Tumor radiotherapy target area automatic sketching system and method

Abstract

The invention relates to the field of intelligent sketching, and particularly discloses an automatic sketching system and method for a tumor radiotherapy target area, and the method comprises the steps: taking an identity label of a patient object as an index, extracting a patient lung parenchyma CT scanning image of a to-be-sketched tumor radiotherapy target area corresponding to the patient from a CT scanning image database, carrying out multi-scale extraction on the patient pulmonary parenchyma CT scanning image by adopting an image analysis and extraction technology based on computer vision to obtain patient pulmonary parenchyma state superficial layer features and patient pulmonary parenchyma state semantic features, and then carrying out hierarchical dominant enhancement on the semantic features; therefore, the automatic sketching result of the tumor radiotherapy target region is automatically obtained based on the interactive coding expression between the lung parenchyma state superficial layer characteristics of the patient and the semantic enhancement expression of the lung parenchyma state of the patient. In this way, the tumor boundary can be identified and drawn more accurately. And the generalization ability and adaptability of the system are enhanced.

Description

Technical Field

[0001] The present application relates to the field of intelligent delineation, and more specifically, to a system and method for automatically delineating a tumor radiotherapy target area. Background Art

[0002] Tumor radiotherapy target delineation refers to the use of medical images (such as CT, MRI, etc.) by doctors or related technicians to determine the location of the tumor that needs to receive radiotherapy and the important organs around it during the radiotherapy planning process. This process is to ensure that the radiation can accurately irradiate the tumor area while minimizing the impact on the surrounding healthy tissues. By accurately defining the tumor boundaries, the treatment effect can be improved and side effects can be reduced.

[0003] Traditionally, techniques such as threshold segmentation, edge detection, and region growing are used to identify and outline tumor boundaries. However, these methods usually require high-quality imaging data as input, and if the image has low contrast, high noise, or is blurred, it may lead to incorrect boundary identification. For example, when the grayscale difference between the tumor and the surrounding tissue is not large, threshold segmentation may fail. In addition, many traditional algorithms (such as Canny edge detection) require setting multiple parameters, and the selection of these parameters directly affects the quality of the results. Finding the right combination of parameters often requires a lot of trials and adjustments. Different parameter settings may be required between different patients or even between different parts of the same patient, which increases the complexity of processing.

[0004] Therefore, an optimized automatic delineation scheme for tumor radiotherapy target volume is desired. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed.

[0006] According to one aspect of the present application, a system for automatically delineating a tumor radiotherapy target area is provided, comprising: A lung parenchyma CT scan image acquisition module is used to extract a lung parenchyma CT scan image of the patient whose tumor radiotherapy target area is to be delineated from a CT scan image database using the patient's identity as an index; A lung parenchyma state extraction module, used for performing multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state superficial features and the patient's lung parenchyma state semantic features; A lung parenchymal state feature enhancement module, used for performing lung parenchymal state semantic hierarchical enhancement on the semantic features of the patient's lung parenchymal state to obtain the semantic enhanced expression features of the patient's lung parenchymal state, wherein the lung parenchymal state feature enhancement module includes: a feature decoupling mapping unit, used for performing feature decoupling and spatial mapping on the semantic features of the patient's lung parenchymal state to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state; a feature weighting unit, used for performing explicit weighting on the semantic features of the patient's lung parenchymal state based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state to obtain the semantic enhanced expression features of the patient's lung parenchymal state; The multi-scale interactive module of lung parenchymal state is used to perform multi-scale semantic interactive encoding of the patient's lung parenchymal state shallow features and the patient's lung parenchymal state semantic enhancement expression features to obtain the lung parenchymal state shallow-semantic fusion encoding features, and based on the lung parenchymal state shallow-semantic fusion encoding features, obtain the automatic delineation result of the tumor radiotherapy target area.

[0007] Furthermore, the lung parenchyma state extraction module is used to: pass the patient's lung parenchyma CT scan image through a lung parenchyma state multi-scale feature extractor based on a hollow pyramid network to obtain a patient's lung parenchyma state shallow feature map as the patient's lung parenchyma state shallow feature and a patient's lung parenchyma state semantic feature map as the patient's lung parenchyma state semantic feature.

[0008] Furthermore, the feature decoupling mapping unit is used to: Performing pixel granularity feature decoupling on the semantic feature map of the patient's lung parenchyma state to obtain a set of semantic pixel granularity decoupling feature vectors of the patient's lung parenchyma state; Hyperbolic space mapping is performed on each patient's lung parenchymal state semantic pixel granularity decoupling feature vector in the set of the patient's lung parenchymal state semantic pixel granularity decoupling feature vectors to obtain a set of patient's lung parenchymal state semantic hyperbolic space modulation pixel granularity decoupling feature vectors as the set of patient's lung parenchymal state semantic hyperbolic space modulation pixel granularity decoupling features.

[0009] Furthermore, the feature weighting unit includes: A lung parenchyma state weight matrix calculation subunit, used to calculate the weight matrix of the set of semantic hyperbolic space modulation pixel granularity decoupling feature vectors of the patient's lung parenchyma state to obtain a patient's lung parenchyma state structural semantic significance distribution weight matrix; The semantic feature enhancement subunit of the lung parenchymal state is used to calculate the position point multiplication of the semantic significance distribution weight matrix of the patient's lung parenchymal state structure and each feature matrix along the channel dimension of the patient's lung parenchymal state semantic feature map to obtain the patient's lung parenchymal state semantic enhancement expression feature map as the patient's lung parenchymal state semantic enhancement expression feature.

[0010] Furthermore, the lung parenchyma state weight matrix calculation subunit is used to: Calculating the fine-grained structural semantic significance measurement coefficient of each patient's lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector in the set of the patient's lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector to obtain a patient's lung parenchyma state structural semantic significance distribution matrix formed by arranging multiple patient's lung parenchyma state fine-grained structural semantic significance measurement coefficients; The semantic significance distribution matrix of the patient's lung parenchymal state structure is subjected to multi-layer explicit hierarchical masking to obtain the semantic significance distribution weight matrix of the patient's lung parenchymal state structure.

[0011] Furthermore, the multi-scale interaction module of lung parenchyma state includes: A shallow semantic interaction encoding unit for lung parenchymal state, used for passing the shallow feature map of the patient's lung parenchymal state and the semantic enhancement expression feature map of the patient's lung parenchymal state through a lung parenchymal state multi-scale semantic interaction encoder to obtain a shallow-semantic fusion encoding feature map of the lung parenchymal state as the shallow-semantic fusion encoding feature of the lung parenchymal state; A lung parenchymal state fusion feature pooling unit, used for performing global maximum pooling along the channel dimension on the lung parenchymal state shallow-semantic fusion encoding feature map to obtain a lung parenchymal state shallow-semantic saliency fusion matrix; A delineation unit is used to obtain the automatic delineation result of the tumor radiotherapy target area based on the shallow-semantic saliency fusion matrix of the lung parenchyma state.

