Liver blood vessel segmentation method and system based on blood vessel diameter weight space optimization

By introducing a Dice loss function optimized by vascular diameter weight space in liver vascular segmentation, combined with sliding window method and morphological algorithm, the problem of poor segmentation effect in the prior art is solved, and higher segmentation accuracy is achieved .

CN118968059APending Publication Date: 2024-11-15GENERAL HOSPITAL OF PLA

Patent Information

Application Number
CN202410988796.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

In prior art In liver vascular segmentation, the Dice loss function only takes into account the overlapping information, resulting in poor segmentation effect in the thinner vascular area.

Method used

The Dice loss function based on the space optimization of vascular diameter weight is used, and the sliding window method and neural network model are combined with trilinear interpolation and morphological algorithm to calculate the vascular radius weight value to improve the prediction accuracy of the model for thinner vascular areas.

Benefits of technology

The segmentation accuracy in the finer vascular region is significantly improved, and performs better in this region than the traditional Dice loss function model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968059A_ABST
    Figure CN118968059A_ABST
Patent Text Reader

Abstract

The invention provides a liver blood vessel segmentation method and system based on blood vessel diameter weight space optimization. The method comprises the following steps: acquiring an epigastric CT image; preprocessing the upper abdomen CT image, and dividing the preprocessed upper abdomen CT image into a training set and a test set; constructing a liver blood vessel segmentation model based on the neural network model; training the liver blood vessel segmentation model by using the training set and based on a Dice loss function of a weight space; and inputting the test set into the trained liver blood vessel segmentation model, and obtaining a prediction result of the liver blood vessel segmentation model through a sliding window method. The method has the advantage that segmentation in a thin blood vessel area is more accurate in the liver blood vessel segmentation process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a liver blood vessel segmentation method and system based on spatial optimization of blood vessel diameter weights. Background Art

[0002] In clinical practice, a full understanding of the structural and morphological information of hepatic vessels is essential for accurate liver disease diagnosis and surgical planning. Traditional clinical methods involve experts manually annotating hepatic vessels on each slice of CT images, which consumes a lot of time and labor costs. In addition, manual annotation is limited by the experience and skills of experts and has a certain degree of subjectivity.

[0003] Therefore, there is an urgent need for an automatic and highly accurate liver vascular segmentation method to improve work efficiency and diagnostic accuracy. In the prior art, the global information supervision module GIS is introduced to improve the vascular segmentation around the edge of the liver. However, most of the existing deep learning-based liver vascular segmentation models rely on the Dice loss function as the main optimization loss function. However, the Dice loss function only considers the overlap information of the prediction and annotation results. Although a high similarity can be obtained, poor prediction often occurs in the thinner vascular areas of the liver vessels. Summary of the invention

[0004] The present invention provides a liver blood vessel segmentation method and system based on spatial optimization of blood vessel diameter weights, so as to solve the problem that the segmentation effect is poor in the thinner blood vessel area of ​​the liver blood vessel.

[0005] In a first aspect, the present invention provides a method for liver blood vessel segmentation based on spatial optimization of blood vessel diameter weights, the method comprising the following steps:

[0006] Acquire CT images of the upper abdomen;

[0007] Preprocessing the upper abdomen CT images, and dividing the preprocessed upper abdomen CT images into a training set and a test set;

[0008] Construct a liver vascular segmentation model based on a neural network model;

[0009] Using the training set and based on the Dice loss function in the weight space, training the liver blood vessel segmentation model;

[0010] The test set is input into the trained liver vascular segmentation model, and the prediction result of the liver vascular segmentation model is obtained by a sliding window method.

[0011] Optionally, the preprocessing of the upper abdomen CT image comprises the following steps:

[0012] Counting the voxel spacing of all the upper abdominal CT images;

[0013] The median of all the voxel spacings is taken as the standard voxel spacing;

[0014] Performing a trilinear interpolation operation on the upper abdomen CT image based on the standard voxel spacing, and performing an interpolation operation on the binary mask annotation using a nearest neighbor algorithm;

[0015] The upper abdomen CT image is normalized.

