Two-dimensional ionization chamber detection matrix super-resolution reconstruction method and system
The dose image super-resolution reconstruction network model trained by deep learning technology solves the problem of low resolution of the two-dimensional ionization chamber detection matrix, achieves high-precision dose image reconstruction, and improves the quality assurance of radiotherapy plans.
Patent Information
- Application Number
- CN202510778072.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-11
- Publication Date
- 2025-09-23
AI Technical Summary
The existing two-dimensional ionization chamber detection matrix has low resolution and cannot accurately capture complex radiation dose distributions. Especially at the edges and in areas with large dose gradients, traditional super-resolution reconstruction methods have dose errors.
Deep learning technology is used to train the dose image super-resolution reconstruction network model. The local dose microstructure extraction module, the global prior dose pattern extraction module and the dose partition feature extraction module are used. In combination with the attention mechanism and adaptive fusion technology, high-resolution dose images are generated, and the model is optimized through the dynamic weighted L1 loss function.
It improves the reconstruction accuracy and reliability of dose images, increases the Gamma pass rate, meets the quality verification standards of radiotherapy plans, and provides a more reliable means of dose verification.
Smart Images
Figure CN120689207A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of radiotherapy radiation dose measurement, and in particular to a two-dimensional ionization chamber detection matrix super-resolution reconstruction method and system. Background Art
[0002] In radiotherapy, a two-dimensional ionization chamber detector matrix is used to measure radiation dose distribution, and its resolution directly affects the accuracy of the dose distribution. However, due to hardware limitations, the resolution of the detector matrix is often low, making it unable to accurately capture complex dose distributions, especially at edges and in areas with large dose gradients. Traditional super-resolution reconstruction methods rely primarily on algorithms (such as bilinear interpolation and bicubic interpolation), which can lead to dose errors when applied to dose images. Summary of the Invention
[0003] The purpose of this application is to provide a two-dimensional ionization chamber detection matrix super-resolution reconstruction method and system, which can improve the accuracy and reliability of reconstructed dose images.
[0004] This application is implemented as follows:
[0005] In a first aspect, the present application provides a two-dimensional ionization chamber detection matrix super-resolution reconstruction method, comprising:
[0006] Paired low-resolution and high-resolution dose images were obtained from the treatment planning system TPS as training datasets;
[0007] Using the training data set to train a preset dose image super-resolution reconstruction network model, so that the dose image super-resolution reconstruction network model has the ability to reconstruct a high-resolution dose image from a low-resolution dose image;
[0008] The measured low-resolution dose image of the two-dimensional ionization chamber matrix is input into the trained dose image super-resolution reconstruction network model to generate the predicted high-resolution dose image of the ionization chamber matrix;
[0009] The predicted high-resolution dose image of the ionization chamber matrix is compared with the high-resolution image in the treatment planning system TPS for gamma pass rate consistency. If the pass rate index is met, the treatment is verified to be passed.
[0010] Based on the first aspect, after the step of obtaining paired low-resolution and high-resolution dose images from the treatment planning system TPS as a training dataset, the following steps are further included:
[0011] The low-resolution and high-resolution dose images are normalized, and the normalization formula is:
[0012] I norm =[I-min(LR min ,HRmin )] / [max(LR max ,HR max )-min(LR min ,HR min )];
[0013] Among them I norm is the normalized dose value, I is the original dose value, (LR min ,HR min ),(LR max ,HR max ) represent the minimum dose and maximum dose of the low-resolution dose image and the high-resolution dose image, respectively;
[0014] The normalized low-resolution and high-resolution dose images were randomly rotated and flipped to enhance the diversity of data samples.