[0012] Furthermore, the shallow semantic interaction encoding unit of the lung parenchyma state is used to: The shallow feature map of the patient's lung parenchyma state and the semantically enhanced expression feature map of the patient's lung parenchyma state are fused to obtain a multi-scale feature map of the patient's lung parenchyma state; Using convolution layers with convolution kernels of 1×3, 3×1, and 3×3, respectively, to perform convolution encoding on the multi-scale patient lung parenchyma state feature map to obtain a first patient lung parenchyma state convolution encoding feature map, a second patient lung parenchyma state convolution encoding feature map, and a third patient lung parenchyma state convolution encoding feature map; Calculating the mean of corresponding positions in the first patient lung parenchyma state convolutional coding feature map, the second patient lung parenchyma state convolutional coding feature map, and the third patient lung parenchyma state convolutional coding feature map, and performing atrous convolution coding on the obtained multi-scale patient lung parenchyma state convolutional coding feature map to obtain a multi-scale perceived patient lung parenchyma state feature map; The multi-scale patient lung parenchymal state convolutional coding feature map and the multi-scale perceived patient lung parenchymal state feature map are residually connected to obtain the lung parenchymal state shallow-semantic fusion coding feature map.

[0013] Furthermore, the delineation unit is used to: input the shallow-semantic saliency fusion matrix of the lung parenchyma state into an image semantic segmentation delineator based on a Softmax function to obtain an automatic delineation result of the tumor radiotherapy target area.

[0014] According to another aspect of the present application, a method for automatically delineating a tumor radiotherapy target area is provided, comprising: Using the patient's identity as an index, extracting from a CT scan image database a CT scan image of the patient's lung parenchyma corresponding to the tumor radiotherapy target area to be delineated; Performing multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain superficial features of the patient's lung parenchyma state and semantic features of the patient's lung parenchyma state; Performing pulmonary parenchymal state semantic hierarchical enhancement on the semantic features of the patient's pulmonary parenchymal state to obtain semantically enhanced expression features of the patient's pulmonary parenchymal state, including: performing feature decoupling and spatial mapping on the semantic features of the patient's pulmonary parenchymal state to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's pulmonary parenchymal state; based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's pulmonary parenchymal state, performing explicit weighting on the semantic features of the patient's pulmonary parenchymal state to obtain the semantically enhanced expression features of the patient's pulmonary parenchymal state; Multi-scale semantic interactive encoding of the patient's lung parenchymal state is performed on the shallow features of the patient's lung parenchymal state and the semantically enhanced expression features of the patient's lung parenchymal state to obtain shallow-semantic fusion encoding features of the lung parenchymal state, and based on the shallow-semantic fusion encoding features of the lung parenchymal state, the automatic delineation result of the tumor radiotherapy target area is obtained.

[0015] Compared with the prior art, the present application provides a system and method for automatically delineating a tumor radiotherapy target area, which uses the patient's identity identifier as an index to extract the patient's lung parenchyma CT scan image of the patient's relative tumor radiotherapy target area to be delineated from the CT scan image database, and uses computer vision-based image analysis and extraction technology to perform multi-scale extraction of the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic features, and then hierarchically enhances the semantic features, so as to automatically obtain the tumor radiotherapy target area automatic delineation results based on the interactive coding representation between the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic enhancement expression. In this way, rich information from microstructure to macroscopic morphology can be captured, so as to more accurately identify and delineate the tumor boundary. In addition, it can better process image data of different qualities to better adapt to individual differences between different patients and changes between different parts of the same patient, thereby enhancing the generalization ability and adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 It is a block diagram of a system for automatically delineating a tumor radiotherapy target area according to an embodiment of the present application.

[0018] Figure 2 It is a block diagram of a lung parenchyma state feature enhancement module in the automatic delineation system of the tumor radiotherapy target area according to an embodiment of the present application.

[0019] Figure 3 It is a block diagram of a multi-scale interactive module of lung parenchyma state in an automatic delineation system of a tumor radiotherapy target area according to an embodiment of the present application.

[0020] Figure 4 The present invention is a flowchart of a method for automatically outlining a tumor radiotherapy target area according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] The embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, but rather these embodiments are provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0022] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0023] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0024] It is worth noting that in this application, all actions to obtain signals, information or image data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0025] Tumor radiotherapy target delineation refers to the use of medical images (such as CT, MRI, etc.) by doctors or related technicians to determine the location of the tumor that needs to receive radiotherapy and the important organs around it during the radiotherapy planning process. This process is to ensure that the radiation can accurately irradiate the tumor area while minimizing the impact on the surrounding healthy tissues. By accurately defining the tumor boundaries, the treatment effect can be improved and side effects can be reduced.

[0026] Traditionally, techniques such as threshold segmentation, edge detection, and region growing are used to identify and outline tumor boundaries. However, these methods usually require high-quality imaging data as input, and if the image has low contrast, high noise, or is blurred, it may lead to incorrect boundary identification. For example, when the grayscale difference between the tumor and the surrounding tissue is not large, threshold segmentation may fail. In addition, many traditional algorithms (such as Canny edge detection) require setting multiple parameters, and the selection of these parameters directly affects the quality of the results. Finding the right combination of parameters often requires a lot of trials and adjustments. Different parameter settings may be required between different patients or even between different parts of the same patient, which increases the complexity of processing.

[0027] Therefore, in response to the above technical problems, the technical concept of the present application is to use the patient's identity as an index, extract the patient's lung parenchyma CT scan image of the patient's relative tumor radiotherapy target area to be delineated from the CT scan image database, and use computer vision-based image analysis and extraction technology to perform multi-scale extraction of the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic features, and then perform hierarchical explicit enhancement on the semantic features, so as to automatically obtain the tumor radiotherapy target area automatic delineation result based on the interactive coding representation between the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic enhancement expression. In this way, rich information from microstructure to macroscopic morphology can be captured, so as to more accurately identify and delineate the tumor boundary. In addition, it can better process image data of different qualities to better adapt to individual differences between different patients and changes between different parts of the same patient, thereby enhancing the generalization ability and adaptability of the system.

[0028] Figure 1 FIG. 1 is a block diagram of a system for automatically delineating a tumor radiotherapy target area according to an embodiment of the present application. Figure 1 As shown, according to the automatic delineation system 100 of the tumor radiotherapy target area of ​​the embodiment of the present application, it includes: a lung parenchyma CT scan image acquisition module 110, which is used to extract the patient's lung parenchyma CT scan image of the tumor radiotherapy target area to be delineated from the CT scan image database with the patient's identity identifier as an index; a lung parenchyma state extraction module 120, which is used to perform multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow feature and the patient's lung parenchyma state semantic feature; a lung parenchyma state feature enhancement module 130, which is used to perform lung parenchyma state semantic hierarchical enhancement on the patient's lung parenchyma state semantic feature to obtain the patient's lung parenchyma state semantic enhancement expression feature; a lung parenchyma state multi-scale interaction module 140, which is used to perform lung parenchyma state multi-scale semantic interaction encoding on the patient's lung parenchyma state shallow feature and the patient's lung parenchyma state semantic enhancement expression feature to obtain the lung parenchyma state shallow-semantic fusion encoding feature, and obtain the tumor radiotherapy target area automatic delineation result based on the lung parenchyma state shallow-semantic fusion encoding feature.