[0016] Optionally, the training of the liver blood vessel segmentation model using the training set and based on the Dice loss function in the weight space comprises the following steps:

[0017] Obtaining a binary mask corresponding to the training set;

[0018] Inputting the training set into the liver vascular segmentation model, and obtaining an initial prediction probability map through forward propagation calculation of the liver vascular segmentation model;

[0019] Calculating the data intersection and data union between the binary mask and the initial prediction probability map;

[0020] Calculating a radius weight value of a liver blood vessel based on the binary mask;

[0021] Based on the Dice loss function in the weight space, the loss function value is calculated by combining the data intersection, the data union and the radius weight value;

[0022] The loss function value is back-propagated in the liver vascular segmentation model, and the model parameters of the liver vascular segmentation model are updated through training cycle iterations.

[0023] Optionally, the step of calculating the radius weight value of the liver blood vessel based on the binary mask comprises the following steps:

[0024] Based on the training set and using a domain detection method, all contour voxel coordinates of the binary mask in three-dimensional space are calculated;

[0025] Calculating the blood vessel centerline of the liver blood vessel corresponding to the binary mask using a morphological algorithm;

[0026] Traversing all the contour voxel coordinates, and calculating the minimum Euclidean distance between each contour voxel coordinate and the centerline of the blood vessel;

[0027] Obtaining an initial weight mask by copying the binary mask;

[0028] Assigning values ​​to the voxels at the current coordinates of each of the initial weight masks based on the minimum Euclidean distance to obtain a basic weight mask;

[0029] Filling the basic weight mask based on the values ​​on the contour line using a three-dimensional convolution kernel to obtain a final weight mask;

[0030] The final weight mask is normalized to obtain the radius weight value of the liver blood vessel.

[0031] Optionally, the loss function value is calculated as follows:

[0032]

[0033] Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth represents the radius weight value, δ>0, δ represents the adjustment parameter.

[0034] In a second aspect, the present invention further provides a liver vessel segmentation system based on spatial optimization of vessel diameter weights, the system comprising:

[0035] An image acquisition module, used for acquiring upper abdominal CT images;

[0036] An image preprocessing module, used for preprocessing the upper abdomen CT image, and dividing the preprocessed upper abdomen CT image into a training set and a test set;

[0037] A model building module, used to build a liver vascular segmentation model based on a neural network model;

[0038] A model training module, used for training the liver blood vessel segmentation model using the training set and based on a Dice loss function in a weight space;

[0039] The model prediction module is used to input the test set into the trained liver vascular segmentation model, and obtain the prediction result of the liver vascular segmentation model through a sliding window method.

[0040] Optionally, the image preprocessing module includes:

[0041] A spacing statistics submodule, used for counting the voxel spacing of all the upper abdominal CT images;

[0042] A spacing standard submodule, used for taking the median of all the voxel spacings as the standard voxel spacing;

[0043] An interpolation submodule, configured to perform a trilinear interpolation operation on the upper abdomen CT image based on the standard voxel spacing, and to perform an interpolation operation on the binary mask annotation using a nearest neighbor algorithm;

[0044] The normalization submodule is used to normalize the upper abdomen CT image.

[0045] Optionally, the model training module includes:

[0046] A mask acquisition submodule, used to acquire a binary mask corresponding to the training set;

[0047] An initial prediction submodule, used for inputting the training set into the liver vascular segmentation model, and obtaining an initial prediction probability map through forward propagation calculation of the liver vascular segmentation model;

[0048] A data calculation submodule, used for calculating the data intersection and data union between the binary mask and the initial prediction probability map;

[0049] A weight calculation submodule, used for calculating the radius weight value of the liver blood vessels based on the binary mask;

[0050] A loss function calculation submodule, which is used for calculating the loss function value based on the Dice loss function in the weight space and combining the data intersection, the data union and the radius weight value;

[0051] The model training submodule is used to back-propagate the loss function value in the liver vascular segmentation model, and iteratively update the model parameters of the liver vascular segmentation model through a training cycle.

[0052] Optionally, the weight calculation submodule includes:

[0053] A coordinate calculation unit, used for calculating the coordinates of all contour voxels of the binary mask in the three-dimensional space based on the training set and using a field detection method;

[0054] A centerline determination unit, configured to calculate the blood vessel centerline of the liver blood vessel corresponding to the binary mask using a morphological algorithm;

[0055] A distance calculation unit, used for traversing all the contour voxel coordinates and calculating the minimum Euclidean distance between each contour voxel coordinate and the centerline of the blood vessel;

[0056] A mask copying unit, used for obtaining an initial weight mask by copying the binary mask;

[0057] A first weight calculation unit assigns values ​​to the voxels of the current coordinates of each of the initial weight masks based on the minimum Euclidean distance to obtain a basic weight mask;

[0058] A second weight calculation unit, using a three-dimensional convolution kernel and filling the basic weight mask based on the value on the contour line to obtain a final weight mask;

[0059] The weight normalization unit normalizes the final weight mask to obtain the radius weight value of the liver blood vessel.

