Low-quality magnetic resonance image reconstruction system, method based on latent space focusing

Abstract

The application provides a low-quality magnetic resonance image reconstruction system and method based on hidden space focusing. A data preprocessing module is configured to preprocess an input magnetic resonance image to obtain three-dimensional image data of a standard magnetic resonance image. A joint distribution optimization module is configured to train a joint probability distribution model using the three-dimensional image data of the standard magnetic resonance image. The joint probability distribution model is configured to construct a joint probability distribution of a low-quality image and a high-quality image. The joint probability distribution model is trained through a low-rank adaptation technique. In the training process, a loss function is constructed using a mean square error loss and a hidden space blood vessel focusing loss, and a flow matching algorithm is used to optimize a generated trajectory. An inference module is configured to map a low-quality image to a high-dimensional hidden space based on model parameters of the trained joint probability distribution model, and to perform deterministic trajectory evolution along a quality manifold defined by flow matching to generate an enhanced high-quality image.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, specifically to a low-quality magnetic resonance image reconstruction system and method based on latent space focusing, and particularly to a technical solution for reconstructing low-field-strength, low-spatial-resolution magnetic resonance images into high-quality magnetic resonance images by introducing a latent space focusing mechanism. Background Technology

[0002] Magnetic resonance imaging (MRI), as a non-invasive medical imaging technique, has been widely used in various fields of clinical diagnosis due to its excellent soft tissue contrast and multi-parameter imaging capabilities. High-quality MRI images are of great significance for the accurate diagnosis of diseases and the formulation of treatment plans.

[0003] However, obtaining high-quality magnetic resonance imaging typically relies on two key factors: high-field magnets and high spatial resolution scanning sequences. These two factors lead to high equipment costs and excessively long scanning times, limiting its application in primary healthcare institutions and in emergency and pediatric settings.

[0004] While some methods exist for enhancing the quality of magnetic resonance images, they generally suffer from weak theoretical foundations, insufficient utilization of prior information, limited 3D data processing capabilities, and poor controllability of reconstruction results. In particular, directly mapping from low-quality images to high-quality images is essentially an underdetermined problem, which easily leads to unstable reconstruction results or artifacts. Furthermore, it often results in the loss of crucial subtle anatomical structures, such as microvessels, due to the singular optimization objective. Summary of the Invention

[0005] To address the aforementioned technical problems, this application proposes a low-quality magnetic resonance image reconstruction system, method, apparatus, and computer storage medium based on latent space focusing.

[0006] To address the aforementioned technical problems, this application proposes a low-quality magnetic resonance image reconstruction system based on latent space focusing. The system includes a data preprocessing module, a joint distribution optimization module, and an inference module. The data preprocessing module preprocesses the input magnetic resonance image to obtain standard 3D image data of the magnetic resonance image. This preprocessing includes spatial registration, brain tissue image extraction, and / or intensity normalization and histogram matching. The joint distribution optimization module trains a joint probability distribution model using the standard 3D image data of the magnetic resonance image. This model constructs a joint probability distribution between the low-quality and high-quality images. The joint probability distribution model is trained using a low-rank adaptation technique. During training, a loss function is constructed using mean squared error loss and latent space vessel focusing loss, and a flow matching algorithm is used to optimize the generated trajectory. The inference module maps the low-quality image to a high-dimensional latent space based on the model parameters of the trained joint probability distribution model and performs deterministic trajectory evolution along the quality manifold defined by flow matching to generate an enhanced high-quality image.

[0007] The process of training the joint probability distribution model by the joint distribution optimization module is as follows: the parameters of the pre-trained generation model are efficiently adjusted based on low-rank adaptation technology to adapt it to the target image enhancement task; during this adaptation process, a composite objective function that integrates mean square pixel error and latent space vascular structure focusing loss is simultaneously applied for optimization; and a flow matching algorithm is used, guided by the model parameters adjusted by the low-rank adaptation technology and the composite objective function, to directly optimize the generation trajectory from noise distribution or low-quality input to high-quality target image, thereby constructing a joint probability distribution of the low-quality image domain and the high-quality image domain.

[0008] The latent space vessel focusing loss method identifies high-response regions in latent space features and assigns them high weight coefficients to guide the model to focus on the reconstruction of vascular details in the image, thereby solving the problem of loss of microstructures.

[0009] The calculation method for the hidden space vascular focusing loss is as follows: on the hidden space feature channel of the joint probability distribution model, the difference between the generated feature map and the target feature map in the vascular structure region is calculated, and the difference is weighted and focused using a predefined vascular structure binary mask.

