Low-dose PET image denoising method based on prompt learning and line integral projection constraint

Abstract

The invention provides a low-dose PET image denoising method and system based on prompt learning and line integral projection constraint. The method comprises the following steps: constructing a PET image denoising model fused with a magnetic resonance multi-mode prompt learning mechanism; constructing a multi-domain loss function with physical consistency constraints; constructing a pairing training data set of a low-dose PET image, a standard-dose PET image and an MR image spatially corresponding to the PET image, and performing end-to-end training on the PET image denoising model by using the multi-domain loss function; and taking the acquired low-dose PET image and the MR image corresponding to the low-dose PET image as input, and carrying out denoising recovery on the low-dose PET image by utilizing the trained PET image denoising model to obtain a denoised PET image. According to the invention, on the premise of reducing the dosage of the radioactive tracer agent, the structural accuracy and the physical interpretability of the PET image are effectively improved, and a solution with high efficiency, low risk and high reliability is provided for clinical low-dosage PET examination.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a low-dose PET image denoising method based on cue learning and line integral projection constraints. More particularly, it relates to a low-dose PET image denoising method and system based on magnetic resonance multimodal cue learning and line integral projection constraints. Background Technology

[0002] In recent years, deep learning technology has demonstrated exceptional modeling capabilities in low-dose medical image reconstruction, learning complex and highly nonlinear mappings between degraded and high-quality images based on large amounts of paired training data. Low-dose PET imaging, a crucial development in molecular imaging, significantly reduces the radiation risk and potential biological side effects for patients by substantially decreasing the injected dose of radiopharmaceuticals, greatly improving the safety and clinical acceptability of the imaging process. It also helps reduce medical costs and promotes the widespread application of PET technology in routine screening and follow-up. Therefore, how to further reduce PET imaging dose while maintaining image quality has become a critical issue that urgently needs to be addressed in the field of medical imaging.

[0003] To address the common problems of high noise and structural blurring in low-dose PET images, numerous studies in recent years have focused on developing efficient image reconstruction and denoising methods. In particular, with the widespread adoption of simultaneous PET / MRI acquisition technology, more and more research is attempting to utilize multimodal paired data from PET and MRI for joint modeling to fully leverage the advantages of MRI images in spatial resolution and soft tissue contrast, thereby improving the reconstruction and denoising quality of low-dose PET images. However, most existing multimodal fusion methods employ only simple feature stitching or weighting strategies, failing to deeply explore the high-resolution anatomical information contained in MRI images and the complex semantic relationships between PET images. This results in limited fusion effects, insufficient structural recovery capabilities, and an inability to meet the actual clinical demand for high-quality low-dose PET images.

[0004] Magnetic resonance imaging (MRI) offers significant advantages such as being radiation-free, providing high soft tissue contrast, and high spatial resolution. It can provide crucial structural prior information for PET images, which is essential for the accurate reconstruction of anatomical details in PET images. Therefore, designing a more expressive and adaptive multimodal fusion mechanism to fully leverage the potential of MRI in structure-guided imaging has become a critical technical challenge that urgently needs to be overcome in the field of multimodal medical image processing.

[0005] Prompt learning, a deep learning mechanism that has emerged in recent years, originated in the field of natural language processing. Its core idea is to guide the model to understand and process input content in a specific way by introducing cue infographics, thereby improving the model's task adaptability and generalization ability. This mechanism has been gradually extended to the field of computer vision, significantly improving cross-modal modeling and structural representation capabilities by guiding the model to focus on specific semantic regions or structural features through cues. This invention introduces a cue learning mechanism, using MR images as cue sources. A neural network is designed to automatically extract MRI cue maps (MRIPrompt Maps, MPMs), and then deeply fuses them with PET encoded features through a Fusion Denoising Module (FDM), effectively guiding the multi-level feature recovery process of low-dose PET images. This mechanism can fully incorporate cross-modal structural knowledge, improve the model's ability to model anatomical structures, and enhance robustness to distribution differences between PET and MR modalities, providing a new technical path for high-quality denoising of low-dose PET images.

[0006] Furthermore, relying solely on the image domain loss function for optimization makes it difficult to guarantee the realism of the model's output image at the imaging physical level. Since PET images are essentially reconstructions of projection data, neglecting their projection consistency may result in good reconstruction in the image space, but deviations between projections and actual physical measurements, thus affecting clinical usability. To address this, this invention introduces line integral projection loss. This loss calculates the difference between the predicted image and the standard dose PET image in the projection domain by rotating the image at multiple angles and performing linear integral projection, effectively constraining the consistency of the image in the physical domain. The LIP operation is a linear process, preserving the differentiability of the loss function, and can achieve end-to-end optimization in the neural network, thereby improving the dual quality of the reconstructed image at both the physical and visual levels. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a low-dose PET image denoising method and system based on cue learning and line integral projection constraints.

[0008] A low-dose PET image denoising method based on cue learning and line integral projection constraints, provided by the present invention, includes:

[0009] Step S1: Construct a PET image denoising model that integrates a magnetic resonance multimodal cue learning mechanism;

[0010] Step S2: Construct a multi-domain loss function with physical consistency constraints;

[0011] Step S3: Construct a paired training dataset of low-dose PET images, standard-dose PET images, and MR images spatially corresponding to the PET images; and use the multi-domain loss function to perform end-to-end training on the PET image denoising model.

