Training method and system of colorectal cancer radical resection degree prediction model

Abstract

The application provides a training method, a use method and a system of a colorectal cancer CRS radical degree prediction model, and comprises the following steps: obtaining medical record data of colorectal cancer cases; screening peritoneal metastasis related indexes of each colorectal cancer case to obtain a first training set; inputting the first training set into a first deep learning model for training; inputting CT images of each colorectal cancer case as a second training set into a second deep learning model for training; splicing first feature vectors and second feature vectors corresponding to each colorectal cancer case to obtain a third feature vector set; dividing the third feature vector set into a training set and a verification set; inputting the training set into a third deep learning model for training to obtain a converged third deep learning model; inputting the verification set into the converged third deep learning model for verification, and optimizing according to a verification result. The application can effectively score the colorectal cancer CRS radical degree through multi-modal feature prediction.

Description

Technical Field

[0001] This invention relates to the application of artificial intelligence in the field of medical technology, and more specifically, to the training, usage methods, and systems of a predictive model for the degree of radical cure of colorectal cancer CRS. Background Technology

[0002] Colorectal cancer, a common malignant tumor of the digestive tract, poses a serious threat to human health.

[0003] Clinically, cytoreductive surgery (CRS) is typically performed to remove visible tumors in patients with peritoneal metastases from colorectal cancer, prolonging patient survival and reducing mortality. Currently, preoperative imaging is used to assess the suitability of CRS surgery; however, the extent of CRS resection in colorectal cancer cannot be determined until the abdominal cavity is opened. This is because preoperative imaging cannot accurately determine the extent of peritoneal metastasis and therefore the degree of radical resection. The integration of AI with medicine, particularly in the diagnosis and treatment of malignant tumors, has been a key research focus in recent years. Artificial intelligence can learn a general pattern from large amounts of medical data, providing greater assistance to physicians' decision-making. Currently, there is no model using AI to predict the extent of CRS resection in colorectal cancer; therefore, a predictive model for the extent of CRS resection in colorectal cancer patients is urgently needed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a training method and system for a predictive model of the degree of radical resection for colorectal cancer resection (CRS). This invention can accurately predict the degree of radical resection for colorectal cancer CRS.

[0005] It should be noted that the scope of protection of the claims of this invention does not include methods for diagnosing diseases, and this invention does not limit the possible use of the predictive model of this invention for disease diagnosis by doctors.

[0006] In a first aspect, the present invention provides a training method for a colorectal cancer CRS radical cure degree prediction model, comprising:

[0007] Medical record data for colorectal cancer cases, including CT images and peritoneal metastasis-related indicators, were acquired. Peritoneal metastasis-related indicators for each colorectal cancer case were filtered to obtain a first training set. This first training set was then input into a first deep learning model for training, resulting in a converged first deep learning model and an output first feature vector set. The CT images of each colorectal cancer case were used as a second training set and input into a second deep learning model for training, resulting in a converged second deep learning model and an output second feature vector set. The first and second feature vectors corresponding to each colorectal cancer case were concatenated to obtain a third feature vector set. This third feature vector set was divided into a training set and a validation set. The training set was input into the third deep learning model for training, resulting in a converged third deep learning model. The validation set was input into the converged third deep learning model for validation, and optimization was performed based on the validation results.

[0008] This invention acquires medical record data of colorectal cancer cases, including CT images and peritoneal metastasis-related indicators, and uses deep learning technology for modeling and training. For CT images, the primary tumor lesion can be delineated during the venous phase of the CT image.

[0009] Peritoneal metastasis-related indicators were screened to select indicator data for use as input for training the deep learning model. Chi-square analysis can be used to calculate the correlation between each peritoneal metastasis-related indicator and the degree of CRS radical cure, in order to select indicator data for training and use it as the first training set.

