A radiotherapy method based on image processing

Abstract

This invention provides an image processing-based radiotherapy method, belonging to the field of medical image processing and radiotherapy technology. This method acquires multi-dimensional data such as the patient's CT images, physiological state information, body mass index, and gender. After preprocessing the CT images, it utilizes a deep convolutional neural network and multimodal feature fusion technology to extract and fuse image features, physiological state features, weight features, and gender features to generate high-quality feature vectors. Furthermore, through semi-supervised learning, combined with labeled and unlabeled CT images, it adaptively adjusts the radiotherapy dosage, thereby achieving precise radiotherapy. This invention, employing the above-mentioned image processing-based radiotherapy method, can effectively improve the personalization and accuracy of radiotherapy planning, and has significant technical application value.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and radiotherapy technology, and in particular to a radiotherapy method based on image processing. Background Technology

[0002] In the field of radiotherapy, ensuring the accuracy and personalization of radiotherapy plans is crucial. Traditional radiotherapy methods often rely on the physician's experience and manual adjustments, which can lead to inaccurate plans and unnecessary radiation doses to patients. With the rapid development of medical imaging technology, image processing-based radiotherapy methods have gradually become a research hotspot.

[0003] However, existing image processing-based radiotherapy methods still face several challenges. First, patient CT images often contain significant amounts of noise and redundant information, increasing the difficulty of feature extraction. Second, radiotherapy planning requires comprehensive consideration of multi-dimensional patient data, including physiological status, body mass index, and gender, but existing methods often only utilize single-modality data, failing to fully leverage the complementary information between multi-modal data. Furthermore, due to the high cost of radiotherapy data annotation, existing radiotherapy planning training models often face the problem of scarce labeled data, limiting the model's generalization ability and accuracy.

[0004] Therefore, there is an urgent need for an image processing-based radiotherapy method that can overcome the aforementioned challenges in order to improve the accuracy and personalization of radiotherapy planning. This invention is proposed precisely to address this problem. Summary of the Invention

[0005] The purpose of this invention is to provide a radiotherapy method based on image processing, which can effectively improve the personalization and accuracy of radiotherapy planning and has important technical application value.

[0006] To achieve the above objectives, the present invention provides a radiotherapy method based on image processing, comprising the following steps:

[0007] Step S1: Obtain patient data, including the patient's current cycle CT images, physiological status information, body mass index, and gender;

[0008] Step S2: Preprocess the acquired CT images to obtain preprocessed CT images;

[0009] Step S3: Based on the deep convolutional neural network and multimodal feature fusion method, the features, physiological state features, weight features and gender features in the preprocessed CT image are extracted and fused to obtain a multimodal feature map. The multimodal feature map is further processed based on the deep feature enhancement and sparsification method to obtain a high-quality feature vector for radiotherapy planning dosage adjustment.

[0010] Step S4: Introduce the high-quality feature vector obtained in step S3 into a semi-supervised learning method, use labeled and unlabeled CT images for joint training, obtain comprehensive analysis results, and adaptively adjust the radiotherapy plan dosage based on the comprehensive analysis results.

[0011] Preferably, in step S1, the physiological state information includes heart rate, respiratory rate, blood pressure, blood oxygen saturation, and heart rate variability.

[0012] Preferably, in step S1, the formula for calculating body mass index is as follows:

[0013]

[0014] Where BMI stands for Body Mass Index; T' represents weight; and S represents height.

[0015] Preferably, in step S2, the preprocessing includes: random cropping, random flipping, resampling, contrast enhancement, and noise removal.

[0016] Preferably, in step S3, based on the deep convolutional neural network and multimodal feature fusion method, features, physiological state features, weight features, and gender features in the preprocessed CT image are extracted and fused to obtain a multimodal feature map. The multimodal feature map is then further processed based on deep feature enhancement and sparsification methods to obtain a high-quality feature vector for adjusting the radiotherapy dosage. The specific operations are as follows:

[0017] Step S31: Use a deep convolutional neural network to extract features from the preprocessed CT image;

[0018] First convolutional layer F conv1 as follows:

[0019] F conv1 =ReLU(W1*I+b1);

[0020] Where ReLU represents the activation function; W1 represents the convolution kernel; I represents the preprocessed CT image; b1 represents the bias term; and * represents the convolution operation.

