A CT pancreatic tumor automatic segmentation method, system, terminal and storage medium

Abstract

The present application relates to a method, system, terminal and storage medium for automatic segmentation of CT pancreatic tumors. The method comprises: obtaining a three-dimensional CT image to be segmented; preprocessing the three-dimensional CT image; inputting the preprocessed three-dimensional CT image into a trained UMRFormer network model, wherein the UMRFormer network model generates feature maps of different scales, encodes the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the three-dimensional CT image. The present application proposes a novel UMRFormer network model that deeply fuses the global long-range correlation and spatial local information of features. By embedding a Transformer-like module that has the advantage of fusing local and global dependencies of features, the network model learns more feature information of small pancreatic tumor targets and global long-range dependency feature information between the tumor and surrounding tissues, thereby improving the overall segmentation accuracy of the network.

Description

Technical Field

[0001] The present application belongs to the field of medical image processing technology, and in particular relates to a CT pancreatic tumor automatic segmentation method, system, terminal and storage medium. Background Art

[0002] Pancreatic cancer is one of the most lethal malignancies of the digestive tract, characterized by insidious symptoms, delayed diagnosis, difficult treatment, and a high mortality rate. The five-year survival rate after diagnosis is approximately 9%, making it one of the malignant tumors with the worst prognosis. Currently, the best treatment for pancreatic tumors is tumor resection, which requires not only advanced medical equipment such as computed tomography, magnetic resonance imaging, single-photon emission computed tomography, and positron emission tomography to capture tumor location, size, and shape, but also requires diagnosis by experienced physicians. In clinical medicine, pancreatic tumors are primarily manually annotated by professional physicians, which consumes a significant amount of time and effort. Furthermore, manual annotation often involves subjectivity, and multiple delineations of the lesion area within the same image can result in different results. This makes consistency and reproducibility difficult to guarantee, leading to deviations in tumor delineation and potentially misdiagnosis or missed diagnosis. Therefore, improving the segmentation accuracy of pancreatic tumors through computer-assisted analysis and diagnosis is a promising approach.

[0003] Existing automated medical tumor segmentation methods primarily include deep learning-based segmentation methods and Transformer network-based segmentation methods. Deep learning-based segmentation methods are all based on the U-Net architecture, such as U-Net++ and Attention-UNet. Current Transformer network-based segmentation methods typically employ two approaches: constructing an encoder-decoder network using only Transformer modules, or using a single Transformer as an auxiliary module, or fusing convolutional modules with Transformers to construct a segmentation network. However, both approaches have significant drawbacks. Specifically, pancreatic cancer is closely surrounded by tissue of similar intensity, while existing models utilize Transformer modules to focus more on the global features of the tumor and surrounding tissue, ignoring the local features of pancreatic tumors as small targets. Furthermore, existing methods only consider the combination of convolution and Transformer models, ignoring the Transformer's inherent limitation of focusing on spatially local features. This results in a loss of segmentation accuracy when segmenting pancreatic tumors. Summary of the Invention

[0004] The present application provides a CT pancreatic tumor automatic segmentation method, system, terminal and storage medium, aiming to solve at least one of the above-mentioned technical problems in the prior art to a certain extent.

[0005] In order to solve the above problems, this application provides the following technical solutions:

[0006] A CT pancreatic tumor automatic segmentation method, comprising:

[0007] Acquire a three-dimensional CT image to be segmented;

[0008] Preprocessing the three-dimensional CT image;

[0009] The preprocessed three-dimensional CT image is input into a trained UMRFormer network model. The UMRFormer network model has a U-shaped symmetrical structure. The UMRFormer network model uses 3D CNN to generate feature maps of different scales, and uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the three-dimensional CT image.

[0010] The technical solution adopted in the embodiment of the present application further includes: the preprocessing of the three-dimensional CT image is specifically as follows:

[0011] The three-dimensional CT image is enhanced, and the enhanced three-dimensional CT image is cropped to a size that meets the input of the UMRFormer network model.

