Liver segmentation method of abdominal CT image and CT imaging method thereof

Abstract

The present invention discloses a liver segmentation method for abdominal CT images, including obtaining original abdominal CT image data and liver segmentation data; constructing a model training set and a model test set; constructing an original liver segmentation model; constructing a training loss function; using the model training set and the model test set, training and testing the original liver segmentation model based on the training loss function and obtaining a liver segmentation model; using the liver segmentation model to segment the liver of a target abdominal CT image, and performing morphological optimization to obtain a final liver segmentation result of the target abdominal CT image. The present invention also discloses a CT imaging method including the above-mentioned liver segmentation method for abdominal CT images. The present invention can improve the continuity between slices of CT sequence liver segmentation results, realize accurate liver segmentation in end-to-end abdominal CT images, improve boundary accuracy, and has strong robustness, good segmentation effect and good reliability.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to a liver segmentation method of an abdominal CT image and a CT imaging method thereof. Background Art

[0002] With the development of economy and technology and the improvement of people's living standards, people are paying more and more attention to their health.

[0003] The liver is one of the most important organs in the human body. Abdominal CT images are one of the most important means for medical personnel to obtain the liver status. Therefore, it is particularly important to segment the liver completely and accurately from abdominal CT images.

[0004] However, the liver has low contrast with the surrounding organs, weak boundaries, and large differences in liver morphology between cases. Accurate segmentation of the liver in abdominal CT images is still difficult. At the same time, due to the large number of slices in CT images, relying on experts to manually outline the liver is not only time-consuming and labor-intensive, but also dependent on expert experience. Therefore, it is of great significance to study the automatic segmentation method of the liver in abdominal CT sequence images.

[0005] Existing liver segmentation methods based on abdominal CT sequence images mainly include three categories: methods based on traditional image analysis, methods based on machine learning, and methods based on deep learning. Based on traditional image analysis, the segmentation process is complex, time-consuming, and requires manual intervention. It is difficult to achieve fully automatic segmentation, and the segmentation results are not ideal. Based on machine learning, the design requirements for discriminative features are very high, and the complexity of medical images is high, so it is difficult to obtain ideal results by discriminative feature design. The method based on deep learning can automatically learn discriminative features based on the training set to achieve automatic liver segmentation; and the design of neural network structure and loss function is the key to liver segmentation based on deep learning. In terms of neural network structure design, the commonly used two-dimensional convolutional neural network can only obtain convolutional features within CT image slices, completely ignoring the contextual information between CT sequence slices, while the three-dimensional convolutional neural network can obtain convolutional features containing contextual information between slices using three-dimensional convolution kernels, but the number of learnable parameters is too large, which requires a lot of computing resources. In addition, in terms of loss function design, the current common technology is only based on the traditional loss function, that is, regional loss, but the current method has the disadvantage of being not sensitive enough to boundary errors, resulting in low segmentation boundary accuracy. Summary of the invention

[0006] One of the purposes of the present invention is to provide a liver segmentation method for abdominal CT images with high segmentation boundary accuracy, good segmentation effect and good reliability.

[0007] A second object of the present invention is to provide a CT imaging method including the liver segmentation method of the abdominal CT image.

[0008] The liver segmentation method of the abdominal CT image provided by the present invention comprises the following steps:

[0009] S1. Acquire original abdominal CT image data and liver segmentation data;

[0010] S2. Construct a model training set and a model test set based on the data obtained in step S1;

[0011] S3. Construct the original liver segmentation model;

[0012] S4. Construct training loss function;

[0013] S5. Using the model training set and model test set constructed in step S2, based on the training loss function constructed in step S4, the original liver segmentation model constructed in step S3 is trained and tested to obtain a liver segmentation model;

[0014] S6. Using the liver segmentation model obtained in step S5, perform liver segmentation on the target abdominal CT image, and perform morphological optimization to obtain the final liver segmentation result of the target abdominal CT image.

