Image denoising model training method, imaging method, scanning equipment and medium

Abstract

The invention discloses an image denoising model training method, an imaging method, scanning equipment and a medium. The method comprises the following steps: constructing a plurality of initial image pairs based on a reference image and a plurality of low-dose images corresponding to the reference image; the reference image comprises a normal dose reconstructed image; generating a training image block data set based on the sampling mask and the plurality of normalized initial image pairs; inputting the training image block data set into the two-stage model to obtain a noise estimation image and a final de-noised image, and training according to the noise estimation image and the final de-noised image to obtain an image de-noising model; the two-stage model comprises networks respectively corresponding to a noise level estimation stage and a noise removal stage. An image denoising model with higher generalization performance is provided, and the image denoising model is applied to a low-dose image, so that the accuracy of image denoising and the quality of a PET reconstructed image are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to an image denoising model training method, an imaging method, a scanning device and a medium. Background Art

[0002] Positron Emission Tomography (PET) plays an important role in basic scientific research such as cognitive science and apoptosis, as well as in the clinical diagnosis of diseases such as cancer, cardiovascular diseases, and neurological diseases. The photon detection efficiency of a conventional whole-body PET imaging system is very low, and the low signal-to-noise ratio of its detection data seriously affects the PET imaging quality. In order to ensure the quality of the reconstructed image, usually, the scanning time is extended or the dose of the injected tracer is increased to improve the data signal-to-noise ratio, thus bringing discomfort and additional radiation risks to patients. The low-dose PET reconstruction algorithm can reconstruct high-quality images with a small amount of data at the software level without increasing the hardware cost of the PET imaging system. Therefore, it has certain scientific research significance and clinical application value.

[0003] Among various deep learning-based low-dose PET algorithms, the post-processing method does not involve the original detection data and the reconstruction process, and the relevant data is easier to collect. The goal of the post-processing model based on the PET reconstructed image is usually to restore the quality of the normal-dose image from the low-dose image, which is essentially an image denoising problem. Summary of the Invention

[0004] The main technical problem to be solved by the present invention is to propose an image denoising model with higher generalization performance, and apply the image denoising model to low-dose images, thereby improving the accuracy of image denoising and the quality of PET reconstructed images.

[0005] According to a first aspect, in one embodiment, an image denoising model training method is provided, which is applied to a positron emission tomography device and includes:

[0006] Constructing a plurality of initial image pairs based on a preset reference image and a plurality of low-dose images corresponding to the reference image; the reference image includes a reconstructed image with a normal dose; the initial image pair includes one reference image and any one of the plurality of low-dose images;

[0007] Generating a training image block data set based on a pre-constructed sampling mask and the plurality of normalized initial image pairs; the sampling mask is constructed according to the reference image; the normalization is calculated according to the percentile parameter corresponding to the plurality of low-dose images;

[0008] Input the training image patch dataset into a pre - constructed two - stage model to obtain a noise - estimation image and a final denoised image, and train an image denoising model based on the noise - estimation image and the final denoised image; the two - stage model includes networks corresponding to a noise level estimation stage and a noise removal stage respectively.

[0009] In some embodiments, generating the training image patch dataset based on the pre - constructed sampling mask and the normalized multiple initial image pairs includes:

[0010] Randomly sample the images in the normalized multiple initial image pairs using the sampling mask to obtain multiple image patches, and aggregate the multiple image patches into a training image patch dataset.

[0011] In some embodiments, perform data interception at different ratios from the detection data corresponding to the reference image to obtain detection data at different ratios after interception; the detection data corresponding to the reference image includes detection data during the imaging process;

[0012] Reconstruct multiple corresponding low - dose images based on the detection data at different ratios after interception; different low - dose images correspond to detection data at different ratios.

[0013] In some embodiments, inputting the training image patch dataset into a pre - constructed two - stage model to obtain a noise - estimation image and a final denoised image includes:

[0014] Generate a noise - estimation image corresponding to the training image patch dataset according to the first - stage network in the two - stage model; the first - stage network includes the network corresponding to the noise level estimation stage;

[0015] Perform concatenation processing on the noise - estimation map and the training image patch dataset, and input the concatenated data into the second - stage network in the two - stage model to obtain a final denoised image; the second - stage network includes the network corresponding to the noise removal stage.