[0010] The data preprocessing module is equipped with a registration unit, which is used to eliminate inconsistencies in the spatial location, angle and / or scale of magnetic resonance images through an image registration algorithm.

[0011] The data preprocessing module is equipped with a brain image processing unit, which is used to remove skull and non-brain tissue from brain images.

[0012] The brain image processing unit is configured to perform end-to-end pixel-level segmentation on the input magnetic resonance image, identify non-brain tissue, separate non-brain tissue regions, and output a binary mask of the brain parenchyma region.

[0013] To address the aforementioned technical problems, this application proposes a low-quality magnetic resonance image reconstruction method based on latent space focusing. This method utilizes the aforementioned low-quality magnetic resonance image reconstruction system and includes the following steps: Step S1, preprocessing the original magnetic resonance data using a data preprocessing module, wherein the preprocessing includes spatial registration, brain tissue image extraction, and / or intensity normalization and histogram matching; Step S2, training the model using a joint distribution optimization module to obtain a joint probability distribution model, which is used to construct a joint probability distribution between the low-quality and high-quality images; the model training steps are as follows: training the pre-trained model using low-rank adaptation technology; during training, constructing a loss function using mean squared error loss and latent space vessel focusing loss, and optimizing the generated trajectory using a flow matching algorithm; Step S3, inputting the low-quality magnetic resonance image into the joint probability distribution model using an inference module, and reconstructing a high-quality magnetic resonance image using the joint optimization model.

[0014] To address the aforementioned technical problems, this application proposes a low-quality magnetic resonance image reconstruction device based on latent space focusing. The device includes a memory and a processor coupled to the memory. The memory stores program data, and the processor executes the program data to implement the aforementioned low-quality magnetic resonance image reconstruction method based on latent space focusing.

[0015] To address the aforementioned technical problems, this application proposes a computer storage medium for storing program data, which, when executed by a computer, is used to implement the aforementioned low-quality magnetic resonance image reconstruction method based on latent space focusing.

[0016] Compared with existing technologies, the beneficial effects of this application are as follows: The low-quality magnetic resonance image reconstruction system includes a data preprocessing module, a joint distribution optimization module, and an inference module. The data preprocessing module is used to preprocess the input magnetic resonance image to obtain standard three-dimensional image data of the magnetic resonance image. The preprocessing includes spatial registration, brain tissue image extraction, and / or intensity normalization and histogram matching. The joint distribution optimization module is used to train a joint probability distribution model using the standard three-dimensional image data of the magnetic resonance image. This joint probability distribution model is used to construct a joint probability distribution between the low-quality image and the high-quality image. The joint probability distribution model is trained using a low-rank adaptation technique. During training, a loss function is constructed using mean squared error loss and latent space vessel focusing loss, and a flow matching algorithm is used to optimize the generated trajectory. The inference module is used to map the low-quality image to a high-dimensional latent space based on the model parameters of the trained joint probability distribution model, and to perform deterministic trajectory evolution along the quality manifold defined by flow matching to generate an enhanced high-quality image. This system achieves high-fidelity reconstruction from low-quality magnetic resonance images to high-quality images through standardized data preprocessing, low-rank optimized joint probability distribution modeling, and flow-matching-based deterministic trajectory evolution. It significantly improves the ability to restore vascular details and the image signal-to-noise ratio, while also improving model training efficiency and the stability of the generation process. Attached Figure Description

[0017] 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 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. Wherein: Figure 1 This is a schematic flowchart of an embodiment of the low-quality magnetic resonance image reconstruction system based on latent space focusing provided in this application; Figure 2 This is a schematic diagram of the joint distributed optimization module provided in this application; Figure 3 This is another schematic diagram of the joint distributed optimization module provided in this application; Figure 4 This is a flowchart illustrating an embodiment of the low-quality magnetic resonance image reconstruction method based on latent space focusing provided in this application. Figure 5 This is a schematic diagram of an embodiment of the low-quality magnetic resonance image reconstruction device based on latent space focusing provided in this application; Figure 6 This is a schematic diagram of the structure of an embodiment of the computer storage medium provided in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a 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.

[0020] In order to realize a theoretically reliable image reconstruction technology that can make full use of multi-level prior information, is applicable to three-dimensional magnetic resonance data processing, and has controllable reconstruction results, especially a mechanism that can effectively preserve and restore subtle anatomical structures, this application proposes a low-quality magnetic resonance image reconstruction system based on latent space focusing.

