A low-dose positron emission tomography image high-fidelity reconstruction method

Abstract

The present application relates to the technical field of image processing, and particularly relates to a low-dose positron emission tomography image high-fidelity reconstruction method. The method comprises the following steps: acquiring paired data of the same imaging sample, including first and second scanning image data; constructing a cyclic constraint adversarial denoising network containing a generator, a first discriminator and a second discriminator. The first scanning image is input into the generator to obtain a denoised image; the denoised image and the second scanning image are input into the first discriminator to construct an adversarial loss; adaptive noise re-injection is performed on the denoised image to obtain a noise image, which is input into the second discriminator together with the first scanning image to construct a cyclic consistency loss; the network is trained based on the two kinds of losses, and the parameters are optimized to obtain a trained generator for new image denoising. The present application simultaneously suppresses noise, retains lesion details and improves quantitative accuracy in low-dose PET image reconstruction through adaptive noise re-injection and cyclic consistency constraint, and realizes high-fidelity reconstruction effect.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for high-fidelity reconstruction of low-dose positron emission tomography (PET) images. Background Technology

[0002] Positron emission tomography (PET) is an important functional medical imaging technique, but its imaging process involves the injection of radioactive tracers, and high radiation doses can pose potential health risks to patients. Therefore, low-dose PET imaging has become a hot topic in clinical research. However, low-dose PET images contain strong Poisson noise due to insufficient photon counts, which seriously affects image quality and the accuracy of quantitative analysis.

[0003] Traditional denoising methods, such as Gaussian filtering and anisotropic diffusion, struggle to preserve lesion edges and structural details while suppressing noise. In recent years, deep learning-based image reconstruction methods, particularly Generative Adversarial Networks (GANs), have been attempted for PET image denoising. However, existing methods typically employ fixed-distribution noise modeling or simple cyclic consistency constraints, neglecting the spatial variability of noise in PET images and the differences in noise sensitivity between lesion and background regions. This leads to problems such as over-smoothing, artifacts, or loss of lesion details in the reconstructed images, making it difficult to achieve high-fidelity reconstruction of low-dose PET images.

[0004] Chinese Patent Publication No. CN113674377B discloses a system and method for reconstructing positron emission tomography (PET) images. The method includes acquiring PET data of an object. The PET data may correspond to at least two voxels in the reconstructed image domain. The method includes acquiring motion signals of the object. The method includes acquiring motion amplitude data. The motion amplitude data may represent the range of motion of each of the at least two voxels. The method includes determining gating data based at least partially on the motion amplitude data. The gating data may include useful percentage counts, each useful percentage count corresponding to at least one voxel of the at least two voxels. The method includes gating the PET data based on the gating data and the motion signals. The method includes reconstructing a PET image of the object based on the gated PET data.

[0005] Therefore, the existing technology has the following problems: The problem of strong noise caused by insufficient photon count in low-dose PET images has not been solved, nor has an adaptive constraint mechanism for denoising networks and lesion perception in the image domain been introduced. Summary of the Invention

[0006] To address this, the present invention provides a high-fidelity reconstruction method for low-dose positron emission tomography (PET) images, which overcomes the problems in the prior art where the strong noise caused by insufficient photon count in low-dose PET images is not resolved, and where no denoising network for the image domain and adaptive constraint mechanism for lesion perception are introduced.

[0007] To achieve the above objectives, the present invention provides a method for high-fidelity reconstruction of low-dose positron emission tomography (PET) images, comprising: Step S1: Obtain paired data for the same imaging sample, wherein the paired data includes first scan image data and second scan image data; Step S2: Construct a cyclic constraint adversarial denoising network, which includes a generator, a first discriminator, and a second discriminator. Step S3: Input the first scanned image data into the generator and perform denoising processing to obtain denoised scanned image data; Step S4: Input the denoised scanned image data and the second scanned image data into the first discriminator, distinguish the denoised scanned image data and the second scanned image data through the first discriminator, and construct an adversarial loss based on the distinction result; Step S5: Perform adaptive noise re-injection on the denoised scan image data to obtain noisy scan image data. Input the first scan image data and the noisy scan image data into the second discriminator. The second discriminator distinguishes the first scan image data from the noisy scan image data and constructs a cycle consistency loss based on the distinction result. Step S6: Construct a composite loss function that includes the adversarial loss and the cycle consistency loss; Step S7: Train the cyclic constraint adversarial denoising network using the paired data and the composite loss function to optimize the network parameters of the generator, the first discriminator and the second discriminator, and obtain the trained generator; Step S8: Use the trained generator to denoise the new first scan image data and output the reconstructed scan image.

[0008] Further, in step S3, the generator includes an encoder, a bottleneck layer, and a decoder connected in sequence to obtain denoised scan image data, including: Step S31: Input the first scanned image data into the encoder, and after several convolutional layers and downsampling operations, output the first feature map to the bottleneck layer; Step S32: Acquire several scan image data that are spatially adjacent to the first scan image data, extract the structural feature map of each adjacent scan image data, and fuse the structural feature map with the first feature map in the bottleneck layer to obtain a fused feature map; Step S33: Input the fused feature map into the decoder and output the denoised scan image data.

[0009] Furthermore, the fused feature map is obtained, including: Calculate the similarity matrix between the first feature map and the structural feature map; Attention weights are generated based on the similarity matrix; The attention weights are weighted together with the structural feature map, and then concatenated with the first feature map at the channel level to obtain the fused feature map.

[0010] Further, in step S4, the adversarial loss is constructed, including: The denoised scanned image data is used as the first sample, and the second scanned image data is used as the second sample, and input into the first discriminator; The first discriminator outputs the probability value that the denoised scanned image data is a real image; The binary cross-entropy loss is calculated based on the probability value and used as the adversarial loss.

