Unsupervised low-dose photon counting CT reconstruction method and device based on optimal transmission

Abstract

The invention provides an unsupervised low-dose photon counting CT (Computed Tomography) reconstruction method and an unsupervised low-dose photon counting CT reconstruction device based on optimal transmission. According to the method, a low-dose photon counting CT reconstruction problem is modeled as a Cantrovitch problem under a distribution consistency constraint, and the method comprises the following steps: constructing and training an optimal transmission-based unsupervised projection recovery network and constructing and training an optimal transmission-based unsupervised image noise reduction network; performing detail recovery on the photon counting CT low-dose projection by using the trained unsupervised projection recovery network to generate a denoised projection; reconstructing the denoised projection by using an image reconstruction algorithm to generate an intermediate image; and performing noise reduction on the intermediate image by using the trained unsupervised image noise reduction network to obtain a final de-noised image. According to the method, the problem that a high-quality pairing data set is difficult to obtain and a photon counting CT low-dose image is difficult to reconstruct can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of CT image processing, and in particular, to an unsupervised low-dose photon counting CT reconstruction method and device based on optimal transport. Background Art

[0002] Computed tomography (CT) has the advantage of ultra-high spatial resolution compared to other medical imaging technologies, and can present the minute tissue structures and their changes in a patient's body to doctors. With the continuous development of medical CT technology, X-ray energy spectrum CT has come into view. Compared with traditional CT and dual-energy CT systems, X-ray energy spectrum CT based on photon counting detectors can divide the continuous X-ray energy spectrum into multiple energy regions, distinguish different materials by using the attenuation characteristics of X-rays with different energies, can effectively suppress X-ray beam hardening artifacts, and obtain better imaging results.

[0003] With the increasingly wide application of CT in clinical examinations, the harmful effects of ionizing radiation caused by CT scans have received more and more attention. Low-dose CT refers to obtaining the best CT diagnostic images with as low a radiation dose as possible on the basis of no obvious decline in image quality. Currently, the most commonly used method to reduce the X-ray dose is to reduce the radiation intensity on each X-ray path. However, low-intensity X-rays will cause a "photon starvation" phenomenon at the detector end, and the noise intensity increases exponentially.

[0004] Traditional model-driven methods utilize the statistical characteristics of noise and require in-depth understanding of the statistical characteristics of projection noise. Therefore, accurately describing the statistical characteristics of noise is crucial for whether low-signal-to-noise projection data can be reconstructed into high-quality CT images. In 2006, K. Dabov et al. proposed the classical block-matching and 3D filtering (BM3D) method. According to the similarity between image blocks, image blocks with similar structures are combined into a three-dimensional array, and a denoised image is obtained through joint filtering and inverse transformation. In 2008, Wang et al. extended the PWLS algorithm to the wavelet domain and performed multi-scale analysis by applying wavelet transform to the projection image; in 2010, Zhang et al. found that after removing isolated noise points, the projection data of low-dose CT follows a Gaussian distribution, and proposed a segmentation-based adaptive filtering method to significantly improve image quality; in 2012, Zhang et al. realized projection reconstruction by solving the posterior probability of the conditional expectation, and used the expectation maximization (EM) algorithm and Gibbs sampling technology to solve the computational problem; Cho and Fessler used a linear quadratic roughness penalty method in 2015 to solve the artifact problem caused by the interaction of system model constraints in the three-dimensional reconstruction of cone-beam low-dose CT.

[0005] In recent years, using deep learning technology to solve medical imaging problems has become a major focus. In the field of low-dose CT image reconstruction, traditional deep learning methods aim to learn the reconstruction mapping from a large number of (low-dose and normal-dose) scan pairs. Deep data-driven methods can be divided into three main categories: supervised methods, semi-supervised methods, and unsupervised methods.

