A gadolinium-free contrast agent dynamic enhancement magnetic resonance image synthesis method and system

Abstract

The present application relates to the technical field of medical image artificial intelligence, and specifically includes a gadolinium-free contrast agent dynamic enhancement magnetic resonance image synthesis method and system, which comprises the following steps: first, acquiring a multi-parameter magnetic resonance imaging sequence set of the same patient; in the first stage, multi-scale features are extracted by two encoders, and then feature fusion is completed by an adaptive affinity fusion module to obtain a conditional vector; in the second stage, a pre-training vector quantization variational autoencoder is used to construct a quantization latent space, encode the complete imaging sequence into a target latent variable, and construct a starting latent variable after processing the missing sequence position; the starting latent variable, a time variable and the conditional vector are input into a deterministic latent representation flow network to predict a velocity field, the synthesized latent variable is obtained by integrator, and finally a gadolinium-free contrast agent enhanced image is output by the decoder; the present application can process any available magnetic resonance imaging sequence combination and any target missing sequence.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology in medical imaging, and in particular to a method and system for synthesizing dynamic enhanced magnetic resonance images without gadolinium contrast agents. Background Technology

[0002] The differential diagnosis of lesions heavily relies on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Taking hepatocellular carcinoma as an example, in standard examinations, the non-enhancing phase (NC), late arterial phase (LAP), portal venous phase (PVP), and delayed phase (DP) collectively reflect the hemodynamic changes of the lesion. Among these, signs such as high enhancement in the non-circular arterial phase, lack of peripheral clearance, and enhancement capsule are of crucial value in the diagnosis of hepatocellular carcinoma. Current examinations typically require intravenous injection of gadolinium contrast agents, which may be accompanied by allergic reactions, the risk of deposition in the body, and a prolonged examination procedure.

[0003] Synthesizing virtual contrast-enhanced images from non-contrast MRI can provide auxiliary imaging information for patients who cannot or should not use gadolinium contrast agents. However, existing generation methods are mostly based on fixed single-input single-output mappings. When arbitrary sequence loss or changes in scanning protocols exist in clinical practice, multiple models need to be repeatedly trained, making it difficult to adapt to different acquisition combinations. Although some multimodal generation models can receive multi-source inputs, they usually fuse features by direct stitching or element-wise addition, which cannot fully distinguish between diagnostically valuable anatomical information and redundant or conflicting information, thus making it difficult to recover subtle enhancement signs.

[0004] Furthermore, random sampling generation models may produce different results under the same input, and the inference path is complex. In lesion scenarios, these instabilities may lead to structural drift or loss of detail in small lesions, reducing the reproducibility of virtual enhanced images. Current technology lacks a unified method capable of handling arbitrary combinations of input sequences and generating high-fidelity enhanced images using deterministic trajectories. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a gadolinium-free contrast agent-based dynamic enhancement magnetic resonance imaging synthesis method and system. Addressing the task of generating missing enhancement sequences from any available magnetic resonance imaging sequence combination, it employs a two-stage technical process: the first stage extracts and fuses conditional representations related to the patient's anatomical structure from incomplete multiparameter magnetic resonance imaging; the second stage uses these conditional representations to constrain a deterministic flow trajectory within a quantized potential space, outputting a target enhancement phase image; it can process arbitrary input sequence combinations and generate high-fidelity enhancement phase images using a deterministic trajectory.

[0006] On the one hand, a method for synthesizing dynamically enhanced magnetic resonance images without gadolinium contrast agent is provided, including: Obtain a collection of multi-parameter magnetic resonance imaging sequences from the same patient; In the first stage, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features; the adaptive affinity fusion module is used to fuse the multi-scale features to obtain a conditional vector. In the second stage, a pre-trained vector quantization variational autoencoder is used to quantize the latent space; the complete multi-parameter magnetic resonance imaging sequence is mapped into target latent variables through the encoder; the missing sequence positions are processed to construct the initial latent variables; Within the latent space, the initial latent variables, time variables, and the condition vector of the first stage are used as inputs to the deterministic latent representation flow network to predict the velocity field in the latent space; and a solver is used to integrate and solve the velocity field to obtain the synthetic latent variables. The synthesized latent variables are decoded using the decoder in the pre-trained vector quantization variational autoencoder, and the enhanced period image without gadolinium contrast agent is output.

