A method and device for automatic segmentation of liver tumors in abdominal CT images

Abstract

The present invention provides a method and device for automatic segmentation of liver tumors in abdominal CT images, the method comprising: obtaining an abdominal CT image to be segmented, and resampling and normalizing the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented, inputting the preprocessed abdominal CT image to be segmented into a trained liver tumor automatic segmentation model, the liver tumor automatic segmentation model sequentially intercepts image areas of preset sizes from the preprocessed abdominal CT image to be segmented in the form of a sliding window for segmentation, and obtains a predicted tumor mask after all areas are segmented. The liver tumor automatic segmentation model constructed and trained by the present invention can automatically segment liver tumors in abdominal CT images, and overcomes the recognition difficulties caused by the strong heterogeneity and blurred boundaries of liver tumors, and successfully solves the false negative segmentation problem that is easy to occur in the tumor area during the segmentation process, thereby improving the accuracy of liver tumor segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and in particular to an automatic segmentation method for liver tumors in abdominal CT images, an automatic segmentation device for liver tumors in abdominal CT images, an electronic device and a computer-readable medium. Background Art

[0002] The morbidity and mortality of primary liver cancer are high, ranking among the top among all types of cancer, and has become a major threat to health. Computed tomography (CT) is a widely used imaging tool that can be used to evaluate the morphology, texture, and focal lesions of the liver, helping doctors to accurately evaluate liver cancer and formulate treatment plans. However, manual annotation of tumor areas is not only time-consuming and labor-intensive, but also difficult to meet the research needs of multi-center large sample data sets. In addition, liver tumors show high heterogeneity in size, shape, location, and appearance. Existing automatic segmentation algorithms for liver tumors fail to segment tumor areas well, especially when the tumor contains a large area of ​​necrosis. The model is easily misled to only segment the necrotic area, while ignoring the tumor capsule that looks similar to normal tissue. Therefore, in clinical practice, there is an urgent need for a method that can quickly, accurately, and automatically segment tumors from liver tumor CT images.

[0003] Existing liver tumor segmentation methods can be divided into two categories: traditional machine learning-based and deep learning-based. In early studies, due to the limited data available for liver tumor analysis, machine learning methods achieved good tumor segmentation results due to their low demand for computing resources and reduced risk of model overfitting. However, these methods rely heavily on manually designed features, and the segmentation effect is not ideal when facing tumors with fuzzy boundaries and strong intratumor heterogeneity. With the expansion of multi-center data research, liver tumor segmentation methods based on traditional machine learning can no longer meet the needs of modern research. Deep learning methods, especially convolutional neural networks (CNNs), have shown great potential in liver tumor segmentation. These methods can adaptively learn tumor features and can capture the complex texture and structure of tumors. However, deep learning methods also have some limitations. First, deep learning models usually require a large amount of labeled data for training. Although they can automatically extract features, the demand for data volume is very high. Especially in medical imaging, it is often challenging to obtain large-scale, high-quality labeled data. Second, existing deep learning methods often suffer from incomplete tumor area segmentation in liver tumor segmentation. The aggressive growth of liver tumors leads to blurred boundaries between tumors and surrounding normal tissues in CT images, making it difficult for the model to distinguish and accurately segment the complete outline of the tumor. In addition, necrotic areas inside the tumor also increase the difficulty of segmentation. These necrotic areas will mislead the model to be unable to correctly identify and segment the tumor capsule, resulting in incomplete segmentation of the tumor area. In addition, differences in different imaging devices, contrast agent types, and imaging parameter settings lead to inconsistencies in the contrast and quality of CT images, which makes the model unstable when processing images from different sources and difficult to maintain efficient segmentation performance in a multi-center large sample data environment. Therefore, improving the performance of deep learning models in liver tumor segmentation and how to deal with the problem of false negative segmentation in the tumor segmentation process are of great clinical significance. Summary of the invention

[0004] In view of the above problems, the present invention is proposed to provide a method for automatically segmenting liver tumors in abdominal CT images and a corresponding device for automatically segmenting liver tumors in abdominal CT images, an electronic device and a computer-readable medium that overcome the above problems or at least partially solve the above problems.

[0005] The present invention discloses a method for automatically segmenting liver tumors in abdominal CT images, the method comprising:

[0006] Acquire an abdominal CT image to be segmented, and perform resampling and normalization processing on the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented;

[0007] The preprocessed abdominal CT image to be segmented is input into the trained liver tumor automatic segmentation model;

[0008] The liver tumor automatic segmentation model sequentially extracts image regions of preset sizes from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtains a predicted tumor mask after all regions are segmented.

[0009] Optionally, the generation of the liver tumor automatic segmentation model includes:

[0010] Acquire an abdominal CT image dataset; the abdominal CT image dataset includes an original abdominal CT image dataset and a tumor mask corresponding to the original abdominal CT image dataset;

[0011] Using a pre-trained liver segmentation model to perform liver segmentation on the original abdominal CT image dataset to obtain a liver mask;

[0012] multiplying the liver mask by the original abdominal CT image dataset to obtain an image of a liver region of interest of the original abdominal CT image dataset;

[0013] Constructing an automatic liver tumor segmentation network;

[0014] Cutting the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result;

[0015] The tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and the parameters of the liver tumor automatic segmentation network are tuned according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model.