[0012] Step S4: Using the acquired low-dose PET image and the corresponding MR image as input, the trained PET image denoising model is used to denoise and restore the low-dose PET image, resulting in a denoised PET image.

[0013] Preferably, the denoising model includes a prompt generation module, a fusion denoising module, a PET image denoising backbone network, and a line integral projection module;

[0014] The prompt generation module is used to extract structural information from MRI images and convert it into a prompt information map that guides structural function;

[0015] The PET image denoising backbone network is a multi-level encoder-decoder network based on the Transformer structure, used for feature extraction, noise modeling and image restoration of the input low-dose PET images;

[0016] The fusion and denoising module is used to perform guided fusion of MRI structural cue information maps and PET intermediate features;

[0017] The line integral projection module is used to introduce prior constraints of the imaging physical domain, perform Radon transform on the output image to calculate the projection of the output image, compare it with the standard dose image reference projection data, and form a physically consistent supervision signal to provide cross-domain feedback for the training of the PET image denoising model.

[0018] Preferably, the prompt generation module takes the simultaneously acquired MR images as input, first extracts multi-level texture and structural features through a CNN encoder, then forms a global receptive field through a global average pooling operation, and obtains a low-dimensional structural MRI prompt information map by encoding with a fully connected layer and multiplying it with the feature map extracted by the CNN.

[0019] Preferably, the PET image denoising backbone network includes multiple Transformer Blocks, which have the ability to model long-distance dependencies, and learn the spatial and semantic feature changes of the image step by step through hierarchical downsampling and upsampling mechanisms. Features at each level are integrated through skip connections and fusion denoising modules.

[0020] Preferably, the fusion denoising module incorporates an internal fusion channel attention mechanism and Transformer modeling capabilities. The obtained MRI cue information map is applied to the PET feature channel weight adjustment. A fusion mask is generated through Sigmoid activation, and the PET features are reconstructed with weights. The fusion result is then input into the Transformer Block for further modeling.

[0021] Preferably, the loss function includes the mean absolute error in the image domain and the line integral projection error in the projection domain;

[0022] The mean absolute error loss in the image domain is used to constrain the consistency between the low-dose PET denoised image and the standard-dose PET image output by the model in the pixel intensity space, as shown in the following formula:

[0023] L mae (X,Y)=(l / N)∑ N {i＝1} |X i -Y i |

[0024] In the formula, L mae Let X represent the mean absolute error loss function in the image domain, where N represents the number of pixels, and X represents the mean absolute error loss function in the image domain. i and Y i These represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively.

[0025] The line integral projection loss in the projection domain is used to constrain the rationality and accuracy of the reconstructed image in the imaging physical domain. A constraint relationship is established between the Radon projection of the model output image and the standard dose PET measurement projection in the physical domain. By introducing a differentiable line integral projection operator, the image space and projection space are closely related, as shown in the following formula:

[0026]

[0027] Among them, P k X represents the line integral projection operator of the image at the k-th angle. i and Y i These represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively.

[0028] Preferably, the paired training dataset is selected from PET and MR synchronously acquired and registered image pairs covering multiple anatomical sites, multiple scanning conditions, and multiple pathological states;

[0029] The multiple anatomical sites include the heart and brain; the multiple scanning conditions include dose level and magnetic resonance contrast; and the multiple pathological states include healthy tissue and diseased tissue.

[0030] Preferably, step S4 includes:

[0031] Step S4.1: Acquire the low-dose PET image to be processed and the corresponding synchronously acquired MR image, and preprocess the low-dose PET image and the MR image;

[0032] Step S4.2: Input the preprocessed MR image as a structural prior to the prompt generation module, extract structural features through the CNN encoder and generate a structurally guided MRI prompt information map;

[0033] Step S4.3: The preprocessed low-dose PET image is input into the backbone denoising network module. The feature maps of different resolutions are extracted step by step through the Transformer encoding in the backbone denoising network. Each level of encoded features and the corresponding MRI prompt information map are deeply interacted in the fusion denoising module to form a guided enhancement feature map, thereby obtaining the denoised PET image.

[0034] Preferably, the preprocessing includes size adjustment, random angle rotation, and image intensity adjustment;

[0035] The MRI information map encodes the spatial distribution, boundary morphology, and tissue layer relationships of magnetic resonance anatomy, providing prior information for PET denoising modeling.

[0036] A low-dose PET image denoising system based on cue learning and line integral projection constraints, provided by the present invention, includes:

[0037] Module M1: Constructing a PET image denoising model that integrates a multimodal cue learning mechanism based on magnetic resonance imaging;

[0038] Module M2: Constructs a multi-domain loss function with physical consistency constraints;

[0039] Module M3: Constructs a paired training dataset of low-dose PET images, standard-dose PET images, and MR images spatially corresponding to the PET images, and uses the multi-domain loss function to perform end-to-end training on the PET image denoising model;

[0040] Module M4: The acquired low-dose PET image and the corresponding MR image are used as input. The trained PET image denoising model is used to denoise and restore the low-dose PET image to obtain the denoised PET image.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. This invention improves denoising performance through multimodal information fusion and physical consistency constraints. It effectively enhances the structural accuracy and physical interpretability of PET images while reducing the dosage of radioactive tracers, providing a highly efficient, low-risk, and highly reliable solution for clinical low-dose PET examinations, with broad clinical application prospects and promotional value.

