A CT image tumor segmentation method, device and medium based on anatomy guidance

Abstract

The present invention provides a method, device and medium for CT image tumor segmentation based on anatomical guidance, which belongs to the technical field of medical image tumor segmentation. The method comprises: collecting human CT original images, and performing pixel-level labeling on the CT original images; performing data enhancement processing on the CT original images and the labeled images; constructing a CT image tumor segmentation model, wherein the CT image tumor segmentation model comprises a feature sharing encoder, a feature-guided decoder and a pixel sharing output module in sequence; using the data set after data enhancement processing to train the CT image tumor segmentation model, collecting the human CT original images to be segmented and pre-processing them, and then inputting them into the CT image tumor segmentation model for tumor segmentation. The present invention makes full use of the complex pattern of image data, improves the precision of segmentation, and significantly improves the accuracy of segmentation.

Description

Technical Field

[0001] The invention relates to a CT image tumor segmentation method, device and medium based on anatomical guidance, belonging to the technical field of medical image tumor segmentation. Background Art

[0002] Currently, medical image tumor segmentation is a crucial research in the field of modern medical imaging. It plays an important role in improving the accuracy of tumor diagnosis, formulating personalized treatment plans, monitoring treatment effects, and evaluating patient prognosis. By accurately identifying and distinguishing tumor tissue from surrounding normal tissue, doctors can more accurately plan surgery and radiotherapy positioning while reducing damage to healthy tissue. There are currently several methods:

[0003] 1. Methods based on manual labeling

[0004] These traditional methods rely on direct visual assessment and manual labeling by doctors or experts. Although they can provide highly accurate results in some cases, they are time-consuming and labor-intensive and susceptible to subjective judgment, limiting their feasibility in large-scale clinical applications.

[0005] 2. Two-stage machine learning based approach

[0006] These methods typically first extract features using image processing techniques and then apply classifiers for segmentation. Although these approaches have made advances in automation and accuracy, they often require elaborate feature engineering and may not fully exploit the complex patterns of image data.

[0007] 3. Simple Deep Learning Based Methods

[0008] With the development of deep learning technology, some simple deep learning methods based on convolutional neural networks (CNN) have been proposed, which can automatically learn features from images and significantly improve the accuracy and efficiency of segmentation. However, these methods usually require a large amount of labeled data for training, and usually only use the labeled tumor location for training, while ignoring other structures of the human body.

[0009] The prior art has the following defects:

[0010] High Dependency: Many existing methods are highly dependent on high-quality annotated data, which may be difficult to obtain in actual clinical settings, especially in disease areas where data is scarce.

[0011] Limited generalization ability: Although some methods perform well on specific datasets, they may perform poorly on new or different datasets, which limits the generalization ability of these techniques in diverse clinical settings.

[0012] Single data: Simple deep learning methods may only focus on identifying the tumor itself, while ignoring the contextual information of other structures in the human body, which may limit the model's understanding of the relationship between the tumor and surrounding tissues and the accuracy of segmentation.

[0013] Faced with the challenges of traditional tumor classification methods, such as over-reliance on manual annotation, insufficient generalization ability, and single data source, an accurate and reliable CT image tumor segmentation method is urgently needed. Summary of the invention

[0014] The purpose of the present invention is to provide a CT image tumor segmentation method, device and medium based on anatomical guidance, which can significantly improve the accuracy of segmentation.

[0015] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0016] A CT image tumor segmentation method based on anatomy guidance, comprising:

[0017] The original CT images of human body are collected and pixel-level labeled. Each labeled original CT image is divided into a tumor tissue labeled image and a human anatomy tissue labeled image according to the labeled content.

[0018] Data enhancement is performed on the original CT images and labeled images.