[0010] The first and second deep learning models were constructed to extract features of peritoneal metastasis-related indicators and CT image features, respectively. The first and second feature vectors were then concatenated to obtain a third feature vector set, which was divided into a training set and a test set. The third deep learning model was then used for training, and through joint training, a deep learning model that can predict the degree of radical resection of colorectal cancer CRS was obtained.

[0011] Furthermore, the first deep learning model is a multilayer perceptron model, and each layer of the neural network includes neuron weight parameters V. e And a modified linear activation function f, let a be the input of the e-th layer neural network. e V e Initialize based on the first training set, and adjust according to the output during training; the calculation formula for each layer of the neural network is S. e =f(V e a e The input to the first layer is each subset of the first training set.

[0012] The multilayer perceptron model has fully connected layers. Through forward propagation, the model learns indicators of the degree of radical treatment of colorectal CRS, and through backpropagation, it optimizes the weight parameters V of each neuron. e Through continuous iterative learning, the multilayer perceptron model converges, which means that the data on peritoneal metastasis-related indicators are obtained, and the feature vectors of the degree of CRS cure for peritoneal metastasis-related indicators are extracted. Then, the degree of CRS cure for colorectal cancer can be preliminarily judged or joint training can be performed to predict the degree of CRS cure for colorectal cancer through multimodal fusion.

[0013] Furthermore, the second deep learning model is a ResNet3D model; the ResNet3D model includes convolutional layers, several residual blocks, and linear layers.

[0014] By constructing a ResNet3D model and extracting features from CT images, features for predicting the degree of CRS cure in colorectal cancer can be obtained. These features can be used to make a preliminary judgment on the degree of CRS cure in colorectal cancer or for joint training. By constructing a predictive model for the degree of CRS cure in colorectal cancer through multimodal fusion, a more accurate and scientific prediction can be made.

[0015] Furthermore, let the CT image of the m-th colorectal cancer case be x. m ∈R 512×512×384 The input tensor;

[0016] set up Input a tensor for the CT image of the m-th colorectal cancer case, where j, k, and l represent the pixels at three locations in the CT image, respectively;

[0017] The convolutional layer is a 3×3×3×64 three-dimensional convolutional neural network;

[0018] The residual block consists of two 3×3×3×64 three-dimensional convolutional neural networks and a ReLU layer. Let the input of the residual block be x and the output be y. The operation performed by the residual block is represented as follows:

[0019] h = σ(G(x)),

[0020] y = σ(G(x) + x);

[0021] σ() represents the operation of the ReLU layer. Let the input be p.

[0022] The three-dimensional convolutional neural network performs convolution operations, assuming the input is x. m The output is G(x) m )=b j，k，l =x m [j -1,j+1],[k-1,k+1],[l-1,l+1] ·Wc , indicating that for x m Each 3×3×3 small square in the W c After performing the inner product operation, it becomes an element in the output tensor b;

[0023] W c The weight parameters of the ResNet3D model are initialized based on the input tensor and adjusted during training based on the output results.

[0024] The linear layer receives the output of the last residual block as input and performs a linear mapping.

[0025] By constructing a ResNet3D model, feature extraction can be performed on CT images. First, a convolutional layer is used to extract shallow features. After delineation and enhancement, the shallow features of CT images from colorectal cancer cases represent the edge features of the primary tumor lesion. A single convolutional operation effectively extracts this important indicator. By constructing several residual blocks, deep features, including central and edge regions, are extracted, ensuring sufficient feature extraction from the CT images of colorectal cancer cases. A linear layer performs linear mapping on the features to obtain the labeled results, which represent the training and learning results of the ResNet3D model.

[0026] Furthermore, the third deep learning model is a multimodal fusion model, including:

[0027] Several neural network layers are used to extract features from the third training set;

[0028] The feature vector d is obtained by averaging the final outputs of the aforementioned neural network layers.

[0029] The eigenvector d is linearly mapped to obtain the three-dimensional eigenvector s.

