A dual-view industrial CT tomography online reconstruction method based on a latent space conditional diffusion model

Abstract

The present application belongs to the field of CT tomographic reconstruction technology and artificial intelligence, and discloses a double-view industrial CT tomographic online reconstruction method based on a latent space conditional diffusion model. In industrial production, products often need to be quickly and non-destructively detected, but traditional full-angle CT scanning takes a long time and cannot meet the requirements of production efficiency on the production line, and due to the existence of rotating mechanisms, it is difficult to be deployed online. The present application designs an overall network structure based on a latent space conditional diffusion model, and embeds a double-view orthogonal projection input condition embedding module as a conditional embedding to guide the diffusion model that has learned the tomographic prior distribution to generate corresponding high-resolution tomographic images. The present application can generate high-quality tomographic images only by using two orthogonal view projections of 0 degrees and 90 degrees, thereby reducing the complexity of the CT device, greatly speeding up the scanning speed, and facilitating online deployment.

Description

Technical Field

[0001] This invention belongs to the fields of CT tomographic reconstruction technology and artificial intelligence, and particularly relates to a dual-view industrial CT tomographic online reconstruction method based on a latent space conditional diffusion model. Background Technology

[0002] With the success of CT (Computed Tomography) technology in the medical field, its applications in industry are gradually gaining attention and being explored. The industry is beginning to recognize that CT technology can be used not only for medical diagnosis but also for non-destructive testing, quality control, and materials analysis in industrial settings. It offers high resolution, 3D reconstruction, and internal structure analysis capabilities, helping industry identify and resolve defects, cracks, foreign objects, holes, and other problems in components. Furthermore, CT technology can be used for fault diagnosis and analysis, providing guidance for the maintenance and improvement of industrial equipment.

[0003] However, traditional CT technology also has some limitations in industrial applications. First, CT scans generally require a long time to acquire projected images of the scanned object from all angles, which can impact production efficiency. Furthermore, when performing CT inspections on industrial products, a rotating mechanism typically needs to be designed separately based on the product's shape, which undoubtedly increases the complexity of the CT equipment and makes online deployment difficult.

[0004] To reduce CT scan time and improve CT detection efficiency, researchers have proposed sparse angle CT reconstruction methods. These methods involve a larger angle interval between adjacent projections to form a sparser projection sequence, thus reducing CT scan time. However, because this method yields a limited number of projections, the reconstructed tomography often contains artifacts. To address this issue, classic artifact removal methods incorporate manually designed regularization terms into the reconstruction formula, providing prior knowledge of the tomographic images for CT reconstruction. With the development of artificial intelligence, new neural network-based solutions for sparse angle reconstruction have emerged. Patent CN116843825A reports a progressive CBCT sparse angle reconstruction method that uses a three-stage fully convolutional neural network to obtain artifact-free tomography from sparse angle tomography images containing artifacts. Patent CN114004907A reports a fast pipelined computed tomography imaging method based on deep learning, which uses a fully convolutional neural network to predict the complete projection sequence using sparse projection sequences, and then performs tomographic reconstruction using traditional methods. While such CT reconstruction schemes can address artifact issues to some extent, they still require a large number of projections, necessitating rotational scanning and resulting in slower speeds. Furthermore, the aforementioned fully convolutional networks have limitations in detail generation and exhibit limited generalization ability. This invention utilizes extreme orthogonal dual-view projection, requiring significantly less projection information compared to the sparse angle reconstruction methods reported in the aforementioned patents. The CT equipment no longer needs rotational scanning, greatly accelerating the scanning speed and facilitating online deployment. Since CT reconstruction relies heavily on dual-view projection information, this invention designs a latent space conditional diffusion model to achieve dual-view CT reconstruction. Compared to existing convolutional network model schemes, the scheme designed in this invention generates higher image quality and exhibits stronger generalization ability. Summary of the Invention

[0005] This invention provides a dual-view industrial CT tomographic online reconstruction method based on a latent space conditional diffusion model. It reconstructs a single tomographic image using only two fan-beam / parallel-beam projections from the 0-degree and 90-degree orthogonal viewing angles of the object being scanned. Because only two projections are used, there is no need to design a rotation mechanism on the CT equipment, facilitating online deployment on industrial production lines.