[0019] A CT image tumor segmentation model is constructed, which includes a feature sharing encoder, a feature guided decoder and a pixel sharing output module in sequence; the feature sharing encoder uses an improved ResNet-101 network, and the improved ResNet-101 network retains the first three downsampling operations of the ResNet-101 network and removes the fully connected layer; the feature guided decoder includes a backbone decoder, a tissue anatomy decoder and a tumor segmentation decoder; the pixel sharing output module fuses the output of the feature guided decoder through feature map multiplication, uses a pixel-level classifier to determine the category of each pixel, and outputs the segmentation mask of the tumor and anatomical tissue.

[0020] The data set after data enhancement is used to train the CT image tumor segmentation model. The original CT image of the human body to be segmented is collected and preprocessed, and then input into the CT image tumor segmentation model for tumor segmentation.

[0021] Preferably, the backbone decoder includes three convolutional layers and an upsampling module, and the output of the previous convolutional layer and upsampling module serves as the input of the next convolutional layer and upsampling module; the tissue anatomy decoder and tumor segmentation decoder respectively use the feature maps output by each of the three convolutional layers and upsampling operations of the backbone decoder as input.

[0022] Preferably, the tissue anatomy decoder focuses on extracting and refining normal anatomical structure features, and includes three convolution modules, with the output of the previous convolution module serving as the input of the next convolution module.

[0023] The input of the first convolution module of the tissue anatomy decoder is the output features of the first convolution and upsampling modules of the backbone decoder.

[0024] The tissue anatomy decoder second convolution module input also includes the backbone decoder second convolution and upsampling module output features.

[0025] The tissue anatomy decoder third convolution module input also includes the backbone decoder third convolution and upsampling module output features.

[0026] Preferably, the tumor segmentation decoder extracts and enhances features of the tumor area, and includes a convolution module and a guided fusion module, specifically including: a first convolution module of the tumor segmentation decoder, a first guided fusion module, a second convolution module of the tumor segmentation decoder, a second guided fusion module, a third convolution module of the tumor segmentation decoder, and a third guided fusion module, and the output of the previous module is used as the input of the next module.

[0027] The input of the first convolution module of the tumor segmentation decoder is the output features of the first convolution and upsampling modules of the backbone decoder.

[0028] The input of the first guided fusion module also includes the output features of the first convolution module of the tissue anatomy decoder.

[0029] The input of the second convolution module of the tumor segmentation decoder also includes the second convolution of the backbone decoder and the output features of the upsampling module.

[0030] The input of the second guided fusion module also includes the output features of the second convolution module of the tissue anatomy decoder.

[0031] The input of the third convolution module of the tumor segmentation decoder also includes the third convolution of the backbone decoder and the output features of the upsampling module.

[0032] The input of the third guided fusion module also includes the output features of the third convolution module of the tissue anatomy decoder.

[0033] Preferably, the guide fusion module processes the features as follows:

[0034] The key and value are multiplied by the feature map, and the weights of each stage of the segmentation decoder are added. The outputs after the key and feature map multiplication are added. The feature map multiplication and addition are parameter point-to-point operations; the specific formula is as follows:

[0035]

[0036] in, and They are feature maps and In position The value at .

[0037] Preferably, the pixel sharing output module feature processing flow is as follows:

[0038] A shared feature space with the same size as the output image is randomly initialized, and the shared feature space is multiplied with the output feature maps of the tissue anatomy decoder and the tumor segmentation decoder to achieve feature fusion.

[0039] The fused feature maps are respectively input into two pixel-level classifiers, wherein the pixel-level classifiers include a convolutional layer and a SoftMax activation function; the classifiers output a probability map indicating the confidence level of each pixel belonging to the anatomical tissue and a probability map indicating the confidence level of each pixel belonging to the tumor.

[0040] A binary segmentation mask is obtained based on the output probability map to distinguish tumors from normal tissues.

[0041] Preferably, the loss function of the CT image tumor segmentation model adopts a binary cross entropy function, the model optimization adopts an SGD optimizer, the learning rate is set to 0.001, and L2 regularization is introduced during the training process to prevent overfitting.

[0042] Preferably, the weights of the CT image tumor segmentation model are initialized using the Xavier method.