[0016] In some embodiments, training an image denoising model based on the noise - estimation image and the final denoised image includes:

[0017] Construct a corresponding loss function based on the noise - estimation image, the final denoised image, and the image patches corresponding to the reference image in the training image patch dataset;

[0018] Obtain a pre - set training period, a training optimizer, and an initial learning rate, and perform training processing on the two - stage model based on the loss function, the training period, the training optimizer, and the initial learning rate to obtain an image denoising model.

[0019] In some embodiments, percentile value parameters corresponding to the low-dose images in the plurality of initial image pairs are calculated respectively.

[0020] The plurality of initial image pairs are normalized according to the percentile value parameters and a preset normalization formula to obtain the normalized plurality of initial image pairs.

[0021] In some embodiments, the sampling mask is obtained by threshold segmentation of the reference image through a preset empirical threshold.

[0022] According to a second aspect, an imaging method for a positron emission tomography device is provided in an embodiment, including:

[0023] Obtain a three-dimensional low-dose image, and perform image block sampling processing on the three-dimensional low-dose image according to a preset sampling order to obtain a sampled image block.

[0024] Input the sampled image block into a trained image denoising model to obtain a denoised low-dose image; the image denoising model is trained through an image denoising model training method.

[0025] According to a third aspect, a positron emission tomography device is provided in an embodiment, including:

[0026] A sampling component, configured to obtain a three-dimensional low-dose image, and perform image block sampling processing on the three-dimensional low-dose image according to a preset sampling order to obtain a sampled image block.

[0027] A denoising component, configured to input the sampled image block into a trained image denoising model to obtain a denoised low-dose image; the image denoising model is trained through an image denoising model training method.

[0028] According to a fourth aspect, a computer-readable storage medium is provided in an embodiment, characterized in that a program is stored on the medium, and the program can be executed by a processor to implement an image denoising model training method.

[0029] The image denoising model training method, the imaging method of a positron emission tomography device, the positron emission tomography device, and the computer-readable storage medium according to the above embodiments construct a training image patch dataset based on a reference image and a plurality of low-dose images corresponding to the reference image, and train an image denoising model based on the training image patch dataset and a two-stage model. The training image patch dataset contains data with different doses, and at the same time constructs image patches with different noise levels corresponding to the data with different doses, thereby improving the generalization performance of the subsequent model. The two-stage model includes networks corresponding to a noise level estimation stage and a noise removal stage respectively. The two-stage model is trained using the training image patch dataset, and finally accurate denoising of the image denoising model is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flowchart of training an image denoising model according to an embodiment of the present application;

[0031] Figure 2 is a flowchart of training an image denoising model according to an embodiment;

[0032] Figure 3 is a flowchart of training an image denoising model according to an embodiment;

[0033] Figure 4 is a flowchart of training an image denoising model according to an embodiment;

[0034] Figure 5 is an overall structural diagram of a noise level estimation model according to an embodiment;

[0035] Figure 6 is a structural diagram of a backbone network ISS-Unet according to an embodiment;

[0036] Figure 7 is a flowchart of training an image denoising model according to an embodiment;

[0037] Figure 8 is a flowchart of imaging of a positron emission tomography device according to an embodiment;

[0038] Figure 9 is a schematic structural diagram of a positron emission tomography device according to an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present invention will be further described in detail below in conjunction with the specific embodiments and the accompanying drawings. Similar elements in different embodiments are denoted by related similar reference numerals. In the following embodiments, many details are described to enable a better understanding of the present application. However, those skilled in the art can easily recognize that some of the features can be omitted in different situations, or can be replaced by other elements, materials, or methods. In some cases, some operations related to the present application are not shown or described in the specification, which is to avoid the core part of the present application being overwhelmed by excessive description. For those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations based on the description in the specification and the general technical knowledge in the art.

[0040] In addition, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can also be reordered or adjusted in a manner obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for clearly describing a certain embodiment and do not mean that they are the necessary sequences, unless it is stated that a certain sequence must be followed.

[0041] The serial numbers assigned to the components in this article, such as "first", "second", etc., are only used to distinguish the described objects and do not have any sequential or technical meaning. The terms "connection" and "coupling" used in this application, unless otherwise specified, both include direct and indirect connection (coupling).

[0042] Please refer to Figure 1 , some embodiments of the present invention provide an image denoising model training method, including steps S10 to S30, which will be specifically described below.

