Single-scanning DSA imaging method and system based on multi-view feature learning

Abstract

The invention discloses a single-scanning DSA imaging method and system based on multi-view feature learning. The single-scanning DSA imaging method comprises the following steps: S1, collecting multiple groups of digital subtraction angiography projection data and multiple groups of three-dimensional cerebrovascular images; s2, the real pairing training data and the simulation pairing data in the S1 are enhanced respectively, and then the enhanced real pairing training data and the enhanced simulation pairing data are merged to obtain a complete subtraction pairing data set; s3, training a learnable subtraction artifact correction model by using the complete subtraction pairing data set obtained in S2 to obtain a trained learnable subtraction artifact correction model; and S4, inputting the multi-view filling ring projection data into the trained learnable subtraction artifact correction model to obtain blood vessel projection data. According to the method, the blood vessel projection can be accurately recovered from the input projection, and the vast majority of subtraction artifacts are corrected.

Description

Technical Field

[0001] The present invention relates to the technical field of digital subtraction angiography, and in particular to a single-scan DSA imaging method and system based on multi-view feature learning. Background Art

[0002] Digital subtraction angiography (DSA) is a widely used imaging method for assessing vascular structure and function. Three-dimensional DSA is typically acquired using C-arm cone-beam computed tomography (CBCT). The conventional acquisition protocol for 3D cerebral angiography involves two scans from multiple angles. Clear cerebral vascular projection data are obtained by subtracting the projection data from the masked scan, which does not inject contrast agent, from the projection data from the contrast-enhanced filled scan. Ideally, the resulting subtraction image should contain only contrast-enhanced vascular structures. However, factors such as X-ray source power fluctuations, imaging chain noise, and patient swallowing, nodding, and respiratory movement between scans can cause misalignment between the subtracted projection data, resulting in blurring of the acquired vascular projection data. This blurring is known as subtraction artifact. Subtraction artifacts can complicate lesion location determination and cause unnecessary structures to appear in the 3D vascular reconstruction volume, reducing the success rate of subsequent diagnosis and treatment. Therefore, reducing or even completely eliminating subtraction artifacts in vascular projection data has been considered a necessary and challenging task.

[0003] Traditional methods for subtraction artifact correction focus on using exogenous techniques to minimize patient motion during exposure. For example, non-ionic contrast agents are used to suppress patient motion caused by the sudden thermal sensation during contrast injection; immobilization devices are used to stabilize the patient's head to reduce inter-scanning motion; and patients are asked to hold their breath during scanning to reduce respiratory motion. In recent years, with the development of deep learning technology and its demonstrated effectiveness in various tasks, this technology has been gradually applied to subtraction artifact correction. Deep learning-based methods for correcting subtraction artifacts can be roughly divided into two categories: deep learning registration methods and deep learning vascular projection data generation methods. Deep learning-based registration methods use a learnable model to automatically align two subtracted projection data in a supervised manner. Deep learning-based vascular projection data generation methods use a learnable model to learn the mapping between filled-circle scan projection data and artifact-free vascular projection data, thereby directly generating subtracted vascular projection data from filled-circle projection data.

[0004] Traditional methods for correcting subtraction artifacts focus on using exogenous techniques to minimize patient motion during exposure. While these methods can mitigate patient motion to a certain extent, they are limited when patients are unable to cooperate. Deep learning-based registration methods utilize a learnable model to automatically align the two subtracted projection data in a supervised manner. While these registration methods can effectively preserve vascular detail, they require additional mask scans of the projection data, and their effectiveness in suppressing artifacts needs improvement. Deep learning-based methods for generating vascular projection data have two major challenges: First, current methods primarily focus on predicting vascular projection data from a single viewpoint, and their ability to extract vascular projection data in complex anatomical regions requires improvement. Second, existing models primarily utilize supervised learning, using clinically observed vascular projection data with minimal artifacts as reference images to guide model learning. Because traditional subtraction methods cannot generate completely artifact-free vascular projection data as a reference, this approach inevitably introduces minor artifacts in the model output. Currently, no research has addressed both of these limitations simultaneously.

[0005] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a single-scan DSA imaging method and system based on multi-view feature learning to solve the shortcomings of the existing technology. Summary of the Invention

[0006] A first object of the present invention is to overcome the shortcomings of the prior art and provide a single-scan DSA imaging method based on multi-view feature learning. This single-scan DSA imaging method based on multi-view feature learning can accurately recover vascular projections from input projections while correcting most subtraction artifacts.

[0007] The above-mentioned purpose of the present invention is achieved through the following technical measures:

[0008] A single-scan DSA imaging method based on multi-view feature learning is provided, which is performed by the following steps:

[0009] S1. Acquire multiple sets of digital subtraction angiography projection data and multiple sets of three-dimensional cerebral vascular images, obtain real paired training data using the digital subtraction angiography projection data, and obtain simulated paired data using the three-dimensional cerebral vascular images; multiple pairs of the real paired training data constitute a real paired training data set, and multiple pairs of the simulated paired data constitute a subtraction paired data set;

[0010] S2, enhance the real paired training data and the simulated paired data in S1 respectively, and then merge the enhanced real paired training data and the enhanced simulated paired data to obtain a complete subtracted paired data set;

[0011] S3, training the learnable subtraction artifact correction model G using the complete subtraction paired dataset obtained in S2, to obtain the trained learnable subtraction artifact correction model G;

[0012] S4. After inputting the multi-view filled circle projection data into training, the subtraction artifact correction model G can be learned to obtain multi-view blood vessel projection data without subtraction artifacts.

[0013] Preferably, the method for obtaining the above-mentioned real paired training data is:

[0014] A1. Acquire the digital subtraction angiography projection data, wherein the digital subtraction angiography projection data includes mask circle projection data M. r and filled circle projection data F r ;

[0015] A2, the filling circle projection data F r and the mask circle projection data M r Subtraction is performed to obtain the blood vessel projection data V p ;

[0016] A3, make the filled circle projection data F r and the blood vessel projection data V p constitute the real paired training data.