[0043] A CT image tumor segmentation device based on anatomical guidance comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the CT image tumor segmentation method based on anatomical guidance when running the program instructions.

[0044] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the CT image tumor segmentation method based on anatomical guidance.

[0045] The advantages of the present invention are:

[0046] The innovative network architecture of feature-sharing encoder and feature-guided decoder, combined with a pixel-sharing output module, achieves accurate segmentation of tumors and normal tissues in CT images, significantly improving the accuracy and robustness of segmentation.

[0047] The use of simultaneous data augmentation techniques, including random rotation, flipping, noise injection, contrast adjustment, etc., increases the diversity of medical images, allowing the model to adapt to data from different situations. CT images are labeled pixel-wise by professional radiologists or trained annotators using LabelMe software, ensuring the high quality and accuracy of the training data.

[0048] Combining the SGD optimizer with momentum and L2 regularization technology effectively avoids the overfitting problem in the model training process and improves the generalization ability of the model. The Xavier initialization method is used to randomly initialize the model weights within a reasonable range, which accelerates the convergence speed of the model.

[0049] The model's precise segmentation capability can help doctors more accurately identify and distinguish tumor tissue from surrounding normal tissue, thereby formulating more personalized and accurate treatment plans. It has important clinical application value in improving the accuracy of tumor diagnosis, formulating treatment plans, monitoring treatment effects, and evaluating patient prognosis.

[0050] In summary, the anatomically guided CT image tumor segmentation network proposed in the present invention has shown significant beneficial effects in terms of technological innovation, data quality, model optimization, and intelligent auxiliary diagnosis, and is expected to play an important role in future clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0052] Figure 1 It is a schematic diagram of the process of the present invention.

[0053] Figure 2 It is a schematic diagram of the model structure of the present invention.

[0054] Figure 3 This is a structural diagram of the feature-guided decoder of the present invention.

[0055] Figure 4 This is a schematic diagram of the structure of the special guide fusion module of the present invention.

[0056] Figure 5 It is a schematic diagram of the structure of the pixel sharing output module of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Example 1

[0059] like Figure 1-Figure 2 As shown, a CT image tumor segmentation method based on anatomical guidance includes:

[0060] S1: Collect human CT original images, perform pixel-level labeling on the CT original images, and divide each labeled CT original image into a tumor tissue labeling image and a human anatomy tissue labeling image according to the labeling content.

[0061] S2: Perform data enhancement processing on the original CT image and the labeled image.

[0062] S3: Construct a CT image tumor segmentation model, which includes a feature sharing encoder, a feature guided decoder and a pixel sharing output module in sequence; the feature sharing encoder uses an improved ResNet-101 network, and the improved ResNet-101 network retains the first three downsampling operations of the ResNet-101 network and removes the fully connected layer; the feature guided decoder includes a backbone decoder, a tissue anatomy decoder and a tumor segmentation decoder; the pixel sharing output module fuses the output of the feature guided decoder through feature map multiplication, and uses a pixel-level classifier to determine the category of each pixel, and outputs the segmentation mask of the tumor and anatomical tissue.

[0063] S4: Use the data set after data enhancement to train the CT image tumor segmentation model, collect the original CT image of the human body to be segmented, pre-process it, and input it into the CT image tumor segmentation model for tumor segmentation.

[0064] As a refinement of the above embodiment, the data required in step S1 is a human CT image, and the human anatomical tissue and tumor part also need to be pixel-level labeled to ensure that the model can accurately identify and distinguish different tissues and pathological features. These images and labeled data constitute the core of the training set. The required human CT image data comes from medical imaging data collected during clinical diagnosis. It is required to have high resolution to ensure that the details in the image are clearly visible, and cover a wide range of anatomical areas to ensure that the model can be exposed to various possible pathological conditions and tissue types.

