Magnetic resonance imaging super-resolution reconstruction method and device

Abstract

The invention relates to a magnetic resonance imaging super-resolution reconstruction method and device, and is applied to the technical field of medical image processing, and the medical image processing technical method comprises the steps: obtaining at least one low-resolution target contrast magnetic resonance image corresponding to a to-be-reconstructed super-resolution magnetic resonance image, respectively converting at least one high-resolution reference contrast magnetic resonance image and at least one high-resolution reference contrast magnetic resonance image into a preset Fourier space to obtain corresponding Fourier space data; obtaining Fourier space sequence data according to a preset sorting method, and inputting the Fourier space sequence data into the target deep learning model to obtain predicted high-frequency region data; the target deep learning model adopts a coding and decoding structure; and generating a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data. According to the invention, the accuracy and robustness of generating the high-resolution image in the MRI super-resolution method are improved.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a method and device for super-resolution reconstruction of magnetic resonance imaging. Background Art

[0002] Magnetic resonance imaging (MRI) is an important medical imaging technique that provides rich soft tissue contrast information without ionizing radiation. However, MRI imaging is relatively slow, and obtaining high-resolution (HR) images typically requires long scan times, which can cause motion artifacts, patient discomfort, and limit clinical throughput. To shorten scan times, low-resolution (LR) images are often acquired, but this sacrifices image detail and compromises diagnostic accuracy.

[0003] MRI super-resolution (SR) technology aims to reconstruct high-resolution images from low-resolution (LR) images and is a key approach to resolving this contradiction. Existing MRI SR methods include interpolation-based methods, reconstruction-based methods (such as compressed sensing), and learning-based methods (particularly deep learning). However, these methods still fail to fully exploit the characteristics of MRI signals, resulting in a need for improvement in the accuracy and robustness of high-resolution image generation.

[0004] The accuracy and robustness of generating high-resolution images in existing MRI super-resolution methods still need to be improved, and no effective solution has been proposed so far. Summary of the Invention

[0005] In this embodiment, a magnetic resonance imaging super-resolution reconstruction method and apparatus are provided to address the problem in the related art that the accuracy and robustness of generating high-resolution images in MRI super-resolution methods still need to be improved.

[0006] In a first aspect, this embodiment provides a method for super-resolution reconstruction of magnetic resonance imaging, the method comprising:

[0007] Acquiring at least one low-resolution target contrast magnetic resonance image and at least one high-resolution reference contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed;

[0008] Converting the low-resolution target contrast magnetic resonance image and the high-resolution reference contrast magnetic resonance image from original image space to a preset Fourier space to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data;

[0009] Sort the Fourier space data according to a preset sorting method to obtain Fourier space sequence data;

[0010] Input the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model adopting an encoder-decoder structure and is used to generate high-frequency region data with a target contrast in a preset Fourier space;

[0011] Generate a super-resolution reconstructed magnetic resonance image with a target contrast based on the predicted high-frequency region data and the low-resolution target space data.

[0012] In some embodiments, the sorting the Fourier space data according to a preset sorting method to obtain Fourier space sequence data includes:

[0013] Obtain multiple spatial points corresponding to the high-resolution reference space data and the low-resolution target space data; calculate multiple spatial distances of the multiple spatial points from the center point in the preset Fourier space;

[0014] Arrange the multiple spatial points in the form of a vector sequence according to the multiple spatial distances, and determine one-dimensional Fourier space sequence data based on a preset embedding strategy.

[0015] In some embodiments, the sorting the Fourier space data according to a preset sorting method to obtain Fourier space sequence data includes:

[0016] Divide the high-resolution reference space data and the low-resolution target space data into multiple spatial data blocks according to a preset block;

[0017] Calculate multiple spatial block distances between the center points of the multiple spatial data blocks and the center point in the preset Fourier space;

[0018] Arrange the multiple spatial data blocks in the form of a vector sequence according to the multiple spatial block distances, and determine multi-dimensional Fourier space sequence data based on a preset embedding strategy.

[0019] In some embodiments, the method further includes:

[0020] Determine the low-resolution target space data according to a preset sampling strategy.

[0021] In some embodiments, the preset deep learning model includes an encoder and a decoder;

[0022] Inputting the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data, including:

[0023] Inputting the high-resolution reference space data processed according to the preset sorting method into the encoder to generate a context representation;

[0024] Inputting the context representation and the low-resolution target space data processed according to the preset sorting method into the decoder to generate predicted high-frequency region data.

[0025] In some embodiments, generating a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data, including:

[0026] Merging the high-frequency region data and the low-resolution target space data to obtain predicted high-resolution target space data;

[0027] Converting the predicted high-resolution target space data from the preset Fourier space to the original image space to generate a super-resolution reconstructed target contrast magnetic resonance image.

[0028] In some embodiments, the method further includes:

[0029] Extracting target high-frequency region data of real high-resolution target training images in a preset training set; the preset training set further includes low-resolution target training images and high-resolution reference training images;

[0030] Inputting the low-resolution target training images and the high-resolution reference training images into a preset deep learning model to determine predicted high-frequency region data of the target contrast in the Fourier space;

[0031] Based on a preset loss function, determining a loss value between the target high-frequency region data and the predicted high-frequency region data;

[0032] According to a preset gradient descent algorithm, training the preset deep learning model based on the loss value to obtain a target deep learning model.

[0033] In a second aspect, in this embodiment, a magnetic resonance imaging super-resolution reconstruction device is provided, and the device includes: an image acquisition module, a space conversion module, a sorting module, and a reconstruction module;

[0034] The image acquisition module is configured to acquire at least one low-resolution target contrast magnetic resonance image and at least one high-resolution reference contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed;

[0035] The spatial transformation module is configured to transform the low-resolution target contrast magnetic resonance image and the high-resolution reference contrast magnetic resonance image from the original image space to a preset Fourier space respectively, so as to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data;

[0036] The sorting module is configured to perform sorting processing on the Fourier space data according to a preset sorting method, so as to obtain Fourier space sequence data;

[0037] The reconstruction module is configured to input the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model adopting an encoder-decoder structure, and is used to generate high-frequency region data of a target contrast in a preset Fourier space; and is further used to generate a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data.