[0007] Preferably, a dynamic affinity routing unit is set in the adaptive affinity fusion module. This unit calculates affinity weights using the fusion query vector and key-value vectors from different encoders and different MRI sequences. It assigns higher weights to features that are consistent with the current patient's anatomical structure and are conducive to enhancing sign inference, and suppresses redundant or inconsistent features. The weighted aggregated features are then global pooled and projected to form a patient-specific anatomical condition vector.

[0008] Preferably, the step of processing the missing sequence positions to construct initial latent variables involves: applying a dynamic mask to the corresponding positions of the missing sequences marked with binary masks, and replacing the unusable parts with noise variables to form initial latent variables.

[0009] Preferably, the multi-parameter magnetic resonance imaging sequence set includes at least one or more of the non-enhanced phase, late arterial phase, portal venous phase, and delayed phase.

[0010] Preferably, the solver is a deterministic ordinary differential equation solver.

[0011] Preferably, during the training phase, the deterministic latent representation flow network models the process of evolving from the initial latent variable to the synthetic latent variable in the quantized latent space as a continuous trajectory constrained by ordinary differential equations. It starts from the incomplete or perturbed initial latent variable at t=0 and reaches the synthetic latent variable containing the target sequence at t=1.

[0012] Preferably, during the training phase, positive sample pairs are formed using complete sequence representations and random missing subset representations of the same patient, and negative sample pairs are formed using representations of other patients, thereby optimizing the anatomical structure consistency contrast loss.

[0013] On the other hand, a gadolinium-free contrast agent-free dynamic enhancement magnetic resonance imaging synthesis system is provided, comprising: The acquisition module acquires a set of multi-parameter magnetic resonance imaging sequences from the same patient. In the first stage module, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features; the adaptive affinity fusion module is used to fuse the multi-scale features to obtain a conditional vector. In the second stage module, a pre-trained vector quantization variational autoencoder is used to quantize the latent space; the complete multi-parameter magnetic resonance imaging sequence is mapped into target latent variables through the encoder; the missing sequence positions are processed to construct initial latent variables; within the latent space, the initial latent variables, time variables, and the condition vector from the first stage are used as inputs to a deterministic latent representation flow network to predict the velocity field in the latent space; and a solver is used to integrate and solve the velocity field to obtain the synthetic latent variables. The decoding module uses the decoder in the pre-trained vector quantization variational autoencoder to decode the synthetic latent variables and output the enhancement period image without gadolinium contrast agent.

[0014] Furthermore, an electronic device is also provided, including: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0015] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0016] The above technical solution has the following advantages or beneficial effects: This invention addresses the task of generating missing enhancement sequences from any available combination of MRI sequences, employing a two-stage workflow. The first stage extracts and fuses conditional representations related to patient anatomy from incomplete multiparametric MRI. The second stage uses these conditional representations to constrain a deterministic flow trajectory within a quantized latent space, outputting the target enhancement phase image. This allows for processing any combination of available MRI sequences and any target missing sequence within a single model, reducing the dependence of fixed mapping models on complete scan protocols. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 The overall process of a gadolinium-free contrast agent-based dynamic enhancement magnetic resonance imaging synthesis method is shown in Example 1. Figure 1 ; Figure 2 The overall process of a gadolinium-free contrast agent-based dynamic enhancement magnetic resonance imaging synthesis method is shown in Example 1. Figure 2 ; Figure 3 This is a schematic diagram of some non-equilibrium optimal transport constraints and deterministic trajectories in a gadolinium-free contrast agent-based dynamic enhanced magnetic resonance imaging synthesis method according to Example 1. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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 invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0022] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Example 1 Terminology Explanation: Gadolinium-free contrast agents: Gadolinium contrast agents are commonly used intravenously injected contrast agents in clinical dynamic contrast-enhanced magnetic resonance imaging (MRI) to highlight the blood flow characteristics of lesions. Gadolinium-free contrast agents refer to those that do not inject this type of contrast agent, but instead rely on existing non-contrast MRI images to generate equivalent enhanced images through technical means, thus avoiding problems caused by gadolinium contrast agents such as allergies, in vivo deposition, and prolonged examination procedures.