[0016] Optionally, the liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module.

[0017] The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion;

[0018] The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

[0019] Optionally, the image of the liver region of interest is cropped into a plurality of liver region of interest image blocks of a preset size, and the blocks are input into a liver tumor automatic segmentation network for tumor segmentation, and the tumor segmentation result is output, including:

[0020] The encoding layer of the 3D U-Net baseline model is used to perform convolution downsampling operations on the image blocks of the liver region of interest to obtain the encoding layer tumor feature map;

[0021] The multi-scale feature aggregation module uses convolution kernels of different sizes to perform convolution processing on the coding layer tumor feature map to obtain a multi-scale tumor feature map, and performs element-by-element operation fusion on the multi-scale tumor feature map to obtain a multi-scale fusion feature map;

[0022] The decoding layer of the 3D U-Net baseline model is used to upsample the multi-scale fusion feature map layer by layer, and then concatenated with the encoding layer tumor feature map to obtain the decoding layer tumor feature map;

[0023] The camouflage feature mining module downsamples the coding layer tumor feature map to the same size as the decoding layer tumor feature map, adds the downsampled coding layer tumor feature map and the decoding layer tumor feature map, and then convolves to obtain a fusion feature, activates the fusion feature using an activation function, and subtracts the activated fusion feature from the all-one matrix of the same size as the activated fusion feature, and upsamples to a difference feature map of the same size as the coding layer tumor feature map after subtraction, and multiplies the difference feature map with the coding layer tumor feature map to obtain a target tumor feature map;

[0024] The target tumor feature map is converted into a tumor segmentation result through the output layer.

[0025] Optionally, comparing the tumor segmentation result with a tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and tuning the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model, including:

[0026] The tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and a stochastic gradient descent optimizer is used to optimize the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators, and a learning rate linear attenuation strategy is used for the learning rate until the liver tumor automatic segmentation network converges to obtain the liver tumor automatic segmentation model.

[0027] Optionally, the method further comprises:

[0028] Resampling and normalizing the original abdominal CT image data set, randomly flipping or rotating the processed image data set according to Cartesian coordinate axes, and adding Gaussian noise of random size to the image;

[0029] The tumor mask corresponding to the original abdominal CT image data set is subjected to the same flipping or image rotation processing.

[0030] The present invention also discloses an automatic segmentation device for liver tumors in abdominal CT images, the device comprising:

[0031] A preprocessing module is used to obtain an abdominal CT image to be segmented, and to perform resampling and normalization processing on the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented;

[0032] An input module, used for inputting the preprocessed abdominal CT image to be segmented into the trained liver tumor automatic segmentation model;

[0033] The output module is used for the automatic liver tumor segmentation model to sequentially extract image areas of preset sizes from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtain a predicted tumor mask after all areas are segmented.

[0034] Optionally, the device further comprises:

[0035] A training data acquisition module is used to acquire an abdominal CT image data set; the abdominal CT image data set includes an original abdominal CT image data set and a tumor mask corresponding to the original abdominal CT image data set;

[0036] A liver segmentation module, used to perform liver segmentation on the original abdominal CT image data set using a pre-trained liver segmentation model to obtain a liver mask;

[0037] a liver region of interest determination module, configured to multiply the liver mask by the original abdominal CT image dataset to obtain an image of the liver region of interest of the original abdominal CT image dataset;

[0038] Building a module for constructing a liver tumor automatic segmentation network;

[0039] A tumor segmentation module, used for cropping the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result;

[0040] The model parameter optimization module is used to compare the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and to tune the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model.

[0041] Optionally, the liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module.

[0042] The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion;

[0043] The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

[0044] Optionally, the tumor segmentation module includes:

[0045] The encoding submodule is used to perform convolution downsampling operation on the image block of the liver region of interest using the encoding layer of the 3D U-Net baseline model to obtain a tumor feature map of the encoding layer;

[0046] A multi-scale fusion submodule, used for the multi-scale feature aggregation module to perform convolution processing on the coding layer tumor feature map using convolution kernels of different sizes to obtain a multi-scale tumor feature map, and to perform element-by-element operation fusion on the multi-scale tumor feature map to obtain a multi-scale fusion feature map;

[0047] A decoding submodule is used to use the decoding layer of the 3D U-Net baseline model to perform layer-by-layer upsampling operations on the multi-scale fusion feature map, and concatenate it with the encoding layer tumor feature map to obtain a decoding layer tumor feature map;

[0048] A feature mining submodule, used for the camouflage feature mining module to downsample the coding layer tumor feature map to the same size as the decoding layer tumor feature map, add the downsampled coding layer tumor feature map and the decoding layer tumor feature map and then convolve to obtain a fusion feature, activate the fusion feature using an activation function, and subtract the activated fusion feature from the all-one matrix of the same size as the activated fusion feature, upsample to a difference feature map of the same size as the coding layer tumor feature map after subtraction, and multiply the difference feature map with the coding layer tumor feature map to obtain a target tumor feature map;

[0049] The tumor segmentation result output submodule is used to convert the target tumor feature map into a tumor segmentation result through an output layer.

