Method and system for constructing pCR prediction model for esophageal cancer
By constructing a neural network based on CT images and a pCR prediction model of XgBoost model, the problem of inaccurate prediction of pCR after neoadjuvant chemotherapy in patients with esophageal cancer in the prior art is solved, and efficient and accurate pCR prediction is achieved, improving work efficiency.
Patent Information
- Application Number
- CN202510489392.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-18
- Publication Date
- 2025-08-01
AI Technical Summary
The prior art cannot accurately predict whether patients with esophageal cancer achieve complete pathological remission (pCR) after neoadjuvant chemoradiation and therapy, resulting in the success rate of observation waiting strategies that depend on manual judgment and lack of efficient prediction models.
A pCR prediction model based on CT images is constructed, and a neural network and XgBoost model are combined, and CNN features and imaging omics features are used to realize multi-scale feature extraction and feature screening of CT images to generate an accurate pCR prediction model.
Accurate prediction of pCR in patients with esophageal cancer is achieved, work efficiency is improved, dependence on manual judgment is reduced, and prediction accuracy and efficiency is improved.
Smart Images
Figure CN120412977A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and particularly relates to a method and system for constructing a pCR prediction model for esophageal cancer. Background Art
[0002] With the development of the times, esophageal cancer has become one of the common malignant tumors. Among them, in the actual treatment process, the standardized treatment option for existing locally advanced esophageal squamous cell carcinoma (ESCC) is nCRT (neoadjuvant chemoradiotherapy) plus esophagectomy, which has been widely used.
[0003] Among them, 43.2%-49% of ESCC patients have pathological complete response (pCR) in the surgical resection specimens after nCRT. For patients who are expected to achieve pCR after nCRT, an observation and waiting strategy can be selected, which can achieve the purpose of organ preservation and reduce surgical complications. Comparing patients who achieved pCR after nCRT with those who underwent surgical resection after nCRT using active surveillance, the results showed that there was no significant difference in progression free survival (PFS) and overall survival (OS) between the two groups.
[0004] Furthermore, existing technologies have shown that for patients who achieve pCR, even if surgery is performed only when tumor recurrence is detected, it will not have a negative impact on the overall survival rate of the patients. However, the success of the observation and waiting strategy depends on accurately predicting pCR after nCRT to select suitable patients to benefit from organ preservation. But currently, pCR can only be determined by histopathological examination of surgical resection specimens. Therefore, in view of the deficiencies of the existing technologies, it is necessary to provide a method that can construct a pCR prediction model to correspondingly improve work efficiency. Summary of the Invention
[0005] Based on this, the objective of the present invention is to provide a method and system for constructing a pCR prediction model for esophageal cancer to provide a method that can create a pCR prediction model in real time.
[0006] In the first aspect of the embodiments of the present invention, it is proposed that:
[0007] A method for constructing a pCR prediction model for esophageal cancer, wherein the method includes:
[0008] Obtain a number of original sample data, and preprocess the number of original sample data to correspondingly generate a number of target sample data, and split the number of target sample data into a training validation set and a test set correspondingly, where each of the original sample data contains a CT image and its corresponding ROI;
[0009] Based on the first preset rule, train a corresponding image processing model in real time according to the training validation set and a preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs corresponding CNN features in real time;
[0010] Extract corresponding radiomics features in the CT images of the target sample data in real time through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features to generate corresponding training features in real time;
[0011] Based on the second preset rule, perform predictive training on a preset XgBoost model through the training features to correspondingly train a pCR prediction model.
[0012] In the present invention, the first preset rule refers to: during the training process, use the cross-entropy loss function to supervise the learning of the model, and save the model with the highest training accuracy.
[0013] The second preset rule refers to: the technical process of jointly determining the optimal feature subset and model parameters through feature screening and Bayesian optimization, and initializing the XgBoost model for training.
[0014] The beneficial effect of the present invention is: by collecting original sample data in real time and performing corresponding preprocessing to obtain target sample data for training, based on this, an image processing model for subsequent processing can be trained in real time according to the first preset rule immediately, and CNN features for subsequent training can be obtained. Correspondingly, in order to improve the model training efficiency, the required training features can be obtained again based on the CNN features, and the required pCR prediction model can be finally obtained according to the second preset rule and the current training features, so that the process of manual judgment can be omitted, the prediction efficiency of pCR can be greatly improved, and the work efficiency is correspondingly improved.
[0015] The present invention focuses on the radiomics feature model at different time points, and fuses CNN features and traditional radiomics features to make the obtained features more comprehensive, and establishes a pCR prediction model based on three-stage CT images before treatment, after treatment, and before surgery. At the same time, a multi-scale attention module is added to better extract the CNN features of tumors at different scales in the CT images. In addition, steps such as feature screening and parameter search are used to prevent overfitting in training, and further improve the accuracy of model prediction.
[0016] Furthermore, the step of training a corresponding image processing model in real time according to the training verification set and a preset neural network based on the first preset rule includes:
[0017] When the training verification set is obtained in real time, a convolution operation is performed on the training verification set through the convolution kernel in the preset neural network to extract the corresponding initial features contained in the training verification set in real time;
[0018] According to the preset batch normalization layer and the preset activation function, the corresponding initial feature map is generated in real time according to the initial features, and the image processing model is generated according to the initial feature map and the preset neural network.