[0006] The basic idea of ​​the diffusion model is to progressively add noise to the tomographic images in the dataset until the tomographic images approximate Gaussian noise. Then, a noise prediction network learns from the progressively added noise. After learning, the diffusion model can start from a Gaussian noise image, progressively denoise it through the noise prediction network, and finally obtain a tomographic image close to the image in the dataset. Considering the model's running speed, this invention uses a latent space representation; that is, the above diffusion process does not take place in the image space, but in the latent space encoded by the encoder network. To ensure the quality of the final reconstructed tomographic image, this invention uses dual-view projection information as a conditional input to the diffusion model, controlling the output of the noise prediction network, thereby obtaining the tomographic image corresponding to the dual-view projection.

[0007] The technical solution of this invention:

[0008] A dual-view industrial CT tomographic reconstruction method based on a latent space conditional diffusion model is proposed. This method achieves tomographic reconstruction with minimal viewpoints through the fusion embedding of dual-view projections and the latent space conditional diffusion model. The latent space diffusion model comprises four parts: an encoder network, a noise prediction network, a conditional embedding module, and a decoder network, collectively forming a rapid CT tomographic generation network. It can be trained offline on large datasets and then rapidly completes tomographic reconstruction tasks online. The specific steps are as follows:

[0009] Step 1: Construct the tomographic dataset

[0010] Obtain a sufficient number of similar samples to be scanned, perform full-angle CT scans on each sample to obtain tomographic images, and construct a dataset for subsequent parameter training of the CT tomographic rapid generation network.

[0011] Step 2: Construct a latent space conditional diffusion model

[0012] (1) Constructing the encoder network: The encoder network is responsible for generating the latent space representation of the tomographic image, encoding the original tomographic image into feature vectors in the latent space, so that the subsequent noise prediction network can be trained and predicted in the latent space.

[0013] The encoder network mainly consists of convolutional layers, multiple residual blocks, and multiple downsampling layers. The input of the convolutional layer is the original tomographic image x, with 1 input channel and 128 output channels. This convolutional layer increases the number of channels in the input data, enhancing the network's feature representation capability. The residual blocks and downsampling layers are placed alternately, forming a continuous residual downsampling structure in the form of "residual block-downsampling layer-residual block-downsampling layer-residual block-downsampling layer". This structure is used to extract features from the data with increased channel count, and the output is a latent space feature vector z containing the original tomographic image.

[0014] (2) Constructing a noise prediction network and a conditional embedding module: The noise prediction network is responsible for predicting the noise contained in the latent space feature vector of the current time step input, so as to remove the noise from the latent space feature vector of the current time step and obtain the input feature vector required for the next denoising step; The conditional embedding module is responsible for converting the dual-view projection data into the condition c corresponding to the latent space feature vector z.

[0015] The noise prediction network backbone mainly consists of convolutional layers, residual blocks, downsampling layers, attention layers, and upsampling layers. A temporal encoding block is also present at the input for temporal encoding. The convolutional layers take latent space feature vectors and embedding conditions as input, with the number of channels being the sum of the number of channels in the latent space feature vectors and the number of channels in the embedding conditions. The output channel count is set to 256. This convolutional layer increases the number of channels in the input data, enhancing the network's feature representation capability. Residual blocks and downsampling layers are alternately placed to form a continuous residual downsampling structure, while residual blocks and upsampling layers are alternately placed to form a continuous residual upsampling structure. When the data passes through the continuous residual downsampling structure, its resolution... The resolution will decrease, but it will be restored after passing through a continuous residual upsampling structure. The network's representational ability can be improved by adding an attention layer at a specific resolution. The overall form of the noise prediction network backbone is "residual block - downsampling layer - residual block - attention layer - downsampling layer - residual block - attention layer - downsampling layer - residual block - attention layer - upsampling layer - residual block - attention layer - upsampling layer - residual block - upsampling layer". The temporal coding block is located outside the network backbone. The output of each residual block is added to the temporal coding output and then fed into the upsampling and downsampling layers, which work together to affect the subsequent upsampling and downsampling processes.

[0016] The entire noise prediction network inputs the latent space feature vector z. t Where t = 1, 2, 3, ..., T, condition c and the current time step t, and its output is noise ε with the same shape as the input latent space feature vector. θ (z t ,c,t), where ε θ Represents a noise prediction network; based on the predicted noise results, z is used in equation (1.1) to... t Noise removal in the process yields the denoising result z in one step. t-1 :

[0017]

[0018] Where, α t For the parameter to be set, α t The value of is determined by the noise addition process, see equation (1.3); e is the sum of z sampled from the standard normal distribution. t Noise vectors of the same shape; σ tThe effect applied to e is used to add diversity to the denoising process, thereby enhancing the network's expressive power. σ t The choice of value ensures

[0019] The conditional embedding module uses the filtered back projection (FBP) algorithm to coarsely reconstruct the dual-view orthogonal projection data obtained from CT scans. This reconstructed image is the condition c that is input into the noise prediction network along with the latent space feature vector.

