A dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization

Abstract

The present application belongs to the field of medical image processing and biomedical engineering, and particularly relates to a dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization. The steps are as follows: obtaining original image data, performing multi-scale data correction preprocessing on the original data to generate preprocessed image data; extracting a large blood vessel region based on the preprocessed image data as a candidate region of a contrast agent input function; constructing a heterogeneous nonlinear pharmacokinetic model; establishing a cooperative reconstruction mechanism of the heterogeneous nonlinear pharmacokinetic model and the preprocessed image data; integrating the kinetic constraint condition of the heterogeneous nonlinear pharmacokinetic model into the objective function of the preprocessed image data reconstruction as a loss term, and further iteratively updating the intermediate parameters and the initial reconstructed image. The method of the present application improves the accuracy of parameter estimation in the tumor lesion region through the adaptive heterogeneous nonlinear pharmacokinetic model.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical image processing and biomedical engineering, specifically involving a dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization. Background Technology

[0002] Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a core imaging technique for precise tumor diagnosis and efficacy evaluation. It analyzes the dynamic distribution of contrast agents within tissues and combines this with pharmacokinetic models to quantitatively calculate tissue physiological and pathological parameters, providing objective evidence for clinical diagnosis. Currently, pharmacokinetic-based DCE-MRI reconstruction is the mainstream technique in this field. The Tofts model is the most classic first-order pharmacokinetic model, and improved models such as the ExtendedTofts model and the Two-compartment model have also been derived.

[0003] Existing pharmacokinetic models for DCE-MRI reconstruction have clear physiological significance and clinical acceptance, but they are prone to problems such as model assumptions not matching actual physiological conditions, large estimation errors of contrast agent input functions, and poor model adaptability to low temporal resolution data. As a result, existing pharmacokinetic-based DCE-MRI reconstruction methods suffer from low accuracy of quantitative parameters, poor inter-individual repeatability, and insufficient lesion identification accuracy in clinical applications, which limits their application value in precision tumor diagnosis. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, the present invention aims to provide a dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization, in order to solve the problems of low accuracy of quantitative parameters, poor inter-individual repeatability, and insufficient lesion identification accuracy in clinical applications.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention discloses a dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization, the method comprising: S1: Acquire raw image data, and perform multi-scale data correction preprocessing on the raw data to generate preprocessed image data; S2: Extract large blood vessel regions based on the preprocessed image data as candidate regions for the contrast agent input function; correct the signal-time curves of the candidate regions and output the final individualized contrast agent input function; S3: Based on the final individualized contrast agent input function, and by introducing tissue heterogeneity weighting factors, nonlinear transport correction coefficients and intermediate parameters, a heterogeneous nonlinear pharmacokinetic model is constructed. S4: Establish a collaborative reconstruction mechanism between the heterogeneous nonlinear pharmacokinetic model and preprocessed image data; the collaborative reconstruction mechanism is to incorporate the kinetic constraints of the heterogeneous nonlinear pharmacokinetic model as a loss term into the objective function of the preprocessed image data reconstruction to generate an initial reconstructed image, and further iteratively update the intermediate parameters and the initial reconstructed image to finally generate a highly accurate reconstructed image and a distribution map of the optimal intermediate parameters.

[0006] Further, the multi-scale data correction preprocessing in S1 includes: wavelet decomposition of the original data to generate a first high-frequency noise component and a low-frequency structural component; applying an adaptive thresholding method to suppress random noise in the first high-frequency noise component to generate a second high-frequency noise component; combining the second high-frequency noise component and the low-frequency structural component, reconstructing the original data using a reconstruction function to generate a reconstructed image; correcting the reconstructed image using a rigid registration function to generate a corrected image; and applying the Gibbs artifact iterative correction method to the corrected image for deconvolution correction.