[0050] Optionally, the model parameter optimization module includes:

[0051] The model parameter optimization submodule is used to compare the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and use a stochastic gradient descent optimizer to optimize the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators, and use a learning rate linear attenuation strategy for the learning rate until the liver tumor automatic segmentation network converges to obtain the liver tumor automatic segmentation model.

[0052] Optionally, the device further comprises:

[0053] A training data preprocessing and enhancement processing module, used for resampling and normalizing the original abdominal CT image data set, randomly flipping or rotating the processed image data set according to the Cartesian coordinate axis, and adding Gaussian noise of random size to the image;

[0054] The tumor mask enhancement processing module is used to perform the same flipping or image rotation processing on the tumor mask corresponding to the original abdominal CT image data set.

[0055] The present invention also discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0056] The memory is used to store computer programs;

[0057] The processor is used to implement the automatic segmentation method of liver tumors in abdominal CT images as described in the present invention when executing the program stored in the memory.

[0058] The present invention also discloses one or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, enable the processors to perform the method for automatic segmentation of liver tumors in abdominal CT images as described in the present invention.

[0059] The present invention includes the following advantages:

[0060] The automatic segmentation method of liver tumors in abdominal CT images of the present invention obtains an abdominal CT image to be segmented, and resamples and normalizes the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented, and inputs the preprocessed abdominal CT image to be segmented into a trained liver tumor automatic segmentation model, and the liver tumor automatic segmentation model sequentially intercepts image areas of preset sizes from the preprocessed abdominal CT image to be segmented in the form of a sliding window for segmentation, and obtains a predicted tumor mask after all areas are segmented. The liver tumor automatic segmentation model constructed and trained by the present invention can automatically segment liver tumors in abdominal CT images, and overcomes the recognition difficulties caused by the strong heterogeneity and blurred boundaries of liver tumors, and successfully solves the false negative segmentation problem that is easy to occur in the tumor area during the segmentation process, thereby improving the accuracy of liver tumor segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flowchart of the steps of a method for automatically segmenting liver tumors in abdominal CT images provided by an embodiment of the present invention;

[0062] Figure 2 is a schematic diagram of the structure of an MFA module provided in an embodiment of the present invention;

[0063] Figure 3 is a schematic diagram of the structure of a CFM module provided in an embodiment of the present invention;

[0064] Figure 4 is a schematic diagram of the structure of a liver tumor automatic segmentation model provided by an embodiment of the present invention;

[0065] Figure 5 is a flow chart of automatically segmenting liver tumors in abdominal CT images provided by an embodiment of the present invention;

[0066] Figure 6 It is a structural block diagram of an automatic segmentation device for liver tumors in abdominal CT images provided by an embodiment of the present invention;

[0067] Figure 7 is a block diagram of an electronic device provided by an embodiment of the present invention;

[0068] Figure 8 It is a schematic diagram of a computer-readable medium provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Reference Figure 1, shows a flowchart of the steps of a method for automatic segmentation of liver tumors in an abdominal CT image provided in an embodiment of the present invention, which may specifically include the following steps:

[0071] Step 101, obtaining an abdominal CT image to be segmented, and resampling and normalizing the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented;

[0072] Step 102, inputting the preprocessed abdominal CT image to be segmented into a trained liver tumor automatic segmentation model;

[0073] Step 103, the liver tumor automatic segmentation model sequentially extracts image regions of preset sizes from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtains a predicted tumor mask after all regions are segmented.

[0074] The liver tumor automatic segmentation model constructed and trained by the present invention can automatically segment liver tumors in abdominal CT images, overcome the recognition difficulties caused by the strong heterogeneity and blurred boundaries of liver tumors, and successfully solve the false negative segmentation problem that is prone to occur in the tumor area during the segmentation process, thereby improving the accuracy of liver tumor segmentation.

[0075] In one embodiment of the present invention, the generation of the liver tumor automatic segmentation model includes:

[0076] Acquire an abdominal CT image dataset; the abdominal CT image dataset includes an original abdominal CT image dataset and a tumor mask corresponding to the original abdominal CT image dataset;

[0077] Using a pre-trained liver segmentation model to perform liver segmentation on the original abdominal CT image dataset to obtain a liver mask;

[0078] multiplying the liver mask by the original abdominal CT image dataset to obtain an image of a liver region of interest of the original abdominal CT image dataset;

[0079] Constructing an automatic liver tumor segmentation network;

[0080] Cutting the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result;

[0081] The tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and the parameters of the liver tumor automatic segmentation network are tuned according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model.

[0082] In one embodiment of the present invention, the liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module.

[0083] The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion;

[0084] The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

[0085] In one embodiment of the present invention, the image of the liver region of interest is cropped into a plurality of liver region of interest image blocks of a preset size, and is input into a liver tumor automatic segmentation network for tumor segmentation, and the tumor segmentation result is output, including:

[0086] The encoding layer of the 3D U-Net baseline model is used to perform convolution downsampling operations on the image blocks of the liver region of interest to obtain the encoding layer tumor feature map;

[0087] The multi-scale feature aggregation module uses convolution kernels of different sizes to perform convolution processing on the coding layer tumor feature map to obtain a multi-scale tumor feature map, and performs element-by-element operation fusion on the multi-scale tumor feature map to obtain a multi-scale fusion feature map;

[0088] The decoding layer of the 3D U-Net baseline model is used to upsample the multi-scale fusion feature map layer by layer, and then concatenated with the encoding layer tumor feature map to obtain the decoding layer tumor feature map;

[0089] The camouflage feature mining module downsamples the coding layer tumor feature map to the same size as the decoding layer tumor feature map, adds the downsampled coding layer tumor feature map and the decoding layer tumor feature map, and then convolves to obtain a fusion feature, activates the fusion feature using an activation function, and subtracts the activated fusion feature from the all-one matrix of the same size as the activated fusion feature, and upsamples to a difference feature map of the same size as the coding layer tumor feature map after subtraction, and multiplies the difference feature map with the coding layer tumor feature map to obtain a target tumor feature map;

[0090] The target tumor feature map is converted into a tumor segmentation result through the output layer.

