A nuclear medicine image denoising method, device and storage medium

Abstract

This invention provides a method for denoising nuclear medicine images, comprising: acquiring full-dose nuclear medicine projection data with different noise levels and reconstructing the images to obtain full-dose nuclear medicine images with different noise levels; downsampling and reconstructing the images of each full-dose nuclear medicine projection data to obtain low-dose nuclear medicine images; constructing a dataset by correlating the low-dose nuclear medicine images with the full-dose nuclear medicine images; constructing a neural network model with noise level embedding; training the neural network model using the low-dose nuclear medicine images in the dataset as input and the full-dose nuclear medicine images as labels to obtain a denoising model; inputting the low-dose nuclear medicine image to be denoised into the denoising model and outputting the denoised nuclear medicine image. This invention can significantly reduce the noise level of low-dose nuclear medicine images and improve image quality.

Description

Technical Field

[0001] This invention relates to the field of nuclear medicine image processing technology, and in particular to a method, apparatus, electronic device, and storage medium for nuclear medicine image denoising. Background Technology

[0002] Single-photon emission computed tomography (SPECT) and positron emission tomography (PET), as important nuclear medicine imaging techniques, play a crucial role in the diagnosis and localization of bone metastases. These techniques visualize the metabolic distribution of radiopharmaceutical tracers in bone; however, clinical practice has shown a positive correlation between image quality and radiopharmaceutical injection dose and scan time. To reduce patient radiation exposure and improve examination comfort, the medical community follows the ALARA (Lowest Feasible Radical) principle, often employing low-dose (generally 10%–50%, sometimes even ultra-low doses of 5%–10%, combined with efficient denoising or super-resolution reconstruction algorithms to compensate for signal-to-noise ratio losses) injections and shortened scan times. However, this can lead to insufficient projection data, resulting in increased image noise, reduced contrast, and errors in quantitative analysis. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] The purpose of this invention is to provide a method, apparatus, electronic device, and storage medium for nuclear medicine image denoising based on self-attention enhanced residual U-Net. This invention can effectively remove noise in low-dose nuclear medicine images while maintaining the structural and detailed information of the images, providing an accurate basis for subsequent diagnosis.

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

[0006] The first aspect of this invention provides a method for denoising nuclear medicine images, comprising:

[0007] Full-dose nuclear medicine projection data with different noise levels were acquired and image reconstruction was performed to obtain full-dose nuclear medicine images with different noise levels. Low-dose nuclear medicine images were obtained by downsampling and image reconstruction of each full-dose nuclear medicine projection data. The low-dose nuclear medicine images were matched with the full-dose nuclear medicine images to construct a dataset.

[0008] Construct a neural network model with noise level embedding;

[0009] Using low-dose nuclear medicine images from the dataset as input and full-dose nuclear medicine images as labels, the neural network model is trained to obtain a denoising model.

[0010] Input the low-dose nuclear medicine image to be denoised into the denoising model, and output the denoised nuclear medicine image.

[0011] In some embodiments, the nuclear medicine projection data uses SPECT projection data. Full-dose SPECT projection data with different noise levels are obtained by referring to the clinical nuclear medicine image acquisition procedure. The images are then reconstructed, and combined with attenuation correction based on CT images, n iterations and m subset reconstructions are performed to obtain full-dose SPECT images with different noise levels.

[0012] In some embodiments, the downsampling ratio is set to α, where α takes any value in (0,1).

[0013] In some embodiments, the neural network model with noise level embedding employs an improved U-net network, which is a self-attention-enhanced residual U-net network with noise level embedding. It includes an input layer, a first encoding layer, a second encoding layer, a third encoding layer, a first attention block, a bottleneck layer, a second attention block, a first decoding layer, a third attention block, a second decoding layer, a third decoding layer, and an output layer, arranged sequentially. Skip connection layers are provided between the first encoding layer and the third decoding layer, between the second encoding layer and the second decoding layer, and between the third encoding layer and the first decoding layer. Each encoding layer, each decoding layer, and the bottleneck layer employs a corresponding residual block. Multiple noise level metrics are embedded into each residual block through a positional encoding layer and two fully connected layers. These multiple noise level metrics include at least the downsampling ratio, and also the average intensity and standard deviation of the low-dose nuclear medicine image.

