Patent Application
20230221388
This application is based upon and claims the benefit of priority from Chinese Patent Application No. 202210025762.1, filed on Jan. 11, 2022; and Japanese Patent Application No. 2023-001205, filed on Jan. 6, 2023 the entire contents of all of which are incorporated herein by reference.
Embodiments described herein relate generally to a magnetic resonance imaging apparatus, an image reconstruction apparatus, and an image reconstruction method.
In magnetic resonance imaging, a technique is known by which scan time of a magnetic resonance imaging apparatus is accelerated by acquiring only a part of frequency domain information so as to reconstruct an image from undersampled frequency domain data. This technique may be applied to a magnetic resonance imaging apparatus based on an algorithm of parallel imaging or compressed sensing. According to this technique, because the data of a part of the frequency domain information is missing, imaging quality may be degraded. In recent years, as machine learning techniques have been developed, magnetic resonance imaging apparatuses using deep learning are proposed. The machine learning techniques may be applied to correcting data of an image domain or a frequency domain or to structuring an end-to-end reconstruction neural network. By using the machine learning techniques, it is possible to increase the width of the scan acceleration and to improve imaging quality.
A magnetic resonance imaging apparatus according to an aspect of the present disclosure includes a sequence controlling circuit and a processing circuit. The sequence controlling circuit is configured to acquire undersampled frequency domain scan data by executing a pulse sequence while carrying out an undersampling process. The processing circuit is configured: to generate image domain corrected data of the frequency domain scan data, by correcting the frequency domain scan data in an image domain; to generate frequency domain corrected data of the frequency domain scan data by correcting the frequency domain scan data in a frequency domain; to optimize the frequency domain scan data on the basis of the image domain corrected data and the frequency domain corrected data; and to reconstruct image data by using the optimized frequency domain scan data.
Exemplary embodiments of a magnetic resonance imaging apparatus, an image reconstruction apparatus, and an image reconstruction method will be explained in detail below, with reference to the accompanying drawings.
The following will describe the magnetic resonance imaging apparatus, the image reconstruction apparatus, and the image reconstruction method of the present disclosure, with reference to the accompanying drawings.
A magnetic resonance image reconstruction apparatus according to an embodiment of the present disclosure is configured to perform optimization n times in total (where n is an integer of 1 or larger) on frequency domain scan data K0 in a K-space obtained from a scan performed on an examined subject (hereinafter, “patient”) by a magnetic resonance scanning apparatus and to obtain image data IMG on the basis of frequency domain scan data Kn resulting from the optimization performed n times. The magnetic resonance scanning apparatus is configured to obtain the frequency domain scan data K0 of the patient, by transmitting a pulse signal to the patient placed in a magnetic field on which frequency encoding and phase encoding processes have been performed, and further receiving an echo signal caused by specific atomic magnetic resonance from a plurality of receiver coils.
In the present embodiment, the frequency domain scan data K0 is a three-dimensional tensor expressed as a width W×a height H×the number of channels C (the number of coils), where the width direction corresponds to a frequency encoding direction, whereas the height direction corresponds to a phase encoding direction. Generally speaking, in magnetic resonance scans, undersampling is carried out by skipping specific frequency encoding for the purpose of shortening scan time. Thus, a scan is performed by skipping a part of the coordinates in the frequency encoding direction (the width direction) during the scan. For this reason, in the frequency domain scan data K0, the part of the coordinates in the width direction (the frequency encoding direction) has no data, and a zero-padding process is performed for the data in the part of the coordinates. In the K-space, because the data in the vicinity of the center position has a larger impact on contrast of the reconstructed image data IMG, a method commonly used at the time of carrying out the undersampling is to sample, with priority, the data positioned in the vicinity of the center position in the frequency encoding direction, while skipping the data in some positions distant from the center position.
In the present embodiment, a mask M is used for indicating, during a magnetic resonance scan, which frequency encoding was sampled from the frequency domain scan data K0. The mask M is a matrix of a width W×a height H, while the values of the coordinates in the frequency encoding and the phase encoding that were sampled are expressed as 1, whereas the values of the coordinates in the frequency encoding and the phase encoding that were skipped in the sampling are expressed as 0.
