Dynamic optimization method and system for medical blood vessel segmentation based on anatomical prior perception

Abstract

The present invention discloses a dynamic optimization method and system for medical blood vessel segmentation based on anatomical prior perception, which belongs to the field of artificial intelligence and medical image processing technology. The method obtains medical image data and corresponding blood vessel annotation data, constructs a data set, and pre-processes and post-processes the data in the training set; constructs a segmentation model, including a multi-scale feature fusion module and a dynamic snake convolution module; constructs a loss function for the segmentation model, and trains the segmentation model using the training set and blood vessel annotation data; and uses the trained segmentation model to perform blood vessel segmentation on the medical image data in the test set. The present invention effectively solves the problem that existing methods are difficult to balance the fine segmentation of blood vessel structure and the overall topological integrity through the dynamically configured multi-scale feature fusion module and fusion mechanism and the dynamic optimization strategy of topology perception, and significantly improves the robustness and operation efficiency of the medical blood vessel segmentation model.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and medical image processing technology, and in particular to a dynamic optimization method and system for medical blood vessel segmentation based on anatomical prior perception. Background Art

[0002] With the rapid development of medical imaging technology, computer-assisted medical image segmentation has become an important technical tool in clinical diagnosis, lesion analysis, and surgical planning. Vascular segmentation, as a typical task in medical image segmentation, aims to accurately extract a continuous, complete, and topologically well-defined vascular pathmap from complex medical images. However, due to the complex and variable structure of human vascular structures, with significant variations in vascular morphology, spatial distribution, and diameter scale across different anatomical regions, traditional segmentation methods struggle to effectively address scale heterogeneity and topological complexity. This can lead to segmentation results with broken vessels, false connections, and missed small vessels, severely impacting the accuracy and efficiency of clinical diagnosis. Furthermore, currently widely used deep neural networks typically employ a uniform convolution kernel size and static weight configuration when processing vascular data. This ignores the spatial heterogeneity of vascular anatomical features, resulting in inadequate feature extraction and an inability to accurately capture high-level spatial semantic information about vessels, further impacting the model's generalization and segmentation performance. Furthermore, the weights in the loss function are typically statically fixed during model training, making it difficult to converge to a flatter local minimum.

[0003] Therefore, there is an urgent need on the market for a dynamic optimization method and system for medical vascular segmentation based on anatomical prior perception, so as to overcome the defects of existing technologies and models that cannot achieve optimal performance under existing computing resources, do not consider the spatial heterogeneity of vascular anatomical features, do not consider the use of global information to assist three-dimensional local feature learning, and do not consider the balance between multi-task objectives, so as to achieve more accurate and rapid brain vascular segmentation. Summary of the Invention

[0004] Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide a dynamic optimization method and system for medical blood vessel segmentation based on anatomical prior perception, so as to achieve more accurate and rapid cerebral blood vessel segmentation.

[0005] Technical solution: On one hand, the present invention provides a dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception, comprising the following steps:

[0006] Obtain medical imaging data and corresponding vascular annotation data, construct a dataset, divide the dataset into training set and test set in proportion, and perform pre-processing and post-processing on the data in the training set;

[0007] Build a segmentation model, including a multi-scale feature fusion module and a dynamic snake convolution module;

[0008] Construct the loss function of the segmentation model and train the segmentation model using the training set and blood vessel annotation data;

[0009] The trained segmentation model is used to perform blood vessel segmentation on the medical image data in the test set.

[0010] Furthermore, post-processing refers to performing maximum intensity projection on the three-dimensional volume data image blocks obtained after pre-processing in the sagittal, coronal and transverse planes to obtain two-dimensional slice images, reconstructing the two-dimensional slice images to the same scale as the three-dimensional volume data image blocks through a broadcast operation, and obtaining three groups of image blocks. The three-dimensional volume data image blocks and the three groups of image blocks are used as input items and input into the segmentation model.

[0011] Furthermore, the segmentation model includes an encoder and a decoder. The encoder includes bottleneck layers and non-bottleneck layers. All non-bottleneck layers use multi-scale feature fusion modules, and the bottleneck layer uses a dynamic snake convolution module.