[0060] Optionally, the loss function value is calculated as follows:

[0061]

[0062] Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth represents the radius weight value, δ>0, δ represents the adjustment parameter.

[0063] The beneficial effects of the present invention are:

[0064] The Dice loss function based on the blood vessel diameter weight space of the technical solution of the present invention is different from the conventional Dice loss function used in the prior art. When the model is in forward propagation, the penalty for errors in the thinner blood vessel area is increased. Thus, the model is constrained to focus on the prediction of the thinner blood vessel area. Ultimately, the model after training convergence will be more accurate in the segmentation of the thinner blood vessel area than the Dice loss function used in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Schematic diagram of the process of liver vessel segmentation method based on spatial optimization of vessel diameter weights in one embodiment of the present application. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0067] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0068] Figure 1 FIG. 1 is a flow chart of a liver vessel segmentation method based on spatial optimization of vessel diameter weights in one embodiment. Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the above method may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps. Figure 1 As shown, the liver blood vessel segmentation method based on blood vessel diameter weight space optimization disclosed in the present invention specifically includes the following steps:

[0069] S101. Acquire an upper abdominal CT image.

[0070] First, a CT scanner is used to scan the patient's upper abdomen to obtain a CT image of the upper abdomen. Usually, the patient needs to lie on a scanning bed, and then the CT scanner scans the patient's upper abdomen through X-rays to generate a series of image slices.

[0071] S102. Preprocess the upper abdominal CT images, and divide the preprocessed upper abdominal CT images into a training set and a test set.

[0072] In the preprocessing step, a series of operations need to be performed on the upper abdominal CT images to prepare the data for training and testing. These operations may include noise removal, image smoothing, grayscale normalization, etc. Then, the preprocessed images are divided into a training set and a test set. Usually, the data set can be divided according to a certain ratio, for example, 80% of the images are used for training and 20% of the images are used for testing.

[0073] S103. Construct a liver vascular segmentation model based on a neural network model.

[0074] In this step, you need to select a suitable neural network model to build a liver vascular segmentation model. Common choices include U-Net, FCN, etc. These models usually consist of an encoder and a decoder. The encoder is used to extract image features and the decoder is used to generate segmentation results.

[0075] In this embodiment, the liver vascular segmentation model includes a feature encoder module and a feature decoder module. The feature encoder module is used to extract a high-dimensional feature map of the input upper abdominal CT image of the subject, including three encoding layers; each encoding layer processes features of different resolutions, and each encoding layer extracts feature maps through a double convolution module to capture features of different scales; in the double convolution module, each convolution operation is followed by a batch normalization layer and a rectified linear unit layer; the channels of the feature map are doubled after each encoder layer; the feature decoder unit is used to gradually restore the feature map extracted by the feature encoder unit to the resolution of the original input image; the feature decoder module includes three layers of decoders, each layer of the decoder includes a deconvolution and a double convolution module; a jump connection is introduced between the feature encoder and the feature decoder to combine features of different levels. At the end of the liver vascular segmentation model network, a convolution layer and a Sigmoid activation function are used to obtain the final segmentation result.

[0076] The dual convolution module consists of two convolutional layers, a batch normalization layer, and a rectified linear unit layer. The convolution kernel sizes of the two convolutional layers are both (3, 3, 3). The batch normalization layer normalizes the data according to the mean and variance calculated by channel in the current batch. The rectified linear unit layer increases the nonlinear fitting ability of the neural network.

[0077] S104. Train the liver vascular segmentation model using the training set and the Dice loss function based on the weight space.

[0078] In this step, the images of the training set and the corresponding labels (liver vascular segmentation results) are used to train the liver vascular segmentation model. Usually, the Dice loss function is used as the optimization objective to optimize the weight of the model by minimizing the Dice loss. The Dice loss function can measure the similarity between the model prediction results and the true labels.

[0079] S105. Input the test set into the trained liver vascular segmentation model, and obtain the prediction result of the liver vascular segmentation model through the sliding window method.