[0015] The step S1 of obtaining the original abdominal CT image data and the liver segmentation data specifically involves obtaining the original abdominal CT image and the liver mask data.

[0016] The construction of the liver segmentation original model described in step S3 specifically includes the following steps:

[0017] A. Based on the convolutional neural network with pre-trained weights, a convolutional layer and a normalization layer are connected in series to obtain the first module;

[0018] B. Use the spatial pyramid pooling module as the second module;

[0019] C. The attention slice interaction model is used as the third module;

[0020] D. Use the full convolution layer and upsampling layer as the fourth module;

[0021] E. Connect the modules obtained in steps A to D in series to form the original liver segmentation model.

[0022] The first module is obtained by connecting a convolutional layer and a normalization layer in series based on the convolutional neural network with pre-trained weights as described in step A, which specifically includes the following steps:

[0023] A1. Train the ResNet101 network on ImageNet to obtain a convolutional neural network with pre-trained weights;

[0024] A2. Retain the first to fourth convolution modules of the convolutional neural network containing pre-trained weights obtained in step A1, and the pre-trained parameters corresponding to the first to fourth convolution modules; at this time, the number of output channels is 1024-dimensional convolution features;

[0025] A3. After the fourth convolution module obtained in step A2, a 1×1 convolution layer is connected in series, and the number of output channels is now 256-dimensional convolution features;

[0026] A4. A batch normalization layer is connected in series after the 1×1 convolution layer in step A3; at this time, the number of output channels is 256-dimensional convolution features.

[0027] The spatial pyramid pooling module as the second module in step B specifically includes the following steps:

[0028] B1. The spatial pyramid pooling module includes a pyramid pooling layer, an upsampling layer, a feature cascade layer, a convolution layer, and a normalization layer;

[0029] B2. The pyramid pooling layer consists of four parallel pooling layers with pooling kernel sizes of 1×1, 2×2, 3×3, and 6×6. After pooling, four sets of multi-scale features with a total of 256 output channels are obtained.

[0030] B3. The upsampling layer is bilinear interpolation upsampling: for features whose feature spatial resolution is smaller than the input feature spatial resolution, interpolation is performed to obtain 4 sets of upsampled multi-scale features with 256 output channels and a spatial resolution equal to the input feature spatial resolution;

[0031] B4. The feature cascade layer cascades the four groups of upsampled multi-scale features to obtain a multi-scale feature with a channel number of 1024 dimensions;

[0032] B5. The convolution layer is a 1×1 convolution layer, which is used to output multi-scale features with a channel number of 512 dimensions;

[0033] B6. The normalization layer is a batch normalization layer, which is used to output multi-scale features with a channel number of 512 dimensions.

[0034] The attention slice interaction model described in step C is used as the third module, and specifically includes the following steps:

[0035] C1. The attention slice interaction model includes the attention slice interaction layer, convolution layer and normalization layer;

[0036] C2. Input the multi-scale features output by the second module into the third module to obtain the multi-scale features containing the inter-slice context; at the same time, input the output features of the first module into the third module to obtain the convolutional features containing the inter-slice context; wherein the output features of the first module are input into the third module only during the model training, and when the model training is completed, the output features of the first module are no longer input into the third module;

[0037] C3. The attention slice interaction layer is implemented by measuring the correlation between pixels between slices;

[0038] C4. The convolution layer is a 1×1 convolution layer, which is used to output features with 256 channels. The normalization layer is a batch normalization layer, which is used to output features with 256 channels.

[0039] The attention slice interaction layer is implemented by measuring the correlation between pixels between slices, specifically using the following formula:

[0040]

[0041]

[0042]

[0043] Where F Ie(i+1),q F is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the next slice i+1 through the attention slices; Se(i),q is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the current slice i through the attention slices; F Ie(i-1),q The feature obtained by the interaction between the attention slices is the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the next slice i+1; F i,q is the feature vector of pixel q in the feature map of the current slice i; N is the total number of feature map pixels in the slice; is the correlation between pixel q of the current slice i and pixel p of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the current slice i; The calculation formula is The calculation formula is The calculation formula is exp() is the exponential function.

