Radiotherapy dose prediction system based on CT images

By using a deep neural network system based on CT images to integrate image domain and frequency domain features, the problem of accurate radiotherapy dose prediction was solved, improving treatment efficacy and reducing side effects.

CN118512724BActive Publication Date: 2024-11-05AFFILIATED HOSPITAL OF JIANGSU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410723826.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-05
Publication Date
2024-11-05
Estimated Expiration
2044-06-05

AI Technical Summary

Technical Problem

Current technology makes it difficult to accurately predict radiotherapy doses, leading to insufficient or excessive radiation to the tumor area, which affects treatment effectiveness and causes complications.

Method used

The radiotherapy dose prediction system based on CT images comprehensively considers image domain and frequency domain features, and uses deep neural networks for feature extraction and fusion, including image domain convolution and frequency domain transformation, combined with high-pass filtering and pooling operations, to preserve detailed tumor feature information and improve recognition accuracy.

Benefits of technology

It improves the accuracy of radiotherapy dose prediction and treatment efficacy, reduces radiation damage to normal tissues, and lowers the incidence of complications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118512724B_ABST
    Figure CN118512724B_ABST
Patent Text Reader

Abstract

The present application provides a radiotherapy dose prediction system based on CT images, wherein the CT image acquisition unit is used to detect the patient's tumor target area and acquire the CT image; the CT frequency domain image acquisition unit performs frequency domain transformation on the CT image to generate a CT frequency domain image; the tumor feature extraction unit extracts the image domain features of the CT image and the frequency domain features of the CT frequency domain image, and performs feature fusion to generate a tumor feature vector; the radiotherapy dose prediction unit identifies the tumor feature vector and generates a radiotherapy dose prediction result. The present application improves the recognition accuracy of tumor features in CT images, accurately predicts the radiotherapy dose, and improves the therapeutic effect of radiotherapy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of image processing, and in particular to a radiation therapy dose prediction system based on CT images. Background Art

[0002] Cancer, also known as malignant tumor, is a general term for a class of diseases in which certain cells in an organism proliferate uncontrollably and invade adjacent tissues or organs. The main treatments for cancer include surgery, radiotherapy, chemotherapy, hormone therapy, and targeted therapy. Among them, about 50% to 70% of cancer patients will receive radiotherapy during the course of the disease, which is currently one of the most important cancer treatments. This method uses high-energy ionizing radiation to irradiate malignant tumors from multiple specific angles to eliminate tumor cells, alleviate clinical symptoms of patients, and prolong their survival.

[0003] When a patient is diagnosed with cancer and needs to receive radiotherapy, the tumor area needs to be scanned first to obtain a dedicated CT image in order to determine the location of the tumor target area and normal tissues and organs. Then, according to the precise range of the tumor target area and the organs at risk, the radiotherapy dose is determined and the tumor target area is treated with radiotherapy. During radiotherapy, ionizing radiation can not only destroy tumor cells, but also damage the patient's immune system. If the dose in the tumor area is insufficient, cancer cells cannot be effectively eliminated, affecting the treatment effect of cancer; while excessive doses of radiation can damage surrounding healthy tissues and organs, damage the normal functions of the patient's body, and cause complications such as hair loss, anemia, and conjunctivitis. Therefore, how to accurately predict the dose of radiotherapy, ensure that the tumor target area is adequately covered by the dose, achieve accurate evaluation of the treatment effect, improve the treatment effect of radiotherapy, and avoid excessive irradiation, reduce the occurrence of complications and side effects, has become the most important step in radiotherapy. Summary of the invention

[0004] In view of the above problems, the present application is proposed to provide a radiotherapy dose prediction system based on CT images, which comprehensively considers the image domain features and frequency domain features of the CT images, adds and splices them, and at the same time, in the pooling layer, uses the changes in the adjacent features of the rows and columns to perform two iterative poolings, which greatly improves the recognition accuracy of tumor features in CT images, accurately predicts the radiotherapy dose, and improves the therapeutic effect of radiotherapy.