[0017] Preferably, the method for obtaining the above simulation pairing data is:

[0018] B1. Acquire and reconstruct the 3D cerebral vascular image through 3D cerebral vascular imaging modality, and then segment it to obtain the 3D cerebral vascular image V s ;

[0019] B2. Three-dimensional cerebral vascular image V s Surface meshing and surface mesh preprocessing are performed, and then the three-dimensional cerebral vascular image V s Perform voxelization and morphological filling to obtain three-dimensional blood vessel data V with internal voxel values ​​of 1 x ;

[0020] B3. Select a set of mask circle projection data M r , and then according to the scanned geometric parameters and the corresponding three-dimensional blood vessel data V x , simulate cone beam CT projection data to obtain simulated vascular projection data without subtraction artifacts

[0021] B4, the simulated blood vessel projection data The corresponding mask circle projection data M rPerform angle-by-angle fusion. The fusion process includes scaling and translation operations on the vascular projection data to obtain the final simulated filling circle projection data.

[0022] B5, making the simulation filling circle projection data and the simulated blood vessel projection data Constructing the simulation pairing data;

[0023] Preferably, the learnable subtraction artifact correction model G includes a block coding stage, an encoder downsampling stage and a decoder upsampling stage.

[0024] Preferably, the processing method of the block coding stage is: first, the multi-view projection data with input dimensions (H, W, D) is processed with a resolution of (P h , P w , P d ) is flattened into blocks, resulting in multiple where H and W are the dimensions of the projection data and D is the number of input views. The blocks are then mapped into a K-dimensional feature space through a convolutional layer. Finally, a K-dimensional learnable position code is added to each block to obtain the encoded block.

[0025] Preferably, the encoder downsampling stage includes three downsampling modules, and each downsampling module is composed of a dynamic snake convolution module, a downsampling convolution module and an efficient paired attention module connected in sequence from front to back.

[0026] Preferably, the above-mentioned encoder upsampling stage includes three upsampling modules, and each upsampling module is composed of an efficient paired attention module, a dynamic snake convolution module and an upsampling convolution module connected in sequence from front to back.

[0027] Preferably, the encoder upsampling stage outputs the subtraction artifact corrected vascular projection data V with dimensions (H, W, D) g .

[0028] Preferably, the upsampling convolution module uses a transposed convolution operation with a stride of 2.

[0029] Preferably, the downsampling convolution module uses a convolution operation with a stride of 2;

[0030] Preferably, the dynamic snake convolution module calculates the position of a 7x7 convolution kernel in each dimensional feature map in the x-axis and y-axis directions by bilinear interpolation.

[0031] Preferably, the above-mentioned efficient paired attention module is provided with a spatial attention calculation unit, a channel attention calculation unit and an attention feature map fusion unit.

[0032] Preferably, the spatial attention calculation unit and the channel attention calculation unit both calculate the query Q, key K and value V, wherein the query Q and the key K are shared in the efficient pairwise attention module; the query Q is calculated as W Q X, the key K is calculated as W K X; the calculation method of the value V in the spatial attention calculation unit is The calculation method of the median V of the channel attention calculation unit is: X is the feature map input to the efficient pairwise attention module.

[0033] Preferably, the spatial attention A in the above-mentioned spatial attention calculation unit s The calculation formula is expressed by formula (1):

[0034]

[0035] Among them, Softmax(·) is the activation function, proj(·) is the low-dimensional mapping operation, T is the transpose operation, and d is the dimension of the vector in the numerator.

[0036] Preferably, the channel attention A in the above-mentioned channel attention calculation unit c The calculation formula is expressed by formula (2):

[0037]

[0038] Preferably, the attention feature map fusion unit is fused in an element-by-element addition manner, and the fusion of the feature map A is completed through two convolution operations. The calculation formula of the feature map A is expressed by formula (3):

[0039] A=Conv1(Conv3(A s +A c ))……Formula (3).

[0040] Among them, Conv1 is a convolution operation with a convolution kernel size of 1×1, and Conv3 is a convolution operation with a convolution kernel size of 3×3.

[0041] In S3 , the learnable subtraction artifact correction model G is trained using stochastic gradient descent and constrained by a total loss function L .

[0042] Preferably, the total loss function L is composed of the L1 loss function L1, the perceptual loss function L per and a topology-preserving loss function.

[0043] Preferably, the topology preservation loss function converts the blood vessel projection data V g And the corresponding reference vessel projection Vt , the reference blood vessel projection V t is the blood vessel projection data V p and simulated vascular projection data Set of, calculate the corresponding two soft binary outputs B g and B t , and the two soft skeleton outputs S of the soft skeleton image S g and S t , the topology preservation loss function includes the binarization loss L bin , skeletonization loss L ske and centerline dice loss L cldice .

[0044] Preferably, the total loss function L is expressed by formula (4):

[0045] L=λ1L1+λ2L per +λ3L bin +λ4L ske +λ5L cldice ...Formula (4);

[0046] Among them, λ1, λ2, λ3, λ4 and λ5 are empirical coefficients, and λ1 and λ2 are both 1, λ3 and λ4 are both 0.8, and λ5 is 0.6.

[0047] Preferably, the above L1 loss function L1 generates the vascular projection data V by constraining g Compared with the reference vessel projection V t The pixel-by-pixel error, the L1 loss function L1 is expressed by formula (5):

[0048] L1=||V g -V t ||1……Formula (5).

[0049] Preferably, the above-mentioned perceptual loss function L per It is expressed by formula (6):

[0050] L per =||VGG(V g )-VGG(V t )| | 2……Formula (6);

[0051] Where VGG(·) is the feature output by the next activation layer of the pooling layer in VGG-16 pre-trained on ImageNet.

