is a gastric tumor lesion area voxel. The second model dataset includes multiple second model data records; the second model data records include a second training image and a second label semantic map; the three-dimensional dimensions of the second label semantic map are consistent with the three-dimensional dimensions of the second training image; The second label semantic map includes multiple second label voxel points; each second label voxel point corresponds to a binary second label semantic, which takes the value of 0 or 1; when the second label semantic is 0, it indicates that the corresponding first label voxel point is a background voxel point, and when it is 1, it indicates that the corresponding second label voxel point is a remote critical tissue lesion voxel point; The third model dataset includes multiple third model data records; each third model data record includes a first training radiomics feature set and a first label type vector; the first training radiomics feature set includes a first gastric tumor feature set and a first distal key tissue feature set; the first distal key tissue feature set includes multiple first distal key tissue feature sets, each of which corresponds to one distal key tissue; the first gastric tumor feature set and all first distal key tissue feature sets are composed of a set of first-order feature sets, shape feature sets, and texture feature sets of radiomics features; the first label type vector includes two first label type probabilities, each of which corresponds to an M-stage type; the M-stage type includes M0 type and M1 type; The fourth model dataset includes multiple fourth model data records; each fourth model data record includes a first training feature map and a second label type vector; the first training feature map includes multiple first feature voxels; each first feature voxel corresponds to a second voxel feature vector; the second voxel feature vector includes at least the second voxel pixel value and the second voxel semantics. The second voxel point semantic is a ternary semantic value, which takes the value of 0, 1 or 2; when the second voxel point semantic is 0, it means that the corresponding first feature voxel point is a background voxel point; when it is 1, it means that the corresponding first feature voxel point is a gastric tumor lesion voxel point; when it is 2, it means that the corresponding first feature voxel point is a distal critical tissue lesion voxel point; the second label type vector includes two second label type probabilities, and each second label type probability corresponds to one M stage type.

4. The processing method for M-stage classification prediction based on CT images according to claim 3, characterized in that, The four types of model training datasets constructed based on the first original dataset are denoted as the first model dataset, the second model dataset, the third model dataset, and the fourth model dataset, and specifically include: Each of the first original data records in the first original dataset is taken as the corresponding current original data record; and the first sample image and the corresponding first M-period type of the current original data record are taken as the corresponding current sample image and current sample type. The current sample image is used as the corresponding first training image and second training image; and the corresponding first and second label semantic maps are constructed based on the three-dimensional dimensions of the first and second training images; and the first and second label semantics of all first and second label voxels in the first and second label semantic maps are initialized to 0. Based on manual annotation or other machine annotation methods, all first sample voxel points located in the gastric tumor lesion area on the current sample image are marked as Class I voxel points, and all first sample voxel points located in the lesion areas of all distal key tissues on the current sample image are marked as Class II voxel points. When the total number of voxel points of the aforementioned type is zero, a set of all-zero first-order feature sets, shape feature sets, and texture feature sets is set to form a corresponding first gastric tumor feature set; When the total number of voxel points of the first type is greater than zero, the first label semantic of the first label voxel point corresponding to each voxel point of the first type on the first label semantic map is reset to 1; and the three-dimensional structure composed of all voxel points of the first type on the current sample image is recorded as the corresponding first type structure; and the first-order feature set, shape feature set and texture feature set of the radiomics features of the first type structure are calculated to form a corresponding first gastric tumor feature set; When the total number of the two types of voxel points is zero, a set of all-zero first-order feature sets, shape feature sets, and texture feature sets are set for each of the distal critical tissues to form a corresponding first distal critical tissue feature set; and all the obtained first distal critical tissue feature sets are combined to form a corresponding first distal critical tissue feature set. When the total number of the two types of voxels is greater than zero, the second label semantics of the second label voxels corresponding to each of the two types of voxels on the second label semantic map are reset to 1; and based on a preset point clustering algorithm, the two types of voxels on the current sample image corresponding to each of the remote critical tissues are clustered to obtain the corresponding first point clusters; and each of the first point clusters is used as the corresponding current point cluster, and it is identified whether the current point cluster is empty; if the current point cluster is empty, a set of all-zero first-order feature sets is set for the remote critical tissue