[0005] In order to achieve the above-mentioned object, the present application provides a lung tumor recognition system based on a deep neural network, including a CT image acquisition unit, a CT frequency domain image acquisition unit, a tumor feature extraction unit, and a radiotherapy dose prediction unit;

[0006] A CT image acquisition unit, used for detecting a patient's tumor target area and acquiring a CT image;

[0007] A CT frequency domain image acquisition unit receives the CT image, performs frequency domain transformation on the CT image, and generates a CT frequency domain image;

[0008] A tumor feature extraction unit receives the CT image and the CT frequency domain image, extracts the image domain features of the CT image and the frequency domain features of the CT frequency domain image, and performs feature fusion to generate a tumor feature vector;

[0009] The radiotherapy dose prediction unit receives the tumor feature vector, identifies the tumor feature vector, and generates a radiotherapy dose prediction result.

[0010] Preferably, the tumor feature extraction unit specifically includes:

[0011] A first feature extraction module receives the CT image, performs image domain convolution and pooling operations on the CT image, and generates a first image domain feature map;

[0012] A first frequency domain feature extraction module receives the CT frequency domain image, performs frequency domain convolution and pooling operations on the CT image, and generates a first frequency domain feature map;

[0013] A first high-pass filtering module receives the first frequency domain feature map, performs high-pass filtering on the first frequency domain feature map, and obtains a first tumor detail feature map;

[0014] A first addition module receives the first image domain feature map and the first tumor detail feature map, performs element-by-element addition and fusion, and generates a first fused feature map;

[0015] A second feature extraction module receives the first fused feature map, performs image domain convolution and pooling operations on the first fused feature map, and generates a second image domain feature map;

[0016] A second frequency domain feature extraction module receives the first frequency domain feature map, performs frequency domain convolution and pooling operations on the first frequency domain feature map, and generates a second frequency domain feature map;

[0017] A second high-pass filtering module receives the second frequency domain feature map, performs high-pass filtering on the second frequency domain feature map, and obtains a second tumor detail feature map;

[0018] A second adding module receives the second image domain feature map and the second tumor detail feature map, performs element-by-element addition and fusion, and generates a second fused feature map;

[0019] A third feature extraction module receives the second fused feature map, performs image domain convolution and pooling operations on the second fused feature map, and generates a third image domain feature map;

[0020] A third frequency domain feature extraction module receives the second frequency domain feature map, performs frequency domain convolution and pooling operations on the second frequency domain feature map, and generates a third frequency domain feature map;

[0021] A splicing and fusion module receives the third image domain feature map and the third frequency domain feature map, performs feature splicing and fusion, and generates a splicing feature map;

[0022] The fully connected layer receives the concatenated feature map, reduces the dimension of the concatenated feature map, and converts it into a one-dimensional tumor feature vector.

[0023] Preferably, the output of the third feature extraction module is expressed as:

[0024] Output I =P3(C3(P2(C2(P1(C1(img in ))+F1(temp1)))+F2(temp2)))

[0025] temp1=P4(C4(FET(img in )))

[0026] temp2=P5(C5(P4(C4(FET(img in )))))

[0027] Among them, Output I represents the output of the third feature extraction module, C1(·), C2(·), and C3(·) represent the convolution operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, respectively, P1(·), P2(·), and P3(·) represent the pooling operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, respectively, img in represents a CT image, F1(·) and F2(·) represent the filtering operations of the first high-pass filtering module and the second high-pass filtering module respectively, temp1 represents the output of the first frequency domain feature extraction module, and temp2 represents the output of the second frequency domain feature extraction module; C4(·) and C5(·) represent the convolution operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module respectively, P4(·) and P5(·) represent the pooling operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module respectively, and FET represents the frequency domain transformation operation.

[0028] Preferably, the output of the third frequency domain feature extraction module is expressed as:

[0029] Output F =P6(C6(P5(C5(P4(C4(FET(img in )))))))

[0030] Among them, Output Frepresents the output of the third frequency domain feature extraction module, C6(·) represents the convolution operation of the third frequency domain feature extraction module, and P6(·) represents the pooling operation of the third frequency domain feature extraction module.

[0031] Preferably, the output of the splicing and fusion module is expressed as:

[0032] Output C =Cat(Output I ,Output F )

[0033] Among them, Output C represents the output of the splicing and fusion module, and Cat(·) represents the splicing operation.

