Fast Pipeline Computed Tomography Method Based on Deep Learning

Abstract

The present invention discloses a fast pipeline computed tomography (CT) method based on deep learning. The steps are as follows: Place multiple objects on a pipeline CT inspection table, perform synchronous CT scanning based on sparse sampling to obtain a pipeline CT sparse projection sequence; Interpolate the pipeline CT sparse projection sequence into a complete projection sequence with missing information; Use deep learning technology to process the projection sequence with missing information, supplement the projection information, and obtain a projection sequence with complete information and dimensions; Perform segmentation processing on the complete projection sequence to obtain the projection sequence corresponding to each object; Use the filtered back-projection reconstruction algorithm to reconstruct the projection sequence of each object separately to obtain the final tomographic reconstruction image. The present invention greatly shortens the imaging time; At the same time, it ensures the reconstruction quality of the fast pipeline CT imaging technology, making it meet the requirements of high-quality and high-efficiency industrial non-destructive testing.

Description

Technical Field

[0001] The present invention relates to the technical fields of X-ray computed tomography and artificial intelligence, and particularly to a fast pipeline computed tomography method based on deep learning. Background Art

[0002] In an X-ray computed tomography (CT) system, an X-ray source emits X-rays, which pass through a certain area of an object to be measured from different angles. A detector placed opposite the X-ray source receives the X-rays at corresponding angles. Then, according to the different degrees of attenuation of the X-rays at each angle, certain reconstruction algorithms and a computer are used for calculation to reconstruct a mapping image of the ray attenuation coefficient distribution of the scanned area of the object, thereby realizing the reconstruction of an image from projections and nondestructively reproducing the characteristics such as the medium density, composition, and structural form of the object in this area.

[0003] The imaging efficiency has always been one of the main factors restricting the wide application of CT. It is mainly determined by the scanning time and the image reconstruction time. Currently, due to the widespread use of the graphics processing unit (GPU) and its corresponding parallel computing architecture CUDA, the image reconstruction time has been greatly improved. Therefore, to further improve the imaging efficiency, fast scanning technologies need to be developed.

[0004] However, fast scanning technologies usually result in missing projection information. Currently, the most commonly used reconstruction algorithm is the Filtered Back Projection (FBP) algorithm. When applied to complete data, the FBP reconstruction is fast and the obtained image quality is good. But when the projection data is incomplete, serious artifacts and noise will exist in the corresponding FBP reconstruction results. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a fast pipeline computed tomography method based on deep learning. By using a pipeline CT architecture and a sparse sampling strategy, fast pipeline CT scanning is realized, and the imaging time is greatly shortened. At the same time, deep learning technology is used to ensure the reconstruction quality of the fast pipeline CT imaging technology, so as to meet the requirements of high-quality and high-efficiency industrial nondestructive testing.

[0006] The technical solution of the present invention: A fast pipeline computed tomography method based on deep learning, comprising the following steps:

[0007] Step 1: Place multiple objects on a pipeline CT imaging inspection table, perform synchronous sparse sampling scanning to obtain a pipeline CT sparse projection sequence; multiple independent turntables parallel to the detector direction are configured on the pipeline CT imaging inspection table, and can perform independent imaging on multiple objects, so that their projection sequences do not interfere with each other;

[0008] Step 2: Interpolate the pipeline CT sparse projection sequence into a complete projection sequence with missing information; the missing information is due to the loss of projection data at some angles during sparse angle scanning, and interpolation approximation can change the size of the projection sequence but cannot supplement the projection information;

[0009] Step 3: Use deep learning technology to process the projection sequence with missing information, supplement the projection information, and obtain a complete projection sequence with complete and accurate projection information; the complete projection sequence with complete and accurate projection information is obtained by using deep learning technology to process the projection sequence with missing information, and the reconstructed image of the projection sequence no longer has sparse scanning artifacts. The deep learning technology refers to the optimization technology of the pipeline CT sparse projection sequence based on convolutional neural network;

[0010] Step 4: Segment the processed projection sequence above to obtain an independent projection sequence for each object; the independent projection sequence is because the pipeline CT images multiple objects simultaneously, and its projection sequence includes the projections of multiple objects. It is necessary to calculate the positions of each object corresponding to the projection sequence and perform projection segmentation;

[0011] Step 5: Use the filtered back-projection reconstruction algorithm to reconstruct each object's projection sequence above respectively to obtain the final tomographic reconstruction image.

[0012] Further, the pipeline CT imaging in Step 1 is different from the conventional multi-object CT imaging where multiple objects rotate around a common rotation axis. Instead, each object is equipped with an independent rotation axis; in addition, sparse sampling scanning further reduces the CT imaging time.