[0030] Let s k The k-th element in the feature vector s

[0031] The output of the multimodal fusion model is obtained by performing operations on the feature vector s. The formula is expressed as:

[0032]

[0033] A multimodal fusion model is used to analyze multimodal feature vectors that integrate CT images and peritoneal metastasis-related indicators. It can analyze and predict using CT image information and peritoneal metastasis-related indicators, i.e., clinical case characteristics, exhibiting high sensitivity and specificity. After dividing the third feature vector set into training and validation sets, the training set is input into the constructed multimodal fusion model for training, resulting in a converged multimodal fusion model, achieving the effect of joint training. The validation set is input into the constructed multimodal fusion model for validation, and the validation results are optimized to obtain a final converged deep learning model for predicting the degree of radical resection of colorectal cancer CRS. This prediction result has diagnostic and identification value in predicting the degree of radical resection of colorectal cancer CRS, which is beneficial for clinicians to analyze the value and risks of cytoreductive surgery for peritoneal metastatic tumors in colorectal cancer, thereby facilitating individualized and precise treatment for colorectal cancer patients.

[0034] Furthermore, the step of averaging the final outputs of the plurality of neural network layers to obtain the feature vector d is specifically as follows:

[0035] Let the final output of the aforementioned neural network layers be u, and s = avg(u) × W c .

[0036] The final output u of several neural network layers is averaged and then compared with the weight parameters W. c By performing a cross product operation, a linear mapping is obtained, resulting in a feature vector d. This facilitates the transformation of features extracted from several neural network layers into a feature vector d that needs to be labeled, which is beneficial for feature extraction from CT images.

[0037] Furthermore, the loss function of the multimodal fusion model is:

[0038]

[0039] Where n represents the number of training samples in the training set.

[0040] By training with n training samples and continuously using the loss function for backpropagation to optimize the parameters, it is beneficial to obtain a convergent model.

[0041] This invention also provides a method for using a colorectal cancer CRS radical cure prediction model, characterized by comprising:

[0042] Obtain medical record data for colorectal cancer cases to be predicted, including CT images and peritoneal metastasis-related indicators;

[0043] The peritoneal metastasis-related indicators are input into a convergent multilayer perceptron model to obtain the feature vectors of the peritoneal metastasis-related indicators.

[0044] The CT image is input into a converged ResNet3D model to obtain the feature vector of the CT image;

[0045] The feature vectors of CT images are concatenated with the feature vectors of peritoneal metastasis-related indicators to obtain a multimodal feature vector;

[0046] The multimodal feature vectors are input into the multimodal fusion model to obtain the prediction results.

[0047] By utilizing convergent multilayer perceptron models, ResNet3D models, and multimodal fusion models, the degree of radical resection of colorectal cancer CRS can be accurately predicted, which is beneficial for assisting in the clinical implementation of colorectal cancer tumor cytoreductive surgery and for individualized and precise treatment of colorectal cancer cases.

[0048] This invention also provides a training system for a colorectal cancer CRS radical cure prediction model, comprising:

[0049] The training data acquisition module is used to acquire medical record data of colorectal cancer cases, including CT images and peritoneal metastasis-related indicators.

[0050] The screening module is used to screen peritoneal metastasis-related indicators for each colorectal cancer case to obtain the first training set;

[0051] The first training module is used to input the first training set into the first deep learning model for training, and to obtain a converged first deep learning model and the output first feature vector set.

[0052] The second training module is used to take the CT images of each colorectal cancer case as the second training set, input them into the second deep learning model for training, and obtain the converged second deep learning model and the output second feature vector set.

[0053] The splicing module is used to splice the first and second feature vectors corresponding to each colorectal cancer case to obtain a third feature vector set;

[0054] The partitioning module is used to divide the third feature vector set into a training set and a validation set.

[0055] The third training module is used to input the training set into the third deep learning model for training, and obtain a converged third deep learning model.

[0056] The optimization module is used to validate the converged third deep learning model input to the validation set and optimize it based on the validation results.