[0044] The fourth module described in step D is a full convolution layer and an upsampling layer, which specifically includes the following steps:

[0045] The output result obtained in step C is input into the full convolution layer and upsampled by the upsampling layer to obtain the final liver segmentation result; the number of output channels of the full convolution layer is 2, and the upsampling layer is 8 times bilinear interpolation.

[0046] The construction of the training loss function described in step S4 specifically includes the following steps:

[0047] a. Use the following formula to construct the regional term Loss of the loss function region :

[0048] Loss region =BCE(P,GT)

[0049] Where P is the predicted probability of predicting the pixel as liver; GT is the binary label of the pixel; BCE() is the binary cross entropy function;

[0050] b. Use the following formula to construct the boundary term Loss of the loss function BDice :

[0051]

[0052] Where S is the liver segmentation mask; GT is the binary label of the pixel; Canny() is the canny edge detection operator; dilate(,se) is the morphological dilation operation with a dilation factor of se, se is a kernel of size 3×3 pixels and value 1; ε is the smoothing factor;

[0053] c. Build side output loss function Loss side Loss side =Loss region +Loss BDice ;

[0054] d. Construct the final output loss function Loss seg Loss seg =Loss region +Loss BDice ;

[0055] e. Finally, the global liver segmentation loss function Loss is constructed as Loss = αLoss side +βLoss seg ; α and β are weight values.

[0056] The morphological optimization described in step S6 specifically includes filling three-dimensional holes and retaining the maximum three-dimensional connected domain.

[0057] The present invention also provides a CT imaging method including the above-mentioned liver segmentation method of abdominal CT images, further comprising the following steps:

[0058] S7. Mark the liver segmentation result of the final target abdominal CT image obtained in step S6 on the original CT image, thereby outputting a CT image with the liver segmentation result.

[0059] The liver segmentation method of the abdominal CT image and the CT imaging method thereof provided by the present invention propose an attention slice interaction model, obtain context information between CT sequence slices on the basis of a two-dimensional convolutional neural network, and improve the continuity between slices of the CT sequence liver segmentation result with a small amount of learnable parameters; at the same time, the present invention proposes a boundary-based loss function and combines it with a regional loss function, so that the model automatically learns and pays attention to the liver boundary information, thereby improving the boundary accuracy; therefore, the method of the present invention can realize end-to-end accurate liver segmentation in abdominal CT images with a small amount of parameters, and has strong robustness, high accuracy of the segmentation boundary, good segmentation effect and good reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the process flow of the segmentation method of the present invention.

[0061] Figure 2 It is a structural schematic diagram of the liver segmentation original model of the segmentation method of the present invention.

[0062] Figure 3 It is a schematic diagram of the result of an embodiment of the segmentation method of the present invention.

[0063] Figure 4 The figure is a schematic diagram of the method flow of the imaging method of the present invention. DETAILED DESCRIPTION

[0064] like Figure 1 The figure shows a flow chart of the segmentation method of the present invention: the liver segmentation method of the abdominal CT image provided by the present invention comprises the following steps:

[0065] S1. Acquire original abdominal CT image data and liver segmentation data; specifically, acquire original abdominal CT image and liver mask data;

[0066] S2. Construct a model training set and a model test set based on the data obtained in step S1;

[0067] S3. Construct the original liver segmentation model (such as Figure 2 Specifically, the following steps are included:

[0068] A. Based on the convolutional neural network with pre-trained weights, a convolutional layer and a normalization layer are connected in series to obtain the first module; specifically, the steps include:

[0069] A1. Train the ResNet101 network on ImageNet to obtain a convolutional neural network with pre-trained weights;

