Method and system for generating magnetic resonance images, computer readable storage medium

Abstract

Embodiments of the present application provide a method and system for generating a magnetic resonance image, and a computer readable storage medium. The method comprises: acquiring a plurality of quantitative maps; synthesizing a first weighted image based on the plurality of quantitative maps; and converting the first weighted image to a corresponding second weighted image based on a trained deep learning network.

Description

TECHNICAL FIELD

[0001] Embodiments disclosed herein relate to medical imaging technology, and more particularly, to a method and system for generating magnetic resonance images, and a computer readable storage medium. BACKGROUND

[0002] Magnetic resonance imaging (MRI) is a medical imaging modality that can obtain three-dimensional images of a human body without using X-ray or other ionizing radiation.

[0003] MRI utilizes a magnet with a strong magnetic field to generate a static magnetic field B0. When a part of a human body to be imaged is positioned in the static magnetic field B0, the nuclear spins associated with hydrogen nuclei in the tissues of the part of the human body to be imaged are polarized, causing the tissues of the part of the human body to be imaged to have a longitudinal magnetization vector in a macroscopic sense, in an equilibrium state. When a radio frequency field B1 intersecting the direction of the static magnetic field B0 is applied, the direction of the proton rotation changes, the longitudinal magnetization vector decays, and the tissues of the part of the human body to be imaged have a transverse magnetization vector in a macroscopic sense.

[0004] After the radio frequency field B1 is removed, the longitudinal magnetization gradually recovers to the equilibrium state, and the transverse magnetization vector decays in a spiral shape until it recovers to zero. During the recovery of the longitudinal magnetization and the decay of the transverse magnetization, a magnetic resonance signal is generated, which can be acquired, and based on the acquired signal, an image of the tissues of the part of the human body to be imaged can be reconstructed.

[0005] The length of time for the longitudinal magnetization vector to recover to the equilibrium state is generally referred to as the longitudinal relaxation time, i.e., T1. The length of time for the transverse magnetization vector to decay to 0 is generally referred to as the transverse relaxation time, i.e., T2. Different tissues of the human body generally have different T1 and different T2. In addition, different tissues of the human body also generally have different proton densities (PD).

[0006] During a magnetic resonance scan, by adjusting the scan parameters (or scan sequences), images highlighting the T1 contrast between tissues (T1-weighted images, T1WI), images highlighting the T2 contrast between tissues (T2-weighted images, T2WI), and images highlighting the proton density contrast between tissues (PD-weighted images, such as Flair) can be obtained.

[0007] In order to meet the needs of clinical diagnosis, it is often necessary to perform multiple scan sequences for the same imaging site (e.g., the brain) to obtain the above different types of weighted images, respectively. This makes the magnetic resonance examination usually take a long time. SUMMARY

[0008] An aspect of the present disclosure provides a method of generating a magnetic resonance image, including: acquiring a plurality of quantitative maps; synthesizing a first weighted image based on the plurality of quantitative maps; and converting the first weighted image into a corresponding second weighted image based on a trained deep learning network.

[0009] In another aspect, acquiring the plurality of quantitative maps includes calculating the plurality of quantitative maps based on an initial image, the initial image being generated by a magnetic resonance imaging device performing a pre-set scan sequence.

[0010] In another aspect, the plurality of quantitative maps includes a T1 quantitative map, a T2 quantitative map, and a proton density quantitative map.

[0011] In another aspect, synthesizing the first weighted image includes determining the first weighted image according to a pre-set relationship between quantitative values in the plurality of quantitative maps and hypothetical sequence parameters and image signal values.

[0012] In another aspect, the hypothetical sequence parameters include some or all of a hypothetical echo time, a hypothetical repetition time, and a hypothetical inversion recovery time.

[0013] In another aspect, the first weighted image is a T1 weighted image, a T2 weighted image, or a proton density weighted image, the first weighted image being obtained based on the following equation:

[0014] S = PD · exp(-TE / T2) · (1 - exp(-TR / T2));

[0015] where S is an image signal value of the first weighted image, exp is an exponential function with a natural constant e as a base, TE is a hypothetical echo time, TR is a hypothetical repetition time, T1, T2, and PD are a T1 quantitative value, a T2 quantitative value, and a proton density quantitative value, respectively.

[0016] In another aspect, the first weighted image is a T1WI-Flair image, a T2WI-Flair image, a STIR image, or a PSIR image, the first weighted image being obtained based on the following equation:

[0017] S = PD · exp(-TE / T2) · (1 - 2 · exp(-TI / T1) + exp(-TR / T1);

[0018] where S is an image signal value of the first weighted image, exp is an exponential function with a natural constant e as a base, TE is a hypothetical echo time, TR is a hypothetical repetition time, TI is a hypothetical inversion recovery time, T1, T2, and PD are a T1 quantitative value, a T2 quantitative value, and a proton density quantitative value, respectively.

[0019] In another aspect, the step of synthesizing the first weighted image further comprises: performing a convolution operation on the plurality of quantitative maps respectively, and synthesizing the first weighted image based on results of the convolution operation.