[0091] In one embodiment of the present invention, the tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and the parameters of the liver tumor automatic segmentation network are tuned according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model, including:

[0092] The tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and a stochastic gradient descent optimizer is used to optimize the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators, and a learning rate linear attenuation strategy is used for the learning rate until the liver tumor automatic segmentation network converges to obtain the liver tumor automatic segmentation model.

[0093] In one embodiment of the present invention, the method further comprises:

[0094] Resampling and normalizing the original abdominal CT image data set, randomly flipping or rotating the processed image data set according to Cartesian coordinate axes, and adding Gaussian noise of random size to the image;

[0095] The tumor mask corresponding to the original abdominal CT image data set is subjected to the same flipping or image rotation processing.

[0096] In view of the strong heterogeneity and blurred boundaries of liver tumors, as well as the false negative segmentation problem that easily occurs in the tumor area during the segmentation process, the present invention proposes corresponding coping strategies, adopts efficient data processing methods, and proposes an automatic liver tumor segmentation model based on camouflaged feature mining and fusion: Multi-scale Camouflaged FeatureMining and Fusion Network (MCFMFNet). MCFMFNet uses a classic 3D U-Net as a baseline model to extract tumor features. In order to improve the segmentation effect of the baseline model, a multi-scale feature aggregation module MFA module is added to the bottleneck layer of the 3D U-Net to extract tumor features at different scales, and the model's ability to express tumor heterogeneity features is enhanced through the feature fusion method of element-by-element operations. The MFA module is as follows: Figure 2 As shown. The MFA module is mainly divided into two functions: multi-scale feature extraction and local feature fusion. The multi-scale feature extraction part aims to enable the model to learn tumor features from receptive fields of different sizes, thereby enhancing its ability to explore heterogeneous tumor features. The local feature fusion part aims to make full use of tumor information in multi-scale features while reducing the impact of information redundancy on model performance. As for the false negative segmentation problem of tumor areas that is prone to occur during the segmentation process of the baseline model, an improvement plan is made. The camouflage feature mining module CFM module is designed to guide the model to focus on tumor features that are easily misclassified as background, thereby reducing the false negative rate in liver tumor segmentation. The CFM module is as follows Figure 3As shown. In addition, considering the noise interference in the shallow features, the CFM module is only deployed in the last three skip connection layers of MCFMFNet. The comprehensive model combining MFA and CFM modules has been verified to improve the segmentation accuracy of liver tumors and alleviate the false negative segmentation problem of tumor areas to a large extent. The structure diagram of the comprehensive model, i.e., the automatic segmentation model of liver tumors, is shown in Figure 4 shown.

[0097] Specifically, the mathematical forms of the multi-scale feature aggregation module and the camouflage feature mining module are expressed as follows:

[0098] The function of the Multi-scale Feature Aggregation Module (MFA) is to extract heterogeneous tumor features. It mainly includes two parts: multi-scale feature extraction and local feature fusion. In order to cope with the challenges of tumor segmentation caused by differences in size, morphology and growth location of liver tumors, a set of convolution kernels with sizes of 1, 3 and 5 are deployed in the feature extraction process of the baseline model to extract multi-scale tumor features. x For a convolution kernel of size x, the mathematical form of the multi-scale feature extraction part can be expressed as:

[0099]

[0100] In the formula, represents the j-th layer encoding feature of the network that performs multi-scale feature extraction, Concat(·) represents the feature concatenation operation, Represents the original feature map output by the jth layer of the network.

[0101] However, the extracted contains a lot of redundant information. In order to reduce the impact of feature redundancy on the segmentation performance of the model, a local feature fusion method is adopted to extract For further feature fusion, see the MFA module structure diagram. Figure 2 , its mathematical expression is as follows:

[0102]

[0103] Among them, Up represents the upsampling operation. represents an element-by-element multiplication operation. It represents the intermediate variables in the element operation process.

[0104] The role of the Camouflaged Feature Mining Module (CFM) is to guide the network to focus on tumor features that were originally misclassified as background. In the liver tumor segmentation task, the loss of tumor information during the sampling process or the unclear boundary between the tumor and normal tissue may lead to incomplete segmentation of the tumor area and a high false negative rate in the segmentation result. To this end, a camouflaged feature mining strategy is adopted to deploy the CFM module to the jump connection part of the baseline model to mine the tumor features hidden in the background as much as possible. The CFM module receives the features of the encoding and decoding parts of the network at the same time, uses gated attention to highlight key tumor features, and applies feature flipping operations to guide the network to focus on tumor features hidden in the background. See the structure diagram of the CFM module. Figure 3 , its mathematical expression is as follows:

[0105]

[0106] Among them, Down represents the downsampling operation. It represents the intermediate features in the operation process. It shows the tumor feature map after being updated by the CFM module.

