Multi-leaf collimator irradiation error classification method and related device based on deep learning

Abstract

The present application discloses a multi-leaf collimator irradiation error classification method and related device based on deep learning, which relates to the field of radiotherapy equipment quality control. The first resolution radiation measurement dose distribution map is applied to a super-resolution reconstruction model to improve the resolution, and a dose difference map and a gamma distribution map are obtained according to the second resolution radiation measurement dose distribution map with improved resolution and the corresponding radiation calculation dose distribution map; the multi-leaf collimator leaf position and sub-field weight map are obtained according to the radiotherapy plan file; the dose difference map, the gamma distribution map and the multi-leaf collimator leaf position and sub-field weight map are input into an error classification network model including a CNN path and a Transformer path to obtain the type of multi-leaf collimator irradiation error. In the present invention, the introduction of the multi-leaf collimator leaf position and sub-field weight map, the introduction of the super-resolution reconstruction model and the error classification network model improves the accuracy of error classification.

Description

Technical Field

[0001] The present application relates to the technical field of radiotherapy equipment quality control, and in particular to a deep learning-based irradiation error classification method and related device for a medical electron linear accelerator multi-leaf collimator. Background Art

[0002] Intensity Modulated Radiation Therapy (IMRT) is an important treatment method in tumor radiotherapy, which can minimize the dose of surrounding organs at risk on the basis of high-dose irradiation of the tumor. Since many control points are used for dose modulation in IMRT, it is necessary to ensure that the electron straight line accurately executes complex radiotherapy plans for the safety of patients and treatment efficiency. The commonly used method at present is the gamma analysis method based on measurement, but this method has insufficient detection accuracy and cannot classify the sources and types of these errors. These errors include leaf position errors of multi-leaf collimators, output dose errors, multi-leaf collimator parameter errors, etc. If the treatment plan error is detected by using the gamma method, the physicist needs to manually detect the type and source of the error, which has low detection efficiency, further affecting the treatment efficiency and timeliness of treatment.

[0003] In recent years, some research teams have been committed to solving the problem of automatic detection and classification of leaf position errors of multi-leaf collimators in IMRT. Specifically, the article "Deep learning for patient-specific quality assurance: Identifying errors in radiotherapy delivery by radiomic analysis of gamma images with convolutional neural networks" has been published, which proposes a method based on electronic portal imaging device (EPID) and radiomics combined with convolutional neural networks to achieve the classification of leaf position errors of multi-leaf collimators in IMRT. The article "The structural similarity index for IMRT quality assurance: radiomics-based error classification" has been published, which uses EPID and radiomics methods, combined with machine learning algorithms, to achieve the classification of leaf position errors of multi-leaf collimators in IMRT. However, previous studies were based on high-resolution detectors or three-dimensional detectors, such as EPID. However, in current medical institutions, radiotherapy departments generally use low-resolution two-dimensional detectors. Therefore, it is more clinically valuable and meaningful to directly use low-resolution two-dimensional detectors with a larger inventory to identify errors and classification of multi-leaf collimators of medical electron linear accelerators. However, when using low-resolution detectors, existing machine learning or deep learning techniques are not effective in classifying the position errors of multi-leaf collimators of medical electron linear accelerators. Summary of the invention

[0004] The purpose of this application is to provide a multi-leaf collimator irradiation error classification method and related devices based on deep learning, which can improve the accuracy of multi-leaf collimator irradiation error classification.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a multi-leaf collimator irradiation error classification method based on deep learning, comprising:

[0007] Obtaining a first-resolution radiation measurement dose distribution map and a corresponding radiation calculation dose distribution map output by a two-dimensional detector in a radiotherapy plan file;

[0008] Inputting the first-resolution radiation measurement dose distribution map into a trained super-resolution reconstruction model to obtain a second-resolution radiation measurement dose distribution map; the super-resolution reconstruction model includes a generator and a discriminator connected in series; the resolution of the second-resolution radiation measurement dose distribution map is higher than the resolution of the first-resolution radiation measurement dose distribution map;

[0009] According to the second resolution radiation measurement dose distribution map and the corresponding radiation calculation dose distribution map, a standard gamma method is applied to obtain a dose difference map and a gamma distribution map;

[0010] The multi-leaf collimator leaf positions and sub-field weight maps are obtained based on the radiotherapy planning file;