[0034] Preferably, the first feature extraction module, the first frequency domain feature extraction module, the second feature extraction module, the second frequency domain feature extraction module, the third feature extraction module and the third frequency domain feature extraction module have the same structure and are connected by convolutional layers and pooling layers connected in series.

[0035] Preferably, the convolution kernel, pooling window and pooling step size of the first feature extraction module and the first frequency domain feature extraction module are the same; the convolution kernel, pooling window and pooling step size of the second feature extraction module and the second frequency domain feature extraction module are the same.

[0036] Preferably, the pooling layer comprises:

[0037] The pooling window setting module sets the size and step size of the pooling window and divides the output feature map of the convolutional layer into multiple sub-regions;

[0038] The row feature change calculation module calculates the row feature change in the generated sub-region, specifically:

[0039]

[0040] Among them, P H represents the change in row features in the sub-region, N represents the size of the pooling window, and x i,j represents the feature of the i-th row and j-th column in the subregion, x i,j+1 Represents the features of the i-th row and j+1-th column in the subregion;

[0041] The column feature change calculation module calculates the column feature change in the generated sub-region, specifically:

[0042]

[0043] Among them, P V represents the change in column features in the sub-region, x i+1,j Represents the feature of the i+1th row and jth column in the subregion;

[0044] The total feature change calculation module calculates the total feature change in the sub-region according to the row feature change and column feature change in the sub-region, specifically:

[0045]

[0046] Among them, P s represents the total feature change variation in the sub-region;

[0047] The feature map combination module combines and generates pooled feature maps based on the total feature changes in each sub-region.

[0048] Preferably, the convolution kernel size in the convolution layer is set to 3×3, and the pooling window size in the pooling layer is set to 2×2.

[0049] Preferably, the frequency domain transformation adopts two-dimensional Fourier transform or wavelet transform.

[0050] The beneficial effects of this application are:

[0051] (1) The present application provides a radiotherapy dose prediction system based on CT images, which comprehensively considers the image domain features and frequency domain features of CT images, extracts the image domain features and frequency domain features respectively, and adds and concatenates them, which can improve the recognition accuracy of tumor features in CT images, accurately predict radiotherapy doses, and improve the therapeutic effect of radiotherapy.

[0052] (2) The present application also provides a tumor feature extraction unit for a radiotherapy dose prediction system, which utilizes two branches to extract image domain features and frequency domain features of CT images, and simultaneously adds image domain information features through tumor detail features in the frequency domain, thereby compensating for the problem of loss of image domain detail features, strengthening tumor detail features in the image domain, and facilitating easier extraction of complex image features in the image domain, further enhancing the recognition accuracy of tumor features in CT images, and improving the accuracy of radiotherapy dose prediction results.

[0053] (3) The present application also provides a pooling layer for a radiotherapy dose prediction system. The pooling layer performs pooling using the changes in adjacent features of rows and columns, respectively, thereby retaining the degree of change in tumor detail feature information. At the same time, through two iterations of rows and columns, the loss of tumor detail feature information is avoided, thereby greatly improving the recognition accuracy of tumor features in CT images and the accuracy of radiotherapy dose prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A structural block diagram of a radiotherapy dose prediction system based on CT images provided in this application;

[0056] Figure 2 A structural block diagram of a tumor feature extraction unit provided in this application;

[0057] Figure 3 A structural block diagram of the feature extraction module provided in this application;

[0058] Figure 4 A structural block diagram of the pooling layer provided in this application;

[0059] Figure 5 A working example diagram of the pooling layer provided in this application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0061] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms of "a", "said", and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two, but does not exclude the inclusion of at least one.

[0062] First of all, the present application provides a radiotherapy dose prediction system based on CT images, which comprehensively considers the image domain features and frequency domain features of CT images, extracts the image domain features and frequency domain features respectively, and adds and concatenates them, which can improve the recognition accuracy of tumor features in CT images, accurately predict radiotherapy doses, and improve the therapeutic effect of radiotherapy.