[0012] The technical solution adopted in the embodiment of the present application also includes: the UMRFormer network model includes a convolutional layer, a multi-head self-attention mechanism layer, a pooling layer, a normalization layer, and an upsampling layer. The UMRFormer network model has a U-shaped symmetric structure. In the skip connection stage of the U-shaped symmetric network, a two-layer Transformer-like module is embedded;

[0013] The UMRFormer network model uses a five-layer 3D CNN to generate feature maps of different scales to capture spatial and depth features. It then uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the fourth and fifth layer feature maps in the global space. Upsampling layers and convolutional layers are repeatedly superimposed to obtain the segmentation results of the 3D CT image.

[0014] The technical solution adopted in the embodiment of the present application also includes: the UMRFormer network model includes an encoder and a decoder, the encoder includes L MRFormer modules, each MRFormer module includes 1 multi-head self-attention module, 3 deep convolution layers and 4 post-normalization layers, the deep convolution layers and normalization layers are superimposed as a combination, and a residual unit is formed through jump connection. The feedforward network is replaced by a combination of 3 deep convolution layers plus a normalization layer, and the normalization layer in front of the multi-head self-attention module is post-placed, so that the normalization layer is moved from the beginning of each residual unit to the back end.

[0015] The technical solution adopted in the embodiment of the present application also includes: the training process of the UMRFormer network model includes:

[0016] In the encoder part: Linear projection is used to increase the channel size of the first two layers from K1=128, K2=256 to d1=512, d2=1024 respectively. The spatial and deep dimensions of the input 3D feature map are uniformly compressed to one dimension. The compressed feature map is added to the learnable positional encoding to create a feature embedding formula:

[0017] z1=f1+PE1=LP1×M1+PE1

[0018] z2=f2+PE2=LP2×M2+PE2

[0019] Among them, LP1 and LP2 are linear projection operations, and represents the position code, and Represents feature embedding; the output formula of the l-th layer (l∈[1, 2, ..., L]) Transformer-like module is:

[0020]

[0021]

[0022] Among them, LN (*) is layer normalization, z l is the output of the MRFormer module at layer l, DConv (*) is the depthwise convolution.

[0023] The technical solution adopted in the embodiment of the present application also includes: the training process of the UMRFormer network model also includes:

[0024] In the decoder part: the feature sequence output by the two-layer MRFormers module is mapped into a three-dimensional feature map, the three-dimensional feature map is upsampled and convolved, and fused with the encoder feature through long-distance jump connections to obtain semantic and spatial detail information.

[0025] Another technical solution adopted in the embodiment of the present application is: a CT pancreatic tumor automatic segmentation system, comprising:

[0026] Data acquisition module: used to obtain the three-dimensional CT image to be segmented;

[0027] Data preprocessing module: used for preprocessing the three-dimensional CT image;

[0028] Data segmentation module: used to input the preprocessed 3D CT image into the trained UMRFormer network model. The UMRFormer network model has a U-shaped symmetrical structure. The UMRFormer network model uses 3D CNN to generate feature maps of different scales, and uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the 3D CT image.

[0029] Another technical solution adopted by the embodiment of the present application is: a terminal, the terminal including a processor and a memory coupled to the processor, wherein:

[0030] The memory stores program instructions for implementing the CT pancreatic tumor automatic segmentation method;

[0031] The processor is configured to execute the program instructions stored in the memory to control automatic segmentation of CT pancreatic tumors.

[0032] Another technical solution adopted by the embodiment of the present application is: a storage medium storing program instructions executable by a processor, wherein the program instructions are used to execute the CT pancreatic tumor automatic segmentation method.