The static magnetic field magnet 201 is a magnet formed so as to have a hollow substantially circular cylindrical shape and is configured to generate a static magnetic field in a space on the inside thereof. For example, the static magnetic field magnet 201 is a superconductive magnet or the like. In another example, the static magnetic field magnet 201 may be a permanent magnet.
The gradient coil 203 is a coil formed so as to have a hollow substantially circular cylindrical shape and is arranged on the inside of the static magnetic field magnet 201. The gradient coil 203 is formed by combining together three coils corresponding to X-, Y-, and Z-axes orthogonal to one another. By individually receiving a supply of an electric current from the gradient power supply 204, the three coils are configured to generate gradient magnetic fields of which magnetic field intensities change along the X-, Y-, and Z-axes. The gradient magnetic fields generated along the X-, Y-, and Z-axes by the gradient coil 203 are, for example, a slice gradient magnetic field Gs, a phase encoding gradient magnetic field Ge, and a readout gradient magnetic field Gr. The gradient power supply 204 is configured to supply the electric currents to the gradient coil 203.
The couch 205 includes a couchtop 205a on which the patient P is placed and is configured, under control of the couch controlling circuit 206, to insert the couchtop 205a into the hollow (an image taking opening) of the gradient coil 203, while the patient P is place thereon. Usually, the couch 205 is installed so that the longitudinal direction thereof is parallel to the central axis of the static magnetic field magnet 201. Under control of the magnetic resonance image reconstruction apparatus 1 (a computer), the couch controlling circuit 206 is configured to drive the couch 205, so as to move the couchtop 205a in longitudinal directions and up-and-down directions.
The transmitter coil 207 is arranged on the inside of the gradient coil 203 and is configured to generate a radio frequency magnetic field by receiving a supply of a Radio Frequency (RF) pulse from the transmitter circuit 208. The transmitter circuit 208 is configured to supply the RF pulse corresponding to a Larmor frequency determined by the type of targeted atoms and the magnetic field intensity, to the transmitter coil 207.
The receiver coil 209 is arranged on the inside of the gradient coil 203 and is configured to receive a magnetic resonance signal (hereinafter, “MR signal”, as necessary) emitted from the patient P due to influence of the radio frequency magnetic field. Upon receipt of the magnetic resonance signal, the receiver coil 209 is configured to output the received magnetic resonance signal to the receiver circuit 210.
The transmitter coil 207 and the receiver coil 209 described above are merely examples. It is possible to select one or combine two or more from among: a coil having only the transmitting function; a coil having only the receiving function; and a coil having the transmitting and receiving functions.
The receiver circuit 210 is configured to detect the magnetic resonance signal output from the receiver coil 209 and to generate magnetic resonance data on the basis of the detected magnetic resonance signal. More specifically, the receiver circuit 210 is configured to generate the magnetic resonance data by digitally converting the magnetic resonance signal output from the receiver coil 209. Further, the receiver circuit 210 is configured to transmit the generated magnetic resonance data to the sequence controlling circuit 220. Alternatively, the receiver circuit 210 may be provided for a gantry apparatus which includes the static magnetic field magnet 201, the gradient coil 203, and the like. Further, it is also acceptable to provide the receiver coil 209 with a part of the functions of the receiver circuit 210 such as, for example, the function to digitally convert the magnetic resonance signal.
The sequence controlling circuit 220 is configured to image the patient P, by driving the gradient power supply 204, the transmitter circuit 208, and the receiver circuit 210, on the basis of sequence information transmitted from the magnetic resonance image reconstruction apparatus 1 (the computer). In this situation, the sequence information is information defining a procedure for performing the imaging process. The sequence information defines: the magnitude of the electric current to be supplied to the gradient coil 203 by the gradient power supply 204 and the timing with which the electric current is to be supplied; the magnitude of the RF pulse to be supplied to the transmitter coil 207 by the transmitter circuit 208 and the timing with which the RF pulse is to be applied; the timing with which the magnetic resonance signal is to be detected by the receiver circuit 210; and the like. For example, the sequence controlling circuit 220 may be an integrated circuit such as an Application Specific Integrated Circuit (ASIC) or a Field Programmable Gate Array (FPGA), or an electronic circuit such as a Central Processing Unit (CPU) or a Micro Processing Unit (MPU). The sequence controlling circuit 220 is an example of a sequence controlling unit.