[0038] In a third aspect, an electronic device is provided in this embodiment, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the magnetic resonance imaging super-resolution reconstruction method described in the first aspect above is implemented.

[0039] In a fourth aspect, a storage medium is provided in this embodiment, on which a computer program is stored. When the program is executed by a processor, the magnetic resonance imaging super-resolution reconstruction method described in the first aspect above is implemented.

[0040] Compared with the related art, a magnetic resonance imaging super-resolution reconstruction method and device provided in this embodiment transform a low-resolution target contrast image and a high-resolution reference contrast image from the original image space to the Fourier space (K space), and in the Fourier space, combine a deep learning model with an encoder-decoder structure to realize the prediction of high-frequency region data of a target contrast. After the prediction is completed, the predicted data is transformed from the Fourier space to the original image space. By using the frequency domain characteristics of the K space and the method of fusing the information of the target contrast image and the reference contrast image, the generation of high-frequency information of the target contrast is guided, ensuring that the generated high-frequency information is consistent with the low-frequency information, and improving the accuracy and robustness of the reconstructed high-resolution image.

[0041] Details of one or more embodiments of the present application are set forth in the following drawings and description, so that other features, objects, and advantages of the present application become more comprehensible. Description of the Drawings

[0042] The accompanying drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not unduly limit the present application. In the drawings:

[0043] Figure 1 is a hardware structure block diagram of a terminal of the magnetic resonance imaging super-resolution reconstruction method provided by an embodiment of the present application;

[0044] Figure 2 is a flowchart of the magnetic resonance imaging super-resolution reconstruction method provided by an embodiment of the present application;

[0045] Figure 3 is a schematic flowchart of obtaining a vector representation by adding multiple embeddings in an embodiment of the present application;

[0046] Figure 4 is a schematic diagram of spatial high and low frequency region division provided by an embodiment of the present application;

[0047] Figure 5 is a schematic diagram of a magnetic resonance imaging super-resolution computer system provided by this specific embodiment;

[0048] Figure 6 is a flowchart of a multi-contrast magnetic resonance imaging super-resolution method based on a K-space Transformer provided by this specific embodiment;

[0049] Figure 7 is a flowchart of obtaining a predicted target high-frequency Fourier space sequence provided by this specific embodiment. Detailed implementation manners

[0050] To understand the purpose, technical solution, and advantages of the present application more clearly, the present application will be described and illustrated below with reference to the accompanying drawings and embodiments.

[0051] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meanings understood by those with ordinary skills in the technical field to which this application belongs. In this application, words such as "a", "an", "one kind", "the", "these", etc. do not indicate a limitation in quantity, and they can be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent in these processes, methods, products or devices. The terms "connect", "be connected", "couple" and other similar words involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "plurality" involved in this application means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, and B exists alone. Usually, the character " / " indicates that the objects associated before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific sorting of the objects.

[0052] The method embodiment provided in this embodiment can be executed on a terminal, a computer or a similar computing device. For example, running on a terminal, Figure 1 is a hardware structure block diagram of a terminal of the magnetic resonance imaging super-resolution reconstruction method provided by an embodiment of this application. As Figure 1 shown, the terminal may include one or more ( Figure 1 only one is shown in the figure) processors 102 and a memory 104 for storing data. Among them, the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA. The above terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above terminal. For example, the terminal may further include more or fewer components than Figure 1 shown in the figure, or have a different configuration from Figure 1 shown.

[0053] The memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the magnetic resonance imaging super-resolution reconstruction method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above-mentioned method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories can be connected to the terminal through a network. Examples of the above network include but are not limited to the Internet, intranet, local area network, mobile communication network, and combinations thereof.

[0054] The transmission device 106 is used to receive or send data via a network. The above network includes a wireless network provided by the communication provider of the terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus communicate with the Internet. In one instance, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0055] Magnetic Resonance Imaging (MRI) is a medical diagnostic technique that uses the principle of nuclear magnetic resonance to image the internal human body. It can provide detailed in-vivo structure images, especially with very good contrast of soft tissues, so it has a wide range of applications in the diagnosis of diseases of the nervous system, musculoskeletal system, cardiovascular system, etc.

[0056] Super-resolution (SR) refers to a series of techniques used to improve the resolution of images or videos. For MRI images, super-resolution techniques can be used to enhance the clarity and detail level of the images, thereby helping doctors make more accurate diagnoses. These techniques can be implemented by different methods, including but not limited to: interpolation-based methods, reconstruction-based methods, and deep learning-based methods.

[0057] Current deep learning-based methods, especially convolutional neural networks (CNNs), have achieved remarkable success in SR tasks. However, most existing methods process in the image domain and may not fully utilize the inherent characteristics of MRI signals in the Fourier space - K-space (frequency domain). The K-space data of MRI contains the frequency information of the image, where each point represents the information of a specific frequency in the image. The central region represents low-frequency information (the general outline and contrast of the image), and the peripheral region represents high-frequency information (image details and edges).

[0058] In addition, clinical MRI examinations usually acquire multiple contrast images of the same part (such as T1-weighted, T2-weighted, FLAIR, proton density-weighted, etc.). There is inherent anatomical consistency and information complementarity among these images with different contrasts. Some existing multi-contrast SR methods attempt to utilize this complementarity, but often do so in the image domain through simple channel concatenation or shallow fusion strategies, making it difficult to effectively and deeply mine and utilize the complex correlations of different contrast images in the K-space to guide the generation of high-frequency information of the target contrast image.

[0059] To address the problems in the prior art such as insufficient utilization of K-space information, low efficiency of multi-contrast information fusion, and difficulty in generating high-frequency details in magnetic resonance imaging methods, there is an urgent need for a magnetic resonance imaging method for MRI super-resolution that can directly operate in the K-space and efficiently fuse multi-contrast information to accurately generate high-frequency details of the target contrast.