[0012] Furthermore, the working process of the multi-scale feature fusion module includes:

[0013] The input feature map of the segmentation model is denoted as ,in Indicates the batch size, represents the number of input feature channels, are the depth, height and width of the three-dimensional image respectively;

[0014] The feature map After being input into the multi-scale feature fusion module, it undergoes five layers of feature fusion with the same structure;

[0015] In each layer of feature fusion, the feature map is first average pooled to extract global features, and then the weights of the convolution branches of each scale are adaptively generated through the fully connected layer and the Softmax normalization layer. ,in , K represents K scales; then, the global features are parallelized through multiple sets of convolution operations with different convolution kernel sizes, and the convolution results of each branch are compared with their weights After multiplication, sum and fusion are performed to obtain the fused multi-scale feature map, which is expressed as ,in Indicates the number of output feature channels;

[0016] The fused multi-scale feature map is calculated by the following formula:

[0017] ,

[0018] in, Representation feature map With scale Convolution operation performed by the convolution kernel; satisfy

[0019] Furthermore, the working process of the dynamic snake convolution module includes:

[0020] The output feature map of the multi-scale feature fusion module in the last layer is denoted as , , Indicates the number of input feature channels;

[0021] Perform a 3D convolution operation on the feature map y, and the output channel is , where P represents the number of sampling points of the dynamic snake convolution, and then after normalization and nonlinear activation function tanh, the offset constrained in the range [-1,1] is obtained. The offset is divided into two groups, each group has a total of offset, Z, Y, X represent depth, height and width respectively; then based on the original integer grid, generate Equally spaced relative coordinates, equal-spaced relative coordinates are expressed as follows: If there are P points to be offset, then within the integer range from arrive Generate a sequence of length P, accumulate the offsets, and get the three-dimensional floating-point sampling coordinates ;

[0022] Perform bilinear interpolation calculation on the floating-point coordinates in three-dimensional space to obtain the deformation feature map, perform group normalization and ReLU activation on all channels, and obtain the output feature map of the dynamic snake convolution layer, which is expressed as ,in Indicates the number of output feature channels.

[0023] Furthermore, in the dynamic snake convolution module, the convolution kernel size is given as The convolution kernel K, the coordinate position of the convolution kernel center is recorded as , then the position of each element in the convolution kernel K is expressed as ,in is the horizontal distance to the center element, then the element coordinate calculation formula in the x direction is as follows:

[0024] ,

[0025] Where, Indicates an increase in the offset of m in the x direction. It represents the cumulative sum of all offsets from position i to position i+m in the y direction. Represents the cumulative sum of all offsets in the z direction from position i to position i+m.

[0026] Furthermore, the loss function is expressed as:

[0027] ,

[0028] in, Indicates the weight of the corresponding part of the loss function, ; Represents the segmentation loss function, which measures the degree of regional overlap between the predicted results and the true labels; It is a centerline loss function based on skeleton structure perception; Represents the cross entropy loss, which is used for classification probability prediction at the pixel and voxel levels.

[0029] Furthermore, the weight It is obtained by dynamic sampling, from the standard normal distribution with a wider distribution mid-sampling , and use the softmax function as the mapping function to sample the sequence Perform normalization;

[0030] Weight The calculation formula is:

[0031] .

[0032] Furthermore, the steps of preprocessing the data in the training set include:

[0033] Medical imaging data and vascular annotation data were resampled and Z-score normalized respectively;

[0034] The image patch size and batch size of the input segmentation model are automatically configured based on the median size of the resampled data and the memory size of the GPU. The configuration rules include:

[0035] Initialize the image block size based on the median voxel shape after resampling, resample all samples in the training set according to the target spacing, and then calculate the voxel shape of each sample after resampling. Use the median shape on each axis as the initial image block size. If the image block size cannot be segmented on a certain axis, the number of times n corresponding to the downsampling model is 2 n If the axis is divisible, fill it up to the divisible size.

[0036] Iteratively adjust the image patch size based on the GPU memory budget, estimate the memory consumption for a given image patch size and batch size based on the segmentation model network architecture, and if the estimated memory exceeds the device budget, select the axis with the largest deviation from the median shape and reduce 2 along that axis. n voxels and update the network downsampling times and feature map size accordingly.

[0037] On the other hand, the present invention provides a medical blood vessel segmentation dynamic optimization system based on anatomical prior perception, including a processor and a memory, wherein the memory stores a blood vessel segmentation dynamic configuration program, and the processor executes the steps of the medical blood vessel segmentation dynamic optimization method based on anatomical prior perception when running the blood vessel segmentation dynamic configuration program.