[0020] In another aspect, the input data set for training the deep learning network comprises a plurality of first weighted images, and the output data set for training the deep learning network comprises a plurality of real weighted images obtained by performing a magnetic resonance imaging scan and reconstructing images based on scan data.

[0021] In another aspect, the network parameters of the deep learning network are obtained by solving the following equation: min θ || f (θ) - f || 2; where θ is the network parameters, f comprises the real weighted images, f (θ) represents the output of the deep learning network, and min represents minimization.

[0022] In another aspect, the present disclosure also provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed, performs the method of any of the above aspects.

[0023] In another aspect, the present disclosure also provides a system for generating a magnetic resonance image, comprising: a quantitative map obtaining module configured to obtain a plurality of quantitative maps; an image synthesizing module configured to synthesize a first weighted image based on the plurality of quantitative maps; and a conversion module configured to convert the first weighted image to a corresponding second weighted image based on a trained deep learning network.

[0024] In another aspect, the system further comprises: a magnetic resonance imaging device configured to generate an initial image; and a computing module configured to calculate the plurality of quantitative maps based on the initial image.

[0025] It should be understood that the brief summary above is provided merely for purposes of summarizing some concepts of the disclosure. It is not meant to identify key or essential features of the claimed subject matter, the scope of which is defined solely by the claims that follow the detailed description. Furthermore, the claimed subject matter is not limited to implementations that solve any or all disadvantages noted in the background. BRIEF DESCRIPTION OF DRAWINGS

[0026] The present disclosure will be better understood with reference to the following non-limiting examples, which are described in conjunction with the accompanying drawings, wherein:

[0027] Figure 1 A flowchart of a method for generating a magnetic resonance image according to an embodiment of the present disclosure is shown;

[0028] Figure 2 To utilize Figure 1 An example image generated by the method shown;

[0029] Figure 3A structural schematic diagram of a magnetic resonance imaging device is shown;

[0030] Figure 4 A block diagram of a system for generating magnetic resonance images according to an embodiment of the present application is shown;

[0031] Figure 5 , Figure 6 A brain T1WI image generated according to an embodiment of the present application and an actually scanned reconstructed brain T1WI image are shown respectively;

[0032] Figure 7 , Figure 8 A brain T2WI image generated according to an embodiment of the present application and an actually scanned reconstructed brain T2WI image are shown respectively;

[0033] Figure 9 , Figure 10 A brain T2W-Flair image generated according to an embodiment of the present application and an actually scanned reconstructed brain T2W-Flair image are shown respectively;

[0034] Figure 11 , Figure 12 A prostate T2WI image generated according to an embodiment of the present application and an actually scanned reconstructed prostate T2WI image are shown respectively. DETAILED DESCRIPTION

[0035] Various embodiments described below include methods and systems for generating magnetic resonance images, and computer-readable storage media.

[0036] Figure 1 A flowchart of one embodiment of the method is shown, Figure 2 An example image generated using the method is shown. Reference is made to Figure 1 , Figure 2 In step S13, a plurality of quantitative maps are obtained. Specifically, step S13 can include a step of calculating quantitative maps based on an initial image, which is produced by a magnetic resonance imaging device performing a scan sequence, which can be a phantom or a real image, the techniques of performing a scan sequence and reconstructing a magnetic resonance image by a magnetic resonance imaging device will be described below in conjunction with Figure 3 .

[0037] For example, a feature parameter (or quantitative value) of a corresponding voxel can be calculated by a signal value of each pixel in the initial image data, and the distribution of the feature parameter on the image forms its quantitative map. In one embodiment, the quantitative value or feature parameter can include T1, T2 and proton density.

[0038] The scan sequence performed when the initial image is generated for obtaining the quantitative map can be various, for example, can be a spin echo (SE) sequence, a fast spin echo (FSE) sequence, a gradient echo (GE) sequence, an inversion recovery (IR) sequence, a turbo field echo (TFE) sequence, or a combination of two or more sequences selected from the above sequences. Using the equation relationship between the signal value in the obtained initial image pixel and the sequence parameters, quantitative values such as T1, T2, and proton density can be solved.

[0039] T1, T2, and proton density can be calculated based on the magnetic resonance image using known techniques to obtain the corresponding T1 quantitative map, T2 quantitative map, and proton density quantitative map, which will not be described here.

[0040] In other embodiments, known quantitative maps can be stored in a storage space to be able to be called for performing embodiments of the present application.

[0041] With reference to the foregoing description Figure 1 , Figure 2 In step S15, a first weighted image is synthesized based on the plurality of quantitative maps.

[0042] In embodiments of the present application, the first weighted image can be determined according to a preset relationship between the quantitative values at the corresponding positions in the plurality of quantitative maps and the assumed sequence parameters and the weighted image signal values, and the first weighted image can include any one of a T1 weighted image T1WI, a T2 weighted image T2WI, a proton density weighted image PDWI, a T1W-Flair image, a T2W-Flair image, a STIR image, a PSIR image, a PSIR-vossel image. The preset relationship can be described by an image synthesis formula. The assumed sequence parameters can include part or all of an assumed echo time (TE), an assumed repetition time (TR), and an assumed inversion recovery time (TI).

