A quality control method based on an electronic portal imaging device and the maximum tissue phantom ratio

Abstract

The present invention belongs to the technical field of radiotherapy quality control using an electronic portal imaging device, and discloses a quality control method based on an electronic portal imaging device and the maximum tissue phantom ratio. The digital reconstructed radiograph (DRR) image is improved by combining the maximum tissue phantom ratio (TMR) of the radiotherapy device to obtain a TMR-weighted DRR image; the TMR-weighted DRR image and the fluence map generated by the radiotherapy plan are used as data inputs to train an improved U-net network, and then a predicted EPID map is obtained; the predicted EPID map and the EPID map actually generated by the radiotherapy device are analyzed based on the gamma index, and the quality control of the radiotherapy device is carried out according to the gamma index. The present invention uses the maximum tissue phantom ratio data collected during the operation test of the device to convert the three-dimensional CT data into two-dimensional data, and the data needs to be collected additionally. Using an improved U-net network, a relatively high precision can be found after gamma analysis, and the positioning error during verification can be easily found and discriminated.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy quality control using an electronic portal imaging device, and particularly to a quality control method based on an electronic portal imaging device and the maximum tissue phantom ratio. Background Art

[0002] EPID (Electronic Portal Imaging Device), namely an electronic portal imaging system, is an important imaging device in the field of modern radiotherapy. It uses a flat panel detector to receive and record the image of X-rays passing through the patient, thereby realizing the monitoring and position verification of the radiotherapy process. By applying appropriate quality control methods, multiple performance parameters of radiotherapy equipment can be detected using EPID, such as output dose, flatness and symmetry of the radiation field, consistency between the light field and the radiation field, and the in-place accuracy of the multi-leaf collimator, etc. These parameters are crucial for ensuring the accuracy and stability of the operation of radiotherapy equipment. By regularly using EPID for equipment quality control, problems existing in the equipment can be detected and corrected in a timely manner, thereby ensuring the smooth progress of the radiotherapy process. EPID also has the function of real-time monitoring of the radiation dose. During the radiotherapy process, EPID can monitor the change of the radiation dose in real time, thereby helping doctors to detect and correct errors in the radiotherapy process in a timely manner. This helps to improve the accuracy and safety of radiotherapy and reduce the side effects and complications caused by radiotherapy.

[0003] The core of EPID quality control lies in evaluating the accuracy and consistency of the radiotherapy process by comparing the difference between the actual EPID image and the expected or reference image. There are two existing methods for accurately predicting the transmission-type EPID response image based on treatment plan data. One is the simulation method. According to the principles of radiation physics, a physical model between the accelerator, the patient, and the EPID is established. This model should be able to simulate the transmission, attenuation, and scattering processes of the rays in the patient's body, as well as the response of the EPID to the rays; the second is to use machine learning algorithms. The treatment plan data and the corresponding EPID response images are used to train the grid model so that it can learn the mapping relationship between the data.

[0004] The response of EPID shows a linear characteristic with the beam dose, but due to the change of human tissue density, organ movement and volume change, it increases the difficulty of prediction in actual work. At the same time, the image quality of the transmission-type EPID is affected by factors such as ray energy, scattering, background noise, etc., resulting in a low accuracy of the simulation algorithm; the training data of the existing machine learning methods usually adopt treatment plan data and the corresponding EPID response data. A large amount of training data is required to establish a prediction model. However, in actual applications, it is relatively difficult to obtain sufficient training data to support the training of the model. Summary of the Invention

[0005] The present invention proposes a quality control method based on an electronic portal imaging device (EPID) and the maximum tissue phantom ratio, aiming to accurately predict the images of the EPID to evaluate the performance of radiotherapy equipment and the execution effect of radiotherapy plans. This method combines the maximum tissue phantom ratio (TMR) parameter of radiotherapy equipment to improve the generation of digitally reconstructed radiographs (DRRs), and obtains TMR-weighted DRR images. The TMR-weighted DRR images can more accurately simulate the dose distribution during radiotherapy. By generating TMR-weighted DRR images based on CT images, the fluence data generated by radiotherapy plans is merged with the TMR-weighted DRR images into a two-channel data, which serves as the input to the improved U-net network. During the training phase, a training dataset containing a large number of actual radiotherapy cases is constructed, and each case includes two-channel input data and corresponding EPID image data. By training the improved U-net network, it can learn the dose distribution characteristics during radiotherapy and become a prediction model that accurately predicts EPID images. During the prediction phase, the trained improved U-net network is used to predict new radiotherapy plans to generate predicted EPID images. Then, the gamma index analysis method is used to compare the predicted EPID images with the EPID images actually generated by radiotherapy equipment to evaluate the differences between the two. Through the comparison results, it can be determined whether the performance of radiotherapy equipment meets the requirements and whether the radiotherapy plan has been accurately executed.