corresponding to the current point cluster. The shape feature set and the texture feature set form a corresponding first remote critical tissue feature set; if the current point group is not empty, the three-dimensional structure composed of the current point group on the current sample image is recorded as the corresponding second-class structure, and the first-order features, shape features and texture features of the image omics features of the second-class structure are calculated to obtain the corresponding first-order feature set, the shape feature set and the texture feature set form a corresponding first remote critical tissue feature set; and all the obtained first remote critical tissue feature sets form a corresponding first remote critical tissue feature set; The first training image omics feature set is composed of the first gastric tumor feature set corresponding to the current sample image and the first distal key tissue feature set. And take each of the first sample voxel points of the current sample image as the corresponding current sample voxel point; and take the first and second label voxel points on the first and second label semantic maps corresponding to the current sample image as the corresponding first and second matching voxel points; take the pixel value of the first voxel point of the current sample voxel point as a corresponding second voxel point pixel value; and set a corresponding second voxel point semantic value to 0 when the first and second label semantic values ​​corresponding to the first and second matching voxel points are both 0; and set the first label semantic value corresponding to the first matching voxel point to 0. When the semantics of the second voxel point are set to 1, a corresponding second voxel point semantics of 1 is set to 2; and when the semantics of the second label corresponding to the second matching voxel point are 1, a corresponding second voxel point semantics of 2 is set to 2; and a corresponding first feature voxel point is constructed based on the three-dimensional voxel coordinates of the current sample voxel point, denoted as the corresponding current feature voxel point; and a second voxel point feature vector corresponding to the current feature voxel point is formed by the pixel value of the second voxel point corresponding to the current sample voxel point and the semantics of the second voxel point; and a corresponding first training feature map is formed by all the first feature voxel points corresponding to the current sample image. The first and second label type vectors are constructed according to the current sample type. When the current sample type is M0, the probability of the first and second label types corresponding to type M0 in the corresponding first and second label type vectors is 1, and the probability of the first and second label types corresponding to type M1 is 0. When the current sample type is M1, the probability of the first and second label types corresponding to type M1 in the corresponding first and second label type vectors is 0, and the probability of the first and second label types corresponding to type M1 is 1. The first model data record is composed of the first training image and the first label semantic map corresponding to the current original data record; the second model data record is composed of the second training image and the second label semantic map corresponding to the current original data record; the third model data record is composed of the first training image omics feature set and the first label type vector corresponding to the current original data record; and the fourth model data record is composed of the first training feature map and the second label type vector corresponding to the current original data record. The first model dataset is composed of all the first model data records corresponding to the first original dataset; the second model dataset is composed of all the second model data records corresponding to the first original dataset; the third model dataset is composed of all the third model data records corresponding to the first original dataset; and the fourth model dataset is composed of all the fourth model data records corresponding to the first original dataset.

5. The processing method for M-stage classification prediction based on CT images according to claim 3, characterized in that, The training of the first semantic segmentation model based on the first model dataset specifically includes: Step 51: Take the first first model data record of the first model dataset as the corresponding current model data record; Step 52: Input the first training image recorded in the current model data into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first predicted semantic map. Step 53: Input the first predicted semantic map and the first label semantic map of the current model data record into a preset first model loss function to calculate the corresponding first loss value; The first model loss function is implemented based on the L1 loss function or the L2 loss function; Step 54: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current model data record is the last first model data record in the first model dataset; if yes, proceed to step 55; otherwise, take the next first model data record in the first model dataset as the new current model data record and return to step 52 to continue training; if the first loss value does not meet the first loss value range, optimize the first semantic segmentation model parameters in one round based on the preset first model parameter optimizer in the direction of minimizing the first model loss function, and return to step 52 to continue training when the parameter optimization ends. The first model parameter optimizer includes at least an SDG optimizer; Step 55: Solidify the model parameters of the first semantic segmentation model and confirm that the training of the first semantic segmentation model has ended.