[0021] Second convolution F conv2 as follows:

[0022] F conv2 =ReLU(W2*F conv1 +b2);

[0023] Where W2 represents the convolution kernel; b2 represents the bias term;

[0024] Step S32: Introduce a channel attention mechanism to enhance the discriminative power of the features, and obtain image features F. img ;

[0025]

[0026] Where AvgPool and MaxPool represent global average pooling and max pooling operations, respectively; W ca Represents the weight matrix; σ represents element-wise multiplication; b represents the Sigmoid activation function; ca Indicates the bias term;

[0027] Step S33: Extract physiological state features F using the BERT model. SL The gender is converted into a numerical form using a label encoding method, and the gender feature F is extracted. gender Encode the body mass index as a numerical feature F BMI ;

[0028] Step S34: Apply a cross-attention mechanism between features of different modalities to achieve dynamic alignment and information complementarity between modalities;

[0029]

[0030] Among them, F cross Indicates the enhanced features; F i Represents the characteristics of different modes; N represents the number of modes; i represents the index of the mode; α i α represents the weights calculated through the attention mechanism. i The calculation formula is as follows:

[0031]

[0032] Where j represents the index variable, used to iterate over the features F of all N different modalities in the summation operation in the denominator. j W att b represents the weight vector; att Indicates the bias term;

[0033] Step S35: Fuse the enhanced features to obtain the fused feature F. fused ;

[0034] F fused =W fues ×F cross +b fuse ;

[0035] Among them, W fues b represents the weight matrix of the fusion layer; fues Indicates the bias term;

[0036] Step S36: Analyze the fused features F fusedFurther enhancement and sparsification processes are performed to obtain the sparsified feature F. sparse This refers to high-quality feature vectors;

[0037]

[0038] Among them, W sparse b represents the weight matrix; sparse This indicates the bias term.

[0039] Preferably, in step S4, the high-quality feature vector obtained in step S3 is introduced into a semi-supervised learning method, and joint training is performed using labeled and unlabeled CT images to obtain a comprehensive analysis result. The radiotherapy dosage is then adaptively adjusted based on the comprehensive analysis result. The specific operation is as follows:

[0040] Step S41: Further fuse the high-quality input feature vectors using the Multi-Level Feature Fusion (MLFF) module;

[0041]

[0042] in, W represents the sparsified features of the k-th layer; k b represents the weight matrix of the k-th layer; k This represents the bias term of the k-th layer; K represents the number of layers fused.

[0043] Step S42: Extract discriminative local features F through the local voting module MPV. MPV ;

[0044] F MPV =W MPV ×F MLFF +b MPV ;

[0045] Among them, W MPV b represents the weight matrix; MPV Indicates the bias term;

[0046] Step S43: Use the neighborhood anchor discovery algorithm AND all samples to extract features;

[0047]

[0048] Among them, F AND This represents the features extracted by the AND algorithm; Indicates an AND network;

[0049] Step S44: Using a comprehensive strategy that considers both the mean and consistency of sample outputs, filter out pseudo-labels y for unlabeled samples. pscudo ;

[0050]

[0051] Where T represents the number of iterations; f() represents the prediction function of the deep learning model; θ t This represents the model parameters for the t-th iteration;

[0052] Step S45: Define the loss function as a weighted sum of supervised and unsupervised losses;

[0053] L = L labeled +λL unlabeled ;

[0054] Where L represents the total loss; L labeled This indicates a supervisory loss; L unlabeled λ represents the unsupervised loss; λ represents the hyperparameter used to balance supervised and unsupervised losses.

[0055]

[0056] Among them, D labeled D represents the labeled sample set; unlabeled x represents the set of unlabeled samples; i This represents the input features of labeled samples; y i This indicates the actual planned radiotherapy dosage for labeled samples; x j f represents the input features of the unlabeled samples; f() represents the prediction function of the deep learning model;

[0057] Step S46: Update the model parameters θ using gradient descent by minimizing the total loss;

[0058]

[0059] Where θ' represents the updated model parameters; β represents the learning rate; This represents the gradient of the loss function with respect to the model parameters;

[0060] Step S47: Apply the trained model to the new patient data x new To predict and obtain the planned radiotherapy dosage y new ;

[0061] y new =f(x) new ;θ);

[0062] Here, f() represents the prediction function of the deep learning model.