[0043] Since the real scan sequence is not performed in the process of generating the weighted image of each quantitative (or characteristic parameter) using the quantitative map, the weighted image of the quantitative parameter is generated by the preset relationship between the assumed sequence parameters of the scan sequence and the quantitative, and by changing the assumed values of the sequence parameters, images highlighting the contrast of different quantifications can be obtained.

[0044] In one specific embodiment, in step S15, the first weighted image is synthesized based on the following formula (1).

[0045] S = PD • exp(-TE / T2) • (1 - exp(-TR / T2)), (1);

[0046] where S is a signal value of the first weighted image, the S value is calculated for each pixel based on the above formula, and a T1 WI image or a T2 WI image can be obtained, for example, exp is an exponential function with the natural constant e as the base, TE is an assumed echo time, TR is an assumed repetition time, T1, T2, and PD are T1 quantitative value, T2 quantitative value, and proton density quantitative value, respectively.

[0047] When a small TE value (for example, 10 milliseconds) and a small TR value (for example, 500 milliseconds) are set for formula (1), a T1 WI image is obtained. This image can have similar image features to an image reconstructed by a scanning sequence performing T1 weighting, for example, a dark region of a water-containing tissue region such as cerebrospinal fluid.

[0048] When a large TE value (for example, 100 milliseconds) and a large TR value (for example, 4500 milliseconds) are set for formula (1), a T2 WI image is obtained. This image can have similar image features to an image reconstructed by a scanning sequence performing T2 weighting, for example, a high-light region of a water-containing tissue region such as cerebrospinal fluid.

[0049] When a small TE value (for example, 10 milliseconds) and a large TR value (for example, 8000 milliseconds) are set for formula (1), a PD WI image is obtained. This image can have similar image features to an image reconstructed by a scanning sequence performing proton density weighting, for example, the more hydrogen protons a tissue contains, the stronger the image signal.

[0050] There are various scanning sequences for T1 weighting, T2 weighting, and proton density weighting in the prior art, which will not be described here.

[0051] In step S15, the first weighted image can also be synthesized based on the following formula (2), and specifically, a T1W-Flair (FLuid Attenuated Inversion Recovery) image, a T2W-Flair image, a STIR (Short T1 Inversion Recovery) image, a PSIR (Phase Sensitive Inversion Recovery) image, or a PSIR vossel image, etc. can be synthesized.

[0052] S = PD · exp(-TE / T2) · (1-2 · exp(-TI / T1) + exp(-TR / T1), (2);

[0053] where S is the signal value of the first weighted image, exp is the exponential function with base the natural constant e, TE is the assumed echo time, TR is the assumed repetition time, and TI is the assumed inversion recovery time.

[0054] When a small TE value (e.g. 10 milliseconds) and a small TR value (e.g. 2500 milliseconds) are set for equation (2) and a suitable TI value (e.g. 1050 milliseconds) is set, the resulting image is a T1W-Flair. This image can have similar image characteristics as an image reconstructed by performing a T1W-Flair scan sequence.

[0055] When a large TE value (e.g. 100 milliseconds) and a large TR value (e.g. 15000 milliseconds) are set for equation (2) and a suitable TI value (e.g. 3000 milliseconds) is set, the resulting image is a T2W-Flair. This image can have similar image characteristics as an image reconstructed by performing a T2W-Flair scan sequence.

[0056] When a large TE value (e.g. 100 milliseconds) and a large TR value (e.g. 15000 milliseconds) are set for equation (2) and a small TI value (e.g. 300 milliseconds) is set, the resulting image S is an STIR image, which can have similar image characteristics as an image reconstructed by performing an STIR scan sequence, e.g. with fat suppression characteristics.

[0057] When a small TE value (e.g. 10 milliseconds) and a large TR value (e.g. 6000 milliseconds) are set for equation (2) and a small TI value (e.g. 500 milliseconds) is set, the resulting image is a PSIR image, which can have similar image characteristics as an image reconstructed by performing a PSIR scan sequence.

[0058] When a small TE value (e.g. 10 milliseconds) and a large TR value (e.g. 8000 milliseconds) are set for equation (2) and a small TI value (e.g. 10 milliseconds) is set, the resulting image is a PSIR vossel image, which can have similar image characteristics as an image reconstructed by performing a PSIR vossel scan sequence.

[0059] There are also a variety of T1W-Flair, T2W-Flair, STIR, PSIR and PSIR vossel scan sequences in the prior art, which are not described here again.

[0060] Optionally, in order to perform appropriate feature extraction, the T1, T2 and PD quantitative maps obtained in step S13 are further subjected to convolution operation before being input into the image synthesis formula, and in step S15, the T1, T2 and PD quantitative maps after convolution operation are input into the corresponding image synthesis formula to synthesize the first weighted image.

[0061] For example, the obtained T1, T2 and PD quantitative maps have a size of 512*512, the input of the convolution operation can be 512*512*1 (where 1 is the number of input channels or image thickness or number of layers), after convolution operation with a convolution kernel having a size of 5*5*1*64 (where 5, 5 and 1 are the length, width and thickness of the convolution kernel respectively, and 64 is the number of convolution kernels), the obtained feature image is 512*512*64 (where 512, 512 and 64 are the length, width and thickness of the output image respectively), which is input into formula (1) or (2) or other appropriate formula, and after setting the values of the sequence parameters in the formula, the corresponding first weighted image can be output.