[0070] A2. Retain the first to fourth convolution modules of the convolutional neural network containing pre-trained weights obtained in step A1, and the pre-trained parameters corresponding to the first to fourth convolution modules; at this time, the number of output channels is 1024-dimensional convolution features;

[0071] A3. After the fourth convolution module obtained in step A2, a 1×1 convolution layer is connected in series, and the number of output channels is now 256-dimensional convolution features;

[0072] A4. A batch normalization layer is connected in series after the 1×1 convolution layer in step A3; at this time, the number of output channels is 256-dimensional convolution features;

[0073] B. Using the spatial pyramid pooling module as the second module; specifically including the following steps:

[0074] B1. The spatial pyramid pooling module includes a pyramid pooling layer, an upsampling layer, a feature cascade layer, a convolution layer, and a normalization layer;

[0075] B2. The pyramid pooling layer consists of four parallel pooling layers with pooling kernel sizes of 1×1, 2×2, 3×3, and 6×6. After pooling, four sets of multi-scale features with a total of 256 output channels are obtained.

[0076] B3. The upsampling layer is bilinear interpolation upsampling: for features whose feature spatial resolution is smaller than the input feature spatial resolution, interpolation is performed to obtain 4 sets of upsampled multi-scale features with 256 output channels and a spatial resolution equal to the input feature spatial resolution;

[0077] B4. The feature cascade layer cascades the four groups of upsampled multi-scale features to obtain a multi-scale feature with a channel number of 1024 dimensions;

[0078] B5. The convolution layer is a 1×1 convolution layer, which is used to output multi-scale features with a channel number of 512 dimensions;

[0079] B6. The normalization layer is a batch normalization layer, which is used to output multi-scale features with a channel number of 512 dimensions;

[0080] C. The attention slice interaction model is used as the third module; specifically, the following steps are included:

[0081] C1. The attention slice interaction model includes the attention slice interaction layer, convolution layer and normalization layer;

[0082] C2. Input the multi-scale features output by the second module into the third module to obtain the multi-scale features containing the inter-slice context; at the same time, input the output features of the first module into the third module to obtain the convolutional features containing the inter-slice context; wherein the output features of the first module are input into the third module only during the model training, and when the model training is completed, the output features of the first module are no longer input into the third module;

[0083] C3. The attention slice interaction layer is implemented by measuring the correlation between pixels between slices; specifically, it is implemented using the following formula:

[0084]

[0085]

[0086]

[0087] Where F Ie(i+1),q F is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the next slice i+1 through the attention slices; Se(i),q is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the current slice i through the attention slices; F Ie(i-1),q The feature obtained by the interaction between the attention slices is the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the next slice i+1; F i,q is the feature vector of pixel q in the feature map of the current slice i; N is the total number of feature map pixels in the slice; is the correlation between pixel q of the current slice i and pixel p of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the current slice i; The calculation formula is The calculation formula is The calculation formula is exp() is the exponential function

[0088] C4. The convolution layer is a 1×1 convolution layer, which is used to output features with 256 channels. The normalization layer is a batch normalization layer, which is used to output features with 256 channels.

[0089] D. The full convolution layer and upsampling layer are used as the fourth module; specifically, the following steps are included:

[0090] The output result obtained in step C is input into the full convolution layer, and upsampled by the upsampling layer to obtain the final liver segmentation result; the number of output channels of the full convolution layer is 2, and the upsampling layer is 8 times bilinear interpolation;

[0091] E. Connecting the modules obtained in steps A to D in series in sequence to form an original liver segmentation model;

[0092] S4. Construct a training loss function; specifically, the following steps are included:

[0093] a. Use the following formula to construct the regional term Loss of the loss function region :

[0094] Loss region =BCE(P,GT)

[0095] Where P is the predicted probability of predicting the pixel as liver; GT is the binary label of the pixel; BCE() is the binary cross entropy function;

