Magnetic resonance signal adaptive sampling method and device based on deep learning

Abstract

The present invention provides a method and device for adaptive sampling of magnetic resonance signals based on deep learning, comprising: determining a sampling space, a sampling value threshold, and the number of sampling points for a single excitation; selecting an initial region to be sampled based on the number of sampling points; determining a 3D epi-intensity spectroscopy (EPI) sequence corresponding to a single excitation based on phase encoding of the sampling points within the region to be sampled; performing a single radiofrequency excitation based on the 3D EPI sequence to obtain a corresponding gradient echo signal; reconstructing an optimized dMRI image based on the gradient echo signal; inputting the optimized dMRI image into a trained full-sampling dMRI image prediction model to predict a full-sampling dMRI image and obtain a reconstructed image error map; obtaining a sampling value for each sampling point in the sampling space based on the reconstructed image error map; and determining a region to be sampled for the next sampling based on the sampling value of each sampling point and the sampling value threshold. This application further shortens scanning time and improves the accuracy of the generated dMRI images.

Description

Technical Field

[0001] The present invention relates to the field of magnetic resonance imaging technology, and in particular to a method and device for adaptive sampling of magnetic resonance signals based on deep learning. Background Art

[0002] 3D high-resolution diffusion magnetic resonance imaging (dMRI) is a non-invasive imaging technique that reveals the microstructure and connectivity patterns of white matter fiber tracts in the brain by detecting the diffusion direction of water molecules in tissue. This technology is widely used in neuroscience research, for example, in mapping neural pathways between brain regions. 3D high-resolution imaging significantly improves the resolution of fiber crossings. Combined with techniques such as diffusion tensor imaging (DTI) and diffusion spectrum imaging (DSI), it can quantify parameters such as anisotropy, providing a precise basis for exploring brain network mechanisms. Future integration with artificial intelligence will further promote the development of personalized neuroscience research.

[0003] Although 3D high-resolution dMRI offers significant advantages in revealing the microstructure of brain white matter, its long scan time remains a significant bottleneck restricting its widespread application. Achieving high spatial resolution typically requires increasing the number of phase encoding steps, which results in longer total scan times. Longer scan times can increase patient discomfort and make scan results more susceptible to motion artifacts. Therefore, significantly reducing scan times while maintaining both resolution and accuracy is a pressing need in this field.

[0004] K-space undersampling is one of the primary methods currently used to shorten 3D high-resolution dMRI scan times. K-space is the frequency domain space used to store raw signal data in magnetic resonance imaging (MRI), with each point corresponding to information of a different spatial frequency: the central region contains the image's low-frequency components (which determine overall contrast and structural contours), while the peripheral regions correspond to high-frequency details (which influence resolution and edge clarity). While traditional full sampling requires rigorous traversal of all locations in k-space, undersampling exploits the frequency-domain redundancy of magnetic resonance signals, capturing only a portion of the high-frequency data while ensuring the integrity of the core low-frequency information. This strategy directly reduces the number of phase encoding steps required for scanning, thereby proportionally reducing imaging time. It is particularly effective for multidirectional diffusion-weighted dMRI sequences.

[0005] Existing undersampling methods, such as random sampling and radial / spiral sampling, typically rely on preset static sampling templates and are unable to dynamically optimize the sampling path based on individual anatomical structures or signal characteristics. This can result in the omission of critical high-frequency information or the repeated acquisition of redundant low-frequency data, leading to suboptimal scan time compression and low accuracy of the dMRI images generated based on the sampled signals. Therefore, how to further shorten the scan time and how to improve the accuracy of the generated dMRI images are urgent technical challenges to be solved. Summary of the Invention

[0006] In view of this, an embodiment of the present invention provides a method and apparatus for adaptive sampling of magnetic resonance signals based on deep learning to eliminate or improve one or more defects in the prior art.