[0006] Supervised methods are trained with a large number of paired noisy and clear images, aiming to learn the mapping from noisy images to clean images. A classic method in the image domain is FBPConvNet, which solves the traditional convolutional inversion problem by adopting a convolutional neural network (CNNs) that incorporates physical characteristics after direct inversion. Semi-supervised methods are usually trained on unpaired noisy images or generate pseudo-labels. Zeroshot-N2N and Dn-Dp are semi-supervised methods proposed in the past two years, which have certain advantages in terms of data requirements and running speed. However, these methods have limited effects on the complex noise caused by the reduction of the number of photons, and the gap becomes more obvious as the photon noise level increases. Unsupervised denoising methods refer to denoising based on the statistical characteristics, self-similarity, or distribution of data without clean data as a reference. N2N and N2S are two classic unsupervised methods, and these two methods have relatively strict assumptions about noise. N2N assumes that the noise is independently and identically distributed, while N2S requires the simulated noise to match the actual noise. In actual operation, it is difficult to fully achieve this, thus affecting the performance and generalization ability of the model. Summary of the Invention

[0007] To solve the problems of difficult acquisition of high-quality paired datasets and low-dose image reconstruction of photon-counting CT, the present invention provides an unsupervised low-dose photon-counting CT reconstruction method and device based on optimal transport.

[0008] In a first aspect, the present invention provides an unsupervised low-dose photon-counting CT reconstruction method based on optimal transport, which models the low-dose photon-counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints, including:

[0009] Construct and train an unsupervised projection recovery network based on optimal transport and construct and train an unsupervised image denoising network based on optimal transport;

[0010] Use the trained unsupervised projection recovery network to recover the details of the low-dose projection of photon-counting CT to generate a denoised projection;

[0011] Use an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image;

[0012] Use the trained unsupervised image denoising network to denoise the intermediate image to obtain the final denoised image.

[0013] Further, during the training process of the unsupervised projection recovery network, the following formula is used as the loss function:

[0014]

[0015] where γ1 is a weight parameter, is the probability distribution of the denoised output projection, is the probability distribution of the true projection, W1 is the Wasserstein-1 distance, Y is the input noisy projection, is the unsupervised projection recovery network is the denoised output projection

[0016] Further, during the training process of the unsupervised image denoising network, the following formula is used as the loss function:

[0017]

[0018] where γ2 is a weight parameter, p X is the probability distribution of the denoised output image, is the probability distribution of the true image, W1 is the Wasserstein-1 distance, is the intermediate image, is the unsupervised image denoising network is the denoised output image

[0019] Further, both the unsupervised projection recovery network and the unsupervised image denoising network adopt the WGAN-GP framework; wherein, the generator includes a first convolutional layer, a residual channel attention module, a U-Net structure, and a second convolutional layer connected in sequence, and a skip connection is arranged between the input of the first convolutional layer and the output of the second convolutional layer; the discriminator includes a plurality of convolutional blocks stacked in sequence and two fully connected layers; the convolutional block includes a convolutional layer and a ReLU layer cascaded in sequence.

[0020] Further, the encoder in the U-Net structure includes two downsampling CNN layers, and the decoder includes two upsampling CNN layers, and each downsampling CNN layer and upsampling CNN layer are connected by a residual channel attention module.

[0021] In a second aspect, the present invention provides an unsupervised low-dose photon counting CT reconstruction device based on optimal transport, which models the low-dose photon counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints, including:

[0022] A network construction and training module, configured to construct and train an unsupervised projection recovery network based on optimal transport and construct and train an unsupervised image denoising network based on optimal transport;

[0023] A projection denoising module, configured to use the trained unsupervised projection recovery network to recover details of the low-dose projection of photon-counting CT, and generate a denoised projection;

[0024] An image reconstruction module, configured to use an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image;

[0025] An image denoising module, configured to use the trained unsupervised image denoising network to denoise the intermediate image to obtain a final denoised image.

[0026] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method described in the first aspect is implemented.

[0027] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described in the first aspect is implemented.

