A PET imaging noise suppression method based on deep neural networks

Abstract

This invention discloses a PET imaging noise suppression method based on a deep neural network. To address the problem that low-count PET images are prone to noise amplification and artifacts during iterative reconstruction using a point spread function (PSF) model, leading to decreased signal-to-noise ratio and microstructure discernibility, and difficulty in balancing resolution and quantitative accuracy, this invention performs iterative reconstruction of the original PET data using a PSF model to obtain an initial PET image and determine the PSF model parameters. Based on these parameters, a spatially variable condition map is generated that is spatially registered with the initial PET image. This spatially variable condition map is then used to infer a conditional denoising deep neural network to output a noise residual map, a detail compensation map, and a fusion weight map. Finally, based on the fusion weights, the noise residual and detail compensation are weighted and fused at the voxel level, and then combined with the initial PET image at the voxel level to obtain a noise-suppressed PET image. This achieves the technical effect of significantly reducing noise and artifacts and improving image quality while maintaining detail and contrast.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more particularly to a method for noise suppression in PET imaging based on deep neural networks. Background Technology

[0002] Positron emission tomography (PET) can acquire three-dimensional distribution information of radioactive tracers in vivo and is widely used in tumor diagnosis and staging, efficacy evaluation, and research on neurological diseases. To reduce the radiation burden on patients, shorten scan time, and support dynamic or low-activity imaging, low-count acquisition schemes, such as low-dose or short-time acquisition, are often used in practice. Since PET imaging is inherently limited by statistical fluctuations in counts, image noise increases significantly under low-count conditions, thus affecting lesion detection, boundary interpretation, and the stability of quantitative indicators.

[0003] In existing technologies, PET image reconstruction typically employs iterative reconstruction methods such as the maximum likelihood expectation-maximization algorithm and its accelerated form, the ordered subset expectation-maximization algorithm, combined with normalization correction, stochastic coincidence correction, attenuation correction, and scattering correction to improve quantitative accuracy. To improve spatial resolution, the industry often incorporates point spread function models into the system response, and can combine time-of-flight information to enhance fine structures and edge details. Meanwhile, to suppress noise, common methods include post-reconstruction smoothing filtering, penalized likelihood-based regularized reconstruction, and deep learning-based image domain denoising networks.

[0004] However, the aforementioned existing technologies still have the following shortcomings:

[0005] 1. Under low count conditions, iterative reconstruction is prone to noise amplification and ringing artifacts after introducing the point spread function model. It is difficult to simultaneously achieve noise suppression and resolution improvement in reconstruction parameters and post-processing parameters.

[0006] 2. Traditional filtering or regularization methods often use spatially fixed smoothing strength or prior constraints, which are difficult to adapt to the spatially variable resolution characteristics of the point spread function caused by changes in the field of view position. This can easily lead to loss of detail, decreased contrast, or quantitative bias.

[0007] 3. Existing deep learning denoising methods often lack a conditional adaptive mechanism for the parameters of the system's point spread function model, which can easily lead to over-smoothing or structural distortion, and have limited generalization ability under different imaging conditions.

[0008] Therefore, there is a need for a PET imaging noise suppression method that can overcome the shortcomings of the existing technology. Summary of the Invention

[0009] One objective of this invention is to propose a noise suppression method for PET imaging based on deep neural networks. Addressing the problems of existing technologies that easily generate noise amplification and artifacts when using iterative reconstruction with a point spread function (PSF) model under low-count PET imaging conditions, the difficulty of traditional filtering or regularization methods in simultaneously achieving noise suppression, spatial resolution, and quantitative accuracy, and the lack of adaptive capability to spatial variations of the PSF model in conventional deep learning denoising, the following technical solution is proposed: Acquire raw PET data and perform iterative image reconstruction incorporating a PSF model to obtain an initial PET image, while simultaneously determining the PSF model parameters; generate a spatially variable condition map that spatially registers with the initial PET image based on the model parameters; use the spatially variable condition map to infer the conditional denoising deep neural network model, outputting a noise residual map, a detail compensation map, and a fusion weight map. The network employs shared feature extraction and sets noise and detail branches, performs conditional modulation in multi-scale feature layers, and introduces dynamic convolution in at least one convolutional layer to adaptively adjust the receptive field; perform voxel-level weighted fusion of the noise residual and detail compensation based on the fusion weights, and perform voxel-level operations with the initial PET image to obtain a noise-suppressed PET image. The present invention has the technical effect of significantly reducing noise and artifacts, improving image signal-to-noise ratio and the discernibility of microstructures, and improving the adaptability to imaging characteristics of different spatial locations while maintaining detail and contrast.