6. The processing method for M-stage classification prediction based on CT images according to claim 3, characterized in that, The training of the second semantic segmentation model based on the second model dataset specifically includes: Step 61: Take the first second model data record of the second model dataset as the corresponding current model data record; Step 62: Input the second training image recorded in the current model data into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the remote key tissue to obtain the corresponding second predicted semantic map; Step 63: Input the second predicted semantic map and the second label semantic map of the current model data record into a preset second model loss function to calculate the corresponding second loss value; The second model loss function is implemented based on either the L1 loss function or the L2 loss function. Step 64: Identify whether the second loss value meets the preset second loss value range; if the second loss value meets the second loss value range, identify whether the current model data record is the last second model data record in the second model dataset; if yes, proceed to step 65; otherwise, take the next second model data record in the second model dataset as the new current model data record and return to step 62 to continue training; if the second loss value does not meet the second loss value range, optimize the second semantic segmentation model in one round based on the preset second model parameter optimizer in the direction of minimizing the second model loss function, and return to step 62 to continue training when the parameter optimization ends. The second model parameter optimizer includes at least an SDG optimizer; Step 65: Solidify the model parameters of the second semantic segmentation model and confirm that the training of the second semantic segmentation model has ended.

7. The processing method for M-stage classification prediction based on CT images according to claim 3, characterized in that, The training of the first prediction model based on the third model dataset specifically includes: Step 701: Perform a full combination of the hyperparameters of the first prediction model to obtain multiple corresponding first hyperparameter combinations; and perform a hyperparameter setting on the first prediction model based on each first hyperparameter combination to obtain the corresponding first prediction model. Step 702: Set parameter K to a positive integer greater than or equal to 5; and divide the third model dataset into K equal subsets, and denote each subset as the corresponding first subset D. i ; 1 ≤ index i ≤ K; Step 703, take any one of the first subsets D i DB serves as the corresponding first validation dataset i And will exclude the current first subset D i The remaining K-1 data points of the first subset D i Merge into a corresponding first test dataset DA i ; and the obtained first test datasets DA i and the corresponding first verification dataset DB i Form a corresponding first test and verification set; Step 704: Take the first prediction model as the corresponding current prediction model; Step 705: Take the first of the first test verification sets as the corresponding current test verification set; Step 706, the first test dataset DA of the current test verification set. i and the first verification dataset DB i As the corresponding current test dataset and current validation dataset; Step 707: Take the first third model data record of the current test dataset as the corresponding current model data record; Step 708: Input the first training image omics feature set recorded in the current model data into the current prediction model to perform M-period classification prediction to obtain the corresponding first prediction vector; Wherein, the first prediction vector includes two first type prediction probabilities, each of the first type prediction probabilities corresponding to one M-stage type; Step 709: Substitute the first prediction vector and the first label type vector of the current model data record into the preset third model loss function; and perform a round of model parameter optimization on the model parameters of the current prediction model based on the preset third model parameter optimizer in the direction of minimizing the third model loss function; and at the end of this round of parameter optimization, identify whether the current model data record is the last third model data record of the current test dataset. If so, proceed to step 710; otherwise, extract the next third model data record of the current test dataset as the new current model data record and return to step 708 for training. The third model loss function is implemented based on the binary cross-entropy loss function; the third model parameter optimizer includes at least the SDG optimizer. Step 710: Set three counters initially set to 0 and denoted as TP counter, FP counter and FN counter; Step 711: Traverse all the third model data records in the current validation dataset; during traversal, take the currently traversed third model data record as the corresponding current validation record; take the first training radiomics feature set and the first label type vector of the current validation record as the corresponding current radiomics feature set and current label type vector; take the M stage type corresponding to the first label type probability with a probability value of 1 in the current label type vector as the corresponding current label type; input the current radiomics feature set into the current prediction model to perform M stage classification prediction to obtain the corresponding second prediction vector; take the M stage