Method and apparatus for reconstruction of dynamic metabolic imaging signals

Abstract

This application proposes a method and apparatus for reconstructing dynamic metabolic imaging signals. First, raw observation data is obtained and preprocessed. The observation noise level is estimated from the preprocessed observation data. Then, a K-round iterative reconstruction process is performed on the preprocessed observation data until convergence, ultimately yielding the reconstruction result. In each iteration, firstly, at the corresponding noise scale, based on the constructed diffusion prior denoising model, the current estimate is updated to an intermediate denoising result. Then, the intermediate denoising result is combined with the observation data for data consistency projection to obtain the next estimate. The noise scale and iterative reconstruction parameters are determined by the observed noise level. This approach can improve the signal-to-noise ratio while preserving subtle spectral peaks and true dynamic changes, and enhance the stability and reliability of metabolite quantification and subsequent kinetic parameter estimation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for reconstructing dynamic metabolic imaging signals. Background Technology

[0002] The goal of metabolic imaging is not just to examine anatomical structures, but to observe the chemical reactions occurring within tissues (such as glucose uptake and lactate production). Deuterium (²H) metabolic imaging typically involves administering a deuterated substrate (such as deuterated glucose) to the subject, followed by voxel-by-voxel acquisition using deuterium nuclear magnetic resonance spectroscopy. This simultaneously yields both spatial distribution and spectral information, and allows for repeated acquisition over time to create dynamic sequences for observing metabolic flux / kinetic processes. Dynamic metabolic imaging (DMI) is considered a relatively simple and robust approach combining metabolic tracing and MR imaging, and systematic reviews and exemplary studies have been conducted, particularly in the study of metabolism in organs such as the brain.

[0003] In 2H magnetic resonance spectroscopy (MRSI), each voxel acquires not a regular photograph, but a complex time signal, often called free induction decays (FIDs). The spectrum is obtained through Fourier transform, with peaks at different frequencies corresponding to different metabolites (e.g., glucose, lactate, Glx). Dynamic 2H MRSI involves repeated acquisition at multiple time points, so the final data can often be understood as follows: Spatial dimension: voxel location (two-dimensional / three-dimensional); Spectral dimension: frequency sampling points; Temporal dimension: dynamic frame; Complexity: real / imaginary part (phase information is crucial for quantification).

[0004] The core bottleneck faced by this type of data is the low signal-to-noise ratio (SNR):²H has inherently low sensitivity, and the concentration of deuterium-labeled metabolites in vivo is not high; in order to observe dynamic processes, short frame times are required, resulting in more noise per frame. Experimental papers generally emphasize that low SNR directly affects the visibility of metabolic peaks, the stability of spectral fitting, and the reliability of subsequent kinetic parameter estimation.

[0005] Current dynamic deuterium metabolism imaging / dynamic ²H spectroscopy imaging generally faces the problem of insufficient signal-to-noise ratio in clinical and scientific research applications: due to the low ²H gyromagnetic ratio, low concentration of labeled metabolites in vivo, and the need for rapid frame-by-frame acquisition, the measured spectral signal is significantly contaminated by noise; low SNR not only masks subtle spectral features and spatial / temporal heterogeneity, but also transmits high variance and bias to subsequent spectral fitting and kinetic modeling, resulting in unstable or even misleading interpretations of metabolic quantification and kinetic parameter estimation.

[0006] Meanwhile, existing denoising / reconstruction schemes suffer from several key defects: model-driven methods (such as low-rank / TV / subspace methods) are sensitive to hyperparameters such as rank / regularization, which can easily lead to oversmoothing, loss of detail, or even suppression of real focal metabolic changes; pure data-driven deep learning methods are limited by the scarcity of "clean labels" and domain shifts across scanners / coils / field strengths / protocols, and more importantly, they may violate measurement physics and produce "illusionary" pseudostructures, making it difficult to meet the reliability requirements of high-risk clinical interpretations; and existing self-supervised denoising often relies on statistical assumptions such as independent zero-mean noise, which are prone to failure when there is correlated, non-stationary noise and structured artifacts in real data, resulting in oversmoothing, loss of detail, or output bias.

[0007] Therefore, under noisy signal conditions, how to achieve reliable denoising / reconstruction of dynamic²H spectral-time data, while significantly improving SNR and preserving subtle spectral peaks and dynamic metabolic evolution characteristics, and strictly meeting the data consistency of measurement physics, thereby improving the accuracy and stability of metabolite quantification and downstream kinetic parameter estimation, is an urgent problem to be solved. Summary of the Invention

[0008] This application proposes a method and device for reconstructing dynamic metabolic imaging signals, which can solve one of the problems existing in the background art.

[0009] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for reconstructing dynamic metabolic imaging signals is provided, including: Raw observation data is obtained and preprocessed, the raw observation data being obtained from dynamic metabolic imaging (DMI) applied to the target; Observational data obtained from preprocessing Estimated observation noise level ; Furthermore, a K-round iterative reconstruction process is performed on the preprocessed observation data, k=1,2……K, until convergence, finally obtaining the reconstruction result: In each iteration, firstly, at the corresponding noise scale... Based on the constructed diffusion prior denoising model, the current estimate is then... Update to intermediate denoising results Then, the intermediate denoising results With observation data By combining data consistency projection, the next estimate is obtained. Noise scale The parameters for iterative reconstruction are determined by the observed noise level. Sure.

[0010] In one possible design of the first aspect, the diffusion prior denoising model includes: a noise conditional encoder and a noise prediction network, wherein the noise conditional encoder is used to convert the noise scale... The encoding is a conditional embedding vector, which is then injected into the corresponding layer of the noise prediction network.

[0011] In one possible design of the first aspect, the noise conditional encoder includes, in sequence, a logarithmic transformation module, a sinusoidal position encoding module, and a multilayer perceptron (MLP). The logarithmic transformation module is used to perform a logarithmic transformation operation on the noise scale, the sinusoidal position encoding module is used to extract high-dimensional periodic features from the logarithmic values, and the multilayer perceptron is used to map the high-dimensional periodic features into a conditional embedding vector.

[0012] In one possible design of the first aspect, the noise prediction network adopts a U-shaped encoder-decoder structure, which includes, in sequence, an input convolutional layer, a multi-level coding layer, a bottleneck layer, and a multi-level decoding layer, with skip connections between the multi-level coding layer and the multi-level decoding layer. The multi-level coding layer is used to extract local spectral peak morphology features, cross-frequency structural features, and dynamic time evolution features. The bottleneck layer is used to integrate global time-frequency context information and extract high-level spectral structure representations. The multi-level decoding layer is set correspondingly to the multi-level coding layer.

[0013] In one possible design of the first aspect, the conditional embedding vector is injected into the FiLM conditional modulation unit in the coding layer, the bottleneck layer and the decoding layer. The FiLM conditional modulation unit is used to generate scaling parameters and offset parameters of the corresponding layer based on the conditional embedding vector, thereby achieving adaptive adjustment of the denoising intensity.

[0014] In one possible design of the first aspect, the multi-level coding layer includes a first coding layer, a second coding layer, a third coding layer and a fourth coding layer cascaded in sequence. The first coding layer includes several residual blocks. The second coding layer, the third coding layer and the fourth coding layer each include downsampling and several residual blocks. The bottleneck layer includes several residual blocks. The multi-level decoding layer includes a fourth decoding layer, a third decoding layer, a second decoding layer and a first decoding layer cascaded in sequence. The fourth decoding layer, the third decoding layer, the second decoding layer and the first decoding layer each include skip connection splicing and several residual blocks.

[0015] In one possible design of the first aspect, the residual block comprises, in sequence, a group normalization layer, a SiLU activation function, a 3×3 convolutional layer, a FiLM conditional modulation layer, a Dropout layer, and residual connections.