Further, upon receipt of the magnetic resonance data from the receiver circuit 210 as a result of imaging the patient P by driving the gradient power supply 204, the transmitter circuit 208, and the receiver circuit 210, the sequence controlling circuit 220 is configured to transfer the received magnetic resonance data to the magnetic resonance image reconstruction apparatus 1.
The input/output interface 101 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to an input apparatus (not illustrated) and is configured to receive an input operation of a user from the input apparatus and to transfer a signal based on the received input operation to the magnetic resonance image reconstruction apparatus 1. The input/output interface 101 may be a serial bus interface such as a Universal Serial Bus (USB), for example. Examples of the input apparatus include a mouse, a keyboard, a trackball, a switch, a button, a joystick, a touch screen, and a microphone. Further, the input/output interface 101 may have a storage apparatus connected thereto, so as to be able to read and write various types of data from and to the storage apparatus. The storage apparatus may be a Hard Disc Drive (HDD), a Solid State Drive (SSD), or the like.
The display 102 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to a display apparatus (not illustrated) and is configured to transmit data to the display apparatus and to cause the display apparatus to display an image. For example, the display 102 is a picture output interface such as a Digital Visual Interface (DIV) or a High-Definition Multimedia Interface (HDMI (registered trademark)). Examples of the display apparatus include a Liquid Crystal Display (LCD) and an organic Electroluminescence (EL) Display. The display apparatus is configured to display a user interface for receiving input operations from the user and the image data IMG output by the magnetic resonance image reconstruction apparatus 1. Examples of the user interface include a Graphical User Interface (GUI).
The communication interface 103 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to a server (not illustrated) and is configured to be able to transmit and receive various types of data to and from the server. The communication interface 103 is, for example, a network card such as a wireless network card or a wired network card.
The storage unit 104 is configured to store therein the frequency domain scan data K0 for performing image reconstruction and the mask M kept in correspondence therewith. Further, the storage unit 104 is configured to store therein parameters used by the magnetic resonance image reconstruction apparatus 1 at the time of performing the image reconstruction, such as parameters of a neural network, for example. Also, the storage unit 104 is configured store therein neural networks used by the magnetic resonance image reconstruction apparatus 1 and training data used for training other parameters that can be trained through machine learning. Sets of training data include the frequency domain scan data K0, the mask M, and image data ground truth IMGGT. For example, the storage unit 104 is realized by using a storage apparatus such as a Read Only Memory (ROM), a flash memory, a Random Access Memory (RAM), a Hard Disc Drive (HDD), a Solid State Drive (SSD), or a register. The flash memory, the HDD, the SSD, and the like are non-volatile storage media. These non-volatile storage media are realized by other storage apparatuses connected via a network such as a Network Attached Storage (NAS) or an external storage server apparatus. In this situation, examples of the network include the Internet, a Wide Area Network (WAN), a Local Area Network (LAN), a carrier terminal, a wireless communication network, a wireless base station, and a dedicated communication line.