[0063] Therefore, the present invention employs the above-described image processing-based radiotherapy method, and the beneficial technical effects are as follows:

[0064] First, this invention effectively addresses the issues of noise and redundant information in CT images by employing deep convolutional neural networks and multimodal feature fusion technology, thereby improving the accuracy and efficiency of feature extraction. This method fully utilizes multidimensional patient data, including physiological status information, body mass index, and gender, achieving complementary information fusion between multimodal data and providing more comprehensive and accurate information support for radiotherapy planning.

[0065] Secondly, this invention introduces a semi-supervised learning method, using both labeled and unlabeled CT images for joint training, effectively alleviating the problem of high cost in radiotherapy data labeling. Through this method, this invention can improve the model's generalization ability and accuracy with limited labeled data, providing strong technical support for the personalized development of radiotherapy plans. Attached Figure Description

[0066] Figure 1 This is a flowchart of an image processing-based radiotherapy method according to the present invention;

[0067] Figure 2 This is a flowchart for semi-supervised learning training. Detailed Implementation

[0068] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0070] Example 1

[0071] like Figures 1-2 This invention provides a radiotherapy method based on image processing, comprising the following steps:

[0072] Step S1: Obtain patient data, including the patient's current cycle CT images, physiological status information, body mass index, and gender.

[0073] CT images contain detailed information about the location, size, and surrounding tissues of the tumor.

[0074] Physiological information includes heart rate, respiratory rate, blood pressure, blood oxygen saturation, and heart rate variability. These physiological parameters reflect the patient's overall health status and have a significant impact on radiotherapy planning.

[0075] The formula for calculating body mass index is as follows:

[0076]

[0077] Where BMI stands for Body Mass Index; T' represents weight; and S represents height.

[0078] Differences in physiological structure and metabolism between sexes affect the tolerance and efficacy of radiotherapy.

[0079] Step S2: Preprocess the acquired CT images to obtain preprocessed CT images.

[0080] Preprocessing includes: random cropping, random flipping, resampling, contrast enhancement, and noise removal.

[0081] Random cropping: CT images are randomly cropped to generate image fragments of different sizes, enhancing the model's generalization ability.

[0082] Random flipping: Flipping images horizontally or vertically to increase data diversity.

[0083] Resampling: Adjusting the resolution of the image to meet the requirements of the model input.

[0084] Enhance contrast: By adjusting the contrast of the image, the boundary between the tumor and the surrounding tissue becomes clearer.

[0085] Noise Removal: Using filtering algorithms to remove noise from images and improve image quality.

[0086] Step S3: Based on the deep convolutional neural network and multimodal feature fusion method, features, physiological state features, weight features, and gender features in the preprocessed CT image are extracted and fused to obtain a multimodal feature map. The multimodal feature map is then further processed using deep feature enhancement and sparsification methods to obtain a high-quality feature vector for adjusting radiotherapy dosage. The specific operations are as follows:

[0087] Step S31: Use a deep convolutional neural network to extract features from the preprocessed CT image;

[0088] First convolutional layer F conv1 as follows:

[0089] F conv1 =ReLU(W1*I+b1);

[0090] Where ReLU represents the activation function; W1 represents the convolution kernel; I represents the preprocessed CT image; b1 represents the bias term; and * represents the convolution operation.

[0091] Second convolution F conv2 as follows:

[0092] F conv2 =ReLU(W2*F conv1 +b2);

[0093] Where W2 represents the convolution kernel; b2 represents the bias term;

[0094] Step S32: Introduce a channel attention mechanism to enhance the discriminative power of the features, and obtain image features F. img ;

[0095]

[0096] Where AvgPool and MaxPool represent global average pooling and max pooling operations, respectively; W ca Represents the weight matrix; σ represents element-wise multiplication; b represents the Sigmoid activation function; ca Indicates the bias term;

[0097] Step S33: Extract physiological state features F using the BERT model. SL The BERT model can process serialized physiological parameter data and extract feature vectors that reflect the patient's physiological state.

[0098] Gender information is converted into numerical form through label encoding, and gender feature F is extracted. gender For example, 0 is for males and 1 is for females.

[0099] Encode BMI values ​​as numerical features F BMI This is to facilitate integration with other features.