[0057] The training system for the colorectal cancer CRS radical cure prediction model of the present invention is based on the same inventive concept as the training method for the colorectal cancer CRS radical cure prediction model, and will not be described again here.

[0058] This invention also provides a system for using a colorectal cancer CRS radical cure prediction model, comprising:

[0059] The medical record data acquisition module is used to acquire medical record data of colorectal cancer cases to be predicted, including CT images and peritoneal metastasis-related indicators.

[0060] The peritoneal metastasis-related indicator feature extraction module is used to input peritoneal metastasis-related indicators into a convergent multilayer perceptron model to obtain feature vectors of peritoneal metastasis-related indicators.

[0061] The CT image feature extraction module takes a converged ResNet3D model as input and obtains the feature vector of the CT image.

[0062] The vector stitching module is used to stitch together the feature vectors of CT images with the vectors of peritoneal metastasis-related indicators to obtain multimodal data.

[0063] The prediction module is used to input multimodal feature vectors into the multimodal fusion model to obtain prediction results.

[0064] The system for using the colorectal cancer CRS radical cure prediction model of the present invention is based on the same inventive concept as the method of using the colorectal cancer CRS radical cure prediction model, and will not be described again here.

[0065] The beneficial effects of this invention are as follows:

[0066] (1) Construct a prediction model for the degree of radical treatment of colorectal cancer CRS by multimodal fusion to make more accurate and scientific predictions.

[0067] (2) Through joint training, a deep learning model that can predict the degree of radical treatment of colorectal cancer CRS was obtained.

[0068] (3) It has a diagnostic and identification role in the degree of CRS radical treatment of colorectal cancer, which is conducive to clinical staff to analyze the value and benefits of CRS surgery for colorectal cancer, and thus facilitates individualized and precise treatment of colorectal cancer patients. Attached Figure Description

[0069] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0070] Figure 2 This is a schematic diagram of the residual block of the present invention.

[0071] Figure 3 This is a schematic diagram of the structure of the three-dimensional convolutional neural network implemented in this invention.

[0072] Figure 4This is a schematic diagram of the ResNet3D model implemented in this invention.

[0073] Figure 5 This is a schematic diagram of the structure of the multilayer perceptron model implemented in this invention.

[0074] Figure 6 This is a schematic diagram of the structure of the multimodal fusion model implemented in this invention.

[0075] Figure 7 The AUC diagram is used to verify the multimodal fusion model implemented in this invention.

[0076] Figure 8 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0077] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. To better illustrate the following embodiments, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0078] Example 1

[0079] This embodiment provides a training method for a colorectal cancer CRS radical cure degree prediction model, such as... Figure 1 As shown, it includes:

[0080] S1. Obtain medical record data for colorectal cancer cases, including CT images and peritoneal metastasis-related indicators;

[0081] S2. Screen the peritoneal metastasis-related indicators for each colorectal cancer case to obtain the first training set;

[0082] S3. Input the first training set into the first deep learning model for training, and obtain the converged first deep learning model and the output first feature vector set;

[0083] S4. Use the CT images of each colorectal cancer case as the second training set, input them into the second deep learning model for training, and obtain the converged second deep learning model and the output second feature vector set.

[0084] S5. Concatenate the first and second feature vectors corresponding to each colorectal cancer case to obtain the third feature vector set;

[0085] S6. Divide the third feature vector set into a training set and a validation set;

[0086] S7. Input the training set into the third deep learning model for training to obtain a converged third deep learning model.

[0087] S8. Input the validation set into the converged third deep learning model for validation, and optimize based on the validation results.

[0088] This invention acquires medical record data of colorectal cancer cases, including CT images and peritoneal metastasis-related indicators, and uses deep learning technology for modeling and training. For CT images, the primary tumor lesion can be delineated first, which facilitates subsequent feature extraction.

[0089] Peritoneal metastasis-related indicators were screened to select indicator data for use as input for training the deep learning model. Chi-square analysis can be used to calculate the correlation between each peritoneal metastasis-related indicator and the degree of CRS radical cure, in order to select indicator data for training and use it as the first training set.

