Pancreatic tumor segmentation method, device and program product based on CT (Computed Tomography) image
By preprocessing and feature extraction of three-dimensional CT images of pancreatic tumor patients, combining visual cues and SAM2 model fusion, the continuous section memory mechanism is used to solve the problem of insufficient segmentation accuracy between pancreatic tumors and peripheral blood vessels, achieving higher segmentation accuracy and detail processing capabilities.
Patent Information
- Application Number
- CN202510091714.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-05-16
AI Technical Summary
The accuracy of automatic segmentation of pancreatic tumors and peripheral blood vessels is insufficient, and existing methods have limitations in processing small sample data sets and capturing fine anatomical features.
Using a segmentation method based on CT images, preprocessing three-dimensional CT images, using 3D nnU-Net segmentation network for feature extraction, and generating visual cues, combining SAM2 models for fusion, and using continuous slice memory mechanism to model the final segmentation result.
It significantly improves the accuracy and detailed processing ability of pancreatic tumor segmentation, enhances the ability to identify and segment complex structures, and provides important support for clinical diagnosis and treatment.
Smart Images
Figure CN120014271A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a pancreatic tumor segmentation method, device and program product based on CT images. Background Art
[0002] Pancreatic cancer is an extremely aggressive malignant tumor with a 5-year survival rate of less than 10%, and with the increasing incidence and mortality worldwide. Surgical resection is currently the only effective treatment that can achieve long-term survival for patients, and accurate preoperative evaluation is crucial to determine the surgical plan. Enhanced computed tomography (CT) is the preferred preoperative evaluation tool, which can provide important information about the location, size, and relationship of the tumor with surrounding key vascular structures through dual-phase scanning. However, due to individual differences in pancreatic morphology, tumor invasive growth patterns, and challenges in imaging contrast and resolution, automatic segmentation of pancreatic tumors and their surrounding blood vessels is a complex and difficult task. In recent years, the development of deep learning technology, especially convolutional neural networks (CNNs), has brought new opportunities to solve this problem. For example, U-Net and its variant nnU-Net have performed well in a variety of medical image segmentation tasks. Nevertheless, for targets such as pancreatic tumors with high heterogeneity and anatomical complexity, existing methods still have limitations, especially in processing small sample data sets and capturing fine anatomical features. Summary of the invention
[0003] The embodiments of the present invention provide a pancreatic tumor segmentation method, device and program product based on CT images to solve the problem of insufficient accuracy in automatic segmentation of pancreatic tumors and surrounding blood vessels, and significantly improve the accuracy of segmentation and detail processing capabilities.
[0004] In order to achieve the above object, on the one hand, a pancreatic tumor segmentation method based on CT images is provided, the method comprising:
[0005] S1, obtaining a three-dimensional CT image of a pancreatic tumor patient, and preprocessing the three-dimensional CT image;
[0006] S2, slicing the preprocessed three-dimensional CT image to obtain a plurality of first two-dimensional slices; and inputting the preprocessed three-dimensional CT image into a pre-constructed 3D nnU-Net segmentation network for feature extraction to obtain a three-dimensional segmentation result; the three-dimensional segmentation result includes a mask of the pancreatic tumor;
[0007] S3, determining a minimum circumscribed rectangle of the mask of the pancreatic tumor to generate a visual cue of the pancreatic tumor, and then slicing the three-dimensional segmentation result to obtain a plurality of second two-dimensional slices; wherein the visual cue is applied to each second two-dimensional slice, and the first two-dimensional slice corresponds to each two-dimensional slice in the second two-dimensional slices;
[0008] S4, inputting the first two-dimensional slice and the second two-dimensional slice into a pre-built SAM2 model for fusion to obtain a final segmentation result; wherein,
[0009] S41, for the first two-dimensional slices between the first appearance and disappearance of each category label in the first two-dimensional slices, select a first two-dimensional slice corresponding to each category label to obtain a third two-dimensional slice; wherein the category labels include: pancreatic tumor, vein and / or artery;
[0010] S42, encoding the third two-dimensional slice through an image encoder to generate a first embedding vector;
[0011] S43, encoding the visual cue of the second two-dimensional slice through a cue encoder to generate a second embedding vector;
[0012] S44: The first embedding vector and the second embedding vector are modeled through a continuous slice memory mechanism and then decoded to obtain a final segmentation result.