[0025] Missing Enhancement Sequences: Standard clinical dynamic contrast-enhanced MRI examination includes multiple time-series image sequences such as the non-enhanced phase, late arterial phase, portal venous phase, and delayed phase. Missing enhancement sequences refer to one or more phases of dynamic contrast-enhanced images that could not be acquired due to factors such as scanning protocol, equipment, and patient condition. These are the target images that this invention aims to synthesize and complete.

[0026] Dynamic contrast-enhanced magnetic resonance imaging (MRI) is a core imaging technique for the differential diagnosis of diseases in organs such as the liver. After intravenous injection of contrast agent, continuous scanning is performed at different time points to acquire multi-temporal images, including the non-enhancing phase, late arterial phase, portal venous phase, and delayed phase. This directly reflects the hemodynamic changes of the lesion, and the nature of the lesion is determined by imaging signs such as enhancement and clearance.

[0027] In one exemplary embodiment, such as Figure 1-2 As shown, a method for synthesizing dynamically enhanced magnetic resonance images without gadolinium contrast agent is provided.

[0028] S1: Obtain a set of multi-parameter magnetic resonance imaging sequences for the same patient.

[0029] The multi-parameter magnetic resonance imaging sequence set includes at least one or more of the following: non-enhanced phase, late arterial phase, portal venous phase, and delayed phase.

[0030] Acquiring a collection of multi-parameter magnetic resonance imaging sequences from the same patient The multiparameter magnetic resonance imaging sequence set includes at least one or more of the non-enhancing phase (NC), late arterial phase (LAP), portal venous phase (PVP), and delayed phase (DP).

[0031] Multi-center, multi-parameter magnetic resonance imaging (MRI) data were acquired. A set of multi-parameter MRI sequences corresponding to the same patient included NC, LAP, PVP, and DP sequences, as well as lesion region annotations.

[0032] The images of NC, LAP, PVP, and DP sequences are registered, intensity normalized, and region cropped to obtain the preprocessed NC, LAP, PVP, and DP sequences.

[0033] During the training phase, a subset is randomly sampled from the complete sequence set. One or more sequences are randomly selected from the complete sequence set as input subsets, and the remaining sequences are used as targets for synthesis. A binary mask is used to indicate whether each sequence is available. This training method can comprehensively cover various clinical sequence loss scenarios and allows the model to learn the mapping relationship of missing sequences from any combination of available sequences. During the inference phase, only the actually available sequences and their masks need to be input to specify the missing sequences to be synthesized.

[0034] S2: In the first stage, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features; the adaptive affinity fusion module is used to fuse the multi-scale features to obtain a conditional vector.

[0035] In the first stage, two parallel pre-trained medical image encoders are set up, namely the first encoder and the second encoder. The first encoder is used to extract global semantic and contrast-related features of multi-parameter magnetic resonance imaging sequences, and the second encoder is used to extract anatomical and segmentation-related features of multi-parameter magnetic resonance imaging sequences.

[0036] Among them, the contrast-related features refer to the differences in signal intensity, gray / brightness, texture, and local enhancement patterns between lesions and normal tissues.

[0037] Anatomical structures and segmentation-related features include the morphology, location, outline, boundaries, spatial adjacency, and lesion region segmentation response characteristics of the liver, lesions, and surrounding tissues.