[0038] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0039] The present invention achieves better segmentation accuracy for the cerebral vascular segmentation model on existing hardware devices by optimizing the data output size and preprocessing steps, dynamically setting the network structured parameters, dynamically optimizing the loss function, and designing an automatic parameter setting framework for the adaptive graphics computing unit. At the same time, the segmentation model input combines the broadcast of maximum intensity projection results of different axes to provide global information assistance. The adaptive convolution kernel strategy and dynamic serpentine convolution module are introduced into the basic network architecture to integrate anatomical prior knowledge and accurately capture the heterogeneous characteristics of the spatial distribution of vascular structures. The higher segmentation accuracy, topological indicators and lower false positives make the present invention a reference option for vascular segmentation, and in the future it can also be embedded as a reference process in the current clinical auxiliary diagnosis process. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flowchart of the dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception;

[0041] Figure 2 This is a flowchart of a dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception;

[0042] Figure 3 Schematic diagram of the structure of the multi-scale feature fusion module;

[0043] Figure 4 Schematic diagram of the results of the present invention and other comparative methods;

[0044] Figure 5 A schematic diagram of the structure of an image processing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0046] Example 1

[0047] On the one hand, the present invention provides a dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception, combined with Figures 1 to 2 , the method comprises the following steps:

[0048] Step 1: Obtain medical imaging data and corresponding vascular annotation data, construct a dataset, divide the dataset into a training set and a test set in proportion, and perform pre-processing and post-processing on the data in the training set;

[0049] Obtain medical image data on vascular imaging and corresponding vascular annotation data, wherein the vascular annotation data refers to the annotation of the structure and morphological information of normal cerebral blood vessels, and use the vascular annotation data as labels. Preprocess the medical image data and labels to obtain a three-dimensional volume data image block.

[0050] Preprocessing of the training data includes skull removal, resampling, standardization, and automatic configuration of network structure parameters. Skull removal involves using FreeSufer software to remove the skull from the image, improving the data quality at the input level.

[0051] For anisotropic image data, cubic spline interpolation is used within the plane and nearest neighbor interpolation is used outside the plane. For isotropic image data, cubic spline interpolation is used for global resampling. For label data, it is first converted to one-hot encoding. For anisotropic data, linear interpolation is used within the plane and nearest neighbor interpolation is used outside the plane. For isotropic data, global linear interpolation is performed.

[0052] Image-by-image z-score normalization is used to normalize each image using its own mean and standard deviation. This method is very effective for possible variations in imaging intensity between different patients and helps to improve the adaptability of the model.

[0053] The image block size and batch size of the input network are automatically configured according to the median size of the resampled data and the memory size of the graphics processor. Parameters such as network depth, number of feature channels, convolution step size, and convolution layer depth in the network structure are also automatically configured.

[0054] Furthermore, the steps of data preprocessing include:

[0055] Medical imaging data and vascular annotation data were resampled and Z-score normalized respectively;

[0056] The image patch size and batch size of the input segmentation model are automatically configured based on the median size of the resampled data and the memory size of the GPU. The configuration rules include:

[0057] Initialize the image block size based on the median voxel shape after resampling, resample all samples in the training set according to the target spacing, and then calculate the voxel shape of each sample after resampling. Use the median shape on each axis as the initial image block size. If the image block size cannot be segmented on a certain axis, the number of times n corresponding to the downsampling model is 2 n If the axis is divisible, fill it up to the divisible size.

[0058] Iteratively adjust the image patch size based on the GPU memory budget, estimate the memory consumption for a given image patch size and batch size based on the segmentation model network architecture, and if the estimated memory exceeds the device budget, select the axis with the largest deviation from the median shape and reduce 2 along that axis. n voxels and update the network downsampling times and feature map size accordingly.

[0059] Among them, the medical image data and vascular annotation data were resampled using cubic spline interpolation and linear interpolation respectively. The general form of the three-dimensional cubic spline interpolation function is:

[0060] ,

[0061] Where, are coefficients determined by known data points and their derivative conditions, They are the polynomial terms in each direction of the cubic spline interpolation polynomial, where i, j, and k represent the order of the term in the x, y, and z directions.

[0062] Based on the linear combination of the eight adjacent points where the voxel is located, three-dimensional linear interpolation estimates the intermediate value by connecting straight lines between adjacent data points in three directions. Assume that the eight vertex values ​​are , and their corresponding coordinates are: , where each vertex coordinate corresponds to a vertex position of the interpolation unit.