[0080] In this step, the images of the test set need to be input into the trained liver vascular segmentation model. Since the images may be large and cannot be input into the model at once, the sliding window method can be used to process them. The sliding window method divides the image into multiple small blocks and inputs each small block into the model for prediction. The trained liver vascular segmentation model is applied to the image blocks for prediction and inference to obtain the segmentation prediction results of each window. Finally, the prediction results are combined to obtain the liver vascular segmentation result of the entire image.

[0081] The implementation principle of this embodiment is:

[0082] By executing the above steps, the Dice loss function can be re-weighted using the information of the vessel diameter. Compared with the conventional Dice loss function, when the model is propagating forward, the penalty for errors in the thinner vessel area is increased. Thus, the model is constrained to focus on the prediction of the thinner vessel area. Finally, the model after training convergence will be more accurate than the conventional Dice loss function in the segmentation of the thinner vessel area.

[0083] In one embodiment, the preprocessing of the upper abdomen CT image in step S102 specifically includes the following steps:

[0084] The voxel spacing of all upper abdominal CT images was counted;

[0085] The median of all voxel spacings was taken as the standard voxel spacing;

[0086] The upper abdominal CT images were interpolated by trilinear method based on the standard voxel spacing, and the binary mask annotations were interpolated by the nearest neighbor algorithm.

[0087] The upper abdominal CT images were normalized.

[0088] In this embodiment, the upper abdominal CT image in NIfTI format (Neuroimaging Informatics Technology Initiative, nii) is read. Since there are differences in scanning layer thickness and spatial resolution between different data, different data need to be resampled in order to ensure the model effect. In order to ensure the effect of resampling, the voxel spacing of all data is first counted, and the median of the voxel spacing is taken as the unified voxel spacing after resampling. After obtaining a unified voxel spacing, trilinear interpolation operations are performed on all images. At the same time, the annotation mask corresponding to the image data uses the nearest neighbor algorithm to interpolate the annotation mask to ensure the consistency of the image and annotation information. The input CT image is truncated in the HU range, and the truncation range is [-200,300], and then the zero mean (Z-score) normalization method is used to normalize the image data.

[0089] In one embodiment, the training set can also be subjected to data enhancement processing, as follows: the preprocessed training set and its corresponding annotation results are randomly cropped to obtain a number of image blocks p of size (96, 96, 96); then, the image blocks of the training set and the annotation data image blocks have a 50% probability of rotating the images, and the rotation angle is set to +-15 degrees to obtain image_1 and the corresponding mask_1; then, image_1 and mask_1 have a 50% probability of performing a horizontal mirror flip operation to obtain image_2 and mask_2; then, image_2 and mask_2 have a 50% probability of performing an up-down mirror flip operation to obtain image_3 and mask_3. At this point, all spatial enhancement techniques have been applied. Next, image_3 and mask_3 will be randomly Gaussian blurred with a probability of 75%, where the sigma value is set to a random real number in the interval [0.1, 0.9], to obtain image_4 and mask_4; next, image_4 and mask_4 will be randomly noised with a probability of 75%, where kernel_size is set to 5 and sigma is set to 1. All the above probabilities can be adjusted accordingly according to the actual situation.

[0090] In one implementation manner, step S104 specifically includes the following steps:

[0091] Get the binary mask corresponding to the training set;

[0092] The training set is input into the liver vascular segmentation model, and the initial prediction probability map is obtained through forward propagation calculation of the liver vascular segmentation model;

[0093] Calculate the data intersection and data union between the binary mask and the initial prediction probability map;

[0094] The radius weight value of the liver blood vessels is calculated based on the binary mask;

[0095] Based on the Dice loss function in the weight space, the loss function value is calculated by combining the data intersection, data union and radius weight value;

[0096] The loss function value is back-propagated in the liver vascular segmentation model, and the model parameters of the liver vascular segmentation model are updated through training cycle iterations.

[0097] In this embodiment, an image block of size (1, 96, 96, 96) after preprocessing and data enhancement is input, and the constructed liver vascular segmentation model is forward propagated to obtain an initial prediction probability map. The initial prediction probability map and its corresponding cropped annotation result (i.e., binary mask) are used for loss value calculation, and then the Adam optimizer is used for back propagation. With the iteration of the training cycle, the model parameters are continuously updated to make the loss value smaller and smaller. By optimizing the model parameters, the prediction results of the model gradually approach the annotated results.