[0038] Each input sequence is processed by the first encoder and the second encoder to obtain corresponding multi-scale features. All multi-scale features are linearly projected and mapped to a unified dimension before being input into the adaptive affinity fusion (AAF) module.

[0039] The multi-scale features include global semantic and contrast-related features output by the first encoder, and anatomical and segmentation-related features output by the second encoder; these features are obtained by linear projection of feature maps output by the two encoders at different network layers.

[0040] Each sequence of images is input into a parallel medical image encoder to obtain semantic and structural features. These features are then fused using an adaptive affinity fusion module, and finally processed by global pooling and a projection head to form a patient-specific anatomical condition vector. The adaptive affinity fusion module includes a dynamic affinity routing hub (DAR-Hub). This hub calculates affinity weights using the fusion query vector and key-value vectors from different encoders and MRI sequences. Features consistent with the current patient's anatomy and beneficial for enhancing sign inference are assigned higher weights, while redundant or inconsistent features are suppressed. The weighted aggregated features are then global pooled and processed by the projection head to form a patient-specific anatomical condition vector. .

[0041] This invention improves the ability to model the correlation between static anatomical structures and dynamic enhancement signs by forming a patient-specific anatomical condition vector through a collaborative dual encoder and an adaptive affinity fusion module driven by a dynamic affinity routing unit.

[0042] Furthermore, to ensure the conditional vector remains stable even with missing sequences, during the training phase, it is represented using complete sequences from the same patient. and random missing subset representation Positive sample pairs are constructed using representations from other patients, while negative sample pairs are constructed using representations from other patients, optimizing the anatomical consistency contrast loss. This loss keeps the conditional vectors of the same patient close across different sequence availability states, while maintaining the distinguishability of different patient anatomical representations (patient-specific anatomical conditional vectors). After training, the network parameters for the first stage are fixed to stably provide conditional vectors.

[0043] The first-stage network parameters include trainable weights for network layers such as the first encoder, second encoder, adaptive affinity fusion module, and projection head.

[0044] Complete sequence representation and random missing subset representation All sequences are obtained from the first-stage network. The complete sequence set is input into the first-stage network to obtain... After randomly applying a sequence-level mask to the same patient sequence set, the usable sequence subset is input into the same first-stage network to obtain... .

[0045] The patient-specific anatomical condition vector of the present invention is used to characterize the morphology, contour, location and spatial relationship of the liver, lesion and surrounding tissues of the same patient.

[0046] The formula for loss of anatomical structural consistency is: (1) in, For the first A complete sequence representation of a patient, given a complete sequence set as input, is output by the first-stage network. For the same first The random missing subset representation of a patient given a random missing sequence subset input by the output of the first-stage network; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; For indexing patients in the batch training samples, and The time indicates other patient samples; This represents the negative sample corresponding to the i-th other patient.

[0047] This invention maintains stable characterization under both complete and missing sequence inputs by contrasting loss of anatomical structure consistency, which is beneficial to improving the reliability of synthesis under incomplete clinical acquisition conditions.

[0048] S3: In the second stage, a pre-trained vector quantization variational autoencoder is used to quantize the latent space; and the complete multi-parameter magnetic resonance imaging sequence is mapped into target latent variables through the encoder, the missing sequence positions are processed, and the initial latent variables are constructed.

[0049] The process of processing the missing sequence positions to construct initial latent variables involves: applying a dynamic mask to the corresponding positions of the missing sequence marked with a binary mask, and replacing the unusable parts with noise variables to form initial latent variables.

[0050] In the second stage, a pre-trained vector quantization variational autoencoder (VQ-VAE) is used to quantize the latent space. The pre-trained VQ-VAE includes an encoder... and decoder Specifically: encoder First, the MRI sequence is mapped into a continuous latent feature map. For each location in the feature map, the quantified latent variable is obtained by codebook nearest neighbor search.

[0051] Complete target multiparameter MRI sequence via encoder Mapped to target latent variables .