[0016] In one possible design approach of the first aspect, noise scale Corresponding to a diffused noise sequence: ,in, Indicates an initial high noise level. This indicates the termination of the low noise level, where K represents the number of iterations, and the intermediate denoising results are... With observation data By combining data consistency projection, the next estimate is obtained. Specifically, it includes: The intermediate denoising results are adaptively adjusted based on the strength of the noise scale. With observation data The fusion weights; And, using intermediate denoising results With observation data The fusion result is used as the next estimate .

[0017] In one possible design of the first aspect, the method for reconstructing the dynamic metabolic imaging signal further includes: Apply smoothness, monotonicity, peak timing, delayed response, or dynamic model consistency constraints to the reconstruction results.

[0018] In a second aspect, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory to cause the electronic device to perform the dynamic metabolic imaging signal reconstruction method as described in any possible implementation of the first aspect.

[0019] Beneficial effects: Based on the above technical solution, the original observation data is first obtained and preprocessed. The observation noise level is estimated from the preprocessed observation data. Then, K rounds of iterative reconstruction are performed on the preprocessed observation data until convergence, and the reconstruction result is finally obtained. In each iteration, firstly, under the corresponding noise scale, the current estimate is updated to the intermediate denoising result based on the constructed diffusion prior denoising model. Then, the intermediate denoising result is combined with the observation data for data consistency projection to obtain the next estimate. The noise scale and the parameters of iterative reconstruction are determined by the observation noise level. In this way, while improving the signal-to-noise ratio, subtle spectral peaks and real dynamic changes can be preserved as much as possible, and the stability and reliability of metabolite quantification and subsequent kinetic parameter estimation can be improved. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method for reconstructing dynamic metabolic imaging signals provided in Embodiment 1 of this application; Figure 2 This is a block diagram of the dynamic metabolic magnetic resonance signal denoising and reconstruction device provided in Embodiment 2 of this application; Figure 3 This is a flowchart of the dynamic metabolic magnetic resonance signal denoising and reconstruction method provided in the embodiments of this application; Figure 4 This is a block diagram of the noise prediction network structure provided in the embodiments of this application; Figure 5 This is a comparison chart of the denoising results provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for reconstructing dynamic metabolic imaging signals, including: Step S101: Obtain raw observation data and perform data preprocessing. The raw observation data is obtained by dynamic metabolic imaging (DMI) applied to the target. Step S102, from the preprocessed observation data Estimated observation noise level ; And, in step S103, a K-round iterative reconstruction process is performed on the preprocessed observation data, k=1,2……K, until convergence, and the final reconstruction result is obtained: In each round of iteration, firstly, at the corresponding noise scale... Based on the constructed diffusion prior denoising model, the current estimate is then... Update to intermediate denoising results Then, the intermediate denoising results With observation data By combining data consistency projection, the next estimate is obtained. Noise scale The parameters for iterative reconstruction are determined by the observed noise level. Sure.

[0026] Specifically, preprocessing may include one or more of the following: frequency band truncation, zero-order phase correction, first-order phase correction, frequency drift correction, amplitude normalization, baseline correction, channel merging, channel rearrangement, and abnormal frame removal.

[0027] Observation of noise level It can be determined in one of the following ways: from observation data Noise samples are extracted from frequency bands without obvious metabolic peaks, spectral edge bands, or preset noise reference regions, and the standard deviation or variance of the observed noise is calculated to obtain... .

[0028] Based on the observed noise level This can generate a corresponding diffuse noise sequence: ,in, Indicates an initial high noise level. This indicates the termination of the low noise level, and K represents the number of iterations.

[0029] The diffusion process of the diffusion prior denoising model includes a forward noise addition process and a backward denoising process. In the forward process, noise is gradually added to the clean spectrum X0 to form XK; in the backward process, the system gradually denoises from a high-noise state, sequentially recovering the clean spectrum XK. .

[0030] In one possible implementation of this embodiment, the diffusion prior denoising model includes: a noise conditional encoder and a noise prediction network, wherein the noise conditional encoder is used to convert the noise scale... The encoding is a conditional embedding vector, which is then injected into the corresponding layer of the noise prediction network.

[0031] The noise conditional encoder includes a logarithmic transformation module, a sinusoidal position encoding module, and a multilayer perceptron (MLP) cascaded in sequence. The logarithmic transformation module is used to perform a logarithmic transformation operation on the noise scale. The sinusoidal position encoding module is used to extract high-dimensional periodic features from the logarithmic values. The multilayer perceptron is used to map the high-dimensional periodic features into conditional embedding vectors.

[0032] The noise prediction network adopts a U-shaped encoder-decoder structure, which includes a series of cascaded input convolutional layers, multi-level coding layers, bottleneck layers, and multi-level decoding layers. The multi-level coding layers and multi-level decoding layers are connected in a skip connection. The multi-level coding layers are used to extract local spectral peak morphology features, cross-frequency structure features, and dynamic time evolution features. The bottleneck layer is used to integrate global time-frequency context information and extract high-level spectral structure representations. The multi-level decoding layers are set up correspondingly to the multi-level coding layers.

[0033] The conditional embedding vector is injected into the FiLM conditional modulation unit in the coding layer, bottleneck layer and decoding layer. The FiLM conditional modulation unit is used to generate the scaling parameters and offset parameters of the corresponding layer according to the conditional embedding vector, thereby realizing adaptive adjustment of the denoising intensity.

[0034] The multi-level coding layer includes the following cascaded layers: first coding layer, second coding layer, third coding layer and fourth coding layer. The first coding layer includes several residual blocks. The second coding layer, third coding layer and fourth coding layer each include downsampling and several residual blocks. The bottleneck layer includes several residual blocks. The multi-level decoding layer includes the following cascaded layers: fourth decoding layer, third decoding layer, second decoding layer and first decoding layer. The fourth decoding layer, third decoding layer, second decoding layer and first decoding layer each include skip connection splicing and several residual blocks.

[0035] The residual block comprises, in sequence, a group normalization layer, a SiLU activation function, a 3×3 convolutional layer, a FiLM conditional modulation layer, a Dropout layer, and residual connections.

[0036] In one possible implementation of this embodiment, the intermediate denoising result is... With observation data By combining data consistency projection, the next estimate is obtained. Specifically, it includes: The intermediate denoising results are adaptively adjusted based on the strength of the noise scale. With observation data The fusion weights; And, using intermediate denoising results With observation data The fusion result is used as the next estimate .

[0037] In this way, an adaptive balance between fidelity and denoising intensity can be achieved at different iteration stages.

[0038] In one possible implementation of this embodiment, the method for reconstructing the dynamic metabolic imaging signal further includes: Apply smoothness, monotonicity, peak timing, delayed response, or dynamic model consistency constraints to the reconstruction results.

[0039] This can suppress non-physiological shaking and maintain the continuity of metabolic changes.

[0040] This embodiment first obtains the original observation data and performs data preprocessing. The observation noise level is estimated from the preprocessed observation data. Then, a K-round iterative reconstruction process is performed on the preprocessed observation data until convergence, and the reconstruction result is finally obtained. In each iteration, firstly, under the corresponding noise scale, based on the constructed diffusion prior denoising model, the current estimate is updated to the intermediate denoising result. Then, the intermediate denoising result is combined with the observation data for data consistency projection to obtain the next estimate. The noise scale and the parameters of iterative reconstruction are determined by the observation noise level. In this way, while improving the signal-to-noise ratio, subtle spectral peaks and real dynamic changes can be preserved as much as possible, and the stability and reliability of metabolite quantification and subsequent kinetic parameter estimation can be improved.