[0007] Further, the correction process in S2 includes: averaging the signal time curve of the candidate region to obtain a first signal curve; performing nonlinear smoothing fitting based on the first signal curve to obtain a second signal curve; introducing individual physiological indicators to correct the second signal curve to generate an initial individualized contrast agent input function; constructing a reliability assessment model to score the peak concentration, rise time, and fall rate characteristics in the smooth curve of the initial individualized contrast agent input function, and eliminating distorted signals from abnormal perfusion regions in the initial individualized contrast agent input function.

[0008] Furthermore, in step S3, each pixel in the preprocessed image data is processed by a region segmentation function to generate a tissue region label; based on the observed baseline acquisition rate and baseline clearing rate of the tissue region label, an adaptive modulation rule is introduced to generate intermediate parameters, including the adaptive acquisition rate and the adaptive clearing rate.

[0009] Furthermore, the heterogeneous nonlinear pharmacokinetic model in S3 is shown below: in, Contrast agent concentration within the tissue; This is the plasma volume fraction. These are the position coordinates of the pixel; This refers to the concentration of contrast agent in the plasma. For the organization's adaptive uptake rate; For the region heterogeneity matrix; For the tissue's adaptive clearance rate; This is the nonlinear transport correction coefficient; As a weighting factor for organizational heterogeneity; For convolution integral, For integration variables; It is a time variable.

[0010] Furthermore, the objective function in S4 is as follows: in, Total loss; For data consistency items, To preprocess image data, This is the initial reconstructed image; and These are the weighting coefficients; Loss due to dynamic consistency constraints; For spatial structure loss; The initial reconstructed image is iteratively updated using a first iterative function; the intermediate parameters are iteratively updated using a second iterative function; a convergence judgment function is constructed and a preset judgment threshold is introduced to determine the termination of the iterative update process.

[0011] Furthermore, the generation of the initial reconstructed image in S4 is achieved using a conditional diffusion generation model, specifically including: After embedding image blocks into the T1-weighted image of the preprocessed image data output by S1, spatial position coding information is generated by a hybrid position encoder and injected into the feature sequence. The hybrid position encoder is composed of convolutional position coding and frequency domain position coding added and fused element by element. The feature sequence after hybrid positional encoding injection and the noisy reference map features are fed together into the pyramid Transformer conditional encoder to extract hierarchical multi-scale conditional features from high resolution to low resolution through multiple scale stages. The multi-scale conditional features are distributed to K parallel phase decoders. Each phase decoder generates an enhanced image of the corresponding phase through FiLM temporal modulation, upsampling, dense residual blocks, self-attention, and convolutional layers. The K phase decoders simultaneously output complete multi-phase DCE-MRI sequences. The denoising diffusion loss and the dynamic consistency constraint loss in S4 are jointly calculated for the output multi-phase DCE-MRI sequences, and the model parameters are updated by backpropagation until convergence.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: The method of this invention constructs a heterogeneous nonlinear pharmacokinetic model adapted to the tissue, breaking through the idealized assumptions of traditional Tofts and other models based on tissue homogeneity and linear contrast agent transport. It introduces a tissue heterogeneity weighting factor and a nonlinear transport correction coefficient, breaking the homogeneity assumption of traditional models and improving the accuracy of parameter estimation in the tumor lesion area. It also develops an automatic candidate region extraction algorithm for contrast agent input function (AIF) based on vascular semantic segmentation, which integrates individual patient physiological indicators and imaging features to complete the nonlinear fitting and correction of AIF. At the same time, it establishes an AIF credibility assessment and abnormal signal removal mechanism to completely eliminate the individual physiological difference bias of traditional population average AIF and the operator's subjective error caused by manual selection of AIF, providing accurate core input for fitting heterogeneous nonlinear pharmacokinetic models. The method of this invention incorporates pharmacokinetic constraints into the objective function of image reconstruction, allowing the model to guide the dynamic transport of contrast agents in image reconstruction. Simultaneously, high-quality image reconstruction data is fed back to the model fitting stage, achieving precise bidirectional matching between image sequences and kinetic features, and suppressing the superposition of bidirectional errors. A multi-dimensional joint correction scheme is constructed, incorporating wavelet decomposition noise suppression, rigid registration motion correction, and iterative deconvolution artifact elimination. This reduces the impact of various interferences on model fitting and image reconstruction from the data preprocessing stage, significantly improving the stability and repeatability of reconstruction results and parameter estimation. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the specific technical solution of the method of the present invention; Figure 2 This is a diagram of the diffusion forward noise addition process of the method of the present invention; Figure 3 This is a schematic diagram of the framework for a conditional diffusion medical image generation method based on pharmacokinetic constraints. Figure 4 This is a schematic diagram of the parallel phase decoder and hybrid position encoder structure; Figure 5 This is a schematic diagram of hybrid positional encoding. Detailed Implementation