[0006] The technical solution of the present invention is as follows: A quality control method based on an electronic portal imaging device and the maximum tissue phantom ratio, which improves DRR images by combining the maximum tissue phantom ratio TMR of radiotherapy equipment to obtain TMR-weighted DRR images; the TMR-weighted DRR images and the fluence map generated by radiotherapy plans are merged into a two-channel data, which serves as the input to the improved U-net network; the improved U-net network is trained, and a predicted EPID map is obtained through the trained improved U-net network; the predicted EPID map and the EPID map actually generated by radiotherapy equipment are analyzed based on the gamma index, and the quality control of radiotherapy equipment is carried out according to the gamma index.

[0007] When the relative position of the radiation source and the detector of the radiotherapy equipment remains fixed, the depth of the detector in water is changed to obtain a series of measured values that change with depth. The ratio of this series of measured values that change with depth to the maximum measured value is the TMR; the TMR-weighted DRR images are obtained by combining the TMR.

[0008] The steps for obtaining the TMR-weighted DRR images are as follows:

[0009] Step 1: According to the CT value - electron density correspondence table, convert the CT values in the CT image into relative electron densities to generate a volumetric data of three - dimensional relative electron density.

[0010] Step 2: According to the isocenter position of the radiation beam in the radiotherapy plan, adjust the volumetric data of the three - dimensional relative electron density obtained in Step 1 so that its volume center is located at the isocenter position of the radiation beam.

[0011] Step 3: Perform data format conversion and adjustment to make the size of the volumetric data of the three - dimensional relative electron density in Step 2 be 128x128x128, with a voxel size of 3mm; at the same time, generate the fluence map of each radiation field in the radiotherapy plan according to the DCM data of the radiotherapy plan; the size of the data after conversion of the fluence map is 128x128x1, with a pixel size of 3mm.

[0012] Step 4: Select the TMR parameter of the radiotherapy device to be quality - controlled currently; select the TMR parameter under the corresponding 10x10 radiation field; when the radiation field of the radiotherapy plan is 0 degrees, the top layer in the beam direction of the volumetric data of the three - dimensional relative electron density in Step 3 is 0mm deep, and the top layer of the volumetric data of the three - dimensional relative electron density is multiplied by the corresponding TMR parameter value at 0mm to become the first layer of the weighted volumetric data, and so on for the subsequent layers to obtain the weighted volumetric data; accumulate the weighted volumetric data of each layer along the beam direction to obtain the DRR image weighted by TMR.

[0013] When the radiation field in the radiotherapy plan is not 0 degrees, rotate the volumetric data of the three - dimensional relative electron density in Step 3 according to the radiation field angle. The first layer of the rotated volumetric data of the three - dimensional relative electron density is multiplied by the TMR parameter value corresponding to the first layer thickness, and so on for the subsequent layers to obtain the volumetric data weighted by TMR, and accumulate the weighted volumetric data of each layer to obtain the DRR image weighted by TMR.

[0014] The improved U - net network adopts the res U - net ++ network; the res U - net ++ network includes: a backbone block, four residual blocks, two receptive field blocks, and two attention blocks.

[0015] The input data is the fluence map of the radiation field and the DRR map weighted by TMR of the corresponding radiation field, and the output data label is the predicted EPID map after the corresponding beam passes through the phantom at different angles.

[0016] The input data first passes through the backbone block, with a filter size of 16, and the output result of the backbone block is c1.

[0017] c1 passes through the first residual block, with a filter size of 32, and the output result of the first residual block is c2.

[0018] c2 passes through the second residual block with a filter size of 64, and the output of the second residual block is c3;

[0019] c3 passes through the first receptive field block with a filter size of 128, and the output of the first receptive field block is b1;

[0020] b1 and c2 are used as the inputs of the first attention block. The output result is upsampled by a convolution kernel of (2, 2) to obtain d1a. After connecting d1a and c2, it passes through the third residual block with a filter size of 64, and the output result is d1;

[0021] d1 and c1 pass through the second attention block. The output result is upsampled by a convolution kernel of (2, 2) to obtain d2a. After connecting d2a and c1, it passes through the fourth residual block with a filter size of 32, and the output result is d2;

[0022] d2 passes through the second receptive field block with a filter size of 16, and then passes through a convolutional layer with a convolution kernel size of (1, 1) and an output channel of 1 and uses 'ReLu' activation. The activation result is used as the final output result.