[0013] Further, for the pipeline CT sparse sampling projection sequence in Step 2, use the bicubic interpolation algorithm to expand and interpolate it approximately to make its size the same as that of the complete sampling projection sequence.

[0014] Further, the projection optimization technology based on deep learning in Step 3 is shown in formulas (1)-(4):

[0015]

[0016] f(P(ω,φ))=W T ·P(ω,φ)+Bias (2)

[0017]

[0018]

[0019] where P(ω,φ) is the projection sequence with missing information and complete size, is a complete projection sequence with complete and accurate projection information, where (ω, φ) represents the detector element position and rotation angle of the projection sequence; f() and F() represent the encoding network and decoding network based on deep learning technology, which are respectively used to extract features from P(ω, φ) and parse the information missing degree of the projection sequence from the features; Λ represents a non-linear mapping function; Error represents the learning objective of the convolutional neural network, which is used to measure the difference between the network output and the label; W and Bias represent the learning parameters in the convolutional neural network - weights and biases, and the update of the above parameters is achieved by using the gradient descent algorithm to solve the partial derivatives of the learning objective with respect to the parameters; η and respectively represent the learning rate and the network parameters to be learned.

[0020] Furthermore, both the encoding network and the decoding network are composed of multiple levels of convolutional neural network layers. In the two networks, the height and width dimensions of the feature maps at each level are respectively reduced and increased by a factor of two, while the corresponding number of feature maps is opposite. The feature maps with the same height and width dimensions in the encoding network and the decoding network are concatenated and then used as the input feature map of the next-level decoding network.

[0021] Furthermore, the segmentation process described in step 4 needs to be based on the accurate calculation of the position parameters corresponding to each object in the projection sequence, as shown in equations (5)-(6):

[0022]

[0023]

[0024] where, S A and S B respectively represent the left and right positions of the projection of a certain object in the two-dimensional projection sequence, D is the distance from the radiation source to the detector, s is the distance from the projection position of the rotation axis center corresponding to the object on the detector to the center of the detector, r is the radius of gyration of the object, E represents the projection position of the rotation center of the rotation axis where the object is located in the detector and the position of the radiation source, tan and tan -1 respectively represent the tangent operation and the arctangent operation, sin and sin -1 represent the sine operation and the arcsine operation.

[0025] Furthermore, the Filtered Back-projection (FBP) reconstruction algorithm described in step 5 is as shown in equation (7):

[0026]

[0027] where, Denote the complete projection sequence of the network output, \(R(r,\theta)\) denote the reconstructed image, \((r,\theta)\) denote polar coordinates, \(U\) denote the projection weight matrix, \(D\) denote the distance from the ray source to the rotation center of the turntable, \(h\) denote the one-dimensional filter, and \((\omega,\varphi)\) denote the detector element coordinates and the turntable rotation angle respectively.

[0028] Compared with the traditional computed tomography method, the embodiment of the present invention realizes fast pipeline CT scanning through a pipeline CT imaging architecture and a sparse sampling strategy, greatly shortening the imaging time; and uses deep learning technology to maintain the high quality of the reconstruction result to meet the high-quality and high-efficiency production requirements of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of the fast pipeline computed tomography method based on deep learning of the present invention;

[0030] Figure 2 is the pipeline CT imaging architecture of the fast pipeline computed tomography method based on deep learning of the present invention;

[0031] Figure 3 is the network structure diagram of the deep learning technology of the fast pipeline computed tomography reconstruction method based on deep learning of the present invention;

[0032] Figure 4 are the projection sequence of the sparse angle data of the pipeline CT imaging processed by the present invention, the missing information projection sequence obtained by interpolation, the optimized projection sequence after being processed by deep learning technology, and the complete projection sequence image; where a is the pipeline CT sparse angle projection sequence (45×1600); b is the complete projection sequence with missing information after bicubic interpolation (360×1600); c is the optimized complete projection sequence after deep learning technology (360×1600); d is the complete projection sequence (360×1600);

[0033] Figure 5 is the corresponding to Figure 4 the projection sequence of each object corresponding to the projection sequence; a is the projection sequence of the left object; b is the projection sequence of the middle object; c is the projection sequence of the right object;

[0034] Figure 6 is the corresponding to Figure 5 the reconstructed images corresponding to the projection sequence, where a is the reconstructed image of the sparse projection, b is the reconstructed image based on deep learning technology, and c is the reconstructed image of the complete projection sequence. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0036] Figure 1 The figure is a flowchart of a fast pipeline computed tomography reconstruction method based on deep learning provided by an embodiment of the present invention. An embodiment of the present invention provides a reconstruction method based on deep learning for the incomplete projection data of multiple objects obtained by sparse sampling pipeline computed tomography. The specific steps of this method are as follows:

[0037] Step S101: Place multiple objects on the pipeline CT imaging table and perform sparse sampling scanning synchronously; in the pipeline CT imaging table, a cone-beam X-ray source is used as the ray source, and multiple independent turntables are installed parallel to the detector direction, as Figure 2 shown, to obtain a projection sequence of multiple objects that do not affect each other.

