Methods, systems, and media for constructing neural networks suitable for sparse angular CT characterization

Abstract

This invention provides a method, system, and medium for constructing a neural network suitable for sparse angle CT representation, comprising: Step S1: Constructing a deep neural network G1 in the projection sinusoidal domain, and using neural network G1 to repair the projection sinusoidal diagram under sparse angles to obtain a sinusoidal diagram with complete angles; Step S2: Converting the full-angle sinusoidal diagram into a tomographic image using a filtered back-projection algorithm; Step S3: Extracting high-frequency information from the obtained tomographic image; Step S4: Constructing a deep neural network G2 in the tomographic domain, and using convolutional neural network G2 to perform further super-resolution reconstruction of the reconstructed tomographic image; Step S5: Constraining the neural network framework according to the constructed high-frequency information, and training the network using the gradient descent method. This invention can increase the constraint of high-frequency information to improve the network's representation of detailed information.

Description

Technical Field

[0001] This invention relates to the fields of sparse angle CT reconstruction, convolutional neural networks (DNNs), sparse angle CT representation, and CT image super-resolution technology. Specifically, it relates to a high-frequency information-constrained neural network suitable for sparse angle CT representation, and more particularly to a method, system, and medium for constructing a neural network suitable for sparse angle CT representation. Background Technology

[0002] Computed tomography (CT) technology can obtain the three-dimensional structure of materials. It is currently widely used for measurement and characterization in fields such as mechanics, medicine, materials science, and industry. However, in many applications, limitations in imaging system acquisition time or radiation dose necessitate sampling under sparse angle conditions. Due to the inability to meet the integrity requirements of Tuy-Smith data, the reconstruction results of traditional methods (such as the classical filtered back projection method, FBP) are also unsatisfactory. Therefore, it is necessary to investigate reconstruction methods to improve the accuracy of CT characterization under sparse angle conditions.

[0003] In recent years, convolutional neural networks have gradually shown significant advantages in the field of image super-resolution reconstruction. If deep neural networks can be used to optimize the reconstruction results of ultra-sparse sampled CT, they can be applied to accurate CT characterization.

[0004] Existing deep learning approaches primarily focus on optimizing two key information domains (the sine wave domain and the tomographic domain) involved in CT reconstruction, achieving positive results. However, these approaches fail to consider the representation of image details. The sine wave is essentially an integral of the tomographic image along the X-ray direction, thus high-frequency information of local details is easily lost. If high-frequency information cannot be captured in the sine wave domain, the lost details cannot be recovered during the subsequent optimization process in the tomographic domain, leading to distorted reconstruction results. This limits its potential application in accurate CT characterization. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method, system, and medium for constructing a neural network suitable for sparse angular CT characterization.

[0006] According to the present invention, a method, system, and medium for constructing a neural network suitable for sparse angular CT characterization are provided, the scheme is as follows:

[0007] Firstly, a method for constructing a neural network suitable for sparse angular CT representation is provided, the method comprising:

[0008] Step S1: Construct a deep neural network G1 in the projected sinusoidal domain, and use the neural network G1 to repair the projected sinusoidal graph under sparse angles to obtain the sinusoidal graph with complete angles;

[0009] Step S2: Convert the full-angle sine wave diagram into a tomographic diagram using a filtered back-projection algorithm;

[0010] Step S3: Extract high-frequency information from the obtained fault map;

[0011] Step S4: Construct a deep neural network G2 in the fault domain, and use the convolutional neural network G2 to perform further super-resolution reconstruction of the reconstructed fault map;

[0012] Step S5: Based on the constructed high-frequency information constrained neural network framework, train the network using the gradient descent method.

[0013] Preferably, step S1 includes: the input of the neural network G1 is a sparsely sampled projection sine graph, and the projection sine graph is mapped to a full-angle sine graph through the G1 network.

[0014] Preferably, step S3, high-frequency information extraction, includes the following two steps:

[0015] Step S3.1: Perform gradient transformation on the reconstructed fault map;

[0016] Step S3.2: Perform Radon transform on the obtained gradient image to obtain the sinusoidal domain representation of the high-frequency information of the tomographic image.

[0017] Preferably, step S5, training loss calculation, includes:

[0018] Step S5.1: Calculate the sinusoidal content loss l1: the mean square error between the completed angle sinusoidal graph generated by the neural network G1 and the true label;

[0019] Step S5.2: Calculate the high-frequency information loss l2: the mean square error between the high-frequency information extraction result and the true label;

[0020] Step S5.3: Calculate the fault content loss l3: the mean square error between the labels of the fault map generated by the neural network G2 and the real fault map;

[0021] Step S5.4: High-frequency information constrains the final loss function of the neural network. loss The target is defined as: l loss =l1+l2+l3;

[0022] Step S5.5: Continuously calculate l during training. loss The value of is used to update the network weights through gradient descent, and the training is repeated until the loss is minimized.