[0028] The beneficial effects of the present invention are as follows:

[0029] The present invention introduces the optimal transport theory into the low-dose photon-counting CT reconstruction problem, models from the perspective of distribution similarity learning. At the same time, aiming at the problem that it is difficult to obtain a high-quality paired data set, unsupervised networks are designed in the projection domain and the image domain to construct a dual-domain low-dose photon-counting CT reconstruction framework, and the noisy projection is processed in the dual domain to finally obtain a high-quality CT reconstruction image. By taking advantage of the optimal transport theory in distribution similarity learning and dividing the low-dose photon-counting CT reconstruction into two sub-problems of detail recovery and image denoising and solving them separately in the dual domain, the quality of the low-dose reconstructed image can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is one of the flow diagrams of the unsupervised low-dose photon-counting CT reconstruction method based on optimal transport provided by the embodiment of the present invention;

[0031] Figure 2 It is another flow diagram of the unsupervised low-dose photon-counting CT reconstruction method based on optimal transport provided by the embodiment of the present invention;

[0032] Figure 3 It is the structural diagram of the unsupervised optimal transport network provided by the embodiment of the present invention;

[0033] Figure 4 This is a comparison chart of the reconstruction results of the chest CT images by the method of the present invention provided in the embodiments of the present invention under three simulated low photon dose conditions;

[0034] Figure 5 This is a comparison chart of the reconstruction results of the head phantom CT images by the method of the present invention provided in the embodiments of the present invention under three actual low current conditions;

[0035] Figure 6 This is a schematic structural diagram of an unsupervised low-dose photon counting CT reconstruction device based on optimal transport provided in the embodiments of the present invention;

[0036] Figure 7 This is a structural block diagram of an electronic device provided in the embodiments of the present invention. Detailed implementation manners

[0037] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0038] Optimal transport is a way to determine the most effective way to transform one distribution into another considering a specified cost, and it is achieved by measuring the similarity between different distributions. Intuitively, the purpose of CT image reconstruction is to recover the two-dimensional and three-dimensional images of an object from projection data, which represents the line integrals of the object at different angles. This reconstruction process corresponds to a one-to-many transport process. Therefore, it is more appropriate to choose the Kantorovich problem (hereinafter referred to as the KP problem) when implementing reconstruction under the optimal transport framework. The KP problem can be expressed by the following equation:

[0039]

[0040] Among them, represents the marginal distributions of the joint distribution π, which are represented by p and q respectively. In the context of the KP problem, the joint distribution is also called the π transport plan.

[0041] For the key application scenarios of low-dose photon counting CT image reconstruction, in one embodiment, by deeply analyzing the physical process and data characteristics of low-dose CT image reconstruction, the embodiments of the present invention establish a dual-domain joint low-dose reconstruction framework - DDOT-Net. The core of this framework is to introduce distribution consistency constraints and seek the optimal solution of the KP problem. The KP problem subject to distribution consistency constraints can be expressed as follows:

[0042]

[0043] Among them, represents an approximately isometric embedding space δ represents a small constant. represents the cost matrix.

[0044] In practical applications, formula (2) can be transformed into the following optimal transport problem with convex constraints:

[0045]

[0046] Formula (3) can be interpreted as an optimal transport problem with regularization constraints. Among them, the regularization term is The parameter γ is used as a trade-off factor between the convergence rate of the regularization problem and the approximation error of the original optimal transport. Inspired by the optimal transport theory, the inventors found that the low-dose photon-counting CT reconstruction problem can be regarded as a KP problem under the distribution consistency constraint shown in formula (3), where the regularization term is used as a prior constraint and transformed into formula (4) for implementation, and there is:

[0047]

[0048] Based on this, combining Figure 1 and Figure 2 as shown, the unsupervised low-dose photon-counting CT reconstruction method based on optimal transport provided by the embodiments of the present invention includes the following steps:

[0049] S101: Construct and train an unsupervised projection recovery network based on optimal transport and construct and train an unsupervised image denoising network based on optimal transport; among them, the unsupervised projection recovery network refers to a network that, after unsupervised training using projection data, inputs a noisy projection and outputs a detail-restored projection. The unsupervised image denoising network refers to a network that, after unsupervised training using image data, inputs a noisy image and outputs a denoised image.