[0013] Preferably, in the pancreatic tumor segmentation method based on CT images, in step S1, preprocessing the three-dimensional CT image includes one or more of the following:
[0014] Describing a target structure on a three-dimensional CT image of a pancreatic tumor patient and annotating a category label of the target structure;
[0015] Pancreatic tumors with CT slice thickness less than or equal to the predetermined size were selected by initial upper abdominal contrast-enhanced CT;
[0016] Exclude 3D CT images with blurred CT images or CT image artifacts;
[0017] The three-dimensional CT image is adjusted to a pixel matrix of a predetermined size.
[0018] Preferably, in the pancreatic tumor segmentation method based on CT images, in step S2, the 3D nnU-Net segmentation network includes: a first encoder and a first decoder, and the preprocessed three-dimensional CT image is input into the pre-constructed 3D nnU-Net segmentation network for feature extraction, including:
[0019] The first encoder extracts semantic features of a pancreatic tumor and blood vessels surrounding the pancreatic tumor, and the first decoder enhances positioning through skip connections to cover the pancreatic tumor and blood vessels surrounding the pancreatic tumor.
[0020] Preferably, in the pancreatic tumor segmentation method based on CT images, in step S3, determining the minimum bounding rectangle of the mask of the pancreatic tumor to generate a visual cue of the pancreatic tumor includes:
[0021] The four vertices of the minimum bounding rectangle of the mask are used as negative sample points, and the center point of the minimum bounding rectangle is used as the positive sample point.
[0022] Preferably, the pancreatic tumor segmentation method based on CT images, wherein step S3, further comprises: identifying and increasing negative sample points and / or positive sample points of the pancreatic tumor and its surrounding area by an image analysis method, wherein:
[0023] Selecting a plurality of positive sample points in the pancreatic tumor, wherein the positive sample points are evenly distributed in the pancreatic tumor;
[0024] Selecting a plurality of negative sample points outside the pancreatic tumor, wherein the negative sample points are located outside the pancreatic tumor;
[0025] An enlarged frame is defined for the pancreatic tumor, and the positive sample points or the negative sample points are selected within the enlarged frame, wherein the enlarged frame is larger than a minimum circumscribed rectangle.
[0026] Preferably, in the pancreatic tumor segmentation method based on CT images, in step S44, the continuous slice memory mechanism includes: capturing the spatial correlation between the embedded vectors corresponding to adjacent two-dimensional slices through multi-layer self-attention and cross-attention modules, and integrating the contextual information of adjacent two-dimensional slices.
[0027] Preferably, the pancreatic tumor segmentation method based on CT images, wherein in step S4, the first two-dimensional slice and the second two-dimensional slice are input into a pre-built SAM2 model for fusion, further comprising:
[0028] The pre-built SAM2 model generates a plurality of candidate masks for the pancreatic tumor based on the second embedding vector, and calculates a confidence score for each candidate mask;
[0029] The candidate mask with the highest confidence score is selected from all candidate masks of the pancreatic tumor as the final mask of the pancreatic tumor.
[0030] On the other hand, an embodiment of the present invention provides a pancreatic tumor segmentation device based on CT images, which includes a memory and a processor, the memory stores at least one program, and the at least one program is executed by the processor to implement any of the pancreatic tumor segmentation methods based on CT images as described above.
[0031] On the other hand, an embodiment of the present invention provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements any of the above-mentioned pancreatic tumor segmentation methods based on CT images.
[0032] The above technical solution has the following technical effects:
[0033] The embodiments of the present invention ensure the quality of image data by preprocessing three-dimensional CT images of pancreatic tumor patients; use the 3D nnU-Net segmentation network to extract features from the preprocessed three-dimensional CT images to generate a three-dimensional segmentation result containing a mask of the pancreatic tumor; generate visual cues by determining the minimum circumscribed rectangle of the mask of the pancreatic tumor, and apply these cues to the second two-dimensional slices obtained by re-slicing, thereby ensuring the continuity and consistency of visual information; input the first two-dimensional slice and the second two-dimensional slice of the original CT image into the SAM2 model for fusion, and use the powerful segmentation capability of the SAM2 model combined with the visual cues extracted in the previous steps to obtain the final segmentation result, which not only improves the accuracy of segmentation, but also enhances the recognition and segmentation capabilities of complex structures, providing important support for clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 A schematic diagram of a flow chart of a pancreatic tumor segmentation method based on CT images according to an embodiment of the present invention;
[0035] Figure 2 A flow chart of a pancreatic tumor segmentation method based on CT images according to another embodiment of the present invention;
[0036] Figure 3 FIG. 4 is a schematic diagram of the structure of a pancreatic tumor segmentation device based on CT images according to an embodiment of the present invention. DETAILED DESCRIPTION
[0037] To further illustrate the various embodiments, the present invention provides drawings. These drawings are part of the disclosure of the present invention, which are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, a person of ordinary skill in the art should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0038] The present invention will now be further described with reference to the accompanying drawings and specific implementation methods.