A processing circuit (not illustrated) included in the magnetic resonance image reconstruction apparatus 1 has processing functions such as a K-space interpolating function, a sensitivity distribution calculating function, an image domain correcting function, a consistency data calculating function, a frequency domain correcting function, an optimizing function, and a reconstructing function. The processing functions implemented by the K-space interpolating function, the sensitivity distribution calculating function, the image domain correcting function, the consistency data calculating function, the frequency domain correcting function, and the optimizing function are stored in the storage unit 104 in the form of computer-executable programs. The processing circuit is a processor configured to realize the functions corresponding to the programs, by reading and executing the programs from the storage unit 104. In other words, the processing circuit that has read the programs has the abovementioned functions. That is to say, the K-space interpolating function, the sensitivity distribution calculating function, the image domain correcting function, the consistency data calculating function, the frequency domain correcting function, and the optimizing function, and the reconstructing function are examples of the K-space interpolating unit 105, the sensitivity distribution calculating unit 106, the image domain correcting unit 107, the consistency data calculating unit 108, the frequency domain correcting unit 109, the optimizing unit 110 and the reconstruction unit 111, respectively. Further, although
The term “processor” used in the above description denotes, for example, a Central Processing Unit (CPU), a Graphical Processing Unit (GPU), or a circuit such as an Application Specific Integrated Circuit (ASIC) or a programmable logic device (e.g., a Simple Programmable Logic Device (SPLD), a Complex Programmable Logic Device (CPLD), or a Field Programmable Gate Array (FPGA)). The one or more processors are configured to realize the functions by reading and executing the programs saved in the memory 132.
Further, instead of having the programs saved in the storage unit 104, it is also acceptable to directly incorporate the programs in the circuits of the one or more processors. In that situation, the one or more processors realize the functions by reading and executing the programs incorporated in the circuits thereof.
On the basis of the mask M, the K-space interpolating unit 105 is configured to generate frequency domain interpolated data KSI by performing an interpolating process on a zero-padded part of the frequency domain scan data K0. The K-space interpolating unit 105 includes a first neural network 1051 and a computing unit 1052. The first neural network 1051 may be, for example, a feed forward neural network, a convolutional neural network, or a transformer. Preferably, the first neural network 1051 is a convolutional neural network. More preferably, the first neural network 1051 is U-Net. In the present embodiment, the first neural network 1051 is assumed to be a convolutional neural network including an input layer, an output layer, a convolution layer, an activation layer, a pooling layer, a batch normalization layer, and a fully connected layer, while the sizes of the input layer and the output layer are equal to each other. Parameters used by the first neural network 1051 are stored in the storage unit 104. The computing unit 1052 is configured to convert the mask M into an inversed mask IM. The inversed mask IM is a matrix having the same size as that of the mask M and is obtained by replacing is in the mask M with 0s and replacing 0s in the mask M with 1s.
The process performed by the first neural network 1051 is considered as a process of embedding data in the zero-padded part of the frequency domain scan data K0 on the basis of the data in the part of the frequency domain scan data K0 that is not zero-padded. When the neural network data NN1 (K0) is compared with the frequency domain scan data K0, the data has been embedded in the zero-padded part of the frequency domain scan data K0, and the data in the part that is not zero-padded has also changed. Further, the K-space interpolating unit 105 is configured to read the mask M from the storage unit 104 and to cause the computing unit 1052 to calculate the inversed mask IM. After that, the K-space interpolating unit 105 is configured to calculate the frequency domain interpolated data KSI, by adding together the inner product of the neural network data NN1(K0) and the inversed mask IM and the inner product of the frequency domain scan data K0 and the mask M.
When the frequency domain interpolated data KSI is compared with the frequency domain scan data K0, only the zero-padded part of the frequency domain scan data K0 is different, and the original data from the frequency domain scan data K0 is retained in the other part. The frequency domain interpolated data KSI is considered as normally-sampled frequency domain scan data that was estimated by the K-space interpolating unit 105 on the basis of the undersampled frequency domain scan data K0.
The sensitivity distribution calculating unit 106 is configured to generate a sensitivity distribution map SM by estimating sensitivities of a plurality of coils used by a magnetic resonance apparatus, on the basis of the frequency domain scan data K0. The sensitivity distribution calculating unit 106 includes an inverse Fourier transform unit 1061, a second neural network 1062, and a computing unit 1063. The second neural network 1062 may be, for example, a feed forward neural network, a convolutional neural network, or a transformer. Preferably, the second neural network 1062 is a convolutional neural network. More preferably, the second neural network 1062 is U-net. In the present embodiment, the second neural network 1062 is assumed to be a convolutional neural network including an input layer, an output layer, a convolution layer, an activation layer, a pooling layer, a batch normalization layer, and a fully connected layer, while the sizes of the input layer and the output layer are equal to each other. Parameters used by the second neural network 1062 are stored in the storage unit 104. The computing unit 1063 is configured to convert neural network data NN2(X0) output by the second neural network 1062 into the sensitivity distribution map SM.