[0033] Compared with the prior art, the beneficial effects of the embodiments of the present application are as follows: the CT pancreatic tumor automatic segmentation method, system, terminal and storage medium of the embodiments of the present application propose a new UMRFormer network model that deeply integrates global long-distance correlation and spatial local information of features. The network model is based on a U-shaped codec structure. In the jump connection stage of the U-shaped structure, a Transformer-like module with the advantages of integrating local and global dependencies of focused features is embedded, which enables the network model to learn more feature information of small targets of pancreatic tumors and global long-dependency feature information between tumors and surrounding tissues, so as to obtain more effective deep feature context information, thereby improving the existing single Transformer-like network framework with certain defects and improving the overall segmentation accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flowchart of the CT pancreatic tumor automatic segmentation method of the first embodiment of the present application;

[0035] Figure 2 is a flow chart of the method for automatic segmentation of CT pancreatic tumors according to the second embodiment of the present application;

[0036] Figure 3 This is a diagram of the UMRFormer network model architecture of an embodiment of the present application;

[0037] Figure 4 This is a schematic diagram of the MRFormer module structure of an embodiment of the present application;

[0038] Figure 5 This is a schematic diagram of the structure of the CT pancreatic tumor automatic segmentation system according to an embodiment of the present application;

[0039] Figure 6 This is a schematic diagram of the terminal structure of an embodiment of the present application;

[0040] Figure 7 A schematic diagram of the structure of the storage medium of an embodiment of the present application. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0042] The terms "first," "second," and "third" in this application are used only for descriptive purposes and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of such features. In the description of this application, "multiple" means at least two, for example, two, three, etc., unless otherwise specifically defined. All directional indications in the embodiments of this application (such as up, down, left, right, front, back...) are only used to explain the relative positional relationship, movement, etc. between the components under a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications also change accordingly. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products, or devices.

[0043] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0044] See also Figure 1 , is a flow chart of the CT pancreatic tumor automatic segmentation method of the first embodiment of the present application. The CT pancreatic tumor automatic segmentation method of the embodiment of the present application comprises the following steps:

[0045] S100: Acquire a three-dimensional CT image to be segmented;

[0046] S101: Preprocessing the three-dimensional CT image;

[0047] In this step, the three-dimensional CT image is preprocessed specifically by performing enhancement processing on the three-dimensional CT image and cropping the enhanced three-dimensional CT image to a size that meets the input size of the UMRFormer network model.

[0048] S102: The preprocessed 3D CT image is input into the trained UMRFormer network model. The UMRFormer network model has a U-shaped symmetrical structure. The UMRFormer network model uses 3D CNN to generate feature maps of different scales, and uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the 3D CT image.

[0049] In this step, the UMRFormer network model includes convolutional layers, multi-head self-attention mechanism layers, pooling layers, normalization layers, and upsampling layers. The UMRFormer network model has a U-shaped symmetrical structure. In the skip connection stage of the U-shaped symmetrical network, a two-layer Transformer-like module is embedded. The UMRFormer network model uses a five-layer 3D CNN to generate feature maps of different scales to capture spatial and depth features. Then, a two-layer Transformer-like module is used to encode the long-range dependency semantic information of the fourth and fifth layer feature maps in the global space. The upsampling layers and convolution layers are repeatedly superimposed to obtain the segmentation results of the three-dimensional CT image.

[0050] See also Figure 2 , is a flow chart of the CT pancreatic tumor automatic segmentation method of the second embodiment of the present application. The CT pancreatic tumor automatic segmentation method of the second embodiment of the present application comprises the following steps:

[0051] S200: Acquire a CT pancreas dataset, and divide the CT pancreas dataset into a training set, a validation set, and a test set according to a set ratio;

[0052] In this step, the CT pancreas dataset used includes 281 CT cases from the MSD pancreas dataset. The CT pancreas dataset is divided into a ratio of 6:2:2, resulting in 187 training sets, 47 validation sets, and 47 test sets, respectively.

[0053] S201: Preprocessing the CT pancreas dataset and setting model training parameters;

[0054] In this step, data preprocessing specifically involves enhancing the CT pancreatic data and cropping the enhanced data to a size that meets the model input requirements. For example, if the original CT pancreatic data is 512x512x148, the cropped data is sized to 128x128x64 to meet the input requirements of model training. Model training parameters include, but are not limited to, epochs (number of iterations), batch size, and learning rate. Preferably, epochs is set to 1000 in this embodiment, and specific settings can be made based on the application scenario.