[0021] like Figure 1 As shown, Figure 1 This is a schematic flowchart of an embodiment of a low-quality magnetic resonance image reconstruction system based on latent space focusing provided in this application. This application proposes a low-quality magnetic resonance image reconstruction system 800 based on latent space focusing, which includes a data preprocessing module 81, a joint distribution optimization module 82, and an inference module 83.

[0022] The data preprocessing module 81 is used to preprocess the input magnetic resonance image to obtain standard three-dimensional image data of the magnetic resonance image, wherein the preprocessing includes spatial registration, brain tissue image extraction and / or intensity normalization and histogram matching.

[0023] The joint distribution optimization module 82 is used to train a joint probability distribution model using the three-dimensional image data of the standard magnetic resonance image. The joint probability distribution model is used to construct a joint probability distribution of low-quality images and high-quality images. The joint probability distribution model is trained using low-rank adaptation technology. During the training process, a loss function is constructed using mean square error loss and latent space vessel focusing loss, and a flow matching algorithm is used to optimize the generated trajectory.

[0024] The inference module 83 is used to map the low-quality image to a high-dimensional latent space based on the model parameters of the trained joint probability distribution model, and perform deterministic trajectory evolution along the defined quality manifold to generate an enhanced high-quality image.

[0025] Specifically, the main function of the data preprocessing module 81 is to receive raw low-quality and high-quality magnetic resonance image data and process them into a standardized format suitable for network input.

[0026] The data preprocessing module 81 is equipped with a registration unit, which is used to eliminate inconsistencies in the spatial location, angle and / or scale of magnetic resonance images through an image registration algorithm.

[0027] The registration unit performs spatial geometric alignment of low-quality and high-quality images through a resampling algorithm, eliminating displacement deviations caused by different scanning times or patient movement, ensuring voxel-level correspondence, and outputting structurally aligned and uniformly grayscale 3D image data to be provided to the joint distribution optimization module 82.

[0028] The data preprocessing module 81 is equipped with a brain image processing unit, which is used to remove the skull and non-brain tissue from the brain image.

[0029] The brain image processing unit uses automated algorithms to remove non-brain tissue structures, such as the skull and scalp, from images, retaining only the brain region, thereby reducing background noise interference and focusing on the anatomical structures of interest.

[0030] The brain image processing unit performs end-to-end pixel-level segmentation on the input magnetic resonance image, identifies non-brain tissue, separates non-brain tissue regions, and outputs a binary mask of the brain parenchyma region.

[0031] The brain image processing unit is an automated tool for end-to-end pixel-level segmentation. It can scan the entire image pixel by pixel, automatically identify and mark all parts that do not belong to the brain tissue, such as the scalp, skull, meninges, and orbital contents. Finally, it generates a clear binary mask map containing only the brain tissue region, thereby separating the pure brain tissue from the original complex image.

[0032] The brain image processing unit provides clean input for subsequent critical steps such as brain volume measurement and image registration, avoiding interference from non-brain tissue signals, ensuring the reliability of the analysis results, and saving a significant amount of time and manpower required for manual removal of these tissues. Fully automated processing eliminates the subjective variability of manual operation, enabling images generated at different times, on different devices, and even from different institutions to be processed according to a unified standard, greatly improving the consistency and comparability of research data.

[0033] Optionally, the data preprocessing module 81 also includes a normalization and histogram matching unit, which is used to normalize the image intensity and perform histogram matching so that images under different scanning devices or parameters have consistent grayscale distribution characteristics, providing a stable data foundation for subsequent joint distribution optimization.

[0034] The data preprocessing module 81 eliminates the differences in data collection sources through a strict standardized process, enabling the subsequent model to learn more essential features and significantly improving the model's generalization performance and deployment stability across different medical institutions and devices.

[0035] It should be noted that the data preprocessing module 81 is the technical front end, responsible for transforming heterogeneous, raw medical image data into standardized, structured digital information carriers that meet the requirements of subsequent AI model training and inference. Its input does not come directly from the image acquisition equipment, but rather from standardized data packets exported after being archived by the information system and anonymized at the institutional level. The technical operation of the data preprocessing module 81 is based on the publicly available format specifications and numerical matrices contained in these data packets. Its core function is to perform a series of automated and rule-based signal conversion and format standardization operations. The data preprocessing module 81 is compatible with parsers for internationally recognized medical digital image communication standard formats such as DICOM, NIfTI, and PAR / REC, and can automatically read image data and embedded non-identifiable metadata.