[0029] In an embodiment of the present application, the lung parenchyma CT scan image acquisition module 110 is used to extract the patient's lung parenchyma CT scan image of the patient's tumor radiotherapy target area to be delineated from the CT scan image database with the patient's identity as an index. It should be understood that the anatomical structure of each patient is unique, and even the same disease may show different characteristics in different people. Therefore, extracting the individual CT scan image of a specific patient from the database based on the identity of the patient can ensure that the subsequent processing process is carried out for individual conditions. The patient's lung parenchyma CT scan image to be delineated for the tumor radiotherapy target area refers to a specific CT (computed tomography) image that needs to be analyzed and processed to determine the location and range of the tumor. These images are cross-sectional images of the patient's lungs obtained by a CT scanner, which can show the lung tissue and its internal structure, including any abnormal areas such as tumors. The extracted CT image will be used to identify and accurately locate the location of the tumor. This is crucial for formulating an accurate radiotherapy plan, because the radiation needs to be irradiated to the tumor as accurately as possible while minimizing the impact on surrounding healthy tissues.

[0030] To achieve this goal, the system first needs a data management system that can store a large number of CT scan images, and these images must be closely associated with the patient's identity. When it is necessary to delineate the tumor radiotherapy target area of ​​a specific patient, the system will use the patient's unique identity (such as medical record number, name, etc.) as the query keyword to quickly locate and extract the corresponding CT scan image in the database.

[0031] In actual operation, the technical details involved in this step include but are not limited to the construction of a data management system, the uniqueness of identity identification, and efficient data retrieval algorithms. Specifically, data management systems usually use relational databases or more advanced NoSQL databases to organize and store CT scan images and related metadata, such as shooting date, equipment model, examination site, etc. For each patient, all relevant medical imaging data will be assigned a unique identity, which not only facilitates the classification and organization of data, but also ensures that there will be no confusion when retrieving information about a specific patient. In addition, in order to improve retrieval efficiency, the database design will consider the use of appropriate indexing mechanisms, such as B-tree indexes or hash indexes, so that even in the face of massive data, the search task can be completed quickly.

[0032] Once the specific case to be treated is determined, the next step is to obtain the corresponding CT scan image from the database. The "patient-relative CT scan image of the patient's lung parenchyma to be delineated for the tumor radiotherapy target area" referred to here specifically refers to those CT images that contain information about the tumor area and the important organs around it. This type of image is obtained through a CT scanner and can show the detailed structure of the lung cross-section, including any abnormal areas such as tumors. Due to anatomical differences between different patients, even the same disease may show different characteristics, so it is very important to accurately extract individual CT scan images from the database based on the identity of the specific patient. This can not only help doctors identify the location of the tumor more accurately, but also effectively avoid the risk of misdiagnosis due to incorrect matching.

[0033] It is worth noting that data security and privacy protection issues also need to be considered throughout the process. Medical and health information belongs to the category of sensitive data. Therefore, when implementing the above process, relevant laws and regulations must be strictly observed, and necessary encryption measures and technical means must be taken to ensure information security. For example, the SSL / TLS protocol can be used to encrypt the data stream during transmission, and the access control list (ACL) and other permission management mechanisms can be used to restrict only authorized personnel to access the medical records of specific patients. In addition, regular security audits are also an indispensable task, which helps to timely discover potential security vulnerabilities and take corresponding measures to repair them. In summary, using the patient's identity as an index to extract the patient's lung parenchyma CT scan image to be delineated for the tumor radiotherapy target area is one of the foundations for the operation of the entire tumor radiotherapy target area automatic delineation system. This process requires not only an efficient and reliable data management system to support, but also the comprehensive use of a variety of technical means to ensure the accuracy, integrity and security of the data, and ultimately provide clinicians with high-quality imaging data to help formulate more accurate and effective radiotherapy plans.

[0034] In an embodiment of the present application, the lung parenchyma state extraction module 120 is used to perform multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic features. Specifically, in an embodiment of the present application, the lung parenchyma state extraction module 120 is used to: pass the patient's lung parenchyma CT scan image through a lung parenchyma state multi-scale feature extractor based on a hollow pyramid network to obtain a patient's lung parenchyma state shallow feature map as the patient's lung parenchyma state shallow features and a patient's lung parenchyma state semantic feature map as the patient's lung parenchyma state semantic features. Accordingly, considering that the patient's lung parenchyma CT scan image contains a large amount of information about the structure and state of the lungs, such as the edges, textures, and shapes of shallow cardinality visual information, it also contains higher-level understanding and deep semantic information of abstract concepts, that is, shallow features help determine the physical boundaries of the tumor, while deep semantic features help confirm the specific type of the tumor and its biological behavior. Based on this, in order to more comprehensively and meticulously capture the feature information of different scales of the lung structure, so as to use visual information at different levels to improve the accuracy of tumor positioning and segmentation, in the technical solution of the present application, the patient's lung parenchyma CT scan image is passed through a multi-scale feature extractor of the lung parenchyma state based on a hollow pyramid network to obtain a shallow feature map of the patient's lung parenchyma state and a semantic feature map of the patient's lung parenchyma state. It can be understood that the hollow pyramid network combines the ideas of hollow convolution (also called dilated convolution) and feature pyramid network. It can obtain features under different receptive fields by operating hollow convolution (also called dilated convolution) at different expansion rates, which helps to capture local detail information, such as edges, textures, etc., which is very important for identifying subtle changes in a small range. At the same time, the feature pyramid network can extract high-level features with strong semantic meanings, which reflect patterns or objects in a larger range in the image, such as lung nodules, vascular structures, etc., thereby generating multi-scale feature representations.

[0035] Specifically, the specific processing process of the patient's lung parenchyma CT scan image through the lung parenchyma state multi-scale feature extractor based on the hollow pyramid network is as follows: First, the patient's lung parenchyma CT scan image will pass through a series of basic convolutional layers, which are used to preliminarily extract the basic features of the image, such as edges, textures, etc. This stage is similar to the first few layers in a traditional convolutional neural network (CNN), but the following process will be different.

[0036] As the network goes deeper, the concept of dilated convolution is introduced. Compared with standard convolution, dilated convolution can expand the receptive field without increasing the number of parameters by inserting holes between the convolution kernel elements (that is, skipping some pixels). This means that a wider range of information can be captured even in shallower network layers. For example, in the first layer of dilated convolution, the expansion rate can be selected as 2, so that each convolution kernel actually covers a 4x4 area on the original image; in the second layer, the expansion rate can be further increased to 4, at which time each convolution kernel acts on an 8x8 area. This setting enables the network to gradually build a rich feature representation from local details to global patterns.

[0037] Next, to form the feature pyramid, the network usually contains multiple parallel branches, each corresponding to a different scale. These branches are composed of dilated convolutional layers with different dilation rates to generate feature maps with different resolutions and receptive field sizes. For example, one branch may focus on small-scale features and use a smaller dilation rate, while another branch focuses on large-scale features and uses a larger dilation rate. All these branches work together to ensure that information at different scales from micro to macro is fully captured.

