Automatic segmentation method and device for complex-shaped target region in radiotherapy plan
Through the classification and segmentation cutting technology of target area model, combined with deep learning models, the complex target area is characterized and spliced, which solves the problem of complex target area segmentation in the existing technology, and realizes automatic and accurate target area identification and segmentation , reduces computing resource consumption and is suitable for multiple modal images.
Patent Information
- Application Number
- CN202411215764.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-13
AI Technical Summary
The existing technology is difficult to effectively and automatically segment target areas with wide coverage and complex shapes. Traditional methods are time-consuming, subjective and limited accuracy. Deep learning methods are prone to overfitting when processing complex data, and it is difficult to extract areas with severe shape changes. Sufficient features.
The image is initially positioned using the target area model, divided into fixed layer and variable layer feature maps, inference and splicing, using depth separation convolution, multiple 3D Unet, FPN and softmax functions, combined with dice loss function Training with the three-dimensional boundary loss function to achieve automatic and accurate segmentation of the target area.
It realizes automatic and accurate identification and segmentation of complex shape target areas, reduces computing resource consumption, improves recognition effect, is suitable for target areas with arbitrary shapes and wide coverage, and reduces the requirement for labeled data volume.
Smart Images
Figure CN119131060A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical image processing, and in particular to an automatic segmentation method and device for a target area with a complex shape in a radiotherapy plan. Background Art
[0002] In radiotherapy planning, the target segmentation of common tumors such as lung tumors and brain tumors has been relatively complete, but it is still difficult to outline other target areas with wide coverage and complex and changeable shapes. Usually such target areas have the following characteristics:
[0003] 1. Wide coverage, some even cover half or full body.
[0004] 2. The shape is complex and changeable, that is, the shape or size changes dramatically with the position.
[0005] For example, medulloblastoma requires the use of the whole brain and spinal cord target area. Another example is that in order to prevent the spread or metastasis of breast cancer patients, the target area is often expanded to the area between the lower chest and the upper clavicle.
[0006] Commonly used delineation methods in radiation therapy planning include:
[0007] 1. Doctors manually outline complex target areas on imaging data. This method is difficult and time-consuming, highly subjective, and has limited accuracy. In addition, doctors' experience and knowledge may vary from person to person without a unified standard, which can easily lead to errors.
[0008] 2. As the target area is wider and more complex in shape, the segmentation effect of traditional methods in the AI field and deep learning methods is poor. For example, the machine learning SVM algorithm is used to automatically identify and outline complex target areas. The decision boundary of this method is usually a complex hyperplane, which makes it difficult to intuitively explain the model decision process. In addition, since the performance of SVM is highly dependent on hyperparameters, the more parameters there are, the easier it is to overfit when processing complex data, and the worse the segmentation effect. For another example, deep learning networks such as nnunet are used to automatically identify and outline complex target areas. This method uses a three-dimensional sliding window to extract features during network training. The shape of a complex target area usually changes dramatically with position, and the spatial scale of the sliding window is limited. If the sliding window scale is too large, it will exceed the computer's video memory range, increase computer resource consumption and training costs, and if the sliding window size is too small, it will not be able to extract enough full-view features of the target area with complex shapes and dramatic changes, resulting in poor recognition effect.
[0009] In summary, there is currently no targeted segmentation method for complex-shaped targets, and the existing target or organ-at-risk segmentation methods are not suitable for complex-shaped targets. Summary of the invention
[0010] The present invention provides a method and device for automatically segmenting a target area with a complex shape in a radiotherapy plan, aiming to automatically and accurately segment a target area with a complex shape.