[0036] Furthermore, the joint distribution optimization module 82 is the core training part, used to establish a joint probability distribution model between low-quality magnetic resonance images and high-quality magnetic resonance images.

[0037] The process of the joint distribution optimization module 82 training the joint probability distribution model is as follows: the parameters of the pre-trained generation model are efficiently adjusted based on the low-rank adaptation technique to adapt it to the target image enhancement task; during this adaptation process, a composite objective function that integrates mean square pixel error and latent space vascular structure focusing loss is simultaneously applied for optimization; the flow matching algorithm is used, guided by the model parameters adjusted by the low-rank adaptation technique and the composite objective function, to directly optimize the generation trajectory from noise distribution or low-quality input to high-quality target image, and construct the joint probability distribution of low-quality image domain and high-quality image domain.

[0038] In one embodiment of this application, the joint distribution optimization module 82 is based on a pre-trained diffusion Transformer model architecture and is fine-tuned using low-rank adaptation techniques.

[0039] In other embodiments of this application, other neural network deep learning frameworks may also be used, and this application does not make specific limitations.

[0040] The joint distributed optimization module 82 first performs cache generation, the main purpose of which is to establish an efficient data foundation for subsequent cross-modal joint optimization. This process employs a pre-trained variational autoencoder (VAE) to compress and encode the high-dimensional three-dimensional magnetic resonance (MR) volumetric data into a low-dimensional, dense latent space feature vector, serving as a compact mathematical representation of the image content. Simultaneously, features are extracted from textual prompts associated with the image, such as diagnostic descriptions or anatomical markers, to obtain their text embedding vectors. To ensure compatibility between the 3D data and the input format of the VAE model, necessary preprocessing operations are performed during this process, including padding and cropping, to form regular 3D data blocks. Finally, the feature representations of the image and text are stored as a cache for rapid access by subsequent joint optimization modules, thus avoiding the repetitive, time-consuming encoding calculations in each iteration.

[0041] The cache generation step is an efficient preprocessing stage that uses advanced deep learning models to transform images and text into feature forms more suitable for computer processing and joint optimization. It also lays a crucial foundation for the fast and stable operation of the entire joint distributed optimization module 82 by establishing a caching mechanism.

[0042] In this embodiment, the joint distribution optimization module 82 employs a flow matching algorithm, using the latent space features of low-quality images as conditional input to train the network to generate latent space features of high-quality images. During this process, a latent space focusing mechanism is used. This mechanism designs a latent space vessel focusing loss to perform weighted optimization on the features of fine structures such as blood vessels in the latent space.

[0043] Based on the deterministic vector field learned by the flow matching algorithm during the training phase, an ordinary differential equation is constructed from the latent space representation of low-quality images to the latent space representation of high-quality images. The equation is solved by a numerical integrator to obtain the continuous evolution trajectory along the image quality manifold, and finally converges to the latent space point corresponding to the target high-quality image.

[0044] Latent space vessel focusing loss addresses the problem of lost microstructures by identifying high-response regions in latent space features and assigning them high weight coefficients to guide the model to focus on the reconstruction of vascular details in the image.

[0045] The method for calculating the latent space vessel focusing loss is as follows: on the latent space feature channel of the joint probability distribution model, the difference between the generated feature map and the target feature map in the vascular structure region is calculated, and the difference is weighted and focused using a predefined vascular structure binary mask.

[0046] In this embodiment, the training process is divided into two stages. In the first stage, only the mean square error loss is used to restore the basic structure, and in the second stage, the hidden space vessel focusing loss is introduced.

[0047] The first stage is the reconstruction of the basic structure. In this stage, the training objective is to maintain global structural fidelity. By minimizing the mean squared error (MSE) loss, the model is guided to learn the overall mapping relationship from input data, such as low-quality or sparse magnetic resonance images, to high-quality vascular images or labels, ensuring that the macroscopic topology and main contour of the vascular network are correctly restored.

[0048] The second stage focuses on enhancing details and high-frequency features. Once the model possesses basic reconstruction capabilities, training enters the fine-tuning phase, the core of which involves introducing a latent space vessel focusing loss. First, in the low-dimensional latent space feature map generated by the model's encoder, regions with high responses to vascular structures are automatically identified and located. These high-response regions typically correspond to key anatomical features in the latent space, such as vessel edges and bifurcations. Subsequently, when calculating the loss, significantly higher penalty weights are assigned to these high-response regions, while reducing attention to background or homogeneous regions.