[0043] 2. This invention introduces a cue learning mechanism into the PET denoising neural network model. By extracting structural information from high-resolution MRI images into cue information maps and deeply fusing them with features from low-dose PET images, it achieves effective guidance from cross-modal structural priors. This mechanism significantly enhances the structural fidelity and detail restoration capabilities of low-dose PET images, particularly in key areas such as edges and lesions, thereby improving the overall quality of reconstructed images and their clinical diagnostic value.

[0044] 3. This invention introduces physical domain supervision signals such as line integral projection loss (LIP-loss) during network training. By applying a differentiable Radon transform to the output image, it ensures that the reconstruction results are not only highly consistent with standard-dose PET images in the image domain, but also possess good rationality and consistency in the imaging physical domain. This physical consistency constraint effectively improves the reliability and clinical usability of the model output. Attached Figure Description

[0045] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0046] Figure 1 This is a schematic diagram of the working method of the present invention;

[0047] Figure 2 This is a schematic diagram of the multimodal PET image denoising model structure proposed in this invention;

[0048] Figure 3 This is a diagram illustrating the multimodal PET image denoising results of Embodiment 2 of the present invention;

[0049] Figure 4 This is a diagram showing the results of multimodal PET image denoising in Embodiment 3 of the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0051] This invention fully utilizes the high-resolution anatomical information provided by magnetic resonance imaging (MR) images. By introducing a cue learning mechanism, it extracts and fuses MRI cue maps to guide the denoising and detail restoration process of low-dose PET images. Simultaneously, considering the physical processes of PET imaging, it introduces line integral projection loss as a supervision signal in the physical domain during network training, enhancing the model's consistency in both the image and physical projection domains. Through the synergistic effect of these structural priors and physical constraints, this invention can significantly improve the structural fidelity and physical plausibility of low-dose PET images, thereby effectively improving the quality and clinical usability of reconstructed images.

[0052] Example 1

[0053] This invention provides a low-dose PET image denoising method based on cue learning and line integral projection constraints, such as... Figure 1 As shown, it includes:

[0054] Step S1: Construct a PET image denoising model that integrates a multimodal cue learning mechanism based on magnetic resonance imaging. For example... Figure 2 As shown, the denoising model includes a cue generation module, a fusion denoising module (FDM), a PET image denoising backbone network, and a line integral projection module, used to implement supervised constraints in the physical domain. Simultaneously acquired low-dose PET images and corresponding MR images are used as model inputs. The MR images are first input to the cue generation module to extract structural cue information maps; the PET images are input to the backbone denoising network to extract encoded features; subsequently, the cue information maps and the encoded features of the PET images are deeply fused in the fusion denoising module to guide the subsequent image denoising network's modeling process in terms of structure restoration and noise suppression.

[0055] The cue generation module extracts structural information from MRI images and converts it into a cue map that provides structural guidance. This module takes simultaneously acquired MR images as input. First, it extracts multi-level texture and structural features using a CNN encoder. Then, it forms a global receptive field through global average pooling (GAP) and multiplies the feature map extracted by the CNN with the data encoded by a fully connected layer (FC) to obtain a low-dimensional structural MRI cue map (MPM). This cue map aims to provide clues to missing or blurred anatomical structures in high-dose PET images, providing cross-modal prior information support for the denoising process.

[0056] The PET image denoising backbone network is a multi-level encoder-decoder network based on the Transformer structure, used for feature extraction, noise modeling, and image restoration of input low-dose PET images. The network backbone consists of multiple Transformer Blocks, possessing long-distance dependency modeling capabilities, and progressively learns spatial and semantic feature changes of the image through hierarchical downsampling and upsampling mechanisms. Features at each level are integrated through skip connections and a fusion denoising module to enhance feature recovery capabilities.

[0057] The fusion denoising module is used to perform guided fusion of MRI structural cue maps and PET intermediate features. This module utilizes a fusion channel attention mechanism and Transformer modeling capabilities to apply the obtained MRI cue map to PET feature channel weight adjustment. A fusion mask is generated through Sigmoid activation, and the PET features are reconstructed using weights. The fusion result is then input into the Transformer Block for further modeling, effectively improving the structural sensitivity and guidance of the denoising process.

[0058] The line integral projection module is used to introduce prior constraints in the imaging physical domain. This module simulates the forward projection process in PET imaging, calculates the projection of the output image by performing a Radon transform, and compares it with reference projection data of standard dose to form a physically consistent supervision signal that provides cross-domain feedback for network training.