[0060] In this embodiment, a method for super-resolution reconstruction of magnetic resonance imaging is provided. Figure 2 It is a flowchart of the method for super-resolution reconstruction of magnetic resonance imaging provided by the embodiments of the present application, as Figure 2 shown. The process includes the following steps:

[0061] Step S210, obtain at least one low-resolution target contrast magnetic resonance image corresponding to the magnetic resonance imaging to be reconstructed super-resolved, and at least one high-resolution reference contrast magnetic resonance image.

[0062] Among them, when it is necessary to perform high-frequency detail enhancement on the magnetic resonance imaging to be reconstructed super-resolved, it is first necessary to obtain at least one low-resolution target contrast magnetic resonance image and at least one registered high-resolution reference contrast magnetic resonance imaging.

[0063] Specifically, the target contrast refers to the type of magnetic resonance imaging (MRI) that needs to be super-resolved. The reference contrast refers to high-resolution MRIs with different imaging contrasts that are spatially registered and used as supplementary information to guide the reconstruction of the target image. Exemplarily, the low-resolution target contrast MRI is set as a low-resolution T2-weighted image; the high-resolution reference contrast MRIs are set as high-resolution T1-weighted images and high-resolution fluid-attenuated inversion recovery (FLAIR) images. The setting methods of images with different contrasts can be determined according to the actual MRI scenario and are not specifically limited herein.

[0064] Furthermore, T1-weighted imaging generates images based on the longitudinal relaxation time (T1) of tissues. By using a short repetition time (TR) and a short echo time (TE), this technique can emphasize the T1 differences between different tissues. Adipose tissue appears as high signal intensity (bright) on T1-weighted images, while water or cerebrospinal fluid appears as low signal intensity (dark). T1WI is particularly useful for showing anatomical structures.

[0065] T2-weighted imaging depends on the transverse relaxation time (T2). It uses a long TR and a long TE to enhance the contrast of T2 values between tissues. Contrary to T1-weighted imaging, on T2-weighted images, water or cerebrospinal fluid appears as high signal intensity (bright), while fat and other soft tissues show different degrees of brightness depending on their water content. T2-weighted imaging helps detect pathological changes such as edema and inflammation.

[0066] Fluid-attenuated inversion recovery imaging (FLAIR) is a special T2-weighted imaging technique that suppresses the signal of free water by applying a selective inversion pulse, thereby reducing the influence of fluids such as cerebrospinal fluid on the image. This technique can effectively eliminate the high signal around the ventricular system, making lesions near these areas more obvious, and is particularly useful for identifying white matter diseases and subcortical lesions.

[0067] Step S220: Convert the low-resolution target contrast MRI and the high-resolution reference contrast MRI from the original image space to a preset Fourier space respectively to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data.

[0068] Among them, since at least one low-resolution target contrast MRI and at least one high-resolution reference contrast MRI are in the original image space, and being in the image domain, the structural characteristics of the frequency information cannot be determined, nor can high-frequency details be generated.

[0069] Therefore, in this embodiment, at least one low-resolution target contrast magnetic resonance image in the original image space, i.e., the image domain, and at least one high-resolution reference contrast magnetic resonance image are transformed from the original image space to a preset Fourier space to obtain corresponding Fourier space data. The Fourier space data includes high-resolution reference space data corresponding to the high-resolution reference contrast magnetic resonance image, and also includes low-resolution target space data corresponding to the low-resolution target contrast magnetic resonance image.

[0070] The method of transforming from the original image space to the preset Fourier space is specifically set as Fourier transform. In practical applications, Fourier transform and its variants can all achieve the effect of transforming to the frequency domain, and the specific Fourier transform method is not specifically limited here.

[0071] Step S230: Sort the Fourier space data according to a preset sorting method to obtain Fourier space sequence data.

[0072] Among them, after obtaining the Fourier space data with frequency domain characteristics, in the Fourier space, the Fourier space data is sorted to realize the preprocessing and serialization of the Fourier space data, and then the Fourier space sequence data is obtained. The Fourier space sequence data is a sequence suitable for input into a deep learning model for processing. Exemplarily, the preset sorting method includes: one-dimensionally expanding and sorting the spatial points corresponding to the Fourier space data based on frequency, and also includes two-dimensional / three-dimensional serialization processing based on the blocks (Patch) corresponding to the Fourier space data. The specific sorting method is not specifically limited here.

[0073] Step S240: Input the Fourier space sequence data into the target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model with an encoder-decoder structure and is used to generate high-frequency region data of the target contrast in the preset Fourier space.

[0074] Among them, after inputting the preprocessed and serialized Fourier space data into the trained target deep learning model, the predicted high-frequency region data is generated through the encoder-decoder structure in the target deep learning.

[0075] Step S250: Generate a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data.

[0076] Among them, after predicting the high-frequency region data, the predicted high-frequency region data and the low-resolution target space data are integrated according to the sequence to obtain the reconstructed target contrast sequence data. Then, the reconstructed target contrast sequence data is transformed from the Fourier space to the original image space, and finally the reconstructed high-resolution target contrast magnetic resonance image, that is, the super-resolution reconstructed target contrast magnetic resonance image, is obtained.

[0077] Through the above steps, the low-resolution target contrast image and the high-resolution reference contrast image are transformed from the original image space to the Fourier space, and in the Fourier space, combined with a deep learning model with an encoding-decoding structure, the prediction of the high-frequency region data of the target contrast is realized. After the prediction is completed, the predicted data is then transformed from the Fourier space to the original image space. By using the frequency domain characteristics of the Fourier space and the method of fusing the target contrast image and the reference contrast image information, the generation of the high-frequency information of the target contrast is improved, ensuring that the generated high-frequency information is consistent with the low-frequency information, and improving the accuracy and robustness of the reconstructed high-resolution image.

[0078] In some of these embodiments, according to a preset sorting method, the Fourier space data is sorted to obtain Fourier space sequence data, including: obtaining a plurality of spatial points corresponding to the high-resolution reference space data and the low-resolution target space data; calculating a plurality of spatial distances of the plurality of spatial points from the center point in the preset Fourier space; arranging the plurality of spatial points in the form of a vector sequence according to the plurality of spatial distances, and determining the one-dimensional Fourier space sequence data based on a preset embedding strategy.