[0039] Embodiment 1:
[0040] In order to solve the problem of insufficient precision in automatic segmentation of pancreatic tumors and surrounding blood vessels, thereby improving segmentation accuracy and detail processing capability, an embodiment of the present invention provides a pancreatic tumor segmentation method based on CT images. Figure 1 FIG. 1 is a flow chart of a pancreatic tumor segmentation method based on CT images according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0041] S1, obtaining three-dimensional CT images of pancreatic tumor patients and preprocessing the three-dimensional CT images;
[0042] S2, slicing the preprocessed three-dimensional CT image to obtain a plurality of first two-dimensional slices; and inputting the preprocessed three-dimensional CT image into a pre-constructed 3D nnU-Net segmentation network for feature extraction to obtain a three-dimensional segmentation result, wherein the three-dimensional segmentation result includes a mask of the pancreatic tumor;
[0043] S3, determining a minimum circumscribed rectangle of the mask of the pancreatic tumor to generate a visual cue of the pancreatic tumor, and then slicing the three-dimensional segmentation result to obtain a plurality of second two-dimensional slices; wherein the visual cue is applied to each second two-dimensional slice, and the first two-dimensional slice corresponds to each two-dimensional slice in the second two-dimensional slices;
[0044] S4, inputting the first two-dimensional slice and the second two-dimensional slice into the pre-built SAM2 model for fusion to obtain the final segmentation result; wherein,
[0045] S41, for the first two-dimensional slices between the first appearance and disappearance of each category label in the first two-dimensional slice, select a first two-dimensional slice corresponding to each category label to obtain a third two-dimensional slice; wherein the category labels include: pancreatic tumor, vein and / or artery;
[0046] S42, encoding the third two-dimensional slice through an image encoder to generate a first embedding vector;
[0047] S43, encoding the visual cue of the second two-dimensional slice through a cue encoder to generate a second embedding vector;
[0048] S44, the first embedding vector and the second embedding vector are modeled through a continuous slice memory mechanism and then decoded to obtain a final segmentation result.
[0049] Embodiment 2:
[0050] Figure 2The figure is a flow chart of a pancreatic tumor segmentation method based on CT images according to another embodiment of the present invention.
[0051] like Figure 2 As shown, the method includes:
[0052] 1. Obtain three-dimensional CT images of pancreatic tumor patients and pre-process the three-dimensional CT images;
[0053] Preferably, preprocessing the three-dimensional CT image includes:
[0054] The target structure is depicted on the three-dimensional CT image of the pancreatic tumor patient, and the category label of the target structure is annotated; preferably, the category label includes: pancreatic tumor, vein and / or artery.
[0055] Pancreatic tumors with CT slice thickness less than or equal to a predetermined size are selected through initial upper abdominal enhanced CT; preferably, the predetermined size is 5 mm.
[0056] Patients who had received previous treatment for pancreatic tumors and those with blurred CT images or 3D CT images with significant artifacts were excluded; and / or,
[0057] In order to enhance the reflection of clinical heterogeneity and improve the generalizability of the segmentation model, the three-dimensional CT image is adjusted to a pixel matrix of a predetermined size. Preferably, the predetermined size is 512×512.
[0058] 2. Slicing the preprocessed three-dimensional CT image to obtain a plurality of first two-dimensional slices; and inputting the preprocessed three-dimensional CT image into a pre-constructed 3D nnU-Net segmentation network for feature extraction to obtain a three-dimensional segmentation result; preferably, the three-dimensional segmentation result includes a mask of the pancreatic tumor;
[0059] Preferably, the 3D nnU-Net segmentation network is trained with 5-fold cross validation and experiments are performed on an Nvidia RTX 4090 GPU.