[0014] In some embodiments, each residual block includes a connected backbone and a bypass. The backbone contains two 3×3×3 convolutions with ReLU activation, and the bypass contains a 1×1×1 pointwise convolution with ReLU activation. The output of the three convolutions is added to the residual block input channel by channel.

[0015] In some embodiments, each attention block consists of a flattened layer, a multi-head attention mechanism, and an inverse flattened layer connected in sequence.

[0016] In some embodiments, before training the neural network model, a patch-based strategy is used to augment the low-dose nuclear medicine images in the dataset. Then, leave-one-out cross-validation is used to train and test the constructed neural network model with the augmented low-dose nuclear medicine images as input and the full-dose nuclear medicine images as labels to obtain the denoising model.

[0017] In some embodiments, the leave-one-out cross-validation method includes: dividing the dataset into five equal parts, where 80% of the dataset is used as the training dataset and the remaining 20% ​​of the dataset is used as the test dataset, performing five training and validation experiments to generate five training models.

[0018] The training dataset and the test dataset are from different scanned objects.

[0019] A second aspect of the present invention provides a nuclear medicine image denoising apparatus, comprising:

[0020] The first module is configured to acquire full-dose nuclear medicine projection data with different noise levels and perform image reconstruction to obtain full-dose nuclear medicine images with different noise levels. Low-dose nuclear medicine images are obtained by downsampling and image reconstruction of each full-dose nuclear medicine projection data. The low-dose nuclear medicine images are matched with the full-dose nuclear medicine images to construct a dataset.

[0021] The second module is configured to build a neural network model with noise level embeddings;

[0022] The third module is configured to use low-dose nuclear medicine images from the dataset as input and full-dose nuclear medicine images as labels to train the neural network model and obtain a denoising model.

[0023] The fourth module is configured to input the low-dose nuclear medicine image to be denoised into the denoising model and output the denoised nuclear medicine image.

[0024] A third aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to perform a nuclear medicine image denoising method according to any embodiment of the first aspect of the present invention.

[0025] The features and beneficial effects of this disclosure are as follows:

[0026] This invention presents an image denoising method based on Self-Attention Enhanced Residual U-Net (ARUN). By combining residual learning and a self-attention mechanism, it effectively removes noise from low-dose SPECT images while improving image quality and preserving image details. This method can adaptively process images with different noise levels and possesses strong versatility and efficiency. It is not only applicable to SPECT images but can also be extended to denoising other nuclear medicine images. Through optimization of the self-attention mechanism, the network can focus on key regions, improving the denoising effect and ensuring that the structure and details of the denoised image are more complete, ultimately providing higher-quality image support for medical imaging diagnosis. Attached Figure Description

[0027] Figure 1This is an overall flowchart of the nuclear medicine image denoising method provided by the first aspect of the present invention.

[0028] Figure 2 This is a schematic diagram of a U-Net (ARUN) network architecture based on self-attention enhanced residuals, provided as an embodiment of the present invention.

[0029] Figure 3 This is a comparison of the denoising effects before and after low-dose SPECT images.

[0030] Figure 4 A schematic diagram of the structure of an electronic device provided according to a third aspect embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining this application and are not intended to limit this application.

[0032] Conversely, this application covers any alternatives, modifications, equivalent methods, and schemes made within the spirit and scope of this application as defined by the claims. Furthermore, to provide the public with a better understanding of this application, certain specific details are described in detail below. However, this application can be fully understood by those skilled in the art even without these detailed descriptions.

[0033] See Figure 1 The first aspect of the present invention provides a method for denoising nuclear medicine images, comprising:

[0034] Step S100: Obtain full-dose nuclear medicine projection data with different noise levels and perform image reconstruction to obtain full-dose nuclear medicine images with different noise levels. Downsample and reconstruct the images of each full-dose nuclear medicine projection data to obtain low-dose nuclear medicine images. Match the low-dose nuclear medicine images with the full-dose nuclear medicine images to construct a dataset.

[0035] Step S200: Construct a neural network model with noise level embedding;

[0036] Step S300: Using the low-dose nuclear medicine images in the dataset as input and the full-dose nuclear medicine images as labels, train the above neural network model to obtain a denoising model;

[0037] Step S400: Input the low-dose nuclear medicine image to be denoised into the denoising model and output the denoised nuclear medicine image.