[0059] Step S2: Construct a multi-domain loss function with physical consistency constraints. The loss function design adopts a joint multi-domain, multi-scale supervision strategy. The loss function includes the mean absolute error (MAE) in the image domain and the line integral projection error (LIP) in the projection domain, which serve as optimization objectives for model training, taking into account both image quality and physical plausibility.

[0060] The mean absolute error loss (MSE-loss) in the image domain is used to constrain the consistency between the low-dose PET denoised image and the standard-dose PET image output by the model in the pixel intensity space, and the formula is as follows:

[0061] L mae (X, Y) = (l / N)∑ N {i＝1} |X i -Y i |

[0062] In the formula, L mae Let X represent the mean absolute error loss function in the image domain, where N represents the number of pixels, and X represents the mean absolute error loss function in the image domain. i and Y iThese represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively. The mean absolute error loss in the image domain, by averaging the absolute errors of all pixels, effectively measures the deviation of the overall signal, ensuring the accuracy of the denoised image in terms of global structure and intensity distribution. This loss exhibits good numerical stability, suppresses large-scale global errors, and helps maintain the authenticity of the original anatomical structure and tissue signals.

[0063] The line integral projection loss (LIP-loss) in the projection domain is used to constrain the rationality and accuracy of the reconstructed image in the imaging physical domain. This loss term establishes a constraint relationship between the Radon projection of the model output image and the standard dose PET measurement projection in the physical domain. By introducing a differentiable line integral projection operator, the image space and projection space are closely related, realizing the consistency constraint between image and projection, as shown in the following formula:

[0064]

[0065] Among them, P k X represents the line integral projection operator of the image at the k-th angle. i and Y i represents the pixel value of the i-th pixel in the low-dose denoised PET image and the standard-dose PET image output by the model, respectively. This loss can significantly improve the physical interpretability of the model, making the reconstruction results more consistent with the physical generation mechanism of PET imaging, especially in different lesion regions and under low signal-to-noise ratio conditions, effectively improving the accuracy and reliability of reconstruction.

[0066] Step S3: Construct a paired training dataset consisting of low-dose PET images, standard-dose PET images, and spatially corresponding MR images. The PET image denoising model is trained end-to-end using the multi-domain loss function to optimize the network parameters. This step of constructing a diverse joint imaging dataset aims to improve the model's generalization ability and robustness. The paired training dataset selects PET and MR synchronously acquired and registered image pairs covering multiple anatomical sites, scanning conditions, and pathological states to ensure good adaptability of the model in various clinical scenarios. The multiple anatomical sites include the heart and brain; the multiple scanning conditions include dose level and MRI contrast; and the multiple pathological states include healthy tissue and diseased tissue. During acquisition, MR navigation, synchronous acquisition, or post-processing registration techniques are used to determine high-quality registration relationships between PET and MR images, achieving spatial consistency. For acquired samples including those with different genders, ages, body types, and physiological parameters, the model's cross-subject generalization ability is enhanced, and a standard-dose PET image is provided as a supervisory benchmark for each sample. The constructed dataset is used for joint training of the above loss functions. The prompt generation network, the denoising backbone network, and the fusion denoising module are optimized through end-to-end backpropagation, so that the prompt mechanism and the image modeling process are optimized in synergy, and the image restoration quality is gradually improved.

[0067] Step S4: Using the acquired low-dose PET images and their corresponding MR images as input, the trained PET image denoising model is used to denoise and restore the low-dose PET images. By extracting structural cue maps from the MR images and deeply fusing them with the PET feature maps in the fusion denoising module, the denoising and structural restoration of the PET images are guided layer by layer, ultimately outputting PET images with quality close to that of standard dose images. Step S4 includes:

[0068] Step S4.1: Acquire the low-dose PET image to be processed and the corresponding synchronously acquired MR image, and preprocess the low-dose PET image and the MR image. The preprocessing includes size adjustment, random angle rotation, and image intensity adjustment. The purpose of the preprocessing is to ensure that the two are aligned in the spatial coordinate system and avoid the negative impact of fusion error on modeling.

[0069] Step S4.2: The preprocessed MR image is used as a structural prior input to the cue generation module. A CNN encoder extracts structural features and generates a structure-guided MRI cue map (MPM). This MPM encodes the spatial distribution, boundary morphology, and tissue hierarchy of the MRI anatomical structures, providing effective prior information for PET denoising modeling. This MPM is injected into multiple fusion layers of the PET denoising backbone network, enhancing the structural representation and boundary reconstruction capabilities of the PET image.

[0070] Step S4.3: The preprocessed low-dose PET image is input into the backbone denoising network module. Through the Transformer encoding and decoding module within the backbone denoising network, multi-level feature modeling and reconstruction are performed. The MRI cue maps are fused, and the learned physical consistency priors are combined to finally output a high-quality PET image at the standard dose level. In other words, the low-dose PET image is used by the backbone denoising network module to extract feature maps of different resolutions level by level. Each level of encoded features and the corresponding MRI cue maps undergo deep interaction in the fusion denoising module to form a guided enhancement feature map, which guides the next stage of denoising modeling. Through the above feature fusion strategy, the model achieves a structure-guided mechanism utilizing multimodal information. While maintaining the regularity of PET signal distribution, it significantly enhances the accuracy of structure reconstruction, especially in key areas such as edge regions and lesion boundaries, demonstrating stronger reconstruction capabilities. The output image exhibits significant advantages in structural fidelity, detail restoration, and noise suppression, preserving key anatomical information and functional signal features, providing a reliable foundation for clinical diagnosis, lesion detection, and quantitative analysis.