[0041] Example 2 The purpose of this embodiment is to provide a denoising / reconstruction method and corresponding system / storage medium that takes into account both measurement physical consistency and metabolic kinetic temporal regularity under the conditions of low SNR, complex noise / artifacts, and difficulty in obtaining paired clean label data for dynamic DMI / ²H MRSI. This allows the output spectral-time data to improve SNR while preserving subtle spectral peaks and true dynamic changes as much as possible, and to improve the stability and reliability of metabolite quantification and subsequent kinetic parameter estimation.

[0042] To achieve the above objectives, this embodiment aims to achieve the following technical goals: Introducing generative priors: learning the statistical distribution of clean spectral-time data using a diffusion / score-based model; Strengthen data consistency: Explicitly use measurement models to constrain the output to be consistent with the observations during the iterative recovery process, reducing the risk of distortion caused by violations of physics; Incorporating metabolic dynamics time priors (optional): By using the spectrum-time network structure and dynamic regularization / constraints during the training phase, the smoothness and interpretability of metabolic evolution are incorporated into the recovery process, reducing the cumulative effects of oversmoothing and output bias in the time dimension.

[0043] 1. Overall Overview of the Technical Solution This embodiment provides a device and method for denoising and reconstructing dynamic metabolic magnetic resonance signals. It is used to denoise and reconstruct low signal-to-noise ratio complex spectral data acquired through dynamic deuterium metabolism imaging (or dynamic magnetic resonance spectroscopy / spectral imaging), thereby outputting metabolite spectra, metabolite concentration curves, and metabolic parameter maps with higher reliability. This technical solution couples generative diffusion priors with physical data consistency constraints and metabolic kinetic temporal constraints. During the iterative reconstruction process, it alternately performs "prior denoising update" and "likelihood projection update," and supports automatic estimation of noise levels from the signal-free frequency bands of the spectrum, enabling robust applications without paired ground truth data (the overall process and network structure can be found in the method and structure description in the document).

[0044] 2. Device structure and connection relationship As attached Figure 1 As shown, the device in this embodiment is used for physically guided diffusion denoising and reconstruction of dynamic magnetic resonance spectral data. Taking dynamic complex spectral data acquired through magnetic resonance acquisition as input, the device works collaboratively with modules such as data preprocessing, noise estimation, diffusion prior denoising, data consistency projection, and metabolic kinetic constraints. While maintaining the true peak structure and temporal evolution of the metabolic spectrum, it effectively suppresses random noise and acquisition artifacts, ultimately outputting denoised dynamic spectral data, metabolite concentration-time curves, and related metabolic parameter graphs.

[0045] The device in this embodiment includes at least the following components: 2.1 Magnetic Resonance Acquisition Module The magnetic resonance acquisition module is used to acquire dynamic magnetic resonance spectral data of the target object to obtain raw observation data y. The target object can be a human body, animal, isolated tissue, cell sample, or metabolic model system. The acquisition module can employ deuterium nuclear magnetic resonance imaging sequences, three-dimensional chemical shift imaging sequences, dynamic magnetic resonance spectroscopy sequences, or other magnetic resonance acquisition units capable of obtaining "time-frequency" spectral data.

[0046] The data output by this module is typically in complex form, containing amplitude and phase information. Its data dimensions can include time points, spatial locations, receiving channels, and frequency sampling points. For dynamic metabolic imaging scenarios, the acquisition module can continuously acquire multiple frames of spectral data within a given time interval to reflect the uptake, transformation, or clearance of target metabolites over time.

[0047] exist Figure 1 In this section, this module corresponds to the "Input" spectrum section on the left. The input spectrum can be represented as spectral curves at different time points, where the horizontal axis represents frequency or chemical shift, the vertical or stacking direction represents time frames, and the spectral peak regions reflect different metabolite signals. Due to the limited acquisition time of dynamic spectral imaging, the raw observation data usually has a low signal-to-noise ratio, and the spectral peaks may be submerged by noise, thus requiring subsequent denoising and reconstruction processing.

[0048] 2.2 Data Preprocessing Module The data preprocessing module is connected to the magnetic resonance acquisition module and is used to perform format conversion and quality correction on the raw observation data y to form an input tensor suitable for neural network processing. This module may include sub-units such as frequency band truncation, zero-order phase correction, first-order phase correction, frequency drift correction, amplitude normalization, baseline correction, channel merging, channel rearrangement, and outlier frame removal.

[0049] In a preferred embodiment, the preprocessing module splits the complex spectral data into two channels: real and imaginary, and organizes them into a tensor form as follows: [B,T,C,F]. Here, B represents the batch size, T represents the number of dynamic time points, C=2 represents the two channels (real and imaginary), and F represents the number of frequency sampling points. For spectral imaging data containing spatial dimensions, it can be further extended to [B,T,X,Y,Z,C,F] or the spatial positions can be expanded into batch dimensions for processing, where X, Y, and Z are pixel coordinates.

[0050] This module not only performs data format adaptation but also minimizes the impact of non-metabolic factors on subsequent denoising. For example, phase inconsistency can cause information aliasing between the real and imaginary parts of the complex spectrum, frequency drift can cause shifts in the position of metabolic peaks, and differences in amplitude scale can affect the network's judgment of noise levels. Therefore, the preprocessing module can improve the stability of the subsequent diffusion prior denoising module and data consistency projection module.

[0051] 2.3 Noise Estimation Module The noise estimation module is connected to the data preprocessing module and is used to estimate the observed noise level from the preprocessed spectral data. Specifically, this module can extract noise samples from frequency bands without obvious metabolic peaks, spectral edge bands, or preset noise reference regions, and calculate the standard deviation or variance of the observed noise to obtain... .

[0052] The noise estimation results can serve as important inputs for subsequent reconstruction control modules and data consistency projection modules. On the one hand, It can be used to determine the initial noise level, noise scheduling range, and number of iterations in the diffusion denoising process; on the other hand, It can be used to control the fusion weights between observed data and network priors in data consistency projection.

[0053] For example, when the observation noise is high, the system can enhance the role of the diffusion prior denoising module, making the reconstruction results more dependent on the learned spectral structure prior; when the observation noise is low, the system can enhance the data consistency constraint, making the reconstruction results closer to the original observation data and avoiding over-smoothing or spectral peak distortion.

[0054] 2.4 Diffusion Prior Denoising Module The diffusion prior denoising module is one of the core modules of the device in this embodiment. This module is used to perform prior denoising on the current spectrum estimation result under a given noise level, and output the intermediate denoising result. As shown in the upper part of the attached diagram, the diffusion prior denoising module may include "metabolic kinetics prior", "model prior" and the corresponding diffusion model structure.

[0055] This module preferably includes a noise prediction network and a noise conditional encoder. The noise prediction network can adopt a U-shaped encoder-decoder structure, including multi-layer convolution, normalization, non-linear activation, downsampling, upsampling, and skip connections. For dynamic spectral data, the network can also introduce temporal modeling units, such as one-dimensional temporal convolution, recurrent structures, Transformer structures, or temporal self-attention modules, to capture metabolic evolution relationships between different time points.

[0056] The noise condition encoder is used to determine the current diffuse noise level. The noise conditional encoder is encoded as a feature vector and injected into different layers of the noise prediction network. This noise conditional encoder can employ Gaussian Fourier feature maps, multilayer perceptrons, or positional coding structures, enabling the network to adaptively adjust the denoising intensity according to different noise levels.

[0057] In the k-th iteration, the diffusion prior denoising module receives the current estimate. and noise level It predicts noise components or directly predicts clean spectral components and generates intermediate results. This process is equivalent to the stepwise denoising step in the "reverse process" at the bottom of the attached diagram, that is, gradually recovering a clearer signal that is more consistent with the metabolic spectrum structure from a high-noise state.