[0011] The dose difference map, the gamma distribution map and the multi-leaf collimator leaf position and sub-field weight map are input into the trained error classification network model to obtain the type of multi-leaf collimator irradiation error; the error classification network model includes a first Stem module, a CNN path, a Transformer path, a feature fusion module and a first classification head module; the CNN path includes a plurality of first residual channel attention module groups connected in series; each of the first residual channel attention module groups includes a plurality of first residual channel attention modules connected in series; the input of the first first residual channel attention module in the CNN path is the output of the first Stem module; the Transformer path includes a Patch Partition module, a Linear Embedding module, a first SwinT Block module group and a plurality of PMSB module groups connected in series in sequence; the first SwinT Block module group includes a plurality of first SwinT Block modules connected in series in sequence; each of the PMSB module groups includes a Patch Merging module and a second SwinT Block module group connected in series in sequence; the second SwinT Block module group includes a plurality of second SwinT Block module; the input of the PatchPartition module is the output of the first Stem module; the input of the feature fusion module is the output of the last first residual channel attention module in the CNN path and the output of the last PMSB module group in the Transformer path; the input of the first classification head module is the output of the feature fusion module, and the output of the first classification head module is the type of the multi-leaf collimator illumination error.

[0012] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-leaf collimator irradiation error classification method based on deep learning.

[0013] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-leaf collimator irradiation error classification method based on deep learning.

[0014] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the multi-leaf collimator irradiation error classification method based on deep learning.

[0015] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0016] The present application provides a multi-leaf collimator irradiation error classification method and related devices based on deep learning. In the first aspect, a super-resolution reconstruction model is introduced, which can reconstruct a low-resolution radiation measurement dose distribution map into a high-resolution radiation measurement dose distribution map. The high-resolution radiation measurement dose distribution map is used for error classification, which can accurately process the errors of millimeter-level multi-leaf collimator leaves and improve the accuracy of error classification. In the second aspect, an error classification network model including a CNN path and a Transformer path is also introduced, which can combine local semantic features and global semantic features to improve the accuracy of error identification. In the third aspect, when performing error classification, in the input data, on the basis of the dose difference map and the gamma distribution map, the multi-leaf collimator leaf position and sub-field weight map are also introduced, which can accurately guide the position of the multi-leaf collimator leaves, and improve the error classification effect during error classification. In the fourth aspect, the error classification method of the present invention can realize the accurate classification of the position error of the multi-leaf collimator of the medical electron linear accelerator without relying on the relevant knowledge of experts. Fifthly, the error classification method of the present invention uses data from low-resolution detectors with a larger number of detectors, and classifies them through a super-resolution CNN / Transformer hybrid deep learning network structure (i.e., a super-resolution reconstruction model and an error classification network model), which has good classification effect and expands the scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 This is an application environment diagram of a multi-leaf collimator illumination error classification method based on deep learning in one embodiment of the present application;

[0019] Figure 2 A schematic diagram of a flow chart of a multi-leaf collimator illumination error classification method based on deep learning provided in one embodiment of the present application;

[0020] Figure 3 A conceptual diagram of a multi-leaf collimator illumination error classification method based on deep learning provided in one embodiment of the present application;

[0021] Figure 4 A schematic diagram of the structure of an error classification network model provided in one embodiment of the present application;

[0022] Figure 5 A schematic diagram of the structure of a super-resolution reconstruction model provided in one embodiment of the present application;

[0023] Figure 6 A schematic diagram of the structure of a residual channel attention module provided in an embodiment of the present application;

[0024] Figure 7 A schematic diagram of the structure of a SwinT Block module provided in an embodiment of the present application;

[0025] Figure 8 A schematic diagram of the structure of a CBR module provided in one embodiment of the present application;

[0026] Fig. 9 A schematic diagram of the structure of a Stem module provided in one embodiment of the present application;

[0027] Fig.10 A schematic diagram of the structure of a feature fusion module provided in one embodiment of the present application;

[0028] Fig.11 A schematic diagram of the training process of a super-resolution reconstruction model and an error classification network model provided in one embodiment of the present application;

[0029] Fig.12 A schematic diagram of the positions of the leaves of a multi-leaf collimator and the weights of sub-fields provided in one embodiment of the present application;

[0030] Fig.13 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in 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. 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.

[0032] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0033] The multi-leaf collimator illumination error classification method based on deep learning provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the radiotherapy plan file to the server 104. After the server 104 receives the radiotherapy plan file, the server 104 obtains the first resolution radiation measurement dose distribution map output by the two-dimensional detector in the radiotherapy plan file and the corresponding radiation calculation dose distribution map; the first resolution radiation measurement dose distribution map is input into the trained super-resolution reconstruction model to obtain the second resolution radiation measurement dose distribution map; the standard gamma method is applied to obtain the dose difference map and the gamma distribution map according to the second resolution radiation measurement dose distribution map and the corresponding radiation calculation dose distribution map; the multi-leaf collimator leaf position and sub-field weight map are obtained according to the radiotherapy plan file; the dose difference map, the gamma distribution map and the multi-leaf collimator leaf position and sub-field weight map are input into the trained error classification network model to obtain the type of multi-leaf collimator irradiation error. The server 104 may feed back the type of the obtained multi-leaf collimator irradiation error to the terminal 102. In addition, in some embodiments, the multi-leaf collimator irradiation error classification method based on deep learning may also be implemented by the server 104 or the terminal 102 alone, such as the terminal 102 may directly classify the multi-leaf collimator irradiation error for the radiotherapy plan file, or the server 104 may obtain the radiotherapy plan file from the data storage system and classify the multi-leaf collimator irradiation error for the radiotherapy plan file.