[0052] Preferably, the above binarization loss L bin It is expressed by formula (7):

[0053]

[0054] Preferably, the centerline dice loss L cldice It is expressed by formula (8):

[0055]

[0056] Among them, T P is the specificity term, T S is the sensitivity term, and T P for T S for

[0057] Preferably, the above-mentioned skeleton loss L ske It is expressed by formula (9):

[0058] L ske =||S g -S t ||2……Formula (9).

[0059] Preferably, the calculation method of the soft binary image B and the soft skeletonized image S is as follows:

[0060] C1. Calculate the initial soft binary image B (0) and the initial soft skeletonized image S (0) , soft binary image B (0) Calculated by formula (10), the initial soft skeleton image S (0) Calculated by formula (11)-formula (12):

[0061] B (0) =Sigmoid(-α×(IT))…Formula (10);

[0062] R′ (0) =maxpool(minpool(B (0) )……Formula (11);

[0063] S (0) =ReLU(B (0) -B′ (0) )……Formula (12);

[0064] Among them, α is the binary granularity, I is the input blood vessel projection data V g And the corresponding reference vessel projection V t , T is the binarization threshold, Sigmoid(·) is the activation function used to map the output projection intensity to the interval [0, 1], minpool(·) is the minimum pooling operation, maxpool(·) is the maximum pooling operation, ReLU(·) is the activation function, and B′(0) is the image after one opening operation;

[0065] C2, set 0 = i-1, enter C3;

[0066] C3. Calculate the soft binary image B through formula (13) (i) , calculate the soft skeleton image S through equations (14) and (15) (i) ;

[0067] B (i) =minpool(B (i-1) )……Formula (13);

[0068] B′ (i) =maxpool(minpool(B (i) ))……Formula (14);

[0069]

[0070] in, is the element-wise product;

[0071] C4, determine whether i is equal to θ. If not, set i = i + 1 and return to C3; if yes, enter C5;

[0072] C5. Soft binary image B (i) Defined as soft binary image B, the soft skeleton image S (i) Defined as the soft skeletonized image S.

[0073] Preferably, the three-dimensional cerebral vascular imaging modality is three-dimensional digital subtraction angiography 3D-DSA, computed tomography angiography CTA, or magnetic resonance angiography MRA.

[0074] In B2, the three-dimensional cerebral vascular image V is processed by using the trimesh library of Python. s Perform voxelization; use the python ndimage library to voxelize the three-dimensional cerebral vascular image V s Perform morphological filling.

[0075] In B3, cone-beam CT front projection data simulation is performed using the TIGRE toolkit or the Astra toolkit.

[0076] A second object of the present invention is to provide a single-scan DSA imaging system that overcomes the shortcomings of the prior art and can accurately recover blood vessel projections from input projections while correcting most subtraction artifacts.

[0077] The above-mentioned purpose of the present invention is achieved through the following technical measures:

[0078] A single-scan DSA imaging system is provided to execute the above-mentioned single-scan DSA imaging method based on multi-view feature learning.

[0079] The present invention provides a single-scan DSA imaging method and system based on multi-view feature learning, wherein the single-scan DSA imaging method based on multi-view feature learning is performed by the following steps: S1, collecting multiple sets of digital subtraction angiography projection data and multiple sets of three-dimensional cerebral vascular images, obtaining real paired training data through the digital subtraction angiography projection data, and obtaining simulated paired data through the three-dimensional cerebral vascular images; multiple pairs of the real paired training data constitute a real paired training data set, and multiple pairs of the simulated paired data constitute a subtraction paired data set; S2, enhancing the real paired training data and the simulated paired data of S1 respectively, and then merging the enhanced real paired training data and the enhanced simulated paired data to obtain a complete subtraction paired data set; S3, training a learnable subtraction artifact correction model G with the complete subtraction paired data set obtained in S2 to obtain a trained learnable subtraction artifact correction model G; S4, inputting the multi-view filled circle projection data into the trained learnable subtraction artifact correction model G to obtain multi-view vascular projection data without subtraction artifacts. The beneficial effects of the present invention are as follows: 1. It provides a learnable subtraction artifact correction model with multi-view input. The dynamic snake convolution and efficient paired attention module embedded in the learnable subtraction artifact correction model can effectively learn the information between projections from different viewpoints, accurately extract the projections expressing the vascular structure, and overcome the limitations of the previous artifact correction model based on single-view input. 2. It uses a paired data set containing real and simulated data to train the learnable subtraction artifact correction model; wherein the real data can provide the model with complex noise, anatomical structure and other features in the actual scene, ensuring that the model has good adaptability and robustness in practical applications. Moreover, the virtual three-dimensional brain digital subtraction angiography projection simulation process provided by the present invention realizes paired data simulation without subtraction artifacts, overcoming the defect that the previous subtraction artifact correction model based on supervised learning could not obtain paired data without subtraction artifacts from real situations. At the same time, data simulation can provide a large number of diverse training samples for the training model, effectively improving the generalization ability of the model when real clinical data is insufficient. 3. High practicality. The subtraction artifact correction system provided by the present invention does not require an additional scanning mask ring during clinical use. Compared with the traditional three-dimensional digital subtraction angiography process, it reduces the radiation dose by 50%, effectively reducing the X-ray radiation dose borne by patients and equipment operators during the imaging process. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The present invention is further described with reference to the accompanying drawings, but the contents in the drawings do not constitute any limitation to the present invention.

[0081] Figure 1Flowchart of a single-scan DSA imaging method based on multi-view feature learning.

[0082] Figure 2 Figure 2 is the structural diagram of the learnable subtraction artifact correction model.

[0083] Figure 3 is a real clinical data diagram, where Figure 3 (a) is the mask circle scanning data M r , Figure 3 (b) is the scanning data F of the filling circle at the same angle r , Figure 3 (c) is the blood vessel projection data V p .