type corresponding to the second type prediction probability with the highest probability in the second prediction vector as the corresponding current prediction type; compare the current prediction type with the current label type; if the current prediction type and the current label type are both M1 type, then increment the TP counter by 1; if the current prediction type is M1 type and the current label type is M0 type, then increment the FP counter by 1; if the current prediction type is M0 type and the current label type is M1 type, then increment the FN counter by 1. Wherein, the second prediction vector includes two second type prediction probabilities, each second type prediction probability corresponding to one M-stage type; Step 712: After traversing all the third model data records in the current validation dataset, calculate the corresponding first precision and first recall based on the TP counter, the FP counter, and the FN counter; and calculate the corresponding first score F1 based on the first precision and the first recall. i ; Step 713: Identify whether the current test verification set is the Kth first test verification set; if yes, proceed to step 714; if no, take the next first test verification set as the new current test verification set and return to step 706. Step 714, based on the obtained K first scores F1 i The first error E is obtained by calculating the mean square error. Step 715: Identify whether the current prediction model is the last first prediction model; if yes, proceed to step 716; if no, take the next first prediction model as the new current prediction model and return to step 705. Step 716: Take the full model parameter set of the first prediction model corresponding to the first error E with the smallest error as the corresponding optimal full model parameter set. Step 717: Based on the optimal full model parameter set, solidify the model parameters of the first prediction model and confirm that the model training of the first prediction model has ended.

8. The processing method for M-stage classification prediction based on CT images according to claim 3, characterized in that, The training of the second prediction model based on the fourth model dataset specifically includes: Step 81: Take the first fourth model data record of the fourth model dataset as the corresponding current model data record; Step 82: Input the first training feature map of the current model data record into the second prediction model to perform M-period classification prediction to obtain the corresponding third prediction vector; The third prediction vector includes two third-type prediction probabilities, each of which corresponds to one of the M-stage types. Step 83: Input the third prediction vector and the second label type vector of the current model data record into the preset fourth model loss function to calculate the corresponding third loss value; The fourth model loss function is implemented based on the binary classification cross-entropy loss function; Step 84: Identify whether the third loss value meets the preset third loss value range; if the third loss value meets the third loss value range, identify whether the current model data record is the last fourth model data record in the fourth model dataset; if yes, proceed to step 85; otherwise, take the next fourth model data record in the fourth model dataset as the new current model data record and return to step 82 to continue training; if the third loss value does not meet the third loss value range, perform a round of model parameter optimization on the second prediction model based on the preset fourth model parameter optimizer in the direction of minimizing the fourth model loss function, and return to step 82 to continue training when the parameter optimization ends. The fourth model parameter optimizer includes at least the SDG optimizer; Step 85: Solidify the model parameters of the second prediction model and confirm that the training of the second prediction model is complete.

9. The processing method for M-stage classification prediction based on CT images according to claim 1, characterized in that, The step of predicting the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model specifically includes: Step 91: Input the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding first semantic map; The first CT image includes multiple first voxel points; the three-dimensional dimensions of the first semantic map are consistent with the three-dimensional dimensions of the first CT image; the first semantic map includes multiple second voxel points, and the first and second voxel points correspond one-to-one; each second voxel point corresponds to a binary gastric tumor lesion area semantic, with a value of 0 or 1; when the gastric tumor lesion area semantic is 0, it indicates that the corresponding second voxel point is a background voxel point, and when it is 1, it indicates that the corresponding second voxel point is a gastric tumor lesion area voxel point; Step 92: Count the number of second voxel points with semantic value 1 in the gastric tumor lesion area on the first semantic map to obtain the corresponding first number; and identify whether the first number exceeds the preset first number threshold. If it exceeds the threshold, proceed to step 93; if it does not exceed the threshold, set the corresponding first prediction type to type M0 and proceed to step 99. Step 93: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding second semantic map; The second semantic map has the same three-dimensional dimensions as the first CT image. The second semantic map includes multiple third voxel points, and the first and third voxel points correspond