[0019] Furthermore, the step of generating the image processing model according to the initial feature map and the preset neural network includes:
[0020] When the initial feature map is acquired in real time, residual processing is performed on the initial feature map through a preset residual block to output a corresponding first intermediate feature map in real time;
[0021] Perform channel segmentation, global average pooling, and Softmax processing on the first intermediate feature map in sequence through a preset multi-scale attention module to output the corresponding second intermediate feature map in real time;
[0022] Furthermore, the step of generating a second intermediate feature map according to the multi-scale attention module includes:
[0023] Input feature map X∈R C×H×W Different branches are obtained by channel segmentation, and each branch uses a convolution kernel k of different sizes i ×k i and the corresponding number of groups G i , generate a multi-scale feature map F i :
[0024] F i =Conv(k i ×k i ,G i )(X),i=0,1,…,S-1
[0025] The convolution kernel size k i =2×(i+1)+1, number of groups Used to control computational complexity and output feature map F i ∈R C′×H×W , C ′ =C / S, where S is the number of branches.
[0026] For each scale feature map Fi , generate the channel attention weight Z through global average pooling and a fully connected layer i :
[0027] Z i =σ(W1δ(W0(g c )))
[0028] where is global average pooling, W0 ∈ R C′×C′ / r and W1 ∈ R C′ / r×C′ are fully connected layers, r is the compression ratio. δ is the ReLU activation function, and σ is the Sigmoid function.
[0029] Normalize the attention weights Z of different scales through the Softmax function to achieve cross-scale interaction: i
[0030]
[0031] Concatenate the calibration weights of all scales:
[0032]
[0033] Multiply the weight att i channel by channel with the feature map F of the corresponding scale, and concatenate to obtain the final output second feature map: i
[0034] Y i =F i ⊙att i , Out = Cat([Y0, Y1, …, Y S-1 )
[0035] Generate the image processing model according to the second intermediate feature map and the preset neural network, and the first intermediate feature map has the same size as the second intermediate feature map.
[0036] Furthermore, the step of generating the image processing model according to the second intermediate feature map and the preset neural network includes:
[0037] When the second intermediate feature map is obtained in real time, iteratively process the second intermediate feature map through the preset residual block and the preset multi-scale attention module to output the corresponding target feature map in real time;
[0038] Output the prediction value corresponding to the target feature map in real time through a preset fully connected layer, and train the preset neural network through the prediction value and a preset cross-entropy loss to output the image processing model in real time.
[0039] Further, the step of predicting and training the preset XgBoost model with the training features based on the second preset rule to correspondingly train the pCR prediction model includes:
[0040] When the training features are obtained in real time, a first feature subset is screened out from the training features in real time through a preset RFECV algorithm. The RFECV algorithm is a feature selection algorithm. Its principle is to first use the recursive feature elimination algorithm (RFE) to rank the features, and then perform cross-validation (CV) on the ranked features to select the best feature subset. RFECV obtains the importance degree of each feature through the coef_ attribute or feature_importances_ attribute returned by the learner. Then, the least important feature is removed from the current feature set, and this recursive step is continuously repeated on the feature set until the required number of features is finally reached. The first feature subset contains several sub-features;
[0041] The contribution degree generated by each sub-feature for the model prediction of the preset XgBoost model is calculated in real time through a preset SHAP algorithm, and each sub-feature is sorted in real time according to the size of the contribution degree to generate a corresponding feature subset sequence list in real time;
[0042] The pCR prediction model is correspondingly trained according to the feature subset sequence list and the preset XgBoost model.
[0043] Further, the step of correspondingly training the pCR prediction model according to the feature subset sequence list and the preset XgBoost model includes:
[0044] When the feature subset sequence list is obtained in real time, the first n sub-features are screened out from the feature subset sequence list in real time in the order of ranking as different second feature subsets;
[0045] The pCR prediction model is correspondingly trained according to the second feature subset and the preset XgBoost model.
[0046] Further, the step of correspondingly training the pCR prediction model according to the second feature subset and the preset XgBoost model includes:
[0047] When the different second feature subsets are obtained in real time, the preset XgBoost model is initialized and trained according to the second feature subset through a preset Bayesian optimization search algorithm to output several prediction accuracies in real time;
[0048] Detect the highest prediction accuracy in real time, and detect in real time the second feature subset composed of the first n sub-features corresponding to the highest prediction accuracy and the corresponding optimal features obtained through Bayesian parameter search;
[0049] Iteratively train the preset XgBoost model through the optimal features and the optimal parameters to generate the pCR prediction model in real time.