[0107] The training of the liver tumor segmentation network based on camouflage feature mining and fusion can be divided into the following processes:

[0108] S1. Data processing of abdominal CT images from five private centers was performed, and the mask of the largest tumor volume in the liver region was manually outlined;

[0109] S2, using a 3D U-Net trained on a public abdominal CT image dataset to automatically segment the liver region mask on the private central data, and only extract the liver region as the input for the subsequent segmentation task;

[0110] S3, select one center as the external test set, and use the data of the remaining centers as training data sets and perform preprocessing operations;

[0111] S4, performing data enhancement processing such as flipping and adding noise on the pre-processed image;

[0112] S5. Build the MCFMFNet model and pass the data processed in step S4 into MCFMFNet for five-fold cross-training to find the optimal training parameters and save the optimal training weights to obtain the optimal training model.

[0113] S6. After determining the optimal training parameters, all training set data are included in the training process, and the validation data set is no longer divided. The model is retrained to obtain the optimal model;

[0114] S7. Input the pre-divided external test set data into the MCFMFNet trained in step S6 for testing, obtain the segmentation results and compare them with the manually annotated liver tumor mask gold standard to obtain various quantitative evaluation indicators.

[0115] Further S3 includes the following four steps:

[0116] Step 1: The intensity values ​​of all CT images were truncated to the range of [3,138] HU to eliminate the interference of irrelevant information. Step 2: Considering the inconsistency of voxel spacing between data from different centers, all images were resampled to a uniform voxel spacing of 0.67×0.67×1.50 mm. 3 The third step is to calculate the mean and standard deviation of the training set data, and normalize all samples in the training set and test set by subtracting the mean and dividing by the standard deviation. The fourth step is to adopt a patch-based training strategy to randomly extract a region of size 48×192×192 from the CT image processed in the third step as the input of MCFMFNet.

[0117] Further S4 comprises the following two steps:

[0118] Data enhancement is performed on the image processed by step S3, including random inversion and center rotation. Step 1: Randomly flip or rotate the image processed by S3 according to the Cartesian coordinate axis, and add Gaussian noise of random size to the image. Step 2: Perform the same flip or rotation on the tumor mask corresponding to the flipped or rotated image in the first step, so that the tumor mask is consistent with the enhanced data. Step S4 can expand the number of samples, reduce the occurrence of overfitting, and improve the robustness of the model and the accuracy of tumor edge area segmentation.

[0119] Further S5 includes:

[0120] The training set data processed in step S4 is divided into training set and validation set data with a ratio of 4:1 by five-fold cross-data. The MCFMFNet is trained five times, and the validation set results predicted by each fold model are quantitatively analyzed, and the parameters of the model are tuned to finally obtain the optimal training model parameters.

[0121] Further S6 comprises:

[0122] After determining the training parameters of the optimal model, all training set data is used for model training, and the validation set data is no longer divided for model tuning, and the final model training weights are saved.

[0123] Further S7 includes:

[0124] The model weights obtained in step S6 were imported into MCFMFNet and frozen. The pre-divided test set data were input into MCFMFNet. The tumor area in the test set data was segmented block by block using a sliding window strategy. Finally, the tumor mask predicted by the model was obtained and quantitatively analyzed with the manually drawn tumor mask gold standard.

[0125] Experimental parameters:

[0126] This project was implemented using PyTorch and the code was run on a server equipped with an NVIDIA GeForce RTX 2080Ti GPU. In the liver pre-segmentation stage of S2, the liver region was segmented using a 3D U-Net model and trained for 500 epochs using a stochastic gradient descent (SGD) optimizer. In the liver tumor segmentation stage, the proposed MCFMFNet was trained for 1000 epochs using the SGD optimizer. The batch size of both stages was set to 2, the initial learning rate lr was set to 0.01, and a linear learning rate decay strategy was adopted. The learning rate at the tth epoch can be expressed as:

[0127] lr=lr initial ×(1-epoch / total epoch) 0.9

[0128] like Figure 5 As shown in the figure, the specific process of automatically segmenting liver tumors in abdominal CT images is as follows:

[0129] (1) Obtain a CT image that requires pre-segmentation of the liver region and perform data preprocessing operations such as resampling and normalization on it.

[0130] (2) The image preprocessed in step (1) is input into a trained liver segmentation network, and the CT image is segmented by the liver segmentation network to obtain a liver mask, which is then multiplied by the original image to obtain the liver region of interest.

[0131] (3) Randomly crop the image of the tumor region obtained in step (2) to obtain an image block of 48×192×192 containing the tumor region, and perform data preprocessing operations such as resampling and normalization. The resampling operation is expressed as: target image size = current voxel spacing × current image size / target voxel spacing. The normalization operation is expressed as: image grayscale value = (image grayscale value - training set image grayscale mean) / training set image grayscale variance.