[0084] Figure 4 is the simulation data diagram, where Figure 4 (a) is the mask circle scan data M used for simulation r , Figure 4 (b) is the projection data of the simulated filling circle at the 52nd angle Figure 4 (c) is the simulated blood vessel projection data at the 52nd angle

[0085] Figure 5 is the artifact correction result, where Figure 5 (a) Scanning data of the filling circle at a certain angle; Figure 5 (b) Blood vessel projection obtained during the subtraction step of the conventional DSA imaging process; Figure 5 (c) The blood vessel projection obtained by inputting the filled circle scan data into the subtraction artifact correction model; Figure 5 (d) is a three-dimensional vascular image reconstructed by FDK after the vascular projection is corrected by the method of the present invention. DETAILED DESCRIPTION

[0086] The technical solution of the present invention is further described with reference to the following examples.

[0087] Example 1

[0088] A single-scan DSA imaging method based on multi-view feature learning, such as Figure 1 , proceed as follows:

[0089] S1. Acquire multiple sets of digital subtraction angiography projection data and multiple sets of three-dimensional cerebral vascular images, obtain real paired training data using the digital subtraction angiography projection data, and obtain simulated paired data using the three-dimensional cerebral vascular images; multiple pairs of the real paired training data constitute a real paired training data set, and multiple pairs of the simulated paired data constitute a subtraction paired data set;

[0090] S2, enhance the real paired training data and the simulated paired data in S1 respectively, and then merge the enhanced real paired training data and the enhanced simulated paired data to obtain a complete subtracted paired data set;

[0091] S3, training the learnable subtraction artifact correction model G using the complete subtraction paired dataset obtained in S2, to obtain the trained learnable subtraction artifact correction model G;

[0092] S4. After inputting the multi-view filled circle projection data into training, the subtraction artifact correction model G can be learned to obtain multi-view blood vessel projection data without subtraction artifacts.

[0093] The method for obtaining real paired training data is:

[0094] A1. Collecting the digital subtraction angiography projection data, which includes the mask circle projection data M. r and filled circle projection data F r ;

[0095] A2, the filling circle projection data F r and the mask circle projection data M r Subtraction is performed to obtain the blood vessel projection data V p ;

[0096] A3. Project the filled circle data F r and vascular projection data V p Constitute the real paired training data.

[0097] It should be noted that the real clinical digital subtraction angiography projection data refers to the filled circle projection data F obtained clinically by the C-arm system. r With the mask circle projection data M r , that is, 3D-DSA imaging modality. Vascular projection data V p Refers to the filled circle projection data F r With the mask circle projection data M r Projection data obtained after subtraction, where subtraction refers to the direct subtraction of the fill scan and mask scan at the same angle.

[0098] The method for obtaining simulation pairing data is:

[0099] B1. Acquire and reconstruct the 3D cerebral vascular image through 3D cerebral vascular imaging modality, and then segment it to obtain the 3D cerebral vascular image V s ;

[0100] B2. Three-dimensional cerebral vascular image V s Surface meshing and surface mesh preprocessing are performed, and then the three-dimensional cerebral vascular image V sPerform voxelization and morphological filling to obtain three-dimensional blood vessel data V with internal voxel values ​​of 1 x ; Among them, the surface mesh preprocessing operations include pruning the blood vessel region of interest, filling surface holes, and smoothing the blood vessel surface;

[0101] B3. Select a set of mask circle projection data M r , and then according to the scanned geometric parameters and the corresponding three-dimensional blood vessel data V x , simulate cone beam CT projection data to obtain simulated vascular projection data without subtraction artifacts

[0102] B4. Project the simulated blood vessels The corresponding mask circle projection data M r Perform angle-by-angle fusion. The fusion process includes scaling and translation operations on the vascular projection data to obtain the final simulated filling circle projection data.

[0103] B5. Project the simulation fill circle data and simulated vascular projection data Construct simulation pairing data.

[0104] It should be noted that since the three-dimensional images acquired and reconstructed by the three-dimensional cerebral vascular imaging modality usually contain brain structures such as the skull in addition to the vascular structures enhanced by contrast agents, it is necessary to segment these three-dimensional images to obtain the three-dimensional cerebral vascular image V s .

[0105] It should also be noted that a set of digital subtraction angiography projection data refers to projection data collected from one object, and a set of three-dimensional cerebrovascular images refers to a three-dimensional cerebrovascular image. In addition, a set of digital subtraction angiography projection data and a set of three-dimensional cerebrovascular images both include multiple angles, that is, filled circle projection data F r , vascular projection data V p , simulation filling circle projection data and simulated vascular projection data Both include multiple angles.

[0106] The training of the subtraction artifact correction model G can be learned in S3, which refers to randomly extracting real paired training data or simulated paired data from the subtraction paired data set for training.

[0107] like Figure 2 , the learnable subtraction artifact correction model G includes a block coding stage, an encoder downsampling stage, and a decoder upsampling stage;

[0108] The processing method of the block coding stage is: first, the multi-view projection data with the input dimension (H, W, D) is processed with a resolution of (P h , P w , P d ) is flattened into blocks, resulting in multiple where H and W are the dimensions of the projection data and D is the number of input views. The blocks are then mapped into a K-dimensional feature space through a convolutional layer. Finally, a K-dimensional learnable position code is added to each block to obtain the encoded block.

[0109] It should be noted that the number of projections and the projection dimensions H, W of the real paired training data are determined by the scanning model. The number and dimension of the projections of the simulated paired data are determined by the selected mask circle projection data M r The number and dimension of the projections are determined by the number of projections. The D number of multi-view filled circle projection data input by the learnable subtraction artifact correction model G is the projection of D adjacent angles in a data set selected from the paired data set, which is set to 16 in the present invention.

[0110] The encoder downsampling stage includes three downsampling modules, and each downsampling module consists of a dynamic snake convolution module, a downsampling convolution module and an efficient paired attention module connected in sequence from front to back.

[0111] The encoder upsampling stage includes three upsampling modules, and each upsampling module consists of an efficient paired attention module, a dynamic snake convolution module and an upsampling convolution module connected in sequence from front to back.