[0015] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0016] Reference Figure 1 As shown, the present invention provides a dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization, comprising the following steps: Raw DCE-MRI images of the kidneys of a healthy volunteer were acquired using a 3.0T MRI scanner with a fast phase-scrambled gradient echo sequence. The MRI scanner parameters were as follows: TR / TE=4.5 / 1.8ms, flip angle 15°, field of view 380mm×380mm, matrix 192×192, slice thickness 5mm, temporal resolution 2.5 seconds, and a total of 60 time points were acquired.

[0017] S1: Acquire raw image data, and perform multi-scale data correction preprocessing on the raw data to generate preprocessed image data; First, the raw clinical imaging data is acquired and processed by wavelet decomposition. Taking a raw image data size of 192×192 pixels as an example, the Daubechies-4 wavelet basis is used with a decomposition level of 1, which helps to preserve key structures such as renal parenchyma and renal pelvis, while separating random noise caused by respiration, blood flow, etc., generating the first high-frequency noise component and the low-frequency structural component. The wavelet decomposition function is shown below: in, This is the first high-frequency noise component; Low-frequency structural components; For the first The original image data at that moment, It is a time variable; For wavelet transform operators; Random noise is suppressed by an adaptive thresholding method on the first high-frequency noise component to generate a second high-frequency noise component. This enhances the clarity of the boundary between the renal cortex and medulla and avoids deviations in perfusion parameter estimation due to noise interference. The function is shown below: in, This is the second high-frequency noise component; For the first Adaptive threshold at any given time; For the adjustment coefficient, candidate adjustment coefficients are selected within a certain range on several representative images. The peak signal-to-noise ratio (PSNR) affected by each candidate adjustment coefficient is calculated. The adjustment coefficient that best balances denoising effect and detail preservation is selected. For example... ; This is the median absolute deviation function; A symbolic function used to return the sign of the input value; By combining the second high-frequency noise component and the low-frequency structure component, the original image data is reconstructed using a reconstruction function to generate a reconstructed image. The reconstruction function is shown below: in, For the first Reconstructed image at any given moment; This is the inverse wavelet transform; The peak signal-to-noise ratio of the reconstructed image was improved by an average of 6.2 dB; Secondly, the reconstructed image is corrected using a rigid registration function to eliminate motion artifacts and generate a corrected image. The rigid registration function is shown below: in, For the first Registration parameters at any given time; For the first Reconstructed image at any given moment; for The reconstructed image at each time step undergoes spatial transformation according to the registration parameters; The square of the L2 norm is used to measure image differences; For the first Corrected image at time; Finally, the inherent artifacts in the corrected image are corrected by deconvolution using the Gibbs artifact iterative correction method, generating preprocessed image data that can eliminate ringing artifacts caused by truncation in image reconstruction. The function is shown below: in, To preprocess image data; It is a frequency domain filtering operator; For the frequency domain compensation function, for example, using the Tukey window, the cutoff frequency is set to 0.7 times the Nyquist frequency; S2: Extract large blood vessel regions based on the preprocessed image data as candidate regions for the contrast agent input function; correct the signal-time curves of the candidate regions and output the final individualized contrast agent input function; First, a vascular semantic segmentation model is constructed to extract large vessel regions from preprocessed image data, generating candidate regions for the contrast agent input function (AIF). The U-Net model is then used to segment the large vessel regions, including the renal artery trunk and its first-order branches, which are segmented into approximately 1560 pixels. This model can identify the renal artery trunk and its first-order branches, reducing the inclusion of other tissue signals. The vascular semantic segmentation model is shown below: in, This represents the candidate region of AIF, which is represented as a smooth curve. For blood vessel segmentation operators, the U-Net architecture is typically used; To preprocess image data; The first signal curve is obtained by averaging the signal time curves based on the candidate regions. The averaging function is shown below: in, This is the first signal curve; Total number of pixels; These are the position coordinates of the pixel; For the first Real-time preprocessing of image data at the pixel level The grayscale value at that location; Secondly, a nonlinear smoothing fit is performed on the first signal curve to obtain the second signal curve. The smoothing fit function is shown below: in, This is the second signal curve; For example, the number of exponent terms. ; For the first The magnitude coefficients of each exponential term, for example... , ; For the first The decay coefficient of each exponential term, for example... , ; The amplitude coefficient and decay coefficient are obtained by extracting average concentration curves from multiple healthy volunteers and fitting them with an exponential model to minimize the fitting error, thus obtaining population-representative amplitude coefficients and decay coefficients. The second signal curve is corrected