[0055] S202: Input the preprocessed training set into the UMRFormer network model for iterative training, and output the CT pancreatic tumor segmentation model after determining that the number of iterations reaches the set number;

[0056] In this step, the embodiment of the present application constructs a UMRFormer network model by deeply fusing the local information of the convolutional network and the global characteristics of the Transformer module, such as Figure 3 As shown, this is the UMRFormer network model architecture diagram of an embodiment of the present application. The model includes a convolutional layer, a multi-head self-attention mechanism (MHSA) layer, a pooling layer, a normalization layer, and an upsampling layer. The network model has a U-shaped symmetrical structure as a whole and is divided into two stages: encoding and decoding. In the jump connection stage of the U-shaped network, a double-layer Transformer-like module is embedded. Given an input three-dimensional CT image with a spatial resolution of HxW, a depth dimension of D (number of slices), and a number of channels of C (number of modalities), the UMRFormer network model first uses a 5-layer 3D CNN to generate feature maps of different scales to capture spatial and depth features, and then uses a double-layer Transformer-like module to encode the long-range dependency semantic information of the fourth and fifth layer feature maps in the global space, and repeatedly superimposes the upsampling layer and the convolution layer to gradually obtain a high-resolution segmentation result.

[0057] Furthermore, the encoder in the embodiment of the present application is composed of L MRFormer modules, specifically as follows Figure 4 As shown, it is a schematic diagram of the MRFormer module structure of an embodiment of the present application. Each MRFormer module includes a multi-head self-attention (MHSA) module, 3 deep convolutional layers and 4 post-normalization layers. Compared with the existing Transformer structure, the MRFormer module of the embodiment of the present application superimposes the deep convolutional layer and the normalization layer as a combination, forms a residual unit through jump connection, and uses a combination of 3 deep convolutional layers plus a normalization layer to replace the feedforward network (FFN), aiming to focus on the local receptive field in the feature sequence. In addition, the normalization layer in front of the multi-head self-attention module is post-positioned, that is, the normalization layer is moved from the beginning of each residual unit to the back end, so that a milder activation value can be generated on the network layer, making the model training more stable.

[0058] Based on the above network structure, the UMRFormer network model training process of the embodiment of the present application is specifically as follows:

[0059] In the encoder part: Linear projection (3x3x3 convolutional layer) is used to increase the channel size of the first two layers from K1=128, K2=256 to d1=512, d2=1024. Since the two-layer MRFormers module requires a sequence as input, the spatial dimension and deep dimension of the input 3D feature map are compressed to one dimension, resulting in a and The feature maps f1 and f2 are Q1d1-dimensional labels and Q2d2-dimensional labels respectively. After compressing the 3D feature map to one dimension, it is added to the learnable positional encoding to create a feature embedding formula:

[0060] z1=f1+PE1=LP1×M1+PE1 (1)

[0061] z2=f2+PE2=LP2×M2+PE2 (2)

[0062] Among them, LP1 and LP2 are linear projection operations, and represents the position code, and Represents feature embedding. The output formula of the l-th layer (l∈[1, 2, ..., L]) Transformer-like module can be expressed as:

[0063]

[0064]

[0065] Among them, LN (*) is layer normalization, z l is the output of the MRFormer module at layer l, DConv (*) is the depthwise convolution.

[0066] In the decoder part: In order to adapt to the input size of the decoder, the feature sequence output by the double-layer MRFormers module needs to be mapped into a three-dimensional feature map. The double-layer MRFormers module (i.e.: and ) are reshaped into and After feature mapping, the outputs of Z1′ and Z2′ are upsampled and convolved respectively, and then fused with the encoder features through long-distance skip connections to obtain finer semantic and fine-grained information as well as richer spatial details.

[0067] S203: Inputting the validation set into the CT pancreatic tumor segmentation model, verifying the preliminary segmentation results of the CT pancreas / pancreatic tumor by the CT pancreatic tumor segmentation model, and adjusting the model hyperparameters;

[0068] S204: Input the test set into the CT pancreatic tumor segmentation model to test the CT pancreas / pancreatic tumor segmentation performance of the model.