[0050] The second aspect of the embodiments of the present invention proposes:
[0051] A construction system for a pCR prediction model for esophageal cancer, wherein the system includes:
[0052] A splitting module, configured to obtain a plurality of original sample data, preprocess the plurality of original sample data to correspondingly generate a plurality of target sample data, and correspondingly split the plurality of target sample data into a training and validation set and a test set, wherein each of the original sample data contains a CT image and its corresponding ROI;
[0053] A first training module, configured to train a corresponding image processing model in real time based on a first preset rule according to the training and validation set and a preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs corresponding CNN features in real time;
[0054] A processing module, configured to extract corresponding radiomics features in the CT images of the target sample data in real time through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features to generate corresponding training features in real time;
[0055] A second training module, configured to perform prediction training on a preset XgBoost model based on a second preset rule through the training features to correspondingly train a pCR prediction model.
[0056] Further, the first training module is specifically configured to:
[0057] When the training and validation set is obtained in real time, perform a convolution operation on the training and validation set through the convolution kernels in the preset neural network to extract the initial features correspondingly included in the training and validation set in real time;
[0058] Generate corresponding initial feature maps according to the initial features through a preset batch normalization layer and a preset activation function, and generate the image processing model according to the initial feature maps and the preset neural network.
[0059] Further, the first training module is specifically configured to:
[0060] When the initial feature map is obtained in real time, perform residual processing on the initial feature map through a preset residual block to output the corresponding first intermediate feature map in real time;
[0061] Perform channel splitting, global average pooling, and Softmax processing on the first intermediate feature map in sequence through a preset multi-scale attention module to output the corresponding second intermediate feature map in real time;
[0062] Further, the steps of generating the second intermediate feature map according to the multi-scale attention module include:
[0063] Input feature map \(X\in\mathbb{R}\) C×G×W Obtain different branches through channel splitting, and each branch uses a convolution kernel \(k\) i \(\times k\) i and the corresponding number of groups \(G\) i , to generate a multi-scale feature map \(F\) i :
[0064] \(F\) i =\(Conv(k\) i \(\times k\) i ,G\) i )(X), i = 0, 1, …, S - 1
[0065] where the convolution kernel size \(k\) i = 2×(i + 1)+1, the number of groups is used to control the computational complexity, and the output feature map \(F\) i \(\in\mathbb{R}\) C′×H×W , \(C\) ′ = \(C / S\), and \(S\) is the number of branches.
[0066] For each scale of the feature map \(F\) i , generate the channel attention weight \(Z\) through global average pooling and a fully connected layer i :
[0067] \(Z\) i =\(\sigma(W_1\delta(W_0(g\) c )))\)
[0068] where is global average pooling, \(W_0\in\mathbb{R}\) C′×C′ / r and \(W_1\in\mathbb{R}\) C′ / r×C′ are fully connected layers, \(r\) is the compression ratio. \(\delta\) is the ReLU activation function, and \(\sigma\) is the Sigmoid function.
[0069] Normalize the attention weights \(Z\) of different scales through the Softmax function to achieve cross-scale interaction: i
[0070]
[0071] Concatenate the calibration weights at all scales:
[0072]
[0073] Multiply the weight att i with the feature map F of the corresponding scale i channel by channel, and concatenate to obtain the final output second feature map:
[0074] Y i = F i ⊙ att i , Out = Cat([Y0, Y1, …, Y S-1 )
[0075] Generate the image processing model according to the second intermediate feature map and the preset neural network, and the first intermediate feature map has the same size as the second intermediate feature map.
[0076] Furthermore, the first training module is specifically used for:
[0077] When the second intermediate feature map is obtained in real time, perform iterative processing on the second intermediate feature map through the preset residual block and the preset multi-scale attention module to output the corresponding target feature map in real time;
[0078] Output the prediction value corresponding to the target feature map in real time through a preset fully connected layer, and perform model training on the preset neural network through the prediction value and a preset cross-entropy loss to output the image processing model in real time.
[0079] Furthermore, the second training module is specifically used for:
[0080] When the training features are obtained in real time, use a preset RFECV algorithm to screen out the first feature subset in the training features in real time. The RFECV algorithm is a feature selection algorithm. Its principle is to first use the recursive feature elimination algorithm (RFE) to sort the features, and then perform cross-validation (CV) on the sorted features to select the best feature subset. RFECV obtains the importance of each feature through the coef_ attribute or feature_importances_ attribute returned by the learner. Then, remove the least important feature from the current feature set, and continuously repeat this recursive step on the feature set until the required number of features is finally reached. The first feature subset contains several sub-features;
[0081] The contribution degree generated by each of the sub - features when the preset XgBoost model makes a model prediction is calculated in real - time through the preset SHAP algorithm, and each of the sub - features is sorted in real - time according to the magnitude of the contribution degree to generate a corresponding feature subset sequence list in real - time;
[0082] The pCR prediction model is trained according to the feature subset sequence list and the preset XgBoost model.
[0083] Further, the second training module is specifically configured to:
[0084] When the feature subset sequence list is obtained in real - time, the first n sub - features are screened out in the feature subset sequence list in the order of ranking as different second feature subsets;
[0085] The pCR prediction model is trained according to the second feature subset and the preset XgBoost model.
[0086] Further, the second training module is specifically configured to:
[0087] When the different second feature subsets are obtained in real - time, the preset XgBoost model is initialized and trained according to the second feature subsets through the preset Bayesian optimization search algorithm to output a number of prediction accuracies in real - time;
[0088] The highest prediction accuracy is detected in real - time, and the second feature subset composed of the first n sub - features corresponding to the highest prediction accuracy and the corresponding best features obtained through Bayesian parameter search are detected in real - time;
[0089] The preset XgBoost model is iteratively trained through the best features and the best parameters to generate the pCR prediction model in real - time.