using individual physiological indicators to generate an initial individualized AIF. The initial individualized AIF is represented by a smooth curve to address physiological differences among patients and improve the clinical comparability of pharmacokinetic parameters. The correction function is shown below: in, for The initial individualized AIF at time; and To determine the physiological regulatory weights, a regression model was established by collecting data on blood pressure (BP), heart rate (HR), and the corresponding DCE signal enhancement amplitude from multiple groups of individuals, thereby identifying the physiological regulatory weights. For example, , ; For an individual's blood pressure measurement, for example, ; For an individual's heart rate measurement, for example, ; Substituting the data into the formula, we obtain the initial individualized AIF: Finally, a reliability assessment model is constructed to score the peak concentration, rise time, and fall rate characteristics in the smooth curve of the initial individualized AIF. This model removes distorted signals from abnormal perfusion regions in the initial individualized AIF, eliminating abnormal AIF caused by motion and artifacts, ensuring stable and reliable perfusion measurement, and outputting the final high-precision individualized AIF. The reliability assessment model is shown below: in, For the overall score; , and The weighting coefficients for each evaluation indicator are determined by having radiology experts label a large number of perfusion curves as "qualified" or "unqualified." A logistic regression classifier is then trained, and the learned feature weights represent the contribution of each indicator. For example, , and ; For example, the maximum peak concentration. ; For the rise time, for example, ; For the rate of decrease, for example, ; For the final individualized AIF; To determine the final threshold, we observe the visual quality of the reconstruction results and the coefficient of variation of parameter estimates at different thresholds, using a preset quality threshold as an example. ; An empty set represents signals that do not meet quality standards in the initial individualized AIF; Substitute the data into the formula to obtain the overall score: This indicates that all time points are qualified, and all initial individualized AIFs are retained and used as the final high-precision individualized AIF; S3: Based on the final individualized contrast agent input function, and by introducing tissue heterogeneity weighting factors, nonlinear transport correction coefficients and intermediate parameters, a heterogeneous nonlinear pharmacokinetic model is constructed. First, each pixel in the preprocessed image data is processed using a region partitioning function to generate tissue region labels. Since the perfusion dynamics of different regions vary significantly, this process avoids distortion that might occur after uniform modeling. The region partitioning function is shown below: in, For pixels Organizational region label These are the position coordinates of the pixel; For the first Feature weight vectors for each region; For pixels The multimodal feature vector; Each pixel is divided into three categories based on multimodal features: vascular region ( ), high-irrigation area ( ) and low-irrigation areas ( ), to obtain the tissue region label to which the pixel belongs: vascular region renal cortex region and renal medullary region ; Based on the observed baseline uptake rate and baseline clearance rate combined with the tissue region label, an adaptive modulation rule is introduced to generate an adaptive uptake rate and adaptive clearance rate. The adaptive modulation rule is as follows: in, For adaptive ingestion rate; The baseline uptake rate in the tissue region; For pixels Local heterogeneity characteristics; For adaptive clearance rate; The baseline clearance rate for the tissue region; and For the modulation intensity coefficient, by trying multiple values ​​of the modulation intensity coefficient within a certain range on several representative images, the peak signal-to-noise ratio is calculated, and the modulation intensity coefficient that best balances denoising effect and detail preservation is selected. For example... , ; The activation function is typically the Sigmoid function; It is the hyperbolic tangent function; The baseline uptake rate and baseline clearance rate for different regions were summarized to obtain the baseline parameter table, as shown below. A typical pixel in the renal cortex region ( Taking this as an example, substituting the data into the formula yields the adaptive uptake rate and adaptive clearance rate: Secondly, a heterogeneous nonlinear pharmacokinetic model is constructed, which, combined with a final high-precision individualized AIF, outputs the contrast agent concentration within the tissue. This allows for adaptive adjustment strategies of model parameters for different patient sites, improving the model's site adaptability. The heterogeneous nonlinear pharmacokinetic model is shown below: in, Contrast agent concentration within the tissue; This refers to the volume fraction of plasma, for example, in vascular regions. renal cortex region renal medullary region ; This refers to the concentration of contrast agent in the plasma, which is the final high-precision individualized AIF. The tissue's adaptive uptake rate reflects the rate at which contrast agent enters the tissue from the blood vessels; For the region heterogeneity matrix; The tissue's adaptive clearance rate reflects the rate at which the contrast agent flows back from the tissue; This is the nonlinear transport correction coefficient; As a weighting factor for organizational heterogeneity; The convolution integral describes the retention and clearance process of contrast agent in the interstitial space. For integration variables; It is a time variable.