[0015] Based on the first aspect, the dose image super-resolution reconstruction network model sequentially includes a first convolutional layer, multiple feature enhancement networks, a second convolutional layer and upsampling;
[0016] Among them, any feature enhancement network includes multiple attention enhancement networks and a convolutional layer to extract multi-scale dose features from the input of each layer;
[0017] Any attention-enhancing network includes:
[0018] Local dose microstructure extraction module: uses a window attention mechanism to capture the subtle structural features of the dose distribution;
[0019] Global prior dose pattern extraction module: learns a global dose distribution pattern with physical constraints through a cross-attention mechanism;
[0020] Dose partition feature extraction module: realizes feature grouping based on dose value similarity;
[0021] The features output by the local dose microstructure extraction module, the global prior dose pattern extraction module and the dose partition feature extraction module are adaptively fused to form multi-scale dose features, which are then gradually formed into high-level dose distribution features through a cascade network. Finally, after upsampling processing, the output is a dose distribution map that maintains physical properties and has high spatial resolution.
[0022] Based on the first aspect, the local dose microstructure extraction module specifically includes:
[0023] The window segmentation mechanism and multi-head self-attention are used to dynamically focus on the dose gradient mutation characteristics in a limited area. The formula is expressed as:
[0024]
[0025] Among them, the query matrix Q is used to encode the dose gradient characteristics of the current spatial position, the key matrix K is used to store the dose value distribution pattern of adjacent positions in the window, the value matrix V is used to carry the physical value of the dose microstructure, and the scaling factor It is used to stabilize the gradient calculation, the bias term B is used for trainable window position encoding, and Softmax is a probability distribution. These are the features output by the local dose microstructure module.
[0026] Based on the first aspect, the global prior dose pattern extraction module specifically includes:
[0027] By using the updateable physical dose distribution law, a priori knowledge token is constructed to supplement the anatomical structure-related dose distribution pattern for the low-resolution input, which is expressed as follows:
[0028] Q X =XW Q ,K D =DW K ,V D =DW V ;
[0029] A=SoftMax(Sim cos (Q X ,K D ) / τ);
[0030]
[0031] Among them, X is the dose feature output by the previous layer, D is the token dictionary of the physical dose distribution law, and W Q 、W K 、W V are the linear transformations of the query matrix, key matrix, and value matrix respectively; Q X , K D 、V D are the generated query tokens, key tokens, and value tokens respectively; τ is a learnable parameter that adjusts the range of similarity values, Sim cos Indicates the calculation of the cosine similarity between two tokens. The SoftMax function converts the feature similarity map S into an attention map A. Features output by the cross-attention module for the token dictionary.
[0032] Based on the first aspect, the dose partition feature extraction module specifically includes:
[0033] By using the dose value similarity, the features are classified into low-dose area, dose drop area and high-dose area to achieve targeted learning of features in different areas. The formula is expressed as:
[0034] θ i (x,A)={xj |argmax k (A jk )=i};
[0035]
[0036] Among them, A is the attention map obtained by the global prior dose pattern extraction module. If j1 , A j2 ,...,A jM A ji is the largest, pixels x j is classified as θ i ; MSA is multi-head self-attention, It is the feature output by the dose partition feature extraction module;
[0037] Set the dynamic weighted L1 loss function for each region: Let the predicted actual high-resolution dose image D pred and high-resolution images D in the treatment planning system TPS gt The size of the H×W is divided into three areas according to the dose value range:
[0038] Low-dose region (Ω l ):D gt (i,j)<0.1D max ;
[0039] Mid-dose region (Ω m ): 0.1D max <D gt (i,j)≤0.8D max ;
[0040] High-dose region (Ω h ):D gt (i,j)>0.8D max ;
[0041] Among them, D max is the global maximum value of the dose map;
[0042] The L1 loss function of each region is expressed as:
[0043]
[0044] Initialize the weights of the low-dose area, dose drop area, and high-dose area as: W l =W m =W h =1; Set the dynamic adjustment rule. In the tth epoch of training, if K consecutive epochs meet That is, the loss decreases and stagnates, then the adjusted weight is expressed as:
[0045] W m =W m +α*Δw;
[0046] W h =W h +β*Δw(β>α>0);
[0047] Among them, Δw is the basic adjustment step size, α and β are the adjustment coefficients of the dose drop area and high dose area respectively; the dynamic weighted L1 loss function is expressed as:
[0048] L dynamic =W l *L l +W m *L m +W h *L h ;
[0049] Among them, W l 、W m 、W h Represent the weights of low-dose area, dose drop area and high-dose area respectively, L l , L m , L h They represent the losses in the low-dose area, dose drop area, and high-dose area, respectively.
