Automatic positioning method, device and radiotherapy equipment based on anatomical structure detection

Abstract

The present invention discloses an automatic positioning method, device and radiotherapy equipment based on anatomical structure detection. The method includes the following steps: collecting image data, performing data annotation, and training the annotated image data using a deep convolutional neural network to obtain DR image training models and DRR image training models at various angles; respectively sending DR images and DRR images at multiple angles into the DR image training model and DRR image training model at the corresponding angles, detecting the DR anatomical structure and the DRR anatomical structure, subtracting the coordinates of the DR anatomical structure and the DRR anatomical structure at the same angle to obtain the offset at the angle, and the final positioning offset is based on the offset at the angle with the best imaging effect, and automatically aligning the corresponding DR image and DRR image. The present invention uses a deep learning algorithm to automatically detect anatomical structure feature points, automatically align the corresponding DR image and DRR image, reduce the technician's operating time, and improve the effect of tumor radiotherapy.

Description

Technical Field

[0001] The present invention belongs to the technical field of accelerator radiotherapy, and in particular relates to an automatic positioning method and device based on anatomical structure detection, and radiotherapy equipment. Background Art

[0002] Medical linear accelerators are the primary equipment for tumor radiotherapy, and most are equipped with an EPID. When using a medical linear accelerator for tumor radiotherapy, cancer patients typically undergo several radiation exposures, and accurate positioning during treatment is crucial to ensuring effective treatment.

[0003] A linear accelerator equipped with EPID can use DR images taken before the patient's treatment to verify positioning. However, the image quality produced by MV-level rays is relatively average, requiring the operating technician to perform meticulous window adjustment to align with the DRR image. In addition, the image quality of the DRR image is not particularly good due to factors such as the layer thickness of the CT image.

[0004] Each fraction of tumor radiotherapy typically takes about 10 minutes. Technicians manually position the patient using EPID before treatment. They capture two orthogonal DR (Digital Radiography) images, typically at 0, 90, or 270 degrees. These images are then compared with the DRR (Digital Reconstructed Radiography) images generated by CT scans in the corresponding directions. Anatomical structures that are clearly visible in both images are selected for alignment. Manual registration of DR and DRR images is a significant effort. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes an automatic positioning method, device and radiotherapy equipment based on anatomical structure detection, which uses a deep learning algorithm to automatically detect anatomical structure feature points and automatically align the corresponding DR images and DRR images, reducing the technician's operation time and improving the effect of tumor radiotherapy.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] In one aspect, the present invention discloses an automatic positioning method based on anatomical structure detection, comprising the following steps:

[0008] Collecting a number of image data, the image data including: DR images at various angles and CT images and / or DRR images at various angles, where the CT images can generate corresponding DRR images;

[0009] Annotate the collected image data and mark the clear anatomical structures in the DR images and DRR images at each angle;

[0010] Use deep convolutional neural networks to train the labeled image data to obtain DR image training models and DRR image training models for each angle;

[0011] DR images from multiple angles are fed into the DR image training model of the corresponding angles to detect the corresponding DR anatomical structures in the DR images from the multiple angles. DRR images from multiple angles are fed into the DRR image training model of the corresponding angles to detect the corresponding DRR anatomical structures in the DRR images from the multiple angles. The coordinates of the DR anatomical structures obtained by detecting the DR images and DRR images at the same angle are subtracted from the coordinates of the DRR anatomical structures to obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect.

[0012] The corresponding DR and DRR images are automatically aligned according to the final positioning offset.

[0013] On the basis of the above technical solution, the following improvements can be made:

[0014] As a preferred solution, before labeling the image data, the process further includes: cleaning the image data.

[0015] As a preferred solution, the specific steps for training the labeled image data using a deep convolutional neural network are as follows:

[0016] The labeled image data is divided into training set and test set according to the proportion;

[0017] The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value;

[0018] The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

[0019] As a preferred solution, the DR images, CT images and DRR images are files in DICOM format.

[0020] As a preferred solution, the backbone of the deep convolutional neural network uses a feature pyramid network as a feature extractor.

[0021] As a preferred solution, the bottom-up path of the first half of the deep convolutional neural network includes four stages, namely the first stage, the second stage, the third stage and the fourth stage. Each stage uses a ResNet block. In each stage, a feature map of the same size is generated and features are extracted from it. The resolution of the feature map in the first stage is 4 times lower than the input image resolution, the resolution of the feature map in the second stage is 8 times lower than the input image resolution, the resolution of the feature map in the third stage is 16 times lower than the input image resolution, and the resolution of the feature map in the fourth stage is 32 times lower than the input image resolution. For every doubling of the resolution, the number of channels is doubled accordingly.