[0007] One aspect of the present invention provides a method for adaptive sampling of magnetic resonance signals based on deep learning, the method comprising:

[0008] Determining a sampling space, a sampling value threshold, and the number of sampling points for a single shot, and selecting a K-space trajectory of a corresponding number of sampling points at the center of the sampling space as an initial area to be sampled based on the number of sampling points for a single shot;

[0009] Determining a 3D EPI sequence corresponding to the single excitation based on phase encoding of sampling points within the area to be sampled, and performing a single radio frequency excitation based on the 3D EPI sequence corresponding to the single excitation to obtain a corresponding gradient echo signal;

[0010] Determining to reconstruct a dMRI optimized image based on the obtained gradient echo signal;

[0011] Inputting the reconstructed dMRI optimized image into a trained full-sampling dMRI image prediction model to predict a full-sampling dMRI image, and calculating a reconstructed image error map based on the reconstructed dMRI optimized image and the predicted full-sampling dMRI image;

[0012] The sampling value of each sampling point in the sampling space is obtained based on the reconstructed image error map, and the area to be sampled for the next sampling is determined based on the sampling value of each sampling point and the sampling value threshold.

[0013] In some embodiments of the present invention, determining and reconstructing a dMRI optimized image based on the obtained gradient echo signal includes:

[0014] The reconstructed dMRI optimized image is calculated based on the obtained gradient echo signal using a 3D dMRI image reconstruction algorithm.

[0015] In some embodiments of the present invention, the calculation formula of the 3D dMRI image reconstruction algorithm is:

[0016] ;

[0017] in, To optimize the image for dMRI reconstruction, To optimize the reconstruction of dMRI images, is the number of completed excitation samplings, is the number of coil channels of the magnetic resonance system, is the gradient echo signal collected by the j-th coil after the i-th excitation, F is the Fourier transform, is the sensitivity coefficient of the jth coil, is the complex conjugate of the navigator echo phase of the i-th excitation, is the k-space sampling mask for the i-th excitation, and α is the regularization coefficient.

[0018] In some embodiments of the present invention, the method further comprises:

[0019] Determining an initial fully sampled dMRI image prediction model, a model loss function, and a sample data set, wherein the sample data in the sample data set includes reconstructed dMRI optimized image sample data and fully sampled dMRI image sample data;

[0020] The initial full-sampling dMRI image prediction model is pre-trained based on the model loss function and the sample data set to obtain a trained full-sampling dMRI image prediction model.

[0021] In some embodiments of the present invention, the fully sampled dMRI image prediction model includes an encoder and a decoder, the encoder includes an initial convolution block and five downsampling blocks, and the decoder includes five upsampling blocks and an output layer.

[0022] In some embodiments of the present invention, the initial convolution block includes two 3×3×3 convolutional layers, and each of the downsampling blocks includes a 2×2×2 maximum pooling layer and two 3×3×3 convolutional layers.

[0023] In some embodiments of the present invention, calculating a reconstructed image error map based on the reconstructed dMRI optimized image and the predicted fully sampled dMRI image includes:

[0024] Calculating the absolute difference between the voxel signals of the reconstructed dMRI optimized image and the predicted full-sampled dMRI image;

[0025] The reconstructed image error map is determined based on the absolute difference of the signals of each voxel.

[0026] In some embodiments of the present invention, obtaining the sampling value of each sampling point in the sampling space based on the reconstructed image error map includes:

[0027] Performing Fourier transform on the reconstructed image error map to obtain a K-space error map;

[0028] Calculate the average difference in each frequency encoding direction in the k-space error map;

[0029] The average difference values ​​are used as the sampling values ​​of each sampling point.

[0030] According to another aspect of the present invention, a deep learning-based adaptive sampling system for magnetic resonance signals is provided. The system includes a processor, a memory, and a computer program stored in the memory. The processor is used to execute the computer program. When the computer program is executed, the system implements the steps of the method described in any of the above embodiments.

[0031] According to yet another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any of the above embodiments are implemented.