[0058] Compared with traditional filtering methods, this module not only removes noise based on the assumption of local smoothness, but also learns the morphological characteristics, peak width characteristics, complex spectrum phase characteristics, and dynamic time curve characteristics of real metabolic peaks, thereby reducing the risk of peaks being accidentally deleted or metabolic kinetic information being destroyed.

[0059] 2.5 Data Consistency Projection Module The data consistency projection module is connected to the diffusion prior denoising module and the noise estimation module. It is used to constrain the diffusion prior denoising results based on the physical observation model and generate the next reconstruction result. .

[0060] In the accompanying diagram, this module corresponds to the physical constraint part of the "Physically Guided Denoising Framework." The diffusion model provides learned priors, but relying solely on neural networks may introduce illusory spectral peaks, weaken true weak peaks, or alter the true observation structure. Therefore, this embodiment introduces data consistency projection after each diffusion denoising step, ensuring that the reconstruction result remains statistically consistent with the original observation data. Maintain consistency.

[0061] This module can be based on the following observation model: in, The raw observation data obtained by the acquisition module. The true dynamic spectrum to be recovered. For the physical observation operator corresponding to the acquisition system, This represents the noise term. For the case where only spectral domain denoising is performed, It can be an identity operator; for undersampled spectral imaging, k-space acquisition, or multi-channel reconstruction, etc. It can include Fourier transform, sampling mask, coil sensitivity coding, or spectral transform operator.

[0062] In a preferred embodiment, the data consistency projection module employs a closed-loop fusion update to diffuse prior results. With observation data The weighted fusion based on noise level and observed noise intensity, and the consistent projection step can be written as the following optimization problem: This is the joint constraint of "observational consistency" and "diffusion prior consistency". When The observation model is simplified to: The above optimization problem has a closed-form solution: In the high-noise iteration phase, the module can retain more observational information to prevent the reconstruction results from deviating from the actual acquired data; in the low-noise iteration phase, the module can rely more on diffusion priors to obtain a smoother and clearer spectral peak structure. Alternatively, a contrasting or adaptive fusion strategy can be adopted depending on the actual design, i.e., based on... , The weights are dynamically adjusted based on the residual size and the confidence level of the spectral peaks.

[0063] The output of this module will be used as the input for the next diffusion prior denoising, thus forming an alternating iterative structure of "prior denoising - physical projection - re-denoising - re-projection".

[0064] 2.6 Metabolic Kinetics Constraint Module The metabolic kinetics constraint module is used to constrain the evolution of dynamic spectra over time, so that the reconstruction results not only have reasonable peak shapes at a single time point, but also conform to the physical and physiological laws of metabolite concentration changes over time throughout the dynamic process.

[0065] like Figure 1 As shown above, this embodiment not only utilizes model priors but also introduces metabolic kinetic priors. This module can impose constraints on the reconstruction results based on the time curve characteristics of the target metabolite, including smoothness, monotonicity, peak timing, delayed response, or kinetic model consistency. For example, for substrate uptake processes, the time curve can be constrained to show a gradual upward trend or a trend of initial rise followed by stabilization; for metabolite formation processes, a certain time delay relative to the substrate signal can be constrained; and for clearance processes, an exponential decay or multi-compartment model variation pattern can be constrained.

[0066] This module can be used as part of the loss function during the training phase, or as a post-processing or iterative constraint during the inference phase. During training, metabolic kinetic constraints may include time smoothing regularization, peak integral curve consistency loss, and metabolite fitting parameter constraint loss. During inference, this module can perform outlier detection, smoothing correction, or kinetic parameter fitting on the time curve obtained from the integration of each metabolic peak, and feed the results back to the reconstruction control module or the data consistency projection module.

[0067] This module reduces temporal curve jitter caused by random noise, minimizes the impact of single-frame abnormal reconstruction results on metabolic parameter estimation, and improves the stability and reliability of dynamic metabolic analysis.

[0068] 2.7 Rebuild the control module The reconstruction control module coordinates the execution order of the above modules and controls the noise scheduling, iteration count, and stopping conditions during the diffusion denoising process. This module receives the output from the noise estimation module. And generate a diffuse noise sequence according to a preset strategy: in, Indicates an initial high noise level. This indicates the termination of the low noise level, and K represents the number of iterations.

[0069] As shown in Figure 1 below, the diffusion process includes a forward noise addition process and a reverse noise reduction process. In the forward process, noise is gradually added to the clean spectrum X0, forming a new clean spectrum. In the reverse process, the system gradually removes noise from a high-noise state, sequentially recovering the clean spectrum. In this embodiment, the reverse denoising process does not rely solely on neural networks, but combines data consistency projection and metabolic kinetic constraints at each step to achieve physically guided diffusion reconstruction.

[0070] The reconstruction control module can determine whether to stop the iteration based on the following conditions: the preset maximum number of iterations is reached; the change between two adjacent reconstruction results is less than a threshold; the data consistency residual is less than a threshold; the metabolic peak integral change tends to stabilize; or the output quality index meets the preset requirements.

[0071] 2.8 Result Output Module The results output module is connected to the reconstruction control module, data consistency projection module, and metabolic kinetics constraint module, and is used to output the final reconstruction results. This module can output denoised dynamic complex spectrum data, amplitude spectrum data, metabolic peak spectra at each time point, metabolic peak integral results, metabolite concentration-time curves, metabolic kinetic parameters, and metabolic parameter graphs.

[0072] exist Figure 1 In the diagram, this module corresponds to the "Output" spectrum on the right. Compared to the input spectrum, the output spectrum shows significantly reduced random noise, clearer peak structures, and more continuous peak changes over time, which is beneficial for subsequent metabolite identification, peak area integration, concentration quantification, and kinetic analysis.

[0073] 2.9 Storage Module The storage module stores model parameters, noise scheduling tables, acquisition parameters, preprocessing parameters, iteration logs, intermediate reconstruction results, and final output results required during device operation. This module may include local hard drives, databases, network storage units, or cloud storage units.

[0074] In one implementation, the storage module can record the input data identifier, noise estimation results, number of iterations, residuals at each step, output quality indicators, and model version information for each reconstruction task, so as to facilitate subsequent traceability, reproduction, and quality control.

[0075] 2.10 Computing Platform The computing platform provides computing resources for the aforementioned modules. This platform can be a GPU workstation, server, edge computing box, embedded computing board, or processing unit integrated into the magnetic resonance system. For deep diffusion model inference, the computing platform is preferably configured with a graphics processor, tensor computation unit, or other hardware supporting parallel computing. For real-time or near-real-time applications, the computing platform can also employ model compression, half-precision inference, batch parallel processing, or pipelined scheduling to improve operational efficiency.

[0076] 3. Key Reconstruction Algorithm Flow The key reconstruction algorithm flow of the device in this embodiment is as follows: Figure 3 As shown.

[0077] The algorithm first acquires dynamic complex spectral data from the magnetic resonance acquisition system to form raw observation data. This data is then preprocessed, including frequency band truncation, phase correction, amplitude normalization, real / imaginary part channel splitting, and tensor organization, to obtain a spectral tensor suitable for the diffusion model input. Next, the algorithm extracts noise samples from frequency bands without metabolic peaks or spectral edge noise regions, estimates the observed noise level, and initializes the reconstruction parameters, including the initial estimation results, the diffusion noise scheduling sequence, the number of iterations, and the stopping condition.

[0078] After initialization, the algorithm enters the core iterative reconstruction phase. The reconstruction control module first determines whether the current reconstruction result meets the stopping conditions, such as whether the maximum number of iterations has been reached, whether the change between two adjacent reconstruction results is less than a preset threshold, whether the data consistency residual has converged, or whether the metabolic peak integral result has stabilized. If the stopping conditions are not yet met, an alternating update process consisting of diffusion prior denoising, data consistency projection, and dynamic constraints is executed; if the stopping conditions are met, the iteration is terminated and the final reconstruction result is output.