[0011] The present invention provides an automatic segmentation method for complex-shaped targets in radiotherapy plans, the method comprising: using a positioning target model to perform target positioning on an image of a patient containing a complex-shaped target, and obtaining a positioning target feature map of a distribution area of a preliminarily positioned target; based on changes in the complex-shaped target in the positioning target feature map, dividing the positioning target feature map into a fixed-layer feature map with stable target changes and a variable-layer feature map with obvious target changes; segmenting the variable-layer feature map to obtain a plurality of sub-variable-layer feature maps; performing target reasoning on the fixed-layer feature map and the plurality of sub-variable-layer feature maps, respectively, to obtain a fixed-layer target feature map and a plurality of sub-variable-layer target feature maps; splicing the fixed-layer target feature map and the plurality of variable-layer target feature maps to obtain a complex-shaped target feature map.
[0012] Preferably, the positioning target area model includes a depth-separable convolution, multiple series-connected 3D Unets, FPNs, and a softmax function.
[0013] Preferably, the method of dividing the positioning target area feature map into a fixed layer feature map with stable target area changes and a variable layer feature map with obvious target area changes based on the changes of the complex shape target area in the positioning target area feature map includes: dividing the positioning target area feature map into a fixed layer feature map with stable target area changes and a variable layer feature map with obvious target area changes according to at least one of the shape changes and position changes of the complex shape target area in the positioning target area feature map.
[0014] Preferably, the shape change of the complex-shaped target area is determined by the following steps: calculating the number of connected domains in each layer of the thermal map of the positioning target area feature map; determining the part of the target area where the shape change is stable and the part of the target area where the shape change is drastic based on the comparison result of the number of connected domains in each layer of the thermal map and the first threshold, or the change in the number of connected domains in two adjacent layers of the thermal map.
[0015] Preferably, the target position change of the complex-shaped target area is determined by the following steps: calculating the number of pixels whose pixel probability values in each layer of the thermal map of the positioning target area feature map are greater than the second threshold as the target area area; and determining the part where the target position change is stable or the part where the target position change is obvious based on the difference in the target area areas between two adjacent layers of the thermal map.
[0016] Preferably, the segmented cropping of the variable layer feature map to obtain multiple sub-variable layer feature maps includes: cutting in the XY direction of the variable layer feature map with the target area position as the center position; and cutting in the Z direction of the variable layer feature map in the form of upper and lower sub-layer stacking to retain the target area boundary information.
[0017] Preferably, the target area reasoning is performed on the fixed layer feature map and the multiple sub-variable layer feature maps respectively to obtain the fixed layer target area feature map and the multiple sub-variable layer target area feature maps, including: using a fixed layer reasoning model to perform target area reasoning on the fixed layer feature map to obtain a fixed layer target area feature map; using a variable layer reasoning model to perform target area reasoning on the multiple sub-variable layer feature maps respectively to obtain multiple sub-variable layer target area feature maps; wherein the fixed layer reasoning model and the variable layer reasoning model are models with the same structure but different parameters.
[0018] Preferably, the fixed layer inference model and the variable layer inference model both include: a first 2D Unet group, a second 2D Unet group, a 3DUnet group and three convolutional layers connected in series, wherein the first 2D Unet group outputs the two-dimensional features extracted from the input feature map to the first convolutional layer, the second 2D Unet group inputs the two-dimensional feature map extracted from the input feature map to the second convolutional layer, and the 3DUnet group inputs the three-dimensional feature map extracted from the input feature map to the third convolutional layer.
[0019] Preferably, when training the fixed layer reasoning model, the dice loss function and Hausdorff distance are used to calculate the training loss of the fixed layer reasoning model; when training the variable layer reasoning model, the three-dimensional boundary loss function and the three-dimensional sphere surface loss function are used to calculate the training loss of the variable layer reasoning model.
[0020] The present invention also provides an automatic segmentation device for a target area with a complex shape in a radiotherapy plan. The device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, the steps of the automatic segmentation method for a target area with a complex shape in a radiotherapy plan are implemented as described above.