[0023] Secondly, a system for constructing a neural network suitable for sparse angular CT characterization is provided, the system comprising:

[0024] Module M1: Construct a deep neural network G1 in the projected sinusoidal domain. Use the neural network G1 to repair the projected sinusoidal graph under sparse angles to obtain the sinusoidal graph with complete angles.

[0025] Module M2: Converts a full-angle sine wave into a tomographic image using a filtered back-projection algorithm;

[0026] Module M3: Extracts high-frequency information from the obtained tomographic images;

[0027] Module M4: Constructs a deep neural network G2 in the fault domain, and performs further super-resolution reconstruction of the reconstructed fault map through the convolutional neural network G2;

[0028] Module M5: Based on the constructed high-frequency information constraining the neural network framework, the network is trained using the gradient descent method.

[0029] Preferably, the module M1 includes: the input of the neural network G1 is a sparsely sampled projection sine graph, and the projection sine graph is mapped to a full-angle sine graph through the G1 network.

[0030] Preferably, the high-frequency information extraction of module M3 includes the following two steps:

[0031] Module M3.1: Performs gradient transformation on the reconstructed fault map;

[0032] Module M3.2: Performs Radon transform on the obtained gradient image to obtain a sinusoidal domain representation of the high-frequency information of the tomographic image.

[0033] Preferably, the calculation of the training loss of module M5 includes:

[0034] Module M5.1: Calculates the sinusoidal content loss l1: the mean square error between the completed angle sine graph generated by neural network G1 and the true label;

[0035] Module M5.2: Calculates the high-frequency information loss l2: the mean square error between the extracted high-frequency information and the true label;

[0036] Module M5.3: Calculate the fault content loss l3: the mean square error between the labels of the fault map generated by the neural network G2 and the real fault map;

[0037] Module M5.4: The final loss function of the high-frequency information constrained neural network loss The target is defined as: l loss =l1+l2+l3;

[0038] Module M5.5: Continuously calculates l during training. loss The value of is used to update the network weights through gradient descent, and the training is repeated until the loss is minimized.

[0039] Thirdly, a computer-readable storage medium storing a computer program is provided, characterized in that the computer program, when executed by a processor, implements the steps of the method.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. Based on existing deep learning concepts, this invention adds high-frequency information constraints, which effectively improves the quality of CT reconstruction tomography under sparse angle conditions;

[0042] 2. This invention can achieve high-quality CT reconstruction (standardized covariance image similarity index C > 90%) with a very small number of sparse projections (<10);

[0043] 3. The present invention has good effects in detail restoration and artifact suppression. Measured by the image similarity index (C), the reconstruction quality of this method is improved by more than 42% compared with conventional reconstruction methods such as filtered back projection, and by more than 20% compared with conventional neural network reconstruction methods. Attached Figure Description

[0044] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0045] Figure 1 The network flowchart of the high-frequency information constraint neural network provided in the embodiments of the present invention;

[0046] Figure 2 The neural network structure is constrained by high-frequency information;

[0047] Figure 3 Images of randomly generated particle models;

[0048] Figure 4 This is the result of filtered backprojection reconstruction;

[0049] Figure 5 The reconstruction results are based on the optimization approach of conventional deep learning;

[0050] Figure 6 This is the reconstruction result of the method of the present invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0052] This invention provides a method for constructing a neural network suitable for sparse angular CT representation, referring to... Figure 1 As shown, the specific steps include the following:

[0053] Step S1: The source of degradation in sparse angular CT sampling tomographic images is the loss of information in the projected sinusoidal domain. Therefore, this invention first considers repairing the missing information in the sparse angular CT in the projected sinusoidal domain. A deep neural network G1 is deployed in the projected sinusoidal domain to repair the information lost in sparse sampling, thereby obtaining a full-angle sinusoidal image.

[0054] The network's input is a sparsely sampled projected sine wave, which is then mapped to a high-quality sine wave across all angles using a G1 network. The mean squared error (MSE) loss between the sine wave generated by G1 and the ground truth label is called the sinusoidal content loss l1.

[0055] Step S2: The tomographic image is the final presentation result of the CT representation. Therefore, in order to convert the repaired sinusoidal image to the tomographic domain, this invention uses a classic filtered backprojection algorithm. The filtered backprojection algorithm first filters the projections at each acquisition angle, and then backprojects them along various directions to obtain the tomographic image of the object being detected.

[0056] Step S3: Extract high-frequency information from the obtained fault map.