[0038] In addition, in order to strengthen the communication between features of different scales and enhance the overall representation ability, cross-layer connections or fusion mechanisms are added between branches. A common practice is to use residual connections to pass high-level semantic information to low-level layers, or to directly merge feature maps from different branches by element-level addition. Such a design helps to maintain spatial information while enriching the diversity of features, which is beneficial to subsequent tasks such as tumor boundary recognition.

[0039] Finally, there is usually a global pooling layer at the output of the entire network to aggregate features at all scales to form the final feature representation. This representation can be sent to the fully connected layer for classification prediction, or directly used for image segmentation tasks, such as automatic delineation of tumor radiotherapy targets. Therefore, the dilated pyramid network cleverly combines dilated convolution with the idea of ​​feature pyramids to effectively expand the network's receptive field while retaining high-resolution details, providing rich data support for subsequent analysis of the characteristics of the patient's lung parenchyma status.

[0040] In an embodiment of the present application, the lung parenchymal state feature enhancement module 130 is used to perform semantic hierarchical enhancement of the patient's lung parenchymal state semantic features to obtain the patient's lung parenchymal state semantic enhancement expression features. Further, considering that when processing the patient's lung parenchymal state semantic feature map, the lung parenchymal states of different local areas have different key and significant features, and are also affected by interference factors and irrelevant details. These interferences may include image noise, non-pathological changes, and background structures, which may mask or confuse key pathological information. Therefore, in order to be able to understand and analyze more fine-grained, so as to extract and capture the key state semantic information in different local areas, in the technical solution of the present application, the semantic features of the patient's lung parenchymal state are enhanced by semantic hierarchical enhancement of the lung parenchymal state to obtain the patient's lung parenchymal state semantic enhancement expression features. That is, semantic hierarchical enhancement aims to refine the extraction and highlight the key state information, while effectively suppressing interference factors to ensure that the final feature map is clearer and rich in diagnostic value.

[0041] Figure 2 FIG. 4 is a block diagram of a lung parenchyma state feature enhancement module in a tumor radiotherapy target area automatic delineation system according to an embodiment of the present application. Specifically, Figure 2 As shown, the lung parenchymal state feature enhancement module 130 includes: a feature decoupling mapping unit 131, which is used to perform feature decoupling and spatial mapping on the semantic features of the patient's lung parenchymal state to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state; a feature weighting unit 132, which is used to perform explicit weighting on the semantic features of the patient's lung parenchymal state based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state to obtain the semantic enhanced expression features of the patient's lung parenchymal state.

[0042] Specifically, the feature decoupling mapping unit 131 is used to: perform pixel granularity feature decoupling on the semantic feature map of the patient's lung parenchyma state to obtain a set of semantic pixel granularity decoupling feature vectors of the patient's lung parenchyma state; perform hyperbolic space mapping on each semantic pixel granularity decoupling feature vector of the patient's lung parenchyma state in the set of semantic pixel granularity decoupling feature vectors of the patient's lung parenchyma state to obtain a set of semantic hyperbolic space modulation pixel granularity decoupling feature vectors of the patient's lung parenchyma state as the set of semantic hyperbolic space modulation pixel granularity decoupling features of the patient's lung parenchyma state. First, pixel granularity feature decoupling is performed on the semantic feature map of the patient's lung parenchyma state to decompose an entire feature map into multiple local features containing pixel granularity to remove unnecessary correlations between pixels, allowing the features of each pixel to be processed independently, and obtaining a set of semantic pixel granularity decoupling feature vectors of the patient's lung parenchyma state. That is, by decomposing the image into pixel-level features, the features of each pixel point or small area can be captured more meticulously, while removing or weakening the influence of these redundant information, thereby improving the description ability of local details, which is of great significance for identifying small lesions, edge fuzzy areas, etc. Next, each pixel-size decoupled feature is hyperbolically mapped to better process and analyze data with complex hierarchical structures, and a set of semantic hyperbolic space modulation pixel-size decoupled feature vectors of the patient's lung parenchyma state is obtained. It should be understood that hyperbolic space can represent data with tree-like or network-like structures in a more compact way than Euclidean space. For the tissue structure in lung images, this feature helps capture the relationship between different levels (from macro to micro). Moreover, in hyperbolic space, as the distance from the origin increases, the space "expands", which keeps a larger relative distance between points far from the center. This property is conducive to maintaining the uniqueness of local features in high-dimensional space, especially when these features represent specific types of abnormalities or lesions.

[0043] More specifically, the processing process of the feature decoupling mapping unit 131 can be expressed as follows: ; ; in, is the semantic feature map of the patient's lung parenchyma state, For Decouple pixel-level features. , , and They are the first, second, and third semantic pixel-granularity decoupled feature vectors of the patient’s lung parenchyma status. and The semantic pixel-level decoupling feature vector of the lung parenchyma status of each patient, and are the first weight matrix and the second weight matrix respectively, for The corresponding semantic hyperbolic spatial modulation pixel-granularity decoupled feature vector of the patient's lung parenchyma state.

[0044] Specifically, the feature weighting unit 132 includes: a lung parenchymal state weight matrix calculation subunit, which is used to calculate the weight matrix of the set of semantic hyperbolic space modulation pixel granularity decoupling feature vectors of the patient's lung parenchymal state to obtain the patient's lung parenchymal state structural semantic significance distribution weight matrix; a lung parenchymal state semantic feature enhancement subunit, which is used to calculate the point-by-point multiplication of the patient's lung parenchymal state structural semantic significance distribution weight matrix and each feature matrix along the channel dimension of the patient's lung parenchymal state semantic feature map to obtain the patient's lung parenchymal state semantic enhanced expression feature map as the patient's lung parenchymal state semantic enhanced expression feature.

[0045] More specifically, in the embodiment of the present application, the lung parenchyma state weight matrix calculation subunit is used to: calculate the fine-grained structural semantic saliency measurement coefficient of each patient lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector in the set of the patient lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector to obtain a patient lung parenchyma state structural semantic saliency distribution matrix arranged by multiple patient lung parenchyma state fine-grained structural semantic saliency measurement coefficients; perform multi-layer explicit hierarchical masking on the patient lung parenchyma state structural semantic saliency distribution matrix to obtain the patient lung parenchyma state structural semantic saliency distribution weight matrix. It should be understood that the fine-grained structural semantic saliency measurement coefficient of each hyperbolic spatial modulation pixel granularity decoupling feature vector is calculated to quantify the relative importance of each local area in the overall semantic expression, thereby highlighting those areas with high significance, which may contain important pathological information, such as tumors, inflammation or other abnormalities, to obtain the patient lung parenchyma state structural semantic saliency distribution matrix. Subsequently, in order to further refine and optimize the importance representation of features in the saliency distribution matrix to highlight important and significant features, the saliency distribution matrix is ​​subjected to multi-layer explicit hierarchical masking to obtain the semantic saliency distribution weight matrix of the patient's lung parenchyma state structure. That is, through multi-layer explicit hierarchical masking, each value in the saliency distribution matrix can be adjusted at different levels so that each value can be optimized according to its specific importance, which helps to capture important information at different scales from macro to micro, which means that each pixel or local area can obtain a weight that matches its importance, thereby generating a more refined and adaptive weight matrix.