[0022] Another top-down path is used to upsample the feature map, using nearest neighbor upsampling. The upsampled feature map is reduced to 256 channels through a 1×1 convolutional layer, and then merged with the corresponding bottom-up feature map through element-by-element addition;

[0023] The second half of the deep convolutional neural network is used to predict the coordinates of organ anatomical structures. The feature map obtained by the first half of the deep convolutional neural network is first passed through two 3×3 convolutional layers to unify the channel dimension to 128. Then, a nearest neighbor upsampling layer is used to adjust the feature map to the same resolution as the input image, and then it is concatenated into a 512-dimensional feature map.

[0024] Finally, the deep convolutional neural network outputs a heat map of the anatomical structure position, and the bright spots on the heat map represent the position of the anatomical structure. On the other hand, the present invention discloses an automatic positioning device based on anatomical structure detection, comprising:

[0025] An image data collection device for collecting image data of a plurality of patients, wherein the image data includes: DR images at various angles and CT images and / or DRR images at various angles, wherein the CT images can generate corresponding DRR images;

[0026] An image data processing device, used to annotate the collected image data and annotate clear anatomical structures in the DR images and DRR images at various angles;

[0027] An image data training device is used to train the labeled image data using a deep convolutional neural network to obtain DR image training models and DRR image training models for each angle;

[0028] The final positioning offset obtaining device is used to send DR images at multiple angles into the DR image training model of the corresponding angles for training, detect the corresponding DR anatomical structures in the DR images at multiple angles, send DRR images at multiple angles into the DRR image training model of the corresponding angles for training, detect the corresponding DRR anatomical structures in the DRR images at multiple angles, subtract the coordinates of the DR anatomical structure detected after the DR image and the DRR image training at the same angle from the coordinates of the DRR anatomical structure, and obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect.

[0029] The automatic positioning device automatically aligns the corresponding DR image and DRR image according to the final positioning offset.

[0030] As a preferred solution, the image data processing device also performs data cleaning on the image data before performing data labeling on the image data.

[0031] As a preferred solution, the specific steps of the image data training device using a deep convolutional neural network to train the labeled image data are as follows:

[0032] The labeled image data is divided into training set and test set according to the proportion;

[0033] The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value;

[0034] The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

[0035] On the other hand, the present invention discloses a radiotherapy device, which is used to implement any of the above-mentioned automatic positioning methods based on anatomical structure detection, or the radiotherapy device includes any of the above-mentioned automatic positioning devices based on anatomical structure detection.

[0036] The present invention uses a deep learning algorithm to automatically detect anatomical structure feature points and automatically align the corresponding DR images and DRR images, reducing the technician's operation time and improving the effect of tumor radiotherapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 A flowchart of an automatic positioning method based on anatomical structure detection provided in an embodiment of the present invention.

[0039] Figure 2 This is the image registration interface of the IVS (Image Viewing System) software system provided in an embodiment of the present invention.

[0040] Figure 3 A 0-degree DR image of a chest tumor patient with tracheal bifurcation annotation provided in an embodiment of the present invention.

[0041] Figure 4 A 0-degree DRR image of a chest tumor patient with tracheal bifurcation annotation provided in an embodiment of the present invention.

[0042] Figure 5 A 90-degree DR image of a chest tumor patient with sternal angle annotation provided in an embodiment of the present invention.

[0043] Figure 6 A 90-degree DRR image of a chest tumor patient with sternal angle annotation provided in an embodiment of the present invention.

[0044] Figure 7 A schematic diagram of the structure of a deep convolutional neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] The expression “including” an element is an “open” expression, which simply means that corresponding components exist and should not be interpreted as excluding additional components.