[0065] Furthermore, professional radiologists or trained annotators use LabelMe software to classify each pixel in the CT image. Each CT image label corresponds to two labeled images, one for tumor tissue labeling and the other for human anatomy tissue labeling. Tumors include solid tumors, cystic tumors, primary tumors, and metastatic tumors. Human anatomical tissues include but are not limited to the skeletal system, muscle tissue, vascular structure, lungs, liver, kidneys, heart, and brain. These original images and labels constitute a basic data set.

[0066] Finally, in order to improve the robustness and adaptability of the model, a series of data enhancement processing is performed on the collected CT images. Here, three images are processed synchronously, and the processing methods used include random rotation, flipping, noise injection, contrast adjustment, brightness change, blurring, etc. In this embodiment, the random rotation angles used are 90°, 180°, and 270°, the noise used is Gaussian noise with a mean of 0 and a variance of 1, the contrast adjustment method used is histogram equalization, and the blur is a mean filter with a filter kernel size of 5. Each group of images performs the above operations 10 times randomly to generate 10 groups of enhanced images and save them to the data set. These synchronous processing methods ensure the diversity of image data in the model training process and the complexity of the real world. Through the above steps, a large-scale data set containing original data and enhanced data is formed. This provides a basis for subsequent deep learning model training.

[0067] As a refinement of the above embodiment, the CT image tumor segmentation model includes a feature sharing encoder, a feature guided decoder and a pixel sharing output module.

[0068] (1) Feature sharing encoder

[0069] The feature sharing encoder is the front-end part of the network, which is responsible for accepting the enhanced CT image data set as input and extracting features from it. This embodiment uses the ResNet-101 network as an encoder and makes targeted changes. Specifically, the original ResNet-101 network was designed for image classification tasks. Through continuous downsampling layers, the network gradually extracts the features of the image and gradually loses spatial resolution. However, in medical image segmentation tasks, it is more important to retain the detailed information of the image. This embodiment makes necessary modifications to ResNet-101: it streamlines the number of downsampling times of the network and only retains the first three downsampling operations. This not only greatly retains the detailed information in the CT image, but also avoids the over-smoothing problem common in deep networks.

[0070] In addition, considering that the fully connected layer of the original network is not suitable for pixel-level segmentation tasks, it is completely removed. This change allows the output of the encoder to directly serve the subsequent feature-guided decoder without unnecessary dimensionality conversion. With these changes, the feature sharing encoder can output a high-resolution feature map as the input of the subsequent decoder.

[0071] (2) Feature-guided decoder

[0072] The feature-guided decoder proposes a dual decoder structure, including a backbone decoder, a tissue anatomy decoder, and a tumor segmentation decoder. The overall structure is as follows Figure 3 As shown in the figure. Specifically, a backbone decoder with three convolutional layers and upsampling operations is first constructed. The design of this backbone decoder adopts a step-by-step refinement strategy. Through continuous convolution and upsampling layers, the spatial resolution of the image is gradually restored, while the deep features extracted from the encoder are upsampled and refined. Based on the backbone decoder, the tissue anatomy decoder and the tumor segmentation decoder take the feature maps of each stage of the backbone decoder as input respectively. The tissue anatomy decoder focuses on extracting and refining the features of normal anatomical structures, while the tumor segmentation decoder specifically extracts and enhances the features of the tumor area.

[0073] (2-1) Backbone decoder

[0074] The backbone decoder includes three convolutional layers and an upsampling module, and the output of the previous convolutional layer and upsampling module is used as the input of the next convolutional layer and upsampling module; the tissue anatomy decoder and tumor segmentation decoder respectively use the feature maps output by each of the three convolutional layers and upsampling operations of the backbone decoder as input.

[0075] (2-2) Tissue Anatomy Decoder

[0076] The tissue anatomy decoder focuses on extracting and refining normal anatomical structure features and includes three convolution modules, where the output of the previous convolution module serves as the input of the next convolution module.

[0077] The input of the first convolution module of the tissue anatomy decoder is the output features of the first convolution and upsampling modules of the backbone decoder.

[0078] The tissue anatomy decoder second convolution module input also includes the backbone decoder second convolution and upsampling module output features.