[0079] In this embodiment, the Fourier space data is unfolded one-dimensionally based on the frequency to obtain the one-dimensional Fourier space sequence data, where the frequency refers to the distance from the center of the Fourier space. Specifically, a plurality of spatial points corresponding to the high-resolution reference space data and the low-resolution target space data in the Fourier space data are determined, and a plurality of spatial distances of the plurality of spatial points from the center point of the preset Fourier space are calculated, and then sorted in ascending order according to the plurality of spatial distances, or sorted according to the lexicographical order of the coordinates of the spatial points.

[0080] Exemplarily, the spatial distance, i.e., the Euclidean distance r, from each spatial point (kx, ky) or (kx, ky, kz) to the center of the K-space (0, 0) or (0, 0, 0) can be calculated here. All spatial points are sorted in ascending order according to the Euclidean distance r. To ensure the uniqueness of the sorting result, for spatial points with the same distance, such as those located on the same circle or sphere, secondary sorting rules can be adopted, such as sorting in ascending order according to the polar coordinate angle (atan2(ky, kx)), or sorting in lexicographical order of coordinates (kx first, ky second, kz third). Finally, a one-dimensional sequence data of the Fourier space is obtained, i.e., a one-dimensional sequence of K-space points k[0], k[1],..., where the front part of the sequence corresponds to low-frequency information and the back part corresponds to high-frequency information.

[0081] In some of these embodiments, according to a preset sorting method, the Fourier space data is sorted to obtain Fourier space sequence data, which further includes: dividing the high-resolution reference space data and the low-resolution target space data into multiple spatial data blocks according to a preset block; calculating multiple spatial block distances between the center points of the multiple spatial data blocks and the center point in the preset Fourier space; arranging the multiple spatial data blocks in the form of a vector sequence according to the multiple spatial block distances, and determining the multi-dimensional sequence data of the Fourier space based on a preset embedding strategy.

[0082] Among them, in addition to the one-dimensional sequence expansion method based on frequency in the above embodiments to obtain the one-dimensional sequence data of the Fourier space, the two-dimensional or three-dimensional sequence data of the Fourier space can also be obtained based on patches. Specifically, the Fourier space can also be regarded as an "image" and divided into small patches for processing to capture local frequency space correlations and reduce the sequence length.

[0083] First, the size of the patch is predefined, such as two-dimensional PxP (e.g., 4x4) and three-dimensional PxPxP. The high-resolution reference space data and the low-resolution target space data are divided into multiple spatial data blocks according to the preset block. The multiple spatial data blocks are usually non-overlapping (or can have a small amount of overlap) patches. Subsequently, it is necessary to determine the order of these spatial data blocks forming a sequence, calculate the distance from the center of each patch to the center of the Fourier space, and sort the patches according to this distance (similarly, the sorting uniqueness of patches with the same distance needs to be processed). Finally, a patch sequence Patch[0], Patch[1],... is obtained. The spatial data points inside each patch will be processed (such as flattened) and embedded later to obtain the multi-dimensional sequence data of the Fourier space.

[0084] In determining the Fourier space (K-space) sequence data, the sorted Fourier space data points or spatial data patches (Patches) are represented as a sequence of vectors (Tokens). For the one-dimensional Fourier space sequence data obtained by one-dimensional unfolding, each Token usually represents a Fourier space point (including its complex value information); for the Patch-based method, each Token represents a Patch (including the complex value information of all points within the Patch, usually flattened into a vector first). Each Token not only contains spatial data information (complex value or flattened vector), but also needs to incorporate its spatial position / frequency information.

[0085] In some of these embodiments, the complex values (real part, imaginary part) are concatenated with the encoded position information, such as positional embeddings, or mapped to Token vectors through a linear layer.

[0086] Specifically, Figure 3 is a schematic flow diagram of obtaining a vector representation by adding multiple embeddings provided in the embodiments of the present application. Refer to Figure 3 , Data Embedding: The data part of each K-space Token (e.g., complex value or vector after flattening the Patch), that is, the Fourier space sequence data, is mapped to a high-dimensional vector space through a linear layer to obtain a data vector.

[0087] Positional Embedding: Since the position (frequency) of K-space points or the position of Patches is crucial, positional embeddings need to be introduced to encode this spatial / frequency information. For the one-dimensional Fourier space sequence data obtained by one-dimensional unfolding, learnable positional embeddings based on its index in the one-dimensional sequence or positional encodings calculated based on its original K-space coordinates (kx, ky, kz) can be used (such as Fourier feature embeddings, which map the coordinates to a high-dimensional sine / cosine feature space and can better represent the continuity of the frequency domain). For two-dimensional / three-dimensional Patch sequences, positional embeddings capable of representing the two-dimensional / three-dimensional positions of Patches in the original K-space grid need to be designed, such as learnable 2D / 3D positional embeddings or Fourier feature embeddings based on the center coordinates of the Patch, and finally obtain positional vectors.

[0088] Modality Embedding: In order to distinguish the low-frequency information of different reference contrasts and target contrasts, modality embeddings can be introduced, and finally obtain modality vectors. Specifically, a specific embedding vector can be learned for each input modality (such as T1, T2_LF, FLAIR, etc.) and added to the Token embeddings corresponding to the modality.

[0089] The data vectors, position vectors, and optionally modality vectors obtained by data embedding, position embedding, and modality embedding are added or concatenated and then mapped to form the final token sequence input into the deep learning model, and exemplarily input into the Transformer model.

[0090] In some of these embodiments, the above method further includes: determining low-resolution target space data according to a preset sampling strategy. Here, the low-resolution target space data can be understood as the low-frequency region data in the high-resolution target space data.

[0091] Among them, Figure 4 is a schematic diagram of the division of high and low frequency regions in the space provided by the embodiments of the present application. Refer to Figure 4 , the low-resolution target space data, that is, the low-frequency region data in the high-resolution target space data, is usually the data in the central part of the K space, and the size of its data region is determined by the sampling range of the low-resolution target contrast magnetic resonance image. The sampling range is determined according to the actual imaging sampling and will not be specifically limited here.