[0069] Based on the above, the CT pancreatic tumor automatic segmentation method of the embodiment of the present application proposes a new UMRFormer network model that deeply integrates the global long-distance correlation and spatial local information of features. The network model is based on a U-shaped codec structure. In the jump connection stage of the U-shaped structure, a Transformer-like module with the advantages of integrating the local and global dependencies of the focused features is embedded. This enables the network model to learn more feature information of small targets of pancreatic tumors and global long-dependency feature information between the tumor and surrounding tissues, so as to obtain more effective deep feature context information, thereby improving the existing single Transformer-like network framework with certain defects and improving the overall segmentation accuracy of the network.

[0070] See also Figure 5 , is a schematic diagram of the structure of the CT pancreatic tumor automatic segmentation system according to an embodiment of the present application. The CT pancreatic tumor automatic segmentation system 40 according to an embodiment of the present application comprises:

[0071] Data acquisition module 41: used to acquire the three-dimensional CT image to be segmented;

[0072] Data preprocessing module 42: used to preprocess the 3D CT image; wherein, the preprocessing of the 3D CT image specifically includes: enhancing the 3D CT image and cropping the enhanced 3D CT image to a size that meets the input of the UMRFormer network model.

[0073] Data segmentation module 43: used to input the preprocessed three-dimensional CT image into the trained UMRFormer network model. The UMRFormer network model has a U-shaped symmetrical structure. The UMRFormer network model uses 3D CNN to generate feature maps of different scales, and uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the three-dimensional CT image; wherein, the UMRFormer network model includes a convolutional layer, a multi-head self-attention mechanism layer, a pooling layer, a normalization layer and an upsampling layer. The UMRFormer network model has a U-shaped symmetrical structure. In the jump connection stage of the U-shaped symmetrical network, a two-layer Transformer-like module is embedded; the UMRFormer network model uses a 5-layer 3D CNN to generate feature maps of different scales to capture spatial and depth features, and then uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the fourth and fifth layer feature maps in the global space, and repeatedly superimposes the upsampling layer and the convolution layer to obtain the segmentation result of the three-dimensional CT image.

[0074] Based on the above, the CT pancreatic tumor automatic segmentation system of the embodiment of the present application proposes a new UMRFormer network model that deeply integrates the global long-distance correlation and spatial local information of features. The network model is based on a U-shaped codec structure. In the jump connection stage of the U-shaped structure, a Transformer-like module with the advantages of integrating local and global dependencies of attention features is embedded. This enables the network model to learn more feature information of small targets of pancreatic tumors and global long-dependency feature information between tumors and surrounding tissues to obtain more effective deep feature context information, thereby improving the existing single Transformer-like network framework with certain defects and improving the overall segmentation accuracy of the network.

[0075] See also Figure 6 , is a schematic diagram of the terminal structure of an embodiment of the present application. The terminal 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0076] The memory 52 stores program instructions for implementing the above-mentioned CT pancreatic tumor automatic segmentation method.

[0077] The processor 51 is configured to execute program instructions stored in the memory 52 to control automatic CT pancreatic tumor segmentation.

[0078] The processor 51 may also be referred to as a CPU (Central Processing Unit). The processor 51 may be an integrated circuit chip having signal processing capabilities. The processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The general-purpose processor may be a microprocessor or any conventional processor.

[0079] See also Figure 7 , which is a structural diagram of the storage medium of the embodiment of the present application. The storage medium of the embodiment of the present application stores a program file 61 that can implement all the above methods, wherein the program file 61 can be stored in the above storage medium in the form of a software product, including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the various implementation methods of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0080] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to encompass the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A CT pancreatic tumor automatic segmentation method, characterized in that: include: Acquire a three-dimensional CT image to be segmented; Preprocessing the three-dimensional CT image; Inputting the preprocessed 3D CT image into a trained UMRFormer network model, the UMRFormer network model having a U-shaped symmetrical structure, using 3D CNN to generate feature maps of different scales, and using a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputting the tumor segmentation result of the 3D CT image; The UMRFormer network model includes an encoder and a decoder. The encoder includes L MRFormer modules. Each MRFormer module includes a multi-head self-attention module, three deep convolutional layers and four post-normalization layers. The deep convolutional layers and normalization layers are superimposed as a combination, and a residual unit is formed through jump connection. The feedforward network is replaced by a combination of three deep convolutional layers plus a normalization layer, and the normalization layer in front of the multi-head self-attention module is post-positioned, so that the normalization layer moves from the beginning to the back end of each residual unit.