[0018] S4: Establish a collaborative reconstruction mechanism between the heterogeneous nonlinear pharmacokinetic model and preprocessed image data; the collaborative reconstruction mechanism is to incorporate the kinetic constraints of the heterogeneous nonlinear pharmacokinetic model as a loss term into the objective function of the preprocessed image data reconstruction, generate an initial reconstructed image, and further iterate and update the intermediate parameters and the initial reconstructed image to finally generate a highly accurate reconstructed image and a distribution map of the optimal intermediate parameters. First, a collaborative reconstruction mechanism between the heterogeneous nonlinear pharmacokinetic model and preprocessed image data is established. The kinetic constraints of the heterogeneous nonlinear pharmacokinetic model are incorporated into the objective function, as shown below: in, Total loss; For data consistency items, To preprocess image data, The initial reconstructed image was obtained by jointly optimizing the data fidelity term, kinetic constraint term, and spatial constraint term, based on the preprocessed image data and under the feedback guidance of the fitting parameters of the heterogeneous nonlinear pharmacokinetic model. and The weighting coefficients are used to plot the total loss as a function of iterations by trying multiple combinations on a small dataset. The combination that steadily decreases and minimizes the final residual is selected, for example, , ; Loss due to dynamic consistency constraints; For spatial structure loss; Expanding the data consistency item, the function is shown below: in, For the first Initial reconstructed image at time; For time sets; It is a time variable; The loss due to dynamic consistency constraints is expanded as follows: in, Loss due to dynamic consistency constraints; The spatial structure loss is expanded into the following function: in, For the first The gradient of the initial reconstructed image at time step; Secondly, based on the total loss and the initial reconstructed image, the initial reconstructed image is iteratively updated using a first iteration function, which is shown below: in, For the first The reconstructed image from the next iteration; For the first The reconstructed image from the next iteration; To determine the gradient descent step size, try multiple combinations on a small dataset, plot the total loss as a function of iterations, and select the combination that achieves stable descent and minimizes the final residual. For example... ; Based on the intermediate parameters, updates are performed using a second iteration function, which is shown below: in, For the first Intermediate parameters for the next iteration; For the first Intermediate parameters for the next iteration; Intermediate parameters, including adaptive uptake rate Adaptive clearing rate ; To determine the iteration step size, try multiple combinations on a small dataset, plot the total loss as a function of iterations, and select the combination that steadily decreases and minimizes the final residual. For example, ; The gradient of the intermediate parameters; Total loss; Finally, a convergence judgment function is constructed and a preset judgment threshold is introduced to determine the termination of the first and second iteration functions. The convergence judgment function is shown below: in, The goodness of fit; Contrast agent concentration within the tissue; The baseline tissue concentration is [value], and the observed value is [value]. The time-averaged concentration of tissue; It is a time variable; For time sets; A preset threshold is used as the judgment threshold. When the goodness of fit is greater than the threshold, the iteration stops. The coefficient of variation of parameter estimation under different thresholds is observed. If the threshold is too low, artifacts will remain; if it is too high, the iteration will not converge. This determines the final threshold. For example... ; Substituting the data into the formula, after 12 iterations, the goodness of fit is obtained: At this point, the fit is greater than the preset judgment threshold, the iteration stops, and the high-precision reconstructed image and the distribution map of the optimal intermediate parameters are output. The distribution map of the high-precision reconstructed image and the optimal intermediate parameters are shown below: in, For high-precision reconstructed images; This is a distribution diagram of the optimal intermediate parameters.