[0011] Further, in step S5, adaptive noise re-injection is performed on the denoised scan image data to obtain noisy scan image data, including: Step S51: Detect lesion-sensitive regions in the denoised scan image data and generate a spatial attention mask. The spatial attention mask is used to identify lesion candidate regions and non-lesion regions in the denoised scan image data. Step S52: According to the spatial attention mask, apply a first noise intensity to the lesion candidate region in the denoised scan image data and apply a second noise intensity to the non-lesion region, wherein the first noise intensity is less than the second noise intensity; Step S53: The image after applying noise is used as the noise scan image data.

[0012] Furthermore, a spatial attention mask is generated, including: Extract the gradient map of the denoised scanned image data or the first scanned image data; The gradient map is segmented by a threshold to obtain high gradient regions; The high gradient region is subjected to morphological dilation to obtain a binary mask, which is then used as the spatial attention mask. The high gradient region corresponds to the lesion candidate region, and the remaining regions correspond to the non-lesion regions.

[0013] Further, in step S5, the cycle consistency loss is constructed, including: Noise is added to the denoised scan image data to obtain noisy scan image data; The first scanned image data is used as the first sample, and the noisy scanned image data is used as the second sample, which is then input into the second discriminator. The second discriminator outputs the probability value that the noisy scan image data is a real image; The binary cross-entropy loss is calculated based on the probability values, and the cycle consistency loss is obtained.

[0014] Further, in step S6, a composite loss function is constructed, including: Obtain the adversarial loss, the cycle consistency loss, and the additional loss; The adversarial loss, the cycle consistency loss, and the additional loss are weighted and summed to obtain the composite loss function.

[0015] Furthermore, additional losses include: The supervised loss is obtained by calculating the sum of squared differences between corresponding pixels between the denoised scanned image data and the second scanned image data; The second scanned image data is input into the generator to obtain a reconstructed image. The pixel-level difference between the reconstructed image and the second scanned image data is calculated to obtain the identity mapping loss. Calculate the structural similarity index loss between the denoised scanned image data and several adjacent scanned image data to obtain the neighborhood structural similarity loss; The supervision loss, the identity mapping loss, and the neighborhood structure similarity loss are used as additional losses.

[0016] Further, in step S7, the trained generator is obtained, including: The paired data is input into the cyclic constraint adversarial denoising network in batches. The backpropagation algorithm and gradient descent optimizer are used to optimize the network parameters of the generator, the first discriminator and the second discriminator alternately with the goal of minimizing the composite loss function. In each iteration, the first discriminator and the second discriminator are optimized first, and then the generator is optimized. Repeat the above iterative process until the value of the composite loss function no longer decreases or reaches the preset maximum number of iterations. Use the optimized generator at this point as the trained generator.

[0017] Compared with the prior art, the beneficial effects of the present invention are that by introducing the structural feature maps of adjacent multi-slice slices and the first feature map for attention fusion, the generator can make full use of the inter-layer anatomical structure information in the three-dimensional PET image. At the bottleneck layer, attention weights are generated by calculating the similarity matrix, and the anatomical structure features of adjacent slices are adaptively weighted and fused into the current slice feature map. This effectively enhances the ability to preserve lesion edges, tissue interfaces and minute details during the denoising process, avoiding the edge blurring and structure loss problems caused by the lack of spatial context information in traditional single-slice denoising methods, and significantly improving the reconstruction quality of low-dose positron emission tomography images.

[0018] Furthermore, this invention constructs an adversarial loss through a first discriminator to perform pixel-by-pixel discrimination of local regions between the denoised scan image data and the second scan image data. This forces the generator to learn the mapping relationship from low-dose images to high-quality images at multiple spatial scales, enabling the generator to not only focus on the similarity of the overall image distribution but also pay more attention to the realism of local textures, edges, and lesion details. This effectively avoids the problem of excessive image smoothing caused by traditional global discriminators and significantly improves the visual quality and diagnostic usability of the reconstructed images.

[0019] Furthermore, this invention generates a spatial attention mask that can effectively identify lesion candidate regions by performing adaptive threshold segmentation and morphological dilation on the gradient map. This method does not require additional lesion annotation data and is automatically completed based entirely on the gradient information of the image itself. It has the advantages of high computational efficiency and strong generalization ability, and is suitable for image reconstruction tasks of different parts and different lesion types.

[0020] Furthermore, by constructing a cycle consistency loss, this invention forces the denoising mapping learned by the generator to be reversible. That is, the denoised image can be restored to the noise statistical features of the original low-dose image after noise is added again. This effectively avoids the pattern collapse and loss of output diversity problems common in generative adversarial network training, and ensures the stability and reliability of the reconstructed image under different dose conditions.

[0021] Furthermore, this invention constructs a composite loss function that includes adversarial loss, cycle consistency loss, supervision loss, identity mapping loss, and neighborhood structure similarity loss. This constrains the generator's training from multiple dimensions, enabling the generator to learn the distribution characteristics of real images in adversarial games, maintain the invertibility of denoising mapping in cycle consistency constraints, approximate high-quality reference images in supervised learning, maintain structural invariance in identity mapping, and maintain inter-layer continuity in neighborhood structure similarity. This comprehensively improves the fidelity of low-dose positron emission tomography (PET) image reconstruction.

[0022] Furthermore, by employing an alternating training strategy of first optimizing the discriminator and then optimizing the generator, this invention ensures that the generator and discriminator maintain a dynamic balance during the adversarial game, avoiding the problem of training collapse caused by one side becoming too strong. At the same time, by setting reasonable convergence conditions and iteration counts, the efficiency and stability of network training are guaranteed, ultimately obtaining a generator model with excellent denoising performance and good generalization ability. Attached Figure Description

[0023] Figure 1 This is a flowchart of the high-fidelity reconstruction method for low-dose positron emission tomography images in this embodiment; Figure 2 This is a flowchart illustrating the process of obtaining denoised scan image data in the high-fidelity reconstruction method of low-dose positron emission tomography images in this embodiment; Figure 3 This is a flowchart illustrating the noise application process in the high-fidelity reconstruction method for low-dose positron emission tomography images in this embodiment. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] Please see Figure 1 As shown, Figure 1 This is a flowchart of the high-fidelity reconstruction method for low-dose positron emission tomography images in this embodiment.