[0010] This invention provides a PET imaging noise suppression method based on a deep neural network, comprising: S1, acquiring the original PET data of the object to be imaged, performing iterative image reconstruction processing including a point spread function model to obtain an initial PET image, and determining the model parameters corresponding to the point spread function model during the iterative image reconstruction process; S2, generating a spatially variable condition map registered with the initial PET image based on the model parameters, used to characterize the changes of the model parameters at different spatial locations in the initial PET image; S3, performing inference processing of a conditional denoising deep neural network model on the initial PET image using the spatially variable condition map to obtain a noise residual map, a detail compensation map, and a fusion weight map. The conditional denoising deep neural network model includes a feature extraction network and a noise branch and a detail branch connected to the feature extraction network. The noise branch outputs the noise residual map, and the detail branch outputs the detail compensation map; S4, performing voxel-level weighted fusion of the noise residual map and the detail compensation map based on the fusion weight map to obtain a fusion residual map, and performing voxel-level operations on the fusion residual map and the initial PET image to obtain a noise-suppressed PET image.

[0011] Optionally, S1 includes:

[0012] Acquire raw PET data of the object to be imaged, which is collected by the PET detector during the PET imaging process;

[0013] The raw PET data is subjected to data correction processing, which includes at least normalization correction and random compliance correction.

[0014] A system response for iterative image reconstruction processing is established based on the point spread function model, and the point spread function model is introduced into the system response.

[0015] Based on the system response, iterative image reconstruction processing is performed on the raw PET data after the data correction process to obtain the initial PET image;

[0016] The model parameters include parameters used to characterize the spatial broadening characteristics of the point spread function model, which vary with the spatial position of the initial PET image;

[0017] Furthermore, the iterative image reconstruction process is a TOF-PET iterative reconstruction process, and the model parameters also include parameters characterizing the temporal resolution or TOF kernel width. The spatial variable condition map is further used to characterize the change or preset value of the temporal resolution or TOF kernel width at spatial location.

[0018] Optionally, S2 includes:

[0019] Obtain the model parameters and the voxel coordinate system information of the initial PET image;

[0020] For each voxel position in the initial PET image, the broadening parameter of the point spread function model at the voxel position is calculated based on the spatial coordinates of the voxel position in the voxel coordinate system and the model parameters.

[0021] The spatial variable condition map is generated by using the broadening parameters corresponding to each voxel position in the same voxel arrangement as the initial PET image, so that the spatial variable condition map has the same matrix size and voxel spacing as the initial PET image.

[0022] The widening parameters include the full width at half height in a predetermined direction.

[0023] Optionally, S3 includes:

[0024] The initial PET image is subjected to intensity normalization processing to obtain a normalized initial PET image;

[0025] Perform condition feature extraction processing on the spatial variable condition graph to obtain condition features;

[0026] In the feature layers of the conditional denoising deep neural network model at multiple scales, scale modulation parameters corresponding to the scales are generated based on the conditional features. The scale modulation parameters include channel scaling parameters for scaling the channel features and channel bias parameters for translating the channel features. The channel scaling parameters and the channel bias parameters are then used to perform channel-level modulation on the features at the corresponding scales.

[0027] In at least one convolutional layer of the conditional denoising deep neural network model, dynamic convolutional parameters are generated based on the conditional features, and the convolutional operation of the convolutional layer is adaptively adjusted using the dynamic convolutional parameters to change the receptive field of the conditional denoising deep neural network model.

[0028] The shared features obtained by the feature extraction network are input into the noise branch and the detail branch respectively, so that the noise branch outputs the noise residual map and the detail branch outputs the detail compensation map.

[0029] The fusion weight map is generated based on the shared features and the conditional features, and the voxel weight values ​​of the fusion weight map are restricted to 0 to 1 by a bounded activation function.

[0030] Furthermore, the conditional denoising deep neural network model also outputs an uncertainty map, and the fusion weight map is modified according to the uncertainty map so that the weight of the noise residual map corresponding to the voxel position with higher uncertainty is increased or the weight of the detail compensation map is decreased.

[0031] Optionally, S4 includes:

[0032] Numerical truncation is performed on the fused weight map so that the weight values ​​at each voxel position in the fused weight map are between 0 and 1.