[0049] The core of this mechanism is that it forces the model to prioritize the allocation of limited computing resources and optimization budget to the most difficult-to-reconstruct, most information-rich, and most easily lost vascular details during backpropagation optimization. The model is driven to generate vascular structures containing richer high-frequency information and clearer edges.

[0050] The latent space vessel focusing loss identifies high-response regions in the latent space features, corresponding to vascular structures, and applies a significantly higher penalty weight to these regions than to background regions. This mechanism forces the model to prioritize optimizing vascular details during backpropagation, effectively addressing the problem of traditional methods easily losing high-frequency microstructures.

[0051] This application employs a holistic learning process, starting with the overall picture and then moving to the details. The first stage establishes a solid global understanding, preventing the model from getting bogged down in local details and going astray. The second stage refines the model based on this correct understanding. This approach is more stable and efficient than directly using complex loss functions for one-time training.

[0052] Traditional MSE loss treats the error of every pixel in an image equally. Since blood vessels, especially tiny blood vessels, only account for a small proportion of pixels in the entire image, MSE loss is easily dominated by the background area, which occupies a large area. This leads to the optimization direction tending to obtain an average, blurry result that lacks high-frequency details.

[0053] Latent space vascular focusing loss directly overturns this egalitarian approach through a reweighting mechanism. It forcibly shifts the optimization focus back to critical, sparse vascular structures, thus effectively addressing the core pain point of microvessels and blurred edges in traditional methods.

[0054] Unlike operating directly in the image pixel space, identifying high-response regions in the latent space is more fundamental and robust. Latent space features have already been abstracted by the network, eliminating a large amount of irrelevant noise. The high response to blood vessels more directly represents the model's internal understanding of the concept of blood vessels. Applying pressure at this level can more accurately guide the model to optimize its feature extraction and representation capabilities, rather than simply outputting pixel values, which usually leads to better generalization.

[0055] This two-stage training strategy, especially the latent space vessel focusing loss in the second stage, is a clever loss function engineering approach. Through mechanism design, it guides the model to prioritize learning the most critical and difficult parts of the task, thereby significantly improving the performance ceiling on tasks such as vascular imaging that are extremely class-imbalanced and sensitive to high-frequency details. It is a key technical guarantee for obtaining high-precision, high-fidelity vascular reconstruction results.

[0056] In the embodiments of this application, a discrete flow displacement strategy can also be used to optimize time step sampling, further improving the generation quality.

[0057] Furthermore, the inference module 83 is used to receive new low-quality magnetic resonance images and generate high-quality images in practical applications.

[0058] Specifically, the inference module 83 first loads the model weights trained using joint distribution optimization. Next, the low-quality image to be reconstructed, after preprocessing and encoding by a variational autoencoder, is used as conditional input. The low-quality 3D magnetic resonance image to be processed, such as low-resolution, high-noise, or sparsely sampled data, undergoes standardized preprocessing, such as cropping and normalization. Then, it is compressed into a conditional latent space feature vector by a pre-trained variational autoencoder (VAE). This vector encapsulates the core anatomical structural information of the input image, serving as a guiding condition for the generation process.

[0059] The 800 low-quality magnetic resonance image reconstruction system based on latent space focusing uses a learned flow field. Starting with random Gaussian noise or a basic initial state, it iteratively extrapolates under the guidance of a learned probabilistic flow field. This flow field is essentially the inverse transformation path learned by the model from a noise distribution to a sharp image distribution. Guided by the conditional latent vector throughout the process, a numerical integrator gradually and steadily removes noise and injects details along this flow field, ultimately synthesizing a high-quality, high signal-to-noise ratio target latent space feature.

[0060] Finally, the generated latent space features of the target are input into the VAE's decoder. The decoder is responsible for upsampling and reconstructing this detailed, compact representation into high-resolution, well-defined vascular textures and significantly suppressed noise 3D magnetic resonance volume data.

[0061] Furthermore, this application proposes a specific embodiment to describe the implementation process of a low-quality magnetic resonance image reconstruction system 800 based on latent space focusing, as follows: The data preprocessing module 81 is responsible for processing the raw medical image data format. First, the registration unit uses a resampling algorithm to register low-resolution images, such as raw data from time-of-flight magnetic resonance angiography, to a standard space, ensuring that they are consistent with the target high-quality image in anatomical location. The brain extraction unit uses deep learning-based automated tools to strip away non-brain tissue, generating a mask containing only brain parenchyma and blood vessels to eliminate interference from irrelevant structures such as the skull. The normalization unit scales the image voxel intensity to a uniform numerical range and adjusts the grayscale distribution based on a reference histogram to eliminate intensity differences caused by different scanning devices.