[0062] As described above, an initial image can be obtained by a single short-time magnetic resonance scanning procedure, and based on the initial image, a quantitative value related to each pixel can be calculated, in step S15, based on the preset relationship between the quantitative value, the specific weighted image and the sequence parameter, the assumed sequence parameter value can be set, and the first weighted image can be generated. As described before, the generated first weighted image can have similar image features as the reconstructed image obtained after performing a specific scanning sequence, however, due to the complexity of magnetic resonance imaging, the actual magnetic resonance imaging result can be affected by factors other than the above-mentioned sequence parameters, so the first weighted image can still have a large difference from the actual scanning image, making it difficult to be directly used for clinical observation.

[0063] Therefore, as shown in FIG. 1, Figure 1 , Figure 2 the method of the present embodiment further comprises step S17: converting each first weighted image into a corresponding second weighted image based on the trained deep learning network.

[0064] Wherein, the input data set for training the deep learning network can include a plurality of first weighted images obtained according to step S15, and the output data set for training the deep learning network can include real weighted images obtained by performing magnetic resonance imaging scanning and reconstructing images based on scanning data, such as real T1 weighted images, real T2 weighted images, real proton density weighted images, real T1W-Flair, T2W-Flair, STIR, PSIR, etc.

[0065] Specifically, this deep learning network can be trained based on the ADAM (Adaptive Moment Estimation) optimization method or other well-known models. Once the deep learning network is created or trained, simply inputting each of the aforementioned first weighted images into the network will yield a first weighted image that more closely resembles the actual scanned image.

[0066] The aforementioned deep learning network may include an input layer, an output layer, and a processing layer (or hidden layer). The input layer is used to preprocess the input data or image, such as removing the mean, normalizing, or reducing dimensionality. The processing layer may include multiple convolutional layers for feature extraction and an activation layer that performs a nonlinear mapping on the output of the convolutional layers.

[0067] Each convolutional layer includes a number of neurons, and the number of neurons in multiple convolutional layers can be the same or different as needed. Based on the first weighted image (known input) and the true weighted image (desired output), by setting the number of processing layers in the network and the number of neurons in each processing layer, and estimating (or adjusting or calibrating) the network's weights and / or biases, the mathematical relationship between the known input and the desired output and / or the mathematical relationship between the input and output of each layer can be identified and characterized.

[0068] Specifically, when the number of neurons in one layer is n, and the corresponding values ​​of these n neurons are X1, X2, ... X n The number of neurons in the next layer connected to one of the layers is m, and the values ​​of these m neurons are Y1, Y2, ... Y. m Then the relationship between the two adjacent layers can be represented as:

[0069]

[0070] Among them, X i Y represents the value corresponding to the i-th neuron in the previous layer. j W represents the value corresponding to the j-th neuron in the next layer. ji B represents the weight. j This represents the deviation. In some embodiments, the function f is a corrected linear function.

[0071] Therefore, by adjusting the weight W ji and / or deviation B j This allows us to identify the mathematical relationship between the input and output of each layer, causing the loss function to converge, thus training the aforementioned deep learning network.

[0072] In this embodiment, the network parameters of the deep learning network are obtained by solving the following equation (3):

[0073] min θ ||f(θ) - f||2, (3)

[0074] where θ represents the network parameters of the deep learning network, which can include the above-mentioned weights W ji and / or biases B j , f includes the real weighted image, such as the above-mentioned one real weighted image of a person, f(θ) represents the output of the deep learning network, and min represents minimization. The network parameters are set by minimizing the difference between the network output image and the actual scan image to construct the deep learning network.

[0075] In one embodiment, although the configuration of the deep learning network will be guided by the prior knowledge of the estimation problem, the dimensions of the input, output, etc., relying on or specifically according to the input data to achieve the best approximation of the required output data. In various alternative implementations, certain aspects and / or features of the data, imaging geometry, reconstruction algorithm, etc. can be utilized to give explicit meaning to certain data representations in the deep learning network, which can help to speed up the training. Because this creates an opportunity to train (or pre-train) or define certain layers separately in the deep learning network.

[0076] In some embodiments, the above-mentioned trained network is trained based on a training module on an external carrier (e.g., a device other than a medical imaging system). In some embodiments, the training system can include a first module for storing a training data set, a second module for training and / or updating based on a model, and a communication network for connecting the first module and the second module. In some embodiments, the first module includes a first processing unit and a first storage unit, wherein the first storage unit is used to store the training data set, and the first processing unit is used to receive relevant instructions (e.g., to obtain the training data set) and to send the training data set according to the instructions. In addition, the second module includes a second processing unit and a second storage unit, wherein the second storage unit is used to store the training model, and the second processing unit is used to receive relevant instructions, to train and / or update the network, etc. In other embodiments, the training data set can also be stored in the second storage unit of the second module, and the training system can not include the first module. In some embodiments, the communication network can include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0077] Once the data (e.g., trained network) is generated and / or configured, the data can be copied and / or loaded into a medical imaging system (e.g., a magnetic resonance imaging system as will be described below), which can be done in different ways. For example, the model can be loaded through a direct connection or link between the medical imaging system and the computer. In this regard, communication between different elements can be done using available wired and / or wireless connections and / or according to any suitable communication (and / or network) standard or protocol. Alternatively or additionally, the data can be loaded into the medical imaging system indirectly. For example, the data can be stored into a suitable machine-readable medium (e.g., a flash card or the like) and then the data is loaded into the medical imaging system (on-site, such as by a user or authorized personnel of the system) using the medium, or the data can be downloaded into an electronic device (e.g., a laptop or the like) that is capable of communicating locally and then the device is used (e.g., by a user or authorized personnel of the system) on-site to upload the data into the medical imaging system via a direct connection (e.g., a USB connector or the like).