[0090] The first and second deep learning models were constructed to extract features of peritoneal metastasis-related indicators and CT image features, respectively. The first and second feature vectors were then concatenated to obtain a third feature vector set, which was divided into a training set and a test set. The third deep learning model was then used for training, and through joint training, a deep learning model that can predict the degree of radical resection of colorectal cancer CRS was obtained.

[0091] Preferably, such as Figure 5 As shown, the first deep learning model is a multilayer perceptron model, and each layer of the neural network includes neuron weight parameters V. e And a modified linear activation function f, let a be the input of the e-th layer neural network. e V e Initialize based on the first training set, and adjust according to the output during training; the calculation formula for each layer of the neural network is S. e =f(V e a e The input to the first layer is each subset of the first training set.

[0092] The multilayer perceptron model has fully connected layers. Through forward propagation, the model learns indicators of the degree of radical treatment of colorectal CRS, and through backpropagation, it optimizes the weight parameters V of each neuron. e Through continuous iterative learning, the multilayer perceptron model converges, which means that the data on peritoneal metastasis-related indicators are obtained. The feature vectors of the degree of CRS cure for peritoneal metastasis-related indicators are extracted, which can preliminarily determine the degree of CRS cure for colorectal cancer. Alternatively, joint training can be performed to predict the degree of CRS cure for colorectal cancer through multimodal fusion.

[0093] Preferably, such as Figure 4As shown, the second deep learning model is a ResNet3D model, which includes convolutional layers, several residual blocks, and linear layers.

[0094] By constructing a ResNet3D model, features are extracted from CT images to obtain features for predicting the degree of CRS cure in colorectal cancer. These features can be used to judge the degree of CRS cure in colorectal cancer or for joint training. Through multimodal fusion, a predictive model for the degree of CRS cure in colorectal cancer is constructed for more accurate and scientific prediction.

[0095] Preferably, let the CT image of the m-th colorectal cancer case be x. m ∈R 512×512×384 The input tensor;

[0096] set up Input a tensor for the CT image of the m-th colorectal cancer case, where j, k, and l represent the pixels at three locations in the CT image, respectively;

[0097] The convolutional layer is a 3×3×3×64 three-dimensional convolutional neural network;

[0098] like Figure 2 As shown, the residual block consists of two 3×3×3×64 three-dimensional convolutional neural networks and a ReLU layer. Let the input of the residual block be x and the output be y. The operation performed by the residual block is represented as follows:

[0099] h = σ(G(x)),

[0100] y = σ(G(x) + x);

[0101] σ() represents the operation of the ReLU layer. Let the input be p.

[0102] like Figure 3 As shown, the three-dimensional convolutional neural network performs convolution operations, with the input being x. m The output is G(x) m )=b j ，k，l =x m [j-1,j+1],[k-1,k+1],[l-1,l+1] ·W c , indicating that for x m Each 3×3×3 small square in the W c After performing the inner product operation, it becomes an element in the output tensor b;

[0103] W c The weight parameters of the ResNet3D model are initialized based on the input tensor and adjusted during training based on the output results.

[0104] The linear layer receives the output of the last residual block as input and performs a linear mapping.

[0105] like Figure 4 As shown, a ResNet3D model is constructed to extract features from CT images. First, a convolutional layer is used to extract shallow features. After delineation and enhancement, the shallow features of the CT images from colorectal cancer cases represent the edge features of the primary tumor lesion. A single convolutional operation effectively extracts this important indicator. By constructing several residual blocks, deep features, including central and edge regions, are extracted, ensuring sufficient feature extraction from the CT images of colorectal cancer cases. Linear layers linearly map these features to obtain the labeled results, which represent the results of ResNet3D training and learning.