The image domain correcting unit 107 is configured to generate image domain corrected data IRt by correcting frequency domain scan data Kt that is obtained as a result of the optimizing unit 110 optimizing the frequency domain scan data K0 t times, where t is an integer from 0 to n inclusive. The optimizing unit 110 and the optimizing process performed thereby will be explained later. The image domain correcting unit 107 includes an inverse Fourier transform unit 1071, a coil integrating unit 1072, a third neural network 1073, a coil separating unit 1074, and a Fourier transform unit 1075. The coil integrating unit 1072 is configured to convert multi-channel data into single-channel data. The third neural network 1073 may be, for example, a feed forward neural network, a convolutional neural network, or a transformer. Preferably, the third neural network 1073 is a convolutional neural network. More preferably, the third neural network 1073 is U-Net. In the present embodiment, the third neural network 1073 is assumed to be a convolutional neural network including an input layer, an output layer, a convolution layer, an activation layer, a pooling layer, a batch normalization layer, and a fully connected layer, while the sizes of the input layer and the output layer are equal to each other. Parameters used by the third neural network 1073 are stored in the storage unit 104. The coil separating unit 1074 is configured to convert single-channel data into multi-channel data.
The consistency data calculating unit 108 is configured to generate consistency data DCt for constraining consistency between the frequency domain scan data Kt and the frequency domain scan data K0.
The frequency domain correcting unit 109 is configured to generate frequency domain corrected data KRt, by correcting the frequency domain scan data Kt. The frequency domain correcting unit 109 includes a fourth neural network 1091. The fourth neural network 1091 may be, for example, a feed forward neural network, a convolutional neural network, or a transformer. Preferably, the fourth neural network 1091 is a convolutional neural network. More preferably, the fourth neural network 1091 is U-Net. In the present embodiment, the fourth neural network 1091 is assumed to be a convolutional neural network including an input layer, an output layer, a convolution layer, an activation layer, a pooling layer, a batch normalization layer, and a fully connected layer, while the sizes of the input layer and the output layer are equal to each other. Parameters used by the fourth neural network 1091 are stored in the storage unit 104.
The optimizing unit 110 is configured to generate frequency domain scan data Kt+1, by optimizing the frequency domain scan data Kt, on the basis of the frequency domain interpolated data KSI, the image domain corrected data IRt, the consistency data DCt, and the frequency domain corrected data KRt.
The reconstructing unit 111 is configured to generate the image data IMG, on the basis of the frequency domain scan data Kn resulting from performing optimization n times. The reconstructing unit 111 includes an inverse Fourier transform unit 1111 and a channel integrating unit 1112.
The magnetic resonance image reconstruction method according to the present embodiment is configured to carry out the iteration n times in total, so that the undersampled frequency domain scan data K0 gradually approaches normally-sampled frequency domain scan data KNS. In each iteration, the frequency domain scan data K0 is optimized once. In the present embodiment, carrying out the iteration once denotes performing steps S105 and S106 once in the stated order. While the magnetic resonance image reconstruction method according to the present embodiment is implemented, the parameters of the neural networks and the other parameters that can be trained through machine learning are not changed. Preferably, the number of times of iteration is 6 to 12 times.
The sequence controlling circuit 220 is configured to acquire the undersampled frequency domain scan data K0, by executing the pulse sequence while carrying out the undersampling process. At step S101, by using the user interface displayed on the display apparatus, the user selects, via the input apparatus, the frequency domain scan data K0 and the mask M kept in correspondence with the frequency domain scan data K0 that are either stored in the storage unit 104 or input from an external source. The process then proceeds to step S102.
At step S102, the K-space interpolating unit 105 generates the frequency domain interpolated data KSI by performing the interpolating process on the zero-padded part of the frequency domain scan data K0 on the basis of the mask M. The process then proceeds to step S103. The frequency domain interpolated data KSI is used for optimizing the frequency domain scan data Kt in the iteration of n times thereafter.