[0023] In the backbone block, the input data passes through two branches respectively. The first branch is sequentially a convolutional layer, a batch normalization layer, a ReLu activation layer, and a convolutional layer. The second branch is sequentially a convolutional layer and a batch normalization layer. The outputs of the two branches are added together as the output of the backbone block;

[0024] In the residual block, the input data passes through two branches respectively. The first branch is sequentially a batch normalization layer, a ReLu activation layer, a convolutional layer with a convolution kernel of (3, 3), a batch normalization layer, a ReLu activation layer, and a convolutional layer with a convolution kernel of (3, 3). The second branch is sequentially a convolutional layer with a convolution kernel of (1, 1) and a batch normalization layer. The outputs of the two branches are added together as the output of the residual block;

[0025] In the receptive field block, according to the dose deposition characteristics of the beam physics in the tissue, the input data passes through four convolutional layers with a convolution kernel of (3, 3) and batch normalization and then are added together, and then passes through a convolutional layer with a convolution kernel of (1, 1) as the output of the receptive field block; Among the four convolutional layers, the first three convolutional layers are dilated convolutions with dilation rates of 2, 3, and 5 respectively;

[0026] The data input of the attention block is divided into two parameters, which are named g and x respectively within the attention block; g passes through two batch normalization layers, ReLu activation layers, convolutional layers with a convolutional kernel of (3,3), and then passes through a max pooling layer, and the result is labeled as g_pool; the result after x passes through two batch normalization layers, ReLu activation layers, and convolutional layers with a convolutional kernel of (3,3) is added to g_pool, and the result is labeled as gc_sum; gc_sum then passes through a batch normalization layer, ReLu activation layer, convolutional layer with a convolutional kernel of (3,3), and is multiplied by x, and the result is used as the output of the attention block; the inputs of the first attention block are b1 and c2, where b1 corresponds to g within the attention block and c2 corresponds to x within the attention block; the inputs of the second attention block are d1 and c1, where d1 corresponds to g within the attention block and c1 corresponds to x within the attention block.

[0027] The improved U-net network uses the following beam sequences and measurement methods to obtain training data. All beams irradiate the phantom, and the actual EPID images are synchronously obtained using the EPID system:

[0028] Square fields: Include square fields of sizes 1x1, 2x2, 3x3, 4x4, 5x5, 7x7, 10x10, 12x12, 15x15, 18x18. The center of the field is the center of the EPID image; for the first 5 fields, namely 1x1, 2x2, 3x3, 4x4, 5x5, the off-center positions are set to (5,0), (-5,0), (0,5), (0,-5), (5,5), (-5,5), (-5,-5), (5,-5) respectively; 50 MU for each field, a total of 50 fields;

[0029] Rectangular fields: 1x2, 1x5, 1x7, 1x10, 2x5, 2x7, 2x10. The isocenters of the corresponding beams for each field are set at (5,0), (0,0), (-5,0) respectively; 2x1, 5x1, 7x1, 10x1, 5x2, 7x2, 10x2. The isocenters of the corresponding beams for the fields are set at (0,5), (0,0), (0,-5) respectively; 50 MU for each field, a total of 42 fields;

[0030] Intensity-modulated fields: Select intensity-modulated fields with a field size within 24x24 cm. Transplant the radiotherapy plan onto the quality control phantom, a total of 40 cases and 360 beams.

[0031] Based on gamma index analysis, when the gamma index value between the predicted EPID image and the actual EPID image is greater than 93%, the radiotherapy equipment is operating normally; when the gamma index value is between 90 - 93%, the operating state of the equipment is manually confirmed. When the gamma index is less than 90%, the radiotherapy equipment is operating abnormally, and the radiotherapy equipment is stopped and the relevant parameters of the radiotherapy equipment operation are checked.

[0032] Advantages of the present invention: Compared with the prior art, the method of the present invention has low implementation difficulty, fast operation speed, high accuracy, and can quickly and accurately determine the setup error. In this study, the maximum tissue phantom ratio data collected during the equipment operation test was used to convert the three-dimensional CT data into two-dimensional data. The data itself does not require additional collection, and the conversion process is very fast under the existing computer technology. An improved U-net network is used. After the improved U-net network is established, it has a high operation speed. After gamma analysis, it can be found that the accuracy is high, and it is relatively easy to detect and determine the setup error during verification. Description of the Drawings

[0033] Figure 1 It is a flowchart of the quality control method based on the electronic portal imaging device and the maximum tissue phantom ratio;

[0034] Figure 2 It is a predicted EPID image introducing a 5% dose deviation;

[0035] Figure 3 It is a gamma analysis chart of the actual EPID image and the predicted EPID image with a 5% dose deviation;

[0036] Figure 4 It is a dose deviation chart of the predicted EPID image and the actual EPID image introducing a 0.5 - 2 mm deviation of the leaf;

[0037] Figure 5 It is a gamma analysis chart of the actual EPID image and the predicted EPID image of the leaf position deviation;