[0079] The tissue anatomy decoder third convolution module input also includes the backbone decoder third convolution and upsampling module output features.

[0080] (2-3) Tumor segmentation decoder

[0081] The tumor segmentation decoder extracts and enhances the features of the tumor area, and includes a convolution module and a guided fusion module, specifically including: a first convolution module of the tumor segmentation decoder, a first guided fusion module, a second convolution module of the tumor segmentation decoder, a second guided fusion module, a third convolution module of the tumor segmentation decoder, and a third guided fusion module, and the output of the previous module is used as the input of the latter module.

[0082] The input of the first convolution module of the tumor segmentation decoder is the output features of the first convolution and upsampling modules of the backbone decoder.

[0083] The input of the first guided fusion module also includes the output features of the first convolution module of the tissue anatomy decoder.

[0084] The input of the second convolution module of the tumor segmentation decoder also includes the second convolution of the backbone decoder and the output features of the upsampling module.

[0085] The input of the second guided fusion module also includes the output features of the second convolution module of the tissue anatomy decoder.

[0086] The input of the third convolution module of the tumor segmentation decoder also includes the third convolution of the backbone decoder and the output features of the upsampling module.

[0087] The input of the third guided fusion module also includes the output features of the third convolution module of the tissue anatomy decoder.

[0088] Furthermore, in order to achieve effective information exchange between the two decoders, a structure is designed to further fuse the input of the tissue anatomy decoder with the tumor segmentation decoder. At the same time, for better fusion, a guided fusion module is designed in the fusion process. This module draws on the idea of ​​the attention mechanism, uses the output of the tissue anatomy decoder as the key, and uses the output of each stage of the tumor segmentation decoder as the value. At the same time, a jump structure is added. Its specific structure is as follows: Figure 4 As shown in the figure, the key value is first multiplied by the feature map to achieve weighting of each stage of the segmentation decoder, and then the output after the key and feature map multiplication is added to achieve a jump structure, which can avoid gradient explosion and optimize the network convergence speed. The feature map multiplication and addition are parameter point-to-point operations, and the specific formula is as follows:

[0089]

[0090] in, and They are feature maps and In position The value at .

[0091] Through the design of the dual decoder structure and the guided fusion module, this embodiment implements a mechanism for guiding the tumor segmentation results by the knowledge features of tissue anatomy, which can improve the accuracy and robustness of the segmentation of tumors and normal anatomical structures in CT images.

[0092] (3) Pixel sharing output module

[0093] After the feature-guided decoder, the feature maps of the two paths have been restored to the same size as the original image, that is, they can be aligned point by point on a pixel scale. In order to further use tissue anatomy knowledge to guide tumor segmentation, the present invention designs a pixel sharing output module in the final segmentation output part. The module structure is as follows: Figure 5 As shown in the figure, the module structurally designs a shared feature space (in the form of a three-dimensional matrix) with the same size as the output image, which is randomly initialized. Furthermore, the shared feature space is multiplied with the output of the tissue anatomy decoder and the tumor segmentation decoder respectively to achieve feature fusion. Feature map multiplication not only enhances the expressive power of features, but also automatically adjusts the feature contribution from different decoders through weight learning, thereby more accurately capturing the relationship between the tumor and surrounding tissues.

[0094] The fused feature map is then fed into a pixel-level classifier, which contains a convolutional layer and a SoftMax function activation to ultimately determine whether each pixel belongs to the anatomy or tumor region. The output of this classifier is a probability map, indicating the confidence that each pixel belongs to the anatomy or tumor. Finally, a binary segmentation mask can be obtained based on the tumor output probability map to clearly distinguish between tumors and normal tissues.

[0095] As a refinement of the above embodiment, the training and application of the CT image tumor segmentation model are as follows:

[0096] The model proposed in this paper is based on the Pytorch deep learning framework and is a dynamic graph architecture that ensures the flexibility and efficiency of model development. The loss function uses the classic binary cross entropy function. The model is optimized using the SGD optimizer with momentum, where the learning rate is set to 0.001. The detailed formula is as follows:

[0097] ,

[0098] in, represents the model parameters, is the learning rate, is the gradient of the loss function with respect to the parameters, is the current iteration number.