[0106] Preferably, the third deep learning model is a multimodal fusion model, such as... Figure 6 As shown, it includes:

[0107] Several neural network layers are used to extract features from the third training set;

[0108] The feature vector d is obtained by averaging the final outputs of the aforementioned neural network layers.

[0109] The eigenvector d is linearly mapped to obtain the three-dimensional eigenvector s.

[0110] Let s k The k-th element in the feature vector s

[0111] The output of the multimodal fusion model is obtained by performing operations on the feature vector s. The formula is expressed as:

[0112]

[0113] A multimodal fusion model is used to analyze multimodal feature vectors that integrate CT images and peritoneal metastasis-related indicators. It can analyze and predict using CT image information and peritoneal metastasis-related indicators, i.e., clinical case characteristics, exhibiting high sensitivity and specificity. After dividing the third feature vector set into training and validation sets, the training set is input into the constructed multimodal fusion model for training, resulting in a converged multimodal fusion model, achieving the effect of joint training. The validation set is input into the constructed multimodal fusion model for validation, and the validation results are optimized to obtain a final converged deep learning model for predicting the degree of radical resection of colorectal cancer CRS. This prediction result has diagnostic and identification value for the degree of radical resection of colorectal cancer CRS, which is beneficial for clinicians to analyze the value of CRS surgery and for providing individualized and precise treatment for colorectal cancer patients.

[0114] Preferably, the step of averaging the final outputs of the plurality of neural network layers to obtain the feature vector d specifically involves:

[0115] Let the final output of the aforementioned neural network layers be u, and s = avg(u) × W c .

[0116] The final output u of several neural network layers is averaged and then compared with the weight parameters W. c By performing a cross product operation, a linear mapping is obtained, resulting in a feature vector d. This facilitates the transformation of features extracted from several neural network layers into a feature vector d that needs to be labeled, which is beneficial for feature extraction from CT images.

[0117] Preferably, the loss function of the multimodal fusion model is:

[0118]

[0119] Where n represents the number of training samples in the training set.

[0120] By training with n training samples and continuously using the loss function for backpropagation to optimize the parameters, it is beneficial to obtain a convergent model.

[0121] like Figure 7 As shown, the multimodal fusion model in this embodiment is validated after training and convergence to obtain the AUC diagram.

[0122] Example 2

[0123] This embodiment provides a method for using a colorectal cancer CRS radical cure prediction model, such as... Figure 8 As shown, it includes:

[0124] T1. Obtain the medical record data of the colorectal cancer case to be predicted, including CT images and peritoneal metastasis-related indicators;

[0125] T2. Input the peritoneal metastasis-related indicators into the convergent multilayer perceptron model to obtain the feature vector of the peritoneal metastasis-related indicators;

[0126] T3. Input the CT image into the converged ResNet3D model to obtain the feature vector of the CT image;

[0127] T4. Concatenate the feature vectors of the CT images with the feature vectors of peritoneal metastasis-related indicators to obtain a multimodal feature vector;

[0128] T5. Input the multimodal feature vectors into the multimodal fusion model to obtain the prediction results.

[0129] By utilizing convergent multilayer perceptron models, ResNet3D models, and multimodal fusion models, the degree of radical resection of colorectal cancer CRS can be accurately predicted, which is beneficial for clinical support in the treatment of colorectal cancer CRS and for individualized and precise treatment of colorectal cancer cases.

[0130] Example 3

[0131] This embodiment provides a training system for a colorectal cancer CRS radical treatment prediction model, including:

[0132] The training data acquisition module is used to acquire medical record data of colorectal cancer cases, including CT images and peritoneal metastasis-related indicators.

[0133] The screening module is used to screen peritoneal metastasis-related indicators for each colorectal cancer case to obtain the first training set;

[0134] The first training module is used to input the first training set into the first deep learning model for training, and to obtain a converged first deep learning model and the output first feature vector set.

[0135] The second training module is used to take the CT images of each colorectal cancer case as the second training set, input them into the second deep learning model for training, and obtain the converged second deep learning model and the output second feature vector set.

