A 3D medical image segmentation model and its training method and application

By introducing Prototype feature vectors and dynamic convolution kernel generators, the problems of excessive parameters and inflexible expansion in existing technologies are solved, achieving efficient segmentation of organs and tumors in multi-class 3D medical images, and improving segmentation accuracy and model flexibility.

CN114820636BActive Publication Date: 2024-11-15NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210549492.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-20
Publication Date
2024-11-15
Estimated Expiration
2042-05-20

AI Technical Summary

Technical Problem

Existing technologies suffer from problems such as excessive parameters, high computational complexity, and inflexible expansion in the segmentation of three-dimensional medical images of multiple organs and tumors, especially in the case of some labeled datasets where it is difficult to perform multi-task segmentation effectively.

Method used

A task controller module was designed, which introduces Prototype feature vectors and adopts dynamic convolution. Through a dynamic convolution kernel generator and a dynamic segmentation head module, lightweight and flexible multi-class 3D medical image organ and tumor segmentation is achieved.

Benefits of technology

It improves the accuracy of segmentation results and the flexibility of network models, reduces computational complexity, and exhibits a higher Dice coefficient in multi-task segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114820636B_ABST
    Figure CN114820636B_ABST
Patent Text Reader

Abstract

The present invention discloses a three-dimensional medical image segmentation model and its training method and application. The model is based on a deep learning convolutional neural network. The network model consists of a prototype computing module, an encoder-decoder module, a task controller module, a dynamic convolution kernel generator module, and a dynamic segmentation head module. The model designs a task controller and introduces a prototype feature vector to guide the network to perform segmentation task discrimination. It uses dynamic convolution to flexibly handle different segmentation tasks, solves the partial labeling problem in the field of medical image segmentation, achieves the purpose of multi-category three-dimensional medical image organ and tumor segmentation, solves the problem of too many parameters and inflexible task expansion in the current mainstream method, and has a certain improvement in the segmentation results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a three-dimensional medical image segmentation model and a training method and application thereof, belonging to the technical field of image segmentation. Background Art

[0002] Automatic segmentation of organs and tumors using computed tomography (CT) is one of the most basic but also the most challenging tasks in medical image analysis. It plays a vital role in various computer-aided diagnosis tasks, such as analyzing lesion contours, making surgical plans, and performing 3D reconstruction. However, due to the limitations of labor costs and expertise, it is currently difficult to perform voxel-level annotation of multiple organs and tumors in large datasets. Therefore, most basic datasets are currently used for segmentation tasks of only one type of organ or tumor, and all organs and tumors that are not relevant to the task are annotated as background. For example, the LiTS dataset only has annotations for the liver and liver tumors, while the KiTS dataset only provides annotations for the kidneys and kidney tumors. Therefore, one of the major challenges facing the segmentation of multi-organ and tumor CT images is the so-called part labeling problem, that is, how to learn the representation of multiple organs and tumors under the supervision of these partly annotated images.

[0003] In recent years, many researchers have studied the problem of partial labeling of multiple organs and tumors. In the published literature, the mainstream method is to divide the partially labeled dataset into several fully labeled subsets, and then train a specific network on each subset for a specific segmentation task to solve the partial labeling problem (Multi-Nets), such as the network model for pancreatic segmentation proposed by Zhu et al. (Zhu Z, Xia Y, Xie L, et a l. Multi-scale coarse-to-fine segmentation for screening pancreatic ductal adenoc arcinoma[C] / / International conference on medical image computing and compute r-assisted intervention. Springer, Cham, 2019: 3-12.), and the network model for kidney segmentation proposed by Yu et al. (Yu Q, Shi Y, Sun J, et al. Crossbar-net: A novel convolutional neural networkfor kidney tumor segmentation in ct images[J]. IEEE trans actions on image processing, 2019, 28(8): 4060-4074.). Undoubtedly, this intuitive and simple strategy increases the computational complexity greatly.Another commonly used solution is to design a multi-head network, such as the Med3d proposed by Chen et al. (Chen S, Ma K, Zheng Y. Med3d: Transfer learning for 3D medical image analysis [J]. arXiv preprint arXiv: 1904.00625, 2019.), and the PIPO-FAN network model proposed by Fang et al. (Fang X, Yan P. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction [J]. IE EE Transactions on Medical Imaging, 2020, 39 (11): 3619-3629.). They are both composed of a shared encoder and multiple decoding heads for specific tasks. During the training phase, when each partially labeled data is input into the network, only the parameters in a specific corresponding decoding head will be trained. At this time, the parameters of other decoding heads will be frozen. Compared with the multi-network design, the multi-head network reduces the computational complexity to a certain extent. However, in the testing process of the multi-head network, only specific decoding heads will be used, and the training and inference of other decoding heads are wasted. In addition, the architecture of the multi-head network is not easy to extend to newly labeled tasks and lacks flexibility. DoDNet ​​(Zhang J, Xie Y, Xia Y, et al. DoDNet: Learning to segment multi-organ and tumors from multiple partially labeled datasets [C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021: 1195-1204.) proposed by Zhang et al. has flexible extension, but the hard coding method used makes the generalization ability of the network model insufficient.