[0038] Figure 6 It is a network structure diagram of the backbone block of the improved U-net network;

[0039] Figure 7 It is a network structure diagram of the residual block of the improved U-net network;

[0040] Figure 8 It is a network structure diagram of the attention block of the improved U-net network;

[0041] Figure 9 It is a network structure diagram of the receptive field block of the improved U-net network;

[0042] Figure 10 It is a network structure diagram of the improved U-net network. Detailed Embodiments

[0043] In order to accurately predict EPID images based on radiotherapy plans, the present invention designs a quality control method based on an electronic portal imaging device (EPID) and the maximum tissue phantom ratio (TMR). Combining the maximum tissue phantom ratio (TMR, The maximum tissue Phantom ratio) of the device, based on CT images, applying TMR parameters, generating TMR-weighted digital reconstructed radiographs (DRRs), based on an improved U-net network, merging the fluence map generated from the radiotherapy plan and the data of the TMR-weighted DRR images into a two-channel data as the input, and using the image data of the EPID data as the output to construct a training dataset. The trained improved U-net network is used for prediction, and based on gamma index analysis, the difference between the predicted EPID map and the actual EPID map is analyzed, so as to realize the quality control of radiotherapy equipment.

[0044] A quality control method based on an electronic portal imaging device (EPID) and the maximum tissue phantom ratio (TMR) improves the DRR images by combining the maximum tissue phantom ratio (TMR) of the radiotherapy equipment to obtain TMR-weighted DRR images; the TMR-weighted DRR images and the fluence map generated from the radiotherapy plan are merged into a two-channel data as the input of the improved U-net network; the improved U-net network is trained, and a predicted EPID map is obtained through the trained improved U-net network; based on gamma index analysis, the predicted EPID map and the EPID map actually generated by the radiotherapy equipment are analyzed, and the quality control of the radiotherapy equipment is carried out according to the gamma index.

[0045] When the relative positions of the radiation source and the detector of the radiotherapy equipment are fixed, the depth of the detector in water is changed to obtain a series of measured values varying with the depth. The ratio of this series of measured values varying with the depth to the maximum measured value is the TMR; the TMR-weighted DRR images are obtained by combining the TMR.

[0046] The steps for obtaining the TMR-weighted DRR images are as follows:

[0047] Step 1: According to the correspondence table between CT values and electron densities, convert the CT values in the CT image into relative electron densities to generate a three-dimensional volume data of relative electron densities.

[0048] Step 2: According to the beam isocenter position in the radiotherapy plan file, adjust the three-dimensional volume data of relative electron densities obtained in Step 1 so that its volume center is located at the beam isocenter.

[0049] Step 3: Perform data format conversion and adjustment to make the size of the three-dimensional volume data of relative electron densities in Step 2 be 128x128 x 128, and the voxel size be 3 mm; at the same time, generate the fluence maps of each radiation field of the radiotherapy plan according to the DCM data of the radiotherapy plan; the size of the converted fluence map data is 128x128x1, and the pixel size is 3 mm.

[0050] Step 4: Select the TMR parameters of the radiotherapy equipment to be quality controlled currently; select the TMR parameters corresponding to the 10x10 field. When the field of the radiotherapy plan is 0 degrees, the uppermost layer of the beam direction of the volumetric data of the three-dimensional relative electron density in Step 3 is 0 mm deep. Multiply the volumetric data of the uppermost layer of the three-dimensional relative electron density by the corresponding TMR parameter value at 0 mm to obtain the first layer of the weighted volumetric data. And so on for each subsequent layer to obtain the weighted volumetric data; accumulate the weighted volumetric data of each layer along the beam direction to obtain the DRR image weighted by TMR.

[0051] When the field in the radiotherapy plan is not 0 degrees, rotate the volumetric data of the three-dimensional relative electron density in Step 3 according to the field angle. Multiply the first layer of the rotated volumetric data of the three-dimensional relative electron density by the TMR parameter value corresponding to the first layer thickness, and so on for each subsequent layer to obtain the volumetric data weighted by TMR. Accumulate the weighted volumetric data of each layer to obtain the DRR image weighted by TMR.

[0052] The improved U-net network adopts the res U-net ++ network; the res U-net ++ network includes: a backbone block, four residual blocks, two receptive field blocks, and two attention blocks.

[0053] The input data is the fluence map under the field and the DRR map weighted by TMR corresponding to the field, and the output data label is the predicted EPID image after the corresponding beam passes through the phantom at different angles.

[0054] Set the filter size to (16, 32, 64, 128).

[0055] The input data first passes through the backbone block with a filter size of 16, and the output result of the backbone block is c1.

[0056] c1 passes through the first residual block with a filter size of 32, and the output result of the first residual block is c2.