[0063] Secondly, the present application also provides a tumor feature extraction unit for a radiotherapy dose prediction system, which utilizes two branches to extract image domain features and frequency domain features of CT images, and simultaneously adds image domain information features through tumor detail features in the frequency domain, thereby compensating for the problem of loss of image domain detail features, strengthening tumor detail features in the image domain, and facilitating easier extraction of complex image features in the image domain, further enhancing the recognition accuracy of tumor features in CT images, and improving the accuracy of radiotherapy dose prediction results.

[0064] Finally, the present application also provides a pooling layer for a radiotherapy dose prediction system. The pooling layer uses the change in adjacent features of rows and columns for pooling, retaining the degree of change in tumor detail feature information. At the same time, through two iterations of rows and columns, the loss of tumor detail feature information is avoided, which greatly improves the recognition accuracy of tumor features in CT images and improves the accuracy of radiotherapy dose prediction results.

[0065] The present application is further described below in conjunction with the accompanying drawings and specific embodiments.

[0066] First embodiment

[0067] like Figure 1 As shown, the present application provides a radiotherapy dose prediction system based on CT images. The radiotherapy dose prediction system includes: a CT image acquisition unit, a CT frequency domain image acquisition unit, a tumor feature extraction unit, and a radiotherapy dose prediction unit.

[0068] The CT image acquisition unit is used to detect the patient's tumor target area and acquire CT images.

[0069] For example, the CT image may be generated by scanning a patient's tumor target area with a CT device.

[0070] The CT frequency domain image acquisition unit receives the CT image, performs frequency domain transformation on the CT image, and generates a CT frequency domain image.

[0071] In the image domain, CT images have complex and changeable features, especially early tumor images, whose image domain features are more easily ignored. Therefore, the CT frequency domain image acquisition unit performs frequency domain transformation on the CT image to generate a CT frequency domain image, which completely retains the boundary, tumor and background information of the CT image in different frequency components, making it easier for the subsequent tumor feature extraction unit to extract and identify complex image features.

[0072] Exemplarily, the frequency domain transform may adopt a two-dimensional Fourier transform or a wavelet transform.

[0073] The tumor feature extraction unit receives the CT image and the CT frequency domain image, extracts the image domain features of the CT image and the frequency domain features of the CT frequency domain image, performs feature fusion, and generates a tumor feature vector.

[0074] The radiotherapy dose prediction unit receives the tumor feature vector, identifies the tumor feature vector, and generates a radiotherapy dose prediction result.

[0075] Exemplarily, the radiotherapy dose prediction unit can identify the stage, size, and location of the tumor according to the tumor feature vector, and use the Softmax function to generate a corresponding radiotherapy dose prediction result.

[0076] In this embodiment of the present application, the radiotherapy dose prediction system comprehensively considers the image domain features and frequency domain features of the CT image, extracts the image domain features and frequency domain features respectively, and adds and concatenates them, which can improve the recognition accuracy of tumor features in CT images, accurately predict the radiotherapy dose, and improve the therapeutic effect of radiotherapy.

[0077] Second embodiment

[0078] In another embodiment, the present application also provides a tumor feature extraction unit for a radiotherapy dose prediction system, such as Figure 2 As shown, the tumor feature extraction unit specifically includes:

[0079] A first feature extraction module receives the CT image, performs image domain convolution and pooling operations on the CT image, and generates a first image domain feature map;

[0080] A first frequency domain feature extraction module receives the CT frequency domain image, performs frequency domain convolution and pooling operations on the CT image, and generates a first frequency domain feature map;

[0081] A first high-pass filtering module receives the first frequency domain feature map, performs high-pass filtering on the first frequency domain feature map, and obtains a first tumor detail feature map;

[0082] Exemplarily, the first high-pass filtering module may adopt two-dimensional Gaussian filtering.

[0083] In the first frequency domain feature map, the tumor detail feature map is decomposed into high-frequency components, and the background noise can be removed by the first high-pass filtering module to obtain the tumor detail feature map;

[0084] A first addition module receives the first image domain feature map and the first tumor detail feature map, performs element-by-element addition and fusion, and generates a first fusion feature map to supplement and strengthen the tumor detail feature;

[0085] In the image domain, due to the convolution and pooling operations performed by the feature extraction module, the detailed features of the tumor will gradually be lost, resulting in insufficient resolution and accuracy of the extracted tumor features. Therefore, the detailed features of the tumor in the frequency domain are used to supplement and strengthen the information features in the image domain. The elements are added and fused element by element to supplement and strengthen the detailed features of the tumor in the image domain, making it easier to extract complex image features in the image domain.