[0052] For the corresponding positions of the missing sequences marked with binary masks in S1, dynamic masks are applied, and the unusable parts are replaced with noise variables to form the initial latent variables. As shown below: (2) in, Let be the initial latent variable at the i-th sequence position; For complete target multiparameter MRI sequences via encoder The target latent variable at the i-th sequence position obtained through mapping; This is a noise variable used to replace the location of missing sequences; The binary mask label for the i-th sequence position; This indicates that the MRI sequence corresponding to this location is available and should be retained. ; This indicates that the sequence is missing and uses Substitute; if ,in It follows a standard Gaussian distribution.

[0053] Specifically, the usability of each MRI sequence is first determined based on the binary mask in S1. For usable sequences, they are mapped to the quantized latent space via an encoder, and the corresponding latent variables are retained. For missing sequences, a dynamic mask is applied to their corresponding latent positions, and noise variables are sampled to replace the unusable parts. This yields initial latent variables that contain both the acquired real sequence information and the reserved generation space for the missing sequences to be synthesized.

[0054] Therefore, the initial latent variables retain the real sequence information that has been obtained, while setting up a space for completion for the sequence to be generated.

[0055] S4: In the latent space, the initial latent variables, time variables, and the condition vector of the first stage are used as inputs to the deterministic latent representation flow network to predict the velocity field in the latent space; and the solver is used to integrate the velocity field to obtain the synthetic latent variables.

[0056] The solver is a deterministic ordinary differential equation solver.

[0057] First, a deterministic latent representation flow (DLRF) network is established. The deterministic latent representation flow network can be represented as follows: Its input is the current potential state. Given the time variable t and the patient-specific anatomical condition vector v, the output is... Dimensionally consistent velocity field It is used to describe the direction and magnitude of the evolution of latent variables from the initial state to the target state, and to predict the velocity field in the latent space.

[0058] In this embodiment, the deterministic latent representation flow network is not an independent general image generation network, but a conditional neural vector field network used to parameterize the drift velocity field of the ordinary differential equation. Specifically, it includes a latent variable input unit, a time input unit, a conditional input unit, a velocity field prediction unit, and an ordinary differential equation integration unit. The latent variable input unit receives the initial latent variable obtained from a dynamic mask and the intermediate latent state obtained by linear interpolation of the initial and target latent variables. The time input unit receives the time variable within the interval [0,1]. The conditional input unit receives the patient-specific anatomical condition vector output by the first-stage adaptive affinity fusion module. The velocity field prediction unit outputs a drift velocity field with the same dimension as the intermediate latent state, used to approximate a constant drift from the source latent variable to the target latent variable. During inference, the ordinary differential equation integrator unit integrates the drift velocity field from t=0 to t=1 to obtain the synthetic latent variables, and the synthesized magnetic resonance image is reconstructed by the decoder of the pre-trained vector quantized variational autoencoder.

[0059] During training, the correspondence between source latent variables and target latent variables is determined by partially unbalanced optimal transfer, and hard constraints are imposed on the target edge distribution to maintain the diagnostic image features in real dynamic enhanced magnetic resonance images.

[0060] Next, a deterministic ordinary differential equation solver is used to integrate and solve the velocity field to obtain the composite latent variables. Specifically, through ordinary differential equations The potential trajectory is described, and the velocity field is solved by integrating in the interval from t=0 to t=1 using a deterministic ordinary differential equation solver to obtain the synthetic potential variables.

[0061] exist When starting from incomplete or perturbed initial latent variables, in The synthetic latent variable containing the target sequence is obtained at the specified time, as shown in the following formula: (3) in, A deterministic latent representation flow network for predicting the velocity field in the latent space; Let the latent variable state be at time t; To synthesize latent variables.

[0062] During training, for the initial latent variables and target latent variables At random time Constructing intermediate potential states ;Will Inputting a deterministic latent representation flow network with t and conditional vector v, it predicts the velocity field. and with As a velocity-supervised optimization prediction error, the velocity field determines the direction and magnitude of the evolution of latent variables at each time point. The deterministic generated trajectory is obtained by integrating the velocity field using the ODE.