[0021] The present invention provides a method and device for automatically segmenting a target area with a complex shape in a radiotherapy plan. The method performs target area reasoning and feature splicing based on the division and segmentation clipping results of a positioning target area feature map, and can automatically and accurately identify / outline / segment a target area with any shape and / or a complex shape with a wide coverage. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a first flow chart of the method for automatically segmenting a target area with a complex shape in a radiotherapy plan provided by the present invention;
[0023] Figure 2 is a second flow chart of the method for automatically segmenting a target area with a complex shape in a radiotherapy plan provided by the present invention;
[0024] Figure 3 yes Figure 2 The data preprocessing flow chart of the data preprocessing module shown;
[0025] Figure 4 This is a schematic diagram of the layers of the whole brain and spinal cord;
[0026] Figure 5 It is the network structure diagram of the coarse segmentation model;
[0027] Figure 6 It is the network structure diagram of the secondary segmentation model;
[0028] Figure 7a It is the loss function explanation diagram of the fixed layer inference model;
[0029] Figure 7b It is the loss function explanation diagram of the variable layer inference model;
[0030] Figure 8 It is a schematic diagram of the surface loss of a three-dimensional sphere;
[0031] Fig. 9 yes Figure 2 Model prediction flow chart of the model inference module shown;
[0032] Fig.10 It is the schematic diagram of cutting in XY direction of the target center;
[0033] Fig.11 It is a schematic diagram of stacking cutting in the Z direction. DETAILED DESCRIPTION
[0034] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In the subsequent description, the use of suffixes such as "module", "component" or "unit" used to represent elements is only to facilitate the description of the present invention and has no specific meaning in itself. Therefore, "module", "component" or "unit" can be used in a mixed manner.
[0035] The present invention discloses an automatic segmentation technology for complex-shaped target areas (complex target areas for short) in radiotherapy plans. In the training process, the complex target area to be identified (or outlined, segmented) is first divided into a fixed layer and a variable layer according to a set rule / mechanism, and then feature learning is performed on the two groups of data of the fixed layer and the variable layer to obtain a fixed layer model and a variable layer model. When applied, the positioning target area model with learned whole-body features is first used to find the approximate area of the target area, and then the approximate area of the identified target area is divided into a fixed layer and a variable layer through an adaptive module, and then the fixed layer model and the variable layer model are accurately segmented on this basis.
[0036] See also Figure 1 , an automatic segmentation method for a target area with a complex shape in a radiotherapy plan includes:
[0037] Step S1: using the positioning target model, the target image of the patient containing the target of complex shape is positioned to obtain a positioning target feature map of the distribution area of the target of complex shape which has been preliminarily positioned.
[0038] The positioning target area model is a model that has learned whole body features.
[0039] Step S2: Based on the changes of the complex shape target area in the positioning target area feature map, the positioning target area feature map is divided into a fixed layer feature map with stable target area changes and a variable layer feature map with obvious target area changes.
[0040] The change of the complex-shaped target area in the positioning target area feature map includes at least one of the shape change and the position change of the complex-shaped target area in the positioning target area feature map.
[0041] Among them, the number of connected domains in each layer of the heat map of the positioning target area feature map is calculated to determine the shape change of the complex shape target area according to the number of connected domains in each layer of the heat map. For example, the shape change of the complex shape target area is determined according to the comparison result of the number of connected domains in each layer of the heat map and the first threshold. Specifically, if the number of connected domains in each layer of the heat map is greater than the first threshold, it is determined that the target shape of the area changes dramatically, otherwise it is determined that the target shape of the area changes stably. For another example, according to the change in the number of connected domains in the two adjacent layers of the heat map, the part where the target shape changes stably and the part where the target shape changes dramatically are determined. Specifically, if the number of connected domains in the two adjacent layers of the heat map changes too much, it is determined that the target shape of the area with a large number of connected domains changes dramatically, otherwise it is determined that the target shape of the area changes stably.