[0050] Based on the first aspect, it also includes:
[0051] Δw is set to 0.1 to avoid excessive weight changes that may cause model oscillation. The low-dose area is usually set at the outer edge of the target area, where the dose level is extremely low and the model is easy to fit. The dose drop area is set at the edge of the target area and contains most of the dose drop information. The high-dose area is set at the core area of the dose, involving the maximum dose. l Set the weight to remain unchanged to ensure the basic fitting of the low-dose area and give the dose drop area L m and high dose area L h Set a growing weight to make the model pay more attention to the dose distribution in these two areas, the dose drop area W m and high dose area W h The weight range is between 1 and 3 and can be adjusted dynamically during training.
[0052] In a second aspect, the present application provides a radiotherapy ionization chamber matrix super-resolution reconstruction system based on deep learning, comprising:
[0053] A data acquisition module is used to obtain paired low-resolution and high-resolution dose images from the treatment planning system TPS as a training dataset;
[0054] A model training module is used to train a preset dose image super-resolution reconstruction network model using a training data set, so that the dose image super-resolution reconstruction network model has the ability to reconstruct a high-resolution dose image from a low-resolution dose image;
[0055] A high-resolution image generation module is used to input the measured low-resolution dose image of the two-dimensional ionization chamber matrix into the trained dose image super-resolution reconstruction network model to generate a predicted high-resolution dose image of the ionization chamber matrix;
[0056] The pass rate verification module is used to compare the predicted ionization chamber matrix high-resolution dose image with the high-resolution image in the treatment planning system TPS for Gamma pass rate consistency. If the pass rate index is met, the verification is passed.
[0057] In a third aspect, the present application provides an electronic device, comprising:
[0058] a memory for storing one or more programs;
[0059] processor;
[0060] When the one or more programs are executed by the processor, the above method is implemented.
[0061] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned method when executed by a processor.
[0062] Compared with the prior art, this application has at least the following advantages or beneficial effects:
[0063] This application provides a method and system for super-resolution reconstruction of a two-dimensional ionization chamber detection matrix. The system creates a dose image super-resolution reconstruction network model capable of reconstructing a high-resolution dose image from a low-resolution dose image. The measured low-resolution dose image of the two-dimensional ionization chamber matrix is input into a trained dose image super-resolution reconstruction network model to generate a predicted high-resolution dose image of the ionization chamber matrix. The predicted high-resolution dose image of the ionization chamber matrix is compared with the high-resolution image in the treatment planning system (TPS) for gamma pass rate consistency. If the pass rate index is met, the system is verified as passed. Deep learning technology is used to fully learn the physical property mapping relationship between low-resolution and high-resolution dose images, improving the accuracy and reliability of the reconstructed dose image. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.
[0065] Figure 1 This is a flowchart of a two-dimensional ionization chamber detection matrix super-resolution reconstruction method for this application;
[0066] Figure 2A This is a schematic diagram of a dose image super-resolution reconstruction network model in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method of this application;
[0067] Figure 2B This is a schematic diagram of a feature enhancement network framework in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method in this application;
[0068] Figure 2C This is a schematic diagram of the attention enhancement network framework in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method in this application;
[0069] Figure 3 This is a schematic diagram of a local dose microstructure extraction module in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method of this application;
[0070] Figure 4 This is a schematic diagram of a global prior dose pattern extraction module in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method in this application;
[0071] Figure 5 This is a schematic diagram of a dose partition feature extraction module in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method of this application;
[0072] Figure 6 This is a comparison of the effects of the high-resolution dose image predicted by the method of the present application and the traditional linear interpolation algorithm in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method of the present application;
[0073] Figure 7 This is a schematic structural diagram of a two-dimensional ionization chamber detection matrix super-resolution reconstruction system for this application;
[0074] Figure 8 This is a structural diagram of an electronic device of the present application.
[0075] icon:
[0076] 1. Data acquisition module; 2. Model training module; 3. High-resolution image generation module; 4. Pass rate verification module; 5. Processor; 6. Memory; 7. Communication interface. DETAILED DESCRIPTION
[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Generally, the components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.