[0027] This embodiment provides a method for high-fidelity reconstruction of low-dose positron emission tomography (PET) images, including: Step S1: Obtain paired data for the same imaging sample, wherein the paired data includes first scan image data and second scan image data; In this embodiment, the first scanned image data is low-dose positron emission tomography (PET) image data, specifically referring to PET image data acquired using a radioactive tracer at a dose lower than that used in conventional clinical settings. The conventional clinical dose is calculated according to a standardized injection dose formula, specifically, the activity of the injected radioactive tracer equals body weight multiplied by a dose coefficient, where the dose coefficient ranges from 0.15 to 0.25 mCi / kg. The low dose is 10% to 30% of the conventional clinical dose, i.e., the dose coefficient ranges from 0.015 to 0.075 mCi / kg.

[0028] The determination of the dose coefficient range is based on a balance between the physical and statistical properties of positron emission tomography (PET) imaging and clinical image quality requirements. At conventional clinical doses, the coincidence count received by the detector is sufficiently high, ensuring that the signal-to-noise ratio of the reconstructed image meets the basic clinical diagnostic requirement for lesion identifiability. If the dose coefficient is below 0.15 mCi per kilogram of body weight, the variance of Poisson noise in the image will significantly increase, causing small lesions or low-contrast areas to be overwhelmed by noise; conversely, if the dose coefficient is above 0.25 mCi per kilogram of body weight, the radiation dose received by the patient will exceed the clinically normal exposure level, violating the principle of optimal radiation protection.

[0029] The low-dose setting is 10%–30% of the conventional dose. Within this range, patient radiation exposure can be reduced by 70%–90%, significantly improving examination safety, especially suitable for pediatric patients, patients requiring multiple follow-up examinations, and healthy individuals undergoing checkups. Furthermore, this dose level still ensures that the detector collects sufficient coincidence event counts, making it feasible to recover key diagnostic information through subsequent signal processing and image reconstruction algorithms. If the dose coefficient is below 0.015 mCi per kilogram of body weight, the noise in the image will exhibit strong non-stationary characteristics, and the signal will be completely submerged in the noise background, making it difficult to recover clinically valuable image features even using deep learning methods. If the dose coefficient is above 0.075 mCi per kilogram of body weight, the radiation dose reduction is less than 70%, deviating from the original intention of low-dose imaging.

[0030] The PET image data refers to positron emission tomography (PET) image data, which has a low signal-to-noise ratio and significant Poisson noise characteristics. The second scan image data is high-quality PET image data from the same imaging sample and the same anatomical site as the first scan image data. Specifically, it refers to full-count PET image data with high signal-to-noise ratio and clear anatomical details obtained using conventional clinical doses or extended acquisition times. The paired data includes, but is not limited to, the following specific data forms: image pairs consisting of low-dose PET images and conventional-dose PET images acquired from the same patient in the same bed; synthetic image pairs consisting of low-dose images generated by Poisson downsampling simulation of conventional-dose PET images and the original conventional-dose images; two PET scan images acquired from the same patient at different times, one of which is a low-dose protocol scan and the other is a conventional-dose protocol scan; and registered low-dose and conventional-dose PET image pairs from publicly available PET datasets. The paired data needs to undergo image registration preprocessing after acquisition to ensure that the first scan image data and the second scan image data correspond one-to-one in spatial location.

[0031] Step S2: Construct a cyclic constraint adversarial denoising network, which includes a generator, a first discriminator, and a second discriminator. In this embodiment, the generator employs a U-shaped network structure with skip connections, comprising an encoder, a bottleneck layer, and a decoder connected sequentially. The encoder consists of four downsampling blocks, each containing a convolutional layer, a batch normalization layer, and a Leaky ReLU activation function, used to progressively extract multi-scale feature maps of the image. Each downsampling halves the feature map spatial size and doubles the number of channels. The bottleneck layer contains two residual blocks, each consisting of two convolutional layers and skip connections, used to encode the semantic information of the image in the highest-order feature space. The decoder consists of four upsampling blocks, each containing a transposed convolutional layer, a feature concatenation operation, a convolutional layer, and a ReLU activation function. The feature concatenation operation concatenates the feature map of the current layer of the decoder with the feature map of the corresponding layer of the encoder in the channel dimension to recover the spatial detail information of the image. Finally, the decoder outputs a feature map with one channel, which serves as the denoised scanned image data.

[0032] Both the first and second discriminators employ a Patch GAN structure, which is a fully convolutional network consisting of five convolutional layers. Each convolutional layer is followed by a batch normalization layer and a Leaky ReLU activation function. The first four convolutional layers use a 2-stride convolution operation for downsampling, and the last convolutional layer outputs an N-by-N discriminant matrix, where N can be either 16 or 32. Each discriminant unit corresponds to a receptive field region in the input image, used to distinguish local structural differences between the real image and the generated image. The first discriminator distinguishes between denoised scanned image data and second scanned image data, while the second discriminator distinguishes between first scanned image data and noisy scanned image data. The initial network parameters of the cyclic constraint adversarial denoising network are initialized using the Xavier method, and the generator, first discriminator, and second discriminator are alternately optimized during training.

[0033] Step S3: Input the first scanned image data into the generator and perform denoising processing to obtain denoised scanned image data; Step S4: Input the denoised scanned image data and the second scanned image data into the first discriminator, distinguish the denoised scanned image data and the second scanned image data through the first discriminator, and construct an adversarial loss based on the distinction result; Step S5: Perform adaptive noise re-injection on the denoised scan image data to obtain noisy scan image data. Input the first scan image data and the noisy scan image data into the second discriminator. The second discriminator distinguishes the first scan image data from the noisy scan image data and constructs a cycle consistency loss based on the distinction result. Step S6: Construct a composite loss function that includes the adversarial loss and the cycle consistency loss; Step S7: Train the cyclic constraint adversarial denoising network using the paired data and the composite loss function to optimize the network parameters of the generator, the first discriminator and the second discriminator, and obtain the trained generator; Step S8: Use the trained generator to denoise the new first scan image data and output the reconstructed scan image.