[0038] Step S102: Use bicubic interpolation to interpolate the pipeline CT projection sequence based on sparse sampling into a complete projection sequence with missing information. Since the complete projection sequence interpolates the sparse sampling projection sequence, its completeness only reflects in the angle of the projection sequence, but the problem of missing information is not solved.

[0039] Step S103: Use deep learning technology to process the projection sequence with missing information, supplement the projection information, and obtain a complete and accurate projection sequence with complete projection information. The complete and accurate projection sequence with complete projection information is obtained by using deep learning technology to process the projection sequence with missing information. The obtained projection sequence contains projection data under complete sampling angles, and its projection quantity and information are complete.

[0040] Figure 3 The figure is a structural diagram of an example of deep learning technology for the fast pipeline computed tomography reconstruction method based on deep learning of the present invention. As Figure 3 shown, it consists of a 5-level encoding layer and a 4-level decoding layer composed of a convolutional neural network. In each level of the encoding layer and the decoding layer, the height and width dimensions of the feature map are respectively reduced and increased by a factor of two, and the corresponding number of feature maps is opposite. The feature maps with the same height and width dimensions in the encoding network and the decoding network are spliced and then used as the input feature map of the next-level decoding network.

[0041] Step S104: Calculate the accurate position parameters of each object corresponding to the projection sequence, and perform segmentation processing on the processed projection sequence with the position parameters to obtain the projection sequence corresponding to each object in the pipeline CT scan. The projection sequence corresponding to each object is because the pipeline CT simultaneously images multiple objects, and its projection sequence includes the projections of multiple objects, so projection segmentation is required.

[0042] Step S105: Use the filtered back-projection reconstruction algorithm to reconstruct each object's projection sequence respectively to obtain the final tomographic reconstruction image.

[0043] The present invention has two advantages compared with the traditional CT method: 1) By means of a pipeline CT imaging architecture and a sparse sampling strategy, a fast pipeline CT imaging technology is realized, and the imaging time is greatly shortened; 2) A reconstruction framework based on deep learning technology solves the problem of information loss in the projection sequence obtained by the above architecture and optimizes the imaging quality of the reconstructed image.

[0044] To prove the effects of the above embodiments, the present invention conducted the following experiments, and the experimental steps are as follows:

[0045] (1) Conduct fast pipeline CT experiments on multiple objects. Synchronously sparse scan multiple objects through a pipeline CT imaging device, ensuring that the information of each object in the projection sequence is independent of each other and does not interfere with each other while greatly accelerating the imaging. The experimental conditions are set as follows: Simultaneously scan three objects using pipeline CT imaging, set the sampling factor to 4, and obtain 360° projections at 90 sampling angles. At this time, the acquisition condition for complete projections is that the sampling factor is set to 1.

[0046] (2) Use bicubic interpolation to interpolate and approximate the sparse projection sequence into an angularly complete projection sequence.

[0047] (3) According to Figure 3 and formulas (1)-(4), process the complete projection sequence with missing information to obtain a complete projection sequence with complete information and dimensions.

[0048] (4) According to formulas (5) and (6), divide the projection sequence into projection sequences of three objects.

[0049] (5) Use the FBP reconstruction algorithm to obtain the final reconstruction result.

[0050] Figure 4 a to d in Figure 5 are respectively the sparse sampling projection sequences, the complete projection sequences with missing information, the projection sequences optimized by deep learning, and the complete projection sequence images of the pipeline CT imaging of the three objects processed in the embodiments of the present invention. Figure 4 a to c in Figure 6 are respectively the projection sequences of each object corresponding to the Figure 5 c projection sequence divided in the embodiments of the present invention. Figure 6 is the reconstructed image corresponding to the Figure 5 a projection sequence processed in the embodiments of the present invention, and

[0051] From Figures 4 - 6 it can be seen that Figure 6 b in Figure 6The a in it eliminates the sparse reconstruction artifacts, indicating that the fast pipeline computed tomography reconstruction method based on deep learning can effectively process the data reconstruction in the case of sparse sampling pipeline CT.