[0078] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features thereof may be combined with each other.
[0079] Example
[0080] The embodiments of the present application provide a two-dimensional ionization chamber detection matrix super-resolution reconstruction method and system, which can improve the accuracy and reliability of reconstructed dose images.
[0081] Please refer to Figure 1 , Figure 1 This is a flow chart of a two-dimensional ionization chamber detection matrix super-resolution reconstruction method for this application; the method includes the following steps:
[0082] S1: Paired low-resolution and high-resolution dose images are obtained from the treatment planning system TPS as a training dataset;
[0083] Specifically, low-resolution dose images are rapidly generated by TPS using a sparse computational grid (e.g., 5-10mm voxels), sacrificing detail to preserve overall dose trends. High-resolution dose images, on a fine grid (e.g., 1mm-2.5mm voxels), are generated by TPS's precise dose calculations and include microscopic dose gradient details. Furthermore, low-resolution dose images are dose images consistent with the resolution of a two-dimensional ionization chamber matrix detector, while high-resolution dose images strictly adhere to clinical radiotherapy planning specifications (e.g., the ≤2.5mm grid accuracy requirement recommended by AAPM TG-53).
[0084] Furthermore, the resolution of the low-resolution dose images used for training is consistent with the actual probe resolution of the two-dimensional ionization chamber matrix, and the resolution of the high-resolution dose images is not lower than the computational grid accuracy required for treatment planning.
[0085] Furthermore, to ensure the absolute dose consistency of the dose images, the training data is normalized, requiring the low-resolution dose images and high-resolution dose images to have the same normalization range. The normalization formula takes the low-resolution and high-resolution dose images for joint normalization. The normalization formula is:
[0086] I norm =[I-min(LR min ,HR min )] / [max(LR max ,HR max )-min(LR min ,HR min )];
[0087] Among them I norm is the normalized dose value, I is the original dose value, (LR min ,HR min ),(LR max ,HR max ) represent the minimum dose and maximum dose of the low-resolution dose image and the high-resolution dose image, respectively;
[0088] The normalized low-resolution and high-resolution dose images were randomly rotated and flipped to enhance the diversity of data samples.
[0089] S2: Using the training dataset, a pre-set dose image super-resolution reconstruction network model is trained to enable the dose image super-resolution reconstruction network model to have the ability to reconstruct a high-resolution dose image from a low-resolution dose image; Figure 2A This is a schematic diagram of a dose image super-resolution reconstruction network model in a two-dimensional ionization chamber detection matrix super-resolution reconstruction method of this application;
[0090] Please refer to Figure 2A , Figure 2A Schematic diagram of a dose image super-resolution reconstruction network model; specifically, the dose image super-resolution reconstruction network model includes a first convolutional layer, multiple feature enhancement networks, a second convolutional layer, and upsampling in sequence;
[0091] Among them, any feature enhancement network includes multiple attention enhancement networks and a convolutional layer to extract multi-scale dose features from the input of each layer;
[0092] Any attention-enhancing network includes:
[0093] Local dose microstructure extraction module: uses a window attention mechanism to capture the subtle structural features of the dose distribution;
[0094] Global prior dose pattern extraction module: learns a global dose distribution pattern with physical constraints through a cross-attention mechanism;
[0095] Dose partition feature extraction module: realizes feature grouping based on dose value similarity;
[0096] The features output by the local dose microstructure extraction module, the global prior dose pattern extraction module and the dose partition feature extraction module are adaptively fused to form multi-scale dose features, which are then gradually formed into high-level dose distribution features through a cascade network. Finally, after upsampling processing, the output is a dose distribution map that maintains physical properties and has high spatial resolution.
[0097] The model has an asymmetric encoder-decoder architecture. The encoder is formed by multiple feature enhancement networks in series, and the decoder consists of an upsampling module. For a schematic diagram of the feature enhancement network framework, please refer to Figure 2B .
[0098] Furthermore, based on the aforementioned scheme, the encoder is used to map the low-resolution image to a latent representation, which is cascaded by multiple feature enhancement networks. Each feature enhancement network framework contains four attention enhancement networks and one convolutional layer to extract multi-scale dose features from the input of each layer.