[0043] Step S10: Construct a plurality of initial image pairs based on a preset reference image and a plurality of low-dose images corresponding to the reference image; the reference image includes a reconstructed image with a normal dose; the initial image pair includes one reference image and any one of the plurality of low-dose images.

[0044] In some embodiments, the reference image refers to a normal-dose PET reconstructed image generated from complete detection data during positron emission tomography (PET) imaging. Here, the normal-dose PET reconstructed image can be understood as an image with 100% dose. The multiple low-dose images corresponding to the reference image can be understood as low-dose images with different noise levels generated by reconstructing a part intercepted from the complete detection data during PET imaging. Here, the intercepted part can be 50%, and in this embodiment, it can include low-dose images of 25%, 10%, 5%, 2%, and 1%.

[0045] In some embodiments, the reference image and the multiple low-dose images corresponding to the reference image can come from different medical center sources.

[0046] Please refer to Figure 2 , in some embodiments, it further includes steps S11 to S12, which are specifically described below.

[0047] Step S11: Intercept data in different proportions from the detection data corresponding to the reference image to obtain detection data with different proportions after interception; the detection data corresponding to the reference image includes the detection data during the imaging process.

[0048] In some embodiments, the different proportions include 25%, 10%, 5%, 2%, and 1%.

[0049] Step S12: Reconstruct corresponding multiple low-dose images according to the detection data with different proportions after interception; different low-dose images correspond to detection data with different proportions.

[0050] In some embodiments, the reference image and the corresponding multiple low-dose images have the same size. Here, size parameters of 360×360×640 can be set with reference to the three dimensions of x, y, and z, and the pixel size is 1.667×1.667×2.886 mm 3 .

[0051] In some embodiments, for the reference image and the corresponding multiple low-dose images, each low-dose image can form a group of initial image pairs with the reference image. The initial image pair can be denoted as {I L ,I H}, where I L represents the low-dose image and I H represents the reference image. Taking the 25% low-dose image as an example, its corresponding initial image pair can be denoted as {I 25％ ,I 100％}.

[0052] Step S20: Generate a training image patch dataset based on a pre-constructed sampling mask and multiple normalized initial image pairs; the sampling mask is constructed according to a reference image; the normalization is calculated according to the percentile value parameters corresponding to multiple low-dose images.

[0053] Please refer to Figure 3 , in some embodiments, step S20 further includes steps S21 to S22, which are specifically described below.

[0054] Step S21: Calculate the percentile value parameters corresponding to the low-dose images in multiple initial image pairs respectively.

[0055] In some embodiments, the 99.9% percentile value parameter IL corresponding to different low-dose images is calculated respectively -99.9％ . Among them, the percentile value parameter IL corresponding to the low-dose image -99.9％ can be recorded and saved for restoring the denoised image to the scale of the original pixel value later.

[0056] Step S22: Perform normalization processing on multiple initial image pairs according to the percentile value parameters and a preset normalization formula to obtain multiple normalized initial image pairs.

[0057] In some embodiments, performing normalization processing on multiple initial image pairs includes performing image normalization processing on the reference image and the corresponding low-dose image in the initial image pair respectively, where the normalization formula during normalization processing is as follows:

[0058]

[0059]

[0060] Among them, I L-norm represents the normalized low-dose image, I H-norm represents the normalized reference image, I L represents the low-dose image, I H represents the reference image, I L-99.9％ represents the percentile value parameter corresponding to the low-dose image.

[0061] In some embodiments, a sampling mask is obtained by threshold segmentation of the reference image according to a preset empirical threshold. Among them, the preset empirical threshold is 30, that is, the normal-dose image is threshold-segmented using the preset empirical threshold 30 to exclude the area where the pixel value is close to 0, and the general area of the human body is obtained and used as the sampling mask.

[0062] In some embodiments, generating a training image patch dataset based on the sampling mask and multiple normalized initial image pairs includes:

[0063] Randomly sample the images in multiple normalized initial image pairs using a sampling mask to obtain multiple image patches, and aggregate the multiple image patches into a training image patch dataset.

[0064] In some embodiments, the image patch pairs in the training image patch dataset can be represented as {P L , P H}, where P L represents a low-dose image patch, and P H represents a reference image patch.

[0065] In some embodiments, randomly sample image patches on multiple normalized initial image pairs using a sampling mask, crop to obtain multiple image patches with a size of 128×128×128, aggregate the multiple image patches into a training image patch dataset, and use it for subsequent model training. Among them, based on the sampling mask for random sampling, sampling can be performed in the area where the pixel value is higher than the preset empirical threshold, thereby improving the efficiency of model training.