[0092] In some of these embodiments, generating a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data includes: merging the high-frequency region data and the low-resolution target space data to obtain predicted high-resolution target space data; converting the predicted high-resolution target space data from the preset Fourier space to the original image space to generate a super-resolution reconstructed target contrast magnetic resonance image.

[0093] Among them, when the predicted high-frequency region data is obtained by predicting through the target deep learning model, the high-frequency region data and the low-resolution target space data are merged to form complete predicted high-resolution target space data.

[0094] In addition, the low-resolution target space data is used to replace any output that may be generated in the low-frequency region in the predicted high-frequency region data to ensure the accuracy of the low-frequency information.

[0095] After that, an inverse Fourier transform is performed on the predicted high-resolution target space data, such as a two-dimensional or three-dimensional fast inverse Fourier transform, to convert the predicted high-resolution target space data from the Fourier space to the original image space. That is, the frequency domain data is converted into the image domain, and finally a super-resolution reconstructed target contrast magnetic resonance image is obtained.

[0096] In some of these embodiments, the preset deep learning model includes an encoder and a decoder; inputting the Fourier space sequence data into the target deep learning model to obtain the predicted high-frequency region data, including: inputting the high-resolution reference space data processed according to the preset sorting method into the encoder to generate a context representation; inputting the context representation and the low-resolution target space data processed according to the preset sorting method into the decoder to generate the predicted high-frequency region data.

[0097] Among them, the preset deep learning model is based on an encoder-decoder structure and includes an encoder and a decoder. Specifically, a Transformer model with an Encoder-Decoder structure is adopted. This model can be based on the standard Transformer architecture or can be adaptively modified for the complex number characteristics of the K-space data, such as using complex number operations or specific activation functions.

[0098] The input of the encoder is the preprocessed and embedded reference contrast K-space data sequence, that is, the sequence corresponding to the high-resolution reference space data. Through a multi-layer self-attention mechanism and a feed-forward network, the complex dependence relationships and information representations inside the high-resolution reference space data and between different reference contrasts are learned to generate context representations (Contextual Representations) containing rich complementary information.

[0099] The input of the decoder is the context representation output by the encoder and the known low-frequency K-space data sequence of the target contrast, that is, the sequence corresponding to the low-resolution target space data, as the "context" or "prompt".

[0100] The decoder can adopt an autoregressive method, generating a sequence of HF Tokens corresponding to a high-frequency region data one by one depending on the previously generated Tokens; or a non-autoregressive method, generating all the sequences of HF Tokens corresponding to the high-frequency region data at once or in batches, which is usually faster but may sacrifice some accuracy. In either case, the generation process needs to be consistent with the serialization order of the predefined high-frequency region in the K-space. For example, if sorted by frequency, it is generated from low to high by frequency; if sorted by Patch, it is generated in the Patch sequence order.

[0101] The decoder includes a self-attention layer and a cross-attention layer, where the self-attention layer (Self-AttentionLayer) is used to process the sequence corresponding to the low-resolution target space data and the sequence corresponding to the predicted high-frequency region data already generated by the decoder, and the internal dependence relationships between the two.

[0102] The Cross-Attention Layer is used to interact the current state of the decoder (based on the sequence corresponding to the low-resolution target space data and the sequence corresponding to the predicted high-frequency region data generated by the decoder) with the reference contrast context representation output by the encoder, and "translate" or extract information from the reference contrast information that helps generate the target high-frequency region data.

[0103] The output of the decoder is the sequence corresponding to the predicted high-resolution target space data, that is, the Token sequence corresponding to the target contrast high-frequency K-space data k_tar_HF_pred.

[0104] In some of these embodiments, the method for training a preset deep learning model to obtain a target deep learning model includes: extracting the target high-frequency region data of the real high-resolution target training images in the preset training set; the preset training set also includes low-resolution target training images and high-resolution reference training images; inputting the low-resolution target training images and the high-resolution reference training images into the preset deep learning model to determine the predicted high-frequency region data of the target contrast in the Fourier space; determining the loss value between the target high-frequency region data and the predicted high-frequency region data based on a preset loss function; and training the preset deep learning model based on the loss value according to a preset gradient descent algorithm to obtain the target deep learning model.

[0105] Among them, paired low-resolution target training images - LR target images, high-resolution reference training images - HR reference images, and real high-resolution target training images - real HR target images (Ground Truth, I_tar_HR_gt) are used for supervised learning. The real high-resolution target training image I_tar_HR_gt is transformed into the K space to obtain k_tar_HR_gt, and the target high-frequency region data k_tar_HF_gt of its high-frequency part is extracted, corresponding to the high-frequency region to be predicted.

[0106] The loss function is used to define a loss function to measure the difference between the predicted high-frequency region data of the target contrast, that is, the predicted high-frequency K-space data k_tar_HF_pred and the target high-frequency region data of the real high-resolution target training image, that is, the real high-frequency data k_tar_HF_gt.

[0107] Exemplarily, the loss function includes: K-space loss: preferably L1 loss, which is more robust to outliers than L2, and calculates the difference between the predicted high-frequency and the real high-frequency: Loss_k = ||k_tar_HF_pred - k_tar_HF_gt||_1. This loss only acts on the high-frequency region that the model needs to generate.

[0108] Image domain loss: L1 or L2 loss, which calculates the difference between the reconstructed image I_tar_HR_pred and the true high-resolution target training image I_tar_HR_gt: Loss_img = ||I_tar_HR_pred - I_tar_HR_gt||_1 or Loss_img = ||I_tar_HR_pred - I_tar_HR_gt||_2^2, where Loss_img represents the image domain loss.

[0109] Combined loss: Usually, the K-space loss and the image domain loss are weighted and combined: Loss_total = lambda_k × Loss_k + lambda_img × Loss_img, where Loss_k represents the K-space loss, Loss_img represents the image domain loss, and lambda_k and lambda_img represent the weight factors.

[0110] Furthermore, an adversarial loss (GAN Loss) or a perceptual loss (Perceptual Loss) can also be selectively added to enhance the visual realism and details of the image.