[0050] S102: Use the trained unsupervised projection recovery network to perform detail restoration on the low-dose projection of the photon-counting CT to generate a denoised projection;

[0051] S103: Use an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image;

[0052] S104: Use the trained unsupervised image denoising network to denoise the intermediate image to obtain the final denoised image.

[0053] The unsupervised low-dose reconstruction method for photon-counting CT based on optimal transport provided by the embodiments of the present invention constructs a dual-domain joint framework (abbreviation: DDOT-Net) to reconstruct low-dose projections into high-quality images. This framework splits the low-dose photon-counting CT reconstruction problem into two sub-problems: detail restoration and noise reduction. First, for the detail restoration sub-problem, an unsupervised optimal transport network is constructed in the projection domain (i.e., an unsupervised projection restoration network based on optimal transport, the DDOT projection domain network) to process the noisy projections; then, an image reconstruction algorithm is used to reconstruct the processed projections to obtain an intermediate image after detail restoration. Finally, for the noise reduction sub-problem, an unsupervised optimal transport network is constructed in the image domain (i.e., an unsupervised image noise reduction network based on optimal transport, the DDOT image domain network) to process the reconstructed intermediate image to obtain the final denoised image. The low-dose projections are reconstructed into high-quality images through this dual-domain joint framework.

[0054] The embodiments of the present invention innovatively introduce the optimal transport theory into the low-dose photon-counting CT reconstruction problem, model from the perspective of distribution similarity learning, and at the same time, aiming at the problem that it is difficult to obtain high-quality paired data sets, unsupervised networks are designed in the projection domain and the image domain to construct a dual-domain low-dose photon-counting CT reconstruction framework, and the noisy projections are processed in the dual domain to finally obtain high-quality CT reconstruction images. By utilizing the advantages of the optimal transport theory in distribution similarity learning and splitting the low-dose photon-counting CT reconstruction into two sub-problems: detail restoration and image noise reduction, and solving them separately in the dual domain, the quality of the low-dose reconstructed images can be significantly improved.

[0055] In one embodiment, the projection domain optimization method f is parameterized by a deep neural network and trained using noisy projections. According to formula (4), the solution of the projection domain optimal transport network f(·; θ1 ∈ Θ1) can be expressed as:

[0056]

[0057] where γ1 is the weight parameter, is the probability distribution of the output denoised projection, is the probability distribution of the true projection, W1 is the Wasserstein-1 distance, Y is the input noisy projection, is the denoised projection output by the unsupervised projection restoration network and

[0058] Similarly, in one embodiment, the image domain optimization method g is parameterized by a deep neural network and uses the intermediate image Training is carried out. The solution of the image domain optimal transport network \(g(\cdot;\theta_2\in\Theta_2)\) can be expressed as:

[0059]

[0060] where \(\gamma_2\) is the weight parameter, \(p\) X is the probability distribution of the output denoised image, is the probability distribution of the real image, \(W_1\) is the Wasserstein-1 distance, is the intermediate image, is the denoised image output by the unsupervised image denoising network Output denoised image,

[0061] In one embodiment, the structure of the unsupervised DDOT projection recovery network is as Figure 3 shown, designed based on the WGAN-GP framework. Among them, the generator includes a first convolutional layer, a residual channel attention (RCAB) module, a U-Net structure, and a second convolutional layer connected in sequence, and a skip connection is set between the input of the first convolutional layer and the output of the second convolutional layer, as Figure 3 (a) shown; the discriminator includes a plurality of convolutional blocks stacked in sequence and two fully connected layers; the convolutional block includes a convolutional layer and a ReLU layer cascaded in sequence, as Figure 3 (b) shown. Compared with the original WGAN-GP framework, the discriminator in this embodiment removes the batch normalization layer and the Sigmoid activation function. The loss function of this network is formula (5), and the norm of the input projection and the output projection is used for regularization constraint in this loss function.