[0048] In order to achieve the purpose of the present invention, in some embodiments of the automatic positioning method, device and radiotherapy equipment based on anatomical structure detection, as Figure 1 As shown, the present invention discloses an automatic positioning method based on anatomical structure detection, which includes the following steps:

[0049] Collecting a number of image data, the image data including: DR images at various angles and CT images and / or DRR images at various angles, where the CT images can generate corresponding DRR images;

[0050] Annotate the collected image data and mark the clear anatomical structures in the DR images and DRR images at each angle;

[0051] Use deep convolutional neural networks to train the labeled image data to obtain DR image training models and DRR image training models for each angle;

[0052] DR images from multiple angles are fed into the DR image training model of the corresponding angles to detect the corresponding DR anatomical structures in the DR images from the multiple angles. DRR images from multiple angles are fed into the DRR image training model of the corresponding angles to detect the corresponding DRR anatomical structures in the DRR images from the multiple angles. The coordinates of the DR anatomical structures obtained by detecting the DR images and DRR images at the same angle are subtracted from the coordinates of the DRR anatomical structures to obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect.

[0053] The corresponding DR and DRR images are automatically aligned according to the final positioning offset.

[0054] Among them, the DR images and CT images at each angle and / or the DRR images at each angle can be 0-degree DR images, 90-degree / 270-degree DR images and 0-degree DRR images, 90-degree / 270-degree DRR images, and the 90-degree / 270-degree DRR images can be generated from 90-degree / 270-degree CT images.

[0055] It is worth noting that the final positioning offset includes: positioning direction and specific offset.

[0056] In order to further optimize the implementation effect of the present invention, in some other implementations, the remaining feature technologies are the same, except that before data labeling of the image data, it also includes: data cleaning of the image data.

[0057] The specific contents of data cleaning include one or more of the following: checking whether the data images are compatible and whether they are clinical patient images, removing images that do not meet the requirements, and classifying patient images according to the categories of head, neck, chest, abdomen, and pelvis.

[0058] In order to further optimize the implementation effect of the present invention, in some other embodiments, the remaining feature technologies are the same, except that the specific steps of using a deep convolutional neural network to train the labeled image data are as follows:

[0059] The labeled image data is divided into training set and test set according to the proportion;

[0060] The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value;

[0061] The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

[0062] Furthermore, the DR image, CT image, and DRR image are files in DICOM format.

[0063] DICOM: Digital Imaging and Communication in Medicine, is an international standard for medical images and related information (ISO 12052), which defines a medical image format that can be used for data exchange with quality that meets clinical needs.

[0064] Furthermore, the image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed including:

[0065] Convert image data from DICOM format to PNG format, crop the middle part of the image in a unified way and transform it to the size of network input.

[0066] In order to further optimize the implementation effect of the present invention, in some other embodiments, the other feature technologies are the same, except that the backbone of the deep convolutional neural network uses a feature pyramid network as a feature extractor. Figure 7As shown in Figure 1, the backbone of this network uses a Feature Pyramid Network (FPN) as a feature extractor. The bottom-up path (denoted by C) in the first half of the deep convolutional neural network consists of four stages, each using a ResNet block. At each stage, the network produces feature maps of the same size and extracts features from them, representing {C1, C2, C3, C4}. The resolution of the feature maps is reduced by a factor of (4, 8, 16, 32) compared to the input image resolution. For every doubling of the resolution, the number of channels increases by a factor of two. Another top-down path (denoted by P) upsamples the feature maps {P2, P3, P4} using nearest neighbor upsampling. The upsampled feature maps are then passed through a 1×1 convolutional layer to reduce the number of channels to 256. They are then merged with the corresponding bottom-up feature maps through element-wise addition. Furthermore, the bottom-up feature maps are passed through a 1×1 convolutional layer before addition to maintain the same number of channels as the top-down feature maps.

[0067] The second half of this network is used for organ bifurcation point prediction. The feature map of the backbone network is first passed through two 3×3 convolutional layers to unify the channel dimension to 128, and then passed through a nearest neighbor upsampling layer to adjust the feature map to the same resolution as the input image. Represent this process, and then connect them into a 512-dimensional feature map (denoted by D). This architecture combines low-resolution, semantically strong features with high-resolution, semantically weak features through hierarchical stages and lateral connections.

[0068] Finally, the network outputs a heat map of the anatomical structure location, where the bright spots on the heat map represent the location of the anatomical structure.

[0069] The above multiple implementations can be implemented in parallel.