[0060] In a specific embodiment, a small amount of in-domain knowledge is learned to perform an initial rough segmentation of the anatomical structure of the pancreas, and the rough areas of pancreatic tumors and blood vessels are quickly identified. The 3D nnU-Net segmentation network includes: a first encoder and a first decoder, and the pre-processed three-dimensional CT image is input into a pre-built 3DnnU-Net segmentation network for feature extraction, including:
[0061] (1) Coding process;
[0062] The first encoder encodes the input 3D CT image to extract feature information, including the texture, shape, edge, etc. of the image, which is an important basis for subsequent segmentation. In this process, the neural network gradually transforms the original 3D CT image pixel information into a more abstract and semantically meaningful feature representation. The encoding process usually involves a multi-layer convolutional neural network (CNN) to gradually extract features at different levels.
[0063] Specifically, the first encoder extracts semantic features of pancreatic tumors and blood vessels surrounding pancreatic tumors.
[0064] (2) decoding process;
[0065] The decoding process is the opposite of the encoding process, and its purpose is to gradually restore the feature information extracted in the encoding process to a segmentation result of the same size as the original 3D CT image.
[0066] Preferably, the first decoder enhances positioning through skip connections to cover the pancreatic tumor and blood vessels around the pancreatic tumor. Specifically, through skip connections, feature maps of the same resolution in the encoding process are fused with corresponding feature maps in the decoding process, thereby reintroducing some detail information lost in the encoding process into the decoding process, which helps to better restore the local structure and boundary information of the image.
[0067] (3) Local fusion segmentation results;
[0068] During the decoding process, the segmentation results at different levels are locally fused to generate a preliminary segmentation result, which ensures the accuracy of the segmentation result and can capture local details.
[0069] (4) Cutting out local areas;
[0070] Based on the preliminary segmentation results, the local area of interest is cut out to obtain the final 3D segmentation result. The rough segmentation result roughly outlines the location and range of the pancreatic tumor and related blood vessels, but is not accurate enough in the boundary details and segmentation of small blood vessels.
[0071] 3. Generate prompts for the 3D segmentation results, and then perform slicing to obtain a rough segmentation result at the first stage;
[0072] Specifically, a minimum circumscribed rectangle of a mask of a pancreatic tumor is determined to generate a visual cue of the pancreatic tumor, and then the three-dimensional segmentation result is sliced to obtain a plurality of second two-dimensional slices, i.e., a first-stage rough segmentation result; wherein the visual cue is applied to each second two-dimensional slice, and the first two-dimensional slice corresponds to each two-dimensional slice in the second two-dimensional slices;
[0073] Preferably, the minimum bounding rectangle of the mask of the pancreatic tumor is determined to generate a visual cue of the pancreatic tumor, including: using the four vertices of the minimum bounding rectangle of the mask as negative sample points (background cue) and the center point of the minimum bounding rectangle as a positive sample point (foreground cue). These cue information will be used to guide the model in the fine segmentation stage to help it more accurately locate and segment the target area.
[0074] In a specific embodiment, identifying and increasing negative sample points and / or positive sample points of a pancreatic tumor and its surrounding area by an image analysis method includes:
[0075] Selecting multiple positive sample points in the pancreatic tumor, wherein the positive sample points are evenly distributed in the pancreatic tumor;
[0076] Selecting multiple negative sample points outside the pancreatic tumor, wherein the negative sample points are located outside the pancreatic tumor; and / or,
[0077] A magnification box is defined for the pancreatic tumor, and positive sample points or negative sample points are selected within the magnification box, wherein the magnification box is larger than the minimum circumscribed rectangle.
[0078] 4. The first two-dimensional slice and the second two-dimensional slice are input into the pre-built SAM2 model for fusion to obtain the final segmentation result, i.e., the second-stage refined segmentation result;
[0079] In a specific embodiment, through massive out-of-domain knowledge learning, fine segmentation is performed on the rough segmentation result of the first stage to improve the segmentation accuracy. The first two-dimensional slice and the second two-dimensional slice are input into a pre-built SAM2 model for fusion, including:
[0080] (1) for the first two-dimensional slices between the first appearance and disappearance of each category label in the first two-dimensional slices, selecting a first two-dimensional slice corresponding to each category label to obtain a third two-dimensional slice; wherein the category labels include: pancreatic tumor, vein and / or artery;
[0081] (2) encoding the third two-dimensional slice by using an image encoder to generate a first embedding vector; in a specific implementation, the image encoder is an image encoder provided by the SAM2 model encoder;
[0082] (3) encoding the visual cue of the second two-dimensional slice by a cue encoder to generate a second embedding vector; in a specific implementation, the cue encoder is a cue encoder provided by the SAM2 model encoder;
[0083] (4) The first embedding vector and the second embedding vector are modeled through a continuous slice memory mechanism and then decoded to obtain the final segmentation result.