[0086] A second feature extraction module receives the first fused feature map, performs image domain convolution and pooling operations on the first fused feature map, and generates a second image domain feature map;

[0087] A second frequency domain feature extraction module receives the first frequency domain feature map, performs frequency domain convolution and pooling operations on the first frequency domain feature map, and generates a second frequency domain feature map;

[0088] A second high-pass filtering module receives the second frequency domain feature map, performs high-pass filtering on the second frequency domain feature map, and obtains a second tumor detail feature map;

[0089] Exemplarily, the second high-pass filtering module may adopt two-dimensional Gaussian filtering.

[0090] A second addition module receives the second image domain feature map and the second tumor detail feature map, performs element-by-element addition and fusion, and generates a second fused feature map to supplement and strengthen the tumor detail feature;

[0091] A third feature extraction module receives the second fused feature map, performs image domain convolution and pooling operations on the second fused feature map, and generates a third image domain feature map;

[0092] The output of the third feature extraction module can be expressed as:

[0093] Output I =P3(C3(P2(C2(P1(C1(img in ))+F1(temp1)))+F2(temp2)))

[0094] temp1=P4(C4(FET(img in )))

[0095] temp2=P5(C5(P4(C4(FET(img in )))))

[0096] Among them, Output Irepresents the output of the third feature extraction module, C1(·), C2(·), and C3(·) represent the convolution operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, respectively, P1(·), P2(·), and P3(·) represent the pooling operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, respectively, img in represents a CT image, F1(·) and F2(·) represent the filtering operations of the first high-pass filtering module and the second high-pass filtering module respectively, temp1 represents the output of the first frequency domain feature extraction module, and temp2 represents the output of the second frequency domain feature extraction module; C4(·) and C5(·) represent the convolution operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module respectively, P4(·) and P5(·) represent the pooling operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module respectively, and FET represents the frequency domain transformation operation;

[0097] A third frequency domain feature extraction module receives the second frequency domain feature map, performs frequency domain convolution and pooling operations on the second frequency domain feature map, and generates a third frequency domain feature map;

[0098] The output of the third frequency domain feature extraction module can be expressed as:

[0099] Output F =P6(C6(P5(C5(P4(C4(FET(img in )))))))

[0100] Among them, Output F represents the output of the third frequency domain feature extraction module, C6(·) represents the convolution operation of the third frequency domain feature extraction module, and P6(·) represents the pooling operation of the third frequency domain feature extraction module;

[0101] A splicing and fusion module receives the third image domain feature map and the third frequency domain feature map, performs feature splicing and fusion, and generates a splicing feature map;

[0102] The output of the splicing and fusion module can be expressed as:

[0103] Output C =Cat(Output I ,Output F )

[0104] Among them, Output C represents the output of the splicing and fusion module, and Cat(·) represents the splicing operation;

[0105] The fully connected layer receives the concatenated feature map, reduces the dimension of the concatenated feature map, and converts it into a one-dimensional tumor feature vector.

[0106] Furthermore, the structures of the first feature extraction module, the first frequency domain feature extraction module, the second feature extraction module, the second frequency domain feature extraction module, the third feature extraction module and the third frequency domain feature extraction module are the same, such as Figure 3 As shown in Figure 1, each feature extraction module is connected by convolutional layers and pooling layers connected in series.

[0107] Exemplarily, the convolution kernel size of the convolution layer and the pooling window of the pooling layer can be set comprehensively according to the feature extraction speed and the feature extraction resolution. For example, the convolution kernel size is set to 3×3 and the pooling window size is set to 2×2.

[0108] Furthermore, the convolution kernel, pooling window, and pooling step size of the first feature extraction module and the first frequency domain feature extraction module are the same, and the convolution kernel, pooling window, and pooling step size of the second feature extraction module and the second frequency domain feature extraction module are the same, so that the first image domain feature map and the first frequency domain feature map have the same size, and the second image domain feature map and the second frequency domain feature map have the same size, and thus addition operations can be performed without adding additional pooling layers and convolution layer size transformations, thereby reducing the complexity of the tumor feature extraction unit and improving processing efficiency.