[0032] The deep learning-based adaptive sampling method for magnetic resonance signals disclosed in the above embodiments of the present invention determines an initial region to be sampled in a sampling space, obtains a corresponding gradient echo signal based on a single excitation, reconstructs an optimized dMRI image based on the acquired gradient echo signal, further predicts the fully sampled dMRI image corresponding to the reconstructed dMRI optimized image using a fully sampled dMRI image prediction model, and calculates a reconstructed image error map based on the reconstructed dMRI optimized image and the fully sampled dMRI image. Furthermore, the sampling value of a sampling point is determined using the reconstructed image error map, and the region to be sampled for the next sampling is determined based on the sampling value of the sampling point. Based on the acquired magnetic resonance signals, this method evaluates and prioritizes the regions in k-space that are expected to contribute most to the reconstruction error, reduces invalid sampling steps, further compresses scan time, reduces the risk of artifacts caused by user motion or physiological noise, and improves the accuracy of the generated dMRI images, providing greater robustness and efficiency for the application of 3D high-resolution dMRI in complex scenarios. Therefore, this deep learning-based adaptive sampling method for magnetic resonance signals not only improves the accuracy of the generated dMRI images, but also further shortens scan time and improves scan efficiency.

[0033] Additional advantages, objects, and features of the present invention will be set forth in part in the following description and will become apparent to those skilled in the art upon examination of the following or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained by the structures particularly pointed out in the description and drawings.

[0034] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present invention are not limited to the above specific descriptions, and the above and other purposes that can be achieved by the present invention will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are intended to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention. The components in the drawings are not drawn to scale, but are merely for the purpose of illustrating the principles of the present invention. To facilitate the illustration and description of certain portions of the present invention, corresponding portions in the drawings may be exaggerated, that is, may be larger than other components in an exemplary device actually manufactured according to the present invention. In the drawings:

[0036] Figure 1 This is a flowchart of a deep learning-based adaptive sampling method for magnetic resonance signals according to an embodiment of the present application.

[0037] Figure 2 FIG. 1 is a timing diagram of a multi-shot 3D EPI sequence according to an embodiment of the present application.

[0038] Figure 3 FIG. 1 is a schematic diagram of the architecture of a full-sampling dMRI image prediction model according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0040] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, the accompanying drawings only show structures and / or processing steps closely related to the solutions according to the present invention, while other details that are not closely related to the present invention are omitted.

[0041] It should be emphasized that the term "include / comprises" when used herein refers to the existence of features, elements, steps or components, but does not exclude the existence or addition of one or more other features, elements, steps or components.

[0042] It should also be noted here that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection with an intermediary, and not only to a wired connection but also to a wireless connection, and the specific connection can be changed based on the actual application scenario.

[0043] To address the technical issues of existing 3D high-resolution dMRI methods, such as long scanning times, susceptibility to user head motion artifacts, and low accuracy of generated dMRI images, the present application provides a method and apparatus for adaptive sampling of magnetic resonance signals based on deep learning. During the sampling process, this adaptive sampling method evaluates and prioritizes the areas in k-space that are expected to contribute most to reconstruction error based on the acquired magnetic resonance signals, selecting them as the next sampling areas. This method not only improves the accuracy of reconstructed dMRI images at the same sampling rate, but also avoids missing high-frequency information or repeatedly acquiring redundant low-frequency data, further shortening scanning time.

[0044] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0045] Figure 1 FIG. 1 is a flow chart of a method for adaptively sampling magnetic resonance signals based on deep learning according to an embodiment of the present application. Figure 1 As shown, the deep learning-based magnetic resonance signal adaptive sampling method includes steps S10 to S50.

[0046] Step S10: Determine the sampling space, sampling value threshold, and number of sampling points for a single shot, and select a K-space trajectory of a corresponding number of sampling points at the center of the sampling space as the initial area to be sampled based on the number of sampling points for a single shot.

[0047] In this step, a sampling space of equal size can be set based on the resolution of the two phase encoding dimensions in the k-space of the target 3D MRI image. Each sampling point in the sampling space corresponds to a k-space trajectory along the frequency encoding direction. The target 3D MRI image can be understood as the 3D MRI image to be generated based on the sampling signal. The sampling value threshold can be set to T, and the number of sampling points in a single shot can be set to n. The number of sampling points in a single shot can also be understood as the number of k-space trajectories acquired in a single shot, that is, the number of k-space trajectories acquired in a single shot is also n. In this case, the k-space trajectories corresponding to the n sampling points are selected from the center of the sampling space as the initial region to be sampled.