[0136] The splicing module is used to splice the first and second feature vectors corresponding to each colorectal cancer case to obtain a third feature vector set;

[0137] The partitioning module is used to divide the third feature vector set into a training set and a validation set.

[0138] The third training module is used to input the training set into the third deep learning model for training, and obtain a converged third deep learning model.

[0139] The optimization module is used to validate the converged third deep learning model input to the validation set and optimize it based on the validation results.

[0140] The training system for the colorectal cancer CRS radical cure prediction model in this embodiment is based on the same inventive concept as the training method for the colorectal cancer CRS radical cure prediction model, and will not be described again here.

[0141] Example 4

[0142] This embodiment provides a system for using a colorectal cancer CRS radical cure prediction model, including:

[0143] The medical record data acquisition module is used to acquire medical record data of colorectal cancer cases to be predicted. The medical record data includes CT images and peritoneal metastasis-related indicators.

[0144] The peritoneal metastasis-related indicator feature extraction module is used to input peritoneal metastasis-related indicators into a convergent multilayer perceptron model to obtain feature vectors of peritoneal metastasis-related indicators.

[0145] The CT image feature extraction module is used to input CT images into a converged ResNet3D model to obtain the feature vectors of the CT images.

[0146] The vector splicing module is used to splice the feature vectors of CT images with the feature vectors of peritoneal metastasis-related indicators to obtain multimodal feature vectors;

[0147] The prediction module is used to input multimodal feature vectors into the multimodal fusion model to obtain prediction results.

[0148] The system for using the colorectal cancer CRS radical cure prediction model in this embodiment is based on the same inventive concept as the method for using the colorectal cancer CRS radical cure prediction model, and will not be described again here.

[0149] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A training method for a colorectal cancer CRS radical cure prediction model, characterized in that, include: Obtain medical record data for colorectal cancer cases, including CT images and peritoneal metastasis-related indicators; Peritoneal metastasis-related indicators for each colorectal cancer case were screened to obtain the first training set; The first training set is input into the first deep learning model for training, resulting in a converged first deep learning model and the first set of output feature vectors. The CT images of each colorectal cancer case are used as the second training set and input into the second deep learning model for training, resulting in a converged second deep learning model and the output second feature vector set. The first and second feature vectors corresponding to each colorectal cancer case are concatenated to obtain the third feature vector set; The third feature vector set is divided into a training set and a validation set; The training set is input into the third deep learning model for training, resulting in a converged third deep learning model. The validation set is input into the converged third deep learning model for validation, and optimization is performed based on the validation results; The first deep learning model is a multilayer perceptron model, where each layer of the neural network includes neuron weight parameters. V e and modified linear activation function f Let the first e The input of the layer neural network is a e , V e Initialize based on the first training set, and adjust according to the output during training; The calculation formula for each layer of the neural network is as follows: S e = f(V) e a e ） The input to the first layer is each subset of the first training set; The second deep learning model is the ResNet3D model; The ResNet3D model includes convolutional layers, several residual blocks, and linear layers; Let the first m CT images of a colorectal cancer case are The input tensor; set up For the first m Input tensors for CT images of a colorectal cancer case. j, k, l These represent pixels at three locations in the CT image; The convolutional layer is a 3×3×3×64 three-dimensional convolutional neural network; The residual block consists of two 3×3×3×64 three-dimensional convolutional neural networks and a ReLU layer. Let the input of the residual block be... x The output is y The operations performed by the residual block are represented as follows: h = σ(G(x)) ， y = σ(G(x) + x) ； σ() This represents the operation of the ReLU layer, assuming the input is... p , ; The three-dimensional convolutional neural network performs a convolution operation, assuming the input is... x m The output is G ( x m )= b j，k，l =x m [j -1,j+1],[k-1,k+1],[l-1,l+1] ·W c , indicating to x m Each 3×3×3 small square in the and W c After performing the inner product operation, it is used as the output tensor. b One of the elements; W c The weight parameters of the ResNet3D model are initialized based on the input tensor and adjusted during training based on the output results. The linear layer receives the output of the last residual block as input and performs linear mapping; The third deep learning model is a multimodal fusion model, including: Several neural network layers are used to extract features from the third training set; The feature vector is obtained by averaging the final outputs of the aforementioned neural network layers. d ; The eigenvector d is obtained by linear mapping to obtain the three-dimensional eigenvector. s ; Let s k The eigenvector s is the th k One element, The output of the multimodal fusion model is obtained by performing operations on the feature vector s. The formula is expressed as: 。 2. The training method for a colorectal cancer CRS radical cure prediction model according to claim 1, characterized in that: The step of averaging the final outputs of the plurality of neural network layers to obtain the feature vector d is specifically as follows: Let the final output of the aforementioned neural network layers be u , s = avg(u) × W c 。 3. The training method for a colorectal cancer CRS radical cure prediction model according to claim 2, characterized in that: The loss function of the multimodal fusion model is: Where n represents the number of training samples in the training set.