[0078] As discussed herein, deep learning techniques (also referred to as deep machine learning, hierarchical learning, or deep structured learning, among others) employ artificial neural networks for learning. Deep learning methods are characterized by the use of one or more network architectures to extract or model data of interest. Deep learning methods can be accomplished using one or more processing layers (e.g., input layers, output layers, convolutional layers, normalization layers, sampling layers, and so on, which can have different numbers and functionalities depending on the different deep network models) where the configuration and number of layers allow the deep network to handle complex information extraction and modeling tasks. The particular parameters (which can also be referred to as "weights" or "biases") of the network are typically estimated through a so-called learning process (or training process). The learned or trained parameters typically result in (or output) a network corresponding to different levels of layers, so that the output of a previous layer or different aspects of the initial data that are extracted or modeled can typically represent the hierarchy or cascade of layers. In the context of image processing or reconstruction, this can be characterized as different layers with respect to different feature levels in the data. Thus, the processing can be done hierarchically, i.e., earlier or higher level layers can correspond to extracting "simple" features from the input data, followed by layers that combine these simple features into features that exhibit higher complexity. In practice, each layer (or more specifically, each "neuron" in a layer) can employ one or more linear and / or non-linear transformations (so-called activation functions) to process the input data into an output data representation. The number of "neurons" can be constant between layers, or can vary from layer to layer.

[0079] As discussed in this paper, as part of the initial training of a deep learning process for solving a specific problem, the training dataset includes known input values ​​and the expected (target) output value of the final output of the deep learning process. In this way, the deep learning algorithm can process this training dataset (in a supervised or guided manner, or in an unsupervised or unguided manner) until it identifies the mathematical relationship between the known inputs and the expected output and / or identifies and represents the mathematical relationship between the inputs and outputs of each layer. The learning process typically utilizes (partial) input data and creates a network output for that input data, then compares the created network output with the expected output of the dataset, and then iteratively updates the network parameters (weights and / or biases) using the difference between the created and expected outputs. The network parameters can typically be updated using the Stochastic Gradient Descent (SGD) method; however, those skilled in the art will understand that other methods known in the art can also be used to update the network parameters. Similarly, a separate validation dataset can be used to validate the trained network, where both the known input and the desired output are known. The network output can be obtained by feeding the known input to the trained network, and then the network output is compared with the (known) desired output to validate previous training and / or prevent overtraining.

[0080] Continue to refer to Figure 1 , Figure 2 In one example, when the first T1 weighted image with a size of 512*512*64 is obtained using formula (1), it is input into the trained deep learning network mentioned above. Multiple convolution operations are performed using a convolution kernel with a size of 3*3*64*64, and then a convolution operation is performed using a convolution kernel with a size of 3*3*64*1, resulting in a second T1 weighted image with a size of 512*512*1.

[0081] Similarly, when the first Flair image with a size of 512*512*64 is obtained using formula (2), it is input into the trained deep learning network described above. Multiple convolution operations are performed using a convolution kernel with a size of 3*3*64*64, and then a convolution operation is performed using a convolution kernel with a size of 3*3*64*1, resulting in a second Flair weighted image with a size of 512*512*1.

[0082] The aforementioned second T1-weighted image and second Flair image have optimized image quality, and can be used directly for clinical observation, for example, because they are closer to the actual scan images.

[0083] Figure 3A schematic diagram of a structure of a magnetic resonance imaging apparatus is shown. The system, as one example, can be used to execute a scan sequence to generate the initial image described above, can also be used to execute a scan sequence to generate a true weighted image, and can also be used to store or transfer the generated image to other systems.

[0084] The magnetic resonance imaging apparatus 300 includes a scanner 310. The scanner 310 is used to perform a magnetic resonance scan on a subject (e.g., a human body) 36 to generate image data of a region of interest of the subject 36, which can be a predetermined imaging site or imaging tissue.

[0085] The magnetic resonance imaging apparatus 300 can also include a controller 320 coupled to the scanner 310 for controlling the scanner 310 to perform the procedure of the magnetic resonance scan described above. Specifically, the controller 320 can send sequence control signals to relevant components of the scanner 310 (e.g., a radio frequency generator, a gradient coil driver, etc. to be described below) through a sequence generator (not shown) so that the scanner 310 performs a predetermined scan sequence.