[0063] During training, the deterministic latent representation flow network models the process of evolving from initial latent variables to synthetic latent variables within the quantized latent space as a continuous trajectory constrained by ordinary differential equations. When starting from incomplete or perturbed initial latent variables, in The synthetic latent variable containing the target sequence is reached at that time.

[0064] This invention uses DLRF to construct deterministic ordinary differential equations to generate trajectories, enabling the same input to produce reproducible outputs and reducing the risk of structural drift caused by random diffusion sampling.

[0065] To avoid incorrectly transmitting irrelevant background or noisy features to critical diagnostic regions, this invention employs partially unbalanced optimal transport (PUOT) to optimize the pairing relationship between the source and target potential distributions, such as... Figure 3 As shown; by Figure 3 As can be observed, using standard optimal transmission (OT) forces a matching of all source qualities of both the source potential and the target potential, which can easily introduce irrelevant background or noise; using unbalanced optimal transmission (UOT) allows relaxation on both the source potential and the target potential, which may weaken the target diagnostic constraints; using the method proposed in this invention, partially unbalanced optimal transmission (PUOT) preserves the key constraints on the target side (target potential) and discards or downweights irrelevant qualities on the source side (source potential).

[0066] Therefore, the constraints proposed in this invention allow the source side to discard the quality of signs that cannot support diagnosis, while retaining the constraints related to the enhancement signs of the lesion on the target side, thereby enabling the trajectory to prioritize the recovery of information such as high enhancement, clearance, and capsule in the non-circular arterial phase.

[0067] The source latent distribution comes from the initial latent variable set constructed from the missing sequence; the target latent distribution comes from the target latent variable set obtained by encoding and quantizing the complete target sequence using VQ-VAE.

[0068] Initiating latent variables Conditional vector and random time points While inputting a deterministic latent representation stream network, training pairings are determined based on partially unbalanced optimal transport, optimizing velocity prediction error and consistency constraints, enabling the deterministic latent representation stream network to learn a linear deterministic trajectory from incomplete latent variables to complete augmented sequence latent variables.

[0069] This invention selectively discards irrelevant source-side features and protects key diagnostic areas on the target side by using PUOT, which can improve the preservation of small lesions and subtle enhancement signs.

[0070] S5: The decoder in the pre-trained vector quantization variational autoencoder is used to decode the synthetic latent variables and output the enhancement period image without gadolinium contrast agent.

[0071] Synthesize latent variables Input the pre-trained VQ-VAE decoder and output the target enhancement period image, i.e., the virtual enhancement period image under the condition of no gadolinium contrast agent injection.

[0072] For the patient to be examined, the available NC or partially enhanced sequences and target sequence identifiers are input. In the first stage, a conditional vector is generated. In the second stage, the latent representation of the target sequence is obtained through ODE integration, and LAP, PVP, or DP images are generated through decoding. The synthesized results can be given to radiologists for image review, or input into downstream lesion diagnostic models to assist in the identification of lesions such as hepatocellular carcinoma, hemangioma, and focal nodular hyperplasia.

[0073] The present invention can also input the synthesized image into the lesion classification, segmentation or imaging feature evaluation module to form an auxiliary diagnostic process for focal lesions under gadolinium-free contrast agent conditions.

[0074] Example 2 This embodiment provides a gadolinium-free contrast agent-free dynamic enhancement magnetic resonance imaging synthesis system, including: The acquisition module acquires a set of multi-parameter magnetic resonance imaging sequences from the same patient. In the first stage module, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features; the adaptive affinity fusion module is used to fuse the multi-scale features to obtain a conditional vector. In the second stage module, a pre-trained vector quantization variational autoencoder is used to quantize the latent space; the complete multi-parameter magnetic resonance imaging sequence is mapped into target latent variables through the encoder; the missing sequence positions are processed to construct initial latent variables; within the latent space, the initial latent variables, time variables, and the condition vector from the first stage are used as inputs to a deterministic latent representation flow network to predict the velocity field in the latent space; and a solver is used to integrate and solve the velocity field to obtain the synthetic latent variables. The decoding module uses a decoder to decode the synthetic latent variables and outputs an enhancement-period image without gadolinium contrast agent.