[0112] The encoder upsampling stage outputs the subtraction artifact corrected vascular projection data V with dimensions (H, W, D) g .

[0113] The upsampling convolution module uses a transposed convolution operation with a stride of 2. The downsampling convolution module uses a convolution operation with a stride of 2.

[0114] The dynamic snake convolution module calculates the position of a 7x7 convolution kernel in each dimensional feature map in the x-axis and y-axis directions using bilinear interpolation.

[0115] The efficient paired attention module is equipped with a spatial attention calculation unit, a channel attention calculation unit and an attention feature map fusion unit.

[0116] The efficient pairwise attention module can be learned by optimizing the weights of the value V in parallel. and and the weight W of the shared optimized query Q Q and the weight W of key K K .

[0117] The spatial attention calculation unit and the channel attention calculation unit both calculate the query Q, key K and value V, where the query Q and the key K are shared in the efficient pairwise attention module; the query Q is calculated as W Q X, the key K is calculated as W K X; the calculation method of the value V in the spatial attention calculation unit is The calculation method of the median V of the channel attention calculation unit is: X is the feature map input to the efficient pairwise attention module.

[0118] Spatial attention A in the spatial attention calculation unit s The calculation formula is expressed by formula (1):

[0119]

[0120] Among them, Softmax(·) is the activation function, proj(·) is the low-dimensional mapping operation implemented by matrix multiplication, T is the transpose operation, and d is the dimension of the vector in the numerator.

[0121] Channel attention A in the channel attention calculation unit c The calculation formula is expressed by formula (2):

[0122]

[0123] The attention feature map fusion unit fuses the features in an element-by-element manner and completes the fusion of the feature map A through two convolution operations. The calculation formula of the feature map A is expressed by formula (3):

[0124] A=Conv1(Conv3(A s +A c ))……Formula (3).

[0125] Among them, Conv1 is a convolution operation with a convolution kernel size of 1×1, and Conv3 is a convolution operation with a convolution kernel size of 3×3.

[0126] In S3, the learnable subtraction artifact correction model G is trained using stochastic gradient descent and constrained by a total loss function L.

[0127] The total loss function L is composed of the L1 loss function L1, the perceptual loss function L per and a topology-preserving loss function.

[0128] It should be noted that the L1 loss function L1 is used to constrain the generated blood vessel projection data V g And the corresponding reference vessel projection V tThe pixel-by-pixel error of . The topology preservation loss function is used to constrain the continuity and integrity of the generated vascular structure. The perceptual loss function L per Used to enhance texture details in images.

[0129] The topology preservation loss function transforms the blood vessel projection data V g And the corresponding reference vessel projection V t , reference vessel projection V t is the blood vessel projection data V p and simulated vascular projection data Set of, calculate the corresponding two soft binary outputs B g and B t , and the two soft skeleton outputs S of the soft skeleton image S g and S t , the topology preservation loss function includes the binarization loss L bin , skeletonization loss L ske and centerline dice loss L cldice .

[0130] The total loss function L is expressed by formula (4):

[0131] L=λ1L1+λ2L per +λ3L bin +λ4L ske +λ5L cldice ...Formula (4);

[0132] Among them, λ1, λ2, λ3, λ4 and λ5 are empirical coefficients, and λ1 and λ2 are both 1, λ3 and λ4 are both 0.8, and λ5 is 0.6.

[0133] L1 loss function L1 is generated by constraining the vascular projection data V g Compared with the reference vessel projection V t The pixel-by-pixel error of , the L1 loss function L1 is expressed by formula (5):

[0134] L1=||V g -V t ||1……Formula (5).

[0135] Perceptual loss function L per Used to enhance the texture details of the image, expressed by formula (6):

[0136] L per =|VGG(V g )VGG(V t )||2……Formula (6);

[0137] Where VGG(·) is the feature output by the next activation layer of the pooling layer in VGG-16 pre-trained on ImageNet.

[0138] Binarization loss L bin It is expressed by formula (7):

[0139]

[0140] Centerline dice loss L cldice It is expressed by formula (8):

[0141]

[0142] Among them, T P is the specificity term, T S is the sensitivity term, and T P for τ S for

[0143] Skeletonization loss L ske It is expressed by formula (9):

[0144] L ske =||S g -S t ||2……Formula (9).

[0145] The calculation method of the soft binary image B and the soft skeleton image S is as follows:

[0146] C1. Calculate the initial soft binary image B (0) and the initial soft skeletonized image S (0) , soft binary image B (0) Calculated by formula (10), the initial soft skeleton image S (0) Calculated by formula (11)-formula (12):

[0147] B (0) =Sigmoid(-α×(IT))…Formula (10);

[0148] B′ (0) =maxpool(minpool(B (0) ))……Formula (11);

[0149] S (0) =ReLU(B (0) -B′ (0) )……Formula (12);

[0150] Among them, α is the binary granularity, I is the input blood vessel projection data V g And the corresponding reference vessel projection Vt , T is the binarization threshold, Sigmoid(·) is the activation function used to map the output projection intensity to the interval [0, 1], minpool(·) is the minimum pooling operation, maxpool(·) is the maximum pooling operation, ReLU(·) is the activation function, and B′(0) is the image after one opening operation. The image after the opening operation is an image that is first eroded and then dilated. It contains some coarse blood vessel regions, which are used to further extract the blood vessel skeleton.

[0151] C2, set 0 = i-1, enter C3;

[0152] C3. Calculate the soft binary image B through formula (13) (i) , calculate the soft skeleton image S through equations (14) and (15) (i) ;

[0153] B (i) =minpool(B (i-1) )……Formula (13);

[0154] B′ (i) =maxpool(minpool(B (i) ))……Formula (14);

[0155]

[0156] in, is the element-wise product;

[0157] C4, determine whether i is equal to θ. If not, set i=i+1 and return to C3; if yes, enter C5; in this embodiment, θ is 40;

[0158] C5. Soft binary image B (i) Defined as soft binary image B, the soft skeleton image S (i) Defined as the soft skeletonized image S.