[0086] Those skilled in the art can understand that the "scan sequence" described above refers to a combination of pulses with specific power, amplitude, width, direction, and timing applied when performing a magnetic resonance imaging scan, which can typically include, for example, radio frequency pulses and gradient pulses. The radio frequency pulses can include, for example, radio frequency transmit pulses for exciting protons in the human body to resonate. The gradient pulses can include, for example, slice selection gradient pulses, phase encoding gradient pulses, frequency encoding gradient pulses, etc. Typically, a plurality of scan sequences can be pre-set in a magnetic resonance system to enable selection of a sequence that is suitable for a clinical detection requirement, which can include, for example, an imaging site, an imaging function, etc.

[0087] In one example, the scanner 310 can include a main magnet assembly 311, a bed 312, a radio frequency generator 313, a radio frequency transmit coil 314, a gradient coil driver 315, a gradient coil assembly 316, a radio frequency power amplifier 319, and a data acquisition unit 317.

[0088] The main magnet assembly 311 typically includes a ring-shaped superconducting magnet defined within an outer housing, which is mounted within a ring-shaped vacuum container. The ring-shaped superconducting magnet and its outer housing define a cylindrical space that encircles the subject 36, as shown by the scan bore 318. The main magnet assembly 311 generates a constant magnetic field, i.e., a B0field, along the Z direction of the scan bore 318. Typically, a more uniform portion of the B0field is formed in a central region of the main magnet. Figure 3

[0089] ​The bed 312 is used to carry the object 36 and is controlled by the controller 320 to move along the Z direction to go in and out of the scanning cavity 318 described above, for example, in one embodiment, the imaging volume of the object 36 can be positioned to the central region of the scanning cavity where the magnetic field strength is more uniform, so as to facilitate the scanning imaging of the imaging volume of the object 36.

[0090] The magnetic resonance imaging device 300 uses the formed B0 field to transmit a static magnetic pulse signal to the object 36 located in the scanning cavity, so that the precession of the protons in the resonance volume in the object 36 is ordered, and a longitudinal magnetization vector is generated.

[0091] The radio frequency generator 313 and the radio frequency power amplifier 319 can be used as part of a radio frequency transmission link, wherein the radio frequency generator 313 is used to generate a radio frequency pulse in response to a control signal of the controller 320, which is usually a small radio frequency signal with small power, which can be amplified by the radio frequency power amplifier 319 and then applied to the radio frequency transmission coil 314.

[0092] The radio frequency transmission coil 314 can be connected to a transmit / receive (T / R) switch, and by controlling the T / R switch, the body coil can be switched between transmission and reception modes. In the transmission mode, the radio frequency transmission coil 314 is used to respond to the radio frequency excitation pulse described above to transmit a radio frequency field B1 orthogonal to the static magnetic field B0 to the object 36 to excite the atomic nucleus in the object 36, so that the longitudinal magnetization vector is converted into a transverse magnetization vector. In the reception mode, the body coil can be used to receive the magnetic resonance signal from the object 36.

[0093] When the radio frequency excitation pulse ends, a free induction decay signal is generated during the process in which the transverse magnetization vector of the object 36 gradually recovers to zero.

[0094] The gradient coil driver 315 is used to provide appropriate current / power to the gradient coil assembly 316 in response to the gradient pulse control signal or the shimming control signal issued by the controller 320.

[0095] The gradient coil assembly 316 is used to form a varying magnetic field in the imaging space to provide three-dimensional position information for the magnetic resonance signal described above, and is also used to generate a compensation magnetic field for the B0 field to shim the B0 field.

[0096] The gradient coil assembly 316 can include three gradient coils for producing magnetic field gradients that are tilted into three spatial axes (e.g., X, Y and Z axes) that are perpendicular to each other, respectively. More specifically, the gradient coil assembly 316 applies a magnetic field gradient in the slice selection direction (e.g., Z direction) to select a slice in the imaging volume. Those skilled in the art appreciate that the slice is any one of a plurality of two-dimensional slices distributed along the Z direction in a three-dimensional imaging volume, and the "Z direction" is usually the direction that extends from the head to the feet when a patient is positioned on the bed 312. The radio frequency transmit coil 314 transmits radio frequency excitation pulses to the slice of the imaging volume and excite the slice when the imaging is being scanned. The gradient coil assembly 316 applies a magnetic field gradient in the phase encoding direction (e.g., Y direction) to phase encode the magnetic resonance signals of the excited slice. The gradient coil assembly 316 applies a magnetic field gradient in the frequency encoding direction (e.g., X direction) of the subject 36 to frequency encode the magnetic resonance signals of the excited slice.

[0097] The data acquisition unit 317 is configured to acquire the magnetic resonance signals (e.g., received by the body coil or the surface coil) in response to the data acquisition control signals from the controller 320. In one embodiment, the data acquisition unit 317 can include, for example, a radio frequency preamplifier for amplifying the magnetic resonance signals, a phase detector for phase detecting the amplified magnetic resonance signals, and an analog / digital converter for converting the phase detected magnetic resonance signals from analog signals to digital signals.