[0100] Step S34: Apply a cross-attention mechanism between features of different modalities to achieve dynamic alignment and information complementarity between modalities;

[0101]

[0102] Among them, F cross Indicates the enhanced features; F i Represents the characteristics of different modes; N represents the number of modes; i represents the index of the mode; α i α represents the weights calculated through the attention mechanism. i The calculation formula is as follows:

[0103]

[0104] Where j represents the index variable, used to iterate over the features F of all N different modalities in the summation operation in the denominator. j W att b represents the weight vector; att Indicates the bias term;

[0105] Step S35: Fuse the enhanced features to obtain the fused feature F. fused ;

[0106] F fused =W fues ×F cross +b fuse ;

[0107] Among them, W fues b represents the weight matrix of the fusion layer; fues Indicates the bias term;

[0108] Step S36: Further enhance and sparsify the fused features to obtain the sparsified features F. sparse This refers to high-quality feature vectors;

[0109]

[0110] Among them, W sparse b represents the weight matrix; sparse This indicates the bias term.

[0111] The high-quality feature vectors obtained through the above steps can comprehensively reflect the multi-dimensional information of patients, providing a more comprehensive and accurate basis for the formulation of radiotherapy plans.

[0112] Step S4: Introduce the high-quality feature vector obtained in Step S3 into a semi-supervised learning method, and perform joint training using labeled and unlabeled CT images to obtain a comprehensive analysis result. Adaptively adjust the radiotherapy plan dosage based on the comprehensive analysis result. The specific operation is as follows (the deep learning model used below is not specifically limited in this invention, that is, all deep learning models can be applied to this method):

[0113] Step S41: Further fuse the high-quality input feature vectors using the Multi-Level Feature Fusion (MLFF) module;

[0114]

[0115] in, W represents the sparsified features of the k-th layer; k b represents the weight matrix of the k-th layer; k This represents the bias term of the k-th layer; K represents the number of layers fused.

[0116] Step S42: Extract discriminative local features F through the local voting module MPV. MPV ;

[0117] F MPV =W MPV ×F MLFF +b MPV ;

[0118] Among them, W MPVb represents the weight matrix; MPV Indicates the bias term;

[0119] Step S43: Use the Anchor Neighborhood Discovery (AND) algorithm to extract features from all samples and mine the similarities and differences between samples;

[0120]

[0121] Among them, F AND This represents the features extracted by the AND algorithm; Indicates an AND network;

[0122] Step S44: Using a comprehensive strategy that considers both the mean and consistency of sample outputs, filter out pseudo-labels y for unlabeled samples. pscudo ;

[0123]

[0124] Where T represents the number of iterations; f() represents the prediction function of the deep learning model; θ t This represents the model parameters for the t-th iteration;

[0125] Step S45: Define the loss function as a weighted sum of supervised and unsupervised losses;

[0126] L = L labeled +λL unlabeled ;

[0127] Where L represents the total loss; L labeled This indicates a supervisory loss; L unlabeled λ represents the unsupervised loss; λ represents the hyperparameter used to balance supervised and unsupervised losses.

[0128]

[0129] Among them, D labeled D represents the labeled sample set; unlabeled x represents the set of unlabeled samples; i This represents the input features of labeled samples; y i This indicates the actual planned radiotherapy dosage for labeled samples; x j f represents the input features of the unlabeled samples; f() represents the prediction function of the deep learning model;

[0130] Step S46: Update the model parameters θ using gradient descent by minimizing the total loss;

[0131]

[0132] Where θ' represents the updated model parameters; β represents the learning rate; This represents the gradient of the loss function with respect to the model parameters;

[0133] Step S47: Apply the trained model to the new patient data x new To predict and obtain the planned radiotherapy dosage y new ;

[0134] y new =f(x) new ;θ);

[0135] Here, f() represents the prediction function of the deep learning model.

[0136] Therefore, the radiotherapy method based on image processing described above can effectively improve the personalization and accuracy of radiotherapy planning, and has important technical application value.