[0057] c2 passes through the second residual block with a filter size of 64, and the output result of the second residual block is c3.

[0058] c3 passes through the first receptive field block with a filter size of 128, and the output result of the first receptive field block is b1.

[0059] b1 and c2 are used as the input of the first attention block. The output result is upsampled by a convolutional kernel of (2, 2) to be d1a. After connecting d1a and c2, it passes through the third residual block with a filter size of 64, and the output result is d1.

[0060] d1 and c1 pass through the second attention block, and the output result is upsampled with a convolution kernel of (2, 2) to obtain d2a. After connecting d2a and c1, it passes through the fourth residual block with a filter size of 32, and the output result is d2;

[0061] d2 passes through the second receptive field block with a filter size of 16, and then passes through a convolutional layer with a convolution kernel size of (1, 1) and an output channel of 1 and uses 'ReLu' activation. The activation result is used as the final output result.

[0062] In the backbone block, the input data passes through two branches respectively. The first branch is sequentially a convolutional layer, a batch normalization layer, a ReLu activation layer, and a convolutional layer. The second branch is sequentially a convolutional layer and a batch normalization layer. The outputs of the two branches are added together as the output of the backbone block;

[0063] In the residual block, the input data passes through two branches respectively. The first branch is sequentially a batch normalization layer, a ReLu activation layer, a convolutional layer with a convolution kernel of (3, 3), a batch normalization layer, a ReLu activation layer, and a convolutional layer with a convolution kernel of (3, 3). The second branch is sequentially a convolutional layer with a convolution kernel of (1, 1) and a batch normalization layer. The outputs of the two branches are added together as the output of the residual block;

[0064] In the receptive field block, according to the dose deposition characteristics of the beam physics in the tissue, the input data passes through four convolutional layers with a convolution kernel of (3, 3) and batch normalization and then is added together, and then passes through a convolutional layer with a convolution kernel of (1, 1) as the output of the receptive field block; Among the four convolutional layers, the first three convolutional layers are dilated convolutions with dilation rates of 2, 3, and 5 respectively;

[0065] The data input of the attention block is two parameters, so they are respectively named g and x in the attention block; g passes through two batch normalization layers, ReLu activation layers, convolutional layers with a convolution kernel of (3, 3), and then passes through a max pooling layer, and the result is marked as g_pool; the result after x passes through two batch normalization layers, ReLu activation layers, convolutional layers with a convolution kernel of (3, 3) is added to g_pool, and the result is marked as gc_sum; gc_sum then passes through a batch normalization layer, a ReLu activation layer, a convolutional layer with a convolution kernel of (3, 3), and then is multiplied by x, and the result is used as the output of the attention block. The input of the first attention block is b1 and c2, where b1 corresponds to g in the attention block and c2 corresponds to x in the attention block. The input of the second attention block is d1 and c1, where d1 corresponds to g in the attention block and c1 corresponds to x in the attention block.

[0066] The improved U-net network uses the following beam sequence and measurement method to obtain the data required for model training. All beams irradiate the quality control phantom, and the actual EPID images are synchronously obtained using the EPID system:

[0067] Square fields: Include field sizes of 1x1, 2x2, 3x3, 4x4, 5x5, 7x7, 10x10, 12x12, 15x15, 18x18. The center of the field is the center of the EPID image. For the first 5 fields, i.e., 1x1, 2x2, 3x3, 4x4, 5x5, the off - center positions are set to (5,0), (-5,0), (0,5), (0, -5), (5,5), (-5,5), (-5,-5), (5, -5) respectively. Each field is 50 MU, and there are a total of 50 fields.

[0068] Rectangular fields: 1x2, 1x5, 1x7, 1x10, 2x5, 2x7, 2x10. The beam isocenters corresponding to each field are at (5,0), (0,0), (-5,0) respectively; for 2x1, 5x1, 7x1, 10x1, 5x2, 7x2, 10x2, the beam isocenters corresponding to the fields are set at (0,5), (0,0), (0,-5) respectively. Each field is 50 MU, and there are a total of 42 fields.

[0069] Intensity - modulated fields: Select intensity - modulated fields with field sizes within 24x24 cm. Transplant the radiotherapy plan onto the quality control phantom. There are a total of 40 cases and 360 beams.

[0070] Based on gamma - index analysis, when the gamma - index value between the predicted EPID image and the actual EPID image is greater than 93%, the radiotherapy equipment is operating normally. When the gamma - index is between 90 - 93%, manually confirm the equipment operating status. When the gamma - index is less than 90%, the radiotherapy equipment is not operating normally, stop the radiotherapy equipment and check the relevant parameters of the radiotherapy equipment operation.