[0099] Please refer to Figure 2C , Figure 2C Schematic diagram of the attention enhancement network framework; the attention enhancement network is formed by the parallel combination of the local dose microstructure extraction module, the dose partition feature extraction module and the global prior dose pattern extraction module.
[0100] Please refer to Figure 3 , Figure 3 Schematic diagram of the local dose microstructure extraction module. Further, the local dose microstructure extraction module specifically includes:
[0101] The window segmentation mechanism and multi-head self-attention are used to dynamically focus on the dose gradient mutation characteristics in a limited area. The formula is expressed as:
[0102]
[0103] Among them, the query matrix Q is used to encode the dose gradient characteristics of the current spatial position, the key matrix K is used to store the dose value distribution pattern of adjacent positions in the window, the value matrix V is used to carry the physical value of the dose microstructure, and the scaling factor It is used to stabilize the gradient calculation, the bias term B is used for trainable window position encoding, and Softmax is a probability distribution. These are the features output by the local dose microstructure module.
[0104] Please refer to Figure 4 , Figure 4is a global a priori dose pattern extraction module; further, the global a priori dose pattern extraction module specifically includes:
[0105] By using the updateable physical dose distribution law, a priori knowledge token is constructed to supplement the anatomical structure-related dose distribution pattern for the low-resolution input, which is expressed as follows:
[0106] Q X =XW Q ,K D =DW K ,V D =DW V ;
[0107] A=SoftMax(Sim cos (Q X ,K D ) / τ);
[0108]
[0109] Among them, X is the dose feature output by the previous layer, D is the token dictionary of the physical dose distribution law, and W Q 、W K 、W V are the linear transformations of the query matrix, key matrix, and value matrix respectively; Q X , K D 、V D are the generated query tokens, key tokens, and value tokens respectively; τ is a learnable parameter that adjusts the range of similarity values, Sim cos Represents the calculation of the cosine similarity between two tokens. The SoftMax function converts the feature similarity map S into an attention map A, X out Features output by the cross-attention module for the token dictionary.
[0110] Please refer to Figure 5 , Figure 5 is a dose partition feature extraction module; further, the dose partition feature extraction module specifically includes:
[0111] By using the dose value similarity, the features are classified into low-dose area, dose drop area and high-dose area to achieve targeted learning of features in different areas. The formula is expressed as:
[0112] θ i (x,A)={x j |argmax k (A jk )=i};
[0113]
[0114] Among them, A is the attention map A obtained by the global prior dose pattern extraction module. If j1 , A j2 ,...,A jM A ji is the largest, pixels x j is classified as θ i ; MSA is multi-head self-attention, X out It is the feature output by the dose partition feature extraction module;
[0115] Set the dynamic weighted L1 loss function for each region: Let the predicted actual high-resolution dose image D pred and high-resolution images D in the treatment planning system TPS gt The size of the H×W is divided into three areas according to the dose value range:
[0116] Low-dose region (Ω l ):D gt (i,j)<0.1D max ;
[0117] Mid-dose region (Ω m ): 0.1D max <D gt (i,j)≤0.8D max ;
[0118] High-dose region (Ω h ):D gt (i,j)>0.8D max ;
[0119] Among them, D max is the global maximum value of the dose map;
[0120] The L1 loss function of each region is expressed as:
[0121]
[0122] Initialize the weights of the low-dose area, dose drop area, and high-dose area as: W l =W m =W h =1; Set the dynamic adjustment rule. In the tth epoch of training, if K consecutive epochs meet That is, the loss decreases and stagnates, then the adjusted weight is expressed as:
[0123] W m =W m +α*Δw;
[0124] W h =W h +β*Δw(β>α>0);
[0125] Among them, Δw is the basic adjustment step size, α and β are the adjustment coefficients of the dose drop area and high dose area respectively; the dynamic weighted L1 loss function is expressed as:
[0126] L dynamic =W l *L l +W m *L m +W h *L h ;
[0127] Among them, W l 、W m 、W h Represent the weights of low-dose area, dose drop area and high-dose area respectively, L l , L m , L h They represent the losses in the low-dose area, dose drop area, and high-dose area, respectively.