[0033] For each voxel position of the noise residual map and the detail compensation map, the noise residual map and the detail compensation map are summed at the voxel level according to the weight value of the fusion weight map at the voxel position to obtain the fusion residual map;

[0034] For each voxel position of the initial PET image, the residual value of the fused residual map at the voxel position is compared with the voxel value of the initial PET image at the voxel position. The voxel-level operation includes voxel-level subtraction of the initial PET image to suppress noise and voxel-level addition of details to compensate for noise, to obtain the noise-suppressed PET image.

[0035] Furthermore, after obtaining the noise-suppressed PET image, the process further includes: using the noise-suppressed PET image as a priori initial value or regularization reference, performing iterative image reconstruction processing containing the point spread function model again to obtain a further optimized PET image.

[0036] Optionally, the conditional denoising deep neural network model is obtained through the following training process: acquiring low-count PET raw data and a reference PET image for training, wherein the reference PET image is reconstructed from high-count PET raw data or from high-count data generated by Monte Carlo simulation; performing iterative image reconstruction processing with a point spread function model on the low-count PET raw data to obtain an initial PET image for training; determining the model parameters corresponding to the point spread function model; generating a spatially variable condition map that is spatially registered with the initial PET image for training based on the model parameters; performing forward computation of the conditional denoising deep neural network model on the initial PET image for training using the spatially variable condition map to obtain a training output PET image; constructing a loss function based on the difference between the training output PET image and the reference PET image, and updating the network parameters of the conditional denoising deep neural network model according to the loss function until a preset training termination condition is met;

[0037] Furthermore, the loss function also includes a data consistency loss, which is constructed based on the difference between the projected data obtained by forward projecting the training output PET image according to the system response and the original low-count PET data.

[0038] The beneficial effects of this invention are:

[0039] 1. By generating a spatially variable condition map from the point spread function model parameters and spatially registering it with the initial PET image, the denoising network is driven to adaptively adjust the denoising intensity and detail preservation strategy at different spatial locations. This effectively suppresses noise amplification and ringing artifacts under low count conditions, thereby improving the image signal-to-noise ratio and overall visual consistency.

[0040] 2. By performing conditional modulation in multi-scale feature layers and introducing dynamic convolution in at least one convolutional layer to adaptively change the receptive field, the network can match the spatial variation characteristics of the point spread function, reducing noise while minimizing over-smoothing, maintaining edge sharpness and the discernibility of microstructures, and improving contrast performance.

[0041] 3. By setting noise branches and detail branches to output noise residual maps and detail compensation maps respectively, and performing voxel-level weighted fusion based on the fusion weight map, and then performing voxel-level operations with the initial PET image, the decoupling and fine control of noise suppression and detail enhancement can be achieved, which is beneficial to balance resolution and quantitative accuracy and reduce the risk of structural distortion. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart of a PET imaging noise suppression method based on a deep neural network proposed in this invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0045] refer to Figure 1 A PET imaging noise suppression method based on deep neural networks includes: S1, acquiring the original PET data of the object to be imaged, performing iterative image reconstruction processing including a point spread function model to obtain an initial PET image, and determining the model parameters corresponding to the point spread function model during the iterative image reconstruction process; S2, generating a spatially variable condition map registered with the initial PET image based on the model parameters, used to characterize the changes of the model parameters at different spatial locations in the initial PET image; S3, performing inference processing of a conditional denoising deep neural network model on the initial PET image using the spatially variable condition map to obtain a noise residual map, a detail compensation map, and a fusion weight map. The conditional denoising deep neural network model includes a feature extraction network and noise and detail branches connected to the feature extraction network. The noise branch outputs the noise residual map, and the detail branch outputs the detail compensation map; S4, performing voxel-level weighted fusion of the noise residual map and the detail compensation map based on the fusion weight map to obtain a fusion residual map, and performing voxel-level operations on the fusion residual map and the initial PET image to obtain a noise-suppressed PET image.

[0046] In this specific embodiment, S1 includes:

[0047] The coincidence event list data of the object to be imaged is acquired using a PET detector, and binned according to a predetermined angle index, radial index, axial index and TOF time window index to obtain cue coincidence projection data. The cue coincidence projection data is then subjected to normalization correction and random coincidence correction in sequence. The normalization correction is based on the detection channel efficiency coefficient obtained by normalization scanning to compensate for the channel gain of each response line count. The random coincidence correction is based on the delay window method to estimate the random coincidence count and subtract it bin by bin in the projection domain, thereby obtaining low-count TOF projection data for reconstruction.