[0063] The interpolation scale factors in the three directions are defined as , the calculation formulas are:

[0064] ,

[0065] Based on the above scale factors, three-dimensional linear interpolation can be expressed as:

[0066] ,

[0067] The calculation process of Z-score standardization is expressed as:

[0068] ,

[0069] in, is the original data, is the mean of the original data, is the standard deviation of the original data.

[0070] Furthermore, post-processing refers to performing maximum intensity projection on the three-dimensional volume data image blocks obtained after pre-processing in the sagittal, coronal and transverse planes to obtain two-dimensional slice images, reconstructing the two-dimensional slice images to the same scale as the three-dimensional volume data image blocks through a broadcast operation, and obtaining three groups of image blocks. The three-dimensional volume data image blocks and the three groups of image blocks are used as input items and input into the segmentation model.

[0071] The maximum intensity projection can be expressed as:

[0072] ,

[0073] Where, The three-dimensional volume data is in voxels The intensity value at are the pixel coordinates on the projection screen, is The intensity value at .

[0074] In this example, maximum intensity projection (MIP) images and three-dimensional volume data are combined for supervised training to integrate global topological information, strengthen the modeling capabilities of long-range dependencies, and further improve the integrity of vascular topology, thereby better meeting the actual needs of clinical medical applications.

[0075] Step 2: Build a segmentation model, including a multi-scale feature fusion module and a dynamic snake convolution module.

[0076] Furthermore, the segmentation model includes an encoder and a decoder. The encoder includes a bottleneck layer and multiple non-bottleneck layers. All non-bottleneck layers use a multi-scale feature fusion module. The bottleneck layer uses a dynamic snake convolution module. All non-bottleneck layers are connected in sequence, and the last non-bottleneck layer is connected to the bottleneck layer. The decoder is UNet.

[0077] In non-bottleneck layers, convolution kernels of varying sizes are selected based on anatomical prior knowledge to extract and fuse features, capturing more complete contextual structural information. Multi-scale features enhance the model's expressiveness and aid its understanding of complex tubular features. By introducing a serpentine convolution module into the semantically abstract bottleneck layer, the model dynamically perceives the spatial distribution heterogeneity of blood vessels, effectively capturing the overall characteristics of vascular structure in a high-level abstract space. A deformable convolution sampling grid is used to effectively perceive irregular tubular features.

[0078] Combine Figure 3 ,Furthermore, the working process of the multi-scale feature fusion module includes:

[0079] The input feature map of the segmentation model is denoted as ,in Indicates the batch size, represents the number of input feature channels, are the depth, height and width of the three-dimensional image respectively;

[0080] The feature map After being input into the multi-scale feature fusion module, it undergoes five layers of feature fusion with the same structure;

[0081] In each layer of feature fusion, the feature map is first average pooled to extract global features, and then the weights of the convolution branches of each scale are adaptively generated through the fully connected layer and the Softmax normalization layer. ,in , K represents K scales, The calculation formula is as follows:

[0082] ,

[0083] in, is the feature vector output by the average pooling layer, and It is The fully connected layer weights and biases corresponding to the branches, t represents the total number of branches, is the adaptive weight of the k-th branch obtained by Softmax normalization. At the same time, That is, the sum of the weights of all branches is 1;

[0084] Subsequently, the global features are parallelized through multiple sets of convolution operations with different convolution kernel sizes, and the convolution results of each branch are compared with their weights. After multiplication, sum and fusion are performed to obtain the fused multi-scale feature map, which is expressed as ,in Indicates the number of output feature channels;

[0085] The fused multi-scale feature map is calculated by the following formula:

[0086] ,

[0087] in, Representation feature map With scale Convolution operation performed by the convolution kernel; satisfy .

[0088] Furthermore, the working process of the dynamic snake convolution module includes:

[0089] The input feature map of the multi-scale feature fusion module is denoted as , , Indicates the number of input feature channels;

[0090] Perform a 3D convolution operation on the feature map y, and the output channel is , where P represents the number of sampling points of the dynamic snake convolution, and then after normalization and nonlinear activation function tanh, the offset constrained in the range [-1,1] is obtained. The offset is divided into two groups, each group has a total of offset, Z, Y, X represent depth, height and width respectively; then based on the original integer grid, generate Equally spaced relative coordinates, equal-spaced relative coordinates are expressed as follows: If there are P points to be offset, then within the integer range from arrive Generate a sequence of length P, accumulate the offsets, and get the three-dimensional floating-point sampling coordinates ;

[0091] Perform bilinear interpolation calculation on the floating-point coordinates in three-dimensional space to obtain the deformation feature map, perform group normalization and ReLU activation on all channels, and obtain the output feature map of the dynamic snake convolution layer, which is expressed as ,in Indicates the number of output feature channels.