[0128] Preferably, Δw is set to 0.1 to avoid excessive weight changes that may cause model oscillations. The low-dose area is usually located at the outer edge of the target area, where the dose level is extremely low and the model is easy to fit. The dose drop area is set at the edge of the target area and contains most of the dose drop information. The high-dose area is set at the core area of the dose, involving the maximum dose. l Set the weight to remain unchanged to ensure the basic fitting of the low-dose area and give the dose drop area L m and high dose area L h Set a growing weight to make the model pay more attention to the dose distribution in these two areas, the dose drop area W m and high dose area W h The weight range is between 1 and 3 and can be adjusted dynamically during training.
[0129] This step innovatively achieves super-resolution reconstruction of dose images while ensuring the physical accuracy of the reconstruction results.
[0130] S3: Input the measured low-resolution dose image of the two-dimensional ionization chamber matrix into the trained dose image super-resolution reconstruction network model to generate the predicted high-resolution dose image of the ionization chamber matrix;
[0131] Specifically, through the trained dose image super-resolution reconstruction network model, the low-resolution dose image is mapped to a high-resolution dose image and then output, reconstructing the microscopic dose details lost due to hardware limitations.
[0132] S4: The predicted high-resolution dose image of the ionization chamber matrix is compared with the high-resolution image in the treatment planning system TPS for gamma pass rate consistency. If the pass rate index is met, the verification is passed.
[0133] Specifically, through verification, the dose image super-resolution reconstruction network model can compensate for the inherent resolution defects of physical detection equipment while retaining the TPS calculation accuracy, providing a reliable dose verification method for radiotherapy plan execution. Figure 6 , Figure 6 To compare the effects of the high-resolution dose image predicted by the method of this application and the traditional linear interpolation algorithm, it can be seen from the Profile dose curve comparison results that the profile dose curve (red line) interpolated by the method of this application is highly consistent with the true high-resolution dose image curve (green line). In contrast, the traditional linear interpolation results are significantly different in the dose drop area. The gamma pass rate (3% / 2mm) evaluation results have also been significantly improved, that is, the gamma pass rate has increased from 82.59% to 96.25%. The clinical verification standard is that it needs to be greater than 95% (according to the National Cancer Center's "Intensity Modulated Radiotherapy Dose Verification Practice Guidelines", the gamma pass rate standard must meet the 3% / 2mm gamma pass rate verification result must be greater than 95%). Therefore, the use of this method can improve the gamma pass rate of the intensity modulated radiotherapy plan verification of the two-dimensional ionization chamber matrix, and can provide more reliable quality assurance for radiotherapy.
[0134] Please refer to Figure 7 , Figure 7 This is a schematic diagram of the structure of a radiotherapy ionization chamber matrix super-resolution reconstruction system based on deep learning in this application; the system includes:
[0135] A data acquisition module 1 is used to acquire paired low-resolution and high-resolution dose images from a treatment planning system TPS as a training dataset;
[0136] Model training module 2, used to train a preset dose image super-resolution reconstruction network model using a training data set, so that the dose image super-resolution reconstruction network model has the ability to reconstruct a high-resolution dose image from a low-resolution dose image;
[0137] The high-resolution image generation module 3 is used to input the measured low-resolution dose image of the two-dimensional ionization chamber matrix into the trained dose image super-resolution reconstruction network model to generate a predicted high-resolution dose image of the ionization chamber matrix;
[0138] The pass rate verification module 4 is used to compare the predicted ionization chamber matrix high-resolution dose image with the high-resolution image in the treatment planning system TPS for Gamma pass rate consistency. If the pass rate index is met, the verification is passed.
[0139] Specifically, for the specific implementation of the radiotherapy ionization chamber matrix super-resolution reconstruction system based on deep learning, please refer to the specific implementation of a two-dimensional ionization chamber detection matrix super-resolution reconstruction method, which will not be elaborated here.