[0084] Preferably, the continuous slice memory mechanism includes: capturing the spatial correlation between the embedding vectors corresponding to adjacent two-dimensional slices through multi-layer self-attention and cross-attention modules, and integrating the contextual information of adjacent two-dimensional slices.
[0085] In a specific embodiment, the first two-dimensional slice and the second two-dimensional slice are input into a pre-built SAM2 model for fusion, including:
[0086] The pre-built SAM2 model generates multiple candidate masks for pancreatic tumors based on the second embedding vector and calculates the confidence score of each candidate mask;
[0087] The candidate mask with the highest confidence score among all candidate masks of the pancreatic tumor is selected as the final mask of the pancreatic tumor.
[0088] In summary, the first stage effectively reduces the computational pressure of subsequent fine segmentation by quickly locating the approximate area of the tumor and the main vascular structure, laying the foundation for the efficient completion of the entire segmentation task. This stage relies on the automated configuration and standardized training process of the 3D-nnUNet segmentation network to achieve rapid and accurate coverage of pancreatic tumors and vascular anatomical structures, ensuring that the main target areas are fully presented in the initial segmentation. However, due to the limited ability of the decoder to restore boundary details, this stage still has certain limitations when dealing with complex tumor boundaries or small-volume blood vessels.
[0089] To make up for the shortcomings of the first stage, the second stage introduced the SAM2 model based on natural scene pre-training, and further optimized the segmentation results through out-of-domain knowledge transfer, especially in boundary detail processing and small target segmentation. The SAM2 model combines the continuous slice memory mechanism and uses the contextual information of adjacent slices to effectively enhance the accuracy and completeness of segmentation. This mechanism gives full play to the spatial correlation advantages of 3D medical images and significantly improves the segmentation accuracy of the tumor and its surrounding vascular interaction area. Especially when dealing with fuzzy boundaries and small volume targets, the model can flexibly adjust the segmentation strategy and accurately identify complex anatomical structures. In addition, the cross-domain knowledge of the natural scene pre-training model is used to overcome the problem of insufficient feature learning caused by small sample data, providing a new solution for the intelligent analysis of medical images.
[0090] Embodiment three:
[0091] The present invention also provides a pancreatic tumor segmentation device based on CT images, such as Figure 3As shown, the device includes a processor 301, a memory 302, a bus 303, and a computer program stored in the memory 302 and executable on the processor 301. The processor 301 includes one or more processing cores. The memory 302 is connected to the processor 301 via the bus 303. The memory 302 is used to store program instructions. When the processor 301 executes the computer program, the steps in the above method embodiment of the first embodiment of the present invention are implemented.
[0092] Further, as an executable solution, the pancreatic tumor segmentation device based on CT images can be a computer unit, which can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The computer unit may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the composition structure of the above-mentioned computer unit is only an example of a computer unit and does not constitute a limitation on the computer unit. It may include more or less components than the above, or a combination of certain components, or different components. For example, the computer unit may also include input and output devices, network access devices, buses, etc., which are not limited in the embodiments of the present invention.
[0093] Further, as an executable solution, the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, and various interfaces and lines are used to connect the various parts of the entire computer unit.
[0094] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the computer unit by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.
[0095] Embodiment 4:
[0096] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the steps of the method described above are implemented.
[0097] Although the present invention has been specifically shown and described in conjunction with the preferred embodiments, it should be understood by those skilled in the art that various changes may be made to the present invention in form and details without departing from the spirit and scope of the present invention as defined by the appended claims, all of which are within the scope of protection of the present invention.