[0048] A system response for iterative image reconstruction processing is established based on the low-count TOF projection data, and a point spread function model is introduced into the system response. The iterative image reconstruction processing uses TOF-OSEM and the number of iterations is fixed. And the fixed number of subsets is In each forward and back projection calculation, geometric mapping, attenuation effect, point spread function broadening, and TOF kernel constraint are considered simultaneously, thus achieving the desired result in the first... At the end of the next iteration, the initial PET image is output. Simultaneously, during the initialization phase of the iterative reconstruction process, model parameters are read from the system calibration file and fixed for use throughout the entire iterative reconstruction process. These model parameters include point spread function model parameters. and TOF kernel width parameter ,in Used to characterize the spatial broadening properties of the point spread function model and how it varies with the spatial location of the initial PET image. The TOF kernel width is used to characterize the system's time resolution and is set to a fixed value in this embodiment;

[0049] The system response satisfies the following statistical imaging model:

[0050] ;

[0051] in Poisson represents the TOF projection data vector after normalization correction and random coincidence correction. Represents the Poisson distribution. This represents the system response matrix that includes both the point spread function model and the Time-of-Flight (TOF) model. Let represent the image vector of the radioactivity distribution to be solved, and let its iteration result be the th... The output of the next iteration is the initial PET image. This represents the scattering coincidence projection estimation vector. This represents the random conformal projection estimation vector. Represents the normalized correction matrix. Indicates the TOF kernel width parameter The determined TOF weighting matrix is ​​taken in this embodiment. And it is used to limit the width of the event location probability distribution along the response line direction. The point spread function modeling matrix represents the point spread function, and its effect on the image domain is equivalent to a spatially dependent 3D anisotropic Gaussian broadened convolution kernel. This represents the geometric projection matrix determined by the detector geometry and binning rules. This represents the attenuation factor matrix, which is obtained by line integration of the attenuation plot.

[0052] The point spread function model parameters According to voxel location Organize and record as follows:

[0053] ;

[0054] FWHM and Representing the voxel positions In direction, direction and The full width at half maximum (FWHM) parameters of the direction and the linear mapping relationship obtained through system point source calibration are derived from the voxel radial distance. The monotonically varying constraints of 4.0 mm at the center of the field of view and 7.0 mm at the edge of the field of view are calculated and satisfied. This ensures that the spatial broadening characteristics of the point spread function vary with spatial location and are fully incorporated into the forward and backward operators of the iterative reconstruction. As a result, the initial PET image is output under low count conditions for subsequent steps to generate a spatially variable condition map. and the corresponding model parameters and .

[0055] In this specific embodiment, S2 includes:

[0056] Read the initial PET image and point spread function model parameters and read The voxel coordinate system information is used to ensure that the subsequently generated spatial variable condition diagram is consistent with... Spatial registration, wherein the voxel coordinate system information includes matrix dimensions. and voxel spacing The origin of the image coordinates is defined at the center of the matrix, thus ensuring a unique correspondence between the voxel index and the physical coordinates.

[0057] against Each voxel position Calculate its physical coordinates And calculate the radial distance based on the physical coordinates. , and then combine Determine the broadening parameters at the voxel location and write them into the spatial variation condition diagram. Among them, spatial variable condition diagram Adopted and Consistent voxel arrangement order and set as three-channel volume data, with the three channels corresponding to... and ,and and Having the same matrix size Voxel spacing Thus, upon completion of the construction, it is in harmony with... Achieve spatial registration;

[0058] The above calculation and writing process satisfies the following deterministic definition:

[0059] ,

[0060] , , , ,

[0061] ;

[0062] in express Directional voxel index with values ​​ranging from 0 to express Directional voxel index with values ​​ranging from 0 to express Directional voxel index with values ​​ranging from 0 to They represent Matrix size in three dimensions These represent the voxel spacing in three dimensions, with units of 1. Voxel representation Physical coordinates in a voxel coordinate system, with units of Indicates by The constructed voxel position vector, Indicates voxel position The radial distance relative to the origin within the cross-section, in units of In this embodiment, the imaging field of view radius is taken as... Indicates the intended direction. Indicates voxel position Point spread function in direction The half-height and full-width parameters are given above and are in mm, and the above definition makes... hour and hour This forms a spatially variable broadening representation that monotonically changes with spatial location. After completion, the spatial variable condition map will be generated. With The same data dimensions and spatial metadata are saved and output to provide spatially variable conditional inputs for the inference process of conditionally denoised deep neural networks.