[0066] Step S30: Input the training image patch dataset into a pre-constructed two-stage model to obtain a noise estimation image and a final denoised image, and train an image denoising model based on the noise estimation image and the final denoised image; the two-stage model includes networks corresponding to the noise level estimation stage and the noise removal stage respectively.

[0067] Please refer to Figure 4 In some embodiments, step S30 inputs the training image patch dataset into a pre-constructed two-stage model to obtain a noise estimation image and a final denoised image, including steps S31 to S32, which are specifically described below.

[0068] Step S31: Generate a noise estimation image corresponding to the training image patch dataset according to the first-stage network in the two-stage model; the first-stage network includes the network corresponding to the noise level estimation stage.

[0069] Please refer to Figure 5, in some embodiments, the pre-built two-stage model refers to a noise-level estimation model (Noise-Level-Estimation Net, NLE-Net). Among them, the noise-level estimation model consists of a noise-level estimation stage and a noise removal stage. The main structures of the two stages both use the same backbone network ISS-Unet. The first-stage network can be regarded as the network corresponding to the noise-level estimation stage, which is used to generate a noise estimation image reflecting the noise levels corresponding to different low-dose images. Among them, in the natural image blind denoising model, the noise map is generated by simulation. By controlling the addition of different levels of noise to the reference image, noise maps and accurate noise estimation maps with different noise levels can be obtained. However, applying the natural image blind denoising model to the blind denoising scenario of PET images cannot obtain an accurate noise estimation map. Therefore, the difference between the input image x and the preliminary denoising result x' corresponding to the training image patch dataset is calculated through the first-stage network as an approximation of the actual noise estimation map, that is, the noise estimation image n. Among them, the input image x is a low-dose image.

[0070] Please refer to Figure 6, in some embodiments, the backbone network ISS-Unet has a symmetric structure. Its left contraction path includes a feature encoding module and multiple downsampling modules with the same structure, and each module is connected in sequence. First, a pixel reshuffle layer PixeUnshuffle rearranges a five-dimensional tensor with the shape of [batchsize, 1, 128, 128, 128] into a tensor with the shape of [batchsize, 8, 64, 64, 64]. Here, batchsize is the batch size during the training process. In this embodiment, batchsize is 16. Secondly, a linear convolutional layer Conv module is used for feature encoding. The Conv module consists of 32 three-dimensional convolutional layers with a kernel size of 3, a stride of 1, and a padding of 1, which can be expressed as 3×3×3Conv(stride 1). Through the linear convolutional layer Conv module, the tensor with the shape of [batchsize, 8, 64, 64, 64] is encoded into a tensor with the size of [batchsize, 32, 64, 64, 64]. Then, convolutional modules and downsampling modules with the same structure are used for processing. Among them, the convolutional module consists of a linear convolutional layer Conv and an activation layer LeakyRelu, which can be expressed as [Conv, LeakyRelu, Conv, LeakyRelu]. The number of kernels in the linear convolutional layer is 32, the stride is 1, and the edge padding is 1. The downsampling module consists of a linear convolutional layer with 64 kernels, a stride of 2, and an edge padding of 1 and a LeakyRelu activation layer, so it can be expressed as 3×3×3Conv(stride 2)+LeakyReLU. Whenever a downsampling is performed, the number of kernels in the subsequent convolution is doubled, and the size of the feature map is reduced by half in each dimension. Among them, when processed by the convolutional module and the downsampling module, it can be expressed as [3×3×3Conv(stride1)+LeakyReLU]×2. After four downsampling processes, a feature map with the size of [batchsize, 512, 4, 4, 4] is finally obtained. At the bottom of this backbone network, a convolutional module is still used, and then it can enter the right expansion path. In some embodiments, the right expansion path maintains multiple upsampling modules and feature decoding modules with the same structure. Among them, the first upsampling module consists of a transposed convolutional layer ConvTranspose with 256 kernels and a stride of 2 and a LeakyRelu activation layer. When the first upsampling module is used for data processing, it can be expressed as 3×3×3ConvTranspose(stride2)+LeakyReLU. The tensor obtained by the upsampling module has the same shape as the input tensor of the corresponding downsampling module on the left, both of which are [batchsize, 256, 8, 8, 8].The output of the upsampling module is concatenated with the input tensor of the corresponding downsampling module on the left along the second channel dimension to form a tensor of [batchsize, 512, 8, 8, 8], which serves as the input to the subsequent convolutional decoding module. This convolutional decoding module is also composed of connected linear convolutional layers and activation layers, denoted as [Conv, LeakyRelu, Conv, LeakyRelu]. The number of kernels in its convolutional layers is 256, with a size of 3, a stride of 1, and a padding of 1. After each downsampling, the number of subsequent convolutional kernels is reduced to half of the previous one, and the size of the feature map expands to twice the previous size in each dimension. After a total of four upsamplings, a feature map of size [batchsize, 32, 64, 64, 64] is finally obtained. Through an addition operation, the features encoded on the left are short-circuited with the features here to achieve feature fusion, and the size of the output feature map remains [batchsize, 32, 64, 64, 64]. Then, a linear convolutional layer with 8 kernels, a size of 3, a stride of 1, and a padding of 1 is used to decode the feature map into [batchsize, 8, 64, 64, 64], and finally, a pixel shuffle layer is used to convert the feature map into a size of [batchsize, 1, 128, 128, 128].