[0098] The magnetic resonance imaging apparatus 300 includes an image reconstruction unit 330 that can reconstruct a series of two-dimensional slice images (i.e., the image sequence) of the imaging volume of the subject 36 based on the digitized magnetic resonance signals. Specifically, the reconstruction unit can communicate with the controller 320 to perform the image reconstruction as described above.

[0099] The magnetic resonance imaging apparatus 300 includes a processing unit 340 that can perform any desired image processing, such as image correction, determining display parameters of the images, etc., on any of the images in the image sequence. The image processing as described above can be any improvement or adaptation of the images in any aspect, such as contrast, uniformity, sharpness, brightness, etc. Specifically, the processing unit 340 can communicate with the controller 120 to perform the image processing as described above.

[0100] The magnetic resonance imaging apparatus 300 can include a display unit 350 that can be used to display the operating interface and various data or images generated during the data processing.

[0101] The magnetic resonance imaging apparatus 300 further includes a console 360, which can include user input devices such as a keyboard and a mouse, etc. The controller 320 can communicate with the scanner 310, the image reconstruction unit, the processing unit 340, the display unit 350, etc. in response to control commands generated by a user based on operation of the console 360 or an operation panel / keypad provided on the main magnet housing, etc.

[0102] In an embodiment, the controller 320, the image reconstruction unit 330, the processing unit 340 can respectively or collectively include a computer and a storage medium on which a predetermined control program, a data processing program to be executed by the computer are recorded, for example, the storage medium can store programs for implementing imaging scanning, image reconstruction, image processing, etc., for example, programs for implementing the method of generating a magnetic resonance image according to an embodiment of the present application can be stored. The above-mentioned storage medium can include, for example, a ROM, a floppy disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, or a non-volatile memory card.

[0103] The above-mentioned magnetic resonance imaging apparatus 300 is described only as an example, and in other embodiments, the apparatus can have various modifications as long as it can acquire image data from an imaging object.

[0104] Reference Figure 4 In addition, an embodiment of the present application can also provide a system for generating a magnetic resonance image, which includes a quantitative map acquisition module 43, an image synthesis module 45, and an optimization module 47.

[0105] The quantitative map acquisition module 43 is used to acquire a quantitative map. In an example, the quantitative map acquisition module 41 can communicate with a reconstruction module of a magnetic resonance imaging apparatus to receive an initial image, which is generated by performing a preset scanning sequence by a magnetic resonance imaging apparatus (for example, the apparatus shown in the figure). In other examples, the quantitative map acquisition module 41 can communicate with a data storage module to retrieve a quantitative map stored in advance therein. Figure 3

[0106] The image synthesis module 45 is used to synthesize a plurality of first weighted images corresponding to the plurality of quantitative maps.

[0107] The optimization module 47 is used to convert any first weighted image into a corresponding second weighted image based on a trained deep learning network.

[0108] The principles and examples have been described in detail in the description of the method of generating a magnetic resonance image according to an embodiment of the present application, and the system for generating a magnetic resonance according to an embodiment of the present application can be used to execute the above-mentioned method, both of which have the same inventive concept.

[0109] ​Another embodiment of the present application also provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed, performs the method of generating magnetic resonance images according to the embodiments of the present application.

[0110] Figure 5 、 Figure 6 respectively show a brain T1WI image generated according to an embodiment of the present application and a brain T1WI image reconstructed from actual scanning, Figure 7 、 Figure 8 respectively show a brain T2WI image generated according to an embodiment of the present application and a brain T2WI image reconstructed from actual scanning, Figure 9 、 Figure 10 respectively show a brain T2W-Flair image generated according to an embodiment of the present application and a brain T2W-Flair image reconstructed from actual scanning, Figure 11 、 Figure 12 respectively show a prostate T2WI image generated according to an embodiment of the present application and a prostate T2WI image reconstructed from actual scanning. By comparing the various weighted images generated according to the embodiments of the present application with the images reconstructed from actual scanning, it is found that the images generated by the method of the present application are very close to the imaging results of actual scanning, and thus the method can replace the existing scanning process to perform magnetic resonance imaging.

[0111] The various embodiments of the present application first generate a first weighted image by using a quantitative map, and then use the first weighted image as the input of a deep learning network which has been trained to be able to optimize the first weighted image to obtain a second weighted image. Compared with the prior art, there is no need to perform real scanning to obtain the image, thus greatly saving the time of magnetic resonance examination; the second weighted image has very close image features and image quality to the actual scanning reconstructed image, so that the reliability of medical diagnosis is ensured while realizing fast imaging, and, since the first weighted image as an intermediate result is obtained according to the quantitative map first, and then used as the input of the deep learning network, the efficient use of prior information can obtain a simplified neural network structure, which is easier to adjust parameters and optimize network structure and other details, and the efficiency is improved, and complex image post-processing operations can also be avoided.