[0057] First, a gradient transformation is performed on the tomographic image, that is, the partial derivatives of the tomographic image along the x and y directions are taken respectively, and then the magnitude is calculated to obtain the gradient image. Next, the Radon transform is performed on the obtained gradient image, that is, the line integral of the gradient image along all projection angles on the plane is performed. The MSE loss between the obtained high-frequency information extraction result and the true label is called the high-frequency information loss l2.

[0058] Step S4: To further optimize the quality of the image converted to the tomographic domain, this invention constructs a deep neural network G2 in the tomographic domain. The convolutional neural network G2 is used to perform super-resolution reconstruction of the reconstructed tomographic image, thereby eliminating noise introduced during the filtered back-projection reconstruction process and improving the reconstruction quality.

[0059] The goal of neural network G2 is to further optimize the filtered backprojection reconstruction results and finally output the reconstructed result. The MSE loss between the reconstruction result output by neural network G2 and the true label is called the tomographic content loss l3.

[0060] Step S5: Based on the constructed high-frequency information constrained neural network framework, train the network using the gradient descent method.

[0061] Training loss calculation includes:

[0062] Step S5.1: Calculate the sinusoidal content loss l1: the mean square error between the completed angle sinusoidal graph generated by neural network G1 and the true label. Mathematically, the sinusoidal content loss can be expressed as:

[0063]

[0064] Where m is the input sparse angular projection sine curve. It is a true full-angle projected sine curve, where G1 represents the mapping of the sine domain and N represents the number of training data pairs.

[0065] Step S5.2: Calculate the high-frequency information loss l2: the mean square error between the extracted high-frequency information and the true label. Mathematically, the high-frequency information loss can be expressed as:

[0066]

[0067] Where m is the input sparse angular projection sine curve. These are real labels; G1 represents the mapping of the sine domain, and S... hf For high-frequency information extraction operations, N represents the number of training data pairs.

[0068] Step S5.3: Calculate the fault content loss l3: the mean square error between the labels of the fault map generated by the neural network G2 and the actual fault map. Mathematically, the fault content loss can be expressed as:

[0069]

[0070] Where m is the input sparse angular projection sine curve. This is a true fault map, where G2 represents the mapping of the fault domain, and f bp This represents filtered backprojection reconstruction, and N represents the number of training data pairs.

[0071] Step S5.4: High-frequency information constrains the final loss function of the neural network. loss The target is defined as: l loss = l1 + l2 + l3.

[0072] Step S5.5: Continuously calculate l during training. loss The value of is used to update the network weights through gradient descent, and the training is repeated until the loss is minimized.

[0073] The present invention will now be described in more detail.

[0074] This invention provides a method for constructing a neural network suitable for sparse angular CT representation, referring to... Figure 1 As shown, the details are as follows:

[0075] (1) Assuming m is the input sparse angle sine graph, mapping G1 converts the sparse angle sine graph into a fully sampled sine graph m′∶m′=G1(m);

[0076] Where m represents the sparse angular projection sine graph of the input; m′ represents the full-angle sine graph of the G1 mapping;

[0077] (2) The full-angle sine curve m′ is transformed onto the fault domain using a filtered back-projection algorithm to obtain the reconstruction result n′∶n′=f bp (m′). Where m′ represents the full-angle sine curve of the G1 mapping; n′ represents the result mapped to the fault domain; f bp (.) indicates filtered back projection reconstruction;

[0078] (3) Next, extract high-frequency information from n′: m′1∶m′1=S hf (n′). Where m′1 represents the high-frequency information extraction result of the sine wave; S hf (.) indicates a high-frequency information extraction operation;

[0079] (4) Then the neural network G2 performs super-resolution reconstruction of the tomographic image: n = G2(f bp (n′)). Where n′ represents the result mapped to the fault domain; n represents the output super-resolution reconstructed fault map;

[0080] (5) Finally, the loss of the high-frequency information constrained neural network consists of three parts: sinusoidal content loss l2, high-frequency information loss l2, and fault content loss l2. G1 and G2 are updated relative to each other using gradient descent.

[0081] The deep neural network results used in this invention are as follows: Figure 2As shown, G1 and G2 are composed of convolutional neural networks. The encoder consists of 5 convolutional layers with a kernel size of 4×4 and a stride of 2. The number of channels in each layer is 64, 128, 256, 512, and 512, respectively. The activation function for all five convolutional layers is ReLU. The decoder consists of 5 deconvolutional layers. The first four convolutional layers have a kernel size of 4×4 and a stride of 2, with the number of channels being 512, 256, 128, and 64, respectively. The fifth layer has a kernel size of 4×4, a stride of 2, and 3 channels. The activation function for all five convolutional layers is ReLU. The output size of the high-frequency information constrained neural network is the same as the input size, both being 256×256×3.