[0109] In this embodiment of the present application, the tumor feature extraction unit utilizes two branches to extract image domain features and frequency domain features of the CT image, and at the same time adds the image domain information features through the tumor detail features in the frequency domain, thereby compensating for the problem of loss of image domain detail features, strengthening the tumor detail features in the image domain, and facilitating easier extraction of complex image features in the image domain, further enhancing the recognition accuracy of tumor features in CT images, and improving the accuracy of radiotherapy dose prediction results.

[0110] Third embodiment

[0111] In another embodiment, the present application also provides a pooling layer for a radiation therapy dose prediction system, such as Figure 4 As shown, the pooling layer specifically includes:

[0112] The pooling window setting module sets the size and step size of the pooling window and divides the output feature map of the convolutional layer into multiple sub-regions;

[0113] like Figure 5 As shown, exemplarily, the size of the pooling window is set to 2×2 and the step size is set to 2;

[0114] The row feature change calculation module calculates the row feature change in the generated sub-region, specifically:

[0115]

[0116] Among them, P Hrepresents the change in row features in the sub-region, N represents the size of the pooling window, and x i,j represents the feature of the i-th row and j-th column in the subregion, x i,j+1 Represents the features of the i-th row and j+1-th column in the subregion;

[0117] When the size of the pooling window is set to 2×2, the change in row features in the sub-region can be expressed as:

[0118]

[0119] The column feature change calculation module calculates the column feature change in the generated sub-region, specifically:

[0120]

[0121] Among them, P V represents the change in column features in the sub-region, x i+1,j Represents the feature of the i+1th row and jth column in the subregion;

[0122] The total feature change calculation module calculates the total feature change in the sub-region according to the row feature change and column feature change in the sub-region, specifically:

[0123]

[0124] Among them, P s represents the total feature change variation in the sub-region;

[0125] The feature map combination module generates a pooled feature map based on the total feature change in each sub-region.

[0126] The pooling layer in the prior art uses the maximum pooling method or the average pooling method, which can easily cause the loss of tumor detail feature information while reducing the spatial dimension of the feature map. However, in this embodiment of the present application, the pooling layer uses the change amount of adjacent features of rows and columns for pooling, retaining the degree of change of tumor detail feature information, and at the same time, through two iterations of rows and columns, avoiding the loss of tumor detail feature information, greatly improving the recognition accuracy of tumor features in CT images, and improving the accuracy of radiotherapy dose prediction results.

[0127] The above description shows and describes several preferred embodiments of the present application, but as mentioned above, it should be understood that the present application is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present application, and should be within the scope of protection of the claims attached to the present application.