[0063] In this specific embodiment, S3 includes:

[0064] Initial PET image With spatial variation condition diagram A common-input conditional denoising deep neural network model is used to perform inference processing, where It is a single-channel three-dimensional volume data and is related to Alignment is performed voxel-by-voxel in the same voxel coordinate system. It is a three-channel three-dimensional volume data, and the three channels correspond to respectively and This characterizes the spatial variation of the point spread function model parameters;

[0065] right Intensity normalization was performed to obtain a normalized initial PET image. The normalization method is to... Divide each voxel value by The maximum value of all primes and constants The sum of them, thus making The numerical scale is stable and division by zero is avoided;

[0066] Spatial variable condition diagram Conditional feature extraction is performed to obtain conditional features. In this embodiment, the conditional feature extraction network is a three-dimensional convolutional encoder consisting of four scales, with the four scales sequentially using convolutional kernel sizes of [size missing]. Furthermore, the 3D convolutional layer with a stride of 1 and the downsampled 3D convolutional layer with a stride of 2 generate conditional features. And make the spatial dimensions of each level of conditional features consistent with the spatial dimensions of the same level feature layer of the backbone feature extraction network;

[0067] right A feature extraction network is used to obtain shared features. This feature extraction network is a four-level scale 3D U-shaped encoder-decoder structure containing cross-scale skip connections, and each scale feature layer at both the encoder and decoder ends is based on conditional features of the same level. Generate scale modulation parameters and ,in The channel scaling parameter is set, and the number of channels is consistent with the number of channels in the main features of the same level. The channel offset parameter is set, and the number of channels is consistent with the number of feature channels of the same level backbone. and From the Applying a core size of The three-dimensional convolution is obtained, and then... and Channel-level modulation is achieved by applying the channel-by-channel scaling and then channel-by-channel translation to the same-level backbone feature map.

[0068] A dynamic convolutional layer is set at the bottleneck scale of the feature extraction network to adaptively adjust the receptive field. The dynamic convolutional layer consists of three parallel three-dimensional convolutional branches, and the kernel size of the three branches is [missing information]. Different receptive fields were formed by using dilated convolutions with dilation rates of 1, 2, and 3, respectively, based on bottleneck-scale conditional features. The fusion coefficients of the three branches are obtained through global average pooling and fully connected mapping. After softmax normalization, the outputs of the three branches are weighted and summed to obtain the dynamic convolution output features. The fusion coefficients constitute the dynamic convolution parameters and are derived from the spatially variable conditional graph. This allows the receptive field to adapt to differences in PSF widening at spatial locations;

[0069] The shared features obtained through the feature extraction network are input into the noise branch and the detail branch, respectively. The noise branch consists of a kernel with a size of [missing information]. 3D convolutional layer output noise residual map The detailed branches are based on a kernel size of 3D convolutional layer output detail compensation map and make and Matrix size and Consistent to satisfy subsequent voxel-level operations;

[0070] Generate the original map of fusion weights based on shared features and conditional features. ,in From the core size The output of the 3D convolutional layer and with Both are of the same size, and both output uncertainty maps from the uncertainty header. ,in From the core size The output of the 3D convolutional layer is activated by softplus to ensure that the uncertainty at each voxel position is non-negative, and Used to characterize the self-evaluation of the reliability of the network for detail compensation at the corresponding voxel position;

[0071] Based on the uncertainty diagram For the original graph of fusion weights The fused weight map is modified and constrained to between 0 and 1 using a bounded activation function. The modification and limiting process satisfies the following:

[0072] ;

[0073] in This represents the output fusion weight map, which is used for subsequent voxel-level weighted fusion. This represents a clipping operator that truncates the input to 0 to 1 per voxel. This represents the sigmoid bounded activation function. This represents the original graph with fused weights. Representing uncertainty graphs, This represents the uncertainty correction factor, and in this embodiment, it is taken as... Furthermore, the aforementioned corrections reduce the fusion weight values ​​corresponding to voxel positions with higher uncertainty, thereby decreasing the proportion of detail compensation and relatively increasing the proportion of noise residuals, resulting in an output noise residual map. Detailed compensation diagram fusion weight graph and uncertainty diagram .