[0079] In each iteration, the algorithm first performs a diffusion prior denoising step. Specifically, it performs denoising on the current estimated spectral data. Current diffused noise level The necessary noise condition encoding is input into the noise prediction network, which uses a diffusion model to predict noise components or the recovered clean spectral components, thus obtaining intermediate denoising results. This step utilizes prior knowledge of spectral structure and metabolic kinetics obtained through training, which can enhance the true metabolic peaks, suppress random noise, and maintain the continuous change characteristics of the dynamic spectrum in the time dimension.

[0080] Subsequently, the algorithm performs a data consistency projection step. This step denoises the intermediate results obtained from the diffusion prior. Compared with the original observation data The data is then combined and updated with constraints based on the physical observation model and observation noise level to obtain a new reconstruction estimate. This step avoids the diffusion model from producing overly smoothed results or spurious spectral peaks that deviate from the actual acquired data, ensuring that the reconstruction results after each iteration remain statistically consistent with the actual observation data. For undersampling, multi-channel acquisition, or spectral imaging scenarios, this step can also incorporate physical operators such as sampling masks, Fourier transforms, and coil sensitivity encoding to complete the projection update.

[0081] Furthermore, the algorithm introduces kinetic constraints during the iteration process, applying time smoothing or metabolic evolution consistency constraints to the updated dynamic spectral data. This step can be based on metabolic peak integral curves, concentration-time curves, or preset metabolic kinetic models to suppress abnormal time points, abrupt noise, and unreasonable temporal fluctuations, ensuring that the reconstructed results not only have clear spectral peaks in a single frame but also conform to the laws of metabolite uptake, generation, transformation, or clearance in the continuous time dimension.

[0082] The aforementioned diffusion prior denoising, data consistency projection, and dynamic constraint steps are executed cyclically under the scheduling of the reconstruction control module. After each iteration, the algorithm updates the current reconstruction result, noise level, and iteration count, and re-evaluates the stopping condition. As the noise level gradually changes from high to low, the reconstruction result gradually approximates the clean dynamic spectrum from coarse to fine.

[0083] 4. Noise estimation network like Figure 4 As shown, this embodiment presents a preferred network structure for the diffusion prior denoising module. This network is used to address noise at a given noise scale. Under these conditions, noise prediction or denoising reconstruction is performed on noisy dynamic complex spectrum data. The input data is a noisy spectral tensor. Its dimensions can be expressed as ,in The number of channels, "2", corresponds to the real and imaginary parts of the complex spectrum, respectively. The dimension "24" represents the number of dynamic time points or time frames, and the dimension "256" represents the number of frequency sampling points. The network output is prediction noise. Its size is consistent with the input, that is This enables point-by-point noise estimation for each time point, each frequency point, and each complex channel.

[0084] In one implementation, the diffusion prior denoising module employs a U-shaped encoder-decoder structure. An input convolutional layer is first set at the input end, and this input convolutional layer can be... The convolutional kernel maps the input two-channel complex spectral features to 64 feature channels. The network then sequentially enters multiple coding layers. Each coding layer may include downsampling units and multiple residual blocks to progressively reduce the temporal-frequency resolution of the feature maps and increase the number of channels. Figure 4 An example of a four-level coding structure is given, where the feature size of coding layer 1 is... The number of channels is 64; the feature size of the coding layer 2 after downsampling is... The number of channels is 128; the feature size of coding layer 3 is... The number of channels is 256; the feature size of coding layer 4 is... The number of channels is 512. Through the above step-by-step encoding, the network can simultaneously extract local spectral peak morphology features, cross-frequency structural features, and dynamic temporal evolution features.

[0085] The residual block in each coding layer may include a normalization layer, a SiLU activation function, and The network consists of convolutional layers, FiLM conditional modulation layers, Dropout layers, and residual connections. Normalization layers stabilize feature distributions across different batches and noise levels; the SiLU activation function enhances the network's nonlinear expressive power; convolutional layers extract time-frequency two-dimensional structural features; Dropout layers improve the network's generalization ability; and residual connections alleviate gradient decay during deep network training while preserving original spectral features. Through residual block stacking, the network can learn noise distribution and spectral structure priors without disrupting the true metabolic peak structure.

[0086] like Figure 4 As shown above, the network also includes a noise conditional coding branch. This branch receives the noise scale during the diffusion process. First, perform a logarithmic transformation on it to obtain This is done to compress the dynamic range of the noise scale and improve numerical stability at different noise levels. Subsequently, The input sinusoidal position encoding module generates high-dimensional periodic features that characterize the noise level; these features are then mapped to a conditional embedding vector via a multilayer perceptron (MLP). This conditional embedding vector is fed into the FiLM conditional modulation unit in each encoding layer, bottleneck layer, and decoding layer to generate the scaling and offset parameters for the corresponding layer, thereby achieving conditional modulation of the intermediate features.

[0087] In its implementation, FiLM conditional modulation performs an affine transformation on network features as follows: scaling and offset coefficients are generated based on the noise conditional embedding vector, and channel-by-channel modulation is applied to the convolutional features. This allows the network to adaptively adjust the denoising intensity at different noise scales. For example, at higher noise scales, the network can rely more on global spectral structure and dynamic priors to suppress strong random perturbations; at lower noise scales, the network can use a weaker denoising amplitude, focusing on preserving details such as weak metabolic peaks, peak widths, peak positions, and complex phases. This conditional modulation mechanism enables the same network to adapt to multiple noise levels in the diffusion reverse process.

[0088] A bottleneck layer is set after the lowest layer of the encoder. The bottleneck layer may include two residual blocks, with 512 channels and a feature size maintained at [value missing]. This layer, located at the deepest part of the network, is used to integrate global time-frequency context information and extract high-level spectral structure representations. Because dynamic magnetic resonance spectral data exhibits continuous changes in metabolite concentration over time and correlations in peak position and shape over frequency, the bottleneck layer can synthesize a large amount of metabolic spectral information within the receptive field, providing global priors for subsequent decoding and reconstruction.

[0089] The decoder is configured with multiple decoding layers corresponding to the encoder. Each decoding layer may include an upsampling unit, a skip-connection splicing unit, and multiple residual blocks. The upsampling unit is used to progressively recover the time-frequency resolution; the skip-connection splicing unit is used to receive the shallow features output from the corresponding coding layer and splice them with the currently decoded features. Figure 4 The example provides a four-level decoding structure, where decoding layer 4 upsamples the output of the bottleneck layer and concatenates it with the features of coding layer 4, with a channel count of 512; decoding layer 3 restores to... The number of channels is 256; decoding layer 2 is restored to The number of channels is 128; decoding layer 1 is restored to The network has 64 channels. Through symmetrical upsampling and skip connections, the network can simultaneously utilize deep global semantic features and shallow local spectral peak details, thereby improving the fidelity of the denoising results.

[0090] The skip connections not only transmit spatial or spectral details but also reduce the loss of weak peak information during deep coding. For low signal-to-noise ratio dynamic spectral data, some weak metabolite peaks may only appear in local frequency bands or some time frames. By directly transmitting features at the same scale at the encoding end to the decoding end, the network can reuse the peak position, peak width, phase, and local texture information extracted earlier in the recovery stage, reducing the risk of over-smoothing or peak distortion.

[0091] After the decoder outputs, the network maps the 64-channel features back to the two-channel complex spectral space through the output convolutional layer to obtain the predicted noise. The predicted noise can be used for noise subtraction, clean spectrum estimation, or subsequent data consistency projection during the diffusion reversal process. Since the output size is the same as the input size, this network can be directly embedded into a progressive denoising iterative framework at each noise level. The following section performs conditional noise prediction on the current noisy estimation.