[0046] More specifically, the processing of the lung parenchyma state weight matrix calculation subunit can be expressed as follows: ; ; in, for The corresponding patient lung parenchyma state semantic hyperbolic space modulation pixel granularity decoupling feature vector, To calculate the square of the length of a vector, represents the logarithmic function value with base 2, is the semantic saliency distribution matrix of the patient's lung parenchymal state structure The corresponding semantic saliency coefficient of the fine-grained structure of the patient's lung parenchyma state, is the preset threshold, For masking operation, is the semantic significance distribution weight matrix of the patient's lung parenchymal state structure The corresponding semantic significance distribution weight value of the patient's lung parenchymal state structure.

[0047] Specifically, the pulmonary parenchymal state semantic feature enhancement subunit is used to calculate the semantic significance distribution weight matrix of the patient's pulmonary parenchymal state structure and the position point multiplication of each feature matrix of the patient's pulmonary parenchymal state semantic feature map along the channel dimension to obtain the patient's pulmonary parenchymal state semantic enhancement expression feature map as the patient's pulmonary parenchymal state semantic enhancement expression feature. In other words, the weight matrix is ​​calculated and the position point multiplication of each feature matrix of the patient's pulmonary parenchymal state semantic feature map along the channel dimension is used to embed the significance information directly into each feature channel, so that the information of these key areas is more prominent, and the patient's pulmonary parenchymal state semantic enhancement expression feature map with characterization ability is obtained.

[0048] More specifically, the processing of the above lung parenchyma state semantic feature enhancement subunit can be expressed as: ; in, is the semantic feature map of the patient's lung parenchyma state, is the semantic significance distribution weight matrix of the patient's lung parenchymal state structure, In order to use the matrix to weight each feature matrix along the channel dimension of the feature map, It is a semantically enhanced expression feature map of the patient's lung parenchymal state.

[0049] In an embodiment of the present application, the lung parenchymal state multi-scale interaction module 140 is used to perform multi-scale semantic interaction encoding of the patient's lung parenchymal state shallow features and the patient's lung parenchymal state semantic enhancement expression features to obtain the lung parenchymal state shallow-semantic fusion coding features, and based on the lung parenchymal state shallow-semantic fusion coding features, obtain the automatic delineation result of the tumor radiotherapy target area. Figure 3 FIG. 4 is a block diagram of a multi-scale interaction module of lung parenchyma state in a tumor radiotherapy target area automatic delineation system according to an embodiment of the present application. Specifically, Figure 3 As shown, the lung parenchymal state multi-scale interaction module 140 includes: a lung parenchymal state shallow semantic interaction encoding unit 141, which is used to pass the patient's lung parenchymal state shallow feature map and the patient's lung parenchymal state semantic enhancement expression feature map through a lung parenchymal state multi-scale semantic interaction encoder to obtain a lung parenchymal state shallow-semantic fusion encoding feature map as the lung parenchymal state shallow-semantic fusion encoding feature; a lung parenchymal state fusion feature pooling unit 142, which is used to perform global maximum pooling along the channel dimension on the lung parenchymal state shallow-semantic fusion encoding feature map to obtain a lung parenchymal state shallow-semantic significant fusion matrix; a delineation unit 143, which is used to obtain the automatic delineation result of the tumor radiotherapy target area based on the lung parenchymal state shallow-semantic significant fusion matrix.

[0050] More specifically, the shallow semantic interaction encoding unit 141 of the lung parenchymal state is used to: fuse the shallow feature map of the patient's lung parenchymal state and the semantic enhanced expression feature map of the patient's lung parenchymal state to obtain a multi-scale patient's lung parenchymal state feature map; use convolution layers with convolution kernels of 1×3, 3×1 and 3×3 to convolutionally encode the multi-scale patient's lung parenchymal state feature map to obtain a first patient's lung parenchymal state convolutional encoding feature map, a second patient's lung parenchymal state convolutional encoding feature map and a third patient's lung parenchymal state convolutional encoding feature map; calculate the means of corresponding positions in the first patient's lung parenchymal state convolutional encoding feature map, the second patient's lung parenchymal state convolutional encoding feature map and the third patient's lung parenchymal state convolutional encoding feature map, and perform atrous convolution encoding on the obtained multi-scale patient's lung parenchymal state convolutional encoding feature map to obtain a multi-scale perceived patient's lung parenchymal state feature map; perform residual connection on the multi-scale patient's lung parenchymal state convolutional encoding feature map and the multi-scale perceived patient's lung parenchymal state feature map to obtain the shallow-semantic fusion encoding feature map of the lung parenchymal state. It should be understood that, considering that the shallow features of the patient's lung parenchymal state usually contain relatively basic and local structural information in the image (such as edges, textures, etc.), the semantically enhanced expression features of the patient's lung parenchymal state after semantic enhancement carry higher-level abstract information (such as organ boundaries, lesion locations). Therefore, in order to fully capture the subtle changes in lung structure and enhance adaptability to different imaging conditions and different patients, in the technical solution of the present application, the shallow feature map of the patient's lung parenchymal state and the semantically enhanced expression feature map of the patient's lung parenchymal state are passed through a multi-scale semantic interactive encoder of the lung parenchymal state to obtain a shallow-semantic fusion encoding feature map of the lung parenchymal state. Specifically, first, the two features are fused to combine feature information at different levels to form an initial fusion representation. Then, the feature map after the initial fusion is convolved using convolution kernels of different sizes (1×3, 3×1, and 3×3 mentioned here) in order to capture spatial information in different directions and features within receptive fields of different sizes. For example, 1×3 and 3×1 convolution kernels help extract edge information in the horizontal and vertical directions, while 3×3 convolution kernels are more suitable for extracting overall features of local areas. Then, for the feature maps obtained by the above three convolutional codes, the average of their pixel values ​​at the same position is calculated to integrate feature information at different scales, and the dilated convolution is applied to these averaged feature maps to increase the receptive field without reducing the feature map resolution, which helps the model obtain broader contextual information while maintaining high resolution. Finally, a residual connection is established between the averaged multi-scale convolutional coding feature map and the multi-scale perceptual feature map processed by dilated convolution.In particular, residual connections are a method that allows gradients to flow more easily through the deep layers of the network, which helps solve the gradient vanishing problem that may occur during deep neural network training and promotes the effective transfer of information. In this way, the feature representation can be further optimized to form the final shallow-semantic fusion encoding feature map of the lung parenchyma state.

[0051] Specifically, the lung parenchymal state fusion feature pooling unit 142 is used to perform global maximum pooling along the channel dimension on the lung parenchymal state shallow-semantic fusion coding feature map to obtain a lung parenchymal state shallow-semantic significant fusion matrix. That is, in order to select the most representative features from the fusion coding features and reduce the amount of data while retaining key information, in the technical solution of the present application, the lung parenchymal state shallow-semantic fusion coding feature map is subjected to global maximum pooling along the channel dimension to further emphasize the importance of the strongest signal in each channel, and obtain a lung parenchymal state shallow-semantic significant fusion matrix.