[0140] Please refer to Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device of the present application; the electronic device includes:
[0141] Memory 7, for storing one or more programs;
[0142] Processor 6; processor 6 is connected to memory 7 via a communication interface 8;
[0143] When the one or more programs are executed by the processor 6 , all or part of the above methods are implemented.
[0144] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements all or part of the above-mentioned method when executed by the processor 6.
[0145] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present application is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.
Claims
1. A two-dimensional ionization chamber detection matrix super-resolution reconstruction method, characterized in that: include: Paired low-resolution and high-resolution dose images were obtained from the treatment planning system TPS as training datasets; Using the training data set to train a preset dose image super-resolution reconstruction network model, so that the dose image super-resolution reconstruction network model has the ability to reconstruct a high-resolution dose image from a low-resolution dose image; The measured low-resolution dose image of the two-dimensional ionization chamber matrix is input into the trained dose image super-resolution reconstruction network model to generate the predicted high-resolution dose image of the ionization chamber matrix; The predicted high-resolution dose image of the ionization chamber matrix is compared with the high-resolution image in the treatment planning system TPS for gamma pass rate consistency. If the pass rate index is met, the treatment is verified to be passed.
2. A two-dimensional ionization chamber detection matrix super-resolution reconstruction method according to claim 1, characterized in that: After the step of obtaining paired low-resolution and high-resolution dose images from the treatment planning system TPS as a training data set, the method further includes: The low-resolution and high-resolution dose images are normalized, and the normalization formula is: I norm =[I-min(LR min ,HR min )] / [max(LR max ,HR max )-min(LR min ,HR min )]; Among them I norm is the normalized dose value, I is the original dose value, (LR min ,HR min ),(LR max ,HR max ) represent the minimum dose and maximum dose of the low-resolution dose image and the high-resolution dose image, respectively; The normalized low-resolution and high-resolution dose images were randomly rotated and flipped to enhance the diversity of data samples.
3. The two-dimensional ionization chamber detection matrix super-resolution reconstruction method according to claim 1, characterized in that: The dose image super-resolution reconstruction network model sequentially includes the first convolutional layer, multiple feature enhancement networks, the second convolutional layer and upsampling; Among them, any feature enhancement network includes multiple attention enhancement networks and a convolutional layer to extract multi-scale dose features from the input of each layer; Any attention-enhancing network includes: Local dose microstructure extraction module: uses a window attention mechanism to capture the subtle structural features of the dose distribution; Global prior dose pattern extraction module: learns a global dose distribution pattern with physical constraints through a cross-attention mechanism; Dose partition feature extraction module: realizes feature grouping based on dose value similarity; The features output by the local dose microstructure extraction module, the global prior dose pattern extraction module and the dose partition feature extraction module are adaptively fused to form multi-scale dose features, which are then gradually formed into high-level dose distribution features through a cascade network. Finally, after upsampling processing, the output is a dose distribution map that maintains physical properties and has high spatial resolution.
4. A two-dimensional ionization chamber detection matrix super-resolution reconstruction method according to claim 3, characterized in that: The local dose microstructure extraction module specifically includes: The window segmentation mechanism and multi-head self-attention are used to dynamically focus on the dose gradient mutation characteristics in a limited area. The formula is expressed as: Among them, the query matrix Q is used to encode the dose gradient characteristics of the current spatial position, the key matrix K is used to store the dose value distribution pattern of adjacent positions in the window, the value matrix V is used to carry the physical value of the dose microstructure, and the scaling factor It is used to stabilize the gradient calculation, the bias term B is used for trainable window position encoding, and Softmax is a probability distribution. These are the features output by the local dose microstructure module.
5. The method for super-resolution reconstruction of a two-dimensional ionization chamber detection matrix according to claim 3, characterized in that: The global prior dose pattern extraction module specifically includes: By using the updateable physical dose distribution law, a priori knowledge token is constructed to supplement the anatomical structure-related dose distribution pattern for the low-resolution input, which is expressed as follows: Q X =XW Q ,K D =DW K ,V D =DW V ; A=SoftMax(Sim cos (Q X ,K D ) / t); Among them, X is the dose feature output by the previous layer, D is the token dictionary of the physical dose distribution law, and W Q 、W K 、W V are the linear transformations of the query matrix, key matrix, and value matrix respectively; Q X , K D 、V D are the generated query tokens, key tokens, and value tokens respectively; τ is a learnable parameter that adjusts the range of similarity values, Sim cos Indicates the calculation of the cosine similarity between two tokens. The SoftMax function converts the feature similarity map S into an attention map A. Features output by the cross-attention module for the token dictionary.