[0074] In this specific embodiment, S4 includes:

[0075] Received noise residual diagram Detailed compensation diagram fusion weight graph and initial PET images and the fusion weight map Numerical truncation is performed to ensure the numerical stability of subsequent voxel-level weighted fusion. This truncation process assigns a weight value less than 0 to 0 and a weight value greater than 1 to 1 at each voxel location. The fused weight map after truncation is still denoted as [image of truncation]. To maintain consistency in terminology and symbols;

[0076] In voxel coordinate system and Perform volumetric weighted fusion and Perform voxel-by-voxel operations to obtain noise-suppressed PET images The effect on noise residuals is manifested in the following way: Voxel-level subtraction is performed to suppress noise, and its effect on detail compensation is reflected in the... Voxel-level fusion is performed to compensate for details, and negative voxels in the output image are truncated to 0 by a non-negativity constraint to satisfy the physical meaning of non-negativity for PET activity. The above voxel-level weighted fusion and voxel-level operations are determined in one step by the following formula:

[0077] ;

[0078] in This indicates a noise-suppressed PET image. This indicates a voxel-wise non-negative truncation operator whose output is not less than 1 / 2. Represents the initial PET image. This indicates a fused weighted graph where the weight values ​​of each voxel are between 0 and 1. This represents the element-wise multiplication operator. This represents a detail compensation map and its voxel values ​​are used for... Enhance the details. Represents the noise residual plot and its voxel values ​​are used for... To perform noise suppression, 1 indicates that... All-in-one data with the same matrix size and each voxel value always being 1, thus Represents the voxel weights of the noise residual term;

[0079] In obtaining After that, use As prior initial values, the same set of corrected low-count TOF projection data from step S1 is used again for TOF-OSEM iterative image reconstruction processing with a point spread function model, and the system response and model parameters used are consistent with those in step S1, with the number of iterations fixed. And the fixed number of subsets is And set the initial image value for the 0th iteration to This allows for the suppression of iterative reconstruction artifacts while maintaining consistency in PSF modeling, and the output of further optimized PET images. This is the final output of this step.

[0080] In this specific embodiment, the conditional denoising deep neural network model is obtained through a training process;

[0081] When constructing the training dataset, a set of low-count PET raw data and a set of high-count PET raw data are obtained for the same object to be imaged. The data correction process, consistent with step S1, is then performed on the two sets of PET raw data to obtain the corresponding low-count TOF projection data. And high-count TOF projection data;

[0082] Based on the TOF-OSEM iterative reconstruction process involving the point spread function model described in step S1, the low-count TOF projection data is reconstructed to obtain the initial PET image for training. And determine the point spread function model parameters. With TOF kernel width parameter and the corresponding system response matrix At the same time, obtain with Same-dimensional scattering coincides with projection estimation vector and random coincidence projection estimation vector ;

[0083] Based on the method described in step S2, Generation and Spatial registration spatial variable condition diagram ;

[0084] Based on the same system response modeling method as in step S1, perform TOF-OSEM iterative reconstruction processing on the high-count TOF projection data and fix the number of iterations. And the fixed number of subsets is To obtain reference PET images and to and Perform intensity normalization consistent with step S3 to ensure that the network input and the monitoring signal are on the same numerical scale;

[0085] Will and The input-conditional denoising deep neural network model performs forward computation to obtain the noise residual map. Detailed compensation diagram fusion weight graph With uncertainty diagram And according to the voxel-level weighted fusion and voxel-level operation rules described in step S4, by , , and The training output PET image is calculated. ;

[0086] based on and Image domain loss is constructed based on image domain differences, and data consistency loss is constructed based on data consistency constraints. The data consistency loss is achieved by... Vectorization And through the system response matrix After forward projection generates predictive projection data, it is compared with low-count TOF projection data. The difference is calculated, and the overall loss function is defined as follows:

[0087] ;

[0088] in This represents the total loss used for training and optimization. Represents the image domain loss weights and takes Represents the data consistency loss weight and takes This represents the summation of the absolute values ​​of all voxels in the 3D volume data. Norm, This represents the square of the sum of the squares of the elements of the vector. Norm, This represents the PET image output from the training process. Indicates a reference PET image. This represents the parameters of the point spread function model. With TOF kernel width parameter The system response matrix is ​​jointly determined and consistent with step S1. Indicates will The image vector is obtained by flattening it in voxel order consistent with the system matrix. This represents the scattering coincidence projection estimation vector. This represents the random conformal projection estimation vector. This represents the low-count TOF projection data vector obtained after data correction processing;