[0004] In summary, there are still some problems in the field of medical image segmentation, such as lack of high-quality labeled data, high labor costs, small data volume, small segmentation targets, and partial labeling. For the partial labeling problem, the current mainstream methods have the disadvantages of too many parameters and inflexible expansion. Summary of the invention

[0005] The purpose of the present invention is to propose a three-dimensional medical image segmentation model, by designing a task controller module and introducing the concept of Prototype to guide the network to perform segmentation task discrimination, using dynamic convolution to flexibly handle different segmentation tasks, thereby achieving the purpose of multi-category three-dimensional medical image organ and tumor segmentation, and solving the problem of too many parameters and inflexible task expansion in the current mainstream method. The present invention makes the entire network model more lightweight, robust and flexible by using dynamic convolution and introducing Prototype feature vectors, and at the same time the segmentation results are improved to a certain extent.

[0006] The main contents of the technical solution of the present invention are as follows:

[0007] In a first aspect, the present invention provides a training method for a three-dimensional medical image segmentation model, the training method comprising:

[0008] Information encoding is performed according to a segmentation task of a three-dimensional medical image as task encoding information;

[0009] Acquire a single annotated data set of multiple categories consistent with the segmentation task, wherein the single annotated data set is a set of at least 50 three-dimensional medical images and single annotation information, i.e., labels. It should be noted that the single annotation information refers to the annotation information of a specific single organ and a corresponding organ tumor accurately annotated by an expert based on the corresponding three-dimensional medical image, and unify the pixels and formats of all the three-dimensional medical images contained in the single annotated data set as the original image;

[0010] Calculate the Prototype feature vector corresponding to a single labeled dataset of each category;

[0011] The encoder performs downsampling operations to extract original image features and high-level semantic information as downsampling information; the decoder receives the feature vectors of each corresponding level in the encoder as input, then upsamples to restore the spatial information of the original image and outputs a pre-segmented feature map that is consistent with the input size;.

[0012] After global pooling of the task encoding information and down-sampling information, the GAP feature vector and the Prototype feature vector corresponding to each category of data set are input into the 3D convolution layer for feature extraction, and the output is a feature vector that can guide the network model to process a specific task, that is, a feature map containing task information;

[0013] Using a dynamic convolution kernel to generate convolution kernel parameters from the feature map containing task information, wherein the convolution kernel parameters are used for the convolution layer of segmentation;

[0014] A dynamic segmentation head is used to take the convolution kernel parameters as parameters of the convolution operation, and the pre-segmented image is subjected to final image segmentation to obtain a segmentation result;

[0015] The loss is calculated for the segmentation result and the label image, and the parameters are trained through the gradient descent strategy. Finally, after at least several hundred rounds of training, a network model with better segmentation results is obtained.