Claims
1. A pancreatic tumor segmentation method based on CT images, characterized in that: include: S1, obtaining a three-dimensional CT image of a pancreatic tumor patient, and preprocessing the three-dimensional CT image; S2, slicing the preprocessed three-dimensional CT image to obtain a plurality of first two-dimensional slices; and inputting the preprocessed three-dimensional CT image into a pre-constructed 3D nnU-Net segmentation network for feature extraction to obtain a three-dimensional segmentation result; the three-dimensional segmentation result includes a mask of the pancreatic tumor; S3, determining a minimum circumscribed rectangle of the mask of the pancreatic tumor to generate a visual cue of the pancreatic tumor, and then slicing the three-dimensional segmentation result to obtain a plurality of second two-dimensional slices; wherein the visual cue is applied to each second two-dimensional slice, and the first two-dimensional slice corresponds to each two-dimensional slice in the second two-dimensional slices; S4, inputting the first two-dimensional slice and the second two-dimensional slice into a pre-built SAM2 model for fusion to obtain a final segmentation result; wherein, S41, for the first two-dimensional slices between the first appearance and disappearance of each category label in the first two-dimensional slices, select a first two-dimensional slice corresponding to each category label to obtain a third two-dimensional slice; wherein the category labels include: pancreatic tumor, vein and / or artery; S42, encoding the third two-dimensional slice through an image encoder to generate a first embedding vector; S43, encoding the visual cue of the second two-dimensional slice through a cue encoder to generate a second embedding vector; S44: The first embedding vector and the second embedding vector are modeled through a continuous slice memory mechanism and then decoded to obtain a final segmentation result.
2. The pancreatic tumor segmentation method based on CT images according to claim 1, characterized in that: In step S1, preprocessing the three-dimensional CT image includes one or more of the following: Describing a target structure on a three-dimensional CT image of a pancreatic tumor patient and annotating a category label of the target structure; Pancreatic tumors with CT slice thickness less than or equal to the predetermined size were selected by initial upper abdominal contrast-enhanced CT; Exclude 3D CT images with blurred CT images or CT image artifacts; The three-dimensional CT image is adjusted to a pixel matrix of a predetermined size.
3. The pancreatic tumor segmentation method based on CT images according to claim 1, characterized in that: In step S2, the 3D nnU-Net segmentation network includes: a first encoder and a first decoder, and inputting the preprocessed three-dimensional CT image into the pre-constructed 3D nnU-Net segmentation network for feature extraction includes: The first encoder extracts semantic features of a pancreatic tumor and blood vessels surrounding the pancreatic tumor, and the first decoder enhances positioning through skip connections to cover the pancreatic tumor and blood vessels surrounding the pancreatic tumor.
4. The pancreatic tumor segmentation method based on CT images according to claim 1, characterized in that: In step S3, determining the minimum bounding rectangle of the mask of the pancreatic tumor to generate a visual cue of the pancreatic tumor includes: The four vertices of the minimum bounding rectangle of the mask are used as negative sample points, and the center point of the minimum bounding rectangle is used as the positive sample point.
5. The pancreatic tumor segmentation method based on CT images according to claim 4, characterized in that: Step S3 also includes: identifying and increasing negative sample points and / or positive sample points of the pancreatic tumor and its surrounding area by an image analysis method, wherein: Selecting a plurality of positive sample points in the pancreatic tumor, wherein the positive sample points are evenly distributed in the pancreatic tumor; Selecting a plurality of negative sample points outside the pancreatic tumor, wherein the negative sample points are located outside the pancreatic tumor; An enlarged frame is defined for the pancreatic tumor, and the positive sample points or the negative sample points are selected within the enlarged frame, wherein the enlarged frame is larger than a minimum circumscribed rectangle.
6. The pancreatic tumor segmentation method based on CT images according to claim 1, characterized in that: In step S44, the continuous slice memory mechanism includes: capturing the spatial correlation between the embedding vectors corresponding to adjacent two-dimensional slices through multi-layer self-attention and cross-attention modules, and integrating the context information of adjacent two-dimensional slices.
7. The pancreatic tumor segmentation method based on CT images according to claim 1, characterized in that: In step S4, the first two-dimensional slice and the second two-dimensional slice are input into a pre-built SAM2 model for fusion, and the step further includes: The pre-built SAM2 model generates a plurality of candidate masks for the pancreatic tumor based on the second embedding vector, and calculates a confidence score for each candidate mask; The candidate mask with the highest confidence score is selected from all candidate masks of the pancreatic tumor as the final mask of the pancreatic tumor.
8. A pancreatic tumor segmentation device based on CT images, characterized in that: The invention comprises a memory and a processor, wherein the memory stores at least one program, and the at least one program is executed by the processor to implement the pancreatic tumor segmentation method based on CT images as claimed in any one of claims 1 to 7.
9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the pancreatic tumor segmentation method based on CT images as claimed in any one of claims 1 to 7 is implemented.