[0099] In addition, to avoid overfitting during model training, L2 regularization technology was introduced to impose constraints on model weights. The dataset was divided into training subsets and validation subsets in a ratio of 8:2 to ensure good generalization performance of the model. In terms of hardware, NVIDIA Tesla V100 GPU under Linux platform was used. During training, the batch size was set to 32, which helps to improve video memory efficiency. In terms of initialization of model weights, the Xavier method was used to initialize the network parameters. The formula is as follows:

[0100] ,

[0101] in, represents uniform distribution, is the number of input features, ensuring that the weights are randomly initialized within a reasonable range.

[0102] In terms of application, the network proposed in the present invention can be integrated into CT diagnostic instruments due to its high precision and efficient processing capabilities, becoming its built-in intelligent analysis module. This integrated design allows CT scan data to be transmitted to the network model immediately after acquisition for real-time processing and analysis. Doctors and technicians can directly access the analysis results provided by the model through the instrument's user interface, including tumor location, segmentation, and malignancy probability prediction.

[0103] It should be noted that:

[0104] The present invention constructs an innovative neural network architecture by integrating human tissue anatomical labeled images and tumor labeled images. The core of the network includes a feature sharing encoder, a feature guided decoder, and a pixel sharing output module, which work together to enable the network to use anatomical information to assist in the accurate segmentation of tumors.

[0105] This method can significantly improve the accuracy of segmentation when processing complex medical images, especially in the key areas of distinguishing tumors from their adjacent tissues. This method not only brings a new perspective to the field of medical image processing, but also provides a strong technical support for future clinical practice. It is expected to play a key role in improving the quality and efficiency of tumor diagnosis and treatment.

[0106] Example 2

[0107] The disclosed embodiment also provides a CT image tumor segmentation device based on anatomical guidance, including a processor and a memory. Optionally, the device may also include a communication interface and a bus. The processor, the communication interface, and the memory may communicate with each other through the bus. The communication interface may be used for information transmission. The processor may call the logic instructions in the memory to execute the CT image tumor segmentation method based on anatomical guidance of the above embodiment.

[0108] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0109] The memory, as a computer-readable storage medium, can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes the function application and data processing by running the program instructions / modules stored in the memory, that is, the CT image tumor segmentation method based on anatomical guidance in the above embodiment is realized.

[0110] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0111] An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned CT image tumor segmentation method based on anatomical guidance.

[0112] The computer-readable storage medium mentioned above may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0113] The technical solution of the embodiment of the present disclosure can be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for enabling a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in the embodiment of the present disclosure. The aforementioned storage medium may be a non-transient storage medium, including: a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, and other media that can store program codes, or a transient storage medium.

[0114] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A CT image tumor segmentation method based on anatomy guidance, characterized in that: include: Collect human CT original images, perform pixel-level labeling on the CT original images, and divide each labeled CT original image into a tumor tissue labeling image and a human anatomy tissue labeling image according to the labeling content; Perform data enhancement processing on the original CT images and labeled images; A CT image tumor segmentation model is constructed, which includes a feature sharing encoder, a feature guided decoder and a pixel sharing output module in sequence; the feature sharing encoder uses an improved ResNet-101 network, and the improved ResNet-101 network retains the first three downsampling operations of the ResNet-101 network and removes the fully connected layer; the feature guided decoder includes a backbone decoder, a tissue anatomy decoder and a tumor segmentation decoder; the pixel sharing output module fuses the output of the feature guided decoder through feature map multiplication, and uses a pixel-level classifier to determine the category of each pixel, and outputs a segmentation mask of the tumor and the anatomical tissue; The data set after data enhancement is used to train the CT image tumor segmentation model. The original CT image of the human body to be segmented is collected and preprocessed, and then input into the CT image tumor segmentation model for tumor segmentation.

