Multi-branch attention prior parameterized limited projection fast fluorescence tomography reconstruction method

Abstract

The application provides a multi-branch attention prior parameterization finite projection fast fluorescence tomographic reconstruction method, comprising: generating and collecting fluorescence projection data, constructing a multi-branch attention prior-based parameterization finite projection reconstruction model, training the parameterization finite projection reconstruction model through the fluorescence projection data, obtaining an optimal parameter model, and saving the optimal parameter model; and performing three-dimensional tomographic reconstruction with the aid of the optimal parameter model. The multi-branch attention prior parameterization finite projection fast fluorescence tomographic reconstruction method can solve the problems and challenges of the existing conventional regularization method-based complex iterative calculation process, time-consuming solution efficiency and low reconstruction accuracy caused by manual selection of regularization parameters in the finite projection reconstruction process.

Description

Technical Field

[0001] This invention relates to the field of optical molecular imaging technology, and in particular to a fast fluorescence tomography reconstruction method using multi-branch attention prior parameterized finite projection. Background Technology

[0002] Fluorescence molecular tomography (FMT) is a promising optical molecular imaging technique with significant advantages in high sensitivity, specificity, and depth resolution. It can quantitatively monitor the three-dimensional (3D) distribution of fluorescent molecular probes within biological tissues. Compared to other anatomical imaging techniques such as CT and MRI, FMT can detect disease at the molecular or cellular level. Over the past few decades, FMT has become increasingly popular in numerous physiological, preclinical, and clinical studies, and is widely used in various small animal model-based preclinical and clinical research, such as tumor diagnosis, drug development, and treatment evaluation.

[0003] However, due to the high scattering characteristics of photons in biological tissues and the lack of surface fluorescence measurement data, the reconstruction process of FMT is considered a seriously ill-conditioned inverse problem, severely affecting the reconstruction quality. To improve the reconstruction performance of FMT, researchers have proposed various regularization algorithms based on the sparsity characteristics of fluorescence sources, including L2 regularization, L1 regularization, L0 regularization, L2,1, and Lasso. Secondly, to obtain better reconstruction results, these algorithms typically employ full-angle projection to reconstruct the three-dimensional distribution of fluorescent molecular probes. In full-angle FMT imaging systems, 36-72 projections, or even more, are generally required to obtain higher reconstructed image quality. However, acquiring full-angle sampling data usually takes a long time (approximately 5-45 minutes), which is not suitable for rapid visualization of in vivo biological processes. Therefore, some researchers have proposed traditional iterative algorithms utilizing finite projection reconstruction to achieve rapid FMT reconstruction by shortening the experimental data acquisition time. While these methods shorten data acquisition time to some extent, they are ultimately based on iterative regularization methods. Their reconstruction process involves complex iterative calculations and time-consuming solution processes (approximately hundreds of seconds). Furthermore, relying on manually selected regularization parameters may result in low reconstruction accuracy, hindering the rapid dynamic reconstruction of biological processes and severely limiting the widespread application of FMT technology in preclinical and clinical settings. Therefore, designing a parameterized finite projection fast fluorescence tomography reconstruction method with multi-branch attention prior is essential. Summary of the Invention

[0004] The purpose of this invention is to provide a fast fluorescence tomography reconstruction method based on multi-branch attention prior parameterization of finite projection, which can solve the problems and challenges of the existing traditional regularization method in the finite projection reconstruction process, such as complex iterative calculation process, time-consuming solution efficiency, and low reconstruction accuracy due to reliance on manual selection of regularization parameters.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A fast fluorescence tomography reconstruction method based on multi-branch attention prior parameterized finite projection includes the following steps:

[0007] Step 1: Generate and acquire fluorescence projection data;

[0008] Step 2: Construct a parameterized finite projection reconstruction model based on multi-branch attention priors;

[0009] Step 3: Train the parametric finite projection reconstruction model using fluorescence projection data to obtain the optimal parameter model and save it;

[0010] Step 4: Perform three-dimensional fault reconstruction using the optimal parameter model.