[0090] In the third aspect of the embodiments of the present invention, there is provided:
[0091] A computer, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method for constructing a pCR prediction model for esophageal cancer as described above is implemented.
[0092] In the fourth aspect of the embodiments of the present invention, there is provided:
[0093] A readable storage medium, on which a computer program is stored, wherein when the program is executed by a processor, the method for constructing a pCR prediction model for esophageal cancer as described above is implemented.
[0094] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Brief Description of the Drawings
[0095] Figure 1 Flow chart of the construction method of the pCR prediction model for esophageal cancer provided in the first embodiment of the present invention;
[0096] Figure 2 Schematic diagram of the work flow of the construction method of the pCR prediction model for esophageal cancer provided in the second embodiment of the present invention;
[0097] Figure 3 Schematic diagram of the work flow of the multi-scale attention module in the construction method of the pCR prediction model for esophageal cancer provided in the second embodiment of the present invention;
[0098] Figure 4 Block diagram of the structure of the construction system of the pCR prediction model for esophageal cancer provided in the third embodiment of the present invention.
[0099] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. Specific Embodiments
[0100] For ease of understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided so that the disclosure of the present invention is thorough and complete.
[0101] It should be noted that when an element is referred to as being "fixedly provided on" another element, it can be directly on the other element or there can also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.
[0102] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.
[0103] Please refer to Figure 1, which shows the construction method of the pCR prediction model for esophageal cancer provided by the first embodiment of the present invention. The construction method of the pCR prediction model for esophageal cancer provided by this embodiment can objectively and accurately complete the prediction of patients' pCR, correspondingly improving the work efficiency.
[0104] Specifically, this embodiment provides:
[0105] A construction method of a pCR prediction model for esophageal cancer, specifically including the following steps:
[0106] Step S10, obtain a number of original sample data, and preprocess the number of original sample data to correspondingly generate a number of target sample data, and split the number of target sample data into a training validation set and a test set correspondingly. Among them, each original sample data contains a CT image and its corresponding ROI;
[0107] Step S20, based on the first preset rule, train a corresponding image processing model in real time according to the training validation set and the preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs the corresponding CNN features in real time;
[0108] Step S30, extract the corresponding radiomics features in the CT images of the target sample data in real time through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features to generate corresponding training features in real time;
[0109] Step S40, based on the second preset rule, perform prediction training on the preset XgBoost model through the training features to correspondingly train the pCR prediction model.
[0110] Further, the step of training a corresponding image processing model in real time according to the training validation set and the preset neural network based on the first preset rule includes: ?
[0111] When the training validation set is obtained in real time, perform a convolution operation on the training validation set through the convolution kernel in the preset neural network to extract the initial features contained in the training validation set in real time;
[0112] Generate corresponding initial feature maps according to the preset batch normalization layer and the preset activation function according to the initial features, and generate the image processing model according to the initial feature maps and the preset neural network.
[0113] Further, the step of generating the image processing model according to the initial feature maps and the preset neural network includes:
[0114] When the initial feature map is obtained in real time, perform residual processing on the initial feature map through a preset residual block to output the corresponding first intermediate feature map in real time;
[0115] Perform channel splitting, global average pooling, and Softmax processing on the first intermediate feature map in sequence through a preset multi-scale attention module to output the corresponding second intermediate feature map in real time;
[0116] Further, the steps of generating the second intermediate feature map according to the multi-scale attention module include:
[0117] Input feature map \(X\in\mathbb{R}\) C×H×W Obtain different branches through channel splitting, and each branch uses a convolution kernel \(k\) i \(\times k\) i and the corresponding number of groups \(G\) i , to generate a multi-scale feature map \(F\) i :
[0118] \(F\) i =\(\text{Conv}(k\) i \(\times k\) i ,G\) i )(X), i = 0, 1, …, S - 1
[0119] where the convolution kernel size \(k\) i = 2×(i + 1)+1, the number of groups is used to control the computational complexity, and the output feature map \(F\) i \(\in\mathbb{R}\) C′×H×W , \(C\) ′ = C / S, and S is the number of branches.
[0120] For each scale of the feature map \(F\) i , generate the channel attention weight \(Z\) through global average pooling and a fully connected layer i :
[0121] \(Z\) i =\(\sigma(W_1\delta(W_0(g\) c )))\)
[0122] where is global average pooling, \(W_0\in\mathbb{R}\) C′×C′ / r and \(W_1\in\mathbb{R}\) C′ / r×C′ are fully connected layers, \(r\) is the compression ratio. \(\delta\) is the ReLU activation function, and \(\sigma\) is the Sigmoid function.