[0038] In some embodiments, the nuclear medicine projection data used in step S100 is SPECT projection data.

[0039] In some embodiments, step S100 specifically includes:

[0040] First, full-dose SPECT projection data at different noise levels were obtained according to clinical SPECT acquisition procedures. SPECT imaging used the default protocol, employing a standard circular rotational track covering 360 degrees, with 30 steps every 6 degrees, and each step lasting 20 seconds. The energy window was set to 126 kEV to 156 kEV, and data was recorded in list mode. The pixel size of the projection data was 4.1 × 4.1 mm. 2 The list pattern data was reconstructed using existing image reconstruction algorithms (such as the OSEM algorithm), combined with attenuation correction based on CT images for n iterations and reconstruction of m subsets, where n and m are natural numbers greater than 0. The CT images were obtained through CT scans using a clinical dual-head SPECT system (Chengdu Novel Medical Equipment Ltd.'s Insight NM / CT Pro), resulting in full-dose SPECT images with different noise levels. The reconstructed image size was 128×128×128, and each voxel in the reconstructed image was 4.1×4.1×4.1mm. 3 The cross-sectional area of ​​a single voxel in a full-dose nuclear medicine image needs to be consistent with the pixel size of the full-dose nuclear medicine projection data.

[0041] Then, using software provided by the supplier, low-dose SPECT projection data is generated from the list pattern data based on different downsampling ratios. Let the downsampling ratio be α, where α takes any value in (0,1). In a specific embodiment of the present invention, the downsampling ratios α are 0.1, 0.2, and 0.33, respectively, to obtain low-dose SPECT projection data with counts of 10%, 20%, and 33.3%.

[0042] Finally, low-dose SPECT images are obtained by reconstructing the low-dose SPECT projection data, which is used to train the denoising model. Meanwhile, full-dose SPECT images are obtained by reconstructing the SPECT projection data with 100% count, which is used as a label for model training.

[0043] In some embodiments, the neural network model with noise level embedding constructed in step S200 is an improved U-net network, specifically a self-attention enhanced residual U-net network with noise level embedding. See also Figure 2The improved U-net network proposed in this embodiment is based on the traditional U-net network. It replaces the two standard convolutional layers used in the original coding layer, decoding layer, and bottleneck layer with residual blocks, and passes multiple noise level metrics through the position coding layer (see [link to documentation]). Figure 2 In the middle PosEncoding (in this embodiment, a sinusoidal position encoding layer) and two fully connected layers (see PosEncoding), Figure 2 The FC (Functionally Selective) layer is embedded into each residual block. Attention blocks are added after the third coding layer, the bottleneck layer, and the first decoding layer, respectively. Among the multiple noise level metrics, at least the downsampling ratio α is included, and the average intensity and standard deviation of the low-dose nuclear medicine image may also be included. Specifically, the U-net network constructed in this embodiment includes an input layer, a first coding layer, a second coding layer, a third coding layer, a first attention block, a bottleneck layer, a second attention block, a first decoding layer, a third attention block, a second decoding layer, a third decoding layer, and an output layer arranged sequentially. The input layer includes a first convolutional layer and a first activation layer connected to it. The activation function of the first activation layer is ReLU. The input layer is used to perform preliminary feature extraction on the noisy low-dose nuclear medicine image, resulting in a feature map with 32 channels and a spatial size of 48×48×48. Skip connection layers are set between the first encoding layer and the third decoding layer, between the second encoding layer and the second decoding layer, and between the third encoding layer and the first decoding layer. Each encoding layer, each decoding layer, and the bottleneck layer uses a corresponding residual block. Each residual block includes a connected backbone and bypass. The backbone contains two ReLU activation layers. The active 3×3×3 convolution has a 1×1×1 pointwise convolution with ReLU activation in the bypass (this pointwise convolution is used to achieve channel or size alignment). The output of this cubic convolution is added to the residual block input channel by channel to achieve cross-layer residual learning, which is beneficial for noise suppression and feature fusion while preserving structural details. Each attention block is composed of a flattened layer, a multi-head attention mechanism, and an unflattened layer connected in sequence. Attention blocks are used to introduce global context dependencies and enhance the U-net network's ability to focus on long-range structures.