[0019] The above S1-S4 describe the core algorithm flow of the collaborative reconstruction mechanism based on a heterogeneous nonlinear pharmacokinetic model. In a preferred embodiment of the present invention, the generation process of the initial reconstructed image in S4 is specifically implemented using a conditional diffusion generation model, and its overall data flow is as follows: Step 1 (Input and Preprocessing): Use the flat scan T1 weighted image from the preprocessed image data output by S1 as the conditional input image A, and add Gaussian noise to the reference enhancement period image to generate a noisy reference image as the input image B. Step 2 (Hybrid Location Encoding Generation): After image patch embedding of the input image A, spatial location encoding information is generated through a hybrid location encoder. The hybrid location encoder is composed of convolutional location encoding (PEG) and frequency domain location encoding (FPEG). PEG generates local location biases on the feature map through k×k convolution, while FPEG generates global location biases in the frequency domain through two-dimensional Fourier transform and 1×1 convolution. The two are fused element-wise and then injected into the subsequent Transformer module to provide spatial location information with both local translational equivariance and global structural consistency for feature extraction. The previous step of the hybrid location encoding is image patch embedding, and the next step is multi-scale feature fusion and extraction. Step 3 (Multi-scale Feature Fusion and Extraction – Pyramid Conditional Encoder): The token sequence after hybrid positional encoding injection is fed into the Pyramid Transformer Conditional Encoder. This encoder contains multiple scale stages. In each stage, image patch embedding, hybrid positional encoding injection, and Transformer self-attention calculation are performed sequentially. As the stages deepen, the image patch size gradually increases, thereby constructing a hierarchical representation from high-resolution local features to low-resolution global features. Input image A and input image B are concatenated along the channel dimension and then fed into the encoder to output a multi-scale conditional feature set. The previous step of the Pyramid Conditional Encoder is hybrid positional encoding generation, and the next step is a phase decoding network. Step 4 (Phase Decoding Network - Parallel Phase Decoder): The multi-scale conditional features output by the pyramid conditional encoder are distributed to K parallel phase decoder branches, each branch corresponding to a DCE-MRI enhanced phase. Each phase decoder sequentially includes a FiLM temporal modulation module, an upsampling module, a dense residual block, a self-attention module, a SimAM attention module, a convolutional layer, and an activation layer, ultimately generating the enhanced image of that phase through the output layer. The FiLM temporal modulation module maps the diffusion time step and phase time point to scaling and bias parameters using Time-MLP, performs affine transformation on the intermediate features, and achieves temporal conditional modulation. The previous step of the phase decoding network is the multi-scale feature output of the pyramid conditional encoder, and the next step is to output multi-phase DCE-MRI sequences and calculate the loss function. Step 5 (Loss Calculation and Iterative Optimization): For the multi-phase DCE-MRI sequences output in Step 4, calculate the denoising diffusion loss (i.e., the data consistency term in S4) and the pharmacokinetic consistency loss described in S3-S4 (i.e., the parameter consistency constraints estimated based on the heterogeneous nonlinear pharmacokinetic model within the tumor ROI region), respectively. Calculate the total loss together and backpropagate to update the model parameters until the model converges.