[0011] Optionally, in step 1, generating and acquiring fluorescence projection data specifically involves:

[0012] Finite element meshes were constructed based on the specifications of the phantom or mouse model used in the experiment to generate computer simulation projection data. The generated computer simulation projection data was divided into training and validation sets at a ratio of 9:1. In addition, real phantom experimental data and live mouse experimental data were collected to form fluorescence projection data and generate a test set.

[0013] Optionally, in step 2, a parameterized finite projection reconstruction model based on multi-branch attention priors is constructed, specifically as follows:

[0014] A parameterized finite projection reconstruction model based on multi-branch attention prior is constructed, wherein the parameterized finite projection reconstruction model includes a parameterized network and a discriminant network, and is represented as follows:

[0015] (1)

[0016] Where G represents the parameterized network and K represents the discriminant network. A parameterized finite projection reconstruction model with multi-branch attention prior is constructed; the parameterized network includes a downsampling module, a fully connected module, an upsampling module, a multi-branch attention prior module, and a parameterized skip connection module.

[0017] The downsampling module consists of five downsampling sub-modules, used to reduce location information, display environmental information, and extract abstract features. Each downsampling sub-module has the same structure, including a convolutional layer, a BatchNorm layer, a ReLU activation function, and a Max Pooling layer with a stride of 2. Each convolutional layer has a small convolutional kernel, and the number of convolutional kernels gradually doubles. After each convolution operation, the number of feature channels doubles. Furthermore, no padding or Max Pooling is set, and the stride is 2.

[0018] A fully connected module is a 2048-dimensional feature composed of fully connected layers;

[0019] The upsampling module includes five upsampling sub-modules for recovering position information. Each upsampling sub-module has the same structure, including an upsampling operation layer, a convolutional layer, a BatchNorm layer, and a ReLU activation function layer. Each convolutional layer has a small convolutional kernel, and the number of convolutional kernels gradually decreases by half. After each upsampling, the number of feature channels is halved, and the size of the feature map is doubled.

[0020] The multi-branch attention prior module is used to extract four deep feature representations from the upsampling module and provide them to the downsampling module. In the process of providing them, linear interpolation is first performed to upsample the feature map. Then, the feature map is compressed to suppress unrelated information in the channel dimension and reduce the amount of computation. Next, the compressed feature map is concatenated in the channel dimension for feature fusion. Finally, the four feature maps are fused to obtain the final prior result.

[0021] The parameterized skip connection module includes parameterized magnitude factors for five different paths, which are used to introduce the multi-branch attention prior module into the network model to skip connections between multiple branches, and assign different parameters to it as multi-branch attention prior information;

[0022] The discriminant network consists of a ResNet network structure with 5 ResBlocks.

[0023] Optionally, linear interpolation is first performed to upsample the feature map. Then, the feature map is compressed to suppress irrelevant information in the channel dimension and reduce computational cost. Next, the compressed feature maps are concatenated along the channel dimension for feature fusion. Finally, the four feature maps are fused to obtain the final prior result, specifically:

[0024] Linear interpolation is performed to upsample the feature map, resulting in a feature map defined as follows: Where D represents depth, i.e., the number of FMT finite projections, C represents the number of channels, and H and W represent the height and width of the FMT projection or feature map, respectively. The feature map is compressed and fused, and the calculation method is as follows:

[0025] (2)

[0026] (3)

[0027] (4)

[0028] In the formula, Represents the fusion features of X5 and X4. Represents the fusion features of X4 and X3. Represents the fusion features of X3 and X2. , and This represents the corresponding compressed convolution. This represents the fusion convolution of X5 and X4. Represents the fusion convolution of X4 and X3. The representative features are concatenated, and the final output of the multi-branch attention prior module is calculated as follows:

[0029] (5)

[0030] In the formula, Represents the fused convolution of X3 and X2. This represents the output convolution after fusion.