[0044] In some embodiments, step S300 includes:

[0045] First, the low-dose nuclear medicine images in the dataset constructed in step S100 are augmented using a patch-based strategy. Since the downsampling ratio alone cannot fully characterize the noise level, this embodiment of the invention also embeds the average intensity and standard deviation of the low-dose nuclear medicine images (which refers to the average intensity and standard deviation of the voxel gray values ​​of a single patch, which can be understood as the average intensity and standard deviation of the gray values ​​of a single patch obtained from a single image) into the input features.

[0046] Then, leave-one-out cross-validation was used to train and test the improved U-net network with the enhanced low-dose nuclear medicine images as input and the full-dose nuclear medicine images as labels, to obtain the denoising model (hereinafter referred to as "ARUN network").

[0047] Specifically, in this embodiment of the invention, nuclear medicine images with a 10% dose (i.e., α = 0.1) are labeled G-10, those with a 20% dose (i.e., α = 0.2) are labeled G-20, and those with a 33.3% dose (i.e., α = 0.33) are labeled G-33. Full-dose nuclear medicine images are referred to as labeled images. All three groups of low-dose nuclear medicine images (G-10, G-20, and G-33) are used to train the ARUN network. Leave-one-out cross-validation is used for model training, with 80% of the dataset used for training and the remaining 20% ​​for testing, resulting in five training models. Patients in the training and testing datasets are mutually exclusive to ensure no overlap between training and testing.

[0048] Furthermore, data augmentation employed a patch-based strategy with a patch size of 48×48×48. The batch size was set to 16. The model was optimized using the L1 loss function. The Adam optimizer had an initial learning rate of 0.0005 and a decay factor of 0.02. The network was trained for 55,000 epochs, resulting in the ARUN network. The expression for the L1 loss function is:

[0049]

[0050] in, This represents the true value of the m-th sample, which is the "full-dose" high-quality patch for the m-th sample. Let be the prediction value of the ARUN model for the m-th sample, and M be the total number of samples used to train the improved U-net network.

[0051] Validation of the embodiments of the present invention:

[0052] To verify the denoising effect of the ARUN network obtained in this embodiment, several other models were also trained based on U-Net and Res-U-Net and compared using the same hyperparameters. All trained models are summarized in Table 1. The experiments were conducted on a workstation equipped with an NVIDIA GPU A6000 graphics card.

[0053] Table 1 Description of all trained models

[0054]

[0055] The performance of all trained models was evaluated using standardized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and lesion bias between full-dose nuclear medicine images and denoised low-dose nuclear medicine images. Using full-dose nuclear medicine images as a reference, NRMSE was defined as follows:

[0056]

[0057] Where, x i and x' i represents the value of the i-th voxel in the denoised low-dose nuclear medicine image and the full-dose nuclear medicine image, respectively; N is the total number of voxels in a single full-dose nuclear medicine image.

[0058] The peak signal-to-noise ratio (PSNR) is defined as follows:

[0059]

[0060] SSIM is defined as:

[0061]

[0062] Where, μ x and μ y These are the average values ​​of low-dose and full-dose nuclear medicine images, respectively, σ. x and σ y These are the standard deviations of low-dose and full-dose nuclear medicine images, respectively, σ. xy It is the cross-covariance between low-dose nuclear medicine images and full-dose nuclear medicine images. c1 and c2 are two small constants set to avoid the denominator being too small.

[0063] To quantitatively assess lesion bias, this embodiment of the invention manually delineated regions of interest (ROIs) on full-dose nuclear medicine images, obtaining ROIs for a total of 162 nodules (441.1 ± 665.1 voxels). Simultaneously, this embodiment of the invention plotted a 5×5×5 voxel ROI inside the lungs of each subject as the background. The normalized bias between the lesion and the background was calculated using the following formula:

[0064]

[0065] Where, λ n and λ' n These represent the average radioactivity of the regions of interest (ROIs) in the denoised image (i.e., the denoised low-dose nuclear medicine image) and the reference image (i.e., the full-dose nuclear medicine image), respectively.