[0123] Normalize the attention weights \(Z\) of different scales through the Softmax function to achieve cross-scale interaction: i
[0124]
[0125] Concatenate the calibration weights at all scales:
[0126]
[0127] Multiply the weight att i with the feature map F at the corresponding scale i channel by channel, and concatenate to obtain the final output second feature map:
[0128] Y i = F i ⊙ att i , Out = Cat([Y0, Y1, …, Y S-1 )
[0129] Generate the image processing model according to the second intermediate feature map and the preset neural network, and the first intermediate feature map and the second intermediate feature map have the same size.
[0130] Further, the step of generating the image processing model according to the second intermediate feature map and the preset neural network includes:
[0131] When the second intermediate feature map is obtained in real time, perform iterative processing on the second intermediate feature map through the preset residual block and the preset multi-scale attention module to output the corresponding target feature map in real time;
[0132] Output the prediction value corresponding to the target feature map in real time through a preset fully connected layer, and perform model training on the preset neural network through the prediction value and a preset cross-entropy loss to output the image processing model in real time.
[0133] Further, the step of predicting and training the preset XgBoost model based on the second preset rule through the training features to correspondingly train the pCR prediction model includes:
[0134] When the training features are obtained in real time, use the preset RFECV algorithm to screen out the first feature subset in the training features in real time. The RFECV algorithm is a feature selection algorithm. Its principle is to first use the recursive feature elimination algorithm (RFE) to sort the features, and then perform cross-validation (CV) on the sorted features to select the best feature subset. RFECV obtains the importance of each feature through the coef_ attribute or feature_importances_ attribute returned by the learner. Then, remove the least important feature from the current feature set, and continuously repeat this recursive step on the feature set until the required number of features is finally reached. The first feature subset contains several sub-features;
[0135] The contribution degree generated by each of the sub - features when the preset XgBoost model makes a model prediction is calculated in real - time through the preset SHAP algorithm, and each of the sub - features is sorted in real - time according to the magnitude of the contribution degree to generate a corresponding feature subset sequence list in real - time;
[0136] The pCR prediction model is trained according to the feature subset sequence list and the preset XgBoost model.
[0137] Further, the step of training the pCR prediction model according to the feature subset sequence list and the preset XgBoost model includes:
[0138] When the feature subset sequence list is obtained in real - time, the first n sub - features are screened out in the feature subset sequence list in the order of ranking as different second feature subsets;
[0139] The pCR prediction model is trained according to the second feature subset and the preset XgBoost model.
[0140] Further, the step of training the pCR prediction model according to the second feature subset and the preset XgBoost model includes:
[0141] When the different second feature subsets are obtained in real - time, the preset XgBoost model is initialized and trained according to the second feature subset through the preset Bayesian optimization search algorithm to output a number of prediction accuracies in real - time;
[0142] The highest prediction accuracy is detected in real - time, and the second feature subset composed of the first n sub - features corresponding to the highest prediction accuracy and the corresponding best features obtained through Bayesian parameter search are detected in real - time;
[0143] The preset XgBoost model is iteratively trained through the best features and the best parameters to generate the pCR prediction model in real - time.
[0144] Specifically, for the sake of easy understanding, in this embodiment, it should be noted that in order to objectively and effectively train the required pCR prediction model, the present invention needs to first obtain a certain number of original sample data. Based on this, all sample data are divided into a training - validation group and a test group. There are 88 sample data before treatment, after treatment, and before surgery respectively, and they are randomly divided into a training - validation group and a test group according to 7:3. Among them, the training - validation group contains 60 data; the test group contains 28 data. Each data contains a CT image and an ROI (gray - level histogram).
[0145] The original input image is augmented. The original image undergoes synchronous transformation through two methods: random horizontal flipping and random rotation. The probability of horizontal flipping is set to 0.5, and the rotation angle is randomly selected from -30° to 30°.
[0146] For each augmented sample, the CT image and ROI are cropped. When cropping, the tumor centroid is used as the center, and the cropping size is 224×224×32, so that the image can adapt to the input of the 3D convolutional neural network model, while ensuring that the tumor area is not cropped off.
[0147] The training and validation set images after data preprocessing are input into the ResNet18_3D network for training. First, a 3D convolutional kernel of 15 is used to perform convolution operations on the input data to extract the initial features of the input data, and then through the batch normalization layer (BN layer) and the activation function (ReLu). The size of the output feature map is 112×112×16, that is, the above-mentioned initial feature map is generated in real time.
[0148] The obtained initial feature map enters the residual block. First, through convolutional layer 1, batch normalization layer, and ReLu activation 1, then through convolutional layer 2, batch normalization layer, and ReLu activation 2, and then through a skip connection to complete a residual unit. After two residual units, the operation of a residual block is completed. After passing through a residual block, the size of the feature map is 56×56×8, that is, the above-mentioned first intermediate feature map is generated in real time.
[0149] The first intermediate feature map output by the residual block enters the multi-scale attention module. The input first intermediate feature map is split by channels, and convolutional operations of different scales are performed, and then the multi-scale feature maps are concatenated. Then, global average pooling is performed, and the linear information between channels is combined through a fully connected layer. Then, Softmax is performed, and the obtained weights and the corresponding feature maps are multiplied element by element. The size of the feature map after passing through the multi-scale attention module remains unchanged. After passing through the first multi-scale attention module, the size of the feature map is 56×56×8, that is, the above-mentioned second intermediate feature map is generated in real time.