[0089] The Adam optimizer is used to iteratively update the network parameters of the conditional denoising deep neural network model. The first-order momentum coefficient of Adam is set to 0.9, the second-order momentum coefficient to 0.999, and the learning rate is set to... And the weight decay coefficient is taken as The training batch size is set to 2. Divided into sliding windows of size based on voxel alignment. The three-dimensional blocks with a stride of 64 are used to form batch inputs. The total number of training rounds is 100, and the network parameters are output as the final training result after the 100th round.

[0090] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0091] This invention addresses the technical problems of low-count PET images, which are prone to noise amplification and artifacts under iterative reconstruction with the introduction of a point spread function (PSF) model, and which struggle to balance resolution and quantitative accuracy. It employs a combined process of "iterative reconstruction including a PSF model plus conditional depth denoising": First, an initial PET image is obtained through iterative reconstruction, and the PSF model parameters are simultaneously determined. Then, the model parameters are expanded according to voxel positions to generate a spatially variable condition map that is spatially registered with the initial PET image, enabling the network to perceive differences in imaging blur and noise characteristics at different spatial locations. Subsequently, a conditional denoising network outputs noise residuals, detail compensation, and fusion weights. Based on the fusion weights, voxel-level weighted fusion is performed, and voxel-level operations are executed with the initial PET image. This process suppresses statistical noise and reconstruction artifacts while preserving edge and microstructural details and maintaining contrast performance, thereby improving image quality.

[0092] Regarding the network structure, this invention makes targeted improvements to address the characteristic that "different denoising strategies are needed at different locations" caused by the spatial variation of the point spread function: First, channel-level modulation is performed in multi-scale feature layers using spatially variable condition maps, enabling structural information at different scales to be adaptively enhanced or suppressed based on location; Second, dynamic convolutional parameters generated by conditional features are introduced in at least one convolutional layer to adaptively change the receptive field, match the spatially variable blurring characteristics, and reduce the risk of over-smoothing; Third, noise branches and detail branches are set, and fusion weights are generated from conditional information to achieve decoupled control of noise suppression and detail compensation, further improving the ability to balance noise and detail, thereby better achieving the technical effects of noise reduction, detail preservation, and quantitative stability improvement under low-count PET conditions.

Claims

1. A deep neural network based PET imaging noise suppression method, characterized in that, include: S1. Acquire the raw PET data of the object to be imaged, perform iterative image reconstruction processing including a point spread function model to obtain an initial PET image, and determine the model parameters corresponding to the point spread function model during the iterative image reconstruction process; S2. Generate a spatially variable condition map that is spatially registered with the initial PET image based on the model parameters, which is used to characterize the changes of the model parameters at different spatial locations in the initial PET image; S3. Perform inference processing of the conditional denoising deep neural network model on the initial PET image using the spatially variable condition map to obtain a noise residual map, a detail compensation map, and a fusion weight map. The conditional denoising deep neural network model includes a feature extraction network and noise and detail branches connected to the feature extraction network. The noise branch outputs the noise residual map, and the detail branch outputs the detail compensation map; S4. Perform voxel-level weighted fusion of the noise residual map and the detail compensation map based on the fusion weight map to obtain a fusion residual map. Perform voxel-level operations on the fusion residual map and the initial PET image to obtain a noise-suppressed PET image.

2. The PET imaging noise suppression method based on deep neural network according to claim 1, characterized in that, S1 includes: Acquire raw PET data of the object to be imaged, which is collected by the PET detector during the PET imaging process; The raw PET data is subjected to data correction processing, which includes at least normalization correction and random compliance correction. A system response for iterative image reconstruction processing is established based on the point spread function model, and the point spread function model is introduced into the system response. Based on the system response, iterative image reconstruction processing is performed on the raw PET data after the data correction process to obtain the initial PET image; The model parameters include parameters used to characterize the spatial broadening properties of the point spread function model, which vary with the spatial position of the initial PET image. 3.The PET imaging noise suppression method based on deep neural network according to claim 1, characterized in that, S2 includes: Obtain the model parameters and the voxel coordinate system information of the initial PET image; For each voxel position in the initial PET image, the broadening parameter of the point spread function model at the voxel position is calculated based on the spatial coordinates of the voxel position in the voxel coordinate system and the model parameters. The spatial variable condition map is generated by using the broadening parameters corresponding to each voxel position in the same voxel arrangement as the initial PET image, so that the spatial variable condition map has the same matrix size and voxel spacing as the initial PET image. The widening parameters include the full width at half height in a predetermined direction.