[0137] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A radiotherapy method based on image processing, characterized in that, Includes the following steps: Step S1: Obtain patient data, including the patient's current cycle CT images, physiological status information, body mass index, and gender; Step S2: Preprocess the acquired CT images to obtain preprocessed CT images; Step S3: Based on the deep convolutional neural network and multimodal feature fusion method, the features, physiological state features, weight features and gender features in the preprocessed CT image are extracted and fused to obtain a multimodal feature map. The multimodal feature map is further processed based on the deep feature enhancement and sparsification method to obtain a high-quality feature vector for radiotherapy planning dosage adjustment. Step S31: Use a deep convolutional neural network to extract features from the preprocessed CT image; First convolutional layer as follows: ； in, Indicates the activation function; Represents the convolution kernel; This represents the preprocessed CT image; Indicates the bias term; Indicates the convolution operation; Second convolutional layer as follows: ； in, Represents the convolution kernel; Indicates the bias term; Step S32: Introduce a channel attention mechanism to enhance the discriminative power of features and obtain image features. ; ； in, , These represent global average pooling and max pooling operations, respectively. Represents the weight matrix; This indicates element-wise multiplication; This represents the Sigmoid activation function; Indicates the bias term; Step S33: Extract physiological state features using the BERT model. The label encoding method is used to convert gender into numerical form and extract gender features. Encode body mass index as a numerical feature ; Step S34: Apply a cross-attention mechanism between features of different modalities to achieve dynamic alignment and information complementarity between modalities; ； in, Indicates the enhanced features; Features representing different modes; Indicates the number of modes; Index representing the modality; This represents the weights calculated using the attention mechanism. The calculation formula is as follows: ； in, This represents an index variable used to iterate over all values ​​in the summation operation within the denominator. Features of different modalities ; Represents the weight vector; Indicates the bias term; Step S35: Fuse the enhanced features to obtain the fused features. ; ； in, This represents the weight matrix of the fusion layer; Indicates the bias term; Step S36: Analyze the fused features Further enhancement and sparsification processes are performed to obtain the sparsified features. This refers to high-quality feature vectors; ； in, Represents the weight matrix; Indicates the bias term; In step S4, the high-quality feature vector obtained in step S3 is introduced into a semi-supervised learning method. Joint training is performed using labeled and unlabeled CT images to obtain a comprehensive analysis result. The radiotherapy dosage is then adaptively adjusted based on this result. The specific operation is as follows: Step S41: Further fuse the high-quality input feature vectors using the Multi-Level Feature Fusion (MLFF) module; ； in, Indicates the first Features of the layer after sparsification; Indicates the first Layer weight matrix; Indicates the first Layer bias terms; Indicates the number of fusion layers; Step S42: Extract discriminative local features using the local voting module MPV. ; ； in, Represents the weight matrix; Indicates the bias term; Step S43: Use the neighborhood anchor discovery algorithm AND all samples to extract features; ； in, This represents the features extracted by the AND algorithm; Indicates an AND network; Step S44: Filter out pseudo-labels for unlabeled samples using a comprehensive strategy that considers both the mean and consistency of sample outputs. ; ； in, Indicates the number of iterations; Represents the prediction function of a deep learning model; Indicates the first Model parameters for the next iteration; Step S45: Define the loss function as a weighted sum of supervised and unsupervised losses; ； in, Indicates the total loss; This indicates a loss due to oversight; This indicates unsupervised loss; This represents a hyperparameter used to balance supervised and unsupervised losses; ； ； in, This represents a set of labeled samples; Represents the set of unlabeled samples; This represents the input features of labeled samples; This indicates the actual radiotherapy treatment dosage for labeled samples; This represents the input features of unlabeled samples; Represents the prediction function of a deep learning model; Step S46: Update the model parameters using gradient descent by minimizing the total loss. ; ； in, This represents the updated model parameters; Indicates the learning rate; This represents the gradient of the loss function with respect to the model parameters; Step S47: Apply the trained model to the new patient data. To make predictions and obtain the planned dosage for radiotherapy. ; ； in, This represents the prediction function of a deep learning model.

2. The radiotherapy method based on image processing according to claim 1, characterized in that, In step S1, physiological status information includes heart rate, respiratory rate, blood pressure, blood oxygen saturation, and heart rate variability.

3. The image processing-based radiotherapy method according to claim 1, characterized in that, In step S1, the formula for calculating body mass index is as follows: ； in, Indicates body mass index; represents weight; Indicates height.

4. The image processing-based radiotherapy method according to claim 3, characterized in that, In step S2, the preprocessing includes: random cropping, random flipping, resampling, contrast enhancement, and noise removal.

Images