[0150] The output second intermediate feature map enters the repeated residual blocks and multi-scale attention modules. A total of four repeated structures are passed through to obtain the output of the ResNet18_3D network. Each time passing through a residual block, the size of this second intermediate feature map is reduced by 2 times. After passing through four residual blocks and multi-scale attention modules, the size of the output feature map is 7×7×1, that is, the above-mentioned target feature map is generated in real time.
[0151] The output of the ResNet18_3D network is used to obtain predicted values through a fully connected layer, and cross-entropy loss is used to supervise the model training. The number of training epochs is set to 65, the optimizer is Adam, the initial learning rate is 0.0001, and it decays to 1 / 2 of the original every 10 epochs. The model with the highest training accuracy is saved.
[0152] The CT images in the target sample data are input into the trained 3D ResNet18 model to obtain CNN features with a length of 512.
[0153] The original mask and CT images are fed into the Radiomics feature extractor to extract radiomics features, obtaining radiomics features R with a length of 1688. The obtained radiomics features are preliminarily screened to delete features with the same values for all samples, and then the spatial resolution of the supplementary feature scan images is added. Finally, the length of the pre-treatment radiomics features is 1539, the length of the post-treatment radiomics features is 1509, and the length of the pre-surgery radiomics features is 1510. It should be noted that the radiomics features disclosed in the present invention include various quantitative indicators such as shape, texture, and signal intensity distribution. The radiomics features are extracted by pyradiomic, and all image types and all feature categories are enabled during the extraction process. The radiomics image types used include ①Original: the original image; ②Wavelet: wavelet filtering; ③LoG: Laplacian of Gaussian filter, an edge enhancement filter that emphasizes regions of gray-scale change. The sigma parameter defines the texture roughness to be emphasized. A lower value emphasizes finer textures, and a higher value emphasizes coarser textures; ④Square: square, that is, obtaining the square of the image intensity value; ⑤SquareRoot: square root, that is, obtaining the square root of the image intensity value; ⑥Logarithm: logarithm, obtaining the logarithm of the absolute image intensity + 1; ⑦Exponential: exponential, obtaining the exponential of the image intensity value; ⑧Gradient: gradient, obtaining the magnitude of the local gradient of the image; ⑨LBP: Local Binary Pattern local binary model. A total of 8 types of radiomics features are extracted, as shown in Table 1 specifically.
[0154] Table 1 Radiomics Feature Types
[0155] Abbreviation Full name Remarks Firstorder First Order Statistics First-order statistic Shape Shape-based(3D) Shape Shape Shape-based(2D) Shape GLCM Gray Level Cooccurence Matrix Gray-level co-occurrence matrix GLRLM Gray Level Run Length Matrix Gray-level run length matrix GLSZM Gray Level Size ZoneMatrix Gray-level size zone matrix NGTDM Neighbouring Gray Tone Difference Matrix Neighbouring gray tone difference matrix GLDM Gray Level Dependence Matrix Gray-level dependence matrix
[0156] According to the requirements, the CNN features and radiomics features in each period are concatenated, the Pearson correlation coefficient and distance correlation coefficient between each feature and the prognosis result are calculated, and then the correlation coefficients between the features are calculated. The results show that some of the features we extracted have a strong correlation with the result. The features with extremely high correlation are excluded to obtain the preliminarily screened features S, that is, the above-mentioned training features are generated in real time.
[0157] From the features S, the best feature subset S1 is selected using the recursive feature elimination with cross-validation (RFECV) method. On the feature subset S1, the XgBoost model is initialized and trained using Bayesian optimization to search for the best initial parameters, and the SHAP (SHapley Additive exPlanations for machine learning model interpretation) method is used to calculate the contribution degree of each feature in S1 to the model prediction. The features in S1 are re-sorted from high to low according to the contribution degree to obtain the ordered feature subset S2, that is, the above-mentioned training feature sequence list is generated in real time; successively select the first n features from the feature subset S2 to form the feature subset S3_n, use Bayesian optimization to search for the best initial parameters to initialize and train the XgBoost model, and select S3_n with the highest prediction accuracy as the best feature, and the parameters of the model at this time as the best parameters.
[0158] We constructed seven models respectively: a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on pre-treatment single-phase CT images, a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on post-treatment single-phase CT images, a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on pre-operative single-phase CT images, a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on two-phase CT images of pre-treatment + post-treatment, a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on two-phase CT images of pre-treatment + pre-operative, a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on two-phase CT images of post-treatment + pre-operative, and a model for predicting pCR of neoadjuvant chemoradiotherapy for esophageal cancer based on three-phase CT images of pre-treatment + post-treatment + pre-operative. ROIpr (n = 7), ROIpo (n = 12), ROIpre (n = 6), ROIpr+po (n = 7), ROIpr+pre (n = 6), ROIpo+pre (n = 7), ROIpr+po+pre (n = 8) are selected as the most predictive features
[0159] for model construction. The number of features extracted and the number of features after screening in each group of models are shown in Table 2.