[0071] In some embodiments, regarding the structural parameters of the above-mentioned backbone network ISS-Unet, for example, parameters such as the number of convolutional layers, the number of convolutional kernels, and the kernel size can be adjusted according to the actual situation. The number of downsamplings and upsamplings can be synchronously increased to five or decreased to three, and the number of kernels in the first encoding layer can be adjusted from 32 in the embodiment to 16 or 64. In such cases, the backbone network ISS-Unet can still achieve similar functions.

[0072] Step S32: Concatenate the noise estimation map and the training image patch dataset, and input the concatenated data into the second-stage network in the two-stage model to obtain the final denoised image; the second-stage network includes the network corresponding to the noise removal stage.

[0073] In some embodiments, the noise estimation image n is concatenated with the corresponding input image x in the training image patch dataset, and the concatenated data is input into the second-stage model for data processing to obtain the final denoised image x * 。

[0074] Please refer to Figure 7 , in some embodiments, step S30 trains an image denoising model based on the noise estimation image and the final denoised image, including steps S33 to S34, which are specifically described below.

[0075] Step S33: Construct a corresponding loss function based on the noise estimation image, the final denoised image, and the image patches corresponding to the reference images in the training image patch dataset.

[0076] In some embodiments, the loss function is as follows:

[0077]

[0078] where L(P H , f1(P L ), f(P L )) represents the loss function, P L represents the low-dose image patch corresponding to it in the training image patch dataset, P H represents the reference image patch corresponding to it in the training image patch dataset, f1(P L ) represents the preliminary denoising result corresponding to the noise estimation image, f(P L ) represents the final denoised image, f1 represents the first-stage network, f represents the two-stage model, and ||·||1 represents the L1 norm.

[0079] In some embodiments, f1(P L ) represents the preliminary denoising result x' corresponding to the noise estimation image, and f(P L ) represents the final denoised image, that is, the final denoised image x * .

[0080] Step S34: Obtain a preset training cycle, a training optimizer, and an initial learning rate, and perform training processing on the two-stage model based on the loss function, the training cycle, the training optimizer, and the initial learning rate to obtain an image denoising model.

[0081] In some embodiments, traversing all the image pairs in the training image patch dataset once is regarded as one cycle. The training cycle is preset to 20 cycles, the training optimizer is Adam, and the initial learning rate is set to 2e -4 . During the training process, the initial learning rate will gradually decrease to 2e -5 . This two-stage model will be trained on a GPU with 24GB of memory.

[0082] In some embodiments, by mixing image data from different sources and at different dose levels, and at the same time constructing a training image patch dataset in the form of image patches to uniformly train a single model, the finally obtained trained image denoising model can be specifically trained and processed according to the noise level differences between different individuals and different body regions, and finally blind denoising of PET images is achieved. Among them, for image data at different dose levels, only the corresponding normal-dose images need to be provided to obtain the corresponding different low-dose image data. The model can be trained without providing specific dose levels, and can receive any low-dose level noise image as input during prediction, with stronger generalization ability. During the model training process, only a single two-stage model is trained, and there is no need to manually design noise level estimation parameters. Therefore, the model training process is simplified. At the same time, the network structure of the two-stage model is simple and the computational load is low, enabling fast image processing.