[0034] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0035] In an exemplary embodiment, Figure 2 and Figure 3 As shown, a multi-leaf collimator irradiation error classification method based on deep learning is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or can be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used for explanation, and the steps include the following steps 201 to 205.

[0036] Step 201 : obtaining a low-resolution radiation measurement dose distribution map (ie, a first-resolution radiation measurement dose distribution map) output by a two-dimensional detector in a radiotherapy plan file and a corresponding radiation calculation dose distribution map.

[0037] The patient's radiotherapy plan file (DICOM-RTplan file) is transferred to the medical electron linear accelerator for irradiation, and the PTW 729 two-dimensional low-resolution detector is used to complete the measurement to obtain a low-resolution radiation measurement dose distribution map. The two-dimensional low-resolution detector can be the PTW 729 two-dimensional low-resolution detector produced by PTW, and can also be extended to detectors produced by other companies, without any limitation here.

[0038] Step 202, input the low-resolution radiation measurement dose distribution map into the trained super-resolution reconstruction model to obtain a high-resolution radiation measurement dose distribution map (i.e., a second-resolution radiation measurement dose distribution map); the super-resolution reconstruction model includes a generator and a discriminator connected in series.

[0039] Generally, the image output by the PTW 729 two-dimensional low-resolution detector has a low resolution, usually between 0.5cm and 1cm. Through super-resolution reconstruction, its resolution can be increased to 8 times the original resolution, that is, 0.0625cm to 0.125cm, reaching the standard of high-resolution radiation dose distribution map.

[0040] Step 203 : applying a standard gamma method according to the high-resolution radiation measurement dose distribution map and the corresponding radiation calculation dose distribution map to obtain a dose difference map and a gamma distribution map.

[0041] Step 204: derive the leaf positions and sub-field weight map of the multi-leaf collimator according to the radiotherapy plan file.

[0042] In the existing error classification method, the patient's radiotherapy plan to be verified is not involved in the detection link. Since the patient's radiotherapy plan file contains the accurate position information of the leaves of the electron linear accelerator multi-leaf collimator, if the patient's radiotherapy plan file is added on the basis of the radiation measurement dose and the radiation calculation dose, the accuracy of the error classification of the electron linear accelerator multi-leaf collimator will be improved. Therefore, the present invention introduces a radiotherapy plan file, based on which the multi-leaf collimator leaf position and sub-field weight map can be obtained, so that the leaf position can be guided during classification, improving the error classification effect.

[0043] Step 205 , input the dose difference map, the gamma distribution map, and the multi-leaf collimator leaf position and sub-field weight map into the trained error classification network model to obtain the type of multi-leaf collimator irradiation error.

[0044] Load the trained network weight parameters, input the input data to be classified into the error classification network model, and use the Softmax function to obtain the probability value of the input data to be classified on each error type of the multi-leaf collimator. Keep the largest probability value, and the category corresponding to the largest probability value is the final prediction category.

[0045] In an exemplary embodiment, Figure 4 As shown, the error classification network model includes a first Stem module (corresponding to Figure 4 Stem-1 in ), CNN path, Transformer path, feature fusion module and the first classification head module (corresponding to Figure 4 The CNN path includes a plurality of first residual channel attention module groups connected in series; each of the first residual channel attention module groups includes a plurality of first residual channel attention modules connected in series (corresponding to Figure 4 RCA-1 in the CNN path); the input of the first residual channel attention module in the CNN path is the output of the first Stem module; the Transformer path includes a Patch Partition module, a LinearEmbedding module, a first SwinT Block module group and a plurality of PMSB module groups connected in series in sequence; the first SwinT Block module group includes a plurality of first SwinT Block modules connected in series in sequence (corresponding to Figure 4SwinT Block-1 in the PMSB module group); each of the PMSB module groups includes a Patch Merging module and a second SwinT Block module group connected in series in sequence; the second SwinT Block module group includes a plurality of second SwinT Block modules connected in series (corresponding to Figure 4 SwinTBlock-2 in the Patch Partition module); the input of the Patch Partition module is the output of the first Stem module; the input of the feature fusion module is the output of the last first residual channel attention module in the CNN path and the output of the last PMSB module group in the Transformer path; the input of the first classification head module is the output of the feature fusion module, and the output of the first classification head module is the type of the multi-leaf collimator illumination error.