[0071] The purpose of this invention is to address the shortcomings of existing low-dose PET image denoising methods in terms of structural fidelity and imaging physical consistency, and to achieve high-quality denoising and restoration of low-dose PET images while reducing the radiation dose.

[0072] Furthermore, in conjunction with the appendix Figure 3 The low-dose PET image denoising method based on magnetic resonance multimodal cue learning and line integral projection constraint of the present invention is described in detail below:

[0073] First, an end-to-end deep neural network architecture is used to construct a PET image denoising model, which includes a multimodal cue generation module, a cue fusion module, a backbone denoising network, and a line integral projection constraint module. The specific settings of each module are as follows:

[0074] (1) Cue Generation Module: This module aims to extract structural cue features for subsequent image reconstruction from paired magnetic resonance (MR) brain T1-weighted images. The input is a 256×256 MR image, which is first expanded to 32 channels using a 3×3 convolutional layer. Subsequently, the feature map is processed through three residual blocks and three downsampling operations, where the number of feature channels increases to 32, 64, 128, and 256 respectively, and the spatial resolution decreases to 256×256, 128×128, 64×64, and 32×32 respectively. Each residual block contains two standard convolutional layers, equipped with batch normalization and ReLU activation functions respectively, to enhance nonlinear expressiveness and alleviate the gradient vanishing problem. The downsampling operation uses a 3×3 convolutional layer with a stride of 2, combined with batch normalization and ReLU activation, to achieve spatial compression and feature extraction. After each downsampling, the spatial resolution of the feature map is halved while the number of channels is doubled, effectively preserving local structural information and improving feature representation capabilities. Following the backbone network, the module extracts global features across the channel dimensions using global average pooling and connects to a fully connected layer to generate channel attention weights. This weighted fusion of multi-scale features further enhances the representation of structural information. The final output feature map incorporates key structural information from the original MR image and can serve as structural cues for subsequent multimodal image denoising tasks.

[0075] (2) Fusion and Denoising Module: This module effectively fuses MRI cue features with PET backbone features to enhance structural fidelity in reconstructed images. Taking the decoding stage as an example, the inputs are feature maps from the low-dose PET backbone network of the brain and cue feature maps from the MR cue generation module. First, the MR cue features are processed through a 1×1 convolutional layer to adjust their channel count to match that of the PET backbone features, facilitating subsequent fusion operations. Subsequently, the adjusted MR cue features and PET backbone features are fused channel-wise with weighted fusion to achieve complementary enhancement of structural and metabolic information. The fused feature map is further input into a 3×3 convolutional layer to extract local spatial context information. Finally, the fused features are upsampled or downsampled using bilinear interpolation to match the spatial resolution required by each decoding layer of the backbone network, thereby achieving cross-scale feature alignment and structural information enhancement.

[0076] (3) Backbone Denoising Network: The backbone network takes low-dose brain PET images as input. It employs a multi-level Transformer structure, primarily consisting of an encoder, decoder, and a multi-scale feature fusion module, to achieve high-quality denoising and reconstruction of low-dose PET images. The encoder first extracts initial features from the input image through block embedding operations. These embedding operations utilize 3×3 convolutional layers to map the input low-dose PET image into an initial feature map. Subsequently, the feature map undergoes three downsampling operations and multi-level Transformer block processing. Each Transformer module consists of multiple basic blocks, each integrating a multi-head self-attention mechanism, a feedforward neural network, and a normalization layer to capture long-range dependency information and local structural features of the image. The decoder is symmetrical to the encoder structure, gradually restoring spatial resolution through upsampling operations implemented using pixel shuffle. After each decoding stage, the decoder output features are fused with the corresponding encoder layer features via skip connections, thus preserving multi-scale structural information. After each level of decoding feature generation, MR cue features of the corresponding scale are introduced and fused through a cue fusion module to enhance the ability to recover structural details. Finally, the feature map output by the decoder is mapped to a PET image of the same size as the original input image through a 3×3 convolutional layer, achieving noise suppression and image quality improvement.

[0077] (4) Line integral projection constraint module: During the model training phase, to further improve the physical consistency and quantitative accuracy of low-dose PET image reconstruction, this invention introduces a line integral projection constraint module into the loss function. Specifically, the denoised PET image output by the model is not only aligned with the standard-dose PET image in the pixel domain, but also mapped to the projection domain by Radon transform. Subsequently, the error between the two in the projection domain is calculated, and this error is used as one of the loss terms to participate in the end-to-end optimization of the model.

[0078] Then, a multi-domain loss function is employed to comprehensively consider the reconstruction quality of both the image domain and the projection domain. The loss function includes:

[0079] The mean absolute error (MAE) in the image domain measures the pixel-level difference between the denoised PET and the standard PET. This loss, by averaging the absolute errors of all pixels, effectively reflects the overall image reconstruction accuracy. MAE loss is insensitive to outliers and exhibits good numerical stability. It helps the model suppress large-scale systematic biases while maintaining global structure and signal intensity distribution, thereby improving the overall quality and reliability of the reconstructed image.