At step S103, the sensitivity distribution calculating unit 106 generates the sensitivity distribution map SM by estimating the sensitivities of the plurality of coils used by the magnetic resonance apparatus, on the basis of the frequency domain scan data K0. The process then proceeds to step S104. The sensitivity distribution map SM is used for calculating the frequency domain corrected data KRt in the iteration of n times thereafter. At step S104, the magnetic resonance image reconstruction apparatus 1 sets the number of times of iteration IT to 0. The process then proceeds to step S105.
Step S105 includes steps S1051, S1052, and S1053. Steps S1051, S1052, and S1053 may be performed in parallel to one another or may be performed sequentially. In the present embodiment, the example in which steps S1051, S1052, and S1053 are performed in parallel to one another will be explained. At step S1051, the image domain correcting unit 107 generates the image domain corrected data IRt, by correcting the frequency domain scan data Kt on the basis of the sensitivity distribution map SM and outputs the image domain corrected data IRt to the optimizing unit 110. At step S1052, the consistency data calculating unit 108 generates the consistency data DCt for constraining consistency between the frequency domain scan data Kt and the frequency domain scan data K0, on the basis of the frequency domain scan data Kt and the frequency domain scan data K0 and further outputs the consistency data DCt to the optimizing unit 110. At step S1053, the frequency domain correcting unit 109 generates the frequency domain corrected data KRt by correcting the frequency domain scan data Kt and further outputs the frequency domain corrected data KRt to the optimizing unit 110. When all the processes at steps S1051, S1052, and S1053 are completed, the process proceeds to step S106.
Returning to the description of
At step S107, the magnetic resonance image reconstruction apparatus 1 judges whether or not the number of times of iteration IT is equal to n. When the judgment result is “YES”, the process proceeds to step S108. When the judgment result is “NO”, the process proceeds to step S105. At step S108, the reconstructing unit 111 generates the image data IMG on the basis of the frequency domain optimized data Kn which is obtained by optimizing the frequency domain scan data K0 n times, and the process is thus ended. At step S109, the magnetic resonance image reconstruction apparatus 1 increments the number of times of iteration IT by 1 and carries out the next iteration.
In the above explanation, the magnetic resonance image reconstruction apparatus and the magnetic resonance image reconstruction method according to the present embodiment are configured to use the first to the fourth neural networks and the parameter ηt. The neural networks and the parameter do not operate properly unless being trained. In the following sections, a method for training the abovementioned neural networks and parameter will be explained.
To begin with, a plurality of sets of training data that are stored in advance are read from the storage unit 104. The sets of training data include the undersampled frequency domain scan data K0 obtained by the magnetic resonance apparatus and the mask M kept in correspondence therewith which serve as input data, as well as the image data ground truth IMGGT which serves as output data.
Further, the plurality of sets of training data are divided into a training set and a test set. Examples of the ratio of the training set to the test set may include 80% to 20% and 90% to 10%. For example, when the total quantity of the sets of training data is 10,000, the training data numbered from #1 to #10,000 may be divided into the training set with the data numbered from #1 to #8,000; and the test set with the data numbered from #8,001 to #10,000. In this situation, the input data in the sets of training data within the training set is input to the magnetic resonance image reconstruction apparatus 1, further the image data IMG is calculated by implementing the magnetic resonance image reconstruction method according to the present embodiment, the difference value between the image data IMG and the image data ground truth IMGGT is calculated, and a backpropagation is carried out on the basis of the difference value, so as to change the parameters of the neural networks and the other parameters that can be trained through the machine learning in such a manner that the difference value between the image data IMG output by the magnetic resonance image reconstruction apparatus 1 and the image data ground truth IMGGT is reduced. The abovementioned process is repeatedly performed on a large part of the data in the training set until the difference value between the image data IMG output by the magnetic resonance image reconstruction apparatus 1 and the image data ground truth IMGGT becomes smaller than a threshold value set in advance. After that, it is determined that the training of the neural networks and the parameters has been completed.