[0016] Furthermore, the calculation of the Prototype feature vector corresponding to the single annotated data set of each category includes the following specific operation steps: using the segmentation network model as a pre-trained model, inputting the original image for segmentation processing, taking the feature map of the last layer, recorded as the segmentation feature map, using the segmentation feature map as the feature map of the original image in the latent space, and then performing trilinear interpolation on the segmentation feature map to obtain an interpolation feature map so that its size is consistent with the original image; at the same time, using the image label as a mask Mask of the organ and tumor; then, multiplying the mask Mask with the interpolation feature map to obtain the Prototype information corresponding to each original image in the single annotated data set of this category, and finally accumulating the Prototype information corresponding to each original image in this category and averaging it to obtain the Prototype feature vector that highly represents the single annotated data set of this category; it should be emphasized that the Prototype feature vector obtained here is the feature vector corresponding to a class of data sets, not the feature vector corresponding to each image in the data set.

[0017] Preferably, the loss function adopts cross entropy loss, and the gradient descent strategy specifically adopts stochastic gradient descent with momentum of 0.99, and the learning rate is initially 0.01 and decays with continuous training.

[0018] In a second aspect, the present invention provides a three-dimensional medical image segmentation model, the model comprising:

[0019] Prototype calculation module: used to calculate the prototype feature vector corresponding to a single labeled data set of each category;

[0020] The encoder-decoder module of the U-shaped network structure: the encoder is used to extract the original image features and high-level semantic information as down-sampling information; the decoder receives the feature vectors of each corresponding layer in the encoder as input, up-samples to restore the spatial information of the original image, and outputs a pre-segmented feature map consistent with the input size;

[0021] Task controller module: uses a 3D convolutional layer to extract features from the input task encoding information, the GAP feature vector obtained after global pooling of the down-sampling information, and the Prototype feature vector corresponding to each type of data set, and outputs a feature vector that can guide the network model to process a specific task, that is, a feature map containing task information;

[0022] Dynamic convolution kernel generator module: used for receiving the feature map containing task information and generating convolution kernel parameters in the convolution layer for segmentation;

[0023] Dynamic segmentation head module: used to receive the convolution kernel parameters and use them as parameters of the convolution operation, and perform final image segmentation on the pre-segmented image output by the decoder to obtain a segmentation result.

[0024] In a third aspect, the present invention provides an application of using the above segmentation model to perform segmentation processing on medical images. The method uses the above segmentation model to perform segmentation processing on medical images and outputs segmentation results of different organs and corresponding tumors.

[0025] Beneficial Effects

[0026] The 3D medical image segmentation model provided by the present invention introduces Prototype, a feature vector that highly represents category information, to guide the network to perform segmentation task discrimination and accurate guidance of multi-task segmentation. It uses dynamic convolution to flexibly handle different segmentation tasks, thereby improving the segmentation results and achieving the purpose of multi-category 3D medical image organ and tumor segmentation. It solves the problem of too many parameters and inflexible task expansion in the current mainstream methods, and at the same time improves the accuracy of the segmentation results to a certain extent.

[0027] The present invention provides a three-dimensional medical image segmentation model based on a deep learning convolutional neural network. The deep learning convolutional neural network is an end-to-end network model that reduces the complexity of network design. The designed task controller module enables it to guide the network to perform image segmentation for specific tasks. Dynamic convolution is introduced, and a dynamic convolution kernel generator module and a dynamic segmentation head module are proposed. Unlike traditional fixed multiple segmentation heads, the dynamic segmentation head of the present invention can make adaptive adjustments to input and assigned tasks, greatly improving efficiency and flexibility.

[0028] The method for 3D medical image segmentation using the above 3D medical image segmentation model provided by the present invention has a higher Dice coefficient value. The specific implementation method of the present invention has also verified the effectiveness and accuracy of the model algorithm by conducting relevant experiments on multiple data sets, and has conducted comparative experiments with the current mainstream algorithms, such as Figure 6 The following are some results of the experiment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a framework diagram of the overall processing flow of the segmentation model described in the present invention;

[0030] Figure 2 This is a processing flow chart of the Prototype calculation module of the present invention;

[0031] Figure 3 The encoder-decoder module sampling flow chart of the present invention;

[0032] Figure 4 This is a processing flow chart of the task controller module of the present invention;

[0033] Figure 5 It is a processing flow chart of the dynamic filter generator module and the dynamic header segmentation module of the present invention;

[0034] Figure 6 This is a diagram of the segmentation experiment results of the model described in the present invention;

[0035] Figure 7 This is a comparison table of indicators between the segmentation experimental effect of the model described in the present invention and the effect of the existing segmentation model. DETAILED DESCRIPTION