[0160] Table 2 Number of features extracted and screened in each group of models
[0161]
[0162] Among them, pr: (prior treatment); po: (post treatment); pre (pre-operation); pr + po: (before treatment + after treatment); pr + pre: (before treatment + before operation); po + pre: (after treatment + before operation); pr + po + pre (before treatment + after treatment + before operation).
[0163] Initialize the XgBoost model using the best parameters after screening, and then train the XgBoost model on the training set using the best features. Test the trained XgBoost model on the test set. Through the above method, a model for predicting pCR can be objectively and effectively trained, thus eliminating the need for manual judgment and correspondingly improving work efficiency.
[0164] Please refer to Figure 4 , the third embodiment of the present invention provides:
[0165] A construction system for a pCR prediction model for esophageal cancer, wherein the system includes:
[0166] A splitting module, configured to obtain a plurality of original sample data, preprocess the plurality of original sample data to correspondingly generate a plurality of target sample data, and split the plurality of target sample data into a training validation set and a test set, wherein each of the original sample data includes a CT image and its corresponding ROI;
[0167] A first training module, configured to train a corresponding image processing model in real time based on a first preset rule according to the training validation set and a preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs corresponding CNN features in real time;
[0168] A processing module, configured to extract corresponding radiomics features in the CT images of the target sample data in real time through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features to generate corresponding training features in real time;
[0169] A second training module, configured to perform predictive training on a preset XgBoost model based on a second preset rule through the training features to correspondingly train a pCR prediction model.
[0170] Further, the first training module is specifically configured to:
[0171] When the training and validation set is obtained in real time, perform a convolution operation on the training and validation set through the convolutional kernel in the preset neural network to extract the initial features contained in the training and validation set in real time;
[0172] According to the preset batch normalization layer and the preset activation function, generate the corresponding initial feature map in real time based on the initial features, and generate the image processing model according to the initial feature map and the preset neural network.
[0173] Further, the first training module is specifically used for:
[0174] When the initial feature map is obtained in real time, perform residual processing on the initial feature map through the preset residual block to output the corresponding first intermediate feature map in real time;
[0175] Perform channel splitting, global average pooling, and Softmax processing on the first intermediate feature map in sequence through the preset multi-scale attention module to output the corresponding second intermediate feature map in real time;
[0176] Generate the image processing model according to the second intermediate feature map and the preset neural network. The first intermediate feature map and the second intermediate feature map have the same size.
[0177] Further, the first training module is specifically used for:
[0178] When the second intermediate feature map is obtained in real time, perform iterative processing on the second intermediate feature map through the preset residual block and the preset multi-scale attention module to output the corresponding target feature map in real time;
[0179] Output the prediction value corresponding to the target feature map in real time through the preset fully connected layer, and perform model training on the preset neural network through the prediction value and the preset cross-entropy loss to output the image processing model in real time.
[0180] Further, the second training module is specifically used for:
[0181] When the training features are obtained in real time, screen out the first feature subset in the training features through the preset RFECV algorithm. The first feature subset contains several sub-features;
[0182] Calculate the contribution degree generated by each sub-feature when the preset XgBoost model performs model prediction in real time through the preset SHAP algorithm, and perform sorting processing on each sub-feature in real time according to the size of the contribution degree to generate the corresponding feature subset sequence table in real time;
[0183] The pCR prediction model is trained according to the feature subset sequence list and the preset XgBoost model.
[0184] Further, the second training module is specifically configured to:
[0185] When the feature subset sequence list is obtained in real time, the first n sub-features are screened out in real time from the feature subset sequence list in the order of ranking to form the second feature subset;
[0186] The pCR prediction model is trained according to the second feature subset and the preset XgBoost model.
[0187] Further, the second training module is specifically configured to:
[0188] When the second feature subset is obtained in real time, the preset XgBoost model is initialized and trained according to the second feature subset through a preset Bayesian optimization search algorithm to output a number of prediction accuracies in real time;
[0189] The highest prediction accuracy is detected in real time, and the best features and best parameters corresponding to the highest prediction accuracy are detected in real time;
[0190] The preset XgBoost model is iteratively trained through the best features and the best parameters to generate the pCR prediction model in real time.
[0191] A fourth embodiment of the present invention provides a computer, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method for constructing a pCR prediction model for esophageal cancer as described above is implemented.
[0192] A fifth embodiment of the present invention provides a readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method for constructing a pCR prediction model for esophageal cancer as described above is implemented.
[0193] In summary, the method and system for constructing a pCR prediction model for esophageal cancer provided in the above embodiments of the present invention can objectively and accurately predict the pCR of patients, correspondingly improving the work efficiency.
[0194] It should be noted that the above-mentioned each module can be a functional module or a program module, and can be implemented by software or by hardware. For the modules implemented by hardware, the above-mentioned each module can be located in the same processor; or the above-mentioned each module can also be located in different processors in any combination form.
[0195] The logic and / or steps represented in the flowchart or otherwise described herein can be considered, for example, as a definite sequence list of executable instructions for implementing a logical function, which can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in conjunction with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.
[0196] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection portion (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then storing it in a computer memory.
[0197] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.
[0198] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0199] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.