[0046] The input of the first Stem module is the dose difference map, the gamma distribution map and the multi-leaf collimator leaf position and sub-field weight map. Figure 4 As shown, before each image is input into the first Stem module, each image is first spliced. The splicing operation corresponds to Figure 4 C in.

[0047] Artificial intelligence technology has developed rapidly, especially deep learning has developed rapidly. As a popular deep learning network, convolutional neural network (CNN) has been widely used in various fields, such as detection, classification and segmentation. Transformer was first applied to natural language processing (NLP) as a potential deep learning network and achieved state-of-the-art performance. At present, the Transformer model has been applied in computer vision and performs well in many machine vision tasks. Although CNN has a strong ability in local feature extraction, it is unable to model long-term dependencies and capture global context representations. On the contrary, Transformer performs well in establishing global representations, but has limitations in local feature extraction. Therefore, in the error classification network model, the present invention adopts a post-fusion framework, combining two independent feature extraction paths based on convolutional neural network (CNN) and Swin Transformer (SwinT), such as Figure 4 As shown in Figure 2. The CNN path performs feature encoding and dimension compression through multiple residual channel attention modules (RCA-1). The RCA-1 module extracts high-level semantic features with lower spatial resolution and larger feature dimensions from these low-level feature maps. To reduce computational cost, the RCA-1 module adopts a bottleneck structure and expands the feature dimension through CBR module operations with a convolution kernel size of 1×1.

[0048] In the Transformer path, the input feature map is first split into smaller blocks through the Patch Partition module, and then these blocks are embedded into a high-dimensional space through the Linear Embedding module. Next, the feature map passes through multiple first SwinT Blocks in turn to model global context dependencies to enhance semantic representation. Different from the standard Transformer structure, the first SwinT Block module introduces a non-overlapping local window mechanism and realizes information interaction between windows through shift operations, thereby reducing computational complexity. The feature map processed by multiple first SwinT Blocks is processed by the Patch Merging module and the second SwinT Block module.

[0049] The outputs of these two feature extraction paths are merged in the feature fusion module. In the error classification network model, the global semantic representation of the Transformer path and the local detail information of the CNN path are enhanced. The fused feature map is further processed by feature fusion and finally sent to the decision stage. In the decision stage, the first classification head module maps the fused features to the prediction probability to determine the error type. Compared with the simple CNN or Transformer structure, the error classification network model reduces the dependence on a large amount of data by combining local and global semantic features, and improves the recognition efficiency and accuracy of multi-leaf collimator errors.

[0050] Implementing the above-mentioned steps 201 to 205, firstly, due to the limitation of low-resolution detectors, existing classification methods are difficult to accurately handle the error identification of millimeter-level multi-leaf collimator leaves. The present invention uses super-resolution reconstruction before classification, which improves the resolution of the error image and increases the resolution of the image to the millimeter level. Secondly, most of the current error classification methods use dose difference maps and gamma distribution maps as input. On this basis, the present invention creatively uses the multi-leaf collimator leaf position and sub-field weight map generated by the patient's radiotherapy plan file. Since the multi-leaf collimator leaf position and sub-field weight map can accurately guide the position of the leaves, it is effective for classifying the errors of the multi-leaf collimator leaf position.

[0051] Although many deep learning networks based on dose maps (including radiation measurement dose distribution maps and radiation calculation dose distribution maps) or gamma distribution maps can identify the error types of multi-leaf collimators, they are not very effective in dealing with complex and subtle error types. Based on the above considerations, in order to improve the accuracy of the network in multi-leaf collimator error identification, on the one hand, the input data is expanded to add the multi-leaf collimator leaf position information and sub-field weight data in the radiotherapy plan file, that is, the multi-leaf collimator leaf position and sub-field weight map. On the other hand, a super-resolution multi-channel hybrid neural network is designed, which integrates super-resolution reconstruction and classification functions. The super-resolution multi-channel hybrid neural network includes a super-resolution reconstruction model and an error classification network model.