[0080] Line integral projection error (LIP) in the projection domain measures the physical consistency between denoised PET and standard PET in the Radon projection domain. This loss directly constrains the physical generation mechanism of the image by performing Radon projection on the image at multiple angles and calculating the mean square error between the corresponding projection curves. LIP loss can effectively improve the model's ability to recover structural details and physical properties, ensuring that the reconstruction result not only visually approximates the real image, but also conforms to the physical laws of PET imaging in the projection domain.

[0081] Next, a paired training dataset was constructed consisting of low-dose PET images of the brain, standard-dose PET images, and spatially corresponding MR images. The first PET image was represented using Standardized Uptake Value (SUV) to directly retain its original physical values ​​and maintain the medical interpretability of metabolic information. The MRI images were T1-weighted images of the brain, which are greatly affected by scanning parameters and equipment. To eliminate amplitude differences between images, linear normalization was used to scale their pixel values ​​to the [0,1] range, facilitating subsequent multimodal fusion processing. Considering the diverse sources of the original image data and the inconsistencies in spatial resolution and image size, all images were adjusted to a standard size of 256×256 to ensure the uniformity of the model input. For larger images, center cropping was used to retain the main anatomical structures and minimize the interference of background areas on model training.

[0082] To improve the model's generalization ability, in constructing the dataset, in addition to introducing real degraded image samples (such as low-dose, noisy, or blurry images generated during equipment acquisition), some standard-dose PET images or high-quality MRI images were also artificially degraded. Specific methods included adding Gaussian noise and simulating low-dose conditions to enhance the network's robustness to various degradation types and improve the adaptability of the fusion denoising module. All preprocessed image data were divided into training, validation, and test sets in an 8:1:1 ratio. This ensured a balanced data distribution while reducing the impact of sample bias on model performance evaluation, thus guaranteeing the scientific nature of the training process and the reliability of reconstruction performance.

[0083] Next, during the model training phase, to further improve the model's generalization ability and robustness, various data augmentation strategies were employed for the preprocessed medical image dataset. Specifically, each input image was first rotated at a random angle to simulate various orientation changes that might occur during actual clinical acquisition. Based on this data augmentation, the deep learning model was jointly trained using preprocessed PET and MRI multimodal medical image data. During training, the model calculated the loss according to the output of each relevant module, following the loss function defined in step S2. Specifically, the loss function design considered not only the image domain but also the projection domain error. The image domain error primarily measures the pixel-level difference between the denoised image and the real image, while the projection domain error compares the reconstructed image mapped to the projection space via Radon transform with standard-quality PET image projection data, thus constraining the model's accuracy in the projection space. This dual constraint effectively improves the model's overall performance in image denoising tasks, ensuring that the denoising result is not only visually highly consistent with the real image but also maintains physical consistency in the projection space. By minimizing the loss function and using the backpropagation algorithm to continuously update the learnable parameters of the model, the model's reconstruction ability and feature representation ability are gradually improved.

[0084] To ensure efficiency and convergence during training, this embodiment uses the Adam optimizer for parameter optimization. The initial learning rate is set to 0.0002, and after every 10 iterations, the learning rate is reduced to 0.5 times its original value to ensure more stable convergence in the later stages of training. The entire training process consists of 80 iterations, with performance evaluation performed on both the training and validation sets in each iteration. Once the loss function on the validation set reaches the convergence criterion, the final trained deep learning model is obtained.

[0085] Finally, low-dose PET and paired MR images are input into the trained model. The model automatically extracts MR structural cues and fuses them with PET features, guiding denoising and structural recovery layer by layer, ultimately outputting high-quality denoised PET images. The output images are highly consistent with standard-dose PET in both the pixel and projection domains, with rich details and clear structures, meeting clinical diagnostic needs. A schematic diagram of the results is shown below. Figure 3 As shown.

[0086] This invention enables high-quality, artifact-free PET image denoising and reconstruction for any given low-dose PET image and its corresponding multimodal magnetic resonance (MR) image. Based on the multimodal cue learning and line integral projection constraint mechanism proposed in this invention, the model can fully mine the rich structural information contained in the MR image and use it as a cue information map to guide feature modeling and structural recovery of the PET image. Through end-to-end joint training, the model not only has the ability to adapt to different anatomical structures and various imaging contrasts, but also effectively suppresses noise and artifacts caused by low-dose acquisition, significantly improving the diagnostic value and physical reliability of the reconstructed image.

[0087] Furthermore, the MR images in the above examples were replaced with LGE reconstructed images, and the resulting schematic diagram is shown in Figure 4.

[0088] Example 2

[0089] This invention also provides a low-dose PET image denoising system based on cue learning and line integral projection constraints. The low-dose PET image denoising system based on cue learning and line integral projection constraints can be implemented by executing the process steps of the low-dose PET image denoising method based on cue learning and line integral projection constraints. That is, those skilled in the art can understand the low-dose PET image denoising method based on cue learning and line integral projection constraints as a preferred embodiment of the low-dose PET image denoising system based on cue learning and line integral projection constraints.