Subsequently, the test data (i.e., the data numbered from #8,000 to #10,000) serving as input data is input to the trained magnetic resonance image reconstruction apparatus 1 so as to calculate a difference value between the image data IMG output by the magnetic resonance image reconstruction apparatus 1 and the image data ground truth IMGGT as evaluation data.
Next, advantageous effects of the magnetic resonance image reconstruction apparatus and the magnetic resonance image reconstruction method according to the present embodiment will be explained.
In a conventional magnetic resonance apparatus, to shorten the scan time, frequency domain information in a K-space is undersampled so as to generate undersampled frequency domain data. Because the reconstruction of a scan image using the undersampled frequency domain data is an ill-posed problem, there are countless solutions, and it is not possible to identify an accurate scan image. In contrast, the magnetic resonance image reconstruction apparatus and the magnetic resonance image reconstruction method according to the present embodiment are configured to solve the problem on the basis of a compressive sensing algorithm. The compression sensing algorithm is able to determine one appropriate solution in one-to-one correspondence, by adding a constraint.
The magnetic resonance image reconstruction method based on the compressive sensing is configured to calculate the normally-sampled frequency domain scan data KNS by iteratively optimizing the undersampled frequency domain scan data K0. It is possible to express a relationship between the frequency domain scan data Kt+1 resulting from the optimization performed t+1 times and the frequency domain scan data Kt resulting from the optimization performed t times, by using expression (1) presented below.
K
t+1
=K
t−ηt·DCt+G(Kt) (1)
In Expression (1), ηt denotes a learning parameter that can be trained through machine learning, whereas G(Kt) is a mathematical function of Kt. An additional constraint is expressed.
In the present embodiment, while the image domain corrected data IRt, the frequency domain corrected data KRt, and the frequency domain interpolated data KSI are used as the mathematical function G(Kt), a sparsity constraint is added to the image reconstruction.
The image domain corrected data IRt is corrected data obtained by correcting the frequency domain scan data Kt in the image domain, while employing the third neural network NN3. The process of adding together the frequency domain scan data Kt and the image domain corrected data IRt is considered as processes of eliminating noise and eliminating artifacts from the frequency domain scan data Kt in the image domain, so as to cause the frequency domain scan data Kt to approach the normally-sampled frequency domain scan data KNS. The frequency domain corrected data KRt is corrected data obtained by correcting the frequency domain scan data Kt in the frequency domain, by employing a fourth neural network NN4. The process of adding together the frequency domain scan data Kt and the frequency domain corrected data KRt is considered as processes of eliminating noise and eliminating artifacts from the frequency domain scan data Kt in the frequency domain, so as to cause the frequency domain scan data Kt to approach the normally-sampled frequency domain scan data KNS. The frequency domain interpolated data KSI is data generated by the first neural network 1051 on the basis of the undersampled frequency domain scan data K0 and is considered as normally-sampled frequency domain scan data estimated by the first neural network 1051. By adding the frequency domain interpolated data KSI multiplied by the learning parameter λt to the frequency domain scan data Kt, the frequency domain scan data Kt is caused to approach the normally-sampled frequency domain scan data KNS.
In an example of 4X undersampling, the image data IMG generated by the magnetic resonance image reconstruction apparatus 1 according to the present embodiment and the image data ground truth IMGGT exhibit the following values: MSE=2.372e−11±2.292e−11; NMSE=0.003668±0.003615; PSNR=40.5±3.687; and SSIM=0.9514±0.3687. In an example of 8× undersampling, the image data IMG generated by the magnetic resonance image reconstruction apparatus 1 according to the present embodiment and the image data ground truth IMGGT exhibit the following values: MSE=9.638e−11±1.146e−10; NMSE=0.01261±0.00547; PSNR=34.77±4.609; and SSIM=0.9064±0.06189.