[0020] Reference Figure 2 As shown, firstly, through the forward diffusion process in the diffusion model, the original multi-phase DCE-MRI images are... Gaussian noise is gradually added to generate a series of intermediate states. , , This continues until a noisy image that approximately follows a standard normal distribution is finally obtained. During this process, the structural information of the image is gradually destroyed, and texture and edge details gradually disappear. This forward diffusion process provides the foundation for subsequent inverse denoising generation, enabling the model to learn the ability to recover the distribution of the target image from random noise.

[0021] Reference Figure 3 As shown, the unweighted images are first... Noise image under the current diffusion step The images are fused and fed into the feature encoding module as conditions. This encoding module employs a pyramid Transformer structure, using image patch embedding, hybrid positional encoding, and a Transformer module to extract multi-scale feature representations to capture anatomical information and enhancement-related features of the image. Subsequently, the encoded features are input into a parallel phase decoder, with each decoding branch corresponding to a different temporal phase, thus generating a complete multi-phase enhanced image sequence in one forward propagation. This parallel structure effectively improves the structural consistency and temporal continuity between different temporal phases. During backdiffusion, a conditional probability distribution is learned. Gradually restore the noisy image to the target image. This enables the generation of high-quality medical images from noise. Furthermore, this invention introduces a pharmacokinetic constraint mechanism. By extracting time-intensity curves from the generated and real sequences within the tumor region of interest, and estimating key parameters based on a heterogeneous nonlinear pharmacokinetic model, a pharmacokinetic consistency loss is constructed to constrain the rationality of the generated results in terms of physiological kinetics, thereby improving the clinical credibility of the generated images. Reference Figure 4 As shown, these are the two core sub-modules of this invention. The left side is the parallel phase decoder, which takes input features as input. After being modulated by the time-conditional parameters generated by Time-MLP, the images are sequentially passed through upsampling, residual blocks, self-attention, convolution and activation layers, and finally the output layer generates a single-phase DCE-MRI image, realizing the parallel generation of multi-phase images and phase feature modeling. Reference Figure 5 As shown, hybrid positional encoding, through the fusion of PEG local positional encoding and FPEG global flow field positional encoding, provides Transformer with positional information that combines local details with global consistency, thereby improving the model's ability to model the spatial position of images and ensuring the phase consistency and anatomical stability of the generated images.

[0022] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic contrast-enhanced magnetic resonance imaging reconstruction method based on pharmacokinetic model optimization, characterized in that, The steps are as follows: S1: Acquire raw image data, and perform multi-scale data correction preprocessing on the raw data to generate preprocessed image data; S2: Extract large blood vessel regions based on the preprocessed image data as candidate regions for the contrast agent input function; correct the signal-time curves of the candidate regions and output the final individualized contrast agent input function; S3: Based on the final individualized contrast agent input function, and by introducing tissue heterogeneity weighting factors, nonlinear transport correction coefficients and intermediate parameters, a heterogeneous nonlinear pharmacokinetic model is constructed. S4: Establish a collaborative reconstruction mechanism between the heterogeneous nonlinear pharmacokinetic model and preprocessed image data; the collaborative reconstruction mechanism is to incorporate the kinetic constraints of the heterogeneous nonlinear pharmacokinetic model as a loss term into the objective function of the preprocessed image data reconstruction to generate an initial reconstructed image, and further iteratively update the intermediate parameters and the initial reconstructed image to finally generate a highly accurate reconstructed image and a distribution map of the optimal intermediate parameters.