[0090] A low-dose PET image denoising system based on cue learning and line integral projection constraints, provided by the present invention, includes:

[0091] Module M1: Constructs a PET image denoising model that integrates a multimodal cues learning mechanism based on magnetic resonance imaging (MRI). The denoising model includes a cues generation module, a fusion denoising module, a PET image denoising backbone network, and a line integral projection module. The cues generation module extracts structural information from MRI images and converts it into cues that guide structural development. Taking simultaneously acquired MR images as input, the module first extracts multi-level texture and structural features using a CNN encoder. Then, it forms a global receptive field through global average pooling and multiplies the feature map extracted by the CNN with the encoded data from a fully connected layer to obtain a low-dimensional structural MRI cues. The PET image denoising backbone network is a multi-level encoder-decoder network based on a Transformer structure, used for feature extraction, noise modeling, and image restoration of the input low-dose PET images. The PET image denoising backbone network includes multiple Transformer Blocks, possessing long-distance dependency modeling capabilities. It learns the spatial and semantic feature changes of the image step-by-step through hierarchical downsampling and upsampling mechanisms. Features at each level are integrated with the fusion denoising module through skip connections. The fusion denoising module is used to guide the fusion of MRI structural cue maps and PET intermediate features. Internally, the module utilizes a fusion channel attention mechanism and Transformer modeling capabilities to apply the obtained MRI cue map to PET feature channel weight adjustment. A fusion mask is generated through Sigmoid activation, and the PET features are reconstructed using weighted methods. The fusion result is then input into the Transformer Block for further modeling. The line integral projection module introduces prior constraints from the imaging physical domain. It performs a Radon transform on the output image to calculate its projection, which is compared with reference projection data from a standard dose image. This forms a physically consistent supervisory signal, providing cross-domain feedback for training the PET image denoising model.

[0092] Module M2: Constructs a multi-domain loss function with physical consistency constraints; the loss function includes the mean absolute error in the image domain and the line integral projection error in the projection domain; the mean absolute error loss in the image domain is used to constrain the consistency between the low-dose PET denoised image and the standard-dose PET image output by the model in the pixel intensity space, as shown in the following formula:

[0093] L mea (X, Y) = (1 / N)∑ N {i＝1} |X i -Y i |

[0094] In the formula, L mae Let X represent the mean absolute error loss function in the image domain, where N represents the number of pixels, and X represents the mean absolute error loss function in the image domain. i and Y iLet represent the pixel value of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively. The line integral projection loss in the projection domain is used to constrain the rationality and accuracy of the reconstructed image in the imaging physical domain. In the physical domain, a constraint relationship is established between the Radon projection of the model output image and the standard-dose PET measurement projection. By introducing a differentiable line integral projection operator, the image space and the projection space are closely related, as shown in the following formula:

[0095]

[0096] Among them, P k X represents the line integral projection operator of the image at the k-th angle. i and Y i These represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively.

[0097] Module M3: Constructs a paired training dataset of low-dose PET images, standard-dose PET images, and spatially corresponding MR images. The PET image denoising model is trained end-to-end using the multi-domain loss function. The paired training dataset selects PET and MR synchronously acquired and registered image pairs covering multiple anatomical sites, multiple scanning conditions, and multiple pathological states. The multiple anatomical sites include the heart and brain, the multiple scanning conditions include dose level and magnetic resonance contrast, and the multiple pathological states include healthy tissue and diseased tissue.

[0098] Module M4: Takes the acquired low-dose PET image and its corresponding MR image as input, and uses a trained PET image denoising model to denoise and restore the low-dose PET image, resulting in a denoised PET image. Module M4 includes: Module M4.1: Acquires the low-dose PET image to be processed and the corresponding synchronously acquired MR image, and preprocesses both the low-dose PET image and the MR image; Module M4.2: Inputs the preprocessed MR image as a structural prior to the prompt generation module, extracts structural features through a CNN encoder, and generates a structurally guided MRI prompt information map; Module M4.3: Inputs the preprocessed low-dose PET image to the backbone denoising network module, where Transformer encoding extracts feature maps of different resolutions at each level. Each level of encoded features interacts deeply with the corresponding MRI prompt information map in the fusion denoising module to form a guided enhanced feature map, thus obtaining the denoised PET image. The preprocessing includes size adjustment, random angle rotation, and image intensity adjustment; the MRI information map encodes the spatial distribution, boundary morphology, and tissue layer relationships of magnetic resonance anatomical structures, providing prior information for PET denoising modeling.

[0099] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0100] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A low-dose PET image denoising method based on cue learning and line integral projection constraints, characterized in that, include: Step S1: Construct a PET image denoising model that integrates a magnetic resonance multimodal cue learning mechanism; Step S2: Construct a multi-domain loss function with physical consistency constraints; Step S3: Construct a paired training dataset of low-dose PET images, standard-dose PET images, and MR images spatially corresponding to the PET images; and use the multi-domain loss function to perform end-to-end training on the PET image denoising model. Step S4: Using the acquired low-dose PET image and the corresponding MR image as input, the trained PET image denoising model is used to denoise and restore the low-dose PET image, resulting in a denoised PET image.

2. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 1, characterized in that, The denoising model includes a prompt generation module, a fusion denoising module, a PET image denoising backbone network, and a line integral projection module. The prompt generation module is used to extract structural information from MRI images and convert it into a prompt information map that guides structural function; The PET image denoising backbone network is a multi-level encoder-decoder network based on the Transformer structure, used for feature extraction, noise modeling and image restoration of the input low-dose PET images; The fusion and denoising module is used to perform guided fusion of MRI structural cue information maps and PET intermediate features; The line integral projection module is used to introduce prior constraints of the imaging physical domain, perform Radon transform on the output image to calculate the projection of the output image, compare it with the standard dose image reference projection data, and form a physically consistent supervision signal to provide cross-domain feedback for the training of the PET image denoising model.

3. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 2, characterized in that, The prompt generation module takes the simultaneously acquired MR images as input. First, it extracts multi-level texture and structural features through a CNN encoder. Then, it forms a global receptive field through a global average pooling operation. Finally, it obtains a low-dimensional structural MRI prompt information map by encoding with a fully connected layer and multiplying the feature map extracted by the CNN.

4. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 2, characterized in that, The PET image denoising backbone network includes multiple Transformer Blocks, which have the ability to model long-distance dependencies. It learns the spatial and semantic feature changes of the image step by step through hierarchical downsampling and upsampling mechanisms. Features at each level are integrated through skip connections and fusion denoising modules.

5. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 2, characterized in that, The fusion denoising module incorporates an internal fusion channel attention mechanism and Transformer modeling capabilities. The obtained MRI cue information map is applied to the PET feature channel weight adjustment. A fusion mask is generated through Sigmoid activation, and the PET features are reconstructed with weights. The fusion result is then input into the Transformer Block for further modeling.

6. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 1, characterized in that, The loss function includes the mean absolute error in the image domain and the line integral projection error in the projection domain; The mean absolute error loss in the image domain is used to constrain the consistency between the low-dose PET denoised image and the standard-dose PET image output by the model in the pixel intensity space, as shown in the following formula: L mae (X, Y) < (1 / N) Σ N {i＝1} |X i -Y i | In the formula, L mae Let X represent the mean absolute error loss function in the image domain, where N represents the number of pixels, and X represents the mean absolute error loss function in the image domain. i and Y i These represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively. The line integral projection loss in the projection domain is used to constrain the rationality and accuracy of the reconstructed image in the imaging physical domain. A constraint relationship is established between the Radon projection of the model output image and the standard dose PET measurement projection in the physical domain. By introducing a differentiable line integral projection operator, the image space and projection space are closely related, as shown in the following formula: Among them, P k X represents the line integral projection operator of the image at the k-th angle. i and Y i These represent the pixel values ​​of the i-th pixel in the low-dose PET denoised image and the standard-dose PET image output by the model, respectively.

7. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 1, characterized in that, The paired training dataset is selected from PET and MR synchronously acquired and registered image pairs covering multiple anatomical sites, multiple scanning conditions, and multiple pathological states. The multiple anatomical sites include the heart and brain; the multiple scanning conditions include dose level and magnetic resonance contrast; and the multiple pathological states include healthy tissue and diseased tissue.

8. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 1, characterized in that, Step S4 includes: Step S4.1: Acquire the low-dose PET image to be processed and the corresponding synchronously acquired MR image, and preprocess the low-dose PET image and the MR image; Step S4.2: Input the preprocessed MR image as a structural prior to the prompt generation module, extract structural features through the CNN encoder and generate a structurally guided MRI prompt information map; Step S4.3: The preprocessed low-dose PET image is input into the backbone denoising network module. The feature maps of different resolutions are extracted step by step through the Transformer encoding in the backbone denoising network. Each level of encoded features and the corresponding MRI prompt information map are deeply interacted in the fusion denoising module to form a guided enhancement feature map, thereby obtaining the denoised PET image.

9. The low-dose PET image denoising method based on cue learning and line integral projection constraints according to claim 8, characterized in that, The preprocessing includes size adjustment, random angle rotation, and image intensity adjustment; The MRI information map encodes the spatial distribution, boundary morphology, and tissue layer relationships of magnetic resonance anatomy, providing prior information for PET denoising modeling.

10. A low-dose PET image denoising system based on cue learning and line integral projection constraints, characterized in that, include: Module M1: Constructing a PET image denoising model that integrates a multimodal cue learning mechanism based on magnetic resonance imaging; Module M2: Constructs a multi-domain loss function with physical consistency constraints; Module M3: Constructs a paired training dataset of low-dose PET images, standard-dose PET images, and MR images spatially corresponding to the PET images, and uses the multi-domain loss function to perform end-to-end training on the PET image denoising model; Module M4: The acquired low-dose PET image and the corresponding MR image are used as input. The trained PET image denoising model is used to denoise and restore the low-dose PET image to obtain the denoised PET image.

Images