[0132] (4) The image block preprocessed in step (3) is used as the input of MCFMFNet for liver tumor segmentation. In the encoding part of the network, five encoding layer feature maps containing tumor semantic information can be obtained due to the convolution downsampling operation, which are denoted as

[0133] (5) In order to more fully capture the heterogeneous tumor characteristics, a set of convolution kernels with sizes of 1, 3, and 5 were used. and The three feature maps are processed to extract tumor features under different receptive fields, and the processed feature maps are recorded as and

[0134] (6) To reduce The influence of redundant information in the feature map on the segmentation performance of the model is further fused by using an element-by-element feature fusion method to further fuse the feature map obtained in (5). The specific steps are as follows: first, the upsampled and Multiply it by the upsampled Perform feature concatenation to obtain intermediate variables Then, the upsampled and Multiply to get the intermediate variable Third, the upsampled and Multiply to get the intermediate variable Finally, and Perform feature concatenation and use a convolution kernel of size 1 to reduce the dimension of the concatenated features to obtain the fused feature F i Through the above operations, the model can fully obtain the information of tumors in different locations, shapes and sizes, enhance the model's feature extraction ability for heterogeneous tumors, and alleviate the impact of information redundancy on the model's segmentation performance, thereby improving the accuracy of liver tumor segmentation.

[0135] (7)F i As the bottleneck layer feature of the tumor segmentation network, the feature extraction part has been completed. Then, it is upsampled layer by layer and compared with the encoded feature Concatenate to get the output of the network decoding part, which is recorded as

[0136] (8) Note the encoding features The incomplete segmentation of the tumor area may be caused by insufficient extraction of tumor features. A camouflage feature mining strategy is needed to guide the network to focus on tumor features that are misclassified as background. The specific operations are as follows: First, in order to reduce the impact of more noise in the shallow feature map of the network, select and Three deep network features are processed. Downsample to Same size, then Add together and pass through a convolution kernel of size 1 to get the intermediate variable Next, Activate through an activation function to get the intermediate variable Then add one with Matrices of the same size and all 1s Subtract, and upsample the subtracted feature map to After the same size, Multiply them together and finally get the feature map with updated tumor information Through the operation of step (8), the model can learn more sufficient tumor information, which greatly alleviates the false negative segmentation problem in the network segmentation results.

[0137] (9) In order to further improve the segmentation performance of the network and accelerate the convergence of the network, the deep supervision strategy is enabled for all decoding layers of the network. First, the output of the network decoding part through the activation function is recorded as Then, the manually annotated tumor mask G i Downsample to s i The corresponding output size is recorded as Third, for each group and Calculate the segmentation loss and assign different weights μ j , the total loss function and weight calculation formula are as follows:

[0138] (10) After the liver tumor segmentation network converges, the CT image to be segmented is preprocessed and then input into the network in the form of a sliding window. Regions of the original image with a size of 48×192×192 are sequentially cut out for segmentation. After all regions are segmented, the liver tumor segmentation result is finally obtained.

[0139] The automatic segmentation method of liver tumors in abdominal CT images implemented according to the present invention has at least the following beneficial effects:

[0140] First, the baseline model 3D U-Net used in this project was quantitatively evaluated for liver tumor segmentation performance, and its visual segmentation results were organized. Secondly, in order to address the problem of different sizes, morphologies and growth locations of liver tumors, the MFA module was added to the baseline model. The specific operation was to integrate it into the bottleneck layer of the baseline model, and replace the bottleneck layer features of the baseline model with the fused multi-scale features of the tumor. Then, in order to alleviate the false negative problem in the tumor segmentation process, the CFM module was integrated into the skip connections of the last three layers of the baseline model to mine the tumor features that were previously misclassified as background by the skip connection features, thereby verifying the effectiveness of the CFM module. Finally, the MFA and CFM modules were integrated into the baseline model to verify the segmentation effect of the comprehensive model MCFMFNet. As shown in Table 1, after adding the MFA module to the baseline model, most of the quantitative evaluation indicators were optimized, among which DSC (Dice similarity coefficient) increased by 2.8% and HD95 (95% Hausdorff distance) decreased by 17.37. These improvements show that the MFA module can effectively capture and integrate tumor features of different scales. In addition, the CFM module proposed in the project alleviates the false negative errors in the segmentation results to a certain extent. After adding the CFM module to the baseline model, the FNR (false negative rate) was reduced by 2.5%. This result shows that the segmentation accuracy of liver tumors can be effectively improved by guiding the model to focus on tumor features that are easily misclassified as background. In addition, integrating the above two modules into the baseline model achieved the best segmentation effect in terms of comprehensive quantitative evaluation indicators. These results show that the two modules proposed in this project can improve the segmentation performance of liver tumors to varying degrees. In particular, in terms of alleviating the false negative segmentation problem of the model, the FNR of the comprehensive model is reduced by 3.4% compared with the baseline model.

[0141] Table 1 is an ablation experiment conducted to verify the effectiveness of the MFA and CFM modules proposed in this project.

[0142]

[0143] The arrows indicate the direction of the ideal value of the quantitative evaluation index: ↑ means the higher the value, the better, and ↓ means the lower the value, the better.

[0144] In order to further evaluate the effectiveness of the proposed MCMFNet, the present invention is compared with several common medical image segmentation methods, including U-Net++, SegResNet, nnFormer and SwinUNETR. In addition, the present invention also selects several methods designed specifically for liver tumor segmentation tasks, such as RMAU-Net, APAUNet, DHT-Net and SBCNet. The specific comparative experimental results are as follows:

[0145] Table 2 Comparative experiments with other liver tumor segmentation methods on an independent center dataset

[0146]

[0147] The arrows indicate the direction of the ideal value of the quantitative evaluation index: ↑ means the higher the value, the better, and ↓ means the lower the value, the better.