[0042] The number of pixels whose pixel probability value in each layer of the thermal map of the positioning target area feature map is greater than the second threshold is calculated as the target area area, so as to determine the part where the target area position changes stably or the part where the target area position changes significantly according to the difference in the target area area between two adjacent layers of the thermal map. Specifically, if the difference in the target area area between two adjacent layers of the thermal map is too large, it is determined that the target area position of the large target area changes significantly, otherwise it is determined that the target area position of the area changes stably.
[0043] Step S3: segmentally cropping the variable layer feature map to obtain multiple sub-variable layer feature maps.
[0044] In the XY direction of the variable layer feature map, cutting is performed with the target area position as the center position, and in the Z direction of the variable layer feature map, cutting is performed in a stacked manner of upper and lower sub-layers to retain the target area boundary information.
[0045] Step S4: performing target area inference on the fixed layer feature map and the plurality of sub-variable layer feature maps respectively to obtain a fixed layer target area feature map and a plurality of sub-variable layer target area feature maps.
[0046] A fixed layer inference model (referred to as the fixed layer model) is used to perform target area inference on the fixed layer feature map to obtain a fixed layer target area feature map. A variable layer inference model (referred to as the variable layer model) is used to perform target area inference on multiple sub-variable layer feature maps to obtain multiple sub-variable layer target area feature maps. The fixed layer model and the variable layer model are respectively a model that has learned fixed layer features and a model that has learned variable layer features, and both models have the same structure but different parameters.
[0047] Step S5: splicing the fixed layer target area feature map and a plurality of the variable layer target area feature maps to obtain a complex shape target area feature map, that is, an image of the complex shape target area that has been identified / outlined / segmented.
[0048] The present invention can automatically and accurately identify / delineate / segment a target area of arbitrary shape and / or complex shape with a wide coverage.
[0049] Below through Figures 2 to 11 The present invention is described in detail.
[0050] The main innovation of the method of this patent application is to design an automatic segmentation scheme for complex target areas in radiotherapy plans, so that the segmentation of complex target areas is more accurate and the radiotherapy plan is more complete.
[0051] This method mainly involves three parts: data preprocessing module, modular cascade network and model inference module. Figure 2 , the process is as follows:
[0052] Step S10: The data preprocessing module mainly performs data denoising, data normalization, resolution resampling, direction unification, manually divides the data into fixed layer and variable layer and labels them, and randomly rotates the data according to a certain ratio.
[0053] See also Figure 3 , step S10 may include: inputting a manually annotated data set (step S101), in which the image modality may be CT or MR. Using a Gaussian filter to smooth the image for denoising (step S102). Considering that most target areas are soft tissues, the CT image is truncated according to the window width and window position of the soft tissue so that the value is within a certain range, and then normalized using a normal distribution (step S103). Resolution resampling (step S104), specifically, calculating the average resolution of all data to obtain an average resolution, and then resampling all data to the average average to facilitate training. Unify the coordinate system directions of all data (step S105). Divide the complex target area into a variable layer and a fixed layer (step S106), and the principle of distinguishing the target area stratification is: the first principle is to separate the part with severe deformation from the part with stable deformation. For example, in the target area of breast cancer, the shape of the breast part changes slightly, while the part from the subclavian to the supraclavicular part changes dramatically. Since there are many critical organs nearby, this target area will have multiple connected domains with irregular shapes. Therefore, the part from the subclavian to the supraclavicular part should be recorded as the variable layer, and the breast part should be recorded as the fixed layer. The second principle is to separate the part of the target area with obvious position changes from the part with relatively fixed position. For example, the whole brain and spinal cord, see Figure 4 , the brain area is a fixed layer, and the spinal cord area is a variable layer because it is curved in the z direction. According to the above principles, the complex target area can be manually divided into a fixed layer and a variable layer and labeled for model learning. The data is randomly rotated (0-180) degrees (step S107). The processed data is output (step S108).