[0031] Optionally, in step 3, the parameterized finite projection reconstruction model is trained using fluorescence projection data to obtain the optimal parameter model, which is then saved. Specifically:

[0032] The parametric finite projection reconstruction model was trained using a training set, and the generalization ability of the optimal parameter model was tested using a validation set. The loss function for network training consisted of three parts: adversarial loss, L1 loss, and prior attention loss. The adversarial loss was used to prevent blurring of the FMT image, the L1 loss was used to remove sharp parts of the FMT image, and the prior attention loss was used to recover the morphology of the fluorescence source. The total loss function of the parametric finite projection reconstruction model with multi-branch attention prior is as follows:

[0033] (6)

[0034] (7)

[0035] In the formula, and Here, are the loss functions for the parameterized network and the discriminant network, respectively, and E is the expectation operator. For the gradient operator, P g For the distribution of the generated FMT image, P rFor the distribution of real FMT images, P x To uniformly sample along a straight line connecting the paired generated FMT images and the real FMT images, The penalty parameter is a constant weight. The network parameters are updated by continuously feeding fluorescence projection data into the reconstruction model in batches and training the network. After repeating the network parameter update operation N times, the parameters are considered to be the optimal parameters when the mean square error of each pixel obtained in the Nth and N-1th times is <1e-3. The optimal parameter model is saved as a pth file for testing.

[0036] Optionally, in step 4, 3D reconstruction is performed using the optimal parameter model, specifically as follows:

[0037] The collected real phantom experimental data and real live experimental data are used as inputs into the optimal parameter model to perform three-dimensional tomographic reconstruction of the fluorescent target, which mainly involves three processes: feature extraction, image fusion and image reconstruction.

[0038] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides a fast fluorescence tomography reconstruction method with multi-branch attention prior parameterized finite projection, the method comprising generating and acquiring fluorescence projection data; constructing a parameterized finite projection reconstruction model based on multi-branch attention prior; performing network training on the parameterized finite projection reconstruction model using fluorescence projection data to obtain an optimal parameter model and saving it; and performing three-dimensional tomography reconstruction using the optimal parameter model. This method employs an end-to-end reconstruction scheme based on deep learning. It does not require prior knowledge such as optical parameters, but instead continuously learns and adjusts model parameters through numerous data mapping cases from internal fluorescence distribution to boundary fluorescence signal distribution, directly establishing a nonlinear mapping relationship between the internal and boundary fluorescence signal distributions. A parameterized multi-branch attention prior mechanism provides parameterized weighted sparse prior information for reconstructing the fluorescence source, effectively mitigating the severe ill-conditioning of the finite projection inverse problem and significantly improving the reconstruction accuracy of FMT. Ultimately, excellent reconstruction performance can be achieved using only three projections. Constructing a parameterized finite projection reconstruction model based on multi-branch attention priors avoids the complex iterative calculation process in traditional regularization algorithms, with a reconstruction time of only about 0.18 seconds, greatly accelerating the 3D visualization process. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the process of the fast fluorescence tomography reconstruction method with multi-branch attention prior parameterization of finite projection according to an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the parameterized reconstruction model structure for multi-branch attention priors;

[0042] Figure 3 This is a schematic diagram of a parameterized network structure;

[0043] Figure 4 This is a schematic diagram of the multi-branch attention prior module structure;

[0044] Figure 5 A schematic diagram for identifying network structure. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The purpose of this invention is to provide a fast fluorescence tomography reconstruction method for finite projection with multi-branch attention prior parameterization, which can solve the problems and challenges of the complex iterative calculation process, time-consuming solution efficiency, and low reconstruction accuracy caused by relying on manual selection of regularization parameters in the existing finite projection reconstruction based on traditional regularization methods.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown in the figure, the fast fluorescence tomography reconstruction method with multi-branch attention prior parameterized finite projection provided by the present invention includes the following steps:

[0049] Step 1: Generate and acquire fluorescence projection data;

[0050] Step 2: Construct a parameterized finite projection reconstruction model based on multi-branch attention priors;

[0051] Step 3: Train the parametric finite projection reconstruction model using fluorescence projection data to obtain the optimal parameter model and save it;

[0052] Step 4: Perform three-dimensional fault reconstruction using the optimal parameter model.