[0148] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0149] Reference Figure 6 , shows a structural block diagram of an automatic segmentation device for liver tumors in abdominal CT images provided in an embodiment of the present invention, which may specifically include the following modules:

[0150] A preprocessing module 601 is used to obtain an abdominal CT image to be segmented, and to perform resampling and normalization processing on the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented;

[0151] An input module 602 is used to input the pre-processed abdominal CT image to be segmented into a trained liver tumor automatic segmentation model;

[0152] The output module 603 is used for the automatic liver tumor segmentation model to sequentially extract image regions of a preset size from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtain a predicted tumor mask after all regions are segmented.

[0153] Optionally, the device further comprises:

[0154] A training data acquisition module is used to acquire an abdominal CT image data set; the abdominal CT image data set includes an original abdominal CT image data set and a tumor mask corresponding to the original abdominal CT image data set;

[0155] A liver segmentation module, used to perform liver segmentation on the original abdominal CT image data set using a pre-trained liver segmentation model to obtain a liver mask;

[0156] a liver region of interest determination module, configured to multiply the liver mask by the original abdominal CT image dataset to obtain an image of the liver region of interest of the original abdominal CT image dataset;

[0157] Building a module for constructing a liver tumor automatic segmentation network;

[0158] A tumor segmentation module, used for cropping the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result;

[0159] The model parameter optimization module is used to compare the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and to tune the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model.

[0160] Optionally, the liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module.

[0161] The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion;

[0162] The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

[0163] Optionally, the tumor segmentation module includes:

[0164] The encoding submodule is used to perform convolution downsampling operation on the image block of the liver region of interest using the encoding layer of the 3D U-Net baseline model to obtain a tumor feature map of the encoding layer;

[0165] A multi-scale fusion submodule, used for the multi-scale feature aggregation module to perform convolution processing on the coding layer tumor feature map using convolution kernels of different sizes to obtain a multi-scale tumor feature map, and to perform element-by-element operation fusion on the multi-scale tumor feature map to obtain a multi-scale fusion feature map;

[0166] A decoding submodule is used to use the decoding layer of the 3D U-Net baseline model to perform layer-by-layer upsampling operations on the multi-scale fusion feature map, and concatenate it with the encoding layer tumor feature map to obtain a decoding layer tumor feature map;

[0167] A feature mining submodule, used for the camouflage feature mining module to downsample the coding layer tumor feature map to the same size as the decoding layer tumor feature map, add the downsampled coding layer tumor feature map and the decoding layer tumor feature map and then convolve to obtain a fusion feature, activate the fusion feature using an activation function, and subtract the activated fusion feature from the all-one matrix of the same size as the activated fusion feature, upsample to a difference feature map of the same size as the coding layer tumor feature map after subtraction, and multiply the difference feature map with the coding layer tumor feature map to obtain a target tumor feature map;

[0168] The tumor segmentation result output submodule is used to convert the target tumor feature map into a tumor segmentation result through an output layer.

[0169] Optionally, the model parameter optimization module includes:

[0170] The model parameter optimization submodule is used to compare the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and use a stochastic gradient descent optimizer to optimize the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators, and use a learning rate linear attenuation strategy for the learning rate until the liver tumor automatic segmentation network converges to obtain the liver tumor automatic segmentation model.

[0171] Optionally, the device further comprises:

[0172] A training data preprocessing and enhancement processing module, used for resampling and normalizing the original abdominal CT image data set, randomly flipping or rotating the processed image data set according to the Cartesian coordinate axis, and adding Gaussian noise of random size to the image;

[0173] The tumor mask enhancement processing module is used to perform the same flipping or image rotation processing on the tumor mask corresponding to the original abdominal CT image data set.

[0174] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0175] In addition, an embodiment of the present invention further provides an electronic device, such as Figure 7 As shown, it includes a processor 701, a communication interface 702, a memory 703 and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704.

[0176] Memory 703, used for storing computer programs;

[0177] The processor 701 is used to implement the automatic segmentation method of liver tumors in abdominal CT images as described in the above embodiment when executing the program stored in the memory 703.

[0178] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0179] The communication interface is used for communication between the above terminal and other devices.

[0180] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0181] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0182] like Figure 8 As shown, in another embodiment provided by the present invention, a computer-readable storage medium 801 is also provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the automatic segmentation method for liver tumors in abdominal CT images described in the above embodiment.

[0183] In another embodiment of the present invention, a computer program product comprising instructions is provided. When the computer program product is run on a computer, the computer executes the method for automatic segmentation of liver tumors in abdominal CT images described in the above embodiment.