[0020] (3) Constructing a decoder network: The decoder network is responsible for decoding the latent space feature vector z after continuous denoising steps from the latent space to restore the tomographic image;

[0021] The composition of the decoder network is basically the same as that of the encoder network. The difference is that the downsampling layer is replaced by the upsampling layer. The latent space feature vector z is passed through a continuous residual upsampling structure to obtain high-resolution data with 128 channels. Then, it passes through a separate convolutional layer to reduce the number of channels to 1, thus restoring the tomographic image.

[0022] Step 3: Train the latent space conditional diffusion model

[0023] In the latent space conditional diffusion model, the encoder network, noise prediction network, and decoder network are three trainable parts with network parameters. These three parts have a corresponding training order: the encoder network and decoder network are trained first. After training, the parameters of the encoder network and decoder network are fixed and used as the basis for training the noise prediction network.

[0024] (1) Training the encoder network and decoder network

[0025] Since the encoder network and the decoder network encode the tomographic image into a latent space feature vector and recover the tomographic image from the latent space feature vector, respectively, that is, to obtain the accurate latent space feature vector from the tomographic image and to accurately recover the corresponding tomographic image from the latent space feature vector; the encoder network and the decoder network are combined for training, and the combination method is that the latent space feature vector output by the encoder network is used as the input latent space feature vector of the decoder network.

[0026] Training process: First, randomly select an image x from the tomographic images obtained in step 1 and input it into the combined encoder and decoder network. The output image of the combined network is named q, and the supervision of the output image is still x. The loss function used consists of MSE loss and KL divergence loss, as shown in equation (1.2):

[0027] LOSS1=MSE(x,q)+KL(x,q) (1.2) Using the same tomographic image as both input and supervision, after training, the decoder network can accurately reconstruct the original tomographic image from the feature vectors generated by the encoder network.

[0028] Using a large number of tomographic images in the dataset established in step 1, repeat the above training process to train the encoder network and decoder network together; when LOSS1 converges, the network training is completed. At this time, the parameters of the encoder network and decoder network remain unchanged and are used for the training of the subsequent noise prediction network.

[0029] (2) Training the noise prediction network

[0030] The noise prediction network is responsible for predicting the noise contained in the latent space feature vector of the input at the current time step, so as to remove the noise from the feature vector of the input at the current time step and obtain the latent space feature vector needed for the next denoising step. Due to the progressive addition and progressive denoising characteristics of the latent space conditional diffusion model, a total number of steps T needs to be specified before training begins. The value of T is selected as 1000.

[0031] Training process: First, randomly select a tomographic image x from the tomographic images obtained in step 1, and input the tomographic image x into the encoder network trained in the previous step to obtain the latent space feature vector z corresponding to the tomographic image x; randomly select a time step t from 1 to T, and use equation (1.3) to add noise to the latent space feature vector z to obtain the noisy feature vector z. t ;

[0032]

[0033] in, α here t In equation (1.1), α t It is the same; given a total number of steps T, the sequence α1, α2, α3, ..., α T Given a pre-defined string of positive numbers, and it must satisfy the following conditions: Therefore, the sequence is set as an arithmetic descending sequence; e is a noise vector with the same shape as the latent space feature vector z, sampled from the standard normal distribution; simultaneously, by embedding the tomographic image x into the conditional module, its corresponding dual-view orthogonal projection condition c can be obtained; z t The input noise prediction network, c,t, yields the predicted output ε. θ (z t ,c,t),ε θ Represents a noise prediction network; the noise component e is compared with the noise prediction network output ε. θ (z t The MSE loss shown in equation (1.4) is used to control network training.

[0034] LOSS2=MSE(e,ε θ (z t ,c,t)) (1.4)

[0035] The noise prediction network is trained using the tomographic images from the dataset established in step 1, and the above training process is repeated. After training, the noise prediction network can predict the input feature vector z with relatively high accuracy. t The noise components contained in it; thus, equation (1.1) is used to determine z. t Denoising to obtain z t-1 ;

[0036] Step 4: Online Operation Process

[0037] When deploying online, first, obtain the dataset of tomographic images of the sample to be scanned according to step 1. Second, build the latent space conditional diffusion model according to steps 2 and 3, and use the dataset obtained in step 1 to train the parameters of the three network parts in the latent space conditional diffusion model. After training, the network parameters are fixed and used for online CT reconstruction.