[0066] In addition, the contrast-to-noise ratio (CNR) of the lesion is calculated as follows:

[0067]

[0068] Where, λ lesion and λ background δ represents the mean radioactivity of the lesion and background in the image (referring to low-dose nuclear medicine image or full-dose nuclear medicine image), respectively. lesion and δ background These represent the activity standard deviations of the lesions and the background in the image, respectively.

[0069] To compare the NRMSE, SSIM, and PSNR of different models, this embodiment of the invention uses a two-tailed paired t-test for datasets from the same group and a two-tailed independent t-test for datasets from different groups. A p-value less than 0.05 is considered statistically significant.

[0070] like Figure 3 As shown, all models significantly suppress noise compared to the original low-dose nuclear medicine images. For the G-20 dataset, the denoised images reconstructed by UNet-G20 and RNet-G20 show excellent performance in terms of structure preservation and noise reduction; however, when directly applied to the G-10 and G-33 datasets, the denoising effects show significant divergence: noise suppression is insufficient in the G-10 dataset, with residual noise particularly noticeable in the bias maps; while the G-33 dataset exhibits a slight tendency towards over-smoothing. In contrast, UNet-All and RNet-All, trained on the entire dataset, show limited improvement in denoising performance on G-10 and G-33, failing to further improve the problems of residual noise or over-smoothing.

[0071] In contrast, the ARUN-G20 network, trained solely on the G-20 dataset and embedding noise level metrics, not only further reduced noise residue in the bias metrics but also outperformed traditional UNet and RNet in overall denoising performance. When this ARUN model was jointly trained on all low-count datasets (G-10, G-20, and G-33), its denoising performance on G-10 and G-33 was significantly enhanced.

[0072] Quantitative indicators

[0073] Table 2 NRMSE

[0074]

[0075]

[0076] Table 3 SSIM

[0077]

[0078] Table 4 PSNR

[0079]

[0080] The quantitative results shown in Tables 2-4 above reveal significant fluctuations in various evaluation metrics for low-dose nuclear medicine images: the NRMSE for the G-10 dataset is 0.363±0.194, while that for the G-33 dataset is 0.105±0.056. Simultaneously, SSIM and PSNR also show an upward trend with increasing dose. Compared to the original low-dose nuclear medicine images, all models significantly reduce NRMSE and improve SSIM and PSNR. Traditional models based on UNet and RNet show similar performance across various quantitative metrics, with only slight fluctuations in SSIM; furthermore, the performance of UNet-G20, UNet-All, RNet-20, and RNet-All varies across different datasets, making it difficult to determine a clear winner.

[0081] Among all methods, ARUN-All consistently performed best. Taking G-10 as an example, it had the lowest NRMSE (0.089±0.031), the highest SSIM (0.902±0.022), and the best PSNR (15.1±2.9), showing a significant advantage over the other five models (p<0.05). On the G-20 and G-33 datasets, ARUN-All achieved SSIMs of 0.913±0.020 and 0.919±0.017, respectively. Notably, ARUN-20 also performed excellently on the G-20 dataset, achieving an SSIM of 0.912±0.021, while maintaining competitive NRMSE and PSNR.

[0082] ROIs bias

[0083] Table 5 Lesion and background biases

[0084]

[0085] The results in Table 5 show that the lesion and background biases in low-dose nuclear medicine images decrease significantly with increasing dose level: in the G-10 dataset, the lesion bias was 2.4% ± 5.2% and the background bias was 38.9% ± 32.8%; while in the G-33 dataset, they decreased to 0.6% ± 4.5% and 35.3% ± 23.8%, respectively (p < 0.05), consistent with the trend that noise decreases with increasing count.

[0086] The UNet-20 and RNet-20 models trained solely on G-20 data exhibit comparable denoising capabilities for lesion and background deviations, maintaining similar performance even when applied to G-10 or G-33 (noise level mismatch). While the UNet-All and RNet-All models, trained jointly on all low-dose nuclear medicine images, show significant improvement in background deviation, lesion deviations show only slight changes.

[0087] In comparison, the ARUN series models proposed in this embodiment of the invention perform better:

[0088] ARUN-20 reduced lesion bias to -4.6% ± 5.5% and background bias to -15.7% ± 21.0% on the G-10 dataset, significantly outperforming UNet-20 and RNet-20.