[0083] Reference Figure 8 , in some embodiments, an imaging method for a positron emission tomography device includes steps S40 to S50, which will be specifically described below.

[0084] Step S40: Obtain a three-dimensional low-dose image, and perform image patch sampling processing on the three-dimensional low-dose image according to a preset sampling order to obtain sampled image patches.

[0085] In some embodiments, the preset sampling order is from left to right and from top to bottom. The three-dimensional low-dose image is sampled for image patches in sequence according to the sampling order from left to right and from top to bottom. The size of the image patch is 128×128×128, and the sampling step is 96. Therefore, during the actual sampling process, there is an overlap of 32 pixels in length between adjacent image patches, and edge padding is used to achieve integer multiple sampling, where the padding value is 0.

[0086] Step S50: Input the sampled image patches into the trained image denoising model to obtain the denoised low-dose image; the image denoising model is trained through the image denoising model training method.

[0087] In some embodiments, after the image patch sampling is completed, it is input into the trained image denoising model for prediction. The prediction results are stitched together in the input order, and the overlapping part of the image patches is weighted using a Hann window function, and finally a denoised low-dose image with the same size as the input three-dimensional low-dose image is formed to achieve blind denoising of low-dose PET images.

[0088] Among them, the trained image denoising model is trained through the following model training method, which will be specifically described below.

[0089] Construct a plurality of initial image pairs based on a preset reference image and a plurality of low-dose images corresponding to the reference image. Among them, the reference image includes a reconstructed image with a normal dose, and the initial image pair includes one reference image and any one of the plurality of low-dose images. Generate a training image patch dataset based on a pre-constructed sampling mask and the normalized plurality of initial image pairs. Among them, the sampling mask is constructed according to the reference image, and the normalization is calculated according to the percentile value parameters corresponding to the plurality of low-dose images. Input the training image patch dataset into a pre-constructed two-stage model to obtain a noise estimation image and a final denoised image, and train an image denoising model according to the noise estimation image and the final denoised image. Among them, the two-stage model includes networks corresponding to a noise level estimation stage and a noise removal stage respectively.

[0090] Reference Figure 9 , in some embodiments, a positron emission tomography device includes a sampling component 10 and a denoising component 20, which will be specifically described below.

[0091] The sampling component 10 is used to obtain a three-dimensional low-dose image, and perform image patch sampling processing on the three-dimensional low-dose image according to a preset sampling order to obtain a sampled image patch.

[0092] In some embodiments, the preset sampling order is from left to right and from top to bottom. The three-dimensional low-dose image is sampled in image patches in sequence according to the sampling order from left to right and from top to bottom. The size of the image patch is 128×128×128, and the sampling step is 96. Therefore, in the actual sampling process, there is an overlap of 32 pixels in length between adjacent image patches, and edge padding is used to achieve integer multiple sampling, where the padding value is 0.

[0093] The denoising component 20 is used to input the sampled image patch into a trained image denoising model to obtain a denoised low-dose image; the image denoising model is trained through an image denoising model training method.

[0094] In some embodiments, after the image patch sampling is completed, it is input into a trained image denoising model for prediction. The prediction results are stitched together in the input order. The overlapping part of the image patches is weighted by a Hann window function, and finally a denoised low-dose image with the same size as the input three-dimensional low-dose image is formed to achieve blind denoising of low-dose PET images.

[0095] Among them, the trained image denoising model is trained through the following model training method, which will be specifically described below.

[0096] Construct multiple initial image pairs based on a preset reference image and multiple low-dose images corresponding to the reference image, where the reference image includes a reconstructed image with a normal dose, and the initial image pair includes one reference image and any one of the multiple low-dose images. Generate a training image patch dataset based on a pre-constructed sampling mask and the normalized multiple initial image pairs, where the sampling mask is constructed according to the reference image, and the normalization is calculated according to the percentile parameter corresponding to the multiple low-dose images. Input the training image patch dataset into a pre-constructed two-stage model to obtain a noise estimation image and a final denoised image, and train an image denoising model based on the noise estimation image and the final denoised image, where the two-stage model includes networks corresponding to a noise level estimation stage and a noise removal stage respectively.