[0052] In the super-resolution reconstruction of dose distribution map, the architecture adopts the design of generator and discriminator. The generator generates high-resolution dose distribution, and the discriminator judges the authenticity of the generated high-resolution dose distribution map, so as to achieve high-resolution reconstruction of dose distribution map. In the generator part, the low-resolution dose distribution map is first input (during training, it corresponds to the low-resolution radiation calculated dose distribution map, and after training, it corresponds to the low-resolution radiation measured dose distribution map). Figure 5 The primary feature map is then processed by multiple second residual channel attention modules (corresponding to Figure 5 As an example, RCA-2 in Figure 1 can be selected as 4 second residual channel attention modules to extract more advanced semantic features. The second residual channel attention module adopts a bottleneck design to save computing resources. Then, the feature map passes through the Patch Embedding module to split the feature map into smaller fixed-size blocks, and each block is embedded in a high-dimensional space to facilitate feature map segmentation and conversion into a form suitable for self-attention mechanism and multi-layer perceptron processing. Next, the feature map passes through multiple third SwinTBlock modules (corresponding to Figure 5 The feature map processed by the third SwinTBlock module needs to be adjusted for feature channels so that it can be added to the primary feature map. Finally, the first upsampling module (corresponding to Figure 5 The upsampling in -1) increases the resolution of the feature map and generates a high-resolution dose distribution map (during training, it corresponds to a high-resolution radiation calculation dose distribution map, and after training, it corresponds to a high-resolution radiation measurement dose distribution map).

[0053] like Figure 5As shown, the generator includes a second Stem module, a second residual channel attention module group, a Patch Embedding module, a third SwinT Block module group, a feature map channel adjustment module, a first addition layer and multiple first upsampling modules connected in series; the second residual channel attention module group includes multiple second residual channel attention modules connected in series; the third SwinT Block module group includes multiple third SwinTBlock modules connected in series; the input end of the first addition layer is also connected to the output end of the second Stem module. Among them, the feature map channel adjustment module includes a Reshape module, a CBR module with a convolution kernel size of 1x1 and an upsampling module connected in series to achieve the adjustment of the feature channel.

[0054] In an exemplary embodiment, the first residual channel attention module and the second residual channel attention module in the present invention may adopt the same structure, which are collectively referred to as residual channel attention modules. Figure 6 As shown, the residual channel attention module (RCA) includes a second CBR module (corresponding to Figure 6 CBR-2 in), the third CBR module (corresponding to Figure 6 CBR-3 in), the fourth CBR module (corresponding to Figure 6 CBR-4 in), Avepool layer, Linear layer, second activation layer (corresponding to Figure 6 Relu-2 in), Sigmod layer, multiplication layer, second addition layer and third activation layer (corresponding to Figure 6 The input end of the multiplication layer is also connected to the output of the fourth CBR module, and the input end of the second addition layer is also connected to the input end of the second CBR module.

[0055] In an exemplary embodiment, the first SwinT Block module, the second SwinT Block module and the third SwinT Block module in the present invention may adopt the same structure, which are collectively referred to as SwinT Block modules. Figure 7 As shown in Figure 1, the SwinT Block module includes two SwinT Block submodules. Each SwinT Block submodule includes two LayerNorm layers, a window multi-head self-attention module (WMSA), two addition layers, and a multi-layer perceptron (MLP) consisting of two linear layers and a GELU activation function. Specifically, the SwinT Block submodules include the first LayerNorm layer (corresponding to Figure 7LayerNorm-1 in ), window multi-head self-attention layer, third addition layer, second LayerNorm layer (corresponding to Figure 7 The input end of the third addition layer is also connected to the input end of the first LayerNorm layer, and the input end of the fourth addition layer is also connected to the output end of the third addition layer.

[0056] In the discriminator part, the input is the real and high-resolution dose distribution generated by the generator, which is first passed through the third Stem module (corresponding to Figure 5 Then, the feature map is sequentially processed through multiple first CBR modules (corresponding to Figure 5 As an example, 4 first CBR modules can be selected. The feature map processed by the first CBR module is passed through a second classification head module (corresponding to Figure 5 The classification head in -2) makes true or false judgments.

[0057] like Figure 5 As shown, the discriminator includes a third Stem module, a CBR module group and a second classification head module connected in series. The CBR module group includes a plurality of first CBR modules connected in series.

[0058] like Figure 8 As shown, the first CBR module includes first two-dimensional convolutional layers (corresponding to Figure 8 Conv2D-1 in ), the first batch normalization layer (corresponding to Figure 8 BatchNorm-1 in ) and the first activation layer (corresponding to Figure 8 Relu-1 in .

[0059] In an exemplary embodiment, the first Stem module, the second Stem module and the third Stem module in the present invention may adopt the same structure, and are collectively referred to as Stem modules herein. Fig. 9 As shown, the Stem module includes a second two-dimensional convolutional layer (corresponding to Fig. 9 Conv2D-2 in ), the second batch normalization layer (corresponding to Fig. 9 BatchNorm-2 in ) and the fourth activation layer (corresponding to Fig. 9 The structures of the Stem module and the CBR module are similar, except that the convolution kernel sizes are different.