[0034] Please see Figure 2 As shown, Figure 2 This is a flowchart illustrating the process of obtaining denoised scan image data in the high-fidelity reconstruction method of low-dose positron emission tomography images in this embodiment.

[0035] Specifically, in step S3, the generator includes an encoder, a bottleneck layer, and a decoder connected in sequence to obtain denoised scan image data, including: Step S31: Input the first scanned image data into the encoder, and after several convolutional layers and downsampling operations, output the first feature map to the bottleneck layer; Step S32: Acquire several scan image data that are spatially adjacent to the first scan image data, extract the structural feature map of each adjacent scan image data, and fuse the structural feature map with the first feature map in the bottleneck layer to obtain a fused feature map; Step S33: Input the fused feature map into the decoder and output the denoised scan image data.

[0036] Specifically, the fused feature map is obtained, including: Calculate the similarity matrix between the first feature map and the structural feature map; Attention weights are generated based on the similarity matrix; The attention weights are weighted together with the structural feature map, and then concatenated with the first feature map at the channel level to obtain the fused feature map.

[0037] In this embodiment, the first scanned image data is input into the encoder, and after passing through several convolutional layers and downsampling operations, a first feature map is output to the bottleneck layer. The encoder consists of four downsampling blocks, each containing a convolutional layer with a stride of 2, a batch normalization layer, and a LeakyReLU activation function. After the first scanned image data passes through the four downsampling blocks sequentially, the spatial resolution is gradually reduced to 1 / 16 of the original size, and the number of channels gradually increases to 512, outputting the first feature map.

[0038] A plurality of scanned image data spatially adjacent to the first scanned image data are acquired, and structural feature maps of each adjacent scanned image data are extracted. These structural feature maps are then fused with the first feature map in the bottleneck layer to obtain a fused feature map. Specifically, the plurality of adjacent scanned image data refers to the image data of one to two adjacent slices spatially adjacent to the current slice in a three-dimensional PET image sequence of the same imaging sample. A shared-weight feature extraction network performs convolution operations on each adjacent scanned image data to extract their respective structural feature maps, which contain anatomical structural information from the adjacent slices.

[0039] Calculate the similarity matrix between the first feature map and the structural feature map; generate attention weights based on the similarity matrix; weight the attention weights with the structural feature map, and then concatenate them with the first feature map at the channel level to obtain the fused feature map. The similarity matrix is ​​obtained by calculating the cosine similarity or dot product between the feature vector at each position in the first feature map and the corresponding feature vector at the structural feature map. The attention weights are obtained by normalizing the similarity matrix using the Softmax function. After concatenating the weighted structural feature map with the first feature map at the channel dimension, it is then passed through a 3×3 convolutional layer for channel fusion, outputting the fused feature map.

[0040] The fused feature map is input into the decoder, and the denoised scanned image data is output. The decoder consists of four upsampling blocks, each containing a transposed convolutional layer, a feature concatenation operation, a convolutional layer, and a ReLU activation function. The fused feature map passes through the four upsampling blocks sequentially. Each upsampling block doubles the feature map's spatial size and halves the number of channels. The feature concatenation operation then concatenates the current layer's feature map with the corresponding layer's feature map from the encoder to recover image details. Finally, the image passes through a 1×1 output convolutional layer, mapping the feature map to a single-channel image, and the denoised scanned image data is output.

[0041] This invention introduces the structural feature maps of adjacent multi-slice slices and performs attention fusion with the first feature map, enabling the generator to fully utilize the inter-slice anatomical structure information in the 3D PET image. At the bottleneck layer, attention weights are generated by calculating a similarity matrix, and the anatomical structure features of adjacent slices are adaptively weighted and fused into the current slice feature map. This effectively enhances the ability to preserve lesion edges, tissue interfaces, and minute details during the denoising process, avoiding the edge blurring and structure loss problems caused by the lack of spatial context information in traditional single-slice denoising methods, and significantly improving the reconstruction quality of low-dose positron emission tomography images.

[0042] Specifically, in step S4, the adversarial loss is constructed, including: The denoised scanned image data is used as the first sample, and the second scanned image data is used as the second sample, and input into the first discriminator; The first discriminator outputs the probability value that the denoised scanned image data is a real image; The adversarial loss is obtained by calculating the binary cross-entropy loss based on the probability value.

[0043] In this embodiment, the denoised scanned image data is used as the first sample, and the second scanned image data is used as the second sample, both input to the first discriminator. The first discriminator outputs the probability value that the denoised scanned image data is a real image. Based on the probability value, a binary classification cross-entropy loss is calculated to obtain the adversarial loss. More specifically, the first discriminator adopts a Patch GAN structure and outputs an N×N discrimination matrix, where the value of each discrimination unit represents the probability that the corresponding local region in the input image is a real image. The second scanned image data is used as a positive sample, and its corresponding discrimination target is a matrix of all 1s; the denoised scanned image data is used as a negative sample, and its corresponding discrimination target is a matrix of all 0s. The adversarial loss is calculated as follows: the adversarial loss is equal to the negative binary classification cross-entropy loss. The positive sample discrimination output is the output probability value of the first discriminator for the second scanned image data, and the negative sample discrimination output is the output probability value of the first discriminator for the denoised scanned image data. This adversarial loss prompts the denoised scanned image data generated by the generator to approximate the real conventional dose image as closely as possible in terms of local structural distribution, while simultaneously prompting the first discriminator to continuously improve its ability to distinguish between real and fake images.

