Target region automatic segmentation method and system based on deep learning

Abstract

The invention relates to the technical field of medical image segmentation, in particular to a target region automatic segmentation method and system based on deep learning, and the method comprises the steps: S1, preparing a medical image source file to be sketched, and copying the independent medical image source file into two recognition image copies; s2, performing feature recognition on the recognized image copy by a target region recognition module and a benign recognition module, and performing range labeling on the medical image source file after recognition to form a primary feature recognition region; s3, comparing the primary feature recognition regions by using a deep learning model, and performing integrated sketching in combination with a map database to sketch a primary target region; s4, marking the medical image resource files one by one to form a primary target area. According to the method, human assistance can be involved during automatic sketching, error correction is facilitated, the sketching precision can be accurately improved after single training, the accuracy of maps in a database is ensured, and repeated sketching errors are avoided.

Description

Technical Field

[0001] The present invention relates to a target area segmentation method, in particular to an automatic target area segmentation method and system based on deep learning, and belongs to the technical field of medical image segmentation. Background Art

[0002] In the treatment of tumors, radiation is often used for treatment. Radiotherapy is to destroy local cancer cells. In order to avoid damaging the healthy tissues and organs of the human body during radiotherapy, it is usually necessary to take a CT localization map of the local area of ​​cancer cells before treatment, and then manually outline the area. The area of ​​these cancer cells is the target area. The segmentation of the target area is very important, which directly affects the difficulty, means and treatment effect of the treatment. In order to improve the outline accuracy as much as possible and reduce the workload of doctors, intelligent systems are often used in the existing technology to automatically outline the target area, and then put it into use after being reviewed by doctors, which can greatly improve work efficiency.

[0003] For example, the Chinese invention patent application document with publication number CN115762724A provides a method, device and system for automatically outlining the target area of ​​a medical image.

[0004] The automatic target area delineation method first uses a trained target area positioning model to locate the target area of ​​the medical image to be delineated to obtain multi-channel information of the target area; then, based on the multi-channel information, a trained target area segmentation model is used to segment the medical image to be delineated to obtain the target target area delineation result of the medical image to be delineated.

[0005] The invention can realize an end-to-end process with high delineation efficiency and strong versatility. Moreover, the entire delineation process does not require human participation, which can reduce the differentiation problems that may be caused by human factors and can better assist doctors in improving the quality and efficiency of radiotherapy plan design.

[0006] In the above document, the image is automatically outlined through the target area positioning model. However, the applicant believes that during the outline process, due to the different locations of lesions and the different spread of cancer cells, the outlined area is variable, and human intervention is required during the specific execution to perform intervention operations such as target area changes.

[0007] In view of this, the present invention is proposed. Summary of the invention

[0008] The purpose of the present invention is to provide a method and system for automatic target area segmentation based on deep learning in order to solve the above problems. It can intervene with human assistance during automatic outlining to facilitate error correction. After a single training, it can accurately improve the outlining accuracy, ensure the accuracy of the atlas in the database, and avoid repeated outlining errors.

[0009] The present invention achieves the above-mentioned object through the following technical solutions: a method and system for automatic target area segmentation based on deep learning, comprising step S1: preparing a medical image source file to be delineated, and copying a separate medical image source file into two recognition image copies;

[0010] Step S2: The target area recognition module and the benign area recognition module respectively perform feature recognition on the recognition image copy, and after recognition, range marking is performed on the medical image source file to form a primary feature recognition area;

[0011] Step S3: using the deep learning model to compare the initial feature recognition area, and integrating and outlining it in combination with the atlas library to outline the initial target area;

[0012] Step S4: Mark each medical image resource file one by one to form an initial target area, involve a deep learning module during marking, monitor changes in the target area, and make reference marks for the monitored area to form a reference graph;

[0013] Step S5: After forming the initial feature recognition area, a feature update graph can be formed, and after outlining the initial target area, a primary target area update graph is formed. The feature update graph, the primary target area update graph, and the reference graph are stored in the update module, and the doctor can view and modify them through the operation terminal. The modified data graph can replace the original data;

[0014] Step S6: The extracted parts of the graphics of each stage stored in the update module are stored in the extraction module, and the doctor checks the graphics of each stage through the operation terminal and can change the graphic data synchronously;

[0015] Step S7: The completed reference figure is finally formed into a solid figure file, which is then stored in the data module.

[0016] Furthermore, in step S1, the medical image source file is composed of continuous and gradually deepening medical graphics, and there are more than 30 medical graphics.

[0017] Furthermore, in step S2, the target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy and call different atlas libraries. The target recognition module only marks the target area, while the benign recognition module recognizes and marks the area except the cancer cell area.

[0018] Furthermore, the atlas library in step S3 stores the outlined atlas resources, and only the atlas corresponding to the location of the reference lesion is called during the outline.