[0062] The joint distribution optimization module 82 is built based on the diffusion Transformer model architecture. First, a cache generation step is performed, reading the preprocessed volumetric data file and encoding it into latent variables using a 3D variational autoencoder. For data that does not meet the temporal or spatial dimension requirements, the low-quality magnetic resonance image reconstruction system 800 based on latent space focusing automatically performs edge filling or cropping. Simultaneously, corresponding low-resolution control latent variables are also generated. Subsequently, network training is performed using a mixed-precision training method, and an optimizer is used to update the parameters.

[0063] During training, a low-rank adaptation technique is applied, fine-tuning only the parameters of the injected low-rank matrix to reduce computational resource consumption. For flow matching settings, discrete flow displacement parameters are used to optimize the noise-to-data transmission path. First, mean squared error loss is used for full-image optimization to establish the overall image structure. In subsequent stages, the core latent space vessel focusing loss is introduced. This loss function calculates high-response regions in the latent features, corresponding to the vascular structure, and assigns them significantly higher weights, forcing the model to prioritize optimizing vascular details during backpropagation, thereby greatly improving the reconstruction quality of fine structures.

[0064] The inference module 83 is used to generate the final image. The low-quality magnetic resonance imaging reconstruction system 800 based on latent space focusing receives a new low-quality magnetic resonance angiography image and loads trained weights. First, conditional features of the image are extracted using a variational autoencoder. Then, iterative denoising and flow transformation operations are performed in the latent space using a flow matching scheduler. Finally, the decoder restores the processed latent space features to high-resolution magnetic resonance angiography volume data, which can be directly used for clinical diagnosis.

[0065] Furthermore, this application also proposes an embodiment for the process and result of reconstructing high-resolution magnetic resonance angiography from low-resolution synthetic data using a low-quality magnetic resonance image reconstruction system 800 based on latent space focusing, as detailed below: During training, the low-quality magnetic resonance image reconstruction system 800 based on latent space focusing first undergoes a pre-training phase based on mean squared error loss to learn the basic anatomical structures of the images. Subsequently, the pre-trained weights are loaded and latent space focusing mode is enabled, applying focusing weights to highly activated regions in the latent space features.

[0066] Compared to models trained using only mean squared error loss, the model incorporating latent space focusing mechanism reconstructs a more continuous vascular tree, significantly reduces the breakage of terminal microvessels, and effectively suppresses background noise. Objective metrics evaluation show a significant improvement in the peak signal-to-noise ratio and structural similarity of the reconstructed images, demonstrating the crucial role and effectiveness of this latent space focusing method in preserving anatomical details.

[0067] This invention employs a unique latent space focusing mechanism, which significantly improves the ability to reconstruct microvascular structures in magnetic resonance angiography images by mitigating latent space vascular focusing loss. It effectively solves the problem that traditional methods and general diffusion models are prone to losing high-frequency information, thus ensuring high fidelity of details.

[0068] This invention constructs a complete end-to-end reconstruction pipeline through explicit preprocessing, optimization, and reasoning partitioning, thus achieving process systematization.

[0069] This invention employs a caching mechanism and a low-rank adaptive fine-tuning strategy, significantly reducing training memory usage and time costs. Combined with a staged training strategy, it ensures the stability of model convergence, achieving efficient and stable training. Finally, this invention processes 3D volumetric data based on a video generation architecture, effectively utilizing the spatial continuity information between slices, resulting in highly consistent generated images in 3D space.

[0070] like Figures 2-3 As shown, Figure 2 This is a schematic diagram of the joint distributed optimization module provided in this application; Figure 3 This is another schematic diagram of the joint distributed optimization module provided in this application.

[0071] In this embodiment, the low-resolution image and the high-resolution image are respectively encoded by a variational autoencoder to obtain corresponding latent representations Y and Z. Then, noise is added to Z to obtain a noisy latent representation Zt, and Zt is concatenated with Y channels to form a diffusion model vθ. The system takes the following conditions as input: vθ is then subjected to a stream transform on the stitched latent representation, and the output is split into a denoised latent representation and intermediate variables used for loss calculation. Next, the latent spatial fidelity supervised loss is calculated using Z^ and the true Z, and the stream loss is calculated using the stream model output. The two are then added together to obtain the total loss. Finally, the model weights are updated through backpropagation based on the total loss, completing one round of training iteration, thereby achieving accurate transition from low-resolution medical images to high-resolution images.