[0060] In order to ensure that the feature map of the Transformer path is aligned with the feature map of the CNN path in the channel dimension, the feature map of the Transformer path will first undergo a reshape operation, and then pass through a CBR module with a 1×1 convolution kernel to adjust the channel. Then, an upsampling operation is performed to make the spatial resolution of the feature map consistent with the feature map of the CNN path. After completing these operations, the feature map of the Transformer path and the feature map of the CNN path will be spliced ​​and processed again through a CBR module with a 1×1 convolution kernel. In an exemplary embodiment, Fig.10 As shown, the feature fusion module includes a Reshape module, a fifth CBR module (corresponding to Fig.10 CBR-5 in), the second upsampling module (corresponding to Fig.10 Upsampling-2 in the splicing module (corresponding to Fig.10 C) and the sixth CBR module (corresponding to Fig.10 CBR-6 in the figure); the input of the Reshape module is the output of the Transformer path; the input end of the splicing module is also connected to the output end of the CNN path.

[0061] As an optional implementation, the first CBR module, the second CBR module, the third CBR module, the fourth CBR module, the fifth CBR module and the sixth CBR module in the present invention may adopt the same structure, such as Figure 8 The structure shown in . Depending on the requirements, the convolution kernel size of the two-dimensional convolution layer can be different.

[0062] In another exemplary embodiment of the present application, before executing step 202 "inputting the low-resolution radiation measurement dose distribution map into the trained super-resolution reconstruction model", it also includes: training the super-resolution reconstruction model, such as Fig.11 As shown, specifically including:

[0063] (1) Obtain low-resolution radiation calculated dose distribution map samples (i.e., first-resolution radiation calculated dose distribution map samples) and corresponding high-resolution radiation calculated dose distribution map samples (i.e., second-resolution radiation calculated dose distribution map samples).

[0064] First, the normal radiotherapy plan of the patient is collected in the radiotherapy planning system, and the radiation calculated dose is derived at high resolution and low resolution to form a super-resolution network training database. For example, the patient's radiotherapy plan is derived from the Monaco radiotherapy planning system developed by Elekta using the Digital Image Storage and Communication in Medicine (DICOM) standard to form a DICOM-RTplan file and two types of radiation calculated dose distributions at high resolution and low resolution. The planning system is the (Monaco) planning system developed by Elekta, which can be extended to other planning systems.

[0065] (2) Using the low-resolution radiation calculated dose distribution map samples as input and the high-resolution radiation calculated dose distribution map samples as labels, the super-resolution reconstruction model is trained until the loss error of the super-resolution reconstruction model converges or the number of model training times reaches the corresponding maximum number of iterations, thereby obtaining a trained super-resolution reconstruction model.

[0066] Fig.11 In the text, low-resolution calculated dose and high-resolution calculated dose refer to low-resolution radiation calculated dose distribution map samples and high-resolution radiation calculated dose distribution map samples, respectively.

[0067] In another exemplary embodiment of the present application, before executing step 205 "inputting the dose difference map, the gamma distribution map, and the multi-leaf collimator leaf position and sub-field weight map into the trained error classification network model", it also includes: training the error classification network model, such as Fig.11 As shown, specifically including:

[0068] (1) Obtaining a radiation calculation dose distribution map sample and a corresponding low-resolution radiation measurement dose distribution map sample (i.e., a first-resolution radiation measurement dose distribution map sample) in a radiotherapy plan file sample and a modified radiotherapy plan file sample; the modified radiotherapy plan file sample refers to a radiotherapy plan file in which the error type is pre-marked in the radiotherapy plan file sample.

[0069] To perform image conversion on the radiation dose distribution map, in order to ensure the invariance of the converted features, it is necessary to set various parameters consistently during the conversion and normalize the converted feature results to an integer range of [0,255].

[0070] By modifying the DICOM-RTplan file, multi-leaf collimator errors are artificially added to all control points of the irradiation field in the patient's radiotherapy plan, including multi-leaf collimator position errors (such as translation errors, random errors, expansion errors, adduction errors, etc.) and multi-leaf collimator parameter errors (such as concave and convex groove modeling errors, leaf gap errors, etc.). For each error, the error type or magnitude added can be artificially defined.

[0071] The patient's radiotherapy plan file (DICOM-RTplan file) is transferred to the tumor information management system, the medical electron linear accelerator is controlled to perform irradiation to execute the radiotherapy plan, and the PTW 729 two-dimensional low-resolution detector is used to complete the measurement to obtain a low-resolution radiation measurement dose distribution map.