[0044] In this embodiment, the optimization goal of the first discriminator is to minimize its own discrimination error between positive and negative samples, that is, to distinguish the denoised scan image data and the second scan image data as correctly as possible; the optimization goal of the generator is to maximize the probability that the first discriminator will identify the denoised scan image data as a real image; through this adversarial game mechanism, the generator gradually learns the mapping relationship from low-dose images to high-quality images.

[0045] This invention constructs an adversarial loss through a first discriminator, performing pixel-by-pixel discrimination of local regions between denoised scanned image data and second scanned image data. This forces the generator to learn the mapping relationship from low-dose images to high-quality images at multiple spatial scales, enabling the generator to not only focus on the similarity of the overall image distribution, but also pay more attention to the realism of local textures, edges, and lesion details. This effectively avoids the problem of excessive image smoothing caused by traditional global discriminators, and significantly improves the visual quality and diagnostic usability of the reconstructed images.

[0046] Please see Figure 3 As shown, Figure 3This is a flowchart illustrating the noise application process in the high-fidelity reconstruction method for low-dose positron emission tomography images in this embodiment.

[0047] Specifically, in step S5, adaptive noise re-injection is performed on the denoised scan image data to obtain noisy scan image data, including: Step S51: Detect lesion-sensitive regions in the denoised scan image data and generate a spatial attention mask. The spatial attention mask is used to identify lesion candidate regions and non-lesion regions in the denoised scan image data. Step S52: According to the spatial attention mask, apply a first noise intensity to the lesion candidate region in the denoised scan image data and apply a second noise intensity to the non-lesion region, wherein the first noise intensity is less than the second noise intensity; Step S53: The image after applying noise is used as the noise scan image data.

[0048] In this embodiment, lesion-sensitive region detection is performed on the denoised scanned image data to generate a spatial attention mask. This spatial attention mask is used to identify lesion candidate regions and non-lesion regions in the denoised scanned image data. Specifically, the gradient map of the denoised scanned image data is extracted, and threshold segmentation is performed on the gradient map to obtain high-gradient regions. Then, morphological dilation is performed on these high-gradient regions to obtain a binary mask, which serves as the spatial attention mask. The high-gradient regions correspond to lesion candidate regions, and the remaining regions correspond to non-lesion regions. Regions with high gradient amplitudes in the gradient map typically correspond to edges, tissue interfaces, and potential lesion boundaries in the image; these regions are identified as sensitive regions requiring protection.

[0049] Based on the spatial attention mask, a first noise intensity is applied to lesion candidate regions in the denoised scan image data, and a second noise intensity is applied to non-lesion regions, wherein the first noise intensity is less than the second noise intensity. Specifically, the first noise intensity can be set to zero, that is, no noise is added to the lesion candidate regions to preserve the original structure and detail information of the region to the greatest extent; the second noise intensity is a positive number, that is, Poisson noise or Gaussian noise that conforms to the physical characteristics of PET imaging is added to the non-lesion regions. Alternatively, the first noise intensity and the second noise intensity can be set as spatially variable noise, the variance of which is negatively correlated with the pixel value of the spatial attention mask, that is, the noise variance corresponding to the lesion candidate regions is less than the noise variance corresponding to the non-lesion regions, so that the noise intensity transitions smoothly in space and avoids abrupt artifacts at the lesion boundaries.

[0050] The image after noise is applied is used as the noisy scan image data. This noisy scan image data is then input together with the first scan image data into a second discriminator, which distinguishes between the two. The second discriminator has the same structure as the first discriminator, employing a Patch GAN architecture. The cycle consistency loss is constructed based on the discrimination result of the second discriminator, and its calculation formula is similar to that of the adversarial loss, i.e., negative. This cycle consistency loss forces the denoised scan image data generated by the generator to be difficult for the second discriminator to distinguish from the original low-dose image after adaptive noise re-injection, thus forming a closed-loop constraint and ensuring that the denoising process does not change the underlying structure and statistical properties of the image.

[0051] Specifically, generating a spatial attention mask includes: Extract the gradient map of the denoised scanned image data or the first scanned image data; The gradient map is segmented by a threshold to obtain high gradient regions; The high gradient region is subjected to morphological dilation to obtain a binary mask, which is then used as the spatial attention mask. The high gradient region corresponds to the lesion candidate region, and the remaining regions correspond to the non-lesion regions.

[0052] In this embodiment, the gradient map of the denoised scan image data or the first scan image data is extracted; the gradient map is segmented by thresholding to obtain high gradient regions; the high gradient regions are subjected to morphological dilation to obtain a binary mask, and the binary mask is used as the spatial attention mask; wherein, the high gradient regions correspond to the lesion candidate regions, and the remaining regions correspond to the non-lesion regions.

[0053] More specifically, the gradient map is obtained by calculating the gradient magnitudes of the image in the horizontal and vertical directions. The gradient can be extracted using the Sobel or Scharr operators. A larger gradient magnitude indicates a more drastic change in pixel value at that location, typically corresponding to edges, tissue interfaces, and potential lesion boundaries in the image. Since lesion areas often exhibit different metabolic activity than surrounding normal tissue, manifesting as abrupt changes in local grayscale values ​​in PET images, high gradient regions can serve as reasonable indicators of lesion candidate regions. The threshold segmentation employs an adaptive thresholding method, dynamically determining the segmentation threshold based on the global statistical characteristics of the gradient map. Specifically, the mean and standard deviation of the gradient magnitudes of all pixels in the gradient map are calculated, and the segmentation threshold is set as a multiple of the mean plus the standard deviation, with the multiple ranging from 1 to 3. Pixels with gradient magnitudes higher than this threshold are marked as high gradient regions, and pixels with gradient magnitudes lower than this threshold are marked as low gradient regions.