[0019] Furthermore, in step S4, a convolutional neural network is used to extract medical graphics in the atlas library, perform feature recognition, and mark ranges on the medical graphics.

[0020] Furthermore, the medical graphics in the update module in step S5 do not participate in the reference range of feature recognition, and medical graphics that are difficult to recognize are also collected in the update module. The standards for difficult-to-recognize graphics are feature update graphics, initial target area update graphics, and reference graphics with a marking completion rate below 95%.

[0021] Furthermore, the extracted part in step S6 is randomly extracted at one-thirtieth, one-fiftieth or one-tenth.

[0022] The system of the automatic target area segmentation method based on deep learning includes a database, in which medical atlases are stored; a deep learning model, which is a convolutional neural network model, and also includes an input module and an operation end, wherein the input module is data-connected to an external data transmission device.

[0023] Furthermore, the deep learning model includes a feature recognition module, an update module and an extraction module.

[0024] Furthermore, the feature recognition module includes a target area recognition module and a benign recognition module.

[0025] The technical effects and advantages of the present invention are as follows: the present invention can intervene with human assistance during automatic outlining, which is convenient for error correction. After a single training, it can accurately improve the outlining accuracy, ensure the accuracy of the atlas in the database, and avoid repeated outlining errors. In actual application, the method can perform dual feature recognition on the medical image source file, and use the call to atlas comparison in the deep learning module to display and compare different features, improve subsequent outlining references, and have more comparison reference features when outlining the initial target area. In addition, the present invention can facilitate external professionals to assist in outlining graphics during the process, improve the accuracy of automatic outlining, and while improving the accuracy, it also reduces the difficulty of training and better reduces the intensity of doctors' outlining work. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the method steps of the present invention;

[0027] Figure 2 It is a schematic diagram of external intervention of the present invention;

[0028] Figure 3 It is a schematic diagram of the system structure of the present invention;

[0029] Figure 4 It is a schematic diagram of the composition of the deep learning model of the present invention;

[0030] Figure 5 It is a schematic diagram of the composition of the feature recognition module of the present invention; DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] See also Figure 1-Figure 5 shown.

[0033] Example 1

[0034] The method for automatic target segmentation based on deep learning includes step S1: preparing a medical image source file to be delineated, and copying a separate medical image source file into two identification image copies, wherein the medical image source file in step S1 is composed of continuous and gradually deepening medical graphics, and there are more than 30 medical graphics, so that the consistency of the graphic data can be guaranteed and a stable reference control can be obtained;

[0035] Step S2: The target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and after recognition, the range is marked on the medical image source file to form the initial feature recognition area. In step S2, the target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and call different atlas libraries. The target recognition module only marks the target area, while the benign recognition module recognizes and marks the area except the cancer cell area.

[0036] Step S3: using the deep learning model to compare the initial feature recognition area, and integrating and outlining it in combination with the atlas library to outline the initial target area. The atlas library in step S3 stores the outlined atlas resources, and only the atlas corresponding to the location of the reference lesion is called during the outlining;

[0037] Step S4: Mark each medical image resource file one by one to form an initial target area. During the marking, a deep learning module is introduced to monitor the changes in the target area, and reference marking is performed on the monitored area to form a reference graph. In step S4, a convolutional neural network is used to extract medical graphics in the atlas library, perform feature recognition, and perform range marking on the medical graphics;

[0038] Step S5: After forming the initial feature recognition area, a feature update graph can be formed. After outlining the initial target area, a primary target area update graph is formed. The feature update graph, the primary target area update graph and the reference graph are stored in the update module. The doctor can view and modify them through the operation terminal. The modified data graph can replace the original data. The medical graph in the update module in step S5 does not participate in the reference range of feature recognition, and the medical graphs that are difficult to identify are also collected in the update module. The standard of the difficult-to-identify graph is the feature update graph, the primary target area update graph and the reference graph with a marking completion degree below 95%;

[0039] Step S6: The extracted parts of the graphics of each stage stored in the update module are stored in the extraction module. The doctor checks the graphics of each stage through the operation terminal and can change the graphic data synchronously. The extracted parts in step S6 are randomly extracted according to one-thirtieth, one-fiftieth or one-tenth;

[0040] Step S7: The completed reference graph is finally formed into a solid graph file, which is then stored in the data module. The solid graph file is a set of files newly added to the atlas library and serves as an independent complete reference when performing target segmentation calls.

[0041] A system for automatic target area segmentation method based on deep learning includes a database storing medical atlases; a deep learning model, which is a convolutional neural network model, and also includes an input module and an operation end. The input module is data-connected to an external data transmission device. The deep learning model includes a feature recognition module, an update module and an extraction module. The feature recognition module includes a target area recognition module and a benign recognition module.