[0111] The parameters of the Transformer model are trained by minimizing the total loss function using a gradient descent optimization algorithm (such as Adam, AdamW) through an optimizer.

[0112] The magnetic resonance imaging super-resolution reconstruction method provided in the embodiments of this application operates directly in the K-space, which is more in line with the MRI imaging physical process and can better utilize the frequency structure characteristics of the K-space to generate high-frequency details.

[0113] The following describes and illustrates this embodiment through specific examples.

[0114] This specific embodiment provides a multi-contrast magnetic resonance imaging super-resolution method based on K-space Transformer, which can be implemented by software and run on a computer system equipped with sufficient computing resources (such as GPU, which needs to support parallel computing frameworks such as CUDA). Figure 5 It is a schematic diagram of the magnetic resonance imaging super-resolution computer system provided in this specific embodiment. Refer to Figure 5 , this system can include the following functional modules: data interface and preprocessing module, Fourier space processing module, Fourier space model module, reconstruction and output module, and model management and training module.

[0115] Among them, the multi-contrast magnetic resonance images are input into the data interface and processing module. The multi-contrast magnetic resonance images here include at least one low-resolution target contrast magnetic resonance image and at least one high-resolution reference contrast magnetic resonance image in the foregoing embodiments.

[0116] Data interface and preprocessing module: Responsible for receiving multi-contrast magnetic resonance image data in formats such as DICOM or NIfTI, performing image registration when needed, and performing preprocessing such as intensity normalization, and passing the processed data to the subsequent Fourier space processing module.

[0117] Fourier space processing module: Receives the images preprocessed by the data interface and preprocessing module, sequentially performs Fast Fourier Transform (FFT), serialization processing, tokenization processing, and embedding processing, converts them to the K space, and obtains Fourier space data; then serializes the K space data according to a preset strategy (such as sorting by frequency or Patch division); performs tokenization, converts the complex-valued data and position information in the Fourier space data into tokens; implements data embedding, position embedding, and modality embedding layers, and outputs a token sequence acceptable to the Transformer model.

[0118] Fourier space model module - K space Transformer model module: Contains a Transformer model based on the Encoder-Decoder structure, that is, a deep learning model containing encoder and decoder structures. This module receives the reference contrast sequence and the target low-frequency sequence output from the Fourier space processing module, that is, the high-resolution reference space data and the low-resolution target space data in the foregoing embodiments, performs encoding and decoding operations, and outputs a predicted target high-frequency K space token sequence, that is, the sequence corresponding to the high-frequency region data. Among them, the reference contrast sequence is input into the encoder Encoder, and the target low-frequency sequence is input into the decoder Decoder.

[0119] Reconstruction and output module: Receives the sequence corresponding to the predicted high-frequency region data, that is, the high-frequency token sequence, performs inverse fast Fourier transform to convert it back to the complex value in the Fourier space after K space integration and achieving data consistency. Specifically, obtain the low-resolution target space data, that is, the original low-frequency data k_tar_LF, from the K space processing module; perform K space data integration (including the data consistency step); perform inverse fast Fourier transform (IFFT) to convert the complete K space sequence data back to the image domain to obtain a high-resolution reconstructed image; output or display the result in an appropriate format (such as DICOM).

[0120] Model management and training module (offline available): responsible for loading pre-trained deep learning model parameters for inference, or when there is a training set including labeled data, training the model and optimizing the parameters according to the aforementioned model training method.

[0121] In the above computer system, the input image is input to the data interface and the preprocessing module, and sequentially passes through the k-space processing module, the k-space Transformer model module, the reconstruction and output module until a high-resolution image is output. The output of the encoder is passed to the cross-attention layer of the decoder.

[0122] This computer system can run as an independent software or be integrated into a PACS system (Picture Archiving and Communication System), a medical image post-processing workstation or an MRI equipment console to provide fast and high-quality MRI super-resolution reconstruction services for clinical diagnosis and research.

[0123] Based on the above computer system, a magnetic resonance imaging super-resolution method is provided. Figure 6 It is a flowchart of the multi-contrast magnetic resonance imaging super-resolution method based on k-space Transformer provided by this specific embodiment. As Figure 6 shown, the multi-contrast magnetic resonance imaging super-resolution method includes the following steps:

[0124] Step S610, input multi-contrast images.

[0125] Specifically, input a low-resolution T2-weighted image (I_T2_LR), that is, the low-resolution target contrast magnetic resonance image in the aforementioned embodiment; and the registered high-resolution T1-weighted image (I_T1_HR) and high-resolution FLAIR image (I_FLAIR_HR), that is, the high-resolution reference contrast magnetic resonance images in the aforementioned embodiment, into the above data interface and preprocessing module.

[0126] Step S620, convert to the Fourier space.

[0127] Specifically, perform 2D fast Fourier transform on three images respectively to obtain Fourier space data k_T1_HR, k_FLAIR_HR including high-resolution reference space data, and Fourier space data k_T2_HR including high-resolution target space data. Subsequently, determine the known low-frequency region data k_T2_LF in the high-resolution target space data k_T2_HR. The known low-frequency region data in the high-resolution target space data here is the data corresponding to the real low-resolution image in the K space in the foregoing embodiments, that is, the low-resolution target space data, and the high-resolution target space data is used as the true value when calculating the loss function during the training of the transformer model. What needs to be generated is the corresponding unknown high-frequency region data in the high-resolution target space data.

[0128] Step S630, preprocessing and serialization.

[0129] Specifically, sort all the K-space points of the high-resolution reference space data k_T1_HR and k_FLAIR_HR, and the K-space points of the low-resolution target space data k_T2_LF according to the distance from the K-space center, and divide them into small Patches (for example, 4x4).

[0130] Flatten the complex values within each Patch and input them into the embedding layer together with the central frequency coordinates of the Patch. Use different modal embedding vectors to distinguish the Tokens of T1, FLAIR, and T2_LF. Use learnable 2D position embeddings to encode the central frequency coordinates of the Patch.