[0075] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0076] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for synthesizing dynamically enhanced magnetic resonance images without gadolinium contrast agent, characterized in that, include: Obtain a collection of multi-parameter magnetic resonance imaging sequences from the same patient; In the first stage, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features; The adaptive affinity fusion module is used to fuse multi-scale features to obtain a conditional vector; In the second stage, a pre-trained vector quantization variational autoencoder is used to quantize the latent space; and the complete multi-parameter magnetic resonance imaging sequence is mapped into target latent variables by the encoder. The missing sequence positions are processed to construct the initial latent variables; Within the latent space, the initial latent variables, time variables, and the condition vector of the first stage are used as inputs to the deterministic latent representation flow network to predict the velocity field in the latent space; and a solver is used to integrate and solve the velocity field to obtain the synthetic latent variables. The synthesized latent variables are decoded using the decoder in the pre-trained vector quantization variational autoencoder, and the enhanced period image without gadolinium contrast agent is output.

2. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, In the adaptive affinity fusion module, a dynamic affinity routing unit is set up. This unit calculates affinity weights using the fusion query vector and key value vectors from different encoders and different MRI sequences. It assigns higher weights to features that are consistent with the current patient's anatomical structure and are conducive to enhancing sign inference, while suppressing redundant or inconsistent features. The weighted aggregated features are then global pooled and projected to form a patient-specific anatomical condition vector.

3. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, The process of processing the missing sequence positions to construct initial latent variables involves: applying a dynamic mask to the corresponding positions of the missing sequence marked with a binary mask, and replacing the unusable parts with noise variables to form initial latent variables.

4. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, The multi-parameter magnetic resonance imaging sequence set includes at least one or more of the following: non-enhanced phase, late arterial phase, portal venous phase, and delayed phase.

5. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, The solver is a deterministic ordinary differential equation solver.

6. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, During the training phase, the deterministic latent representation flow network models the process of evolving from the initial latent variable to the synthetic latent variable in the quantized latent space as a continuous trajectory constrained by ordinary differential equations. It starts from the incomplete or perturbed initial latent variable at t=0 and reaches the synthetic latent variable containing the target sequence at t=1.

7. The method for synthesizing gadolinium-free contrast-enhanced magnetic resonance images according to claim 1, characterized in that, During the training phase, positive sample pairs are formed using complete sequence representations and random missing subset representations of the same patient, while negative sample pairs are formed using representations of other patients, thus optimizing the anatomical structure consistency contrast loss.

8. A gadolinium-free contrast agent-free dynamic enhancement magnetic resonance imaging synthesis system, characterized in that, include: The acquisition module acquires a set of multi-parameter magnetic resonance imaging sequences from the same patient. In the first stage module, the first encoder and the second encoder are used to extract multi-parameter magnetic resonance imaging sequences to obtain multi-scale features. The adaptive affinity fusion module is used to fuse multi-scale features to obtain a conditional vector; In the second phase module, a pre-trained vector quantization variational autoencoder is used to quantize the latent space. The complete multi-parameter magnetic resonance imaging sequence is then mapped into target latent variables via an encoder. The missing sequence positions are processed to construct initial latent variables; within the latent space, the initial latent variables, time variables, and the condition vector of the first stage are used as inputs to the deterministic latent representation flow network to predict the velocity field in the latent space; and a solver is used to integrate and solve the velocity field to obtain the synthetic latent variables. The decoding module uses the decoder in the pre-trained vector quantization variational autoencoder to decode the synthetic latent variables and output the enhancement period image without gadolinium contrast agent.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method described in any one of claims 1-7.

Images