[0048] In addition to setting the above-mentioned sampling value threshold and the number of sampling points for a single excitation, the maximum number of excitations can also be set to M, so that the single RF excitation of the sampling space is stopped when the number of excitations (also called the number of samplings) reaches M times.

[0049] Step S20: determining the 3D EPI sequence corresponding to the single excitation based on the phase encoding of the sampling points in the area to be sampled, and performing a single radio frequency excitation based on the 3D EPI sequence corresponding to the single excitation to obtain a corresponding gradient echo signal.

[0050] In this step, the 3D EPI sequence corresponding to a single excitation is determined based on the sampling points corresponding to the single excitation; for example, in the initial sampling step, the 3D EPI sequence corresponding to a single excitation is determined based on the phase encoding of the sampling points within the initial area to be sampled; in the resampling step (the next sampling step), the 3D EPI sequence corresponding to a single excitation is determined based on the phase encoding of the sampling points within the area to be sampled in the next sampling.

[0051] When acquiring gradient echo signals in the initial area to be sampled, phase encoding is set for each gradient echo based on the n sampling points determined in step S10. A single RF excitation is performed based on the 3D EPI sequence corresponding to the single excitation (the cycle) in the multi-shot 3D EPI sequence to acquire and fill the magnetic resonance signals of n k-space trajectories. After completing this initial sampling, the n sampling points corresponding to the sampling in the sampling space can be marked as sampled.

[0052] Figure 2 This is a timing diagram of a multi-shot 3D EPI sequence according to an embodiment of the present application. Figure 2 , a timing diagram of a multi-shot 3D EPI sequence with a dynamically set phase encoding gradient is shown, that is, the multi-shot 3D EPI sequence is a combination of multiple cycles of single-shot 3D EPI sequences. Figure 2 As shown, this sequence achieves three-dimensional k-space filling based on two phase encoding directions and one frequency encoding direction, and generates multiple gradient echoes with different phase encodings after each excitation, thereby acquiring signals on multiple k-space trajectories along the frequency encoding direction. In addition, to correct the phase differences between the signals acquired for each excitation caused by subject motion or physiological noise, an additional navigator echo can be acquired after each excitation. Navigator echo in magnetic resonance imaging is a technology used to detect and correct motion (such as breathing, heartbeat, or slight patient movement). It is particularly important in dynamic organ imaging (such as the heart and abdomen) or in scenarios requiring long scan times, and can significantly reduce motion-induced artifacts. Specifically, to acquire the navigator echo, additional radio frequency pulses and gradients can be inserted into the imaging sequence to generate a one-dimensional or two-dimensional navigator echo perpendicular to the direction of motion.

[0053] Step S30: Determine and reconstruct a dMRI optimized image based on the obtained gradient echo signal.

[0054] In this step, the reconstructed dMRI optimized image is determined based on the gradient echo signal obtained by sampling. It can be understood that when the gradient echo signal is the sampling signal corresponding to the initial area to be sampled, the reconstructed dMRI optimized image determined is the reconstructed dMRI optimized image corresponding to the initial sampling; and when the gradient echo signal is the sampling signal corresponding to the area to be sampled in the next sampling, the reconstructed dMRI optimized image determined is the under-sampling reconstructed dMRI image corresponding to the next sampling.

[0055] Exemplarily, a 3D dMRI image reconstruction algorithm can be used to reconstruct a 3D dMRI image based on the currently sampled portion of k-space data. Exemplarily, the reconstructed dMRI optimized image can be calculated using a 3D dMRI image reconstruction algorithm based on the obtained gradient echo signal; the reconstruction of the undersampled dMRI image can be optimized by using a gradient descent method based on the following formula:

[0056] ;

[0057] in, To optimize the image for dMRI reconstruction, To optimize the reconstruction of dMRI images, is the number of completed excitation samplings, is the number of coil channels of the magnetic resonance system, is the gradient echo signal collected by the j-th coil after the i-th excitation, F is the Fourier transform, is the sensitivity coefficient of the jth coil, is the complex conjugate of the navigator echo phase of the i-th excitation, is the k-space sampling mask for the i-th excitation, and α is the regularization coefficient. In the above calculation formula, Represents the error between the predicted sampling signal and the actual sampling signal. By minimizing this term, the reconstructed image can be consistent with the actual sampling result; is a regularization term to ensure the sparsity of the reconstruction results in k-space and avoid unstable reconstruction results under under-sampling conditions.