2. The automatic segmentation method for CT pancreatic tumors according to claim 1, characterized in that: The preprocessing of the three-dimensional CT image is specifically as follows: The three-dimensional CT image is enhanced, and the enhanced three-dimensional CT image is cropped to a size that meets the input of the UMRFormer network model.

3. The CT pancreatic tumor automatic segmentation method according to claim 1 or 2, characterized in that: The UMRFormer network model includes a convolutional layer, a multi-head self-attention mechanism layer, a pooling layer, a normalization layer, and an upsampling layer. The UMRFormer network model has a U-shaped symmetrical structure. In the skip connection stage of the UMRFormer network model, a two-layer Transformer-like module is embedded. The UMRFormer network model uses a five-layer 3D CNN to generate feature maps of different scales to capture spatial and depth features. It then uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the fourth and fifth layer feature maps in the global space. Upsampling layers and convolutional layers are repeatedly superimposed to obtain the segmentation results of the 3D CT image.

4. The automatic segmentation method for CT pancreatic tumors according to claim 3, characterized in that: The training process of the UMRFormer network model includes: In the encoder part: Linear projection is used to increase the channel size of the first two layers from K1=128, K2=256 to d1=512, d2=1024 respectively. The spatial and deep dimensions of the input 3D feature map are uniformly compressed to one dimension. The compressed feature map is added to the learnable positional encoding to create a feature embedding formula: z1=f1+PE1=LP1×M1+PE1 z2=f2+PE2=LP2×M2+PE2 Among them, LP1 and LP2 are linear projection operations, and represents the position code, and Represents feature embedding; the output formula of the l-th layer (l∈[1,2,..,L]) Transformer-like module is: z′ l =LN(MHA(z′ l-1 ))+z e-1 With l =LN3(DConv3(LN2(DConv2(LN1(DConv1(z l ))))))+z l Among them, LN (*) is layer normalization, z l is the output of the MRFormer module at layer l, DConv (*) is the depthwise convolution.

5. The automatic segmentation method for CT pancreatic tumors according to claim 4, characterized in that: The training process of the UMRFormer network model also includes: In the decoder part: the feature sequence output by the double-layer MRFormer module is mapped into a three-dimensional feature map, the three-dimensional feature map is upsampled and convolved, and fused with the encoder feature through a long-distance jump connection to obtain semantic and spatial detail information.

6. A CT pancreatic tumor automatic segmentation system using the CT pancreatic tumor automatic segmentation method according to claim 1, characterized in that: include: Data acquisition module: used to obtain the three-dimensional CT image to be segmented; Data preprocessing module: used for preprocessing the three-dimensional CT image; Data segmentation module: used to input the preprocessed 3D CT image into the trained UMRFormer network model. The UMRFormer network model has a U-shaped symmetrical structure. The UMRFormer network model uses 3D CNN to generate feature maps of different scales, and uses a two-layer Transformer-like module to encode the long-range dependency semantic information of the feature maps, and outputs the tumor segmentation result of the 3D CT image.

7. A terminal, characterized in that: The terminal includes a processor and a memory coupled to the processor, wherein: The memory stores program instructions for implementing the CT pancreatic tumor automatic segmentation method according to any one of claims 1 to 5; The processor is configured to execute the program instructions stored in the memory to control automatic segmentation of CT pancreatic tumors.

8. A storage medium, characterized in that: Program instructions executable by a processor are stored, and the program instructions are used to execute the CT pancreatic tumor automatic segmentation method according to any one of claims 1 to 5.

Images