[0072] By combining latent diffusion with normalized flow, the detail fidelity and generation stability of super-resolution images are significantly improved while ensuring structural consistency in medical image reconstruction.

[0073] To address the aforementioned technical problems, this application proposes a low-quality magnetic resonance image reconstruction method based on latent space focusing. This method utilizes the low-quality magnetic resonance image reconstruction system 800 in any embodiment of this application. For details, please refer to the following documentation. Figure 4 , Figure 4 This is a flowchart illustrating an embodiment of a low-quality magnetic resonance image reconstruction method based on latent space focusing provided in this application.

[0074] The low-quality magnetic resonance image reconstruction method based on latent space focusing of this application is applied to a low-quality magnetic resonance image reconstruction device based on latent space focusing. This device can be a server, a terminal device, or a system in which the server and terminal device cooperate. Accordingly, the various components of the low-quality magnetic resonance image reconstruction device based on latent space focusing, such as units, subunits, modules, and submodules, can all be located in the server, all in the terminal device, or separately in the server and terminal device.

[0075] Furthermore, the aforementioned server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules, such as software or software modules used to provide distributed servers, or as a single software program or software module; no specific limitations are made here.

[0076] like Figure 4 As shown, it includes the following steps: Step S1: The raw magnetic resonance data is preprocessed using a data preprocessing module. The preprocessing includes spatial registration, brain tissue image extraction, and / or intensity normalization and histogram matching.

[0077] The process involves acquiring pairs of low-quality and high-quality MRI images, performing spatial resampling and rigid or non-rigid registration, then performing brain tissue extraction to remove the skull, and finally performing intensity normalization and histogram matching to ensure the consistency of the input data.

[0078] Step S2: Use the joint distribution optimization module to train the model and obtain the joint probability distribution model. The joint distribution model is used to construct the joint probability distribution of low-quality images and high-quality images. The model training steps are as follows: use low-rank adaptation technology to train the pre-trained model; during the training process, use mean square error loss and latent space vessel focusing loss to construct a loss function, and use flow matching algorithm to optimize the generated trajectory.

[0079] A pre-trained variational autoencoder model is used to compress preprocessed 3D magnetic resonance data into latent variables in the latent space, and a training dataset containing low-quality conditional latent variables and high-quality target latent variables is constructed. A pre-trained diffusion Transformer model is loaded, the backbone parameters are fixed, and only the low-rank adaptation parameters are updated. In training, preliminary training is first performed based on mean squared error loss to establish the global structure. Then, the key latent space focusing loss is introduced, and targeted weighted optimization is performed for high-value regions such as blood vessels. A flow displacement strategy is applied to adjust the generated trajectory to achieve accurate modeling of the joint distribution.

[0080] Step S3: Input the low-quality magnetic resonance image into the joint probability distribution model using the inference module, and reconstruct a high-quality magnetic resonance image from the low-quality magnetic resonance image using the joint optimization model.

[0081] Low-quality magnetic resonance data is input into the system, and after the aforementioned preprocessing, conditional inference is performed using a trained model to output high-quality three-dimensional magnetic resonance images.

[0082] Specifically, steps S1-S3 are the same as the execution process and method of the low-quality magnetic resonance image reconstruction system focusing in the hidden space in any embodiment of this application, and will not be repeated here.

[0083] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0084] To realize the aforementioned low-quality magnetic resonance image reconstruction method based on latent space focusing, this application also proposes a low-quality magnetic resonance image reconstruction device based on latent space focusing. Please refer to [link to details]. Figure 5 , Figure 5 This is a schematic diagram of an embodiment of the low-quality magnetic resonance image reconstruction device based on latent space focusing provided in this application.

[0085] The low-quality magnetic resonance image reconstruction device 400 based on latent space focusing in this embodiment includes a processor 41, a memory 42, an input / output device 43, and a bus 44.

[0086] The processor 41, memory 42, and input / output device 43 are connected to the bus 44. The memory 42 stores program data, and the processor 41 is used to execute the program data to implement the low-quality magnetic resonance image reconstruction method based on latent space focusing described in the above embodiments.

[0087] In this embodiment, processor 41 can also be referred to as a CPU (Central Processing Unit). Processor 41 may be an integrated circuit chip with signal processing capabilities. Processor 41 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or processor 41 can be any conventional processor.

[0088] This application also provides a computer storage medium; please refer to the following: Figure 6 , Figure 6 This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 600 stores a computer program 61. When the computer program 61 is executed by the processor, it is used to implement the low-quality magnetic resonance image reconstruction method based on latent space focusing in the above embodiment.