[0092] In summary, the network shown in the figure achieves conditional diffusion denoising for dynamic complex spectral data through a structural design of "noise scale encoding—FiLM conditional modulation—U-shaped multi-scale feature extraction—jump connection detail recovery." This structure can both utilize deep networks to learn statistical priors of metabolic spectrum data and adapt to noise levels at different diffusion stages through noise conditional embedding. Simultaneously, the encoder-decoder and jump connection mechanisms help maintain true spectral peaks, complex phases, and temporal dynamic information. Therefore, this diffusion prior denoising module can serve as the core noise prediction network in the physically guided reconstruction algorithm of this embodiment, improving the signal-to-noise ratio, peak fidelity, and reliability of metabolic kinetic parameter estimation in dynamic magnetic resonance spectral reconstruction.

[0093] 5. Result Comparison like Figure 5 As shown in the figure, this embodiment compares the original results of dynamic metabolic imaging data with the denoised results processed by the method of this application. The left side of the figure shows the dynamic changes of various metabolite signals at different time points, including water signals, glucose signals, glutamate / glutamine signals, and lactate signals. In the original data, due to the short dynamic acquisition time and low metabolite concentration, there is obvious random noise in the image, the spatial distribution boundaries of some weak metabolic peaks are unclear, and a certain degree of discontinuity and speckled artifacts can also be observed in the time series. After denoising by the method of this application, the distribution of metabolites at each time point is smoother and more continuous, the background noise is significantly reduced, and the main metabolic regions and the trend of change over time are still maintained.

[0094] Furthermore, Figure 5 The upper right section presents a comparison of the spectra and metabolite time-varying curves. The original data's spectral curves exhibited significant fluctuations and a high noise floor, making weak peaks easily masked by random perturbations, leading to instability in subsequent peak area integrals and concentration estimates. After denoising, the spectral curves became smoother, the peak outlines clearer, and the peak positions and shapes did not shift significantly, indicating that the proposed method can preserve the true metabolic peak characteristics while suppressing noise. The corresponding time curves also transformed from the violent fluctuations in the original data to a more continuous trend consistent with metabolic evolution, which is beneficial for improving the reliability of kinetic parameter estimation.

[0095] Figure 5The lower right section shows the processing effect in scenarios where structural images or spectra are overlaid. It can be seen that the spectral lines in the original data are heavily affected by noise, resulting in unstable local signal display. After denoising, the spectral lines are more regularly arranged, and the signal representation within the target area is clearer. Therefore, the reconstruction method proposed in this application, based on diffusion prior and data consistency constraints, can improve the signal-to-noise ratio of dynamic metabolic imaging while minimizing over-smoothing and spurious structure generation, thereby improving the stability and reliability of metabolite spatial distribution display, spectral analysis, and dynamic quantitative analysis.

[0096] 6. Process Flow and Steps This embodiment provides a dynamic complex spectrum denoising and reconstruction method based on a physically guided diffusion prior. This method can be executed by the aforementioned apparatus or by a computing device with a processor and memory. The method includes at least the following steps.

[0097] S1: Collect dynamic complex spectrum data.

[0098] The target object is dynamically acquired using a magnetic resonance acquisition module to obtain raw observation data. The observation data Complex spectral data can include multiple time points, multiple spatial locations, multiple receiving channels, and multiple frequency sampling points, used to characterize the signal of different metabolite spectral peaks of the target object changing over time during dynamic processes.

[0099] S2: Preprocess the observation data and construct the network input tensor.

[0100] Regarding the observation data The process involves frequency band truncation, phase correction, frequency drift correction, amplitude normalization, baseline correction, and channel merging or rearrangement. Subsequently, the complex spectrum data is split into two channels, real and imaginary, and rearranged. The input tensor is of the form, where, Indicates batch size. Indicates the number of dynamic time points. This represents the number of complex channels, preferably 2. This indicates the number of frequency sampling points.

[0101] S3: Estimate the observation noise level.

[0102] Noise samples are extracted from preset frequency bands without metabolic peaks, spectral edge bands, or noise reference bands, and the standard deviation or variance of the observed noise is calculated to obtain the observed noise level. The observed noise level is used to determine the subsequent diffusion denoising intensity, data consistency projection weight, and iteration stopping condition.

[0103] S4: Set the parameters for spreading noise scheduling and iteration.

[0104] Based on the observed noise level Collect signal-to-noise ratio, target spectrum data type, or preset empirical parameters, and set the diffusion noise scheduling sequence. and the maximum number of iterations The noise scheduling sequence can gradually transition from a high noise level to a low noise level, or it can adopt a linear, exponential, cosine, or adaptive scheduling method.

[0105] S5: Initialize the spectral data to be reconstructed.

[0106] raw observation data The preprocessed observation data, or the results after smoothing filtering, low-pass filtering, and preliminary fitting, are used as the initial reconstruction results. In a preferred embodiment, let To maintain consistency between the initial estimate and the collected data; in another implementation, let for A slightly smoothed version to reduce the impact of initial random noise on the diffusion reverse process.

[0107] S6: Perform physical-guided diffusion iterative reconstruction.

[0108] for The process iteratively performs prior denoising, data consistency projection, and optional dynamic consistency constraints, specifically including: S6-1: Diffusion prior denoising step.

[0109] Current reconstruction results and current noise level Input diffusion prior denoising module. The diffusion prior denoising module generates a noise condition encoder that is consistent with... The corresponding conditional embedding vectors are used, and a noise prediction network is employed to predict noise components in the current spectral data or estimate clean spectral components, resulting in intermediate denoising results. This step is used to introduce priors on metabolic spectrum structure, complex spectrum phase, and dynamic temporal evolution learned from the training data.

[0110] S6-2: Data Consistency Projection Steps.

[0111] intermediate denoising results Raw observation data and the observed noise level The input consistency projection module, based on the physical observation model, Perform likelihood constraints or closed-loop fusion updates to obtain the next reconstruction result. The physical observation model can be represented as follows: ,in For the observation operator corresponding to the acquisition system, This is the noise term. This step is used to ensure that the reconstruction results are consistent with the original observation data, avoiding false spectral peaks, over-smoothing, or results that deviate from the actual acquired data during the diffusion prior denoising process.

[0112] S6-3: Dynamic consistency constraint steps.

[0113] Apply metabolic kinetic consistency checks or light corrections to x. Specifically, smoothing constraints, outlier suppression, monotonicity constraints, delayed response constraints, or kinetic model consistency constraints can be applied to the metabolic peak integral curves, spectral peak fitting curves, or concentration-time curves at different time points. This step can be performed as an optional step, or only when the residuals are abnormal or the time curve fluctuations exceed a threshold, to improve the temporal continuity and metabolic parameter stability of the dynamic spectral reconstruction results.

[0114] During the loop, the iteration can be terminated early based on the current noise level, the difference between two adjacent reconstruction results, the data consistency residual, the change in spectral peak integral, or a preset threshold. If the stopping condition is met, the loop stops and proceeds to step S7; otherwise, the iteration is updated. And continue to execute the next iteration.

[0115] S7: Output denoised dynamic spectrum and perform metabolic analysis.

[0116] After the iteration is complete, the final denoised dynamic spectrum x is output. Based on the final denoised dynamic spectrum, peak position identification, peak shape fitting, peak area integration, or concentration conversion are performed on the target metabolite peaks to obtain the concentration-time curves of each metabolite. Furthermore, metabolic kinetics can be fitted based on the concentration-time curves to generate metabolic rate, conversion rate, clearance rate, or other metabolic parameter graphs.

[0117] S8: Save reconstruction results and process information.