[0053] In step 1, fluorescence projection data is generated and acquired, specifically as follows:

[0054] Finite element meshes were constructed based on the specifications of the phantom or mouse model used in the experiment, generating computer simulation projection data. This computer simulation projection data was then divided into training and validation sets at a 9:1 ratio. The training set was used to train the network model, and the validation set was used to select the optimal model. In addition, real phantom experimental data and live mouse experimental data were collected to form fluorescence projection data, generating a test set to evaluate the model's reconstruction performance.

[0055] The training and validation sets were generated using the Finite Element Method (FEM) to solve a simplified spherical harmonic (SPN) model for training and validation. The FEM mesh was constructed according to the specifications of the phantom or mouse model used in each experiment. In the actual phantom and mouse experiments, data was generated by randomly adding fluorescent targets of different sizes and optical properties. The optical parameters were randomly varied and confined to a biologically relevant range. For different fluorescent targets, the diameter ranged from 2 to 5 mm, while the height was fixed at 5 mm. These targets were randomly placed within the model, and the horizontal and vertical coordinates of their centers were randomly generated between -1.0 and 1.0 cm. The edge-to-edge distances were also randomly determined. The EEDs were randomly varied between 0.3 and 2.5 mm. The source-detector configuration of the data was set to match the real experimental data. Finally, a total of 10,000 simulated samples were generated, of which 1,000 simulated samples were used as a validation set to determine the optimal model.

[0056] To improve model convergence, data preprocessing is performed, which is a crucial step. Considering that light intensity typically decays exponentially in diffuse optical imaging, the training and validation data are first subtracted from the mean, and then the resulting data is divided by its standard deviation.

[0057] In step 2, a parameterized finite projection reconstruction model based on multi-branch attention priors is constructed, specifically as follows:

[0058] The parameterized model of the multi-branch attention prior aims to directly establish the nonlinear mapping relationship between surface fluorescence measurement and internal fluorescence source, as specifically expressed below:

[0059] (1)

[0060] In the formula, represents the surface fluorescence measurement value, which is the input to the network model; x represents the distribution of the internal fluorescence source to be reconstructed, which is the output of the network model. The parameterized finite projection reconstruction model based on multi-branch attention prior is constructed.

[0061] like Figure 2 As shown, a parameterized finite projection reconstruction model based on multi-branch attention prior is constructed. This model includes a parameterized network and a discriminant network, and is represented as follows:

[0062] (2)

[0063] In the formula, G is the parameterized network, K is the discriminant network, and so on. Figure 3 As shown, the parameterized network includes a downsampling module, a fully connected module, an upsampling module, a multi-branch attention prior module, and a parameterized skip connection module;

[0064] The downsampling module comprises five downsampling sub-modules. Their function is to reduce location information, reveal environmental information, and extract abstract features. Each downsampling sub-module has the same structure, including a convolutional layer, a BatchNorm layer, a ReLU activation function, and a Max Pooling layer with a stride of 2. Each convolutional layer uses... The number of convolutional kernels gradually doubles, i.e., 32, 64, 128, 256, and 512. After each convolution operation, the number of feature channels doubles, and no padding or max pooling is added. The stride is 2. After each repeated structural operation, the size of the feature map will become half of the previous size.

[0065] The fully connected module is a 2048-dimensional feature composed of fully connected layers. It approximates the inter-manifold projection from the perceptual domain to the image domain, thereby realizing the mapping from surface photon density to internal fluorescence light source distribution.

[0066] The upsampling module comprises five upsampling sub-modules, whose function is to gradually recover positional information. Each upsampling sub-module has the same structure, including an upsampling operation layer, a convolutional layer, a BatchNorm layer, and a ReLU activation function layer. Each convolutional layer uses a small convolutional kernel. The number of convolution kernels gradually decreases by half, namely 512, 256, 128, 64, 32, and 16. After each upsampling, the number of feature channels is halved, and the size of the feature map is doubled.