[0062] As an implementable manner, in this U-Net structure, the encoder includes two downsampling CNN layers, the decoder includes two upsampling CNN layers, and each downsampling and upsampling layer is connected through a residual channel attention (RCAB) module. The structure of the U-Net is as Figure 3 (d) shown; the structure of the RCAB module is as Figure 3 (c) shown.

[0063] The projection recovery network in this embodiment is trained using noisy projections. The trained network is used to process noisy projections. The input noisy projection \(Y\) passes through the network After processing, the denoised projection is obtained There is:

[0064]

[0065] In one embodiment, to ensure that all quality optimization effects come from the dual-domain joint framework proposed in the present invention, the simple and effective filtered back projection (FBP) algorithm is selected for reconstruction in step S103. For step S103, there is:

[0066]

[0067] where A represents the projection acquisition process, and A -1 represents the image reconstruction process. In this example, A -1 represents the FBP reconstruction algorithm.

[0068] In one embodiment, the image domain network and the projection domain network adopt the same structural design. The loss function of the unsupervised DDOT image denoising network is formula (6), and the norm of the input image and the output image is used for regularization constraint in this loss function. This image domain denoising network uses the intermediate image for training. The trained network is used to input the intermediate image and after being processed by the network the final reconstructed image X is obtained. For step S104, there is:

[0069]

[0070] The unsupervised low-dose photon-counting CT reconstruction method based on optimal transport provided by the present invention introduces the optimal transport theory into the low-dose CT reconstruction problem, models it from the perspective of distribution similarity learning, and at the same time, aiming at the problem that it is difficult to obtain high-quality paired data sets, designs unsupervised networks in the projection domain and the image domain, constructs a low-dose CT reconstruction framework for dual-domain fusion, and performs dual-domain processing on the noisy projection to finally obtain high-quality CT reconstructed images. After this method is deployed, the unsupervised low-dose CT image reconstruction technology can be realized. The simulation experiment and actual experiment results are as Figure 4 、 Figure 5 shown. Compared with the semi-supervised and unsupervised methods in recent years, the method proposed in the present invention has obvious advantages in both image details and numerical indicators. The PSNR is at least 0.18 higher, and the SSIM is at least 0.05 higher. The simulation experiment and actual experiment results both show that the lower the imaging dose, the more obvious the advantages of the proposed method, which proves the effectiveness and robustness of the method of the present invention in low-dose photon-counting CT image reconstruction.

[0071] Based on the same inventive concept, as Figure 6 shown, the embodiment of the present invention also provides an unsupervised low-dose photon-counting CT reconstruction device based on optimal transport, which is used to model the low-dose photon-counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints, and includes a network construction and training module, a projection denoising module, an image reconstruction module, and an image denoising module.

[0072] The network construction and training module is used to construct and train an unsupervised projection recovery network based on optimal transport and construct and train an unsupervised image denoising network based on optimal transport; the projection denoising module is used to use the trained unsupervised projection recovery network to recover the details of the low-dose projection of photon-counting CT to generate a denoised projection; the image reconstruction module is used to use an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image; the image denoising module is used to use the trained unsupervised image denoising network to denoise the intermediate image to obtain a final denoised image.

[0073] It should be noted that the unsupervised low-dose photon-counting CT reconstruction device provided in the embodiments of the present invention is to implement the above method, and its functions can be specifically referred to in the above method embodiments, which will not be elaborated here.