[0054] Step S20: Design a modular cascade network model to identify complex target areas. The modular cascade network model is divided into a coarse segmentation model and a secondary segmentation model (i.e., a high-precision segmentation model). The positioning target area model adopts the coarse segmentation model, and the fixed layer model and the variable layer model adopt the secondary segmentation model.
[0055] See also Figure 5The network structure of the rough segmentation model is composed of a variety of convolution structures, including the network structure of the positioning target area model and the adaptive allocation module. The positioning target area model includes a depthwise separable convolution (Depthwise Separable Convolution), multiple 3D Unets connected in series, a feature pyramid network (FPN), and a softmax function. Among them, the depthwise separable convolution is a special convolution operation, which consists of two parts: depthwise convolution and pointwise convolution. The reason for using the depthwise separable convolution is that the network has a small amount of calculation and can improve the network reasoning speed. 3D Unet is composed of 3D Convolution convolution, which can extract context information and help detect Z-axis edge information. FPN enables the network to effectively extract features of different scales by constructing a multi-scale feature pyramid. The depthwise separable convolution, 3DUnet, and FPN are combined, and a multi-dimensional feature map is output after a softmax. Then, the feature map separation operation is performed through the adaptive allocation module to output two different feature maps of fixed layers and variable layers.
[0056] Among them, the judgment mechanism of the adaptive allocation module is as follows: 1. Calculate the number of pixels (or area) in each layer of the heat map with a probability greater than a certain threshold. When the area difference between the two layers is 2n (n = 1, 2, ..., 10) unit pixels, it can be considered that the gradient changes greatly, and the layers with large gradient changes are classified. Finally, statistical records are made in the Z direction according to the specific categories of each layer. 2. Calculate the number of connected domains in each layer of the heat map, and make statistical records of the connected domains with unreasonable changes according to specific problems. The standard for unreasonable changes in the number of connected domains is: the current layer has two-dimensional connected domains of n, and according to a certain traversal method (from top to bottom, from bottom to top), if the number of connected domains in the second layer is 2n, it is unreasonable. Finally, the statistical results are summarized according to these two judgment mechanisms, and the data is finally segmented into variable layers and fixed layers.
[0057] See also Figure 6The network structure of the secondary segmentation model consists of an improved 3Dunet and three groups of serial convolutional layers. The improved 3Dunet extracts two-dimensional features from two groups of 2D Unet (i.e., the first 2D Unet group and the second 2D Unet group) and a group of 3DUnet (i.e., the 3D Unet group) to extract context information. The network structure then passes the three features to three serial convolutional layers and finally outputs the features. Specifically, the first 2D Unet group outputs the two-dimensional features extracted from the input feature map to the first convolutional layer, the second 2D Unet group inputs the two-dimensional feature map extracted from the input feature map to the second convolutional layer, and the 3DUnet group inputs the three-dimensional feature map extracted from the input feature map to the third convolutional layer.
[0058] Among them, the input of the network structure of the secondary segmentation model is the fixed layer feature map and the variable layer feature map. Two sets of features are input respectively during model training. Finally, two models, a fixed layer inference model and a variable layer inference model, are trained based on the two sets of features, which are suitable for fixed layer target area reasoning / prediction and variable layer target area reasoning / prediction respectively.
[0059] The present invention designs a cascade loss function to perform specific loss function optimization and specific loss calculation for different models.
[0060] See also Figure 7a When training the fixed layer inference model, the dice loss function and Hausdorff distance are used to calculate the training loss of the fixed layer inference model. The dice loss calculates the three-dimensional mask coverage, and the Hausdorff distance loss calculates the distance between two sets of points to balance the boundary.