[0054] The morphological dilation process uses a structuring element to expand the high-gradient region. The structuring element can be circular or square, with a radius ranging from 2 to 5 pixels. The purpose of the dilation operation is to expand the high-gradient region outwards by a certain range, ensuring that the generated mask completely covers the lesion region and its adjacent edge transition zone. This avoids artificial step artifacts at the lesion edge due to excessively steep mask boundaries during subsequent noise reinjection. In the binary mask obtained after the above processing, pixel values ​​within the lesion candidate region are set to 1, and pixel values ​​within the non-lesion region are set to 0. This binary mask is the spatial attention mask, used to guide the subsequent adaptive noise reinjection step. Using this mask, a smaller first noise intensity can be applied to the lesion candidate region with a value of 1, and a larger second noise intensity can be applied to the non-lesion region with a value of 0, thereby achieving differentiated protection of the lesion region.

[0055] This invention generates a spatial attention mask that can effectively identify candidate lesion regions by performing adaptive threshold segmentation and morphological dilation on the gradient map. This method does not require additional lesion annotation data and is completed automatically based entirely on the gradient information of the image itself. It has the advantages of high computational efficiency and strong generalization ability, and is suitable for image reconstruction tasks of different parts and different lesion types.

[0056] Specifically, in step S5, the cycle consistency loss is constructed, including: Noise is added to the denoised scan image data to obtain noisy scan image data; The first scanned image data is used as the first sample, and the noisy scanned image data is used as the second sample, which is then input into the second discriminator. The second discriminator outputs the probability value that the noisy scan image data is a real image; The binary cross-entropy loss is calculated based on the probability values, and the cycle consistency loss is obtained.

[0057] In this embodiment, noise is added to the denoised scan image data to obtain noisy scan image data; the first scan image data is used as the first sample, and the noisy scan image data is used as the second sample, and input into the second discriminator; the second discriminator outputs the probability value that the noisy scan image data is a real image; based on the probability value, the binary classification cross-entropy loss is calculated to obtain the cycle consistency loss.

[0058] More specifically, the second discriminator has the same structure as the first discriminator, both employing a Patch GAN structure and outputting an N×N discrimination matrix. The value of each discrimination unit represents the probability that the corresponding local region in the input image is a real image. The first scanned image data is used as a positive sample, and its corresponding discrimination target is a matrix of all 1s; the noisy scanned image data obtained by adding noise to the denoised scanned image data is used as a negative sample, and its corresponding discrimination target is a matrix of all 0s. The formula for calculating the cycle consistency loss is: the cycle consistency loss is equal to the negative binary cross-entropy loss, where the positive sample discrimination output is the output probability value of the second discriminator for the first scanned image data, and the negative sample discrimination output is the output probability value of the second discriminator for the noisy scanned image data. This cycle consistency loss ensures that the denoised scanned image data generated by the generator, after being noise-added, is difficult for the second discriminator to distinguish from the original low-dose image, thus forming a closed-loop constraint.

[0059] In this embodiment, the added noise can be of different types depending on actual needs. One implementation is to use Gaussian noise with a fixed distribution, a mean of zero, and a variance preset according to the noise level of the first scanned image data. Another implementation is to use Poisson noise that conforms to the physical characteristics of PET imaging, with the intensity parameter of the Poisson noise dynamically set according to the average count level of the first scanned image data. Yet another implementation is to use the adaptive noise re-injection method described in the preceding claims, that is, to apply noise of differentiated intensity to lesion candidate regions and non-lesion regions based on the spatial attention mask generated by lesion sensitive region detection. The cyclic consistency loss and the aforementioned adversarial loss together constitute the core part of the composite loss function. Among them, the adversarial loss is responsible for driving the generator to learn the mapping from low-dose images to high-quality images, while the cyclic consistency loss is responsible for constraining the stability of this mapping process and preventing the generator from producing arbitrary outputs unrelated to the original input. The joint optimization of the two enables the generator to produce high-quality visual effects while maintaining consistency with the original low-dose image in terms of structure and statistical characteristics.

[0060] This invention constructs a cycle consistency loss, which forces the denoising mapping learned by the generator to be reversible. That is, the denoised image can be restored to the noise statistical features of the original low-dose image after noise is added again. This effectively avoids the pattern collapse and loss of output diversity problems common in generative adversarial network training, and ensures the stability and reliability of the reconstructed image under different dose conditions.

[0061] Specifically, in step S6, the composite loss function is constructed, including: Obtain the adversarial loss, the cycle consistency loss, and the additional loss; The adversarial loss, the cycle consistency loss, and the additional loss are weighted and summed to obtain the composite loss function.

[0062] Specifically, additional losses include: The supervised loss is obtained by calculating the sum of squared differences between corresponding pixels between the denoised scanned image data and the second scanned image data; The second scanned image data is input into the generator to obtain a reconstructed image. The pixel-level difference between the reconstructed image and the second scanned image data is calculated to obtain the identity mapping loss. Calculate the structural similarity index loss between the denoised scanned image data and several adjacent scanned image data to obtain the neighborhood structural similarity loss; The supervision loss, the identity mapping loss, and the neighborhood structure similarity loss are used as additional losses.

[0063] In this embodiment, the supervised loss uses either L1 loss or L2 loss, calculating the sum or sum of squares of the absolute differences between each corresponding pixel in the denoised scanned image data and the second scanned image data. This supervised loss directly constrains the generator's output to approximate a realistic, high-quality image, and is a fundamental loss term ensuring the overall fidelity of the reconstructed image. In this embodiment, L1 loss is preferred because it is less sensitive to outliers than L2 loss and can better preserve image edge information.

[0064] The identity mapping loss is calculated as follows: The second scanned image data (high-quality image) is input into the generator. The generator performs forward propagation on this input and outputs a reconstructed image. Since the input is already a high-quality image, an ideal generator should have the ability to perform identity mapping, meaning the output should be as consistent as possible with the input. The identity mapping loss calculates the pixel-level difference between the reconstructed image and the original second scanned image data, typically using L1 or L2 loss. This loss term constrains the generator's behavior, preventing unnecessary modifications or distortions to the originally clear structure during denoising, and also helps maintain tonal and contrast consistency between the input and output images.