[0052] It should be understood that global maximum pooling is a commonly used dimensionality reduction technique, which is widely used in feature extraction and compression in the field of image processing. The basic idea is to select the maximum value from each feature channel as the representative value of the channel, which ensures that the most important features can be retained even in a large range. For the shallow-semantic fusion encoding feature map of lung parenchymal state, this means finding and recording the maximum value of the channel in the entire feature map on each channel (that is, each specific type of feature). These maximum values ​​together constitute a new representation - the shallow-semantic fusion matrix of lung parenchymal state, which reflects the most prominent feature information in the original feature map.

[0053] Specifically, the specific processing process of performing global maximum pooling along the channel dimension on the shallow-semantic fusion coding feature map of the lung parenchyma state is as follows: The specific method for performing this operation is as follows: Assume that the size of the shallow-semantic fusion encoding feature map of the lung parenchymal state is H×W×C, where H and W represent the height and width respectively, and C represents the number of channels, that is, the number of different types of features contained in the feature map. Then, the global maximum pooling operation is to traverse all channels and find the maximum value in each channel. The final result is a matrix of size 1×1×C, that is, a vector of length C, each element of which corresponds to the maximum value of the corresponding channel in the original feature map. This vector or matrix is ​​the shallow-semantic saliency fusion matrix of the lung parenchymal state.

[0054] It is worth noting that the reason for using global maximum pooling rather than other forms such as average pooling is that maximum pooling is better at emphasizing those features that are widely distributed in space and have higher intensity, which is particularly important for the identification of key areas such as tumor boundaries. In addition, this method can also effectively remove the influence of background noise and other irrelevant details, so that the final significant fusion matrix is ​​more focused on the truly important pathological information.

[0055] In practical applications, in order to further improve the performance of the model, some optimization processing may be performed on the obtained significant fusion matrix. For example, considering the possible lack of spatial structure in high-dimensional space, long-distance feature dependencies may be established based on the spatial structure representation of the overall feature scale relative to the significant fusion matrix, so as to improve the local connectivity of features and the ability to capture spatial ambiguous information, thereby improving the spatial inductive bias perception ability of the feature set and the consistency of semantic segmentation convergence. This step helps to ensure that the final automatic delineation results of the tumor radiotherapy target are more accurate and reliable. In summary, by performing global maximum pooling along the channel dimension on the shallow-semantic fusion encoding feature map of the lung parenchyma state, the most representative feature information can be effectively extracted to form a shallow-semantic significant fusion matrix of the lung parenchyma state. This process not only simplifies the data scale for subsequent processing, but also enhances the model's attention to key pathological features, thereby improving the overall performance of the automatic delineation system for tumor radiotherapy targets.

[0056] More specifically, the delineation unit 143 is used to: input the shallow-semantic saliency fusion matrix of the lung parenchyma state into the image semantic segmentation delineator based on the Softmax function to obtain the automatic delineation result of the tumor radiotherapy target area. In this way, rich information from microstructure to macroscopic morphology can be captured, so as to more accurately identify and delineate the tumor boundary. In addition, it can better process image data of different qualities to better adapt to individual differences between different patients and changes between different parts of the same patient, thereby enhancing the generalization ability and adaptability of the system.

[0057] In particular, in a specific embodiment of the present application, the processing process of inputting the shallow-semantic saliency fusion matrix of the lung parenchyma state into the image semantic segmentation delineator based on the Softmax function is as follows: 1. The shallow-semantic saliency fusion matrix of the lung parenchyma state is used as input data. This matrix contains key information after multi-scale feature extraction and enhancement, and is the basis for subsequent processing.

[0058] 2. The matrix is ​​fed into a pre-trained convolutional neural network (CNN). CNNs are usually composed of multiple convolutional layers, pooling layers, and activation functions to further extract complex features in the encoded image. These layers gradually reduce the spatial dimensions while increasing the abstraction level of the features.

[0059] 3. After the encoding stage, a series of upsampling operations (such as deconvolution or bilinear interpolation) are used to restore the spatial resolution of the image. This step is also called the decoding stage, and its purpose is to restore the low-resolution feature map to the original image size for pixel-level classification. It is worth noting that during the decoding process, skip connections may be added to pass some feature maps in the encoding stage directly to the decoding stage to retain more detail information.

[0060] 4. Apply the Softmax function on the decoded feature map. The Softmax function is a normalized exponential function that converts the feature vector at each pixel position into a probability distribution. Specifically, for each pixel, the Softmax function outputs a vector in which each element represents the probability that the pixel belongs to a certain category. For example, in the task of delineating the target area for tumor radiotherapy, there may be two categories: tumor area and non-tumor area. The Softmax function calculates the probabilities of these two categories for each pixel.

[0061] 5. According to the probability distribution output by the Softmax function, the category with the highest probability is selected as the final prediction result. For example, if the probability of a pixel in the tumor category is higher than the probability of the non-tumor category, the pixel is marked as a tumor area.

[0062] 6. In order to improve the quality of the delineation results, some post-processing steps may be required, such as morphological operations (such as dilation and erosion), connected domain analysis, etc., to remove isolated small areas or fill small holes to make the boundary smoother and more reasonable.

[0063] 7. Based on the category labels generated in the above steps, the automatic delineation result of the tumor radiotherapy target area is formed. This result is usually a binary image in which the tumor area is marked, which can be further used to formulate an accurate radiotherapy plan. Through this series of steps, the image semantic segmentation delineator based on the Softmax function can accurately identify and delineate the tumor radiotherapy target area from complex medical images, thereby providing strong support for clinicians. This method not only improves work efficiency, but also improves the consistency and accuracy of delineation.

[0064] It should be understood that here, when the shallow feature map of the patient's lung parenchymal state and the semantically enhanced expression feature map of the patient's lung parenchymal state respectively represent the image semantic features of different depths and scales of the patient's lung parenchymal CT scan image, after multi-scale semantic interaction encoding is performed on them, the shallow-semantic fusion encoding feature map of the lung parenchyma state is subjected to global maximum pooling along the channel dimension to obtain a shallow-semantic significant fusion matrix of the lung parenchyma state. This will also have differences in the spatial structure of interactive fusion distribution due to the misalignment of the semantic spatial distribution of images across scales and depths, affecting the convergence consistency of semantic segmentation, thereby affecting the accuracy of the automatic delineation results of the tumor radiotherapy target area obtained by the image semantic segmentation delineator based on the Softmax function.