[0131] Step S640, input a deep learning model based on an encoder-decoder structure.

[0132] Specifically, Encoder: Input the Token sequences corresponding to the high-resolution reference space data k_T1_HR and k_FLAIR_HR.

[0133] Decoder: Input the Token sequence corresponding to the low-resolution target space data, that is, the low-frequency region data k_T2_LF in the high-resolution target space data, as the initial context. The decoder generates a predicted Token sequence for the predicted high-frequency region data k_T2_HF in the order of increasing frequency (corresponding to the region of k_T2_HF), using self-attention on k_T2_LF and the generated part, and cross-attention on the encoder output.

[0134] Furthermore, Figure 7 is the flowchart for obtaining the predicted target high-frequency Fourier space sequence provided by this specific embodiment. Refer to Figure 7, the reference contrast Fourier space sequence and the target contrast low-frequency Fourier space sequence are respectively input into the encoder and decoder in the deep learning model for position encoding, and then sequentially pass through the multi-head self-attention mechanism and the add-and-normalize mechanism, and at the same time calculate the residual after position encoding. Subsequently, the reference contrast Fourier space sequence passes through the feed-forward mechanism and outputs the context representation through the encoder. The target contrast low-frequency Fourier space sequence processed by position encoding and the multi-head self-attention mechanism, and the context representation output by the encoder pass through the multi-head cross-attention mechanism and the feed-forward mechanism through the linear output processing unit to output the predicted target high-frequency Fourier space sequence.

[0135] Step S650, integration and reconstruction of magnetic resonance imaging super-resolution.

[0136] Specifically, the predicted high-frequency region data k_T2_HF_pred and the known low-resolution target space data k_T2_LF are merged to obtain the predicted high-resolution target space data k_T2_HR_pred. Subsequently, the predicted high-resolution target space data is subjected to 2D inverse fast Fourier transform to obtain the reconstructed high-resolution T2 image I_T2_HR_pred, that is, the super-resolution reconstructed target contrast magnetic resonance image is obtained.

[0137] Additionally, the K-space L1 loss can be calculated using the real high-frequency region data k_T2_HF_gt and the image domain L1 loss can be added for joint optimization, for example, setting Loss_total = 1.0×Loss_k_L1 + 0.1×Loss_img_L1.

[0138] Additionally, different Transformer variants can be used, such as the idea of Vision Transformer (ViT model), regarding the K-space as an "image", performing Patch division and processing. Different K-space sorting and Tokenization strategies can be explored. This method can be extended to 3D MRI data. This method can be combined with other techniques (such as generative adversarial network GAN) to further improve the image quality.

[0139] In the above specific embodiments, a multi-contrast magnetic resonance imaging super-resolution method based on K-space Transformer is provided, which regards the MRI super-resolution problem as an "information translation" process in the K-space, with the low-frequency information of the target modality as the context constraint, from the source modality to the target modality.

[0140] It has the following beneficial effects: making full use of K-space information: operating directly in the K-space, which is more in line with the MRI imaging physical process and can better utilize the frequency structure characteristics of the K-space to generate high-frequency details.

[0141] Efficient multi-contrast information fusion: The attention mechanism of the Transformer can capture long-range and complex dependencies between K-space data with different contrasts, enabling deep and efficient information fusion, and precisely guiding the generation of high-frequency information of the target contrast using the information of the reference contrast.

[0142] Context-constrained generation: Using the low-frequency information of the target contrast as the context input to the decoder ensures the consistency between the generated high-frequency information and the low-frequency information, improving the accuracy and robustness of the reconstruction.

[0143] Powerful representation ability: Drawing on the powerful ability of the Transformer in sequence modeling, it is expected to learn better data representations in the MRI super-resolution reconstruction task, thereby reconstructing high-resolution images with higher quality and richer details.

[0144] Flexibility and scalability: This framework can conveniently handle different numbers and types of reference contrast inputs, with good flexibility and scalability, and is applicable to various clinical multi-contrast MRI scanning protocols.

[0145] In this embodiment, a magnetic resonance imaging super-resolution reconstruction device is also provided. This device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated here. The following terms such as "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function. Although the device described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0146] The magnetic resonance imaging super-resolution reconstruction device includes: an image acquisition module, a spatial transformation module, a sorting module, and a reconstruction module.

[0147] The image acquisition module is used to acquire at least one low-resolution target contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed, and at least one high-resolution reference contrast magnetic resonance image.

[0148] The spatial transformation module is used to transform the low-resolution target contrast magnetic resonance image and the high-resolution reference contrast magnetic resonance image from the original image space to a preset Fourier space respectively, to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data.

[0149] The sorting module is used to perform sorting processing on the Fourier space data according to a preset sorting method to obtain Fourier space sequence data.

[0150] A reconstruction module for inputting Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model with an encoder-decoder structure, and is used to generate high-frequency region data with a target contrast in a preset Fourier space; it is also used to generate a super-resolution reconstructed magnetic resonance image with the target contrast based on the predicted high-frequency region data and the low-resolution target space data.

[0151] It should be noted that the above-mentioned modules can be functional modules or program modules, and can be implemented either by software or by hardware. For the modules implemented by hardware, the above-mentioned modules can be located in the same processor; or the above-mentioned modules can also be located in different processors in any combined form.

[0152] In this embodiment, an electronic device is also provided, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0153] Optionally, the above-mentioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the above-mentioned processor, and the input / output device is connected to the above-mentioned processor.

[0154] Optionally, in this embodiment, the above-mentioned processor may be configured to execute the following steps through a computer program:

[0155] S1. Obtain at least one low-resolution target contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed, and at least one high-resolution reference contrast magnetic resonance image.

[0156] S2. Convert the low-resolution target contrast magnetic resonance image and the high-resolution reference contrast magnetic resonance image from the original image space to a preset Fourier space respectively to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data.

[0157] S3. Sort the Fourier space data according to a preset sorting method to obtain Fourier space sequence data.

[0158] S4. Input the Fourier space sequence data into the target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model with an encoder-decoder structure, and is used to generate high-frequency region data with a target contrast in a preset Fourier space.