[0097] Those skilled in the art can understand that all or part of the functions of the above methods can be implemented in a hardware manner or in a computer program manner. When all or part of the functions in the above embodiments are implemented in a computer program manner, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, magnetic disk, optical disk, hard disk, etc. The above functions can be realized by a computer executing the program. For example, the program is stored in the memory of the device, and when the processor executes the program in the memory, the above all or part of the functions can be realized. In addition, when all or part of the functions in the above embodiments are implemented in a computer program manner, the program can also be stored in a storage medium such as a server, another computer, magnetic disk, optical disk, flash drive or mobile hard disk, downloaded or copied and saved to the memory of the local device, or the system of the local device is updated. When the processor executes the program in the memory, the above all or part of the functions in the above embodiments can be realized.

[0098] The above uses specific examples to elaborate on the present invention, which is only for helping to understand the present invention and is not used to limit the present invention. For those skilled in the art of the present invention, according to the idea of the present invention, several simple deductions, deformations or substitutions can also be made.

Claims

1. A method for training an image denoising model, applied to a positron emission tomography device, characterized in that: include: constructing a plurality of initial image pairs based on a preset reference image and a plurality of low-dose images corresponding to the reference image; The reference image includes a reconstructed image of a normal dose; the initial image pair includes a reference image and any one of a plurality of low-dose images; Generate a training image patch dataset based on the pre-constructed sampling mask and the normalized multiple initial image pairs; The sampling mask is constructed based on the reference image; The normalization is calculated according to percentile value parameters corresponding to the plurality of low-dose images; The training image block dataset is input into a pre-built two-stage model to obtain a noise estimation image and a final denoised image, and an image denoising model is obtained by training the noise estimation image and the final denoised image; the two-stage model includes networks corresponding to the noise level estimation stage and the noise removal stage, respectively.

2. The method according to claim 1, characterized in that The step of generating a training image block dataset based on the pre-constructed sampling mask and the normalized multiple initial image pairs includes: The images in the normalized multiple initial image pairs are randomly sampled using the sampling mask to obtain multiple image blocks, and the multiple image blocks are aggregated into a training image block data set.

3. The method according to claim 1, characterized in that Extracting data of different proportions from the detection data corresponding to the reference image to obtain the intercepted detection data of different proportions; the detection data corresponding to the reference image includes the detection data in the imaging process; Multiple corresponding low-dose images are reconstructed based on the intercepted detection data of different proportions; different low-dose images correspond to detection data of different proportions.

4. The method according to claim 1, characterized in that The step of inputting the training image block dataset into a pre-built two-stage model to obtain a noise estimation image and a final denoised image comprises: Generate a noise estimation image corresponding to the training image block data set according to the first stage network in the two-stage model; the first stage network includes a network corresponding to the noise level estimation stage; The noise estimation map and the training image block data set are processed in series, and the series data are input into the second stage network in the two-stage model to obtain a final denoised image; the second stage network includes a network corresponding to the noise removal stage.

5. The method according to claim 1, characterized in that The step of training the image denoising model according to the noise estimation image and the final denoised image comprises: Constructing a corresponding loss function according to the noise estimation image, the final denoised image and the image blocks corresponding to the reference image in the training image block data set; Obtain a preset training cycle, a training optimizer and an initial learning rate, and perform training processing on the two-stage model based on the loss function, the training cycle, the training optimizer and the initial learning rate to obtain an image denoising model.

6. The method according to claim 1, characterized in that respectively calculating percentile value parameters corresponding to the low-dose images of the plurality of initial image pairs; The plurality of initial image pairs are normalized according to the percentile value parameter and a preset normalization formula to obtain the plurality of normalized initial image pairs.

7. The method according to claim 1, characterized in that The sampling mask is obtained by performing threshold segmentation on the reference image through a preset empirical threshold.

8. An imaging method of a positron emission tomography device, characterized in that: include: Acquire a three-dimensional low-dose image, and perform image block sampling processing on the three-dimensional low-dose image according to a preset sampling order to obtain a sampled image block; The sampled image block is input into a trained image denoising model to obtain a denoised low-dose image; the image denoising model is trained by the training method according to any one of claims 1 to 7.

9. A positron emission tomography device, characterized in that: include: A sampling component, used for acquiring a three-dimensional low-dose image, and performing image block sampling processing on the three-dimensional low-dose image according to a preset sampling sequence to obtain a sampled image block; A denoising component is used to input the sampled image block into a trained image denoising model to obtain a denoised low-dose image; the image denoising model is trained by the training method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The medium stores a program, which can be executed by a processor to implement the method according to any one of claims 1 to 7.

Images