one-to-one. Each third voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding third voxel point is a background voxel point; when it is 1, it indicates that the corresponding third voxel point is a remote critical tissue lesion voxel point. Step 94, and count the number of third voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding second number; and identify whether the second number exceeds the preset second number threshold. If it exceeds, proceed to step 95; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 99. Step 95: The first voxel points on the first CT image that correspond to the second voxel points with semantic value 1 in each of the aforementioned distal key tissue lesion areas are recorded as the corresponding first lesion points; and the three-dimensional structure composed of all the first lesion points is recorded as the corresponding first structure; and the first-order features, shape features, and texture features of the radiomics features of the first structure are calculated to obtain the corresponding first-order feature set, shape feature set, and texture feature set, which together form a corresponding second gastric tumor feature set; Step 96: Based on a preset point clustering algorithm, cluster the first voxel points on the first CT image corresponding to each of the distal critical tissues to obtain corresponding second point groups; and use each of the second point groups as the corresponding current point groups, and identify whether the current point group is empty; if the current point group is empty, set a set of all-zero first-order feature set, shape feature set, and texture feature set for the distal critical tissue corresponding to the current point group to form a corresponding second distal critical tissue feature set; if the current point group is not empty, record the three-dimensional structure on the first CT image composed of the current point groups as the corresponding second structure, and calculate the first-order feature set, shape feature set, and texture feature set of the radiomics features of the second structure to form a corresponding second distal critical tissue feature set; and use all the obtained second distal critical tissue feature sets to form a corresponding second distal critical tissue feature set; Step 97: A corresponding first radiomics feature set is formed by combining the obtained second gastric tumor feature set and the second distal key tissue feature set; and the first radiomics feature set is input into the first prediction model to perform M-stage classification prediction to obtain the corresponding fourth prediction vector; The fourth prediction vector includes two fourth-type prediction probabilities, each of which corresponds to one of the M-stage types. Step 98: Take the M-stage type corresponding to the fourth type prediction probability with the highest probability in the fourth prediction vector as the corresponding first prediction type; Step 99: Output the obtained first prediction type as the prediction result for this time.

10. The processing method for M-stage classification prediction based on CT images according to claim 1, characterized in that, The step of predicting the corresponding first prediction type based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model specifically includes: Step 101: Input the first CT image into the first semantic segmentation model to perform semantic point segmentation processing of the gastric tumor lesion area to obtain the corresponding third semantic map; The first CT image includes multiple first voxel points; each first voxel point corresponds to a first voxel point pixel value; the three-dimensional size of the third semantic map is consistent with the three-dimensional size of the first CT image; the third semantic map includes multiple fourth voxel points, with a one-to-one correspondence between the first and fourth voxel points; each fourth voxel point corresponds to a binary gastric tumor lesion area semantic, with a value of 0 or 1; when the gastric tumor lesion area semantic is 0, it indicates that the corresponding fourth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fourth voxel point is a gastric tumor lesion area voxel point. Step 102: Count the number of fourth voxel points with semantic value 1 in the gastric tumor lesion area on the third semantic map to obtain the corresponding third number; and identify whether the third number exceeds the preset first number threshold. If it exceeds the threshold, proceed to step 103; if it does not exceed the threshold, set the corresponding first prediction type to type M0 and proceed to step 108. Step 103: Input the first CT image into the second semantic segmentation model to perform semantic point segmentation processing of the lesion area of ​​the distal key tissue to obtain the corresponding fourth semantic map; The fourth semantic map has the same three-dimensional dimensions as the first CT image. The fourth semantic map includes multiple fifth voxel points, with a one-to-one correspondence between the first and fifth voxel points. Each fifth voxel point corresponds to a binary remote critical tissue lesion area semantic, with a value of 0 or 1. When the remote critical tissue lesion area semantic is 0, it indicates that the corresponding fifth voxel point is a background voxel point; when it is 1, it indicates that the corresponding fifth voxel point is a remote critical tissue lesion voxel point. Step 104: Count the number of fifth voxel points with semantic value 1 in the remote key tissue lesion area on the second semantic map to obtain the corresponding fourth number; and identify whether the fourth number exceeds the preset