[0065] Preferably, inputting the shallow-semantically significant fusion matrix of the lung parenchyma state into an image semantic segmentation delineator based on a Softmax function to obtain an automatic delineation result of the tumor radiotherapy target area includes: The absolute value sum of all eigenvalues ​​of the shallow-semantic fusion matrix of the lung parenchyma state is calculated to obtain the first lung parenchyma state shallow-semantic fusion spatial structure value, and the square root of the sum of squares of the shallow-semantic fusion matrix of the lung parenchyma state is calculated to obtain the second lung parenchyma state shallow-semantic fusion spatial structure value, that is: ; ; in, is the eigenvalue of each position in the shallow-semantic saliency fusion matrix of the lung parenchyma state, is the number of eigenvalues ​​in the shallow-semantic saliency fusion matrix of the lung parenchyma state, is the shallow-semantically significant fusion spatial structure value of the first lung parenchymal state, is the shallow-semantically significant fusion spatial structure value of the second lung parenchyma state; For each eigenvalue of the shallow-semantic salient fusion matrix of the lung parenchymal state, the product of the eigenvalue and the total number of eigenvalues ​​of the shallow-semantic salient fusion matrix of the lung parenchymal state is calculated to obtain the first lung parenchymal state shallow-semantic salient fusion long-range dependency value ; in, is the shallow-semantically significant fusion long-range dependency value of the first lung parenchymal state; Calculate the second lung parenchymal state shallow-semantically significant fusion long-range dependency value obtained by multiplying the square root of the total number of eigenvalues ​​by the product of the eigenvalues ​​minus the second lung parenchymal state shallow-semantically significant fusion spatial structure value ; in, is the shallow-semantically significant fusion long-range dependency value of the second lung parenchyma state; The index value calculated by taking the shallow-semantic salient fusion long-range dependency value of the first lung parenchyma state as the exponent of the natural constant and the inverse of the shallow-semantic salient fusion long-range dependency value of the second lung parenchyma state are weighted summed to obtain the optimized eigenvalue corresponding to each eigenvalue ; in, and is the weighting coefficient, is the optimized eigenvalue corresponding to each eigenvalue; The optimized lung parenchyma state shallow-semantic saliency fusion matrix composed of the optimized eigenvalues ​​is input into the image semantic segmentation delineator based on the Softmax function to obtain the automatic delineation result of the tumor radiotherapy target area.

[0066] Based on this, in view of the possible spatial structure loss of the feature set of the shallow-semantic salient fusion matrix of the lung parenchymal state in the high-dimensional space, which leads to the convergence inconsistency of the image semantic segmentation based on the implicit inference of spatial structure information of the features, a long-distance feature dependency relationship is established based on the overall feature scale of the shallow-semantic salient fusion matrix of the lung parenchymal state relative to the spatial structure representation of the shallow-semantic salient fusion matrix of the lung parenchymal state, so as to establish the local connectivity of the features of the shallow-semantic salient fusion matrix of the lung parenchymal state, and capture the spatial ambiguous information of the object feature values ​​through the unstructured eigenvalue point prediction of the shallow-semantic salient fusion matrix of the lung parenchymal state, thereby improving the spatial inductive bias perception ability of the feature set of the shallow-semantic salient fusion matrix of the lung parenchymal state, improving the convergence consistency of semantic segmentation, and improving the accuracy of the automatic delineation result of the tumor radiotherapy target area obtained by inputting the shallow-semantic salient fusion matrix of the lung parenchymal state into the image semantic segmentation delineator based on the Softmax function.

[0067] In summary, the automatic delineation system 100 for the tumor radiotherapy target area based on the embodiment of the present application is explained, which extracts the patient's lung parenchyma CT scan image of the patient's relative tumor radiotherapy target area to be delineated from the CT scan image database by using the patient's identity identifier as an index, and uses computer vision-based image analysis and extraction technology to perform multi-scale extraction of the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic features, and then hierarchically enhances the semantic features, so as to automatically obtain the tumor radiotherapy target area automatic delineation result based on the interactive coding representation between the patient's lung parenchyma state shallow features and the patient's lung parenchyma state semantic enhancement expression. In this way, rich information from microstructure to macroscopic morphology can be captured, so as to more accurately identify and delineate the tumor boundary. In addition, it can better process image data of different qualities to better adapt to individual differences between different patients and changes between different parts of the same patient, thereby enhancing the generalization ability and adaptability of the system.

[0068] As described above, the automatic delineation system 100 for tumor radiotherapy target areas according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an automatic delineation algorithm for tumor radiotherapy target areas. In a possible implementation, the automatic delineation system 100 for tumor radiotherapy target areas according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the automatic delineation system 100 for tumor radiotherapy target areas can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the automatic delineation system 100 for tumor radiotherapy target areas can also be one of the many hardware modules of the wireless terminal.

[0069] Alternatively, in another example, the tumor radiotherapy target area automatic delineation system 100 and the wireless terminal may also be separate devices, and the tumor radiotherapy target area automatic delineation system 100 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0070] Figure 4 FIG. 1 is a flow chart of a method for automatically delineating a tumor radiotherapy target area according to an embodiment of the present application. Figure 4As shown, according to the automatic delineation method of the tumor radiotherapy target area of ​​the embodiment of the present application, the method includes: S110, using the patient object's identity identifier as an index, extracting the patient's lung parenchyma CT scan image relative to the tumor radiotherapy target area to be delineated from the CT scan image database; S120, performing multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state shallow feature and the patient's lung parenchyma state semantic feature; S130, performing lung parenchyma state semantic hierarchical enhancement on the patient's lung parenchyma state semantic feature to obtain the patient's lung parenchyma state semantic enhancement expression feature, including: feature decoupling the patient's lung parenchyma state semantic feature and spatial mapping to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state; based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state, the semantic features of the patient's lung parenchymal state are explicitly weighted to obtain the semantic enhanced expression features of the patient's lung parenchymal state; S140, multi-scale semantic interactive encoding of the patient's lung parenchymal state shallow features and the patient's lung parenchymal state semantic enhanced expression features is performed to obtain the shallow-semantic fusion coding features of the lung parenchymal state, and based on the shallow-semantic fusion coding features of the lung parenchymal state, the automatic delineation result of the tumor radiotherapy target area is obtained.

[0071] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned tumor radiotherapy target area automatic delineation method have been referred to above. Figures 1 to 3 The invention has been described in detail in the description of the automatic delineation system for tumor radiotherapy target area, and therefore, its repeated description will be omitted.

[0072] The above descriptions of various implementations of the present disclosure are exemplary and not exhaustive. They are not limited to the disclosed implementations, and many modifications and changes are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The terms used herein are selected to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other ordinary skill in the art to understand the various practical methods disclosed herein.

Claims

1. Tumor radiotherapy target area automatic delineation system, characterized by: include: A lung parenchyma CT scan image acquisition module is used to extract a lung parenchyma CT scan image of the patient whose tumor radiotherapy target area is to be delineated from a CT scan image database using the patient's identity as an index; A lung parenchyma state extraction module, used for performing multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain the patient's lung parenchyma state superficial features and the patient's lung parenchyma state semantic features; A lung parenchymal state feature enhancement module, used for performing lung parenchymal state semantic hierarchical enhancement on the semantic features of the patient's lung parenchymal state to obtain the semantic enhanced expression features of the patient's lung parenchymal state, wherein the lung parenchymal state feature enhancement module includes: a feature decoupling mapping unit, used for performing feature decoupling and spatial mapping on the semantic features of the patient's lung parenchymal state to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state; a feature weighting unit, used for performing explicit weighting on the semantic features of the patient's lung parenchymal state based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchymal state to obtain the semantic enhanced expression features of the patient's lung parenchymal state; The multi-scale interactive module of lung parenchymal state is used to perform multi-scale semantic interactive encoding of the patient's lung parenchymal state shallow features and the patient's lung parenchymal state semantic enhancement expression features to obtain the lung parenchymal state shallow-semantic fusion encoding features, and based on the lung parenchymal state shallow-semantic fusion encoding features, obtain the automatic delineation result of the tumor radiotherapy target area.