[0071] Further, taking a specific embodiment as an example, its TMR parameters are obtained from Table 1. The 10x10 TMR parameters are converted from the PDD (Percentage Depth Dose) when SSD (Source - Skin Distance) = 100 cm, and the depth is 310 mm.

[0072] Table 1 Dose distribution table varying with depth

[0073]

[0074] The specific correspondence table between CT value and electron density is as follows:

[0075] Table 2 CT value and relative electron density table used in the current measurement

[0076]

[0077] The specific steps of this embodiment are as follows:

[0078] Compared with the conventional DRR projection method, the DRR projection with fused TMR parameters is adopted to obtain the TMR-weighted DRR image.

[0079] The conventional DRR projection method is generally: the 3D CT volume data is generated through a mathematical simulation algorithm to generate a 2D image, which is called a DRR image. The basic method of generating a DRR image is the ray projection method. Define a point as a virtual X-ray source, and the X-ray passes through the object to be measured from the virtual ray source and is projected to the detector plane perpendicular to the central axis of the ray. The intersection of the ray and the plane determines the position of the DRR midpoint; in this process, the intersection of the point source and each CT slice in the CT volume data can be calculated to obtain the corresponding CT value. After a ray is projected, the CT value obtained by accumulating the entire path is the DRR image pixel value of the corresponding point on the detector; repeat the above steps. When all the rays are projected, the CT value of each pixel block of the entire DRR image will be obtained. Mapping its accumulated value to the pixel gray value will obtain a DRR image.

[0080] In this embodiment, when the radiation field of the radiotherapy plan is 0 degrees, the top layer of the beam direction of the three-dimensional relative electron density volume data is 0 mm deep, and the top layer of the three-dimensional relative electron density volume data is multiplied by the corresponding TMR parameter value at 0 mm to become the first layer of weighted volume data, and the following layers are deduced in the same way to obtain weighted volume data, as shown in Formula 1, where the 3x3 matrix represents the relative electron density matrix of the current layer, and 0.95 represents the TMR parameter value at the corresponding depth of the layer obtained from the TMR table; the weighted volume data of each layer are accumulated along the beam direction to obtain a TMR-weighted DRR image;

[0081] (1)

[0082] When the radiation field in the radiotherapy plan is not 0 degrees, the three-dimensional relative electron density volume data is rotated according to the radiation field angle. The first layer of the rotated three-dimensional relative electron density volume data is multiplied by the TMR parameter value corresponding to the thickness of the first layer, and the subsequent layers are deduced in the same way to obtain TMR-weighted volume data. The weighted volume data of each layer are accumulated to obtain the TMR-weighted DRR image.

[0083] The radiotherapy plan file includes: radiotherapy plan file RT Plan, radiotherapy outline file RT Structure, radiotherapy dose file RT Dose and radiotherapy plan CT file. All files are in DICOM format with the file suffix .DCM. The RT Plan file contains the radiation field information of the radiotherapy plan, such as the source-skin distance, the position of the beam isocenter, the intensity modulated radiation field size, and the position information of all multi-leaf collimator leaves.

[0084] Furthermore, the gamma index is a statistical method used to compare the differences between the calculated dose distribution and the actual measured dose distribution. It comprehensively considers the distance between the calculation point and the measurement point (Distance to Agreement, DTA) and the degree of difference between dose values (Dose Difference, DD). In the present invention, if the average value of the gamma index of 2%, 2 mm exceeds 93%, it is considered that the radiotherapy plan data can be used.

[0085] Artificially add errors to the radiotherapy plan. In this embodiment, it is assumed that the output dose is adjusted by 5%. Figure 2 It is the predicted EPID image introducing a 5% dose deviation. Figure 3 It is the gamma index map of the predicted EPID response and the actual EPID image, and its gamma passing rate is 70%, which is much less than the threshold of 90% that requires human intervention in the commonly used clinical plan evaluation. This example shows that the present invention can effectively detect the output dose deviation of radiotherapy equipment.

[0086] Artificially add errors to the treatment plan. In this embodiment, it is assumed that there are deviations of 0.5 - 2 mm in the positions of the leaves. Figure 4 It is the dose deviation map, that is, the dose difference at the corresponding positions of the predicted EPID response and the actual EPID image. Figure 5 It is the gamma index map of the predicted EPID response and the actual EPID image, and its gamma passing rate is 53%, which is much less than the threshold of 90% that requires human intervention in the commonly used clinical plan evaluation. This example shows that the present invention can effectively detect the position deviation of the multi - leaf collimator of radiotherapy equipment.

[0087] In summary, the present invention does not simply use the traditional DRR projection algorithm, but combines the TMR parameters of radiotherapy equipment for optimization. The TMR parameters reflect the dose response characteristics of radiotherapy equipment under different tissue densities. By incorporating them into the DRR image generation method, the dose distribution during radiotherapy can be more accurately simulated, thereby improving the accuracy of dose prediction.