[0089] ARUN-All further optimized the background bias of G-10 to -3.8% ± 22.9%, while maintaining a similar level of lesion and background bias as the aforementioned models on G-20 and G-33.

[0090] Signal-to-noise ratio

[0091] Table 6 SNR

[0092]

[0093]

[0094] Table 6 shows that the signal-to-noise ratio (SNR) of low-dose nuclear medicine images increases continuously with increasing dose level, approaching the average level of full-dose nuclear medicine images (2.89±1.06). Compared with the original low-dose nuclear medicine images, all models significantly improved the SNR, validating their good denoising effect. Among them, ARUN-All achieved the highest SNR (3.43±1.56) on the G-10 dataset, while the other five models performed similarly. For the G-20 and G-33 datasets, no significant difference was observed in SNR improvement among all models.

[0095] In summary, the flexibility, accuracy, and superiority of the ARUN model in handling image noise make it a more effective denoising tool, especially in clinical applications involving low-dose nuclear imaging, which helps improve the accuracy and reliability of image diagnosis.

[0096] A nuclear medicine image denoising apparatus provided in a second aspect embodiment of the present invention includes:

[0097] The first module is configured to acquire full-dose nuclear medicine projection data with different noise levels and perform image reconstruction to obtain full-dose nuclear medicine images with different noise levels. Low-dose nuclear medicine images are obtained by downsampling and image reconstruction of each full-dose nuclear medicine projection data. The low-dose nuclear medicine images are matched with the full-dose nuclear medicine images to construct a dataset.

[0098] The second module is configured to build a neural network model with noise level embeddings;

[0099] The third module is configured to use low-dose nuclear medicine images from the dataset as input and full-dose nuclear medicine images as labels to train the neural network model and obtain a denoising model.

[0100] The fourth module is configured to input the low-dose nuclear medicine image to be denoised into the denoising model and output the denoised nuclear medicine image.

[0101] It should be noted that the foregoing explanation of the embodiments of the nuclear medicine image denoising method also applies to the nuclear medicine image denoising device of this embodiment, and will not be repeated here.

[0102] To implement the above embodiments, this invention also proposes a computer-readable storage medium storing a computer program that is executed by a processor to perform the nuclear medicine image denoising method of the above embodiments.

[0103] The following is for reference. Figure 4The diagram illustrates a structural schematic of an electronic device suitable for implementing embodiments of the present invention. It should be noted that the electronic device in the embodiments of the present invention may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs, desktop computers, and servers. Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0104] like Figure 4 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 101, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 102 or a program loaded from a storage device 108 into a random access memory (RAM) 103. The RAM 103 also stores various programs and data required for the operation of the electronic device. The processing unit 101, ROM 102, and RAM 103 are interconnected via a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.

[0105] Typically, the following devices can be connected to I / O interface 105: input devices 106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, etc.; output devices 107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 108 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0106] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, this embodiment includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device 109, or installed from a storage device 108, or installed from a ROM 102. When the computer program is executed by the processing device 101, it performs the functions defined in the methods of the embodiments of the present invention.

[0107] It should be noted that the computer-readable medium described above in this invention can be a computer-readable signal medium, a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0108] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0109] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the aforementioned nuclear medicine image denoising method.

[0110] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, C++, and Python—as well as conventional procedural programming languages—such as the "C-" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0112] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0113] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0114] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0115] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] Those skilled in the art will understand that implementing all or part of the steps of the methods in the above embodiments can be accomplished by instructing related hardware through a program. The developed program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0117] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0118] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for denoising nuclear medicine images, characterized in that, include: Full-dose nuclear medicine projection data with different noise levels were acquired and image reconstruction was performed to obtain full-dose nuclear medicine images with different noise levels. Low-dose nuclear medicine images were obtained by downsampling and image reconstruction of each full-dose nuclear medicine projection data. The low-dose nuclear medicine images were matched with the full-dose nuclear medicine images to construct a dataset. A neural network model with noise level embedding is constructed. The neural network model with noise level embedding adopts an improved U-net network, which is a self-attention enhanced residual U-net network with noise level embedding. It includes an input layer, a first encoding layer, a second encoding layer, a third encoding layer, a first attention block, a bottleneck layer, a second attention block, a first decoding layer, a third attention block, a second decoding layer, a third decoding layer, and an output layer arranged sequentially. Skip connection layers are respectively provided between the first encoding layer and the third decoding layer, between the second encoding layer and the second decoding layer, and between the third encoding layer and the first decoding layer. Each encoding layer, each decoding layer, and the bottleneck layer adopt a corresponding residual block. Multiple noise level metrics are embedded into each residual block through a position encoding layer and two fully connected layers. The multiple noise level metrics include at least the downsampling ratio and also the mean intensity and standard deviation of the low-dose nuclear medicine image. Using low-dose nuclear medicine images from the dataset as input and full-dose nuclear medicine images as labels, the neural network model is trained to obtain a denoising model. Input the low-dose nuclear medicine image to be denoised into the denoising model, and output the denoised nuclear medicine image.