[0052] Compared with the traditional computed tomography method, the embodiment of the present invention realizes fast pipeline CT scanning through the combination of the pipeline CT imaging architecture and the sparse sampling strategy, greatly shortening the imaging time; compared with the previous reconstruction methods for incomplete data based on deep learning where the object is a single object, the embodiment of the present invention directly acts on the projection sequences of multiple objects, the operation process is simple and clear, and the optimization in the projection domain is more conducive to retaining image details to improve the image quality.

[0053] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fast pipeline computed tomography method based on deep learning, characterized in that, It includes the following steps: Step 1: Place multiple objects on the pipeline CT imaging inspection table and perform sparse sampling scanning synchronously to obtain a pipeline CT sparse projection sequence; multiple independent turntables parallel to the detector direction are installed on the pipeline CT imaging inspection table, which can independently image multiple objects so that their projection sequences do not interfere with each other; Step 2: Interpolate the pipeline CT sparse projection sequence into a projection sequence with missing information but complete dimensions; the missing information means that some angles of projection data are lost in sparse angle scanning, and interpolation approximately changes the dimensions of the projection sequence but cannot supplement the projection information; Step 3: Use deep learning technology to process the projection sequence with missing information and complete dimensions, supplement the projection information, and obtain a complete projection sequence with complete and accurate projection information. The reconstructed image of the complete projection sequence no longer has sparse artifacts. The deep learning technology refers to the optimization technology of the pipeline CT sparse projection sequence based on the convolutional neural network; Step 4: Segment the processed complete projection sequence to obtain an independent projection sequence corresponding to each object; the independent projection sequence is because the pipeline CT images multiple objects simultaneously, and its projection sequence includes the projections of multiple objects. It is necessary to calculate the positions of each object in the projection sequence and perform projection segmentation; Step 5: Use the filtered back-projection reconstruction algorithm to reconstruct each independent projection sequence corresponding to each object respectively to obtain the final tomographic reconstruction image; The fast computed tomography method is realized by combining the pipeline CT imaging architecture and the sparse sampling strategy. The pipeline CT imaging architecture performs non-interfering synchronous scanning of multiple objects, and the sparse sampling scanning reduces the single-scan duration. The combination of the two greatly shortens the scanning time; In Step 2, the bicubic interpolation method is used to interpolate the pipeline CT sparse projection sequence and transform it into a projection sequence with missing information and complete dimensions.

2. The fast pipeline computed tomography method based on deep learning according to claim 1, wherein: In Step 3, the convolutional neural network shown in formulas (1)-(4) is used to process the complete projection sequence with missing information, specifically as follows: (1) (2) (3) (4) Among them, is a projection sequence with missing information and complete dimensions, is a complete projection sequence with complete and accurate projection information, represents the detector element position and turntable rotation angle corresponding to each pixel in the projection sequence; f and F represent the encoding network and decoding network in the convolutional neural network, which are respectively used to extract features from and resolve the missing situation of projection information from the features; represents a non-linear mapping function; Error represents the learning objective of the convolutional neural network, which is used to measure the difference between the output of the convolutional neural network and the label; W and Bias represent the learning parameters in the convolutional neural network, namely weights and biases, and the parameter update is achieved by using the gradient descent algorithm to solve the partial derivative of the learning objective with respect to the parameters; and represent the learning rate and the inherent network parameters of learning respectively.

3. The fast pipeline computed tomography method based on deep learning according to claim 1, wherein: In Step 4, formulas (5)-(6) are used to calculate the positions of each object in the projection sequence for subsequent projection sequence segmentation; (5) (6) wherein, and respectively represent the left and right positions of the projection of an object in a two-dimensional projection sequence, is the distance from the ray source to the detector, is the distance from the projection position of the center of the turntable rotation axis on the detector to the center of the detector, is the radius of gyration of the object, E represents the projection position of the rotation center of the rotation axis where the object is located in the detector and the position of the ray source, tan and tan -1 respectively represent the tangent operation and the arctangent operation, sin and sin -1 represent the sine operation and the arcsine operation.

4. The fast pipeline computed tomography method based on deep learning according to claim 1, characterized in that: In Step 5, the filtered back-projection reconstruction algorithm is as shown in formula (7): (7) Among them, represents the complete projection sequence output by the convolutional neural network, represents the reconstructed image, represents polar coordinates, represents the projection weight matrix, represents the distance from the ray source to the rotation center of the turntable, represents a one-dimensional filter, respectively represent the detector element coordinates, the turntable rotation angle, and the rotation angle.

Images