[0096] b. Use the following formula to construct the boundary term Loss of the loss function BDice :

[0097]

[0098] Where S is the liver segmentation mask; GT is the binary label of the pixel; Canny() is the canny edge detection operator; dilate(,se) is the morphological dilation operation with a dilation factor of se, se is a kernel of size 3×3 pixels and value 1; ε is the smoothing factor, preferably 1e-5;

[0099] c. Build side output loss function Loss side Loss side =Loss region +Loss BDice ;

[0100] d. Construct the final output loss function Loss seg Loss seg =Loss region +Loss BDice ;

[0101] e. Finally, the global liver segmentation loss function Loss is constructed as Loss = αLoss side +βLoss seg ; α and β are weight values, preferably α = 0.4, β = 1;

[0102] S5. Using the model training set and model test set constructed in step S2, based on the training loss function constructed in step S4, the original liver segmentation model constructed in step S3 is trained and tested to obtain a liver segmentation model;

[0103] S6. Using the liver segmentation model obtained in step S5, perform liver segmentation on the target abdominal CT image, and perform morphological optimization to obtain the final liver segmentation result of the target abdominal CT image; wherein the morphological optimization specifically includes three-dimensional cavity filling and retaining the maximum three-dimensional connected domain.

[0104] When the above method is specifically implemented, in step C2, the multi-scale features output by the second module are input into the third module to obtain the multi-scale features containing the inter-slice context; at the same time, the output features of the first module are input into the third module to obtain the convolution features containing the inter-slice context;

[0105] Among them, the output features of the first module are input into the third module only during model training. When the model training is completed, the output features of the first module are no longer input into the third module; because the features output by the first module are input into the third module, after subsequent processing steps (including subsequent attention slice interaction model processing, and full convolution layer processing and upsampling layer processing of the fourth module), the side output segmentation result is obtained; this part of the side output segmentation result is only used during model training (specifically reflected in the loss function part), mainly for supervised learning; after the model training is completed (including specific applications), the features output by the first module no longer need to be input into the third module;

[0106] The features output by the second module are input into the third module and processed through subsequent steps (including the subsequent attention slice interaction model processing, and the full convolution layer processing and upsampling layer processing of the fourth module) to obtain the final segmentation result. Therefore, whether it is during training or in specific applications after training and testing, the features output by the second module must be input into the third module and processed through subsequent steps to obtain the final liver segmentation result.

[0107] Figure 3 The figure shows a schematic diagram of the liver segmentation result of an embodiment of the present invention: wherein the white edge is the gold standard of the liver edge, and the black edge is the edge of the liver segmentation result of the embodiment of the present invention. It can be seen that the method of the present invention can indeed obtain accurate segmentation results, and the segmentation edge quality is high.

[0108] The liver segmentation results of the test set LiTS test set and 3D-IRCADb were evaluated using volumetric overlap error (VOE), relative volume difference (RVD), average symmetric surface distance (ASD), root mean square symmetric surface distance (RMSD), maximum symmetric surface distance (MSD) and Dice coefficient (Dice). The results are shown in Table 1:

[0109] Table 1 Schematic diagram of liver image segmentation results

[0110] Test Set VOE RVD ASD RMSD MSD Dice LiTS test set 0.11±0.04 0.06±0.05 2.37±1.86 4.02±3.10 29.75±13.34 0.942±0.024 3D-IRCADb 0.057±0.012 0.013±0.020 1.11±0.17 1.74±0.67 20.02±9.60 0.971±0.007

[0111] It can be found from Table 1 that for different databases, the method of the present invention has smaller VOE, RVD, ASD, RMSD, MSD mean values ​​and larger Dice mean values, which proves that the method of the present invention has higher accuracy in liver segmentation in abdominal CT images; at the same time, the present invention has smaller standard deviation values ​​for the above evaluation indicators, which proves that the method of the present invention has stronger robustness in liver segmentation in abdominal CT images.