[0058] Step S40: inputting the reconstructed dMRI optimized image into a trained full-sampling dMRI image prediction model to predict a full-sampling dMRI image, and calculating a reconstructed image error map based on the reconstructed dMRI optimized image and the predicted full-sampling dMRI image.

[0059] In this step, the fully sampled dMRI image prediction result corresponding to the reconstructed dMRI optimized image obtained in step S30 is further predicted based on the pre-trained fully sampled dMRI image prediction model. In this embodiment, the reconstructed dMRI optimized image serves as the input data of the fully sampled dMRI image prediction model, and the fully sampled dMRI image serves as the output data of the fully sampled dMRI image prediction model.

[0060] Specifically, to obtain a trained fully sampled dMRI image prediction model, the deep learning-based magnetic resonance signal adaptive sampling method may further include the following steps: determining an initial fully sampled dMRI image prediction model, a model loss function, and a sample dataset, wherein the sample data in the sample dataset includes reconstructed dMRI optimized image sample data and fully sampled dMRI image sample data; and pre-training the initial fully sampled dMRI image prediction model based on the model loss function and the sample dataset to obtain a trained fully sampled dMRI image prediction model. Specifically, the training of the fully sampled dMRI image prediction model is implemented based on paired undersampled and fully sampled 3D dMRI image datasets, and the MSE loss function and Adam optimizer may also be used to guide the model weight optimization.

[0061] Exemplarily, the fully sampled dMRI image prediction model includes an encoder and a decoder. The encoder includes an initial convolution block and five downsampling blocks, and the decoder includes five upsampling blocks and an output layer. Furthermore, the initial convolution block includes two 3×3×3 convolution layers, and each downsampling block includes a 2×2×2 max pooling layer and two 3×3×3 convolution layers.

[0062] Figure 3 FIG. 1 is a schematic diagram of the architecture of a full-sampling dMRI image prediction model according to an embodiment of the present application. Figure 3As shown in the figure, this fully sampled dMRI image prediction model is based on the classic 3D U-Net architecture. It adopts a symmetrical encoder-decoder structure. The encoder consists of an initial convolutional block and five downsampling blocks, and the decoder consists of five upsampling blocks and a final output layer. The initial convolutional block consists of two 3×3×3 convolutional layers, which inputs a 512×512×512 single-channel 3D reconstructed dMRI optimized image and outputs 32-channel features of the same resolution. Each downsampling block consists of a 2×2×2 max pooling layer and two 3×3×3 convolutional layers. With each downsampling block, the number of feature channels doubles and the spatial resolution is halved. After passing through all downsampling blocks, the feature resolution is 16×16×16, with 1024 channels. The features are then input into the decoder. In each upsampling block in the decoder, deconvolution is first used to double the feature resolution and halve the number of channels. The feature is then concatenated with the corresponding layer features of the encoder and passed through two 3×3×3 convolutional layers to halve the number of channels again. The output convolution block consists of two 1×1×1 convolutional layers, reducing the number of channels back to a single channel, which serves as the final output of the model.

[0063] In addition, in this step, a reconstructed image error map is further calculated between the reconstructed dMRI optimized image and the fully sampled dMRI image. Exemplarily, calculating the reconstructed image error map based on the reconstructed dMRI optimized image and the predicted fully sampled dMRI image may include the following steps: calculating the absolute difference between each voxel signal of the reconstructed dMRI optimized image and the predicted fully sampled dMRI image; and determining the reconstructed image error map based on the absolute difference between each voxel signal.

[0064] Step S50: obtaining a sampling value of each sampling point in the sampling space based on the reconstructed image error map, and determining a region to be sampled for the next sampling based on the sampling value of each sampling point and the sampling value threshold.

[0065] In this step, the sampling value of each sampling point is further determined based on the reconstructed image error map. The sampling value is used to evaluate whether each sampling point is sampled in the next sampling cycle.