[0065] The neighborhood structural similarity loss is calculated as follows: several scanned image data that are spatially adjacent to the denoised scanned image data are acquired, the structural similarity index (SSIM) between the denoised scanned image data and each adjacent scanned image data is calculated, and then the average value is taken as the neighborhood structural similarity loss. The structural similarity index comprehensively evaluates the similarity between two images from three dimensions: brightness, contrast, and structure. The closer the value is to 1, the more similar the two images are. Since the anatomical structure of adjacent slices in PET images is continuous, the denoised image should maintain a high degree of structural consistency between adjacent slices. This loss term, by maximizing the neighborhood structural similarity, prompts the generator to preserve the interlayer structural continuity in three-dimensional space during the denoising process, avoiding interlayer discontinuity artifacts caused by processing each slice independently.

[0066] The composite loss function is obtained by weighted summation of adversarial loss, cycle consistency loss, supervision loss, identity mapping loss, and neighborhood structure similarity loss. Specifically, the composite loss equals λ1 multiplied by adversarial loss plus λ2 multiplied by cycle consistency loss plus λ3 multiplied by supervision loss plus λ4 multiplied by identity mapping loss plus λ5 multiplied by neighborhood structure similarity loss, where λ1, λ2, λ3, λ4, and λ5 are the weight coefficients of each loss term, used to balance the contribution of different loss terms to the overall optimization objective.

[0067] In this embodiment, the values ​​of each weight coefficient are as follows: λ1 is 0.5 to 2, λ2 is 0.5 to 2, because both adversarial loss and cycle consistency loss originate from the adversarial feedback of the discriminator. They are of similar magnitude and work together to form the core mechanism of the recurrent constraint adversarial network, hence they are assigned similar weight coefficients. If the values ​​are too large, the discriminator will dominate the training, easily leading to unstable generator training; if the values ​​are too small, the adversarial constraint will fail. λ3 is 5 to 20, because the supervisory loss directly calculates the pixel-level difference between the denoised image and the reference image, and its magnitude is usually much larger than that of the adversarial loss and cycle consistency loss, requiring a higher weight. The weights are set to fully leverage their direct constraint on image fidelity. If the weights are too small, the generator may over-rely on the adversarial loss, resulting in artifacts that do not conform to the distribution of the reference image. λ4 is set to 0.1 to 1 because the identity mapping loss, as an auxiliary constraint, guides the generator to maintain an identity mapping on the already clear input. Its weight should not be too large to avoid interfering with the generator's main task of denoising the low-dose image. λ5 is set to 0.1 to 1 because the neighborhood structure similarity loss is used to maintain the continuity between three-dimensional layers. As an auxiliary constraint, a moderate weight can achieve a balance between maintaining the consistency between layers and preserving the independent details of each slice.

[0068] This invention constructs a composite loss function that includes adversarial loss, cycle consistency loss, supervision loss, identity mapping loss, and neighborhood structure similarity loss. This function constrains the generator's training from multiple dimensions, enabling the generator to learn the distribution characteristics of real images in adversarial games, maintain the invertibility of denoising mapping under cycle consistency constraints, approximate high-quality reference images in supervised learning, maintain structural invariance in identity mapping, and maintain inter-layer continuity in neighborhood structure similarity. This comprehensively improves the fidelity of low-dose positron emission tomography (PET) image reconstruction.

[0069] Specifically, in step S7, the trained generator is obtained, including: The paired data is input into the cyclic constraint adversarial denoising network in batches. The backpropagation algorithm and gradient descent optimizer are used to optimize the network parameters of the generator, the first discriminator and the second discriminator alternately with the goal of minimizing the composite loss function. In each iteration, the first discriminator and the second discriminator are optimized first, and then the generator is optimized. Repeat the above iterative process until the value of the composite loss function no longer decreases or reaches the preset maximum number of iterations. Use the optimized generator at this point as the trained generator.

[0070] In this embodiment, the paired data is processed in batches according to batch size, which ranges from 1 to 16. In each batch, a first scanned image data and a corresponding second scanned image data are randomly selected from the paired data.

[0071] The specific steps of the alternating optimization are as follows: First, fix the network parameters of the generator, input the first scanned image data of the current batch into the generator to obtain denoised scanned image data, and then input the denoised scanned image data and the second scanned image data into the first discriminator to calculate the adversarial loss. Simultaneously, perform adaptive noise re-injection on the denoised scanned image data to obtain noisy scanned image data, and input the first scanned image data and the noisy scanned image data into the second discriminator to calculate the cycle consistency loss. Then, use the backpropagation algorithm to calculate the gradients of the first and second discriminators respectively, and use the gradient descent optimizer to update the network parameters of the first and second discriminators, thereby enhancing their discriminative capabilities. Second, fix the network parameters of the first and second discriminators, input the first scanned image data of the current batch into the generator, and sequentially pass it through forward propagation to obtain denoised scanned image data. Calculate the complete composite loss function (including adversarial loss, cycle consistency loss, and various additional losses), and then use the backpropagation algorithm to calculate the gradient of the generator. Use the gradient descent optimizer to update the generator's network parameters, making the denoised image generated by the generator closer to the real high-quality image and more capable of deceiving the two discriminators. The two steps above constitute a complete iteration.

[0072] The gradient descent optimizer uses the Adam optimizer, with momentum parameters β1 set to 0.5 and β2 set to 0.999. The learning rate of the generator is set to 0.0002, and the learning rates of the first and second discriminators are also set to 0.0002. The learning rate can be gradually reduced during training according to a preset decay strategy, for example, multiplying the learning rate by 0.9 every 100 iterations. The preset maximum number of iterations ranges from 100 to 500 training epochs, with each training epoch referring to traversing all paired data once. During training, after each training epoch, the current composite loss function value is calculated on the validation set. Training is terminated early when the composite loss function value no longer decreases for 10 consecutive training epochs to prevent overfitting.