[0118] The final denoised dynamic spectrum, metabolic peak integral results, concentration-time curves, metabolic parameter plots, noise estimation results, noise scheduling parameters, number of iterations, data consistency residuals, and quality evaluation indicators are saved to the storage module for subsequent display, analysis, verification, report generation, or model performance traceability.

[0119] Through the above steps, the method in this embodiment further introduces observation data consistency constraints and metabolic kinetic consistency constraints on the basis of prior denoising of diffusion model. This enables the reconstruction results to effectively reduce random noise in dynamic complex spectrum data, while maintaining the true peak structure, complex phase information and metabolite temporal evolution law, thereby improving the accuracy and stability of dynamic magnetic resonance spectrum reconstruction and metabolic parameter estimation.

[0120] The key technical points of this embodiment are as follows: 1. Generative diffusion / score-based prior for dynamic²H spectral time-to-time data reconstruction.

[0121] To address the characteristics of dynamic complex spectra changing over time, diffusion models or score-based models are used to learn the peak morphology, complex phase, and temporal evolution distribution of the spectrum. This is applied to the denoising and reconstruction of low signal-to-noise ratio dynamic²H spectral data, reducing the reliance on paired "noisy-clean true" training data.

[0122] 2. An alternating iterative framework of "prior denoising + data consistency projection" is adopted.

[0123] At each noise scale Next, the diffusion prior denoising module first modifies the current estimate. Update to intermediate results Combined with the original observations Perform data consistency projection to obtain This ensures consistency between prior learning and physical observation.

[0124] 3. Based on and An adaptive fusion mechanism.

[0125] Based on the current scale of diffused noise and observation noise level The weights of observations and priors are automatically determined to achieve an adaptive balance between fidelity and denoising intensity at different iteration stages.

[0126] 4. Automatically estimate the observation noise level.

[0127] Noise samples are extracted from frequency bands without metabolic peaks or spectral edge noise bands to estimate observed noise. It is used for consistent projection, weight setting and iterative control, reducing manual parameter tuning and improving cross-device and cross-individual adaptability.

[0128] 5. Introduce time constraints for metabolic kinetics.

[0129] During training or inference, time smoothing, differential penalty, or kinetic consistency constraints are applied to the spectral signals or metabolic peak integral curves at adjacent time points to suppress non-physiological jitters and maintain the continuity of metabolic changes.

[0130] 6. Network structure for complex spectrum-time data.

[0131] The complex spectrum is split into real and imaginary channels, and a spectrum-time joint U-shaped network is used, which combines noise conditional embedding, skip connections and optional time self-attention modules to model the peak shape and time dependence.

[0132] 7. Output serves quantitative metabolic analysis.

[0133] The final output is high-confidence dynamic spectrum data after denoising, and further supports the calculation of metabolic peak integrals, concentration-time curves and estimation of metabolic kinetic parameters, thereby improving the stability and reliability of downstream quantitative analysis.

[0134] Compared with the existing technical approach of "generative diffusion / score-based prior + data consistency for medical imaging inverse problem" (i.e., the scheme with diffusion prior recovery as the core and then adding physical consistency constraints), the advantages of this embodiment are mainly reflected in the following aspects (all explained from the perspective of the technical solution): 1. Better adapted to dynamic²H spectral-time data format, reducing the risk of mismatch caused by "method migration". Existing diffusion reconstruction is multi-directional to 2D / 3D image domains; this embodiment organizes the input as a complex spectrum with "real / virtual dual channels + time dimension", and adopts a spectrum-time joint network structure to simultaneously model the spectral peak shape and time correlation. Therefore, it has a more direct structural matching on the target data type (dynamic²H spectrum / spectral imaging), and theoretically it is less likely to encounter peak shape distortion or time inconsistency problems caused by "hard-fitting" the image domain prior to the spectral domain.

[0135] 2. Data consistency constraints are more explicit and controllable, which helps to reduce the bias introduced by prior generation. This embodiment performs alternating updates of "prior denoising → uniform projection" at each iteration noise scale, and employs fusion / projection weights determined by the noise scale and observation noise, ensuring that the output is always constrained by the observation y. Therefore, it can be deduced that compared to methods relying solely on prior sampling or weak uniformity constraints, this embodiment more easily controls the offset and reduces the risk of pseudo-structures that do not conform to the physical reality of the measurement.

[0136] 3. Improve adaptability across scan conditions through automatic noise estimation and reduce manual parameter tuning. This embodiment automatically estimates noise from the frequency band without metabolic peaks in the spectrum and uses it for consistency projection and weight setting. Since different coils, test subjects, and acquisition protocols can cause differences in noise levels, automatic estimation can enable the algorithm to maintain a relatively consistent "denoising intensity-fidelity" balance under different conditions, thereby reducing dependence on empirical parameters.

[0137] 4. Introducing a priori metabolic kinetics helps improve the stability of the dynamic curve. In dynamic²H applications, the ultimate focus is often on the metabolite concentration-time curve and its kinetic parameters. Existing diffusion inverse problem methods typically constrain spatial or image structures. This embodiment incorporates the principle of "smooth evolution of metabolism over time" into the training or inference process through methods such as temporal difference penalties. Therefore, it can be deduced that, under the same observation noise, this embodiment is more likely to suppress non-physiological temporal jitter, resulting in a more stable fit between downstream metabolic quantification and kinetics.

[0138] Other alternative embodiments 1. Alternative solutions for data collection and input formats The acquisition module in this embodiment is not limited to three-dimensional CSI sequences, but can also use 2D or 3D MRSI, FID-CSI, EPSI, spiral-MRSI, UTE-MRSI or other magnetic resonance acquisition sequences that can obtain spectral information.

[0139] This embodiment is not limited to deuterium dynamic spectrum data, but can also be extended to dynamic magnetic resonance spectra or magnetic resonance imaging data of nuclei such as 1H, 13C, and 31P. As long as the acquired data can be represented as noisy observation spectrum data, or can meet similar physical observation models, it can be applied to the denoising and reconstruction framework of this embodiment.

[0140] For input data formats, complex spectrum data can be split into two channels: real and imaginary parts. Alternatively, it can be represented using amplitude / phase, complex convolution, or other equivalent tensor organization methods. Data dimensions can be expanded to include time, space, channels, and frequency, depending on the actual acquisition method.

[0141] 2. Alternatives to the noise estimation module The noise sample sources in the noise estimation module are not limited to the peakless frequency band at the edge of the spectrum. They can also be selected from the frequency band outside the preset water peak or metabolic peak, the peakless region automatically determined by the peak detection algorithm, or the spatially signalless voxel region as the noise estimation region.

[0142] Observation of noise level The estimation methods are not limited to standard deviation calculation; statistical methods such as root mean square value, variance, median absolute deviation, quantile estimation, and noise floor model fitting can also be used. Furthermore, It can adaptively estimate based on time point, frequency band, spatial voxel, or a combination thereof.

[0143] In another implementation, explicit calculation may not be required. The noise level can be obtained by predicting the noise map or curve using a neural network auxiliary branch, or by using maximum likelihood, EM algorithm, or other online estimation methods during iterative optimization. All of these methods are equivalent implementations of the noise level-based consistency constraint adjustment.

[0144] 3. Alternatives to diffusion or generation of prior modules The diffusion prior in this embodiment is not limited to a specific variance-explosive stochastic differential equation (VE-SDE) framework, but can also employ variance-preserving stochastic differential equation (VP-SDE), denoised diffusion probability model (DDPM), denoised diffusion implicit model (DDIM), consistency model, rectified flow model, flow matching model, or other generative models that can provide noise-scale denoising direction, score function, or velocity field.

[0145] Sampling or update strategies during the inference phase are not limited to prediction-correction sampling. They can be implemented using methods such as Euler-Maruyama, Heun method, ordinary differential equation (ODE) sampling, DDIM deterministic sampling, Langevin dynamics, multi-step fast sampling, distillation accelerated sampling, or few-step sampling.