[0112] like Figure 4 The figure shows a schematic flow chart of the imaging method of the present invention: the CT imaging method provided by the present invention including the liver segmentation method of the abdominal CT image comprises the following steps:

[0113] S1. Acquire original abdominal CT image data and liver segmentation data;

[0114] S2. Construct a model training set and a model test set based on the data obtained in step S1;

[0115] S3. Construct the original liver segmentation model;

[0116] S4. Construct training loss function;

[0117] S5. Using the model training set and model test set constructed in step S2, based on the training loss function constructed in step S4, the original liver segmentation model constructed in step S3 is trained and tested to obtain a liver segmentation model;

[0118] S6. Using the liver segmentation model obtained in step S5, perform liver segmentation on the target abdominal CT image, and perform morphological optimization to obtain a final liver segmentation result of the target abdominal CT image;

[0119] S7. Mark the liver segmentation result of the final target abdominal CT image obtained in step S6 on the original CT image, thereby outputting a CT image with the liver segmentation result.

[0120] The CT imaging method provided by the present invention can be integrated into existing abdominal CT imaging equipment (CT machine, etc.); when used, the equipment first obtains the corresponding abdominal CT image according to the existing technology, and then uses the above-mentioned abdominal CT image liver segmentation method to detect, identify and segment the acquired abdominal CT image, and after segmenting to obtain the liver part, the segmentation result is then marked and re-imaged on the acquired abdominal CT image, and then the final CT image with the liver segmentation result and the marked liver segmentation result is output.

Claims

1. A liver segmentation method for abdominal CT images, comprising the following steps: S1. Acquire original abdominal CT image data and liver segmentation data; S2. Construct a model training set and a model test set based on the data obtained in step S1; S3. Constructing the original liver segmentation model; specifically comprising the following steps: A. Based on the convolutional neural network with pre-trained weights, a convolutional layer and a normalization layer are connected in series to obtain the first module; specifically, the steps include: A1. Train the ResNet101 network on ImageNet to obtain a convolutional neural network with pre-trained weights; A2. Retain the first to fourth convolution modules of the convolutional neural network containing pre-trained weights obtained in step A1, and the pre-trained parameters corresponding to the first to fourth convolution modules; at this time, the number of output channels is 1024-dimensional convolution features; A3. After the fourth convolution module obtained in step A2, a 1×1 convolution layer is connected in series, and the number of output channels is now 256-dimensional convolution features; A4. A batch normalization layer is connected in series after the 1×1 convolution layer in step A3; at this time, the number of output channels is 256-dimensional convolution features; B. Using the spatial pyramid pooling module as the second module; specifically including the following steps: B1. The spatial pyramid pooling module includes a pyramid pooling layer, an upsampling layer, a feature cascade layer, a convolution layer, and a normalization layer; B2. The pyramid pooling layer consists of four parallel pooling layers with pooling kernel sizes of 1×1, 2×2, 3×3, and 6×6. After pooling, four sets of multi-scale features with a total of 256 output channels are obtained. B3. The upsampling layer is bilinear interpolation upsampling: for features whose feature spatial resolution is smaller than the input feature spatial resolution, interpolation is performed to obtain 4 sets of upsampled multi-scale features with 256 output channels and a spatial resolution equal to the input feature spatial resolution; B4. The feature cascade layer cascades the four groups of upsampled multi-scale features to obtain a multi-scale feature with a channel number of 1024 dimensions; B5. The convolution layer is a 1×1 convolution layer, which is used to output multi-scale features with a channel number of 512 dimensions; B6. The normalization layer is a batch normalization layer, which is used to output multi-scale features with a channel number of 512 dimensions; C. The attention slice interaction model is used as the third module; specifically, the following steps are included: C1. The attention slice interaction model includes the attention slice interaction layer, convolution layer and normalization layer; C2. Input the multi-scale features output by the second module into the third module to obtain the multi-scale features containing the inter-slice context; at the same time, input the output features of the first module into the third module to obtain the convolutional features containing the inter-slice context; wherein the output features of the first module are input into the third module only during the model training, and when the model training is completed, the output features of the first module are no longer input into the third module; C3. The attention slice interaction layer is implemented by measuring the correlation between pixels between slices; specifically, it is implemented using the following formula: Where F Ie(i+1),q F is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the next slice i+1 through the attention slices; Se(i),q is the feature obtained by the interaction between the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the current slice i through the attention slices; F Ie(i-1),q The feature obtained by the interaction between the attention slices is the pixel point q in the feature map of the current slice i and the pixel point q in the feature map of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the next slice i+1; F i,q is the feature vector of pixel q in the feature map of the current slice i; N is the total number of feature map pixels in the slice; is the correlation between pixel q of the current slice i and pixel p of the previous slice i-1; is the correlation between pixel q of the current slice i and pixel p of the current slice i; The calculation formula is The calculation formula is The calculation formula is exp() is the exponential function; C4. The convolution layer is a 1×1 convolution layer, which is used to output features with 256 channels. The normalization layer is a batch normalization layer, which is used to output features with 256 channels. D. Use the full convolution layer and upsampling layer as the fourth module; E. Connecting the modules obtained in steps A to D in series in sequence to form an original liver segmentation model; S4. Construct training loss function; S5. Using the model training set and model test set constructed in step S2, based on the training loss function constructed in step S4, the original liver segmentation model constructed in step S3 is trained and tested to obtain a liver segmentation model; S6. Using the liver segmentation model obtained in step S5, perform liver segmentation on the target abdominal CT image, and perform morphological optimization to obtain the final liver segmentation result of the target abdominal CT image.