[0067] like Figure 4As shown, the topology of the multi-branch attention prior module extracts four deep feature representations from the downsampling module and provides them to the upsampling module. Due to the different spatial sizes of the feature maps, linear interpolation is first performed to upsample the feature maps. Then, the feature maps are compressed to suppress irrelevant information in the channel dimension and reduce computational load. Next, the compressed feature maps are concatenated along the channel dimension for feature fusion. Finally, the four feature maps are fused to obtain the final prior result. The specific steps are as follows:

[0068] Linear interpolation is performed to upsample the feature map, resulting in a feature map defined as follows: Where D represents depth, i.e., the number of FMT finite projections, C represents the number of channels, and H and W represent the height and width of the FMT projection or feature map, respectively. The feature map is compressed and fused, and the calculation method is as follows:

[0069] (3)

[0070] (4)

[0071] (5)

[0072] In the formula, Represents the fusion features of X5 and X4. Represents the fusion features of X4 and X3. Represents the fusion features of X3 and X2. , and This represents the corresponding compressed convolution. This represents the fusion convolution of X5 and X4. Represents the fusion convolution of X4 and X3. The representative features are concatenated, and the final output of the multi-branch attention prior module is calculated as follows:

[0073] (6)

[0074] In the formula, Represents the fused convolution of X3 and X2. This represents the output convolution after fusion;

[0075] The parameterized skip connection module includes parameterized magnitude factors for five different paths. This is used to introduce a multi-branch attention prior module into a network model, skipping multiple branches and assigning it different parameters as multi-branch attention prior information, including parameterized amplitude factors. These parameters are set as learnable parameters for the network and updated during backpropagation. This setting can add additional nonlinear capabilities to the network, improving the effectiveness of skipping connections;

[0076] Specific use Y represents the feature map from the encoder, and Y represents the prior attention map and parameterized feature map. As shown below:

[0077] (7)

[0078] An attention prior module is introduced into the parameterized skip connection in the parameterized reconstruction model with multi-branch attention prior, serving as multi-branch attention prior information. The prior attention module extracts rich semantic features from the encoder and generates a spatial attention map to guide finite projection reconstruction;

[0079] like Figure 5 As shown, the discriminative network consists of a ResNet network structure with 5 ResBlocks. Each ResBlock structure is followed by a Max pooling operation, and then an Average pooling layer and a fully connected (FC) layer are used to obtain a discriminative score. Specifically, the input feature map is processed by a convolution kernel of size , then LayerNorm normalization and Selu activation function are applied. This process is repeated twice, and the result is added to the previously input feature map to form a skip connection for easier training. This connection is then used as the input to the Selu activation function to obtain the output of the entire ResBlock structure.

[0080] In step 3, the parameterized finite projection reconstruction model is trained using fluorescence projection data to obtain the optimal parameter model, which is then saved. Specifically:

[0081] The parameterized finite projection reconstruction model was trained and its parameters selected using training and validation sets. The generalization ability of the optimal parameter model was tested using a test set. The training loss function consisted of three parts: adversarial loss, L1 loss, and prior attention loss. The adversarial loss was used to prevent blurring of the FMT image and to extract sharp parts of the FMT image. The L1 loss focused on image details, and the prior attention loss was used to recover the morphology of the fluorescence source. The total loss function of the parameterized finite projection reconstruction model with multi-branch attention prior is as follows:

[0082] (8)

[0083] (9)

[0084] In the formula, and Here, are the loss functions for the generator and discriminator, respectively, and E is the expectation operator. For the gradient operator, P g For the distribution of the generated FMT image, P r For the distribution of real FMT images, P x To uniformly sample along a straight line connecting the generated FMT image and the real FMT image in pairs, The penalty parameter is a constant weight. The network parameters are updated by continuously feeding fluorescence projection data into the system model in batches and training the network. After repeating the network parameter update operation N times, the parameters are considered to be the optimal parameters when the mean square error of each pixel obtained in the Nth and N-1th times is <1e-3. The optimal parameter model is saved as a pth file for testing.