[0038] Whenever a new product of the same type requires a CT scan, the CT equipment is used to obtain the projections of fan-beam or parallel-beam views at two orthogonal angles of 0 degrees and 90 degrees. This projection data is then used to input the conditional embedding module to generate the conditions required by the noise prediction network. After obtaining the required conditions, a Gaussian random noise feature vector z is first generated. T Then, the latent space feature vector, conditional and temporal features are iteratively input into the noise prediction network to remove noise, with a total of T iterations, to obtain a noise-free feature vector z. z is then input into the decoder network to finally obtain a noise-free dual-view orthogonal projection reconstructed tomographic image.

[0039] The beneficial effects of the present invention are as follows: (1) The present invention uses only two orthogonal perspective projections to complete the tomographic reconstruction task, without the need for a rotating mechanism, resulting in a qualitative leap in scanning speed, which can meet the requirements of industrial production for scanning efficiency. (2) Since no rotating mechanism is required, there is no need to design special scanning methods for special products, which facilitates online deployment on the production line. Attached Figure Description

[0040] Figure 1 This is an architecture diagram of the dual-view industrial CT tomographic online reconstruction method based on the latent space conditional diffusion model of this invention.

[0041] Figure 2 This is a structural diagram of the encoder network and decoder network of the present invention.

[0042] Figure 3 This is a structural diagram of the noise prediction network in this invention. Detailed Implementation

[0043] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0044] After being fully trained with tomographic data of a certain product, this invention can directly obtain its complete tomographic image using only the orthogonal projections of the workpiece at 0 degrees and 90 degrees. The specific implementation method is as follows:

[0045] First, construct the tomographic dataset according to the description in Step 1. Contact manufacturers who need their products inspected via CT scan to obtain the number of products they require, approximately 100. Then, perform full-angle CT scans on these products to obtain high-quality, complete tomographic images of each product, which will be used for subsequent parameter training of the network.

[0046] Secondly, the latent space conditional diffusion model is constructed according to the description in step 2. First, an encoder network is built, concatenating residual blocks and downsampling layers into a continuous residual downsampling structure as described in step 2. Then, the output of the convolutional layer is connected to the input of this structure, completing the encoder network. Next, a noise prediction network and a conditional embedding module are constructed. Residual blocks, attention layers, upsampling layers, and downsampling layers are concatenated into the backbone of the noise prediction network as described in step 2. Additionally, a temporal encoding block is needed to encode temporal information. The encoded result is added to the output of each residual block and then input into subsequent network layers. The conditional embedding module uses the FBP algorithm to convert the dual-view orthogonal projection obtained from CT scans into conditions for input to the noise prediction network. Finally, a decoder network is constructed. The decoder network reverses the layer order in the continuous residual downsampling structure of the encoder network and replaces the downsampling layers with upsampling layers to form a continuous residual upsampling structure. The output of this structure is connected to the final convolutional layer, completing the decoder network.

[0047] Next, the latent space conditional diffusion model is trained according to the description in step 3. The latent space conditional diffusion model has three trainable parts with network parameters: the encoder network, the noise prediction network, and the decoder network. These three parts have a corresponding training order: first train the encoder network and the decoder network. After training, the parameters of the encoder network and the decoder network can be fixed, and the noise prediction network is trained based on this. For training the encoder network and the decoder network, the encoder output is first connected to the decoder network input, and the two networks are trained together. Specifically, the data from 100 products is divided into 90 sets as the training set for training the network and 10 sets as the validation set to test the network training effect. The joint training process for the encoder network and the decoder network described in step 3 is used for training. Because the distribution consistency of the fault lines of the same type of workpiece is relatively strong, the training process is relatively fast. After 10 hours of training using an RTX 3090 GPU, the loss function usually tends to converge, at which point the encoder network and the decoder network training is complete.

[0048] After training the encoder and decoder networks, their network parameters can be fixed. Next, train the noise prediction network. Divide the data from 100 artifacts into 90 sets as the training set and 10 sets as the validation set to evaluate the network's training performance. Follow the training procedure for the noise prediction network described in step 3. Training can be stopped when the validation set shows good results and no unreasonable artifacts. Based on experience, on an RTX 3090 GPU, one training task can be completed in 20 hours.