[0066] Exemplarily, obtaining the sampling value of each sampling point in the sampling space based on the reconstructed image error map may include: performing a Fourier transform on the reconstructed image error map to obtain a k-space error map; calculating the average difference in each frequency encoding direction in the k-space error map; and using each of the average differences as the sampling value of each sampling point. In this embodiment, the reconstructed error map is Fourier transformed to obtain the k-space error map; the k-space error map is averaged along the frequency encoding direction, and the calculated result is used as the sampling value of each sampling point in the sampling space.

[0067] After calculating the sampling value of each sampling point, the region to be sampled for the next sampling cycle is further determined based on the actual sampling values ​​of all unsampled sampling points. When the sampling value threshold is T, if the sampling values ​​of all unsampled sampling points in the sampling space are less than T, sampling is stopped, and the reconstructed dMRI optimized image obtained in the previous sampling cycle becomes the final dMRI image. Furthermore, in some embodiments, if the maximum number of excitations is set to M, even if there are sampling points in the sampling space with a sampling value greater than the sampling value threshold T, if the number of completed excitations reaches M, sampling is stopped. Similarly, the reconstructed dMRI optimized image obtained in the previous sampling cycle becomes the final dMRI image. It is understood that if there are sampling points in the sampling space with a sampling value greater than the sampling value threshold T and the number of completed excitations has not reached M, the k-space trajectories corresponding to the n sampling points with the largest sampling values ​​in the sampling space are selected as the region to be sampled for the next sampling cycle, and steps S20 to S50 are repeatedly executed to complete adaptive 3D k-space sampling.

[0068] In the above-described embodiment, the deep learning-based adaptive sampling method for magnetic resonance signals utilizes a deep learning model to evaluate the sampling value of each region of the sampling space based on the sampled portion of the magnetic resonance signal during the acquisition process, and to plan subsequent sampling points in real time. Specifically, during the magnetic resonance signal acquisition process, the present application dynamically plans subsequent sampling strategies in real time to maximize the value of the acquired signals in magnetic resonance image reconstruction, thereby reducing the number of samples required and the scan time for 3D high-resolution dMRI, minimizing the risk of motion artifacts, and improving the acquisition efficiency of 3D high-resolution dMRI.

[0069] In a specific embodiment, the method for adaptive sampling of magnetic resonance signals based on deep learning specifically includes the following steps: (1) determining an initial area to be sampled in a sampling space; (2) sampling the area to be sampled determined in the previous step in a single excitation based on magnetic resonance signal acquisition; (3) reconstructing a magnetic resonance image based on the currently sampled magnetic resonance signal; (4) using a deep learning model to predict a full sampling image based on the magnetic resonance image reconstructed in the previous step, and calculating a reconstruction error map; (5) evaluating the sampling value of each area in the unsampled part of the sampling space based on the reconstruction error map; (6) if the maximum sampling value of each area in the unsampled part of the sampling space is less than a sampling value threshold or the number of sampling times reaches a preset maximum number of excitations, then stopping sampling and outputting the current magnetic resonance image; otherwise, confirming the next area to be sampled in the sampling space based on the sampling value of each area, and returning to step (3) to resample the newly created area to be sampled.

[0070] Through the above examples, it can be seen that the deep learning-based adaptive sampling method of magnetic resonance signals in this application evaluates and prioritizes the regions in k-space that are expected to contribute most to reconstruction error based on the acquired magnetic resonance signals during the sampling process, thereby improving the accuracy of reconstructed dMRI images at the same sampling rate. At the same time, by reducing invalid sampling steps, this method can further shorten scan time and reduce the risk of artifacts caused by user motion or physiological noise, providing greater robustness and efficiency for the application of 3D high-resolution dMRI in complex scenarios.

[0071] Correspondingly, the present invention also provides a deep learning-based adaptive sampling system for magnetic resonance signals, the system comprising a processor, a memory, and a computer program stored in the memory, the processor being used to execute the computer program, and when the computer program is executed, the system implements the steps of the method described in any of the above embodiments.

[0072] Embodiments of the present invention further provide a computer-readable storage medium and a computer program product, each having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the steps of the method described in any of the above embodiments. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the art.