[0085] In step 4, three-dimensional fault reconstruction is performed using the optimal parameter model, specifically as follows:

[0086] The collected real phantom experimental data and real live experimental data are used as inputs into the optimal parameter model to perform three-dimensional tomographic reconstruction of the fluorescent target, which mainly involves three processes: feature extraction, image fusion and image reconstruction.

[0087] The reconstruction was performed using the multi-branch attention prior parameterized finite projection fast fluorescence tomography reconstruction method provided by this invention, and the reconstruction time was compared with that of traditional algorithms, as shown in the table below:

[0088] Table 1: Comparison of Reconstruction Time between Traditional Methods and Deep Learning Methods

[0089]

[0090] The reconstruction using the multi-branch attention prior parameterized finite projection fast fluorescence tomography reconstruction method provided by this invention takes only about 0.18 seconds, greatly accelerating the process of 3D visualization.

[0091] This invention provides a fast fluorescence tomography (FMT) reconstruction method using multi-branch attention prior parameterized finite projection. The method includes generating and acquiring fluorescence projection data, constructing a parameterized finite projection reconstruction model based on multi-branch attention priors, training the model with the fluorescence projection data to obtain and save the optimal parameter model, and performing 3D reconstruction using the optimal parameter model. This method employs a deep learning-based end-to-end reconstruction scheme, eliminating the need for prior knowledge such as optical parameters. Instead, it continuously learns and adjusts model parameters through numerous data mapping cases from internal fluorescence distribution to boundary fluorescence signal distribution, directly establishing a nonlinear mapping relationship between the internal and boundary fluorescence signal distributions. The multi-branch attention prior mechanism provides parameterized weighted sparse prior information for reconstructing the fluorescence source, effectively mitigating the severe ill-conditioning of the finite projection inverse problem and significantly improving the reconstruction accuracy of finite projection FMT. Ultimately, excellent reconstruction performance can be achieved using only three projections. Constructing a parameterized network finite projection reconstruction model based on multi-branch attention priors avoids the complex iterative calculation process in traditional regularization algorithms, with a reconstruction time of only about 0.18 seconds, greatly accelerating the 3D visualization process.

[0092] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A fast fluorescence tomography reconstruction method based on multi-branch attention prior parameterized finite projection, characterized in that, Includes the following steps: Step 1: Generate and acquire fluorescence projection data; Step 2: Construct a parameterized finite projection reconstruction model based on multi-branch attention prior; Step 3: Train the parametric finite projection reconstruction model using fluorescence projection data to obtain the optimal parameter model and save it; Step 4: Perform three-dimensional fault reconstruction using the optimal parameter model; In step 2, a parameterized finite projection reconstruction model based on multi-branch attention priors is constructed, specifically as follows: A parameterized finite projection reconstruction model based on multi-branch attention prior is constructed, wherein the parameterized finite projection reconstruction model includes a parameterized network and a discriminant network, and is represented as follows: （1） Where G represents the parameterized network and K represents the discriminant network. This represents a parameterized finite projection reconstruction model with a constructed multi-branch attention prior; the parameterized network includes a downsampling module, a fully connected module, an upsampling module, a multi-branch attention prior module, and a parameterized skip connection module; The downsampling module consists of five downsampling sub-modules, used to reduce location information, display environmental information, and extract abstract features. Each downsampling sub-module has the same structure, including a convolutional layer, a BatchNorm layer, a ReLU activation function, and a Max Pooling layer with a stride of 2. Each convolutional layer has a small convolutional kernel, and the number of convolutional kernels gradually doubles. After each convolution operation, the number of feature channels doubles. Furthermore, no padding or Max Pooling is set, and the stride is 2. A fully connected module is a 2048-dimensional feature composed of fully connected layers; The upsampling module includes five upsampling sub-modules for recovering position information. Each upsampling sub-module has the same structure, including an upsampling operation layer, a convolutional layer, a BatchNorm layer, and a ReLU activation function layer. Each convolutional layer has a small convolutional kernel, and the number of convolutional kernels gradually decreases by half. After each upsampling, the number of feature channels is halved, and the size of the feature map is doubled. The multi-branch attention prior module is used to extract four deep feature representations from the downsampling module and provide them to the upsampling module. In the process of providing them, linear interpolation is first performed to upsample the feature map. Then, the feature map is compressed to suppress unrelated information in the channel dimension and reduce the amount of computation. Next, the compressed feature map is concatenated in the channel dimension to perform feature fusion. Finally, the four feature maps are fused to obtain the final prior result. The parameterized skip connection module includes parameterized magnitude factors for five different paths, which are used to introduce the multi-branch attention prior module into the network model to skip connections between multiple branches, and assign different parameters to it as multi-branch attention prior information; The discriminant network consists of a ResNet network structure with 5 ResBlocks; First, linear interpolation is performed to upsample the feature map. Then, the feature map is compressed to suppress irrelevant information along the channel dimension and reduce computational cost. Next, the compressed feature maps are concatenated along the channel dimension for feature fusion. Finally, the four feature maps are fused to obtain the final prior result, as follows: Linear interpolation is performed to upsample the feature map, resulting in a feature map defined as follows: Where D represents depth, i.e., the number of FMT finite projections, C represents the number of channels, and H and W represent the height and width of the FMT projection or feature map, respectively. The feature map is compressed and fused, and the calculation method is as follows: （2） （3） （4） In the formula, Represents the fusion features of X5 and X4. Represents the fusion features of X4 and X3. Represents the fusion features of X3 and X2. , and This represents the corresponding compressed convolution. This represents the fusion convolution of X5 and X4. Represents the fused convolution of X4 and X3. The representative features are concatenated, and the final output of the multi-branch attention prior module is calculated as follows: （5） In the formula, Represents the fused convolution of X3 and X2. This represents the output convolution after fusion.