[0159] S5. Generate a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target spatial data.

[0160] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation manners, and will not be elaborated herein.

[0161] In addition, in combination with the magnetic resonance imaging super-resolution reconstruction method provided in the above embodiments, a storage medium can also be provided in this embodiment to implement it. A computer program is stored on the storage medium; when the computer program is executed by a processor, any one of the magnetic resonance imaging super-resolution reconstruction methods in the above embodiments is implemented.

[0162] It should be understood that the specific embodiments described here are only used to explain this application, rather than to limit it. According to the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the protection scope of this application.

[0163] Obviously, the drawings are only some examples or embodiments of this application. For those of ordinary skill in the art, this application can also be applied to other similar situations according to these drawings without creative work. In addition, it can be understood that although the work done during this development process may be complex and time-consuming, for those of ordinary skill in the art, some design, manufacturing, or production changes based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient disclosure of this application.

[0164] The term "embodiment" in this application means that the specific features, structures, or characteristics described in combination with the embodiment may be included in at least one embodiment of this application. The phrase appears in various positions in the specification does not necessarily mean the same embodiment, nor does it mean being independent or alternative to other embodiments and mutually exclusive. Those of ordinary skill in the art can clearly or implicitly understand that the embodiments described in this application can be combined with other embodiments without conflict.

[0165] The above-described embodiments only represent several implementation manners of this application, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of patent protection. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application should be subject to the appended claims.

Claims

1. A method for super-resolution reconstruction of magnetic resonance imaging, characterized in that, The method includes: obtaining at least one low-resolution target-contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed, and at least one high-resolution reference-contrast magnetic resonance image; converting the low-resolution target-contrast magnetic resonance image and the high-resolution reference-contrast magnetic resonance image from the original image space to a preset Fourier space respectively to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data; performing sorting processing on the Fourier space data according to a preset sorting method to obtain Fourier space sequence data; inputting the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model adopting an encoder-decoder structure and is used for generating high-frequency region data of a target contrast in a preset Fourier space; generating a super-resolution reconstructed target-contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data.

2. The magnetic resonance imaging super-resolution reconstruction method according to claim 1, wherein The performing sorting processing on the Fourier space data according to a preset sorting method to obtain Fourier space sequence data includes: obtaining a plurality of spatial points corresponding to the high-resolution reference space data and the low-resolution target space data; calculating a plurality of spatial distances of the plurality of spatial points from the center point in the preset Fourier space; arranging the plurality of spatial points in the form of a vector sequence according to the plurality of spatial distances, and determining one-dimensional Fourier space sequence data based on a preset embedding strategy.

3. The magnetic resonance imaging super-resolution reconstruction method according to claim 1, wherein The performing sorting processing on the Fourier space data according to a preset sorting method to obtain Fourier space sequence data includes: dividing the high-resolution reference space data and the low-resolution target space data into a plurality of spatial data blocks according to a preset block; calculating a plurality of spatial block distances between the center points of the plurality of spatial data blocks and the center point in the preset Fourier space; arranging the plurality of spatial data blocks in the form of a vector sequence according to the plurality of spatial block distances, and determining multi-dimensional Fourier space sequence data based on a preset embedding strategy.

4. The magnetic resonance imaging super-resolution reconstruction method according to any one of claims 2 or 3, characterized in that The method further includes: determining the low-resolution target space data according to a preset sampling strategy.

5. The magnetic resonance imaging super-resolution reconstruction method according to claim 4, wherein The preset deep learning model includes an encoder and a decoder; The inputting the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data includes: inputting the high-resolution reference space data processed according to the preset sorting method into the encoder to generate a context representation; inputting the context representation and the low-resolution target space data processed according to the preset sorting method into the decoder to generate predicted high-frequency region data.

6. The magnetic resonance imaging super-resolution reconstruction method according to claim 4, characterized in that, The generating a super-resolution reconstructed target-contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data includes: Merge the high-frequency region data and the low-resolution target space data to obtain predicted high-resolution target space data; Convert the predicted high-resolution target space data from a preset Fourier space to the original image space to generate a super-resolution reconstructed target contrast magnetic resonance image.

7. The magnetic resonance imaging super-resolution reconstruction method according to claim 5, wherein The method further includes: Extract the target high-frequency region data of the true high-resolution target training images in a preset training set; the preset training set also includes low-resolution target training images and high-resolution reference training images; Input the low-resolution target training images and the high-resolution reference training images into a preset deep learning model to determine the predicted high-frequency region data of the target contrast in the Fourier space; Based on a preset loss function, determine the loss value between the target high-frequency region data and the predicted high-frequency region data; According to a preset gradient descent algorithm, based on the loss value, train the preset deep learning model to obtain a target deep learning model.

8. A magnetic resonance imaging super-resolution reconstruction device, characterized in that, The device includes: an image acquisition module, a space conversion module, a sorting module, and a reconstruction module; The image acquisition module is configured to acquire at least one low-resolution target contrast magnetic resonance image corresponding to the super-resolution magnetic resonance imaging to be reconstructed, and at least one high-resolution reference contrast magnetic resonance image; The space conversion module is configured to convert the low-resolution target contrast magnetic resonance image and the high-resolution reference contrast magnetic resonance image from the original image space to a preset Fourier space respectively to obtain corresponding Fourier space data; the Fourier space data includes high-resolution reference space data and low-resolution target space data; The sorting module is configured to perform sorting processing on the Fourier space data according to a preset sorting method to obtain Fourier space sequence data; The reconstruction module is configured to input the Fourier space sequence data into a target deep learning model to obtain predicted high-frequency region data; the target deep learning model is trained based on a preset deep learning model adopting an encoder-decoder structure, and is configured to generate high-frequency region data of a target contrast in a preset Fourier space; and is further configured to generate a super-resolution reconstructed target contrast magnetic resonance image based on the predicted high-frequency region data and the low-resolution target space data.

9. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to run the computer program to execute the magnetic resonance imaging super-resolution reconstruction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the magnetic resonance imaging super-resolution reconstruction method according to any one of claims 1 to 7 are implemented.

Images