Claims
1. A method for constructing a pCR prediction model for esophageal cancer, characterized in that, The method includes: Obtain a number of original sample data, preprocess the number of original sample data to correspondingly generate a number of target sample data, and correspondingly split the number of target sample data into a training and validation set and a test set. Wherein, each original sample data contains a CT image and its corresponding ROI; Based on a first preset rule, train a corresponding image processing model in real time according to the training and validation set and a preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs corresponding CNN features in real time; Extract corresponding radiomics features in the CT images of the target sample data in real time through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features to generate corresponding training features in real time; Based on a second preset rule, perform prediction training on a preset XgBoost model through the training features to correspondingly train a pCR prediction model.
2. The construction method of the pCR prediction model for esophageal cancer according to claim 1, wherein: The step of training a corresponding image processing model in real time according to the training and validation set and a preset neural network based on a first preset rule includes: When the training and validation set is obtained in real time, perform a convolution operation on the training and validation set through the convolution kernel in the preset neural network to extract the initial features contained in the training and validation set in real time; Generate corresponding initial feature maps in real time according to a preset batch normalization layer and a preset activation function according to the initial features, and generate the image processing model according to the initial feature maps and the preset neural network.
3. The construction method of the pCR prediction model for esophageal cancer according to claim 2, characterized in that: The step of generating the image processing model according to the initial feature maps and the preset neural network includes: When the initial feature maps are obtained in real time, perform residual processing on the initial feature maps through a preset residual block to output corresponding first intermediate feature maps in real time; Perform channel splitting, global average pooling, and Softmax processing on the first intermediate feature maps in sequence through a preset multi-scale attention module to output corresponding second intermediate feature maps in real time; Generate the image processing model according to the second intermediate feature maps and the preset neural network, and the first intermediate feature maps and the second intermediate feature maps have the same size.
4. The construction method of the pCR prediction model for esophageal cancer according to claim 3, characterized in that: The step of generating the image processing model according to the second intermediate feature maps and the preset neural network includes: When the second intermediate feature maps are obtained in real time, perform iterative processing on the second intermediate feature maps through the preset residual block and the preset multi-scale attention module to output corresponding target feature maps in real time; Output a prediction value corresponding to the target feature map in real time through a preset fully connected layer, and perform model training on the preset neural network through the prediction value and a preset cross-entropy loss to output the image processing model in real time.
5. The construction method of the pCR prediction model for esophageal cancer according to claim 1, wherein: The step of performing prediction training on a preset XgBoost model through the training features based on a second preset rule to correspondingly train a pCR prediction model includes: When the training features are obtained in real time, a first feature subset is screened out in real time from the training features through a preset RFECV algorithm, and a number of sub-features are included inside the first feature subset; Through a preset SHAP algorithm, the contribution degree generated by each of the sub-features when the preset XgBoost model makes a model prediction is calculated in real time, and each of the sub-features is sorted in real time according to the magnitude of the contribution degree, so as to generate a corresponding feature subset sequence list in real time; According to the feature subset sequence list and the preset XgBoost model, the pCR prediction model is trained correspondingly.
6. The construction method of the pCR prediction model for esophageal cancer according to claim 5, wherein: The step of training the pCR prediction model according to the feature subset sequence list and the preset XgBoost model includes: When the feature subset sequence list is obtained in real time, the first n sub-features are screened out in real time from the feature subset sequence list in the order of ranking as the second feature subset; According to the second feature subset and the preset XgBoost model, the pCR prediction model is trained correspondingly.
7. The construction method of the pCR prediction model for esophageal cancer according to claim 6, characterized in that: The step of training the pCR prediction model according to the second feature subset and the preset XgBoost model includes: When the second feature subset is obtained in real time, the preset XgBoost model is initialized and trained according to the second feature subset through a preset Bayesian optimization search algorithm, so as to output a number of prediction accuracies in real time; The highest prediction accuracy is detected in real time, and the best feature and the best parameters corresponding to the highest prediction accuracy are detected in real time; The preset XgBoost model is iteratively trained through the best feature and the best parameters, so as to generate the pCR prediction model in real time.
8. A construction system for a pCR prediction model for esophageal cancer, characterized in that, The system includes: A splitting module, configured to obtain a number of original sample data, preprocess the number of original sample data, so as to generate a number of target sample data correspondingly, and split the number of target sample data into a training validation set and a test set correspondingly, wherein each of the original sample data includes a CT image and an ROI corresponding thereto; A first training module, configured to train a corresponding image processing model in real time based on a first preset rule according to the training validation set and a preset neural network, and input the CT images in the target sample data into the image processing model correspondingly, so that the image processing model outputs corresponding CNN features in real time; A processing module, configured to extract corresponding radiomics features in the CT images of the target sample data through a preset feature extractor, and perform real-time splicing processing on the CNN features and the radiomics features, so as to generate corresponding training features in real time; A second training module, configured to perform prediction training on a preset XgBoost model based on a second preset rule through the training features, so as to train a pCR prediction model correspondingly.
9. A computer, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, the method for constructing a pCR prediction model for esophageal cancer according to any one of claims 1 to 7 is implemented.
10. A readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method for constructing a pCR prediction model for esophageal cancer according to any one of claims 1 to 7.