[0092] Furthermore, in the dynamic snake convolution module, the convolution kernel size is given as The convolution kernel K, the coordinate position of the convolution kernel center is recorded as , then the position of each element in the convolution kernel K is expressed as ,in is the horizontal distance to the center element, then the element coordinate calculation formula in the x direction is as follows:

[0093] ,

[0094] Where, Indicates an increase in the offset of m in the x direction. It represents the cumulative sum of all offsets from position i to position i+m in the y direction. Represents the cumulative sum of all offsets in the z direction from position i to position i+m.

[0095] Similarly, in the y direction, when the center coordinate position of the convolution kernel is , then the position of each element in the convolution kernel K is expressed as , the element coordinate calculation formula in the y direction is as follows:

[0096] ,

[0097] In the z direction, when the center coordinate position of the convolution kernel is , then the position of each element in the convolution kernel K is expressed as , then the element coordinate calculation formula in the z direction is as follows:

[0098] ,

[0099] Through the above formula, the discrete coordinate values ​​on the three coordinate axes can be obtained.

[0100] Step 3: Construct the loss function of the segmentation model and train the segmentation model using the training set and vascular annotation data.

[0101] Furthermore, the loss function is expressed as:

[0102] ,

[0103] in, Indicates the weight of the corresponding part of the loss function, ; Represents the segmentation loss function, which measures the degree of regional overlap between the predicted results and the true labels; It is a centerline loss function based on skeleton structure perception; Represents the cross entropy loss, which is used for classification probability prediction at the pixel and voxel levels.

[0104] Among them, the cross entropy loss function expression is:

[0105] ,

[0106] Where, and Denote the predicted probability and the true label respectively, and N denotes the total number of all voxel points in the image.

[0107] The expression of the segmentation loss function is:

[0108] .

[0109] The calculation method of the centerline loss function is:

[0110] The predicted binary segmentation results and the true labels are converted into corresponding skeleton representations through the morphological skeleton algorithm. Skeletonization is a result of thinning, retaining only the center line of the object. Topological accuracy and topology sensitivity The calculation methods are as follows:

[0111] ,

[0112] ,

[0113] Where, and Represent the true segmentation mask and the predicted segmentation result respectively, and Represent the skeleton of the true segmentation mask and the predicted segmentation result, respectively. is the smoothing factor, set to 10 -4 ;

[0114] Combining topological accuracy and topological sensitivity, The calculation formula is:

[0115] .

[0116] Furthermore, the weight It is obtained by dynamic sampling, from the standard normal distribution with a wider distribution mid-sampling , and use the softmax function as the mapping function to sample the sequence Perform normalization;

[0117] Weight The calculation formula is:

[0118] .

[0119] In this example, the loss function uses a dynamic weight strategy to automatically configure the weights of the segmentation loss and the topology loss, making it easier for the segmentation model to escape from local minima during training and converge to a flatter local minimum.

[0120] Step 4: Use the trained segmentation model to perform blood vessel segmentation on the medical imaging data in the test set.

[0121] The present invention introduces a dynamic feature perception mechanism in each layer of the encoding and decoding part of the segmentation model, including a multi-scale feature fusion module and a dynamic snake convolution module. Among them, the multi-scale feature fusion module first extracts features separately by using convolution kernels of different scales in parallel, and then uses the fully connected layer to dynamically generate fusion weights corresponding to the convolution kernels of each scale, thereby realizing the adaptive fusion of multi-scale features. This fusion mechanism can simultaneously give play to the advantages of small-scale convolution kernels that are good at capturing local details and edge information and large-scale convolution kernels that can effectively obtain broader contextual information, thereby obtaining a more comprehensive and rich feature expression; in practical applications, the maximum scale of the multi-scale convolution kernel needs to be reasonably selected according to the spatial resolution of the medical image and the diameter range of the target blood vessel. The dynamic snake convolution module effectively perceives the spatial heterogeneity of the vascular structure by dynamically adjusting the geometric shape and spatial distribution of the convolution kernel, further enhancing the segmentation model's ability to capture complex vascular features.