[0088] The present invention innovatively introduces deep learning technology, especially the U - net network, to learn the dose distribution characteristics during radiotherapy from CT images and the flux data generated by radiotherapy plans. By constructing a training data set containing a large number of actual radiotherapy cases and training an improved U - net network, it can accurately predict EPID images. This data - driven method can capture complex dose distribution patterns and improve the accuracy and robustness of prediction.

[0089] The present invention combines the fluence data generated by radiotherapy planning with the TMR-weighted DRR image data into a two-channel data as the input of the deep learning model. This strategy makes full use of the information of radiotherapy planning and the anatomical structure of the phantom, which helps the model better learn the dose distribution characteristics during the radiotherapy process.

[0090] In the prediction stage, the present invention uses the gamma index analysis method to compare the predicted EPID image with the actual EPID image to quantify the difference between the two. The gamma index is a widely used dose verification tool that can comprehensively consider the dose difference and the position difference, thus providing a comprehensive dose distribution assessment. Through the comparison results, it can be accurately judged whether the performance of the radiotherapy equipment meets the requirements and whether the radiotherapy plan has been accurately executed.

Claims

1. A quality control method based on an electronic portal imaging device and maximum tissue phantom ratio, characterized in that, Improve the DRR image by combining the maximum tissue phantom ratio (TMR) of the radiotherapy device to obtain a TMR-weighted DRR image; merge the TMR-weighted DRR image and the fluence map generated by the radiotherapy plan into a two-channel data as the input of the improved U-net network; train the improved U-net network, and obtain the predicted EPID map through the trained improved U-net network; analyze the predicted EPID map and the actually generated EPID map of the radiotherapy device based on the gamma index, and perform quality control of the radiotherapy device according to the gamma index. The steps for obtaining the TMR-weighted DRR image are as follows: The first step: According to the CT value and electron density correspondence table, convert the CT values in the CT image into relative electron densities to generate a three-dimensional volume data of relative electron densities. The second step: Adjust the three-dimensional volume data of relative electron densities obtained in the first step according to the beam isocenter position of the radiotherapy plan so that its volume center is located at the beam isocenter position. The third step: Perform data format conversion and adjustment to make the size of the three-dimensional volume data of relative electron densities in the second step 128x128x128, with a voxel size of 3 mm; at the same time, generate the fluence map of each beam of the radiotherapy plan according to the DCM data of the radiotherapy plan; the size of the converted fluence map data is 128x128x1, with a pixel size of 3 mm. The fourth step: Select the TMR parameter of the radiotherapy device to be quality controlled currently; select the TMR parameter under the corresponding 10x10 beam; when the beam of the radiotherapy plan is 0 degrees, the uppermost layer in the beam direction of the three-dimensional volume data of relative electron densities in the third step is 0 mm deep, and the uppermost layer of the three-dimensional volume data of relative electron densities is multiplied by the corresponding TMR parameter value at 0 mm to become the first layer of the weighted volume data, and so on for the subsequent layers to obtain the weighted volume data. Accumulate the weighted volume data of each layer along the beam direction to obtain the TMR-weighted DRR image. When the beam in the radiotherapy plan is not 0 degrees, rotate the three-dimensional volume data of relative electron densities in the third step according to the beam angle. The first layer of the rotated three-dimensional volume data of relative electron densities is multiplied by the TMR parameter value corresponding to the first layer thickness, and so on for the subsequent layers to obtain the TMR-weighted volume data. Accumulate the weighted volume data of each layer to obtain the TMR-weighted DRR image.

2. The quality control method based on an electronic portal imaging device and maximum tissue phantom ratio according to claim 1, wherein When the relative position of the radiation source and the detector of the radiotherapy device remains fixed, change the depth of the detector in water to obtain a series of measured values that change with depth. The ratio of this series of measured values that change with depth to the maximum measured value is the TMR; combine the TMR to obtain the TMR-weighted DRR image.

3. The quality control method based on the electronic portal imaging system and the maximum tissue phantom ratio according to claim 1, wherein The improved U-net network adopts the res U-net++ network. The res U-net++ network includes: a backbone block, four residual blocks, two receptive field blocks, and two attention blocks. The input data is the fluence map of the beam and the TMR-weighted DRR map of the corresponding beam, and the output data label is the predicted EPID map of the corresponding beam passing through the phantom at different angles. The input data first passes through the backbone block, with a filter size of 16, and the output result of the backbone block is c1. c1 passes through the first residual block with a filter size of 32, and the output of the first residual block is c2; c2 passes through the second residual block with a filter size of 64, and the output of the second residual block is c3; c3 passes through the first receptive field block with a filter size of 128, and the output of the first receptive field block is b1; b1 and c2 are used as the inputs of the first attention block. After the output is upsampled by a convolutional kernel of (2,2), it becomes d1a. After d1a is concatenated with c2 and passes through the third residual block with a filter size of 64, the output is d1; d1 and c1 pass through the second attention block. After the output is upsampled by a convolutional kernel of (2,2), it becomes d2a. After d2a is concatenated with c1 and passes through the fourth residual block with a filter size of 32, the output is d2; d2 passes through the second receptive field block with a filter size of 16, and then passes through a convolutional layer with a convolutional kernel size of (1,1) and an output channel of 1 and uses 'ReLu' activation. The activation result is used as the final output result.