4. The PET imaging noise suppression method based on a deep neural network according to claim 1, characterized in that, S3 includes: performing intensity normalization processing on the initial PET image to obtain a normalized initial PET image; Perform condition feature extraction processing on the spatial variable condition graph to obtain condition features; In the feature layers of the conditional denoising deep neural network model at multiple scales, scale modulation parameters corresponding to the scales are generated based on the conditional features. The scale modulation parameters include channel scaling parameters for scaling the channel features and channel bias parameters for translating the channel features. The channel scaling parameters and the channel bias parameters are then used to perform channel-level modulation on the features at the corresponding scales. In at least one convolutional layer of the conditional denoising deep neural network model, dynamic convolutional parameters are generated based on the conditional features, and the convolutional operation of the convolutional layer is adaptively adjusted using the dynamic convolutional parameters to change the receptive field of the conditional denoising deep neural network model. The shared features obtained by the feature extraction network are input into the noise branch and the detail branch respectively, so that the noise branch outputs the noise residual map and the detail branch outputs the detail compensation map. The fusion weight map is generated based on the shared features and the conditional features, and the voxel weight values ​​of the fusion weight map are restricted to 0 to 1 by a bounded activation function.

5. The PET imaging noise suppression method based on deep neural networks according to claim 1, characterized in that, S4 include: Numerical truncation is performed on the fused weight map so that the weight values ​​at each voxel position in the fused weight map are between 0 and 1. For each voxel position of the noise residual map and the detail compensation map, the noise residual map and the detail compensation map are summed at the voxel level according to the weight value of the fusion weight map at the voxel position to obtain the fusion residual map; For each voxel position of the initial PET image, the residual value of the fused residual map at the voxel position is compared with the voxel value of the initial PET image at the voxel position. The voxel-level operation includes voxel-level subtraction of the initial PET image to suppress noise and voxel-level addition to compensate for details, to obtain the noise-suppressed PET image.

6. The PET imaging noise suppression method based on a deep neural network according to claim 1, characterized in that, The conditional denoising deep neural network model is obtained through the following training process: acquiring low-count PET raw data and a reference PET image for training, wherein the reference PET image is reconstructed from high-count PET raw data or from high-count data generated by Monte Carlo simulation; performing iterative image reconstruction processing containing the point spread function model on the low-count PET raw data to obtain an initial PET image for training; determining the model parameters corresponding to the point spread function model; and generating a spatially variable condition map that is spatially registered with the initial PET image for training based on the model parameters. The conditional denoising deep neural network model is used to perform forward computation on the initial PET image for training using the spatially variable conditional map to obtain the training output PET image. A loss function is constructed based on the difference between the training output PET image and the reference PET image, and the network parameters of the conditional denoising deep neural network model are updated according to the loss function until the preset training termination condition is met.

7. The PET imaging noise suppression method based on a deep neural network according to claim 2, characterized in that, The iterative image reconstruction process is a TOF-PET iterative reconstruction process, and the model parameters also include parameters characterizing the temporal resolution or TOF kernel width. The spatial variable condition map is further used to characterize the change or preset value of the temporal resolution or TOF kernel width at spatial location.

8. The PET imaging noise suppression method based on a deep neural network according to claim 4, characterized in that, The conditional denoising deep neural network model also outputs an uncertainty map, and the fusion weight map is modified according to the uncertainty map so that the weight of the noise residual map corresponding to the voxel position with higher uncertainty is increased or the weight of the detail compensation map is decreased.

9. A PET imaging noise suppression method based on a deep neural network according to claim 5, characterized in that, After obtaining the noise-suppressed PET image, the process further includes: using the noise-suppressed PET image as a priori initial value or regularization reference, performing iterative image reconstruction processing containing the point spread function model again to obtain a further optimized PET image.

10. A PET imaging noise suppression method based on a deep neural network according to claim 6, characterized in that, The loss function also includes a data consistency loss, which is constructed based on the difference between the projected data obtained by forward projecting the training output PET image according to the system response and the original low-count PET data.

Images