[0036] The present invention is further described below in conjunction with the accompanying drawings. This embodiment takes abdominal CT three-dimensional image segmentation as an example to specifically describe the construction process of a segmentation model applicable to abdominal CT three-dimensional images, the model training process, and the image testing process using the model;

[0037] like Figure 1 As shown in FIG. 1 , the specific construction process of the segmentation model described in this embodiment is as follows:

[0038] Step S1, data set preprocessing and pretraining, specifically includes:

[0039] Step S1.1, according to the segmentation task of the three-dimensional medical image, information is encoded as task encoding information. The specific operation is: information encoding is performed on tasks of different categories. This embodiment takes the abdominal CT three-dimensional image segmentation task as an example. Therefore, there are 6 types of segmentation tasks of different organs and corresponding tumors, namely Liver, Kidney, Hepatic vessel, Pancreas, Colon, Lung. The task information is processed by one-hot encoding, as shown in the following table:

[0040] Serial number Dataset coding 1 Liver [1,0,0,0,0,0] 2 Kidney [0,1,0,0,0,0] 3 Hepatic vessel [0,0,1,0,0,0] 4 Pancreas [0,0,0,1,0,0] 5 Colon [0,0,0,0,1,0] 6 Lung [0,0,0,0,0,1]

[0041] Step S1.2, obtaining a plurality of categories of single annotated data sets consistent with the above segmentation task. In this embodiment, six single annotated data sets are collected, namely Liver, Kidney, Hepatic vessel, Pancreas, Colon, and Lung, and these data sets are divided into a training set and a test set at a ratio of 4:1. The above single annotated data sets are all from public data sets disclosed on the Internet, among which the Liver, Hepatic vessel, Pancreas, Colon, and Lung data sets are specifically the corresponding data sets described in the public document DOI: 10.48550 / arXiv.1902.0 9063, and the Kidney data set is specifically the Grand Challenge KiTS19 kidney tumor segmentation data set. Since the sources of the data sets are different, the three-dimensional medical images in the data sets are firstly cropped to the same size so that the resolution of each three-dimensional image is 64*192*192, and then the cropped images are converted into a nii.gz format suitable for network reading as the original images;

[0042] Step S2, calculate the Prototype feature vector corresponding to the single annotated data set of each category; as shown in Figure 2, the specific operation steps include: using the segmentation network model as a pre-trained model, inputting the original image for segmentation processing, taking the feature map of the last layer, namely the segmentation feature map, using the segmentation feature map as the feature map of the original image in the latent space, and then performing trilinear interpolation on the segmentation feature map to obtain an interpolation feature map, so that its size is consistent with the original image; at the same time, by using the image label information in the data set as the mask Mask of the organ and tumor; then, multiplying the mask Mask with the interpolation feature map, so as to obtain the Prototype information corresponding to each original image in the single annotated data set of this category, and finally accumulating the Prototype information corresponding to each original image in this category and averaging it to obtain the Prototype feature vector that highly represents the single annotated data set of this category;

[0043] Step S3, constructing a codec module, such as Figure 3 As shown in Figure 1, the encoder-decoder is a convolutional neural network with a U-shaped structure, where the encoder consists of multiple 3D residual blocks, each of which contains two convolutional layers with a kernel size of 3×3×3. Group normalization and ReLU activation functions are applied after each convolutional layer. Downsampling operations are performed through the encoder to extract original image features and high-level semantic information as downsampling information. In each downsampling step, a convolution with a step size of 2 is applied to halve the resolution of the input feature map. A total of four downsampling operations are performed in the encoder. When a given image X ij , the output can be expressed as:

[0044] F ij =f E (X ij θ E )

[0045] Among them, θ E represents the parameters of all encoders, f E represents the entire encoder network, F ij is the output of the encoder.