4. The quality control method based on the electronic portal imaging system and the maximum tissue phantom ratio according to claim 3, characterized in that In the backbone block, the input data passes through two branches respectively. The first branch is sequentially a convolutional layer, a batch normalization layer, a ReLu activation layer, and a convolutional layer. The second branch is sequentially a convolutional layer and a batch normalization layer. The outputs of the two branches are added together as the output of the backbone block; In the residual block, the input data passes through two branches respectively. The first branch is sequentially a batch normalization layer, a ReLu activation layer, a convolutional layer with a convolutional kernel of (3,3), a batch normalization layer, a ReLu activation layer, and a convolutional layer with a convolutional kernel of (3,3). The second branch is sequentially a convolutional layer with a convolutional kernel of (1,1) and a batch normalization layer. The outputs of the two branches are added together as the output of the residual block; In the receptive field block, according to the dose deposition characteristics of the beam physics in the tissue, the input data passes through four convolutional layers with a convolutional kernel of (3,3) and batch normalization and then added together, and then passes through a convolutional layer with a convolutional kernel of (1,1) as the output of the receptive field block; among the four convolutional layers, the first three convolutional layers are dilated convolutions with dilation rates of 2, 3, and 5 respectively; The data input of the attention block is divided into two parameters, which are named g and x respectively within the attention block; g passes through two batch normalization layers, ReLu activation layers, convolutional layers with a convolutional kernel of (3,3), and then passes through a max pooling layer, and the result is labeled as g_pool; the result after x passes through two batch normalization layers, ReLu activation layers, convolutional layers with a convolutional kernel of (3,3) is added to g_pool, and the result is labeled as gc_sum; gc_sum then passes through a batch normalization layer, ReLu activation layer, convolutional layer with a convolutional kernel of (3,3) and is multiplied by x, and the result is used as the output of the attention block; the inputs of the first attention block are b1 and c2, where b1 corresponds to g within the attention block and c2 corresponds to x within the attention block; the inputs of the second attention block are d1 and c1, where d1 corresponds to g within the attention block and c1 corresponds to x within the attention block.

5. The quality control method based on an electronic portal imaging device and maximum tissue-air ratio according to claim 1, wherein The improved U-net network uses the following beam sequences and measurement methods to obtain training data. All beams irradiate the phantom, and the actual EPID images are synchronously acquired using the EPID system: Field: It includes fields with sizes of 1x1, 2x2, 3x3, 4x4, 5x5, 7x7, 10x10, 12x12, 15x15, 18x18. The center of the field is the center of the EPID image. For the first 5 fields, namely 1x1, 2x2, 3x3, 4x4, 5x5, the off-center positions are set as (5,0), (-5,0), (0,5), (0, -5), (5,5), (-5,5), (-5,-5), (5, -5) respectively. Each field has 50 MU, and there are a total of 50 fields. Strip field: 1x2, 1x5, 1x7, 1x10, 2x5, 2x7, 2x10. The isocenters of the corresponding beams for each field are set at (5,0), (0,0), (-5,0) respectively; 2x1, 5x1, 7x1, 10x1, 5x2, 7x2, 10x2. The isocenters of the corresponding beams for the fields are set at (0,5), (0,0), (0,-5) respectively. Each field has 50 MU, and there are a total of 42 fields. Intensity-modulated field: Select intensity-modulated fields with field sizes within 24x24 cm. Transplant the radiotherapy plan to the quality control phantom. There are a total of 40 cases and 360 beams.

6. The quality control method based on the electronic portal imaging system and the maximum tissue phantom ratio according to claim 1, characterized in that, Based on gamma index analysis, when the gamma index value between the predicted EPID image and the actual EPID image is greater than 93%, the operating state of the radiotherapy equipment is normal; when the gamma index value is between 90 - 93%, the operating state of the equipment is manually confirmed. When the gamma index is less than 90%, the operating state of the radiotherapy equipment is abnormal, and the radiotherapy equipment is stopped and the relevant parameters of the radiotherapy equipment operation are checked.

Images