[0098] In this embodiment, the calculation formula of the loss function value is as follows:

[0099]

[0100] Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth Represents the radius weight value, δ>0, δ represents the adjustment parameter. δ is generally taken as a very small value to prevent the denominator from having a zero value and causing the calculation to fail. The data intersection is the product of the binary mask and the initial prediction probability map, and the data union is the sum of the binary mask and the initial prediction probability map.

[0101] In one embodiment, the step of calculating the radius weight value of the liver blood vessels based on the binary mask specifically includes the following steps:

[0102] Based on the training set and using the area detection method, all contour voxel coordinates of the binary mask in three-dimensional space are calculated;

[0103] The vascular centerline of the liver blood vessels corresponding to the binary mask is calculated using a morphological algorithm;

[0104] Traverse all contour voxel coordinates and calculate the minimum Euclidean distance between each contour voxel coordinate and the blood vessel centerline;

[0105] The initial weight mask is obtained by copying the binary mask;

[0106] Based on the minimum Euclidean distance, the voxels of the current coordinates of each initial weight mask are assigned values ​​to obtain the basic weight mask;

[0107] Use the 3D convolution kernel to fill the basic weight mask based on the values ​​on the contour line to obtain the final weight mask;

[0108] The final weight mask is normalized to obtain the radius weight value of the liver blood vessels.

[0109] In this embodiment, first, the binary mask in the training set is subjected to a domain detection to find all contour voxels. The domain detection method can traverse each voxel of the binary mask to check whether the eight neighboring voxels around it are background (pixel value is 0). If there is a background voxel, the current voxel is a contour voxel. A morphological corrosion operation is performed on the binary mask to remove noise and make the blood vessels more continuous. Then, a morphological expansion operation is performed on the corroded binary mask to restore the shape of the blood vessel. Finally, a skeleton extraction operation is performed on the expanded binary mask to obtain the center line of the blood vessel.

[0110] For each contour voxel coordinate, calculate its Euclidean distance from all points on the centerline of the vessel. Select the point with the smallest distance as the corresponding point between the contour voxel coordinate and the centerline of the vessel. Copy the binary mask as the initial weight mask, keeping the vascular area in the binary mask unchanged. For each voxel of the initial weight mask, assign a corresponding weight value according to its corresponding minimum Euclidean distance. Linear interpolation or other interpolation methods can be used to calculate the weight value. Use a three-dimensional convolution kernel to perform a convolution operation on the basic weight mask to extend the weight value to the area around the blood vessel. In the convolution operation, the weight value can be filled according to the value on the contour line so that the area around the blood vessel has a higher weight value. The final weight mask is normalized to map the weight value to the range of 0 to 1. Normalization can use linear transformation or other normalization methods.

[0111] The present invention also discloses a liver blood vessel segmentation system based on blood vessel diameter weight space optimization, the system comprising:

[0112] An image acquisition module, used for acquiring upper abdominal CT images;

[0113] An image preprocessing module, used for preprocessing upper abdominal CT images, and dividing the preprocessed upper abdominal CT images into a training set and a test set;

[0114] A model building module, used to build a liver vascular segmentation model based on a neural network model;

[0115] A model training module, used to train a liver vascular segmentation model using a training set and a Dice loss function based on a weight space;

[0116] The model prediction module is used to input the test set into the trained liver vascular segmentation model and obtain the prediction result of the liver vascular segmentation model through the sliding window method.

[0117] The implementation principle of this embodiment is:

[0118] Through the steps performed by the above modules, the information of the vessel diameter can be used to reweight the Dice loss function based on the Dice loss function. Compared with the conventional Dice loss function, when the model is forward propagating, the penalty for errors in the thinner vessel area is increased. Thus, the model is constrained to focus on the prediction of the thinner vessel area. Finally, the model after training convergence will be more accurate than the conventional Dice loss function in the segmentation of the thinner vessel area.

[0119] In one embodiment, the image preprocessing module includes:

[0120] The spacing statistics submodule is used to count the voxel spacing of all upper abdominal CT images;

[0121] A spacing standard submodule is used to take the median of all voxel spacings as the standard voxel spacing;

[0122] An interpolation submodule is used to perform trilinear interpolation operations on the upper abdominal CT images based on the standard voxel spacing, and to interpolate the binary mask annotations using the nearest neighbor algorithm;

[0123] The normalization submodule is used to normalize the upper abdominal CT image.