2. The automatic tumor radiotherapy target area delineation system according to claim 1, characterized in that: The lung parenchyma state extraction module is used to: pass the patient's lung parenchyma CT scan image through a lung parenchyma state multi-scale feature extractor based on a hollow pyramid network to obtain a patient's lung parenchyma state shallow feature map as the patient's lung parenchyma state shallow feature and a patient's lung parenchyma state semantic feature map as the patient's lung parenchyma state semantic feature.

3. The automatic delineation system for tumor radiotherapy target area according to claim 2, characterized in that: The feature decoupling mapping unit is used to: Performing pixel granularity feature decoupling on the semantic feature map of the patient's lung parenchyma state to obtain a set of semantic pixel granularity decoupling feature vectors of the patient's lung parenchyma state; Hyperbolic space mapping is performed on each patient's lung parenchymal state semantic pixel granularity decoupling feature vector in the set of the patient's lung parenchymal state semantic pixel granularity decoupling feature vectors to obtain a set of patient's lung parenchymal state semantic hyperbolic space modulation pixel granularity decoupling feature vectors as the set of patient's lung parenchymal state semantic hyperbolic space modulation pixel granularity decoupling features.

4. The automatic delineation system for tumor radiotherapy target area according to claim 3, characterized in that: The feature weighting unit comprises: A lung parenchyma state weight matrix calculation subunit, used to calculate the weight matrix of the set of semantic hyperbolic space modulation pixel granularity decoupling feature vectors of the patient's lung parenchyma state to obtain a patient's lung parenchyma state structural semantic significance distribution weight matrix; The semantic feature enhancement subunit of the lung parenchymal state is used to calculate the position point multiplication of the semantic significance distribution weight matrix of the patient's lung parenchymal state structure and each feature matrix along the channel dimension of the patient's lung parenchymal state semantic feature map to obtain the patient's lung parenchymal state semantic enhancement expression feature map as the patient's lung parenchymal state semantic enhancement expression feature.

5. The automatic delineation system for tumor radiotherapy target area according to claim 4, characterized in that: The lung parenchyma state weight matrix calculation subunit is used for: Calculating the fine-grained structural semantic significance measurement coefficient of each patient's lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector in the set of the patient's lung parenchyma state semantic hyperbolic spatial modulation pixel granularity decoupling feature vector to obtain a patient's lung parenchyma state structural semantic significance distribution matrix formed by arranging multiple patient's lung parenchyma state fine-grained structural semantic significance measurement coefficients; The semantic significance distribution matrix of the patient's lung parenchymal state structure is subjected to multi-layer explicit hierarchical masking to obtain the semantic significance distribution weight matrix of the patient's lung parenchymal state structure.

6. The automatic delineation system for tumor radiotherapy target area according to claim 5, characterized in that: The multi-scale interaction module of lung parenchyma state includes: A shallow semantic interaction encoding unit for lung parenchymal state, used for passing the shallow feature map of the patient's lung parenchymal state and the semantic enhancement expression feature map of the patient's lung parenchymal state through a lung parenchymal state multi-scale semantic interaction encoder to obtain a shallow-semantic fusion encoding feature map of the lung parenchymal state as the shallow-semantic fusion encoding feature of the lung parenchymal state; A lung parenchymal state fusion feature pooling unit, used for performing global maximum pooling along the channel dimension on the lung parenchymal state shallow-semantic fusion encoding feature map to obtain a lung parenchymal state shallow-semantic saliency fusion matrix; A delineation unit is used to obtain the automatic delineation result of the tumor radiotherapy target area based on the shallow-semantic saliency fusion matrix of the lung parenchyma state.

7. The automatic delineation system for tumor radiotherapy target area according to claim 6, characterized in that: The shallow semantic interaction encoding unit of the lung parenchyma state is used for: The shallow feature map of the patient's lung parenchyma state and the semantically enhanced expression feature map of the patient's lung parenchyma state are fused to obtain a multi-scale feature map of the patient's lung parenchyma state; Using convolution layers with convolution kernels of 1×3, 3×1, and 3×3, respectively, to perform convolution encoding on the multi-scale patient lung parenchyma state feature map to obtain a first patient lung parenchyma state convolution encoding feature map, a second patient lung parenchyma state convolution encoding feature map, and a third patient lung parenchyma state convolution encoding feature map; Calculating the mean of corresponding positions in the first patient lung parenchyma state convolutional coding feature map, the second patient lung parenchyma state convolutional coding feature map, and the third patient lung parenchyma state convolutional coding feature map, and performing atrous convolution coding on the obtained multi-scale patient lung parenchyma state convolutional coding feature map to obtain a multi-scale perceived patient lung parenchyma state feature map; The multi-scale patient lung parenchymal state convolutional coding feature map and the multi-scale perceived patient lung parenchymal state feature map are residually connected to obtain the lung parenchymal state shallow-semantic fusion coding feature map.

8. The automatic delineation system for tumor radiotherapy target area according to claim 7, characterized in that: The delineation unit is used to: input the shallow-semantic saliency fusion matrix of the lung parenchyma state into an image semantic segmentation delineator based on a Softmax function to obtain an automatic delineation result of the tumor radiotherapy target area.

9. A method for automatically delineating a tumor radiotherapy target area, characterized in that: include: Using the patient's identity as an index, extracting from a CT scan image database a CT scan image of the patient's lung parenchyma corresponding to the tumor radiotherapy target area to be delineated; Performing multi-scale feature extraction on the patient's lung parenchyma CT scan image to obtain superficial features of the patient's lung parenchyma state and semantic features of the patient's lung parenchyma state; Performing pulmonary parenchymal state semantic hierarchical enhancement on the semantic features of the patient's pulmonary parenchymal state to obtain semantic enhanced expression features of the patient's pulmonary parenchymal state, including: performing feature decoupling and spatial mapping on the semantic features of the patient's pulmonary parenchymal state to obtain a set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's pulmonary parenchymal state; Based on the set of semantic hyperbolic spatial modulation pixel granularity decoupling features of the patient's lung parenchyma state, explicitly weighting the semantic features of the patient's lung parenchyma state to obtain semantic enhanced expression features of the patient's lung parenchyma state; Multi-scale semantic interactive encoding of the patient's lung parenchymal state is performed on the shallow features of the patient's lung parenchymal state and the semantically enhanced expression features of the patient's lung parenchymal state to obtain shallow-semantic fusion encoding features of the lung parenchymal state, and based on the shallow-semantic fusion encoding features of the lung parenchymal state, the automatic delineation result of the tumor radiotherapy target area is obtained.

Images