6. The method for super-resolution reconstruction of a two-dimensional ionization chamber detection matrix according to claim 3, characterized in that: The dose partition feature extraction module specifically includes: By using the dose value similarity, the features are classified into low-dose area, dose drop area and high-dose area to achieve targeted learning of features in different areas. The formula is expressed as: i i (x,A)={x j |argmax k (A jk )=i}; Among them, A is the attention map obtained by the global prior dose pattern extraction module. If j1 , A j2 ,...,A jM A ji is the largest, pixels x j is classified as θ i ; MSA is multi-head self-attention, It is the feature output by the dose partition feature extraction module; Set the dynamic weighted L1 loss function for each region: Let the predicted actual high-resolution dose image D pred and high-resolution images D in the treatment planning system TPS gt The size of the H×W is divided into three areas according to the dose value range: Low-dose region (Ω l ):D gt (i,j)<0.1D max ; Mid-dose region (Ω m ): 0.1D max <D gt (i,j)≤0.8D max ; High-dose region (Ω h ):D gt (i,j)>0.8D max ; Among them, D max is the global maximum value of the dose map; The L1 loss function of each region is expressed as: Initialize the weights of the low-dose area, dose drop area, and high-dose area as: W l =W m =W h =1; Set the dynamic adjustment rule. In the tth epoch of training, if K consecutive epochs meet That is, the loss decreases and stagnates, then the adjusted weight is expressed as: IN m =In m +α*Δw; W h =W h +β*Δw(β>α>0); Among them, Δw is the basic adjustment step size, α and β are the adjustment coefficients of the dose drop area and high dose area respectively; the dynamic weighted L1 loss function is expressed as: L dynamic =W l *L l +W m *L m +W h *L h ; Among them, W l 、W m 、W h Represent the weights of low-dose area, dose drop area and high-dose area respectively, L l , L m , L h They represent the losses in the low-dose area, dose drop area, and high-dose area, respectively.
7. The method for super-resolution reconstruction of a two-dimensional ionization chamber detection matrix according to claim 6, characterized in that: Also includes: Δw is set to 0.1 to avoid excessive weight changes that may cause model oscillation; The low-dose area is usually set at the outer edge of the target area, where the dose level is extremely low and the model is easy to fit; the dose drop area is set at the edge of the target area and contains most of the dose drop information; the high-dose area is set at the core area of the dose and involves the maximum dose; l Set the weight to remain unchanged to ensure the basic fitting of the low-dose area and give the dose drop area L m and high dose area L h Set a growing weight to make the model pay more attention to the dose distribution in these two areas, the dose drop area W m and high dose area W h The weight range is between 1 and 3 and can be adjusted dynamically during training.
8. A radiotherapy ionization chamber matrix super-resolution reconstruction system based on deep learning, characterized by: include: A data acquisition module is used to obtain paired low-resolution and high-resolution dose images from the treatment planning system TPS as a training dataset; A model training module is used to train a preset dose image super-resolution reconstruction network model using a training data set, so that the dose image super-resolution reconstruction network model has the ability to reconstruct a high-resolution dose image from a low-resolution dose image; A high-resolution image generation module is used to input the measured low-resolution dose image of the two-dimensional ionization chamber matrix into the trained dose image super-resolution reconstruction network model to generate a predicted high-resolution dose image of the ionization chamber matrix; The pass rate verification module is used to compare the predicted ionization chamber matrix high-resolution dose image with the high-resolution image in the treatment planning system TPS for Gamma pass rate consistency. If the pass rate index is met, the verification is passed.
9. An electronic device, characterized in that: include: a memory for storing one or more programs; processor; When the one or more programs are executed by the processor, the method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.