[0112] In the various embodiments above, the processing unit, control unit comprises circuitry configured to perform one or more of the tasks, functions or steps discussed herein. In various embodiments, the processing unit can be integrated with the data processing unit 120 of the magnetic resonance imaging system, the control unit can be integrated with the control unit 130 of the magnetic resonance imaging system. The "processing unit", "control unit" as used herein is not intended to necessarily limit to a single processor or computer. For example, the processing unit, control unit can comprise multiple processors, ASICs, FPGAs and / or computers which can be integrated into a common housing or unit, or can be distributed in various units or housings. The depicted processing unit, control unit comprises a memory. The memory 130 can comprise one or more computer readable storage media. For example, the memory 130 can store information about system characteristics (e.g. information about spatial gradients), images (e.g. standard human body images), algorithms or procedures for performing any of the embodiments described above, etc. Further, the process flows and / or flowcharts discussed herein (or aspects thereof) can represent one or more sets of instructions stored in the memory for directing scan control or pre-scan control.

[0113] As used herein, an element or step recited in the singular and preceded with the word "a" or "an" should be understood as not excluding plural of said elements or steps, unless such exclusion is explicitly recited. Furthermore, references to "one embodiment" of the present application are not intended to be interpreted as excluding the existence of additional embodiments that also incorporate the recited features. Moreover, unless explicitly stated to the contrary, embodiments "comprising," "including," "containing," "having" or "with" an element or a plurality of elements having a particular property are intended to include additional such elements not having that property. The terms "including" and "in which" are used as the plain language equivalents of the respective terms "comprising" and "wherein." Moreover, in the following claims, the terms "first," "second," and "third," etc. are used merely as labels, and are not intended to impose numerical requirements on their objects.

[0114] This written description uses examples to disclose the invention, including the best mode, and also to enable any person skilled in the art to practice the invention, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the invention is defined by the claims, and can include other examples that occur to those skilled in the art. Such other examples are intended to fall within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insubstantial differences from the literal languages of the claims.

Claims

1. A method for generating magnetic resonance images, comprising: Obtain multiple quantitative graphs; A first weighted image is synthesized based on the multiple quantitative maps; as well as, The trained deep learning network converts the first weighted image into the corresponding second weighted image; The step of synthesizing the first weighted image includes: The first weighted image is determined based on a preset relationship between the quantitative values ​​in the plurality of quantitative images and the hypothesized sequence parameters and image signal values. The hypothesized sequence parameters include part or all of the hypothesized echo time, hypothesized repetition time, and hypothesized inversion recovery time.

2. The method as described in claim 1, wherein, Obtaining multiple quantitative graphs includes: The plurality of quantitative maps are calculated based on an initial image, which is generated by performing a preset scanning sequence using a magnetic resonance imaging device.

3. The method as described in claim 1, wherein, The multiple quantitative maps include the T1 quantitative map, the T2 quantitative map, and the proton density quantitative map.

4. The method of claim 1, wherein, The first weighted image is a T1-weighted image, a T2-weighted image, or a proton density-weighted image, and is obtained based on the following formula: S=PD·exp(-TE / T2)·(1-exp(-TR / T2)); Where S is the image signal value of the first weighted image, exp is an exponential function with the natural constant e as the base, TE is the assumed echo time, TR is the assumed repetition time, and T1, T2, and PD are the quantitative values ​​of T1, T2, and proton density, respectively.

5. The method of claim 1, wherein, The first weighted image is a T1WI-Flair image, a T2WI-Flair image, a STIR image, or a PSIR image, and the first weighted image is obtained based on the following formula: S=PD·exp(-TE / T2)·(1-2·exp(-TI / T1)+exp(-TR / T1); Where S is the image signal value of the first weighted image, exp is an exponential function with the natural constant e as the base, TE is the assumed echo time, TR is the assumed repetition time, TI is the assumed inversion recovery time, and T1, T2, and PD are the quantitative values ​​of T1, T2, and proton density, respectively.

6. The method of claim 3, wherein, The step of synthesizing the first weighted image further includes: performing convolution operations on the plurality of quantitative images respectively, and synthesizing the first weighted image based on the convolution operation results.

7. The method of claim 1, wherein, The input dataset for training the deep learning network includes multiple first weighted images, and the output dataset for training the deep learning network includes multiple real weighted images obtained by performing magnetic resonance imaging scans and reconstructing images based on the scan data.

8. The method of claim 7, wherein, The network parameters of the deep learning network are obtained by solving the following equation: minθ||f(θ)-f||2 Where θ is the network parameter, f includes the real weighted image, f(θ) represents the output of the deep learning network, and min represents minimization.

9. A computer-readable storage medium comprising a stored computer program, wherein, When the computer program is run, it performs the method according to any one of claims 1 to 8.

10. A system for generating magnetic resonance images, comprising: The quantitative graph acquisition module is used to acquire multiple quantitative graphs; An image synthesis module is used to synthesize a first weighted image based on the plurality of quantitative images; as well as, A conversion module is used to convert the first weighted image into a corresponding second weighted image based on a trained deep learning network; The image synthesis module is used for: The first weighted image is determined based on a preset relationship between the quantitative values ​​in the plurality of quantitative images and the hypothesized sequence parameters and image signal values. The hypothesized sequence parameters include part or all of the hypothesized echo time, hypothesized repetition time, and hypothesized inversion recovery time.

11. The system of claim 10, wherein, Also includes: Magnetic resonance imaging equipment, used to generate initial images; as well as, A calculation module for calculating the plurality of quantitative maps based on the initial image.

Images