[0122] In addition, the present invention also combines maximum intensity projection (MIP) and uses the original 3D volume data and the broadcast results of its 2D MIP projections generated along different axes to jointly participate in model training, assisting in providing global topological information and further improving the model's overall perception of vascular structure.

[0123] The effects of the technical solution of the present invention are further verified and explained below in conjunction with specific experiments.

[0124] To verify the accuracy and superiority of the method presented in this paper, we conducted field experiments using TOF-MRA imaging data collected by the Biomedical Image Analysis Group at Imperial College London from three hospitals in London, UK. Data from two hospitals with the same resolution and voxel spacing were selected as the experimental dataset. A comparison of the experimental data is shown in Table 1 below.

[0125] Table 1

[0126]

[0127] The segmentation prediction results of the model of the present invention and the sample segmentation results of the classic segmentation methods DSCNet (dynamic snake convolution method based on topological geometric constraints for tubular structure segmentation), nnUNet (adaptive segmentation network) and UNETR (UNet method based on Transformers structure) are shown in the figure. Figure 4The figure shows the correlation results of two samples in the test set. The leftmost column shows the original input TOF-MRA image with an image resolution of 384×384×384 and a voxel spacing of (1.0, 1.0, 1.0). The middle four columns show the blood vessel segmentation prediction results obtained by SnakeConv, nnUNet, UNETR, and the method of the present invention, respectively, with red representing blood vessels and black representing background. The rightmost column shows the true value labels corresponding to the two samples, which were manually annotated by professionals. Figure 4 As can be seen in (a), the blood vessel mask obtained by the segmentation method of the present invention is more accurate in the magnified area and maintains a high degree of continuity. Figure 4 As can be seen in (b), the comparison method has different degrees of breakage and false negatives in the magnified area. From an overall observation, there are more false positive predictions than the segmentation results of the method of the present invention. Therefore, the results of the method of the present invention are closer to the corresponding true value labels.

[0128] From a holistic perspective, the blood vessels segmented using the proposed model exhibit relatively complete topological structures. Furthermore, in small, curved areas, this method achieves higher segmentation accuracy and connectivity than the other three methods. This is due to the feature perception capabilities of the dynamic snake convolution and multi-scale feature fusion modules. Furthermore, compared to SnakeConv, nnUNet, and UNETR, this method exhibits lower false positives and higher specificity, which has a positive impact on the use of automated segmentation models in assisting clinical diagnosis.

[0129] Example 2

[0130] The medical blood vessel segmentation dynamic optimization system based on anatomical prior perception described in this embodiment includes a processor and a memory, wherein the memory stores a dynamic configuration program for blood vessel segmentation, and the processor executes the steps of the medical blood vessel segmentation dynamic optimization method based on anatomical prior perception when running the dynamic configuration program for blood vessel segmentation.

[0131] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

[0132] Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0133] At this point, the dynamic optimization method and system for medical blood vessel segmentation based on anatomical prior perception of the present invention are completed.