2. The method for CT image tumor segmentation based on anatomy guidance according to claim 1, characterized in that: The backbone decoder includes three convolutional layers and an upsampling module, and the output of the previous convolutional layer and upsampling module is used as the input of the next convolutional layer and upsampling module; the tissue anatomy decoder and tumor segmentation decoder respectively use the feature maps output by each of the three convolutional layers and upsampling operations of the backbone decoder as input.

3. The method for CT image tumor segmentation based on anatomy guidance according to claim 2, characterized in that: The tissue anatomy decoder focuses on extracting and refining normal anatomical structure features, and includes three convolution modules, where the output of the previous convolution module serves as the input of the next convolution module; The input of the first convolution module of the tissue anatomy decoder is the first convolution of the backbone decoder and the output features of the upsampling module; The second convolution module input of the tissue anatomy decoder also includes the second convolution of the backbone decoder and the output features of the upsampling module; The tissue anatomy decoder third convolution module input also includes the backbone decoder third convolution and upsampling module output features.

4. The method for CT image tumor segmentation based on anatomy guidance according to claim 3, characterized in that: The tumor segmentation decoder extracts and enhances features of the tumor region, including a convolution module and a guided fusion module, specifically including: a first convolution module of the tumor segmentation decoder, a first guided fusion module, a second convolution module of the tumor segmentation decoder, a second guided fusion module, a third convolution module of the tumor segmentation decoder, and a third guided fusion module, wherein the output of the previous module is used as the input of the latter module; The input of the first convolution module of the tumor segmentation decoder is the first convolution of the backbone decoder and the output features of the upsampling module; The input of the first guided fusion module also includes the output features of the first convolutional module of the tissue anatomy decoder; The input of the second convolution module of the tumor segmentation decoder also includes the output features of the second convolution of the backbone decoder and the upsampling module; The input of the second guided fusion module also includes the output features of the second convolution module of the tissue anatomy decoder; The input of the third convolution module of the tumor segmentation decoder also includes the output features of the third convolution of the backbone decoder and the upsampling module; The input of the third guided fusion module also includes the output features of the third convolution module of the tissue anatomy decoder.

5. The method for CT image tumor segmentation based on anatomy guidance according to claim 4, characterized in that: The guide fusion module processes the features as follows: The key and value are multiplied by the feature map, and the weights of each stage of the segmentation decoder are added. The outputs after the key and feature map multiplication are added. The feature map multiplication and addition are parameter point-to-point operations; the specific formula is as follows: in, and They are feature maps and In position The value at .

6. The method for CT image tumor segmentation based on anatomy guidance according to claim 1, characterized in that: The pixel sharing output module feature processing flow is as follows: A shared feature space with the same size as the output image is randomly initialized, and the shared feature space is multiplied with the output feature maps of the tissue anatomy decoder and the tumor segmentation decoder to achieve feature fusion; The fused feature maps are respectively input into two pixel-level classifiers, wherein the pixel-level classifiers include a convolution layer and a SoftMax activation function; the classifiers output a probability map indicating the confidence level of each pixel belonging to the anatomical tissue and a probability map indicating the confidence level of each pixel belonging to the tumor; A binary segmentation mask is obtained based on the output probability map to distinguish tumors from normal tissues.

7. The method for CT image tumor segmentation based on anatomy guidance according to claim 1, characterized in that: The loss function of the CT image tumor segmentation model adopts the binary cross entropy function, the model optimization adopts the SGD optimizer, the learning rate is set to 0.001, and L2 regularization is introduced during the training process to prevent overfitting.

8. The method for CT image tumor segmentation based on anatomy guidance according to claim 1, characterized in that: The Xavier method is used to initialize the weights of the CT image tumor segmentation model.

9. A CT image tumor segmentation device based on anatomy guidance, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the CT image tumor segmentation method based on anatomical guidance according to any one of claims 1 to 6 when running the program instructions.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method described in any one of claims 1 to 6 is implemented.

Images