[0073] After training, the generator parameters are saved to obtain the trained generator. This trained generator has learned the optimal mapping function from low-dose PET images to high-quality PET images and can be directly used in the subsequent inference stage. The paired data is divided into a training set and a validation set according to the proportion, with the training set accounting for 80% to 90% of all paired data and the validation set accounting for 10% to 20%. The training set is used to update the network parameters, and the validation set is used to monitor the training process and determine the convergence condition. The two are independent of each other to ensure that the trained generator has good generalization ability.

[0074] This invention employs an alternating training strategy of first optimizing the discriminator and then optimizing the generator, ensuring a dynamic balance between the generator and discriminator during the adversarial game and avoiding the problem of training collapse caused by one side becoming too strong. At the same time, by setting reasonable convergence conditions and iteration counts, the efficiency and stability of network training are guaranteed, ultimately obtaining a generator model with excellent denoising performance and good generalization ability.

[0075] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for high-fidelity reconstruction of low-dose positron emission tomography (PET) images, characterized in that, include: Step S1: Obtain paired data for the same imaging sample, wherein the paired data includes first scan image data and second scan image data; Step S2: Construct a cyclic constraint adversarial denoising network, which includes a generator, a first discriminator, and a second discriminator. Step S3: Input the first scanned image data into the generator and perform denoising processing to obtain denoised scanned image data; Step S4: Input the denoised scanned image data and the second scanned image data into the first discriminator, distinguish the denoised scanned image data and the second scanned image data through the first discriminator, and construct an adversarial loss based on the distinction result; Step S5: Perform adaptive noise re-injection on the denoised scan image data to obtain noisy scan image data. Input the first scan image data and the noisy scan image data into the second discriminator. The second discriminator distinguishes the first scan image data from the noisy scan image data and constructs a cycle consistency loss based on the distinction result. Step S6: Construct a composite loss function that includes the adversarial loss and the cycle consistency loss; Step S7: Train the cyclic constraint adversarial denoising network using the paired data and the composite loss function to optimize the network parameters of the generator, the first discriminator and the second discriminator, and obtain the trained generator; Step S8: Use the trained generator to denoise the new first scan image data and output the reconstructed scan image.

2. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S3, the generator comprises an encoder, a bottleneck layer, and a decoder connected in sequence to obtain denoised scan image data, including: Step S31: Input the first scanned image data into the encoder, and after several convolutional layers and downsampling operations, output the first feature map to the bottleneck layer; Step S32: Acquire several scan image data that are spatially adjacent to the first scan image data, extract the structural feature map of each adjacent scan image data, and fuse the structural feature map with the first feature map in the bottleneck layer to obtain a fused feature map; Step S33: Input the fused feature map into the decoder and output the denoised scan image data.

3. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 2, characterized in that, The fused feature map is obtained, including: Calculate the similarity matrix between the first feature map and the structural feature map; Attention weights are generated based on the similarity matrix; The attention weights are weighted together with the structural feature map, and then concatenated with the first feature map at the channel level to obtain the fused feature map.

4. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S4, the adversarial loss is constructed, including: The denoised scanned image data is used as the first sample, and the second scanned image data is used as the second sample, and input into the first discriminator; The first discriminator outputs the probability value that the denoised scanned image data is a real image; The binary cross-entropy loss and the adversarial loss are calculated based on the probability values.

5. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S5, adaptive noise re-injection is performed on the denoised scan image data to obtain noisy scan image data, including: Step S51: Detect lesion-sensitive regions in the denoised scan image data and generate a spatial attention mask. The spatial attention mask is used to identify lesion candidate regions and non-lesion regions in the denoised scan image data. Step S52: According to the spatial attention mask, apply a first noise intensity to the lesion candidate region in the denoised scan image data and apply a second noise intensity to the non-lesion region, wherein the first noise intensity is less than the second noise intensity; Step S53: The image after applying noise is used as the noise scan image data.

6. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 5, characterized in that, Generate a spatial attention mask, including: Extract the gradient map of the denoised scanned image data or the first scanned image data; The gradient map is segmented by a threshold to obtain high gradient regions; The high gradient region is subjected to morphological dilation to obtain a binary mask, which is then used as the spatial attention mask. The high gradient region corresponds to the lesion candidate region, and the remaining regions correspond to the non-lesion regions.

7. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S5, the cycle consistency loss is constructed, including: Noise is added to the denoised scan image data to obtain noisy scan image data; The first scanned image data is used as the first sample, and the noisy scanned image data is used as the second sample, which is then input into the second discriminator. The second discriminator outputs the probability value that the noisy scan image data is a real image; The binary cross-entropy loss is calculated based on the probability values, and the cycle consistency loss is obtained.

8. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S6, the composite loss function is constructed, including: Obtain the adversarial loss, the cycle consistency loss, and the additional loss; The adversarial loss, the cycle consistency loss, and the additional loss are weighted and summed to obtain the composite loss function.

9. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 8, characterized in that, Additional losses include: The supervised loss is obtained by calculating the sum of squared differences between corresponding pixels between the denoised scanned image data and the second scanned image data; The second scanned image data is input into the generator to obtain a reconstructed image. The pixel-level difference between the reconstructed image and the second scanned image data is calculated to obtain the identity mapping loss. Calculate the structural similarity index loss between the denoised scanned image data and several adjacent scanned image data to obtain the neighborhood structural similarity loss; The supervision loss, the identity mapping loss, and the neighborhood structure similarity loss are used as additional losses.

10. The method for high-fidelity reconstruction of low-dose positron emission tomography images according to claim 1, characterized in that, In step S7, the trained generator is obtained, including: The paired data is input into the cyclic constraint adversarial denoising network in batches. The backpropagation algorithm and gradient descent optimizer are used to optimize the network parameters of the generator, the first discriminator and the second discriminator alternately with the goal of minimizing the composite loss function. In each iteration, the first discriminator and the second discriminator are optimized first, and then the generator is optimized. Repeat the above iterative process until the value of the composite loss function no longer decreases or reaches the preset maximum number of iterations. Use the optimized generator at this point as the trained generator.

Images