[0082] Finally, numerical experiments verified the effectiveness of the invention. Figure 3 The results are from a 180° full-angle fault reconstruction. Figure 4 This is the reconstruction result of the sparse angle filter back projection algorithm. Figure 5 The results are reconstructed using conventional deep learning methods. Figure 6 The image shows the reconstruction results of this invention. It can be seen that this invention effectively maintains the accuracy of detail information while suppressing sparse angular artifacts.

[0083] This invention provides a method, system, and medium for constructing a neural network suitable for sparse angle CT characterization. Addressing the reconstruction problem of sparse angle CT, it optimizes the quality of reconstruction results by constraining the neural network with high-frequency information, thereby making it suitable for precise CT characterization applications.

[0084] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0085] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for constructing a neural network suitable for sparse angular CT representation, characterized in that, include: Step S1: Construct a deep neural network G1 in the projected sinusoidal domain, and use the neural network G1 to repair the projected sinusoidal graph under sparse angles to obtain the sinusoidal graph with complete angles; Step S2: Convert the full-angle sine wave diagram into a tomographic diagram using a filtered back-projection algorithm; Step S3: Extract high-frequency information from the obtained fault map; Step S4: Construct a deep neural network G2 in the fault domain, and use the convolutional neural network G2 to perform further super-resolution reconstruction of the reconstructed fault map; Step S5: Based on the constructed high-frequency information constrained neural network framework, train the network using the gradient descent method; The training loss calculation in step S5 includes: Step S5.1: Calculate the sinusoidal content loss The mean square error between the completed angle sine curve generated by the G1 neural network and the true label; Step S5.2: Calculate the high-frequency information loss The mean squared error between the extracted high-frequency information and the actual labels; Step S5.3: Calculate the content loss of the fault. The mean square error between the labels of the tomographic images generated by the G2 neural network and the actual tomographic images; Step S5.4: High-frequency information constrains the final loss function of the neural network The target is defined as: ; Step S5.5: Continuously calculate during training. The value of is used to update the network weights through gradient descent, and the training is repeated until the loss is minimized.

2. The method for constructing a neural network suitable for sparse angular CT representation according to claim 1, characterized in that, Step S1 includes: the input of the neural network G1 is a sparsely sampled projection sine graph, and the projection sine graph is mapped to a full-angle sine graph through the G1 network.

3. The method for constructing a neural network suitable for sparse angular CT representation according to claim 1, characterized in that, The high-frequency information extraction step S3 includes the following two steps: Step S3.1: Perform gradient transformation on the reconstructed fault map; Step S3.2: Perform Radon transform on the obtained gradient image to obtain the sinusoidal domain representation of the high-frequency information of the tomographic image.

4. A system for constructing a neural network suitable for sparse angular CT representation, characterized in that, include: Module M1: Construct a deep neural network G1 in the projected sinusoidal domain. Use the neural network G1 to repair the projected sinusoidal graph under sparse angles to obtain the sinusoidal graph with complete angles. Module M2: Converts a full-angle sine wave into a tomographic image using a filtered back-projection algorithm; Module M3: Extracts high-frequency information from the obtained tomographic images; Module M4: Constructs a deep neural network G2 in the fault domain, and performs further super-resolution reconstruction of the reconstructed fault map through the convolutional neural network G2; Module M5: Based on the constructed high-frequency information constraining the neural network framework, the network is trained using the gradient descent method; The training loss calculation for module M5 includes: Module M5.1: Calculate sinusoidal content loss The mean square error between the completed angle sine curve generated by the G1 neural network and the true label; Module M5.2: Calculate high-frequency information loss The mean squared error between the extracted high-frequency information and the actual labels; Module M5.3: Calculate Fault Content Loss The mean square error between the labels of the tomographic images generated by the G2 neural network and the actual tomographic images; Module M5.4: The final loss function of a high-frequency information-constrained neural network The target is defined as: ; Module M5.5: Continuously calculates during training. The value of is used to update the network weights through gradient descent, and the training is repeated until the loss is minimized.

5. The system for constructing a neural network suitable for sparse angular CT characterization according to claim 4, characterized in that, The module M1 includes: the input of the neural network G1 is a sparsely sampled projection sine graph, and the projection sine graph is mapped to a full-angle sine graph through the G1 network.

6. The system for constructing a neural network suitable for sparse angular CT characterization according to claim 4, characterized in that, The high-frequency information extraction of module M3 includes the following two steps: Module M3.1: Performs gradient transformation on the reconstructed fault map; Module M3.2: Performs Radon transform on the obtained gradient image to obtain a sinusoidal domain representation of the high-frequency information of the tomographic image.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images