[0159] Three-dimensional cerebral vascular imaging modalities include three-dimensional digital subtraction angiography (3D-DSA), computed tomography angiography (CTA), and magnetic resonance angiography (MRA);

[0160] In B2, the three-dimensional cerebral vascular image V is processed by the Python trimesh library. s Perform voxelization; use the python ndimage library to voxelize the three-dimensional cerebral vascular image V s Perform morphological filling;

[0161] In B3, cone-beam CT front projection data simulation is performed using the TIGRE toolkit or the Astra toolkit.

[0162] The beneficial effects of the single-scan DSA imaging method based on multi-view feature learning are:

[0163] 1. A learnable subtraction artifact correction model with multi-view input is provided. The dynamic snake convolution and efficient paired attention module embedded in the model can effectively learn the information between projections from different viewpoints, accurately extract the projections expressing the vascular structure, and overcome the limitations of the previous artifact correction model based on single-view input. 2. A paired data set containing real and simulated data is used to train the learnable subtraction artifact correction model. The real data can provide the model with complex noise, anatomical structure and other features in the actual scene, ensuring that the model has good adaptability and robustness in practical applications. Moreover, the virtual three-dimensional brain digital subtraction angiography projection simulation process provided by the present invention realizes paired data simulation without subtraction artifacts, overcoming the defect that the previous subtraction artifact correction model based on supervised learning could not obtain paired data without subtraction artifacts from real situations. At the same time, data simulation can provide a large number of diverse training samples for the training model, effectively improving the generalization ability of the model when real clinical data is insufficient. 3. High practicality. The subtraction artifact correction system provided by the present invention does not require an additional scanning mask ring during clinical use. Compared with the traditional three-dimensional digital subtraction angiography process, it reduces the radiation dose by 50%, effectively reducing the X-ray radiation dose borne by patients and equipment operators during the imaging process.

[0164] Example 2

[0165] A specific application of the single-scan DSA imaging method based on multi-view feature learning as in Example 1:

[0166] S1. Collect multiple sets of real clinical brain digital subtraction angiography projection data, one of which contains mask circle projection data M under 152 scanning angles. r And the filled circle projection data F r , apply logarithmic transformation to all projection data and take the negative, where the mask circle scan data M under the 52nd angle r ,like Figure 3 (a) in the figure corresponds to the filling circle scan F at the same angle. r ,like Figure 3 (b).

[0167] All projection data are normalized using a maximum value of 0.4 and a minimum value of -6. The fill scan and mask scan at the same angle are subtracted from each other, and the subtracted images are normalized using a maximum value of 0.07 and a minimum value of -0.27 to obtain multi-angle vascular projection data V p , where the blood vessel projection data V at the 52nd angle p ,like Figure 3(c) Filled circle projection data F at the corresponding angle r With the blood vessel projection data V p The above preprocessing steps are repeated for the remaining sets of real clinical brain digital subtraction angiography projection data to complete the construction of the real data part of the subtraction paired data set.

[0168] A plurality of three-dimensional blood vessels segmented from existing three-dimensional cerebral angiography modality reconstructed images are acquired, wherein one of the three-dimensional blood vessels is acquired and segmented from a magnetic resonance angiography modality.

[0169] The surface of the three-dimensional blood vessels was preprocessed using Geomagic Warp software, and the trimesh library and ndimage library of Python were used to voxelize and fill the three-dimensional blood vessels with the surface preprocessed, and the simulated three-dimensional blood vessel data V with all internal voxel values ​​​​of 1 were obtained. x .

[0170] Randomly preprocess the three-dimensional blood vessel data V x With a set of mask circle scan data M in step 1 r Pairing, in this example, it is assumed that the three-dimensional blood vessel data V x Example and mask circle projection data M r Example pairing. According to this example, the mask circle projection data M r ,like Figure 4 (a) Geometric parameters and projection matrix, simulated blood vessel projection data at 152 angles using the Tiger toolkit The size of each voxel in the XYZ dimensions is set to [0.3858mm, 0.3858mm, 0.3858mm], the number of pixels of the flat panel detector is 489x489, and the sampling step of the projection data is 0.5. The simulated blood vessel projection data at the 52nd angle like Figure 4 (c).

[0171] The obtained simulated blood vessel projection data The corresponding mask circle projection data M r Perform angle-by-angle fusion. The fusion process includes scaling the blood vessel projection by a factor of 0.8 and performing a 48-pixel translation operation in the Y-axis direction to obtain the final simulated filled circle projection data. The simulation filling circle projection data at the 52nd angle like Figure 4(b) The simulated filled circle projection data and the simulated blood vessel projection data form a pair of simulated paired data. Repeat the above preprocessing steps for the remaining sets of real clinical brain digital subtraction angiography projection data, including the mask scan data and preprocessed 3D blood vessels, to complete the construction of the simulated data portion of the subtraction paired data set.

[0172] S2, enhance the real paired training data and the simulated paired data in S1 respectively, and then merge the enhanced real paired training data and the enhanced simulated paired data to obtain a complete subtracted paired data set;

[0173] S3. Use the subtraction paired dataset to train the subtraction artifact correction model G. After training, the subtraction artifact correction model G can be learned. The dataset contains 5433 sets of paired data with a dimension of [512, 512]. The training epoch is set to 16, the initial learning rate is set to 1e-5, and the Adam optimizer is used to perform stochastic gradient descent strategy optimization model training.

[0174] S4. Obtain a case of brain digital subtraction angiography projection data collected clinically, input the multi-angle filled circle scan data into the trained subtraction artifact correction model, and obtain the corresponding multi-angle subtraction artifact corrected vascular projection, such as Figure 5 .