[0046] The decoder is symmetrical with the encoder. The decoder upsamples the feature map to improve its resolution, restores the spatial information of the original image by upsampling, and outputs a pre-segmented feature map of the same size as the input; the decoder receives the feature vectors of each corresponding level in the encoder as input to better restore the spatial information of the original image; in each upsampling process, the upsampled feature map is first added to the corresponding low-level feature map in the encoder, and then the semantic information is enhanced by the residual block; after the feature map is upsampled four times, the pre-segmented feature map M of the same size as the input can be obtained. ij :

[0047] M ij =f D (F ij θ D )

[0048] Among them, θ D Represents the parameters of all decoders, f D represents the entire decoder network, F ij is the output of the encoder, M ij It is the pre-segmented feature map output by the decoder.

[0049] Step S4, construct a task controller module, such as Figure 4 As shown in the figure, the task controller is a simple 3D convolution layer. The GAP feature vector obtained by global pooling of task encoding information and downsampling information and the Prototype feature vector corresponding to each category of data set are input into the 3D convolution layer for feature extraction. The output is a feature vector that can guide the network model to process a specific task, that is, a feature map containing task information. The specific operations are as follows:

[0050] First, the down-sampling information F in the encoder is ij Perform a pooling operation, then concatenate it with the unique hot encoding of the corresponding task, and then input it into the task controller to obtain the feature map T containing the task information ij :

[0051] T ij =f T (CAT(GAP,Ck ,P k );θ T )

[0052] Among them, θ T represents the parameters of all task encoders, GAP represents the output F of the encoder ij The feature vector obtained by pooling operation, C k represents the one-hot encoding information of the kth task, P k represents the prototype feature vector of the kth task, CAT represents the vector concatenation operation, and f T represents the entire task controller network, T ij Represents the feature graph containing task information output by the task controller module;

[0053] A dynamic segmentation head is used to take the convolution kernel parameters as parameters of a convolution operation, and a final image segmentation is performed on the pre-segmented image to obtain a segmentation result.

[0054] Step S5, construct a dynamic convolution kernel generator module, such as Figure 5 As shown in FIG. 1 , the dynamic convolution kernel generator is composed of a 3D convolution layer, and a dynamic convolution kernel is used to generate convolution kernel parameters from the feature map containing task information, and the convolution kernel parameters are used for the segmented convolution layer; specifically, the dynamic convolution kernel receives the feature map containing task information from the task controller, and generates a parameter w that can dynamically change with the task semantic information. ij :

[0055] w ij =f C (T ij θ C )

[0056] Among them, θ C represents all parameters in the dynamic convolution kernel generator, T ij represents the feature map containing task information output by the task controller module, f C Represents the entire dynamic convolution kernel generator network;

[0057] Step S6, construct a dynamic segmentation head module, such as Figure 5 As shown in Figure 1, the dynamic segmentation head contains three layers of 3D convolution. The number of channels in the first two layers is 8, and the last layer has two channels, one channel for organ segmentation and the other channel for tumor segmentation. The input of the dynamic convolution head comes from the decoder, which contains the semantic information of multiple organs and tumors. ij , unlike ordinary convolution, the convolution kernel in the dynamic convolution head is variable and comes from the dynamic convolution kernel generator, so the segmented image S can be obtained for a specific task. ij :

[0058] S ij =f S (M ij ;w ij )

[0059] Among them, w ij is the convolution kernel parameter from the dynamic convolution kernel generator, M ij is the pre-segmented image from the decoder, f S represents the entire dynamic convolutional head network, S ij Indicates the segmentation result.

[0060] Training process: Take the processed CT images in the training set as input, and get the output segmented image through the constructed neural network including encoder-decoder, task controller, dynamic convolution kernel generator and dynamic convolution head. Then calculate the loss with the label image, train the parameters through the gradient descent strategy, and finally get a network model with good segmentation results after hundreds of rounds of training. This example is calculated on the GeForce GTX2080 Ti processor using the Pytorch library in the Python environment. For network optimization and experimental process, we use cross entropy loss as the loss function in the experiment, and the optimizer uses the SGD optimizer with momentum of 0.99. In order to achieve better convergence, the learning rate is initially set to 0.01 and decays with continuous training.

[0061] Testing process: The CT images of the test set are used as input for testing, and the segmentation results of organs and tumors are obtained through network output. In order to verify the effectiveness of the present invention, multiple current mainstream methods are used to test on the same data set and compared with the experimental results of the present invention.