[0061] See also Figure 7b When training the variable layer inference model, the training loss of the variable layer inference model is calculated using the three-dimensional boundary loss function and the three-dimensional sphere surface loss function. The three-dimensional boundary loss function calculates the mask boundary loss, and the formula is: boundary_loss=dice((mask_t-mask_t1),mask_p); where mask_t is the true value, mask_t1 is the mask after mask_t is expanded, and mask_p is the predicted value; the three-dimensional sphere surface loss function calculates the surface loss of the complex target area, see Figure 8 The schematic diagram of loss calculation is shown. The dotted line is the predicted contour, the solid line is the real contour, and the black area is the loss that needs to be calculated for the deviation. The calculation formula for the surface loss of a three-dimensional sphere is: Among them, S A and S A are the surfaces of objects A and B, respectively. D(B,δ) represents the area where the surface distance of object B is less than the threshold δ. D(A,δ) represents the area where the surface distance of object A is less than the threshold δ.
[0062] Among them, when training different complex target areas, the weight of the loss function can be dynamically adjusted to facilitate learning of special features. For example, when identifying changing layers, some target area boundaries are not clear and difficult to learn. At this time, increasing the weight of the 3D boundary loss will help capture boundary information.
[0063] Step S30: The model reasoning module uses the positioning target area model to find the approximate location of the target area, and then uses the fixed layer model and the variable layer model to reason about the fixed layer and the variable layer of the found target area respectively, and finally splices the reasoning results and outputs them.
[0064] See also Fig. 9 , the segmentation / prediction process of step S30 may include: input data (step S301). Use the positioning target area model to perform target area segmentation and output the segmented feature map (step S302). Call the adaptive allocation module to divide the feature map of step 302 into fixed layer features and variable layer features (step S303). Use the fixed layer model and the variable layer model to perform predictions respectively (step S304). Specifically, since the feature changes of the fixed layer are small, the entire fixed layer feature map is directly sent to the fixed layer model network for direct inference, while the feature changes of the variable layer are large, so it is necessary to segment the variable layer feature map according to certain rules, and then send each segment separately to the variable layer model network for separate prediction. The segmentation rules are: in the two-dimensional XY direction, cutting is performed according to the center position of the target area, and in the Z direction, cutting is performed in a stacked manner between the upper and lower layers, in order to retain the boundary information. Taking the whole brain and spinal cord target area as an example, refer to Fig. 9 Each segment in the XY direction is cut with the target area as the center, and the upper and lower parts in the Z direction are stacked, for example, the upper and lower layers overlap by 50%, to enhance the context connection. The results predicted by the two models are combined into an overall feature through feature splicing (step S305). Output features (step S306).
[0065] The cascade network model designed by the present invention first locates the rough distribution range of the target area on the original image, and then adaptively matches the corresponding high-precision segmentation model according to the rough position and size of the target area. Among them, when training the high-precision segmentation model, different cascade loss functions are used for different features to improve the convergence speed and robustness of the model. When predicting, the sliding window route is intelligently designed according to the position and shape changes of the target area to improve the efficiency of model prediction.
[0066] The present invention can also provide an automatic segmentation device for a target area with a complex shape in a radiotherapy plan, the device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, can implement the steps of the above-mentioned method for automatic segmentation of a target area with a complex shape in a radiotherapy plan.
[0067] Compared with traditional methods and other deep learning algorithms, the present invention has the following advantages: 1. It can process target areas of arbitrary shapes or wide coverage; 2. The cascade model reduces the resource requirements for computer hardware and saves computing overhead; 3. It can achieve high-precision sub-pixel segmentation and recognition; 4. It is applicable to multiple modal images, including: CT, MR, etc.; 5. It greatly reduces the requirement for the amount of labeled data, and is a small sample learning technology.
[0068] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the present invention shall be within the scope of the present invention.