[0175] Figure 5 is an example of artifact correction result of the method of the present invention, Figure 5 From left to right: Scanning data of the filling circle at a certain angle in this embodiment, Figure 5 (a); In this embodiment, a blood vessel projection is obtained by using the subtraction step of the conventional DSA imaging process, Figure 5 (b) Inputting the filled circle scan data of this embodiment into the blood vessel projection obtained by the subtraction artifact correction model constructed by the present invention, Figure 5 (c) The three-dimensional vascular image is reconstructed by FDK using the vascular projection corrected by the method of the present invention. Figure 5 (d).

[0176] pass Figure 5 It can be seen that compared with the direct subtraction method of the traditional imaging process, the method provided by the present invention can effectively correct subtraction artifacts. At the same time, the image reconstructed by the vascular projection corrected by the method of the present invention can clearly display the three-dimensional vascular structure.

[0177] Example 3

[0178] A single-scan DSA imaging system implements the single-scan DSA imaging method based on multi-view feature learning of embodiment 1 or embodiment 2.

[0179] The beneficial effects of the single-scan DSA imaging system are as follows: 1. It provides a learnable subtraction artifact correction model with multi-view input. The dynamic snake convolution and efficient paired attention module embedded in the model can effectively learn the information between projections from different perspectives, accurately extract the projections expressing vascular structures, and overcome the limitations of previous artifact correction models based on single-view input. 2. It uses a paired data set containing real and simulated data to train the learnable subtraction artifact correction model. The real data can provide the model with complex noise, anatomical structure and other features in the actual scene, ensuring that the model has good adaptability and robustness in practical applications. Moreover, the virtual three-dimensional brain digital subtraction angiography projection simulation process provided by the present invention realizes paired data simulation without subtraction artifacts, overcoming the defect that the previous subtraction artifact correction model based on supervised learning could not obtain paired data completely free of subtraction artifacts from real situations. At the same time, data simulation can provide a large number of diverse training samples for the training model, effectively improving the generalization ability of the model when real clinical data is insufficient. 3. High practicality. The subtraction artifact correction system provided by the present invention does not require an additional scanning mask ring during clinical use. Compared with the traditional three-dimensional digital subtraction angiography process, it reduces the radiation dose by 50%, effectively reducing the X-ray radiation dose borne by patients and equipment operators during the imaging process.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A single-scan DSA imaging method based on multi-view feature learning, characterized in that: Proceed as follows: S1. Acquire multiple sets of digital subtraction angiography projection data and multiple sets of three-dimensional cerebral vascular images, obtain real paired training data using the digital subtraction angiography projection data, and obtain simulated paired data using the three-dimensional cerebral vascular images; multiple pairs of the real paired training data constitute a real paired training data set, and multiple pairs of the simulated paired data constitute a subtraction paired data set; S2, respectively enhancing the real paired training data and the simulated paired data of S1, and then merging the enhanced real paired training data and the enhanced simulated paired data to obtain a complete subtracted paired data set; S3, training the learnable subtraction artifact correction model G using the complete subtraction paired dataset obtained in S2, to obtain the trained learnable subtraction artifact correction model G; S4. After inputting the multi-view filled circle projection data into training, the subtraction artifact correction model G can be learned to obtain multi-view blood vessel projection data without subtraction artifacts.

2. The single-scan DSA imaging method based on multi-view feature learning according to claim 1, characterized in that: The method for obtaining the real paired training data is: A1. Acquire the digital subtraction angiography projection data, wherein the digital subtraction angiography projection data includes mask circle projection data M. r and filled circle projection data F r ; A2, the filling circle projection data F r and the mask circle projection data M r Subtraction is performed to obtain the blood vessel projection data V p ; A3, make the filled circle projection data F r and the blood vessel projection data V p constitute the real paired training data.

3. The single-scan DSA imaging method based on multi-view feature learning according to claim 2, characterized in that: The method for obtaining the simulation pairing data is: B1. Acquire and reconstruct the 3D cerebral vascular image through 3D cerebral vascular imaging modality, and then segment it to obtain the 3D cerebral vascular image V s ; B2. Three-dimensional cerebral vascular image V s Surface meshing and surface mesh preprocessing are performed, and then the three-dimensional cerebral vascular image V s Perform voxelization and morphological filling to obtain three-dimensional blood vessel data V with internal voxel values ​​of 1 x ; B3. Select a set of mask circle projection data M r , and then according to the scanned geometric parameters and the corresponding three-dimensional blood vessel data V x , simulate cone beam CT projection data to obtain simulated vascular projection data without subtraction artifacts B4, the simulated blood vessel projection data The corresponding mask circle projection data M r Perform angle-by-angle fusion. The fusion process includes scaling and translation operations on the vascular projection data to obtain the final simulated filling circle projection data. B5, making the simulation filling circle projection data and the simulated blood vessel projection data The simulation pairing data is formed.

4. The single-scan DSA imaging method based on multi-view feature learning according to claim 3, characterized in that: The learnable subtraction artifact correction model G includes a block coding stage, an encoder downsampling stage and a decoder upsampling stage; The processing method of the block coding stage is: first, the multi-view projection data with input dimensions (H, W, D) is processed with a resolution of (P h , P w , P d ) is flattened into blocks, resulting in multiple where H and W are the dimensions of the projection data and D is the number of input views. The blocks are then mapped into a K-dimensional feature space through a convolutional layer. Finally, a K-dimensional learnable positional encoding is added to each block to obtain the encoded block; The encoder downsampling stage includes three downsampling modules, and each downsampling module consists of a dynamic snake convolution module, a downsampling convolution module and an efficient paired attention module connected in sequence from front to back.

5. The single-scan DSA imaging method based on multi-view feature learning according to claim 4, characterized in that: The encoder upsampling stage includes three upsampling modules, and each upsampling module consists of an efficient paired attention module, a dynamic snake convolution module and an upsampling convolution module connected in sequence from front to back; The encoder upsampling stage outputs the subtraction artifact corrected vascular projection data V with dimensions (H, W, D) g .