[0042] The present invention can intervene in human assistance during automatic outlining, which is convenient for error correction. After a single training, it can accurately improve the outlining accuracy, ensure the accuracy of the atlas in the database, and avoid repeated outlining errors. In practical applications, the method can perform dual feature recognition on the medical image source file, and use the call to atlas comparison in the deep learning module to display and compare different features, improve subsequent outlining references, and have more comparison reference features when outlining the initial target area. In addition, the present invention can facilitate external professionals to assist in outlining graphics during the process, improve the accuracy of automatic outlining, and while improving the accuracy, it also reduces the difficulty of training and better reduces the intensity of doctors' outlining work.

[0043] Example 2

[0044] This embodiment uses the same atlas library for initial feature recognition on the basis of Embodiment 1, that is, performs feature comparison twice on the same single atlas. The method for automatic target segmentation based on deep learning includes step S1: preparing a medical image source file to be delineated, and copying the single medical image source file into two copies of the recognition image. In step S1, the medical image source file is composed of continuous and gradually deepening medical graphics, and there are more than 30 medical graphics, so that the consistency of the graphic data can be guaranteed and a stable reference control can be obtained;

[0045] Step S2: The target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and after recognition, the range is marked on the medical image source file to form the initial feature recognition area. In step S2, the target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and call the same atlas library. The target recognition module only marks the target area, while the benign recognition module identifies and marks the area except the cancer cell area. In this way, the same atlas library is used for double recognition, and the overlapping area of ​​the mark can be focused on for detection;

[0046] Step S3: using the deep learning model to compare the initial feature recognition area, focusing on detecting and recalculating the overlapping area, and integrating and outlining in combination with the atlas library to outline the initial target area. The atlas library in step S3 stores the outlined atlas resources, and only the atlas corresponding to the location of the reference lesion is called during the outlining;

[0047] Step S4: Mark each medical image resource file one by one to form an initial target area. During the marking, a deep learning module is introduced to monitor the changes in the target area, and reference marking is performed on the monitored area to form a reference graph. In step S4, a convolutional neural network is used to extract medical graphics in the atlas library, perform feature recognition, and perform range marking on the medical graphics;

[0048] Step S5: After forming the initial feature recognition area, a feature update graph can be formed. After outlining the initial target area, a primary target area update graph is formed. The feature update graph, the primary target area update graph and the reference graph are stored in the update module. The doctor can view and modify them through the operation terminal. The modified data graph can replace the original data. The medical graph in the update module in step S5 does not participate in the reference range of feature recognition, and the medical graphs that are difficult to identify are also collected in the update module. The standard of the difficult-to-identify graph is the feature update graph, the primary target area update graph and the reference graph with a marking completion degree below 95%;

[0049] Step S6: The extracted parts of the graphics of each stage stored in the update module are stored in the extraction module. The doctor checks the graphics of each stage through the operation terminal and can change the graphic data synchronously. The extracted parts in step S6 are randomly extracted according to one-thirtieth, one-fiftieth or one-tenth;

[0050] Step S7: The completed reference graph is finally formed into a solid graph file, which is then stored in the data module. The solid graph file is a set of files newly added to the atlas library and serves as an independent complete reference when performing target segmentation calls.

[0051] A system for automatic target area segmentation method based on deep learning includes a database storing medical atlases; a deep learning model, which is a convolutional neural network model, and also includes an input module and an operation end. The input module is data-connected to an external data transmission device. The deep learning model includes a feature recognition module, an update module and an extraction module. The feature recognition module includes a target area recognition module and a benign recognition module.

[0052] Example 3

[0053] This embodiment, based on the second embodiment, performs a third atlas comparison when outlining, and verifies the accuracy of outlining again. The method for automatic target segmentation based on deep learning includes step S1: preparing a medical image source file to be outlined, and copying a separate medical image source file into two copies of the recognition image. In step S1, the medical image source file is composed of continuous and gradually deepening medical graphics, and there are more than 30 medical graphics, so that the consistency of the graphic data can be guaranteed and a stable reference control can be obtained;

[0054] Step S2: The target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and after recognition, the range is marked on the medical image source file to form the initial feature recognition area. In step S2, the target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy, and call the same atlas library. The target recognition module only marks the target area, while the benign recognition module identifies and marks the area except the cancer cell area. In this way, the same atlas library is used for double recognition, and the overlapping area of ​​the mark can be focused on for detection;

[0055] Step S3: Use the deep learning model to compare the initial feature recognition area, focus on detecting and recalculating the overlapping area, and integrate and outline it in combination with the atlas library. The atlas library used needs to remove the atlas used in the feature recognition area to outline the initial target area. This outline is combined with multiple atlases for targeted comparison, which can further avoid errors or inaccurate outlines. The atlas library in step S3 stores the outlined atlas resources, and only the atlas corresponding to the location of the reference lesion is called during the outline;