[0074] Figure 7 An example of the physical structure diagram of an electronic device is shown as Figure 7 As shown, the electronic device may include: a processor 701, a communication interface 702, a memory 703, and a communication bus 704. Among them, the processor 701, the communication interface 702, and the memory 703 complete mutual communication through the communication bus 704. The processor 701 can call the logical instructions in the memory 703 to execute an unsupervised low-dose photon-counting CT reconstruction method based on optimal transport, and model the low-dose photon-counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints. The method includes: constructing and training an unsupervised projection recovery network based on optimal transport and constructing and training an unsupervised image denoising network based on optimal transport; using the trained unsupervised projection recovery network to recover the details of the low-dose projection of photon-counting CT to generate a denoised projection; using an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image; using the trained unsupervised image denoising network to denoise the intermediate image to obtain a final denoised image.

[0075] In addition, when the logical instructions in the above-mentioned memory 703 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, external hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0076] An embodiment of the present invention further provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the unsupervised low-dose photon counting CT reconstruction method based on optimal transport provided by each of the above method embodiments.

[0077] An embodiment of the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the unsupervised low-dose photon counting CT reconstruction method based on optimal transport provided by each of the above method embodiments.

[0078] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such understanding, the technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disks, optical discs, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0079] 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 described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. An unsupervised low-dose photon-counting CT reconstruction method based on optimal transport, characterized in that, Model the low-dose photon-counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints, including: Construct and train an unsupervised projection recovery network based on optimal transport and construct and train an unsupervised image denoising network based on optimal transport; Use the trained unsupervised projection recovery network to recover the details of the low-dose projection of the photon-counting CT and generate a denoised projection; Use an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image; Use the trained unsupervised image denoising network to denoise the intermediate image to obtain the final denoised image.

2. The unsupervised low-dose photon-counting CT reconstruction method based on optimal transport according to claim 1, wherein During the training process of the unsupervised projection recovery network, use the following formula as the loss function: where γ1 is a weight parameter, is the probability distribution of the denoised output projection, is the probability distribution of the true projection, W1 is the Wasserstein-1 distance, Y is the input noisy projection, is the unsupervised projection recovery network is the denoised output projection 3. The unsupervised low-dose photon-counting CT reconstruction method based on optimal transport according to claim 1, wherein During the training process of the unsupervised image denoising network, use the following formula as the loss function: where γ2 is a weight parameter, p X is the probability distribution of the output denoised image, is the probability distribution of the real image, W1 is the Wasserstein-1 distance, is the intermediate image, is the unsupervised image denoising network outputs the denoised image, 4. The unsupervised low-dose photon counting CT reconstruction method based on optimal transport according to any one of claims 1 to 3, characterized in that, Both the unsupervised projection recovery network and the unsupervised image denoising network adopt the WGAN-GP framework; among them, the generator includes a first convolutional layer, a residual channel attention module, a U-Net structure, and a second convolutional layer connected in sequence, and a skip connection is set between the input of the first convolutional layer and the output of the second convolutional layer; the discriminator includes a plurality of convolutional blocks stacked in sequence and two fully connected layers; the convolutional block includes a convolutional layer and a ReLU layer cascaded in sequence.

5. The unsupervised low-dose photon counting CT reconstruction method based on optimal transport according to claim 4, wherein The encoder in the U-Net structure includes two downsampling CNN layers, and the decoder includes two upsampling CNN layers, and each downsampling CNN layer and upsampling CNN layer are connected by a residual channel attention module.

6. An unsupervised low-dose photon-counting CT reconstruction device based on optimal transport, characterized in that, Model the low-dose photon-counting CT reconstruction problem as a Kantorovich problem under distribution consistency constraints, including: A network construction and training module for constructing and training an unsupervised projection recovery network based on optimal transport and constructing and training an unsupervised image denoising network based on optimal transport; A projection denoising module for using the trained unsupervised projection recovery network to recover the details of the low-dose projection of the photon-counting CT and generate a denoised projection; An image reconstruction module for using an image reconstruction algorithm to reconstruct the denoised projection to generate an intermediate image; An image denoising module for using the trained unsupervised image denoising network to denoise the intermediate image to obtain the final denoised image.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein When the processor executes the program, it implements the method according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method according to any one of claims 1 to 5.

Images