Claims
1. A method for automatically segmenting a target area with complex shapes in radiotherapy planning, characterized in that: The method comprises: Using the positioning target model, the target is positioned on the patient's image containing the target of complex shape, and a positioning target feature map of the distribution area of the target of complex shape has been preliminarily positioned; Based on the change of the complex shape target area in the positioning target area feature map, the positioning target area feature map is divided into a fixed layer feature map with stable target area change and a variable layer feature map with obvious target area change; Cutting the variable layer feature map in sections to obtain a plurality of sub-variable layer feature maps; Performing target area reasoning on the fixed layer feature map and the plurality of sub-variable layer feature maps respectively to obtain a fixed layer target area feature map and a plurality of sub-variable layer target area feature maps; The fixed layer target area feature map and a plurality of variable layer target area feature maps are spliced to obtain a target area feature map with a complex shape.
2. The method according to claim 1, characterized in that The positioning target area model includes sequentially connected depth separable convolutions, multiple series-connected 3D Unets, FPNs, and softmax functions.
3. The method according to claim 1, characterized in that: The method of dividing the positioning target area feature map into a fixed layer feature map with stable target area changes and a variable layer feature map with obvious target area changes based on the changes of the complex shape target area in the positioning target area feature map comprises: According to at least one of the shape change and position change of the complex shape target in the positioning target feature map, the positioning target feature map is divided into a fixed layer feature map with stable target change and a variable layer feature map with obvious target change.
4. The method according to claim 3, characterized in that The shape change of the complex shape target area is determined by the following steps: Calculating the number of connected domains in each layer of the thermal map of the positioning target area feature map; According to the comparison result of the number of connected domains in each layer of the heat map and the first threshold, or the change in the number of connected domains in two adjacent layers of the heat map, the part of the target area with stable shape change and the part of the target area with drastic shape change are determined.
5. The method according to claim 3, characterized in that: The target area position change of the complex shape target area is determined by the following steps: Calculate the number of pixels whose pixel probability values are greater than a second threshold in each layer of the thermal map of the positioning target area feature map as the target area area; According to the difference in the target area area between two adjacent layers of thermal maps, the part where the target area position changes stably or the part where the target area position changes significantly is determined.
6. The method according to claim 1, characterized in that The segmented clipping of the variable layer feature map to obtain a plurality of sub-variable layer feature maps comprises: In the XY direction of the variable layer feature map, cutting is performed with the target area position as the center position; In the Z direction of the variable layer feature map, cutting is performed in a stacked manner of two upper and lower sub-layers to retain the target area boundary information.
7. The method according to claim 1, characterized in that The performing target area reasoning on the fixed layer feature map and the plurality of sub-variable layer feature maps respectively to obtain the fixed layer target area feature map and the plurality of sub-variable layer target area feature maps comprises: Using a fixed layer inference model, target area inference is performed on the fixed layer feature map to obtain a fixed layer target area feature map; Using the variable layer inference model, target area inference is performed on the plurality of sub-variable layer feature maps respectively to obtain a plurality of sub-variable layer target area feature maps; The fixed layer reasoning model and the variable layer reasoning model are models with the same structure but different parameters.
8. The method according to claim 7, characterized in that The fixed layer inference model and the variable layer inference model both include: a first 2D Unet group, a second 2D Unet group, a 3DUnet group and three convolutional layers connected in series, wherein the first 2D Unet group outputs the two-dimensional features extracted from the input feature map to the first convolutional layer, the second 2D Unet group inputs the two-dimensional feature map extracted from the input feature map to the second convolutional layer, and the 3DUnet group inputs the three-dimensional feature map extracted from the input feature map to the third convolutional layer.
9. The method according to claim 7 or 8, characterized in that: When training the fixed layer inference model, using the dice loss function and the Hausdorff distance to calculate the training loss of the fixed layer inference model; When training the variable layer reasoning model, the training loss of the variable layer reasoning model is calculated using a three-dimensional boundary loss function and a three-dimensional sphere surface loss function.
10. An automatic segmentation device for complex target areas in radiotherapy planning, characterized in that: The device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, the steps of the method for automatic segmentation of a target area with a complex shape in a radiotherapy plan as described in any one of claims 1 to 9 are implemented.