Claims

1. A radiotherapy dose prediction system based on CT images, characterized in that: It includes a CT image acquisition unit, a CT frequency domain image acquisition unit, a tumor feature extraction unit, and a radiotherapy dose prediction unit; A CT image acquisition unit, used for detecting a patient's tumor target area and acquiring a CT image; A CT frequency domain image acquisition unit receives the CT image, performs frequency domain transformation on the CT image, and generates a CT frequency domain image; A tumor feature extraction unit receives the CT image and the CT frequency domain image, extracts the image domain features of the CT image and the frequency domain features of the CT frequency domain image, and performs feature fusion to generate a tumor feature vector; A radiotherapy dose prediction unit receives a tumor feature vector, identifies the tumor feature vector, and generates a radiotherapy dose prediction result; The tumor feature extraction unit specifically includes: A first feature extraction module receives the CT image, performs image domain convolution and pooling operations on the CT image, and generates a first image domain feature map; A first frequency domain feature extraction module receives the CT frequency domain image, performs frequency domain convolution and pooling operations on the CT image, and generates a first frequency domain feature map; A first high-pass filtering module receives the first frequency domain feature map, performs high-pass filtering on the first frequency domain feature map, and obtains a first tumor detail feature map; A first addition module receives the first image domain feature map and the first tumor detail feature map, performs element-by-element addition and fusion, and generates a first fused feature map; A second feature extraction module receives the first fused feature map, performs image domain convolution and pooling operations on the first fused feature map, and generates a second image domain feature map; A second frequency domain feature extraction module receives the first frequency domain feature map, performs frequency domain convolution and pooling operations on the first frequency domain feature map, and generates a second frequency domain feature map; A second high-pass filtering module receives the second frequency domain feature map, performs high-pass filtering on the second frequency domain feature map, and obtains a second tumor detail feature map; A second adding module receives the second image domain feature map and the second tumor detail feature map, performs element-by-element addition and fusion, and generates a second fused feature map; A third feature extraction module receives the second fused feature map, performs image domain convolution and pooling operations on the second fused feature map, and generates a third image domain feature map; A third frequency domain feature extraction module receives the second frequency domain feature map, performs frequency domain convolution and pooling operations on the second frequency domain feature map, and generates a third frequency domain feature map; A splicing and fusion module receives the third image domain feature map and the third frequency domain feature map, performs feature splicing and fusion, and generates a splicing feature map; The fully connected layer receives the concatenated feature map, reduces the dimension of the concatenated feature map, and converts it into a one-dimensional tumor feature vector; The output of the third feature extraction module is expressed as: in, represents the output of the third feature extraction module, Respectively represent the convolution operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, Respectively represent the pooling operations of the first feature extraction module, the second feature extraction module, and the third feature extraction module, represents a CT image, Respectively represent the filtering operations of the first high-pass filtering module and the second high-pass filtering module, represents the output of the first frequency domain feature extraction module, represents the output of the second frequency domain feature extraction module; Respectively represent the convolution operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module, They respectively represent the pooling operations of the first frequency domain feature extraction module and the second frequency domain feature extraction module, and FET represents the frequency domain transformation operation.

2. The radiation therapy dose prediction system according to claim 1, characterized in that: The output of the third frequency domain feature extraction module is expressed as: in, represents the output of the third frequency domain feature extraction module, represents the convolution operation of the third frequency domain feature extraction module, Represents the pooling operation of the third frequency domain feature extraction module.

3. The radiation therapy dose prediction system according to claim 2, characterized in that: The output of the splicing and fusion module is expressed as: in, represents the output of the splicing and fusion module, Represents a concatenation operation.

4. The radiation therapy dose prediction system according to claim 1, characterized in that: The first feature extraction module, the first frequency domain feature extraction module, the second feature extraction module, the second frequency domain feature extraction module, the third feature extraction module and the third frequency domain feature extraction module have the same structure and are connected by convolutional layers and pooling layers connected in series.

5. The radiation therapy dose prediction system according to claim 4, characterized in that: The first feature extraction module and the first frequency domain feature extraction module have the same convolution kernel, pooling window, and pooling step size; the second feature extraction module and the second frequency domain feature extraction module have the same convolution kernel, pooling window, and pooling step size.

6. The radiotherapy dose prediction system according to claim 4 or 5, characterized in that: The pooling layer comprises: The pooling window setting module sets the size and step size of the pooling window and divides the output feature map of the convolutional layer into multiple sub-regions; The row feature change calculation module calculates the row feature change in the generated sub-region, specifically: in, represents the change in row features in the sub-region, N represents the size of the pooling window, represents the feature of the i-th row and j-th column in the subregion, Represents the features of the i-th row and j+1-th column in the subregion; The column feature change calculation module calculates the column feature change in the generated sub-region, specifically: in, represents the change in column features in the sub-region, Represents the feature of the i+1th row and jth column in the subregion; The total feature change calculation module calculates the total feature change in the sub-region according to the row feature change and column feature change in the sub-region, specifically: in, Represents the total feature change in the sub-region; The feature map combination module combines and generates pooled feature maps based on the total feature changes in each sub-region.

7. The radiation therapy dose prediction system according to claim 6, characterized in that: The convolution kernel size in the convolution layer is set to 3×3, and the pooling window size in the pooling layer is set to 2×2.

8. The radiation therapy dose prediction system according to claim 1, characterized in that: The frequency domain transformation adopts two-dimensional Fourier transformation or wavelet transformation.