[0070] On the other hand, an embodiment of the present invention further discloses an automatic positioning device based on anatomical structure detection, comprising:

[0071] An image data collection device for collecting image data of a plurality of patients, wherein the image data includes: DR images at various angles and CT images and / or DRR images at various angles, wherein the CT images can generate corresponding DRR images;

[0072] An image data processing device, used to annotate the collected image data and annotate clear anatomical structures in the DR images and DRR images at various angles;

[0073] An image data training device is used to train the labeled image data using a deep convolutional neural network to obtain DR image training models and DRR image training models for each angle;

[0074] The final positioning offset obtaining device is used to send DR images at multiple angles into the DR image training model of the corresponding angles for training, detect the corresponding DR anatomical structures in the DR images at multiple angles, send DRR images at multiple angles into the DRR image training model of the corresponding angles for training, detect the corresponding DRR anatomical structures in the DRR images at multiple angles, subtract the coordinates of the DR anatomical structure detected after the DR image and the DRR image training at the same angle from the coordinates of the DRR anatomical structure, and obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect.

[0075] The automatic positioning device automatically aligns the corresponding DR image and DRR image according to the final positioning offset.

[0076] In order to further optimize the implementation effect of the present invention, in some other implementations, the remaining feature technologies are the same, except that the image data processing device also performs data cleaning on the image data before data labeling.

[0077] In order to further optimize the implementation effect of the present invention, in some other embodiments, the other feature technologies are the same, except that the specific steps of the image data training device using the deep convolutional neural network to train the labeled image data are as follows:

[0078] The labeled image data is divided into training set and test set according to the proportion;

[0079] The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value;

[0080] The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

[0081] On the other hand, the present invention discloses a radiotherapy device, which is used to implement the automatic positioning method based on anatomical structure detection disclosed in any of the above embodiments, or the radiotherapy device includes the automatic positioning device based on anatomical structure detection disclosed in any of the above embodiments.

[0082] The present invention uses a deep learning algorithm to automatically detect anatomical structure feature points and automatically align the corresponding DR images and DRR images, reducing the technician's operation time and improving the effect of tumor radiotherapy.

[0083] The above multiple implementations can be implemented in parallel.

[0084] In order to better understand the technical content of the present invention, a specific embodiment is introduced below.

[0085] In this specific embodiment, in addition to the accelerator and the EPID on the accelerator, an IVS (Image Viewing System) software system developed by a company in the prior art is also applied. Figure 2 This is the image registration interface of the IVS software system. The picture shows a chest patient. Figure 2 The upper left image in the center is a 0-degree DR image, the lower left image is a 90-degree DR image, the upper right image is a 0-degree bone DRR image, and the lower right image is a 90-degree bone DRR image. The two center images are overlaid DR and DRR images from two different angles. From top to bottom, the leftmost area of the interface contains the patient information area, the general tools area, the registration tools area, the offset and approval tools area, the offset area for images at different angles, and the bed offset area.

[0086] The automatic positioning method based on anatomical structure detection disclosed in the present invention specifically includes the following steps:

[0087] Step 1: Collect imaging data of chest tumor patients from the IVS software system. The imaging data include: 0-degree DR image, 90-degree DR image, 0-degree DRR image, and 90-degree DRR image.

[0088] Step 2: Use the IVS software system to clean and annotate the collected image data. In the 0-degree DR image and 0-degree DRR image, the clear anatomical structure is the tracheal bifurcation; in the 90-degree DR image and 90-degree DRR image, the clear anatomical structure is the Louis angle (i.e., the angle formed by the manubrium and the body of the sternum). The clear anatomical structures in the DR images and DRR images at various angles are annotated, such as Figure 3-6 As shown;

[0089] Step 3: Use a deep convolutional neural network to train the labeled image data to obtain a 0-degree DR image training model, a 90-degree DR image training model, a 0-degree DRR image training model, and a 90-degree DRR image training model;

[0090] Step 4: After training the four models, the models are integrated into the IVS software system using an inference process. The input of the inference process is the 0-degree DR image, 90-degree DR image, 0-degree DRR image, and 90-degree DRR image taken during one treatment fraction of a tumor patient.

[0091] Step 4.1: Preprocess the 0-degree DR image, i.e., scale it to the specified size and feed it into the trained corresponding model to detect the tracheal bifurcation feature point in the image. Scale the 0-degree lung window DRR image and feed it into the corresponding model to detect the tracheal bifurcation point. Perform corresponding operations on the 90-degree DR image and the bone window DRR image to detect the sternal angle.

[0092] Step 4.2: Subtract the coordinates of the two tracheal bifurcation points in the 0-degree DR image and the 0-degree DRR image to obtain the offset x in the LAT direction and the offset y1 in the LNG direction. Subtract the coordinates of the two sternal angles in the 90-degree DR image and the 90-degree DRR image to obtain the offset z in the VRT direction and the offset y2 in the LNG direction.