[0089] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A low-quality magnetic resonance image reconstruction system based on latent space focusing, characterized in that, The low-quality magnetic resonance image reconstruction system includes a data preprocessing module, a joint distribution optimization module, and an inference module; The data preprocessing module is used to preprocess the input magnetic resonance image to obtain standard three-dimensional image data of the magnetic resonance image. The preprocessing includes spatial registration, brain tissue image extraction and / or intensity normalization and histogram matching. The joint distribution optimization module is used to train a joint probability distribution model using the three-dimensional image data of the standard magnetic resonance image. The joint probability distribution model is used to construct a joint probability distribution of low-quality images and high-quality images. The joint probability distribution model is trained using low-rank adaptation technology. During the training process, a loss function is constructed using mean square error loss and latent space vessel focusing loss, and a flow matching algorithm is used to optimize the generated trajectory. The inference module is used to map low-quality images to a high-dimensional latent space based on the model parameters of the trained joint probability distribution model, and perform deterministic trajectory evolution along the defined quality manifold to generate enhanced high-quality images.

2. The system according to claim 1, characterized in that, The process by which the joint distribution optimization module trains the joint probability distribution model is as follows: The pre-trained generative model is efficiently adjusted based on low-rank adaptation technology to adapt it to the target image enhancement task. During this adaptation process, a composite objective function that integrates mean square pixel error and latent space vascular structure focusing loss is simultaneously applied for optimization. The flow matching algorithm is adopted, guided by the model parameters adjusted by the low-rank adaptation technique and the composite objective function, to directly optimize the generation trajectory from noise distribution or low-quality input to high-quality target image, and construct the joint probability distribution of low-quality image domain and high-quality image domain.

3. The system according to claim 1, characterized in that, The latent space vascular focusing loss identifies high-response regions in latent space features and assigns them high weight coefficients, guiding the model to focus on the reconstruction of vascular details in the image, thereby addressing the problem of loss of minute structures.

4. The system according to claim 1 or 3, characterized in that, The calculation method for the latent space vascular focusing loss is as follows: on the latent space feature channel of the joint probability distribution model, the difference between the generated feature map and the target feature map in the vascular structure region is calculated, and the difference is weighted and focused using a predefined vascular structure binary mask.

5. The system according to claim 1, characterized in that, The data preprocessing module is equipped with a registration unit, which is used to eliminate inconsistencies in the spatial location, angle and / or scale of magnetic resonance images through an image registration algorithm.

6. The system according to claim 1, characterized in that, The data preprocessing module is equipped with a brain image processing unit, which is used to remove skull and non-brain tissue from brain images.

7. The system according to claim 6, characterized in that, The brain image processing unit is configured to perform end-to-end pixel-level segmentation on the input magnetic resonance image, identify non-brain tissue, separate non-brain tissue regions, and output a binary mask of the brain parenchyma region.

8. A method for reconstructing low-quality magnetic resonance images based on latent space focusing, characterized in that, The low-quality magnetic resonance image reconstruction method, applied to the low-quality magnetic resonance image reconstruction system according to any one of claims 1-7, includes the following steps: Step S1: The raw magnetic resonance data is preprocessed using a data preprocessing module, wherein the preprocessing includes spatial registration, brain tissue image extraction and / or intensity normalization and histogram matching. Step S2: Use the joint distribution optimization module to train the model and obtain the joint probability distribution model. The joint distribution model is used to construct the joint probability distribution of low-quality images and high-quality images. The steps for model training are as follows: The pre-trained model is trained using low-rank adaptation techniques; During training, a loss function is constructed using mean squared error loss and latent space vessel focusing loss, and a flow matching algorithm is used to optimize the generated trajectory. Step S3: Input the low-quality magnetic resonance image into the joint probability distribution model using the inference module, and reconstruct a high-quality magnetic resonance image from the low-quality magnetic resonance image using the joint probability distribution model.

9. A device for reconstructing low-quality magnetic resonance images based on latent space focusing, characterized in that, The low-quality magnetic resonance image reconstruction device based on latent space focusing includes a memory and a processor coupled to the memory; The memory is used to store program data, and the processor is used to execute the program data to implement the low-quality magnetic resonance image reconstruction method based on latent space focusing as described in claim 8.

10. A computer storage medium, characterized in that, The computer storage medium is used to store program data, which, when executed by the computer, is used to implement the low-quality magnetic resonance image reconstruction method based on latent space focusing as described in claim 8.

Images