[0062] As a comparison method, there are multi-network models (Multi-Nets), that is, using multiple segmentation network models, each network model handles a specific task, multi-head network model Med3d (see document DOI:10.48550 / ar Xiv.1904.00625), PIPO-FAN (Pyramid Input and Pyramid Output Feature Abs traction Network, see document DOI:10.1109 / TMI.2020.3001036) and a single network Cond-Dec containing task information (see document: Dmitriev K, Kaufman A E. Learning mult i-class segmentations from single-class datasets [C] / / Proceedings of the IEEE / CV F Conference on Computer Vision and Pattern Recognition. 2019:9501-9511. DOI:10.1109 / CVPR.2019.00973).

[0063] In order to ensure fair comparison of the experiment, all models use the same encoder-decoder module. The Dice coefficient is used as a measure of the segmentation effect. Some experimental results are compared as follows: Figure 6 shown.

[0064] The above is a detailed description of the implementation mode of the present invention in combination with the accompanying drawings, but the present invention is not limited to the above-mentioned implementation mode. Any associative transformation or replacement within the knowledge scope of ordinary technicians in the field should be included in the scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A training method for a three-dimensional medical image segmentation model, characterized in that: The training method comprises: Information encoding is performed according to a segmentation task of a three-dimensional medical image as task encoding information; a single annotated data set of multiple categories consistent with the segmentation task is obtained, the single annotated data set is a collection of at least 50 three-dimensional medical images and single annotated information, i.e., labels, and all three-dimensional medical images contained in the single annotated data set are unified in pixels and format as original images; a Prototype feature vector corresponding to each category of the single annotated data set is calculated; a downsampling operation is performed through an encoder to extract original image features and high-level semantic information as downsampling information; a feature vector of each corresponding level in the encoder is received as input through a decoder, and then upsampled to restore the spatial information of the original image and output a pre-segmentation feature map consistent with the input size; After global pooling of the task encoding information and down-sampling information, the GAP feature vector and the Prototype feature vector corresponding to each category of data set are input into the 3D convolution layer for feature extraction, and the output is a feature vector that can guide the network model to process a specific task, that is, a feature map containing task information; A dynamic convolution kernel is used to generate convolution kernel parameters from the feature map containing task information, and the convolution kernel parameters are used for the convolution layer of segmentation; a dynamic segmentation head is used to use the convolution kernel parameters as parameters of the convolution operation, and the final image segmentation is performed on the pre-segmented image to obtain a segmentation result; The loss is calculated for the segmentation result and the label image, and the parameters are trained using a gradient descent strategy. Finally, the segmentation model is obtained after at least several hundred rounds of training.

2. The training method of a three-dimensional medical image segmentation model according to claim 1, characterized in that: The specific operation steps of calculating the Prototype feature vector corresponding to the single annotated data set of each category include: using the segmentation network model as a pre-trained model, inputting the original image for segmentation processing, taking the feature map of the last layer, namely the segmentation feature map, using the segmentation feature map as the feature map of the original image in the latent space, and then performing trilinear interpolation on the segmentation feature map to obtain an interpolation feature map, so that its size is consistent with the original image; at the same time, by using the image label as the mask Mask of the organ and tumor; then, multiplying the mask Mask with the interpolation feature map, so as to obtain the Prototype information corresponding to each original image in the single annotated data set of this category, and finally accumulating the Prototype information corresponding to each original image in this category and calculating the average value to obtain the Prototype feature vector that highly represents the single annotated data set of this category.

3. The training method of a three-dimensional medical image segmentation model according to claim 1, characterized in that: The loss function uses cross entropy loss, and the gradient descent strategy specifically uses stochastic gradient descent with a momentum of 0.