[0073] It should be understood by those skilled in the art that the various exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether to implement the system in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via a data signal carried in a carrier wave.

[0074] It should be understood that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted. In the above embodiments, several specific steps are described and illustrated as examples. However, the method of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art may make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present invention.

[0075] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or replace features of other embodiments.

[0076] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations to the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for adaptive sampling of magnetic resonance signals based on deep learning, characterized in that: The method comprises: Determining a sampling space, a sampling value threshold, and the number of sampling points for a single shot, and selecting a K-space trajectory of a corresponding number of sampling points at the center of the sampling space as an initial area to be sampled based on the number of sampling points for a single shot; Determining a 3D EPI sequence corresponding to the single excitation based on phase encoding of sampling points within the area to be sampled, and performing a single radio frequency excitation based on the 3D EPI sequence corresponding to the single excitation to obtain a corresponding gradient echo signal; Determining to reconstruct a dMRI optimized image based on the obtained gradient echo signal; Inputting the reconstructed dMRI optimized image into a trained full-sampling dMRI image prediction model to predict a full-sampling dMRI image, and calculating a reconstructed image error map based on the reconstructed dMRI optimized image and the predicted full-sampling dMRI image; Obtaining a sampling value of each sampling point in the sampling space based on the reconstructed image error map, and determining a region to be sampled for the next sampling based on the sampling value of each sampling point and the sampling value threshold; Determining and reconstructing a dMRI optimized image based on the obtained gradient echo signal includes: Calculating the reconstructed dMRI optimized image based on the obtained gradient echo signal using a 3D dMRI image reconstruction algorithm; The calculation formula of the 3DdMRI image reconstruction algorithm is: ； in, To optimize the image for dMRI reconstruction, To optimize the reconstruction of dMRI images, is the number of completed excitation samplings, is the number of coil channels of the magnetic resonance system, is the gradient echo signal collected by the j-th coil after the i-th excitation, F is the Fourier transform, is the sensitivity coefficient of the jth coil, is the complex conjugate of the navigator echo phase of the i-th excitation, is the k-space sampling mask for the i-th excitation, and α is the regularization coefficient.

2. The method for adaptive sampling of magnetic resonance signals based on deep learning according to claim 1, characterized in that: The method further comprises: Determining an initial fully sampled dMRI image prediction model, a model loss function, and a sample data set, wherein the sample data in the sample data set includes reconstructed dMRI optimized image sample data and fully sampled dMRI image sample data; The initial full-sampling dMRI image prediction model is pre-trained based on the model loss function and the sample data set to obtain a trained full-sampling dMRI image prediction model.

3. The method for adaptive sampling of magnetic resonance signals based on deep learning according to claim 2, characterized in that: The fully sampled dMRI image prediction model includes an encoder and a decoder, wherein the encoder includes an initial convolution block and five downsampling blocks, and the decoder includes five upsampling blocks and an output layer.

4. The method for adaptive sampling of magnetic resonance signals based on deep learning according to claim 3, characterized in that: The initial convolution block includes two 3×3×3 convolution layers, and each downsampling block includes a 2×2×2 maximum pooling layer and two 3×3×3 convolution layers.

5. The method for adaptive magnetic resonance signal sampling based on deep learning according to claim 1, characterized in that: A reconstructed image error map is calculated based on the reconstructed dMRI optimized image and the predicted fully sampled dMRI image, including: Calculating the absolute difference between the voxel signals of the reconstructed dMRI optimized image and the predicted full-sampled dMRI image; The reconstructed image error map is determined based on the absolute difference of the signals of each voxel.

6. The method for adaptive sampling of magnetic resonance signals based on deep learning according to claim 1, characterized in that: Obtaining the sampling value of each sampling point in the sampling space based on the reconstructed image error map includes: Performing Fourier transform on the reconstructed image error map to obtain a K-space error map; Calculate the average difference in each frequency encoding direction in the k-space error map; The average difference values ​​are used as the sampling values ​​of each sampling point.

7. A deep learning-based magnetic resonance signal adaptive sampling system, the system comprising a processor, a memory, and a computer program stored in the memory, characterized in that: The processor is configured to execute the computer program. When the computer program is executed, the system implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images