6. The single-scan DSA imaging method based on multi-view feature learning according to claim 5, characterized in that: The upsampling convolution module uses a transposed convolution operation with a stride of 2; The downsampling convolution module uses a convolution operation with a stride of 2; The dynamic snake convolution module calculates the position of a 7x7 convolution kernel in each dimensional feature map in the x-axis and y-axis directions by bilinear interpolation; The efficient paired attention module is provided with a spatial attention calculation unit, a channel attention calculation unit and an attention feature map fusion unit; The efficient paired attention module can be learned by optimizing the weights of the values ​​V in parallel and and the weight W of the shared optimized query Q Q and the weight W of key K K ; The spatial attention calculation unit and the channel attention calculation unit both calculate a query Q, a key K, and a value V, wherein the query Q and the key K are shared in the efficient pairwise attention module; The query Q is calculated as W Q X, the key K is calculated as W K X; the calculation method of the value V in the spatial attention calculation unit is The calculation method of the median V of the channel attention calculation unit is: X is the feature map input to the efficient paired attention module; The spatial attention A in the spatial attention calculation unit s The calculation formula is expressed by formula (1): Among them, Softmax(·) is the activation function, proj(·) is the low-dimensional mapping operation, T is the transpose operation, and d is the dimension of the vector in the numerator; The channel attention A in the channel attention calculation unit c The calculation formula is expressed by formula (2): The attention feature map fusion unit fuses the features in an element-by-element manner and completes the fusion of the feature map A through two convolution operations. The calculation formula of the feature map A is expressed by formula (3): A = Conv1(Conv3(A s + A c )) …… Equation (3); Among them, Conv1 is a convolution operation with a convolution kernel size of 1×1, and Conv3 is a convolution operation with a convolution kernel size of 3×3.

7. The single-scan DSA imaging method based on multi-view feature learning according to claim 6, characterized in that: In S3, the learnable subtraction artifact correction model G is trained using a stochastic gradient descent method and constrained by a total loss function L; The total loss function L is composed of L1 loss function L1, perception loss function L per and topology preservation loss function; The topology preservation loss function converts the blood vessel projection data V g And the corresponding reference vessel projection V t , the reference blood vessel projection V t is the blood vessel projection data V p and simulated vascular projection data Set of, calculate the corresponding two soft binary outputs B g and B t , and the two soft skeleton outputs S of the soft skeleton image S g and S t ; The topology preservation loss function includes a binarization loss L bin , skeletonization loss L ske and centerline dice loss L cldice ; The total loss function L is expressed by formula (4): L=λ1L1+λ2L per +λ3L bin +λ4L ske +λ5L cldice ……expression(4); Among them, λ1, λ2, λ3, λ4 and λ5 are empirical coefficients, and λ1 and λ2 are both 1, λ3 and λ4 are both 0.8, and λ5 is 0.6; The L1 loss function L1 is generated by constraining the blood vessel projection data V g Compared with the reference vessel projection V t The pixel-by-pixel error, the L1 loss function L1 is expressed by formula (5): L1 = ||V g -V t ||1 …… Equation (5); The perceptual loss function L pre It is expressed by formula (6): L per = ||VGG(V g ) - Vgg(V t )||2 …… Equation (6); Where VGG(·) is the feature output by the next activation layer of the pooling layer in VGG-16 pre-trained on ImageNet; The binarization loss L bin It is expressed by formula (7): The centerline dice loss L cldice It is expressed by formula (8): Among them, T P is the specificity term, T S is the sensitivity term, and T P for T S for The skeletonization loss L ske It is expressed by formula (9): L ske =||S g -S t ||2 ...Formula (9).

8. The single-scan DSA imaging method based on multi-view feature learning according to claim 7, characterized in that: The calculation method of the soft binary image B and the soft skeleton image S is as follows: C1. Calculate the initial soft binary image B (0) and the initial soft skeletonized image S (0) , soft binary image B (0) Calculated by formula (10), the initial soft skeleton image S (0) Calculated by formula (11)-formula (12): B (0) = Sigmoid(-α×(I - T)) …… Equation (10); B′ (0) = maxpool(minpool(B (0) )) …… Equation (11); S (0) =ReLU(B (0) -B′ (0) ) ...Formula (12); Among them, α is the binary granularity, I is the input blood vessel projection data V g And the corresponding reference vessel projection V t , T is the binarization threshold, Sigmoid(·) is the activation function used to map the output projection intensity to the [0, 1] interval, minpool(·) is the minimum pooling operation, maxpool(·) is the maximum pooling operation, ReLU(·) is the activation function, B′ (0) is the image after one opening operation; C2, set 0 = i-1, enter C3; C3. Calculate the soft binary image B through formula (13) (i) , calculate the soft skeleton image S through equations (14) and (15) (i) ; B (i) =minpool(B (i-1) ) ...Formula (13); B′ (i) = maxpool(minpool(B (i) ))... Equation (14); in, is the element-wise product; C4, determine whether i is equal to θ. If not, set i = i + 1 and return to C3; if yes, enter C5; C5. Soft binary image B (i) Defined as soft binary image B, the soft skeleton image S (i) Defined as the soft skeletonized image S.

9. The single-scan DSA imaging method based on multi-view feature learning according to claim 8, characterized in that: The three-dimensional cerebral vascular imaging modality is three-dimensional digital subtraction angiography 3D-DSA, computed tomography angiography CTA, and magnetic resonance angiography MRA; In B2, the three-dimensional cerebral vascular image V is processed by using the trimesh library of Python. s Perform voxelization; use the python ndimage library to voxelize the three-dimensional cerebral vascular image V s Perform morphological filling; In B3, cone-beam CT front projection data simulation is performed using the TIGRE toolkit or the Astra toolkit.

10. A single-scan DSA imaging system, characterized in that: Execute the single-scan DSA imaging method based on multi-view feature learning as described in any one of claims 1 to 9.

Images