99. The initial learning rate is 0.01 and decays with continuous training.

4. A three-dimensional medical image segmentation model, characterized in that: The model is obtained by training the model training method described in any one of claims 1 to 3, and the model includes: a prototype calculation module, an encoder module, a decoder module, a task controller module, a dynamic convolution kernel generator module and a dynamic segmentation head module; the prototype calculation module is used to calculate the prototype feature vector corresponding to a single labeled data set of each category; the encoder module is used to extract original image features and high-level semantic information as downsampling information; the decoder is used to receive the feature vectors of each corresponding level in the encoder as input, upsample to restore the spatial information of the original image, and output a pre-segmented image of the same size as the input feature map; the task controller module uses a 3D convolution layer to extract features from the input task encoding information, the GAP feature vector obtained after global pooling of the down-sampling information, and the Prototype feature vector corresponding to each type of data set, and outputs a feature vector that can guide the network model to process a specific task, that is, a feature map containing task information; the dynamic convolution kernel generator module is used to receive the feature map containing task information and generate convolution kernel parameters in the convolution layer for segmentation; the dynamic segmentation head module is used to receive the convolution kernel parameters and use them as parameters of the convolution operation, and perform final image segmentation on the pre-segmented image output by the decoder to obtain a segmentation result.

5. A three-dimensional medical image segmentation model according to claim 4, characterized in that: The encoder module and the decoder module are U-shaped convolutional neural networks; wherein the encoder is composed of multiple 3D residual blocks, each residual block contains two convolutional layers with a convolution kernel size of 3×3×3, and each convolutional layer is followed by group normalization and ReLU activation function. In each downsampling step, a convolution with a step size of 2 is used to halve the resolution of the input feature map, and downsampling is performed four times in total in the encoder; when a given image X ij , the output can be expressed as: F ij =f E (X ij ;θ E ) Among them, θ E represents the parameters of all encoders, f E represents the entire encoder network, F ij is the output of the encoder.

6. A three-dimensional medical image segmentation model according to claim 5, characterized in that: The decoder is symmetrical with the encoder. The decoder upsamples the feature map to improve its resolution. In each upsampling process, the upsampled feature map is first added to the corresponding low-level feature map in the encoder, and then the semantic information is enhanced by the residual block. After upsampling the feature map four times, the pre-segmented feature map M with the same size as the input can be obtained. ij : M ij =f D (F ij ;θ D ) Among them, θ D Represents the parameters of all decoders, f D represents the entire decoder network, F ij is the output of the encoder, M ij is the output of the decoder.

7. A three-dimensional medical image segmentation model according to claim 6, characterized in that: The task controller module is a simple 3D convolutional layer that first performs a ij Perform pooling operation to obtain GAP feature vector, and then compare it with the unique hot encoding C of the corresponding task k And splicing, then input into the task controller to obtain the feature map T containing task information ij : T ij =f T (CAT(GAP,C k ,P k );θ T ) Among them, θ T represents the parameters of all task encoders, GAP represents the output F of the encoder ij The feature vector obtained by pooling operation, C k represents the one-hot encoding information of the kth task, P k represents the prototype feature vector of the kth task, CAT represents the vector concatenation operation, and f T represents the entire task controller network, T ij Represents the feature map containing task information output by the task controller module.

8. A three-dimensional medical image segmentation model according to claim 7, characterized in that: The dynamic convolution kernel generator module is composed of a 3D convolution layer, receives a feature map containing task information from the task controller, and generates a convolution kernel parameter w that changes dynamically with the feature map. ij : w ij =f C (T ij ;θ C ) Among them, θ C represents all parameters in the dynamic convolution kernel generator, T ij represents the feature map containing task information output by the task controller module, f C Represents the entire dynamic convolution kernel generator network.

9. A three-dimensional medical image segmentation model according to claim 8, characterized in that: The dynamic segmentation head module contains three layers of 3D convolution. The number of channels in the first two layers is 8, and the last layer has two channels, one channel for organ segmentation and the other channel for tumor segmentation. The input of the dynamic segmentation head comes from the decoder, which contains the semantic information of multiple organs and tumors. ij , unlike ordinary convolution, the convolution kernel in the dynamic segmentation head is variable and comes from the dynamic convolution kernel generator, so the segmented image S can be obtained for a specific task. ij : S ij =f S (M ij ;w ij ) Among them, w ij is the convolution kernel parameter from the dynamic convolution kernel generator, M ij is the pre-segmented feature map from the decoder, f S represents the entire dynamic convolutional head network, S ij Indicates the segmentation result.

10. Application of the segmentation model of claim 4 in multi-task medical image segmentation processing, characterized in that: The method uses the segmentation model to perform segmentation processing on the medical image, and outputs segmentation results of different organs and corresponding tumors.