[0124] In one embodiment, the model training module includes:

[0125] The mask acquisition submodule is used to obtain the binary mask corresponding to the training set;

[0126] The initial prediction submodule is used to input the training set into the liver vascular segmentation model, and obtain the initial prediction probability map through the forward propagation calculation of the liver vascular segmentation model;

[0127] A data calculation submodule, used to calculate the data intersection and data union between the binary mask and the initial prediction probability map;

[0128] A weight calculation submodule, used for calculating the radius weight value of the liver blood vessels based on the binary mask;

[0129] The loss function calculation submodule is used for the Dice loss function based on the weight space, and the loss function value is calculated by combining the data intersection, data union and radius weight value;

[0130] The model training submodule is used to back-propagate the loss function value in the liver vascular segmentation model and iteratively update the model parameters of the liver vascular segmentation model through the training cycle.

[0131] In this embodiment, the calculation formula of the loss function value is as follows:

[0132]

[0133] Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth Represents the radius weight value, δ>0, δ represents the adjustment parameter. δ is generally taken as a very small value to prevent the denominator from having a zero value and causing the calculation to fail. The data intersection is the product of the binary mask and the initial prediction probability map, and the data union is the sum of the binary mask and the initial prediction probability map.

[0134] In one embodiment, the weight calculation submodule includes:

[0135] A coordinate calculation unit, used for calculating the coordinates of all contour voxels of the binary mask in the three-dimensional space based on the training set and using a domain detection method;

[0136] A centerline determination unit, used for calculating the blood vessel centerline of the liver blood vessel corresponding to the binary mask by using a morphological algorithm;

[0137] A distance calculation unit, used for traversing all contour voxel coordinates and calculating the minimum Euclidean distance between each contour voxel coordinate and the center line of the blood vessel;

[0138] A mask copying unit, used for obtaining an initial weight mask by copying a binary mask;

[0139] A first weight calculation unit assigns values ​​to the voxels of the current coordinates of each initial weight mask based on the minimum Euclidean distance to obtain a basic weight mask;

[0140] A second weight calculation unit uses a three-dimensional convolution kernel and fills the basic weight mask based on the value on the contour line to obtain a final weight mask;

[0141] The weight normalization unit normalizes the final weight mask to obtain the radius weight value of the liver blood vessels.

[0142] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0143] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A liver vascular segmentation method based on vascular diameter weight space optimization, characterized in that: The steps include: Acquire CT images of the upper abdomen; Preprocessing the upper abdomen CT images, and dividing the preprocessed upper abdomen CT images into a training set and a test set; Construct a liver vascular segmentation model based on a neural network model; Using the training set and based on the Dice loss function in the weight space, training the liver blood vessel segmentation model; The test set is input into the trained liver vascular segmentation model, and the prediction result of the liver vascular segmentation model is obtained by a sliding window method.

2. The method for liver vessel segmentation based on spatial optimization of vessel diameter weights according to claim 1, characterized in that: The preprocessing of the upper abdomen CT image comprises the following steps: Counting the voxel spacing of all the upper abdominal CT images; The median of all the voxel spacings is taken as the standard voxel spacing; Performing a trilinear interpolation operation on the upper abdomen CT image based on the standard voxel spacing, and performing an interpolation operation on the binary mask annotation using a nearest neighbor algorithm; The upper abdomen CT image is normalized.

3. The method for liver vessel segmentation based on spatial optimization of vessel diameter weights according to claim 1, characterized in that: The training of the liver blood vessel segmentation model using the training set and based on the Dice loss function in the weight space comprises the following steps: Obtaining a binary mask corresponding to the training set; Inputting the training set into the liver vascular segmentation model, and obtaining an initial prediction probability map through forward propagation calculation of the liver vascular segmentation model; Calculating the data intersection and data union between the binary mask and the initial prediction probability map; Calculating a radius weight value of a liver blood vessel based on the binary mask; Based on the Dice loss function in the weight space, the loss function value is calculated by combining the data intersection, the data union and the radius weight value; The loss function value is back-propagated in the liver vascular segmentation model, and the model parameters of the liver vascular segmentation model are updated through training cycle iterations.