[0049] After all networks have been trained, online CT scans of the new product can be performed following the online operation process described in step 4.

Claims

1. A dual-view industrial CT tomographic online reconstruction method based on a latent space conditional diffusion model, characterized in that, Minimal viewpoint tomographic reconstruction is achieved through dual-view projection fusion embedding and a latent space conditional diffusion model. The latent space diffusion model comprises four parts: an encoder network, a noise prediction network, a conditional embedding module, and a decoder network, collectively forming a rapid CT tomographic generation network. This rapid CT tomographic generation network is trained offline on a large amount of data and then performs rapid tomographic reconstruction online. The specific steps are as follows: Step 1: Construct the tomographic dataset Obtain a sufficient number of similar samples to be scanned, perform full-angle CT scans on each sample to obtain tomographic images, and construct a dataset for subsequent parameter training of the CT tomographic rapid generation network. Step 2: Construct a latent space conditional diffusion model (1) Constructing an encoder network: The encoder network is responsible for generating the latent space representation of the tomographic image, encoding the original tomographic image into a feature vector in the latent space, so that the subsequent noise prediction network can be trained and predicted in the latent space; The encoder network mainly consists of a convolutional layer, multiple residual blocks, and multiple downsampling layers; The input of the convolutional layer is the original tomographic image x, the number of input channels is 1, and the number of output channels is set to 128; The residual blocks and downsampling layers are placed alternately and combined into a continuous residual downsampling structure in the form of "residual block-downsampling layer-residual block-downsampling layer-residual block-downsampling layer", which is used to extract features from the data with increased channel number and output a latent space feature vector z containing the original tomographic image; (2) Constructing a noise prediction network and a conditional embedding module: The noise prediction network is responsible for predicting the noise contained in the latent space feature vector of the current time step input, so as to remove the noise from the latent space feature vector of the current time step and obtain the input feature vector required for the next denoising step. The conditional embedding module is responsible for converting dual-view projection data into condition c corresponding to the latent space feature vector z; The backbone of the noise prediction network mainly consists of convolutional layers, residual blocks, downsampling layers, attention layers, and upsampling layers. A temporal encoding block is also present at the input for temporal encoding. The convolutional layers take latent space feature vectors and embedding conditions as input, with the number of channels being the sum of the number of channels in the latent space feature vectors and the number of channels in the embedding conditions. The number of output channels is set to 256. Residual blocks and downsampling layers are alternately placed to form a continuous residual downsampling structure, while residual blocks and upsampling layers are alternately placed to form a continuous residual upsampling structure. When data passes through the continuous residual downsampling structure, its resolution decreases; when it passes through the continuous residual upsampling structure, its resolution is restored. Attention layers are added at specific resolutions to improve the network's representational power. The time-coded block is located outside the backbone of the noise prediction network. The output of each residual block is added to the output of the time-coded block and then fed into the corresponding upsampling and downsampling layers, which together affect the subsequent upsampling and downsampling processes. The entire noise prediction network inputs the latent space feature vector z. t Where t = 1, 2, 3, ..., T, condition c and the current time step t, and its output is noise ε with the same shape as the input latent space feature vector. θ (z t ,c,t), where ε θ Represents a noise prediction network; based on the predicted noise results, z is used in equation (0.1) to... t Noise removal in the process yields the denoising result z in one step. t-1 : Where, α t For the parameter to be set, α t The value of is determined by the noise addition process, see equation (0.3); e is the sum of z sampled from the standard normal distribution. t Noise vectors of the same shape; σ t The effect applied to e is used to add diversity to the denoising process, thereby enhancing the network's expressive power. σ t The choice of value ensures The conditional embedding module uses the filtered back projection (FBP) algorithm to coarsely reconstruct the dual-view orthogonal projection data obtained from CT scans. This reconstructed image is the condition c that is input into the noise prediction network along with the latent space feature vector. (3) Constructing a decoder network: The decoder network is responsible for decoding the latent space feature vector z after continuous denoising steps from the latent space to restore the tomographic image; The noise prediction network backbone is in the form of "residual block - downsampling layer - residual block - attention layer - downsampling layer - residual block - attention layer - downsampling layer - residual block - attention layer - upsampling layer - residual block - attention layer - upsampling layer - residual block - upsampling layer". The decoder network is basically the same as the encoder network, except that the downsampling layer is replaced by an upsampling layer. The latent space feature vector