[0146] The structure of noise prediction networks is not limited to U-shaped convolutional networks; one-dimensional spectral convolutional networks, temporal convolutional networks, transformer networks based on self-attention mechanisms, convolutional long short-term memory networks (ConvLSTM), U-Net, and other structures can also be used. Normalization layers in the network can employ group normalization, layer normalization, batch normalization, or root mean square normalization (RMSNorm); activation functions can include sigmoid linear units (SiLU), rectified linear units (ReLU), and Gaussian error linear units (GELU); upsampling and downsampling can be achieved using interpolated convolution, deconvolution, strided convolution, and other methods.

[0147] The conditional encoding of the noise scale is not limited to Gaussian Fourier features or sinusoidal positional encoding; it can be replaced by learnable embedding tables, MLP encoding, lookup table encoding, or the noise scale can be directly input into the network as an additional channel. Conditional information can be injected into each residual block, or only into the encoder, decoder, or bottleneck layer.

[0148] The training objective is not limited to denoised score matching. It can also use noise prediction loss, x prediction loss, hybrid loss, perceptual consistency loss, or further add spectral peak fidelity constraints such as peak position, peak area, peak width, and phase consistency.

[0149] 4. Alternatives to the data consistency projection module The observation model in the data consistency projection module is not limited to the simple additive noise model y=x+σ, but can also use a more general physical observation model: .

[0150] Where A can represent a sampling operator, Fourier transform, subsampling mask, point spread function, coil sensitivity coding, spatial convolution operator, or a combination thereof.

[0151] Consistent update methods are not limited to closed-loop weighted fusion; they can also employ proximal gradient updates, conjugate gradients, LSQR solutions, constraint set projection, ADMM, plug-and-play, HQS semi-quadratic splitting, or other equivalent optimization methods. For example, optimization problems of the following form can be solved: Where y represents the observed data, A is the physical observation operator, z is the intermediate estimate obtained by diffusion prior denoising, β is the prior constraint weight, and x is the denoised dynamic spectrum data to be reconstructed.

[0152] The fusion weights or constraint strengths in consistent projection are not limited to analytical calculations of σ, but can be determined by empirical scheduling tables, learning-based gating networks, residual sizes, adaptive rules or frequency bands, and voxel weights. This allows for adaptive data consistency constraints across spatial, frequency, and temporal dimensions.

[0153] 5. Alternative solutions for output and downstream quantitative analysis The output of this embodiment is not limited to denoised dynamic spectral data, but may also include metabolic peak integral results, peak position, peak width, peak area, concentration-time curves, metabolic kinetic parameter plots, and quality evaluation indicators. These quality evaluation indicators may include residuals, signal-to-noise ratio, confidence intervals, goodness of fit, or abnormal time point indications.

[0154] Downstream quantification methods are not limited to a single peak fitting method; they can include frequency domain peak fitting, time domain fitting, Lorentz peak fitting, Gaussian peak fitting, Voigt peak fitting, basis function fitting, or machine learning regression methods. The above outputs and quantification methods are all application forms of the reconstruction results in this embodiment and do not affect the reconstruction framework of this embodiment, which is based on alternating updates of diffusion prior denoising and data consistency projection.

[0155] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0156] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0157] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0158] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0159] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0160] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0161] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0162] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method of reconstruction of dynamic metabolic imaging signals, characterized in that, include: Raw observation data is obtained and preprocessed, the raw observation data being obtained from dynamic metabolic imaging (DMI) applied to the target; Estimating observation noise level from pre-processed observation data ;​ Furthermore, a K-round iterative reconstruction process is performed on the preprocessed observation data, k=1,2……K, until convergence, finally obtaining the reconstruction result: In each iteration, firstly, at the corresponding noise scale... Based on the constructed diffusion prior denoising model, the current estimate is then... Update to intermediate denoising results Then, the intermediate denoising results With observation data By combining data consistency projection, the next estimate is obtained. Noise scale The parameters for iterative reconstruction are determined by the observed noise level. Sure.

2. The method of claim 1, wherein the dynamic metabolic imaging signal is reconstructed by: The diffusion prior denoising model includes a noise conditional encoder and a noise prediction network, wherein the noise conditional encoder is used to convert the noise scale... The encoding is a conditional embedding vector, which is then injected into the corresponding layer of the noise prediction network.

3. The method of claim 2, wherein the dynamic metabolic imaging signal is reconstructed by: The noise conditional encoder includes, in sequence, a logarithmic transformation module, a sinusoidal position encoding module, and a multilayer perceptron (MLP). The logarithmic transformation module is used to perform a logarithmic transformation on the noise scale. The sinusoidal position encoding module is used to extract high-dimensional periodic features from the logarithmic values. The MLP is used to map the high-dimensional periodic features into conditional embedding vectors.

4. The method of claim 2, wherein the dynamic metabolic imaging signal is reconstructed by a method comprising: The noise prediction network adopts a U-shaped encoder-decoder structure, which includes a series of cascaded input convolutional layers, multi-level coding layers, bottleneck layers, and multi-level decoding layers. The multi-level coding layers and multi-level decoding layers are connected in a skip connection. The multi-level coding layers are used to extract local spectral peak morphology features, cross-frequency structure features, and dynamic time evolution features. The bottleneck layer is used to integrate global time-frequency context information and extract high-level spectral structure representations. The multi-level decoding layers are set up correspondingly to the multi-level coding layers. ​ 5. The method of claim 4, wherein the dynamic metabolic imaging signal is reconstructed by a method comprising: The conditional embedding vector is injected into the FiLM conditional modulation unit in the coding layer, bottleneck layer and decoding layer. The FiLM conditional modulation unit is used to generate the scaling parameters and offset parameters of the corresponding layer according to the conditional embedding vector, thereby realizing adaptive adjustment of the denoising intensity. ​ 6. The method of claim 4, wherein the dynamic metabolic imaging signal is reconstructed by a method comprising: The multi-level coding layer includes the following cascaded layers: first coding layer, second coding layer, third coding layer and fourth coding layer. The first coding layer includes several residual blocks. The second coding layer, third coding layer and fourth coding layer each include downsampling and several residual blocks. The bottleneck layer includes several residual blocks. The multi-level decoding layer includes the following cascaded layers: fourth decoding layer, third decoding layer, second decoding layer and first decoding layer. The fourth decoding layer, third decoding layer, second decoding layer and first decoding layer each include skip connection splicing and several residual blocks.

7. The method for reconstructing dynamic metabolic imaging signals as described in claim 6, characterized in that, The residual block comprises, in sequence, a group normalization layer, a SiLU activation function, a 3×3 convolutional layer, a FiLM conditional modulation layer, a Dropout layer, and residual connections.

8. The method of claim 1, wherein the dynamic metabolic imaging signals are reconstructed by a method comprising: Noise scale Corresponding to a diffused noise sequence: ,in, Indicates an initial high noise level. This indicates the termination of the low noise level, where K represents the number of iterations, and the intermediate denoising results are... With observation data By combining data consistency projection, the next estimate is obtained. Specifically, it includes: ​ According to the strength of the noise scale, the intermediate denoising result is adaptively adjusted The fusion weight of the observation data ​ and, with the intermediate denoised result fusion of the observation data as the next estimate .

9. The method for reconstructing dynamic metabolic imaging signals as described in claim 1, characterized in that, The method for reconstructing dynamic metabolic imaging signals also includes: Apply smoothness, monotonicity, peak timing, delayed response, or dynamic model consistency constraints to the reconstruction results.

10. An electronic device, comprising: The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the method for reconstructing dynamic metabolic imaging signals as described in any one of claims 1-9.

Images