4. The method for liver vessel segmentation based on spatial optimization of vessel diameter weights according to claim 3, characterized in that: The step of calculating the radius weight value of the liver blood vessels based on the binary mask comprises the following steps: Based on the training set and using a domain detection method, all contour voxel coordinates of the binary mask in three-dimensional space are calculated; Calculating the blood vessel centerline of the liver blood vessel corresponding to the binary mask using a morphological algorithm; Traversing all the contour voxel coordinates, and calculating the minimum Euclidean distance between each contour voxel coordinate and the centerline of the blood vessel; Obtaining an initial weight mask by copying the binary mask; Assigning values ​​to the voxels at the current coordinates of each of the initial weight masks based on the minimum Euclidean distance to obtain a basic weight mask; Filling the basic weight mask based on the values ​​on the contour line using a three-dimensional convolution kernel to obtain a final weight mask; The final weight mask is normalized to obtain the radius weight value of the liver blood vessel.

5. The method for liver vessel segmentation based on spatial optimization of vessel diameter weights according to claim 4, characterized in that: The calculation formula of the loss function value is as follows: Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth Represents the radius weight value, δ>0, δ represents the adjustment parameter.

6. A liver vessel segmentation system based on spatial optimization of vessel diameter weights, characterized in that: The system comprises: An image acquisition module, used for acquiring upper abdominal CT images; An image preprocessing module, used for preprocessing the upper abdomen CT image, and dividing the preprocessed upper abdomen CT image into a training set and a test set; A model building module, used to build a liver vascular segmentation model based on a neural network model; A model training module, used for training the liver blood vessel segmentation model using the training set and based on a Dice loss function in a weight space; The model prediction module is used to input the test set into the trained liver vascular segmentation model, and obtain the prediction result of the liver vascular segmentation model through a sliding window method.

7. The liver vessel segmentation system based on vessel diameter weight space optimization according to claim 6, characterized in that: The image preprocessing module comprises: A spacing statistics submodule, used for counting the voxel spacing of all the upper abdominal CT images; A spacing standard submodule, used for taking the median of all the voxel spacings as the standard voxel spacing; An interpolation submodule, configured to perform a trilinear interpolation operation on the upper abdomen CT image based on the standard voxel spacing, and to perform an interpolation operation on the binary mask annotation using a nearest neighbor algorithm; The normalization submodule is used to normalize the upper abdomen CT image.

8. The liver vessel segmentation system based on vessel diameter weight space optimization according to claim 6, characterized in that: The model training module includes: A mask acquisition submodule, used to acquire a binary mask corresponding to the training set; An initial prediction submodule, used for inputting the training set into the liver vascular segmentation model, and obtaining an initial prediction probability map through forward propagation calculation of the liver vascular segmentation model; A data calculation submodule, used for calculating the data intersection and data union between the binary mask and the initial prediction probability map; A weight calculation submodule, used for calculating the radius weight value of the liver blood vessels based on the binary mask; A loss function calculation submodule, which is used for calculating the loss function value based on the Dice loss function in the weight space and combining the data intersection, the data union and the radius weight value; The model training submodule is used to back-propagate the loss function value in the liver vascular segmentation model, and iteratively update the model parameters of the liver vascular segmentation model through a training cycle.

9. The liver vessel segmentation system based on vessel diameter weight space optimization according to claim 8, characterized in that: The weight calculation submodule includes: A coordinate calculation unit, used for calculating the coordinates of all contour voxels of the binary mask in the three-dimensional space based on the training set and using a field detection method; A centerline determination unit, configured to calculate the blood vessel centerline of the liver blood vessel corresponding to the binary mask using a morphological algorithm; A distance calculation unit, used for traversing all the contour voxel coordinates and calculating the minimum Euclidean distance between each contour voxel coordinate and the centerline of the blood vessel; A mask copying unit, used for obtaining an initial weight mask by copying the binary mask; A first weight calculation unit assigns values ​​to the voxels of the current coordinates of each of the initial weight masks based on the minimum Euclidean distance to obtain a basic weight mask; A second weight calculation unit, using a three-dimensional convolution kernel and filling the basic weight mask based on the value on the contour line to obtain a final weight mask; The weight normalization unit normalizes the final weight mask to obtain the radius weight value of the liver blood vessel.

10. The liver vessel segmentation system based on vessel diameter weight space optimization according to claim 9, characterized in that: The calculation formula of the loss function value is as follows: Where: L represents the loss function value, P int represents the data intersection, P union represents the data union, W groundtruth represents the radius weight value, δ>0, δ represents the adjustment parameter.

Citation Information

Patent Citations

  • CT image liver artery segmentation method and system based on deep learning

    CN110992383A

  • Liver segmentation method of abdomen CT image, and CT imaging method thereof

    CN114066901A

Cited By

  • Vascular exposure prediction model construction method based on liver resection anatomy

    CN119581042A