[0056] Step S4: Mark each medical image resource file one by one to form an initial target area. During the marking, a deep learning module is introduced to monitor the changes in the target area, and reference marking is performed on the monitored area to form a reference graph. In step S4, a convolutional neural network is used to extract medical graphics in the atlas library, perform feature recognition, and perform range marking on the medical graphics;

[0057] Step S5: After forming the initial feature recognition area, a feature update graph can be formed. After outlining the initial target area, a primary target area update graph is formed. The feature update graph, the primary target area update graph and the reference graph are stored in the update module. The doctor can view and modify them through the operation terminal. The modified data graph can replace the original data. The medical graph in the update module in step S5 does not participate in the reference range of feature recognition, and the medical graphs that are difficult to identify are also collected in the update module. The standard of the difficult-to-identify graph is the feature update graph, the primary target area update graph and the reference graph with a marking completion degree below 95%;

[0058] Step S6: The extracted parts of the graphics of each stage stored in the update module are stored in the extraction module. The doctor checks the graphics of each stage through the operation terminal and can change the graphic data synchronously. The extracted parts in step S6 are randomly extracted according to one-thirtieth, one-fiftieth or one-tenth;

[0059] Step S7: The completed reference graph is finally formed into a solid graph file, which is then stored in the data module. The solid graph file is a set of files newly added to the atlas library and serves as an independent complete reference when performing target segmentation calls.

[0060] A system for automatic target area segmentation method based on deep learning includes a database storing medical atlases; a deep learning model, which is a convolutional neural network model, and also includes an input module and an operation end. The input module is data-connected to an external data transmission device. The deep learning model includes a feature recognition module, an update module and an extraction module. The feature recognition module includes a target area recognition module and a benign recognition module.

[0061] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0062] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for automatic target segmentation based on deep learning, characterized by: The method comprises step S1: preparing a medical image source file to be delineated, and copying a single medical image source file into two recognition image copies; Step S2: The target area recognition module and the benign area recognition module respectively perform feature recognition on the recognition image copy, and after recognition, range marking is performed on the medical image source file to form a primary feature recognition area; Step S3: using the deep learning model to compare the initial feature recognition area, and integrating and outlining it in combination with the atlas library to outline the initial target area; Step S4: Mark each medical image resource file one by one to form an initial target area, involve a deep learning module during marking, monitor changes in the target area, and make reference marks for the monitored area to form a reference graph; Step S5: After forming the initial feature recognition area, a feature update graph can be formed, and after outlining the initial target area, a primary target area update graph is formed. The feature update graph, the primary target area update graph, and the reference graph are stored in the update module, and the doctor can view and modify them through the operation terminal. The modified data graph can replace the original data; Step S6: The extracted parts of the graphics of each stage stored in the update module are stored in the extraction module, and the doctor checks the graphics of each stage through the operation terminal and can change the graphic data synchronously; Step S7: The completed reference figure is finally formed into a solid figure file, which is then stored in the data module.

2. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: In step S1, the medical image source file is composed of continuous and gradually deepening medical graphics, and there are more than 30 medical graphics.

3. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: In step S2, the target recognition module and the benign recognition module respectively perform feature recognition on the recognition image copy and call different atlas libraries. The target recognition module only marks the target area, while the benign recognition module recognizes and marks the area except the cancer cell area.

4. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: The atlas library in step S3 stores the outlined atlas resources, and only the atlas corresponding to the location of the reference lesion is called during the outline.

5. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: In step S4, a convolutional neural network is used to extract medical graphics in the atlas library, perform feature recognition, and mark ranges on the medical graphics.

6. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: The medical graphics in the update module in step S5 do not participate in the reference range of feature recognition, and medical graphics that are difficult to recognize are also collected in the update module. The standards for difficult-to-recognize graphics are feature update graphics, initial target area update graphics, and reference graphics with a marking completion rate below 95%.

7. The method for automatic target area segmentation based on deep learning according to claim 1, characterized in that: The extracted parts in step S6 are randomly selected according to one-thirtieth, one-fiftieth or one-tenth.

8. A system for an automatic target area segmentation method based on deep learning, a system for applying the automatic target area segmentation method according to any one of claims 1 to 7, characterized in that: It includes a database, which stores medical maps; a deep learning model, which is a convolutional neural network model, and also includes an input module and an operation end, and the input module is data-connected to an external data transmission device.

9. The system for automatic target area segmentation method based on deep learning according to claim 8, characterized in that: The deep learning model includes a feature recognition module, an update module and an extraction module.

10. The system for automatic target area segmentation method based on deep learning according to claim 9, characterized in that: The feature recognition module includes a target area recognition module and a benign recognition module.

Images