[0093] Since the imaging quality of the human chest at 0 degrees is better and the detection effect is more accurate, the LNG direction offset y=y1 is selected;

[0094] Step 5: Automatically align the corresponding DR image and DRR image according to the LNG direction offset y=y1.

[0095] Among them: LAT, LNG, VRT are the abbreviations of lateral, longitudinal, and vertical, respectively, that is, the left and right directions, head and feet directions, and front and back directions when the patient is in the supine position.

[0096] In Step 3, four models were trained on 0-degree DR images, 0-degree DRR images, 90-degree DR images, and 90-degree DR images of patients with chest tumors. Taking 0-degree DR images as an example, the input of the training process is several 0-degree DR images of patients with chest tumors and their annotated tracheal bifurcation coordinate points. The specific steps are as follows:

[0097] Step 3.1: Divide the 0-degree DR images and their annotations into training and test sets according to appropriate ratios;

[0098] Step 3.2: Preprocess the images and annotations of the training and test sets, including converting them from DICOM format to PNG format, cropping the center of the image in a unified manner and transforming it to the network input size, and applying a Gaussian blur with a fixed variance to the image containing only the annotation points as the training ground truth. The images and annotations are then fed into the training program to train the model weights.

[0099] Step 3.3: Use the preprocessed test set data to test the offset error between the coordinates of the tracheal bifurcation point detected by the trained model and the true value point; if the iterative stopping condition is not met, such as the error can be further reduced, or the specified number of training iterations has not been reached, continue training in Step 3.2; if the iterative stopping condition is met, stop training and save the trained weights.

[0100] Currently, there are other hardware configurations for automatic registration, such as Cyberknife, BrainLab, and Rayer in China. However, compared to EPID, these systems require additional hardware configurations (such as BrainLab and Rayer) or are not run on a standard linear accelerator (such as Cyberknife). There is no automatic registration system for linear accelerators equipped with EPID, so the present invention is innovative.

[0101] The present invention provides an automatic positioning method, device and radiotherapy equipment based on anatomical structure detection, which uses a deep learning algorithm to automatically detect anatomical structure feature points and automatically align the corresponding DR images and DRR images, reducing the technician's operation time and improving the effect of tumor radiotherapy.

[0102] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable ordinary technicians in this field to understand the content of the present invention and implement it. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. The automatic positioning method based on anatomical structure detection is characterized by: The following steps are involved: Collecting a plurality of image data, wherein the image data includes: DR images at various angles and CT images and / or DRR images at various angles, wherein the CT images can generate corresponding DRR images; Annotate the collected image data and mark the clear anatomical structures in the DR images and DRR images at each angle; Use deep convolutional neural networks to train the labeled image data to obtain DR image training models and DRR image training models for each angle; DR images from multiple angles are fed into the DR image training model of the corresponding angles to detect the corresponding DR anatomical structures in the DR images from the multiple angles. DRR images from multiple angles are fed into the DRR image training model of the corresponding angles to detect the corresponding DRR anatomical structures in the DRR images from the multiple angles. The coordinates of the DR anatomical structures obtained by detecting the DR images and DRR images at the same angle are subtracted from the coordinates of the DRR anatomical structures to obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect. Automatically align the corresponding DR image and DRR image according to the final position offset; The backbone of the deep convolutional neural network uses a feature pyramid network as a feature extractor; The bottom-up path of the first half of the deep convolutional neural network includes four stages, namely the first stage, the second stage, the third stage and the fourth stage. Each stage uses a ResNet block. In each stage, a feature map of the same size is generated and features are extracted from it. The resolution of the feature map in the first stage is 4 times lower than the input image resolution. The resolution of the feature map in the second stage is 8 times lower than the input image resolution. The resolution of the feature map in the third stage is 16 times lower than the input image resolution. The resolution of the feature map in the fourth stage is 32 times lower than the input image resolution. For every doubling of the resolution, the number of channels is doubled accordingly. Another top-down path is used to upsample the feature map, using nearest neighbor upsampling. The upsampled feature map is reduced to 256 channels through a 1×1 convolutional layer, and then merged with the corresponding bottom-up feature map through element-by-element addition; The second half of the deep convolutional neural network is used to predict the coordinates of organ anatomical structures. The feature map obtained by the first half of the deep convolutional neural network is first passed through two 3×3 convolutional layers to unify the channel dimension to 128. Then, the feature map is adjusted to the same resolution as the input image through a nearest neighbor upsampling layer, and then concatenated into a 512-dimensional feature map. Finally, the deep convolutional neural network outputs a heat map of the anatomical structure location, where the bright spots on the heat map represent the locations of the anatomical structures.