second number threshold. If it exceeds, proceed to step 105; if it does not exceed, set the corresponding first prediction type to type M0 and proceed to step 108. Step 105: Each first voxel point in the first CT image is taken as the corresponding current voxel point; the fourth and fifth voxel points on the third and fourth semantic maps corresponding to the current voxel points are recorded as the corresponding third and fourth matching voxel points; the pixel value of the first voxel point of the current voxel point is taken as a corresponding feature voxel point pixel value; when the semantics of the gastric tumor lesion area and the semantics of the distal critical tissue lesion area corresponding to the third and fourth matching voxel points are both 0, a corresponding feature voxel point semantics is set to 0; and when the semantics of the gastric tumor lesion area corresponding to the third matching voxel point is... When the semantics of the feature voxel point are set to 1, a corresponding feature voxel point is set to 2; and when the semantics of the distal critical tissue lesion area corresponding to the fourth matching voxel point are 1, a corresponding feature voxel point is set to 2; and a corresponding second feature voxel point is constructed based on the three-dimensional voxel coordinates of the current voxel point, denoted as the corresponding current feature voxel point; and a feature voxel point feature vector corresponding to the current feature voxel point is formed by the pixel value of the feature voxel point corresponding to the current voxel point and the semantics of the feature voxel point; and a corresponding first feature map is formed by all the second feature voxel points corresponding to the first CT image. Step 106: Input the first feature map into the second prediction model to perform M-period classification prediction to obtain the corresponding fifth prediction vector; The fifth prediction vector includes two fifth-type prediction probabilities, each of which corresponds to one of the M-stage types. Step 107: Take the M-stage type corresponding to the fifth type prediction probability with the highest probability in the fifth prediction vector as the corresponding first prediction type; Step 108: Output the obtained first prediction type as the prediction result for this time.

11. An apparatus for performing the processing method for M-phase classification prediction based on CT images according to any one of claims 1-10, characterized in that, The device includes: a model building module, a model training module, and a model application module; The model building module is used to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of gastric tumor lesions on CT images, denoted as the corresponding first semantic segmentation model; and to construct a semantic segmentation model based on the U-Net model for semantic point segmentation of lesion areas of distal key tissues on the CT images, denoted as the corresponding second semantic segmentation model; and to construct a prediction model based on a machine learning classification model for M-stage classification prediction based on radiomics features, denoted as the corresponding first prediction model; and to construct a prediction model based on a deep learning classification model for M-stage classification prediction based on 3D CT images with lesion point markings, denoted as the corresponding second prediction model; wherein... The CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the distal key tissues include at least the liver, the para-aortic lymph node region, the retropancreatic lymph node region, the root mesenteric lymph node region, and the peritoneum; the types of machine learning classification models include at least classification prediction models based on RF models, classification prediction models based on SVM models, classification prediction models based on GBDT models, and classification prediction models based on GBDT+LR models; the types of deep learning classification models include at least classification prediction models based on ResNet models and classification prediction models based on Transformer models; The model training module is used to train the first semantic segmentation model, the second semantic segmentation model, the first prediction model, and the second prediction model respectively; The model application module is used to receive a first CT image and a first prediction mode after the training of four types of models; and to identify the first prediction mode; if the first prediction mode is a first mode, then a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the first prediction model; if the first prediction mode is a second mode, then a corresponding first prediction type is predicted based on the first CT image, the first semantic segmentation model, the second semantic segmentation model, and the second prediction model; and the obtained first prediction type is output as the prediction result corresponding to the first CT image; wherein, the first CT image is a three-dimensional abdominal CT image with three-dimensional information of the stomach and the distal key tissues; the first prediction mode includes a first mode and a second mode; the first prediction type includes M0 type and M1 type.

12. An electronic device, characterized in that, include: Memory, processor, and transceiver; The processor is configured to be coupled to the memory, read and execute instructions in the memory to implement the method according to any one of claims 1-10; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a computer, cause the computer to perform the method described in any one of claims 1-10.

Images