2. The method according to claim 1, characterized in that, The multi-scale data correction preprocessing in S1 includes: wavelet decomposition of the original data to generate a first high-frequency noise component and a low-frequency structural component; applying an adaptive thresholding method to suppress random noise in the first high-frequency noise component to generate a second high-frequency noise component; combining the second high-frequency noise component and the low-frequency structural component, reconstructing the original data using a reconstruction function to generate a reconstructed image; correcting the reconstructed image using a rigid registration function to generate a corrected image; and applying the Gibbs artifact iterative correction method to the corrected image for deconvolution correction.

3. The method according to claim 1, characterized in that, The correction process in S2 includes: averaging the signal time curve of the candidate region to obtain a first signal curve; performing nonlinear smoothing fitting based on the first signal curve to obtain a second signal curve; introducing individual physiological indicators to correct the second signal curve to generate an initial individualized contrast agent input function; constructing a reliability assessment model to score the peak concentration, rise time, and fall rate characteristics in the smooth curve of the initial individualized contrast agent input function, and eliminating distorted signals from abnormal perfusion areas in the initial individualized contrast agent input function.

4. The method according to claim 1, characterized in that, In step S3, each pixel in the preprocessed image data is processed using a region segmentation function to generate tissue region labels. Based on the observed baseline uptake rate and baseline clearance rate, the tissue region labels are used, and an adaptive modulation rule is introduced to generate intermediate parameters, including the adaptive uptake rate and the adaptive clearance rate.

5. The method according to claim 1, characterized in that, The heterogeneous nonlinear pharmacokinetic model in S3 is shown below: in, Contrast agent concentration within the tissue; This is the plasma volume fraction. These are the position coordinates of the pixel; This refers to the concentration of contrast agent in the plasma. For the organization's adaptive uptake rate; For the region heterogeneity matrix; For the tissue's adaptive clearance rate; This is the nonlinear transport correction coefficient; As a weighting factor for organizational heterogeneity; For convolution integral, For integration variables; It is a time variable.

6. The method according to claim 1, characterized in that, The objective function in S4 is as follows: in, Total loss; For data consistency items, To preprocess image data, This is the initial reconstructed image; and These are the weighting coefficients; Loss due to dynamic consistency constraints; For spatial structure loss; The initial reconstructed image is iteratively updated using a first iterative function; the intermediate parameters are iteratively updated using a second iterative function; a convergence judgment function is constructed and a preset judgment threshold is introduced to determine the termination of the iterative update process.

7. The method according to claim 1, characterized in that, The initial reconstructed image in S4 is generated using a conditional diffusion generation model, specifically including: After embedding image blocks into the T1-weighted image of the preprocessed image data output by S1, spatial position coding information is generated by a hybrid position encoder and injected into the feature sequence. The hybrid position encoder is composed of convolutional position coding and frequency domain position coding added and fused element by element. The feature sequence after hybrid positional encoding injection and the noisy reference map features are fed together into the pyramid Transformer conditional encoder to extract hierarchical multi-scale conditional features from high resolution to low resolution through multiple scale stages. The multi-scale conditional features are distributed to K parallel phase decoders. Each phase decoder generates an enhanced image of the corresponding phase through FiLM temporal modulation, upsampling, dense residual blocks, self-attention, and convolutional layers. The K phase decoders simultaneously output complete multi-phase DCE-MRI sequences. The denoising diffusion loss and the dynamic consistency constraint loss in S4 are jointly calculated for the output multi-phase DCE-MRI sequences, and the model parameters are updated by backpropagation until convergence.

Images