Claims

1. A dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception, characterized by: The following steps are involved: Obtain medical imaging data and corresponding vascular annotation data, construct a dataset, divide the dataset into training set and test set in proportion, and perform pre-processing and post-processing on the data in the training set; Build a segmentation model, including a multi-scale feature fusion module and a dynamic snake convolution module; Construct the loss function of the segmentation model and train the segmentation model using the training set and blood vessel annotation data; Use the trained segmentation model to perform blood vessel segmentation on the medical imaging data in the test set; Post-processing involves performing maximum intensity projection on the 3D volume data blocks obtained after pre-processing in the sagittal, coronal, and transverse planes to obtain 2D slice images. The 2D slice images are then reconstructed to the same scale as the 3D volume data blocks through a broadcast operation to obtain three groups of image blocks. The 3D volume data blocks and the three groups of image blocks are then fed into the segmentation model. The working process of the multi-scale feature fusion module includes: The input feature map of the segmentation model is denoted as ,in Indicates the batch size, represents the number of input feature channels, are the depth, height and width of the three-dimensional image respectively; The feature map After being input into the multi-scale feature fusion module, it undergoes five layers of feature fusion with the same structure; In each layer of feature fusion, the feature map is first average pooled to extract global features, and then the weights of the convolution branches of each scale are adaptively generated through the fully connected layer and the Softmax normalization layer. ,in , K represents K scales; then, the global features are parallelized through multiple sets of convolution operations with different convolution kernel sizes, and the convolution results of each branch are compared with their weights After multiplication, sum and fusion are performed to obtain the fused multi-scale feature map, which is expressed as ,in Indicates the number of output feature channels; The fused multi-scale feature map is calculated by the following formula: ， in, Representation feature map With scale Convolution operation performed by the convolution kernel; satisfy ; The working process of the dynamic snake convolution module includes: The output feature map of the multi-scale feature fusion module in the last layer is denoted as , , Indicates the number of input feature channels; Perform a 3D convolution operation on the feature map y, and the output channel is , where P represents the number of sampling points of the dynamic snake convolution, and then after normalization and nonlinear activation function tanh, the offset constrained in the range [-1,1] is obtained. The offset is divided into two groups, each group has a total of offset, where Z, Y, and X represent depth, height, and width respectively; then, based on the original integer grid, generate Equally spaced relative coordinates, equal-spaced relative coordinates are expressed as follows: If there are P points to be offset, then within the integer range from arrive Generate a sequence of length P, accumulate the offsets, and get the three-dimensional floating-point sampling coordinates ; Perform bilinear interpolation calculation on the floating-point coordinates in three-dimensional space to obtain the deformation feature map, perform group normalization and ReLU activation on all channels, and obtain the output feature map of the dynamic snake convolution layer, which is expressed as ,in Indicates the number of output feature channels.

2. The dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception according to claim 1, characterized in that: The segmentation model includes an encoder and a decoder. The encoder includes bottleneck layers and non-bottleneck layers. All non-bottleneck layers use multi-scale feature fusion modules, and the bottleneck layer uses a dynamic snake convolution module.

3. The dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception according to claim 2, characterized in that: In the dynamic snake convolution module, the convolution kernel size is given as The convolution kernel K, the coordinate position of the convolution kernel center is recorded as , then the position of each element in the convolution kernel K is expressed as ,in is the horizontal distance to the center element, then the element coordinate calculation formula in the x direction is as follows: ， Where, Indicates an increase in the offset of m in the x direction. It represents the cumulative sum of all offsets from position i to position i+m in the y direction. Represents the cumulative sum of all offsets in the z direction from position i to position i+m.

4. The dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception according to claim 1, characterized in that: The expression of the loss function is: ， in, Indicates the weight of the corresponding part of the loss function, ; Represents the segmentation loss function, which measures the degree of regional overlap between the predicted results and the true labels; It is a centerline loss function based on skeleton structure perception; Represents the cross entropy loss, which is used for classification probability prediction at the pixel and voxel levels.

5. The dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception according to claim 4, characterized in that: Weight It is obtained by dynamic sampling, from the standard normal distribution with a wider distribution mid-sampling , and use the softmax function as the mapping function to sample the sequence Perform normalization; Weight The calculation formula is: 。 6. The dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception according to claim 1, characterized in that: The steps for preprocessing the data in the training set include: Medical imaging data and vascular annotation data were resampled and Z-score normalized respectively; The image patch size and batch size of the input segmentation model are automatically configured based on the median size of the resampled data and the memory size of the GPU. The configuration rules include: Initialize the image block size based on the median voxel shape after resampling, resample all samples in the training set according to the target spacing, and then calculate the voxel shape of each sample after resampling. Use the median shape on each axis as the initial image block size. If the image block size cannot be segmented on a certain axis, the number of times n corresponding to the downsampling model is 2 n If the axis is divisible, fill it up to the divisible size. Iteratively adjust the image patch size based on the GPU memory budget, estimate the memory consumption for a given image patch size and batch size based on the segmentation model network architecture, and if the estimated memory exceeds the device budget, select the axis with the largest deviation from the median shape and reduce 2 along that axis. n voxels and update the network downsampling times and feature map size accordingly.

7. A dynamic optimization system for medical vascular segmentation based on anatomical prior perception, characterized by: The method comprises a processor and a memory, wherein the memory stores a dynamic configuration program for blood vessel segmentation, and the processor executes the steps of the dynamic optimization method for medical blood vessel segmentation based on anatomical prior perception as claimed in any one of claims 1 to 6 when running the dynamic configuration program for blood vessel segmentation.

Images