2. The method for liver segmentation of abdominal CT images according to claim 1, characterized in that The fourth module described in step D is a full convolution layer and an upsampling layer, which specifically includes the following steps: The output result obtained in step C is input into the full convolution layer and upsampled by the upsampling layer to obtain the final liver segmentation result; the number of output channels of the full convolution layer is 2, and the upsampling layer is 8 times bilinear interpolation.

3. The method for liver segmentation of abdominal CT images according to claim 1 or 2, characterized in that The construction of the training loss function described in step S4 specifically includes the following steps: a. Use the following formula to construct the regional term Loss of the loss function region : Loss region =BCE(P,GT) Where P is the predicted probability of predicting the pixel as liver; GT is the binary label of the pixel; BCE() is the binary cross entropy function; b. Use the following formula to construct the boundary term Loss of the loss function BDice : Where S is the liver segmentation mask; GT is the binary label of the pixel; Canny() is the canny edge detection operator; dilate(,se) is the morphological dilation operation with a dilation factor of se, se is a kernel of size 3×3 pixels and value 1; ε is the smoothing factor; c. Build side output loss function Loss side Loss side =Loss region +Loss BDice ; d. Construct the final output loss function Loss seg Loss seg =Loss region +Loss BDice ; e. Finally, the global liver segmentation loss function Loss is constructed as Loss = αLoss side +βLoss seg ; α and β are weight values.

4. The method for liver segmentation of abdominal CT images according to claim 1, characterized in that The morphological optimization described in step S6 specifically includes filling three-dimensional holes and retaining the maximum three-dimensional connected domain; the obtaining of original abdominal CT image data and liver segmentation data described in step S1 specifically includes obtaining original abdominal CT image and liver mask data.

5. A CT imaging method comprising the method for liver segmentation of an abdominal CT image according to any one of claims 1 to 4, characterized in that The following steps are also included: S7. Mark the liver segmentation result of the final target abdominal CT image obtained in step S6 on the original CT image, thereby outputting a CT image with the liver segmentation result.

Images