2. The nuclear medicine image denoising method according to claim 1, characterized in that, The nuclear medicine projection data uses SPECT projection data. Full-dose SPECT projection data with different noise levels are obtained by referring to the clinical nuclear medicine image acquisition procedure. The images are then reconstructed, and combined with attenuation correction based on CT images, n iterations and m subset reconstructions are performed to obtain full-dose SPECT images with different noise levels.

3. The nuclear medicine image denoising method according to claim 1, characterized in that, Let the downsampling ratio be α，α Take any value from (0,1).

4. The nuclear medicine image denoising method according to claim 1, characterized in that, Each residual block includes a connected backbone and a bypass. The backbone contains two 3×3×3 convolutions with ReLU activation, and the bypass contains a 1×1×1 pointwise convolution with ReLU activation. The output of the three convolutions is added to the input of the residual block channel by channel.

5. The nuclear medicine image denoising method according to claim 1, characterized in that, Each attention block consists of a flattened layer, a multi-head attention mechanism, and an inverse flattened layer connected in sequence.

6. The nuclear medicine image denoising method according to claim 1, characterized in that, Before training the neural network model, a patch-based strategy is used to augment the low-dose nuclear medicine images in the dataset. Then, leave-one-out cross-validation is used to train and test the constructed neural network model with the augmented low-dose nuclear medicine images as input and the full-dose nuclear medicine images as labels, to obtain the denoising model.

7. The nuclear medicine image denoising method according to claim 6, characterized in that, The leave-one-out cross-validation method includes: dividing the dataset into five equal parts, with 80% of the dataset used as the training dataset and the remaining 20% ​​of the dataset used as the test dataset, and conducting five training and validation experiments to generate five training models. The training dataset and the test dataset are from different scanned objects.

8. A nuclear medicine image denoising device, characterized in that, include: The first module is configured to acquire full-dose nuclear medicine projection data with different noise levels and perform image reconstruction to obtain full-dose nuclear medicine images with different noise levels. Low-dose nuclear medicine images are obtained by downsampling and image reconstruction of each full-dose nuclear medicine projection data. The low-dose nuclear medicine images are matched with the full-dose nuclear medicine images to construct a dataset. The second module is configured to construct a neural network model with noise level embedding. The neural network model with noise level embedding adopts an improved U-net network, which is a self-attention enhanced residual U-net network with noise level embedding. It includes an input layer, a first encoding layer, a second encoding layer, a third encoding layer, a first attention block, a bottleneck layer, a second attention block, a first decoding layer, a third attention block, a second decoding layer, a third decoding layer, and an output layer arranged sequentially. Skip connection layers are respectively provided between the first encoding layer and the third decoding layer, between the second encoding layer and the second decoding layer, and between the third encoding layer and the first decoding layer. Each encoding layer, each decoding layer, and the bottleneck layer adopt a corresponding residual block. Multiple noise level metrics are embedded into each residual block through a position encoding layer and two fully connected layers. The multiple noise level metrics include at least the downsampling ratio and also the average intensity and standard deviation of the low-dose nuclear medicine image. The third module is configured to use low-dose nuclear medicine images from the dataset as input and full-dose nuclear medicine images as labels to train the neural network model and obtain a denoising model. The fourth module is configured to input the low-dose nuclear medicine image to be denoised into the denoising model and output the denoised nuclear medicine image.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the nuclear medicine image denoising method according to any one of claims 1 to 7.

Images