[0072] (2) Inputting the low-resolution radiation measurement dose distribution map sample into the trained super-resolution reconstruction model to obtain a high-resolution radiation measurement dose distribution map sample (i.e., a second-resolution radiation measurement dose distribution map sample). Fig.11 In the present invention, the calculated dose refers to the radiation calculated dose distribution map sample, the low-resolution measured dose and the high-resolution measured dose refer to the low-resolution radiation measured dose distribution map sample and the high-resolution radiation measured dose distribution map sample, respectively.

[0073] (3) Obtaining a sample of multi-leaf collimator leaf positions and sub-field weight maps based on the modified radiotherapy plan file sample.

[0074] The leaf position information in the radiotherapy plan (DICOM-RTplan) is read using the standard DICOM protocol and normalized according to the sub-field dose to form a multi-leaf collimator leaf position and sub-field weight map, such as Fig.12 shown. Fig.12 In the figure, the black area represents the non-irradiated area, and the white area represents the irradiated area. The closer the color is to white, the higher the radiation dose.

[0075] (4) Based on the high-resolution radiation measurement dose distribution map sample and the corresponding radiation calculation dose distribution map sample, a standard gamma method is applied to obtain a dose difference map sample and a gamma distribution map sample.

[0076] (5) Using the dose difference map samples, the gamma distribution map samples, and the multi-leaf collimator leaf position and sub-field weight map samples as input, and using the error types marked in the modified radiotherapy plan file samples as labels to train an error classification network model until the loss error of the error classification network model converges or the number of model training times reaches the corresponding maximum number of iterations, thereby obtaining the trained error classification network model.

[0077] The present application also provides an application scenario, which applies the above-mentioned multi-leaf collimator irradiation error classification method based on deep learning. Specifically: The multi-leaf collimator irradiation error classification method based on deep learning provided in this embodiment can be applied in the radiotherapy equipment quality control scenario. 1. After the doctor formulates an accurate treatment plan, the method of the present invention can be used to predict and verify the settings of the multi-leaf collimator, identify possible errors, and thus adjust the treatment plan to ensure the accuracy of the multi-leaf collimator irradiation. 2. When the multi-leaf collimator performs irradiation control, the action of the multi-leaf collimator is monitored in real time, and the method of the present invention is used to timely detect and classify any possible multi-leaf collimator irradiation errors. In this way, adjustments can be made immediately when errors occur to ensure the accuracy of multi-leaf collimator irradiation. 3. After the multi-leaf collimator completes irradiation control, the method of the present invention is used to evaluate and classify irradiation errors through post-analysis of irradiation data, providing a reference for subsequent optimization of radiotherapy files.

[0078] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.13 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store multi-leaf collimator irradiation error classification data based on deep learning. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a multi-leaf collimator irradiation error classification method based on deep learning is implemented.

[0079] Those skilled in the art will understand that Fig.13 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0080] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0081] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0082] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0083] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0084] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0085] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0086] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A multi-leaf collimator irradiation error classification method based on deep learning, characterized in that: include: Obtaining a first-resolution radiation measurement dose distribution map and a corresponding radiation calculation dose distribution map output by a two-dimensional detector in a radiotherapy plan file; Inputting the first resolution radiation measurement dose distribution map into a trained super-resolution reconstruction model to obtain a second resolution radiation measurement dose distribution map; the super-resolution reconstruction model includes a generator and a discriminator connected in series; The resolution of the second-resolution radiation measurement dose distribution map is higher than the resolution of the first-resolution radiation measurement dose distribution map; According to the second resolution radiation measurement dose distribution map and the corresponding radiation calculation dose distribution map, a standard gamma method is applied to obtain a dose difference map and a gamma distribution map; The multi-leaf collimator leaf positions and sub-field weight maps are obtained based on the radiotherapy planning file; The dose difference map, the gamma distribution map and the multi-leaf collimator leaf position and sub-field weight map are input into the trained error classification network model to obtain the type of multi-leaf collimator irradiation error; the error classification network model includes a first Stem module, a CNN path, a Transformer path, a feature fusion module and a first classification head module; the CNN path includes a plurality of first residual channel attention module groups connected in series; each of the first residual channel attention module groups includes a plurality of first residual channel attention modules connected in series; the input of the first first residual channel attention module in the CNN path is the output of the first Stem module; the Transformer path includes a Patch Partition module, a Linear Embedding module, a first SwinT Block module group and a plurality of PMSB module groups connected in series in sequence; the first SwinT Block module group includes a plurality of first SwinT Block modules connected in series in sequence; each of the PMSB module groups includes a Patch Merging module and a second SwinT Block module group connected in series in sequence; the second SwinT Block module group includes a plurality of second SwinT Block modules connected in series; the Patch The input of the Partition module is the output of the first Stem module; the input of the feature fusion module is the output of the last first residual channel attention module in the CNN path and the output of the last PMSB module group in the Transformer path; the input of the first classification head module is the output of the feature fusion module, and the output of the first classification head module is the type of the multi-leaf collimator irradiation error.

2. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: The generator includes a second Stem module, a second residual channel attention module group, a PatchEmbedding module, a third SwinT Block module group, a feature map channel adjustment module, a first addition layer and a plurality of first upsampling modules connected in series in sequence; the second residual channel attention module group includes a plurality of second residual channel attention modules connected in series in sequence; the third SwinT Block module group includes a plurality of third SwinT Block modules connected in series in sequence; the input end of the first addition layer is also connected to the output end of the second Stem module; The discriminator includes a third Stem module, a CBR module group and a second classification head module connected in series in sequence; the CBR module group includes a plurality of first CBR modules connected in series in sequence; the first CBR module includes a two-dimensional convolutional layer, a batch normalization layer and a first activation layer connected in series in sequence.

3. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: The first residual channel attention module includes a second CBR module, a third CBR module, a fourth CBR module, an Avepool layer, a Linear layer, a second activation layer, a Sigmod layer, a multiplication layer, a second addition layer and a third activation layer connected in series in sequence; The input end of the multiplication layer is also connected to the output of the fourth CBR module, and the input end of the second addition layer is also connected to the input end of the second CBR module.

4. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: The first SwinT Block module includes two SwinT Block submodules connected in series in sequence; each of the SwinTBlock submodules includes a first LayerNorm layer, a window multi-head self-attention layer, a third addition layer, a second LayerNorm layer, a multi-layer perceptron and a fourth addition layer connected in series in sequence; The input end of the third addition layer is also connected to the input end of the first LayerNorm layer, and the input end of the fourth addition layer is also connected to the output end of the third addition layer.

5. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: The feature fusion module includes a Reshape module, a fifth CBR module, a second upsampling module, a splicing module and a sixth CBR module which are connected in series in sequence; the input of the Reshape module is the output of the Transformer path; and the input end of the splicing module is also connected to the output end of the CNN path.

6. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: Before executing the step of "inputting the first resolution radiation measurement dose distribution map into the trained super-resolution reconstruction model", the method further includes: training the super-resolution reconstruction model, specifically including: Acquire a first-resolution radiation calculated dose distribution map sample and a corresponding second-resolution radiation calculated dose distribution map sample; the resolution of the second-resolution radiation calculated dose distribution map sample is higher than the resolution of the first-resolution radiation calculated dose distribution map sample; The super-resolution reconstruction model is trained using the first-resolution radiation calculated dose distribution map samples as input and the second-resolution radiation calculated dose distribution map samples as labels until the loss error of the super-resolution reconstruction model converges or the number of model training times reaches the corresponding maximum number of iterations, thereby obtaining a trained super-resolution reconstruction model.

7. The multi-leaf collimator irradiation error classification method based on deep learning according to claim 1, characterized in that: Before executing the step of "inputting the dose difference map, the gamma distribution map, and the multi-leaf collimator leaf position and sub-field weight map into the trained error classification network model", the step further includes: training the error classification network model, specifically including: Obtaining a radiation calculation dose distribution map sample and a corresponding first resolution radiation measurement dose distribution map sample and a modified radiotherapy plan file sample in the radiotherapy plan file sample; the modified radiotherapy plan file sample refers to a radiotherapy plan file in which an error type is pre-marked in the radiotherapy plan file sample; Inputting the first-resolution radiation measurement dose distribution map sample into the trained super-resolution reconstruction model to obtain a second-resolution radiation measurement dose distribution map sample; the resolution of the second-resolution radiation measurement dose distribution map sample is higher than the resolution of the first-resolution radiation measurement dose distribution map sample; Obtaining a sample of multi-leaf collimator leaf positions and sub-field weight maps according to the modified radiotherapy plan file sample; According to the second resolution radiation measurement dose distribution map sample and the corresponding radiation calculation dose distribution map sample, a standard gamma method is applied to obtain a dose difference map sample and a gamma distribution map sample; The dose difference map samples, the gamma distribution map samples and the multi-leaf collimator leaf position and sub-field weight map samples are used as input, and the error types marked in the modified radiotherapy plan file samples are used as labels to train the error classification network model until the loss error of the error classification network model converges or the number of model training times reaches the corresponding maximum number of iterations, thereby obtaining the trained error classification network model.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-leaf collimator irradiation error classification method based on deep learning as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the deep learning-based multi-leaf collimator irradiation error classification method described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the deep learning-based multi-leaf collimator irradiation error classification method described in any one of claims 1 to 7 is implemented.

Images