4. A method for using a predictive model for the degree of radical resection of colorectal cancer (CRS), characterized in that, The colorectal cancer CRS radical cure prediction model is trained using the training method for a colorectal cancer CRS radical cure prediction model as described in any one of claims 1-3. include: Obtain medical record data for colorectal cancer cases to be predicted, including CT images and peritoneal metastasis-related indicators; The peritoneal metastasis-related indicators are input into a convergent multilayer perceptron model to obtain the feature vectors of the peritoneal metastasis-related indicators. The CT image is input into a converged ResNet3D model to obtain the feature vector of the CT image; The feature vectors of CT images are concatenated with the feature vectors of peritoneal metastasis-related indicators to obtain a multimodal feature vector; The multimodal feature vectors are input into the multimodal fusion model to obtain the prediction results.

5. A training system for a colorectal cancer CRS radical cure prediction model, characterized in that, The colorectal cancer CRS radical cure prediction model is trained using the training method for a colorectal cancer CRS radical cure prediction model as described in any one of claims 1-3. include: The training data acquisition module is used to acquire medical record data of colorectal cancer cases, including CT images and peritoneal metastasis-related indicators. The screening module is used to screen peritoneal metastasis-related indicators for each colorectal cancer case to obtain the first training set; The first training module is used to input the first training set into the first deep learning model for training, and to obtain a converged first deep learning model and the output first feature vector set. The second training module is used to input the CT images of each colorectal cancer case as the second training set into the second deep learning model for training, so as to obtain a converged second deep learning model and the output second feature vector set. The splicing module is used to splice the first and second feature vectors corresponding to each colorectal cancer case to obtain a third feature vector set; The partitioning module is used to divide the third feature vector set into a training set and a validation set. The third training module is used to input the training set into the third deep learning model for training, and obtain a converged third deep learning model. The optimization module is used to validate the converged third deep learning model input to the validation set and optimize it based on the validation results.

6. A system for predicting the degree of radical resection of colorectal cancer (CRS), characterized in that, The colorectal cancer CRS radical cure prediction model is trained using the training method for a colorectal cancer CRS radical cure prediction model as described in any one of claims 1-3. include: The medical record data acquisition module is used to acquire medical record data of colorectal cancer cases to be predicted. The medical record data includes CT images and peritoneal metastasis-related indicators. The peritoneal metastasis-related indicator feature extraction module is used to input peritoneal metastasis-related indicators into a convergent multilayer perceptron model to obtain feature vectors of peritoneal metastasis-related indicators. The CT image feature extraction module is used to input CT images into a converged ResNet3D model to obtain the feature vectors of the CT images. The vector splicing module is used to splice the feature vectors of CT images with the feature vectors of peritoneal metastasis-related indicators to obtain multimodal feature vectors; The prediction module is used to input multimodal feature vectors into the multimodal fusion model to obtain prediction results.

Images