[0184] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0185] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0186] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0187] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for automatic segmentation of liver tumors in abdominal CT images, characterized in that: The method comprises: Acquire an abdominal CT image to be segmented, and perform resampling and normalization processing on the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented; The preprocessed abdominal CT image to be segmented is input into the trained liver tumor automatic segmentation model; The liver tumor automatic segmentation model sequentially extracts image regions of a preset size from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtains a predicted tumor mask after all regions are segmented; The generation of the liver tumor automatic segmentation model includes: Acquire an abdominal CT image dataset; the abdominal CT image dataset includes an original abdominal CT image dataset and a tumor mask corresponding to the original abdominal CT image dataset; Using a pre-trained liver segmentation model to perform liver segmentation on the original abdominal CT image dataset to obtain a liver mask; multiplying the liver mask by the original abdominal CT image dataset to obtain an image of a liver region of interest of the original abdominal CT image dataset; Constructing a liver tumor automatic segmentation network; Cutting the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result; Comparing the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and tuning the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model; The liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module. The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion; The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

2. The method according to claim 1, characterized in that The image of the liver region of interest is cut into a plurality of liver region of interest image blocks of a preset size, and the blocks are input into a liver tumor automatic segmentation network for tumor segmentation, and the tumor segmentation result is output, including: The encoding layer of the 3D U-Net baseline model is used to perform convolution downsampling operations on the image blocks of the liver region of interest to obtain the encoding layer tumor feature map; The multi-scale feature aggregation module uses convolution kernels of different sizes to perform convolution processing on the coding layer tumor feature map to obtain a multi-scale tumor feature map, and performs element-by-element operation fusion on the multi-scale tumor feature map to obtain a multi-scale fusion feature map; The decoding layer of the 3D U-Net baseline model is used to upsample the multi-scale fusion feature map layer by layer, and then concatenated with the encoding layer tumor feature map to obtain the decoding layer tumor feature map; The camouflage feature mining module downsamples the coding layer tumor feature map to the same size as the decoding layer tumor feature map, adds the downsampled coding layer tumor feature map and the decoding layer tumor feature map, and then convolves to obtain a fusion feature, activates the fusion feature using an activation function, and subtracts the activated fusion feature from the all-one matrix of the same size as the activated fusion feature, and upsamples to a difference feature map of the same size as the coding layer tumor feature map after subtraction, and multiplies the difference feature map with the coding layer tumor feature map to obtain a target tumor feature map; The target tumor feature map is converted into a tumor segmentation result through the output layer.

3. The method according to claim 1, characterized in that Comparing the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and tuning the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model, including: The tumor segmentation result is compared with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and a stochastic gradient descent optimizer is used to optimize the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators, and a learning rate linear attenuation strategy is used for the learning rate until the liver tumor automatic segmentation network converges to obtain the liver tumor automatic segmentation model.

4. The method according to claim 1, characterized in that: The method further comprises: Resampling and normalizing the original abdominal CT image data set, randomly flipping or rotating the processed image data set according to Cartesian coordinate axes, and adding Gaussian noise of random size to the image; The tumor mask corresponding to the original abdominal CT image data set is subjected to the same flipping or image rotation processing.

5. An automatic segmentation device for liver tumors in abdominal CT images, characterized in that: The device comprises: A preprocessing module is used to obtain an abdominal CT image to be segmented, and to perform resampling and normalization processing on the abdominal CT image to be segmented to obtain a preprocessed abdominal CT image to be segmented; An input module, used for inputting the preprocessed abdominal CT image to be segmented into the trained liver tumor automatic segmentation model; An output module is used for the automatic liver tumor segmentation model to sequentially intercept image regions of a preset size from the preprocessed abdominal CT image to be segmented in the form of a sliding window, and obtain a predicted tumor mask after all regions are segmented; The device also includes: A training data acquisition module is used to acquire an abdominal CT image data set; the abdominal CT image data set includes an original abdominal CT image data set and a tumor mask corresponding to the original abdominal CT image data set; A liver segmentation module, used to perform liver segmentation on the original abdominal CT image data set using a pre-trained liver segmentation model to obtain a liver mask; a liver region of interest determination module, configured to multiply the liver mask by the original abdominal CT image dataset to obtain an image of the liver region of interest of the original abdominal CT image dataset; Building a module for constructing a liver tumor automatic segmentation network; A tumor segmentation module, used for cropping the liver region of interest image into a plurality of liver region of interest image blocks of a preset size, inputting the blocks into a liver tumor automatic segmentation network for tumor segmentation, and outputting a tumor segmentation result; A model parameter optimization module, used to compare the tumor segmentation result with the tumor mask corresponding to the original abdominal CT image data set to obtain various quantitative evaluation indicators, and to tune the parameters of the liver tumor automatic segmentation network according to the quantitative evaluation indicators to obtain the liver tumor automatic segmentation model; The liver tumor automatic segmentation network includes a 3D U-Net baseline model, a multi-scale feature aggregation module and a camouflage feature mining module. The multi-scale feature aggregation module is deployed in the bottleneck layer of the 3D U-Net baseline model to extract tumor features at different scales and perform local feature fusion; The camouflage feature mining module is deployed in the skip connection layer of the 3D U-Net baseline model to receive encoding features and decoding features, use gated attention to highlight key tumor features, and use feature flipping operations to guide the network to focus on tumor features hidden in the background.

6. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the automatic segmentation method of liver tumors in abdominal CT images as described in any one of claims 1 to 4 when executing the program stored in the memory.

7. One or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, enable the processors to perform the method for automatic segmentation of liver tumors in abdominal CT images according to any one of claims 1 to 4.

Images