2. The fast fluorescence tomography reconstruction method with multi-branch attention prior parameterized finite projection according to claim 1, characterized in that, In step 1, fluorescence projection data is generated and acquired, specifically as follows: Finite element meshes were constructed based on the specifications of the phantom or mouse model used in the experiment to generate computer simulation projection data. The generated computer simulation projection data was divided into training and validation sets at a ratio of 9:

1. In addition, real phantom experimental data and live mouse experimental data were collected to form fluorescence projection data and generate a test set.

3. The fast fluorescence tomography reconstruction method with multi-branch attention prior parameterized finite projection according to claim 1, characterized in that, In step 3, the parameterized finite projection reconstruction model is trained using fluorescence projection data to obtain the optimal parameter model, which is then saved. Specifically: The parametric finite projection reconstruction model was trained using a training set, and the generalization ability of the optimal reconstruction model was tested using a validation set. The loss function for network training consisted of three parts: adversarial loss, L1 loss, and prior attention loss. The adversarial loss was used to prevent blurring of the FMT image, the L1 loss was used to remove sharp parts of the FMT image, and the prior attention loss was used to recover the morphology of the fluorescence source. The total loss function of the parametric finite projection reconstruction model with multi-branch attention prior is as follows: （6） （7） In the formula, and Here, are the loss functions for the parameterized network and the discriminant network, respectively, and E is the expectation operator. For the gradient operator, P g For the distribution of the generated FMT image, P r For the distribution of real FMT images, P x To uniformly sample along a straight line connecting the generated FMT image and the real FMT image in pairs, The penalty parameter is a constant weight. The network parameters are updated by continuously feeding fluorescent projection data into the reconstruction model in batches and training the network. After repeating the network parameter update operation N times, the parameters are considered to be the optimal parameters when the mean square error of each pixel obtained in the Nth and N-1th times is <1e-3. The optimal parameter model is saved as a pth file for testing.

4. The fast fluorescence tomography reconstruction method with multi-branch attention prior parameterized finite projection according to claim 1, characterized in that, In step 4, 3D reconstruction is performed using the optimal parameter model, specifically as follows: The collected real phantom experimental data and real live experimental data are used as inputs and fed into the optimal parameter model to perform three-dimensional tomographic reconstruction of the fluorescent target.

Images