2. The automatic positioning method according to claim 1, characterized in that: Before the image data is labeled, the method further includes: performing data cleaning on the image data.

3. The automatic positioning method according to claim 1, characterized in that: The specific steps for training the labeled image data using a deep convolutional neural network are as follows: The labeled image data is divided into training set and test set according to the proportion; The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value; The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

4. The automatic positioning method according to claim 3, characterized in that: The DR images, CT images and DRR images are files in DICOM format.

5. The automatic positioning device based on anatomical structure detection is characterized by: include: An image data collection device for collecting a plurality of patient image data, wherein the image data includes: DR images at various angles and CT images and / or DRR images at various angles, wherein the CT images can generate corresponding DRR images; An image data processing device, used to annotate the collected image data and annotate clear anatomical structures in the DR images and DRR images at various angles; An image data training device is used to train the labeled image data using a deep convolutional neural network to obtain DR image training models and DRR image training models for each angle; The final positioning offset obtaining device is used to send DR images at multiple angles into the DR image training model of the corresponding angles for training, detect the corresponding DR anatomical structures in the DR images at multiple angles, send DRR images at multiple angles into the DRR image training model of the corresponding angles for training, detect the corresponding DRR anatomical structures in the DRR images at multiple angles, subtract the coordinates of the DR anatomical structure detected after the DR image and the DRR image training at the same angle from the coordinates of the DRR anatomical structure, and obtain the offset at that angle. The final positioning offset is based on the offset at the angle with the best imaging effect. Automatic positioning device, automatically aligning the corresponding DR image and DRR image according to the final positioning offset; The backbone of the deep convolutional neural network uses a feature pyramid network as a feature extractor; The bottom-up path of the first half of the deep convolutional neural network includes four stages, namely the first stage, the second stage, the third stage and the fourth stage. Each stage uses a ResNet block. In each stage, a feature map of the same size is generated and features are extracted from it. The resolution of the feature map in the first stage is 4 times lower than the input image resolution. The resolution of the feature map in the second stage is 8 times lower than the input image resolution. The resolution of the feature map in the third stage is 16 times lower than the input image resolution. The resolution of the feature map in the fourth stage is 32 times lower than the input image resolution. For every doubling of the resolution, the number of channels is doubled accordingly. Another top-down path is used to upsample the feature map, using nearest neighbor upsampling. The upsampled feature map is reduced to 256 channels through a 1×1 convolutional layer, and then merged with the corresponding bottom-up feature map through element-by-element addition; The second half of the deep convolutional neural network is used to predict the coordinates of organ anatomical structures. The feature map obtained by the first half of the deep convolutional neural network is first passed through two 3×3 convolutional layers to unify the channel dimension to 128. Then, the feature map is adjusted to the same resolution as the input image through a nearest neighbor upsampling layer, and then concatenated into a 512-dimensional feature map. Finally, the deep convolutional neural network outputs a heat map of the anatomical structure location, where the bright spots on the heat map represent the locations of the anatomical structures.

6. The automatic positioning device according to claim 5, characterized in that: The image data processing device further performs data cleaning on the image data before performing data labeling on the image data.

7. The automatic positioning device according to claim 5, characterized in that: The specific steps of the image data training device using a deep convolutional neural network to train the labeled image data are as follows: The labeled image data is divided into training set and test set according to the proportion; The image data of the training set and the test set and the corresponding anatomical structure annotations are preprocessed, and the images containing only the anatomical structure annotations are processed to obtain the anatomical structure training truth value; The preprocessed training set image data is sent to the deep convolutional neural network, and the model weights are iteratively trained. The model weights are obtained after each iteration. The preprocessed test set image data is sent to the deep convolutional neural network to test the offset error between the anatomical structure coordinates detected by the model obtained in this training and the true value coordinates of the anatomical structure training. If the number of iterations reaches the training iteration number threshold or the offset error is less than the coordinate offset error threshold, the iteration is stopped and the final model weights are saved.

8. Radiotherapy equipment, characterized in that The radiotherapy equipment is used to implement the automatic positioning method based on anatomical structure detection according to any one of claims 1 to 4, or the radiotherapy equipment includes the automatic positioning device based on anatomical structure detection according to any one of claims 5 to 7.

Images