z is processed through a continuous residual upsampling structure to obtain high-resolution data with 128 channels. Then, it is processed through a separate convolutional layer to reduce the number of channels to 1, thus restoring the tomographic image. Step 3: Train the latent space conditional diffusion model. In the latent space conditional diffusion model, the encoder network, noise prediction network, and decoder network are three trainable parts with network parameters. These three parts have a corresponding training order: the encoder network and decoder network are trained first, and their parameters are fixed after training. The noise prediction network is then trained based on these parameters. (1) Training the encoder network and decoder network Since the encoder network and the decoder network encode the tomographic image into a latent space feature vector and recover the tomographic image from the latent space feature vector, respectively, that is, to obtain the accurate latent space feature vector from the tomographic image and to accurately recover the corresponding tomographic image from the latent space feature vector; the encoder network and the decoder network are combined for training, and the combination method is that the latent space feature vector output by the encoder network is used as the input latent space feature vector of the decoder network. Training process: First, randomly select an image x from the tomographic images obtained in step 1 and input it into the combined network of encoder and decoder networks. The output image of the combined network is named q, and the supervision of the output image is still x. The loss function used consists of MSE loss and KL divergence loss, as shown in equation (0.2): LOSS1 = MSE(x,q) + KL(x,q) (0.2) Using the same tomographic image as both input and supervision, the decoder network, after training, can accurately reconstruct the original tomographic image from the feature vectors generated by the encoder network. Using a large number of tomographic images in the dataset established in step 1, repeat the above training process to train the encoder network and decoder network together; when LOSS1 converges, the encoder network and decoder network training is completed. At this time, the parameters of the encoder network and decoder network remain unchanged and are used for the training of the subsequent noise prediction network. (2) Training the noise prediction network The noise prediction network is responsible for predicting the noise contained in the latent space feature vector of the input at the current time step, so as to remove the noise from the feature vector of the input at the current time step and obtain the latent space feature vector needed for the next denoising step. Due to the progressive addition and progressive denoising characteristics of the latent space conditional diffusion model, a total number of steps T needs to be specified before training begins, and the value of T is selected as 1000. Training process: First, randomly select a tomographic image x from the tomographic images obtained in step 1, and input the tomographic image x into the encoder network trained in the previous step to obtain the latent space feature vector z corresponding to the tomographic image x; randomly select a time step t from 1 to T, and use equation (0.3) to add noise to the latent space feature vector z to obtain the noisy feature vector z. t ; in, α here t α in equation (0.1) t It is the same; given a total number of steps T, the sequence α1, α2, α3, ..., α T Given a pre-defined string of positive numbers, and it must satisfy the following conditions: Therefore, the sequence is set as an arithmetic descending sequence; e is a noise vector with the same shape as the latent space feature vector z, sampled from the standard normal distribution; simultaneously, the tomographic image x is input into the conditional embedding module to obtain its corresponding dual-view orthogonal projection condition c; z t The input noise prediction network, c,t, yields the predicted output ε. θ (z t ,c,t),ε θ Represents a noise prediction network; the noise component e is compared with the noise prediction network output ε. θ (z t The MSE loss shown in equation (0.4) is used to control network training. LOSS2=MSE(e,ε θ (z t ,c,t)) (0.4) The noise prediction network is trained using the tomographic images from the dataset established in step 1, and the above training process is repeated. After training, the noise prediction network can predict the input feature vector z with relatively high accuracy. t The noise components contained in it; thus, equation (0.1) is used to determine z. t Denoising to obtain z t-1 ; Step 4: Online Operation Process When deploying online, first, obtain the dataset of tomographic images of the sample to be scanned according to step 1. Second, build the latent space conditional diffusion model according to steps 2 and 3, and use the dataset obtained in step 1 to train the parameters of the three network parts in the latent space conditional diffusion model. After training, the network parameters are fixed and used for online CT reconstruction. Whenever a new product of the same type requires a CT scan, the CT equipment is used to obtain the projections of fan-beam or parallel-beam views at two orthogonal angles of 0 degrees and 90 degrees. This projection data is then used to input the conditional embedding module to generate the conditions required by the noise prediction network. After obtaining the required conditions, a Gaussian random noise feature vector z is first generated. T Then, the latent space feature vector, conditional and temporal features are iteratively input into the noise prediction network to remove noise, with a total of T iterations, to obtain a noise-free feature vector z. z is then input into the decoder network to finally obtain a noise-free dual-view orthogonal projection reconstructed tomographic image.

Images