In the example of the 4X undersampling, the image data reconstructed by the magnetic resonance image reconstruction apparatus 1 according to the present embodiment restored the image data ground truth almost completely, and also, the processes of the noise elimination and the artifact elimination were performed on the image. In contrast, the image data reconstructed by using the conventional technique is unable to completely restore very small structures in the image, such as blood vessel. In addition, the image has apparent noise and artifacts. In the example of the 8X undersampling, the image data reconstructed by the magnetic resonance image reconstruction apparatus 1 according to the present embodiment roughly restored the image data ground truth, although a part of details of the image was lost. In addition, the processes of the noise elimination and the artifact elimination were performed on the image. In contrast, a large amount of detailed information was lost in the image data reconstructed by using the conventional technique. In addition, serious noise and artifacts occurred.
According to the present embodiment, because the neural networks are used for the K-space interpolating unit 105, the sensitivity distribution calculating unit 106, the image domain correcting unit 107, and the frequency domain correcting unit 109, it is possible to improve the precision level of the reconstruction for the image data IMG. In addition, it is also possible to reduce the number of times of iteration required before the image data IMG is reconstructed.
According to the present embodiment, the frequency domain scan data Kt is corrected by using the image domain corrected data IRt and the frequency domain corrected data KRt. It is therefore possible to reconstruct the image data IMG more accurately and in a more robust manner. Further, the image domain corrected data IRt and the frequency domain corrected data KRt are calculated in parallel to each other. It is therefore possible to enhance efficiency of the reconstruction and to improve the precision level of the reconstruction for the image data IMG.
According to the present embodiment, because the frequency domain scan data Kt is optimized by using the frequency domain interpolated data KSI, it is possible to reconstruct the image data IMG more accurately and in a more robust manner. Further, in the backpropagation during the training, it is possible to simplify the training process.
According to the present embodiment, because the image domain corrected data IRt is calculated by using the sensitivity distribution map SM, it is possible to further improve the precision level of the reconstruction for the image data IMG.
Next, a magnetic resonance image reconstruction apparatus and a magnetic resonance image reconstruction method according to a second embodiment will be explained. In the second embodiment, differences from the first embodiment will primarily be explained. Explanations of some of the features that are the same as those in the first embodiment will be omitted. In the description of the second embodiment, some of the elements that are the same as those in the first embodiment will be referred to by using the same reference characters.
The optimizing unit 110B is configured to generate frequency domain scan data Kt+1, by optimizing the frequency domain scan data Kt, on the basis of the image domain corrected data IRt, the consistency data DCt, and the frequency domain corrected data KRt.
According to the present embodiment, because the frequency domain scan data Kt is corrected by using the image domain corrected data IRt and the frequency domain corrected data KRt, it is possible to reconstruct the image data IMG more accurately and in a more robust manner.
In the embodiments described above, the sensitivity distribution calculating unit 106 is configured to calculate the coil sensitivity distribution map SM on the basis of the neural network. However, the sensitivity distribution calculating unit 106 may be configured to calculate the coil sensitivity distribution map SM on the basis of an ESPiRiT algorithm. Further, it is also acceptable to reconstruct frequency domain scan data (a single channel) resulting from a scan performed by a single-coil magnetic resonance scanning apparatus. In addition, it is also acceptable to improve the precision level and resolution of the image data IMG by performing a super-resolution process on the frequency domain scan data K0.
While a number of embodiments of the present disclosure have been described, these embodiments are presented by way of examples only, and are not intended to limit the scope of the inventions. The novel embodiments described herein may be carried out in a variety of other forms. It is also possible to make various omissions, substitutions, and changes without departing from the gist of the inventions. The embodiments and modifications thereof are covered by the scope and the gist of the inventions and are also covered by the present inventions and equivalents thereof. Further, it is also possible to carry out any of the embodiments described above in combination.
According to at least one aspect of the embodiments described above, it is possible to improve quality of the images.
While certain embodiments have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the inventions. Indeed, the novel embodiments described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutions and changes in the form of the embodiments described herein may be made without departing from the spirit of the inventions. The accompanying claims and their equivalents are intended to cover such forms or modifications as would fall within the scope and spirit of the inventions.
| Number | Date | Country | Kind |
|---|---|---|---|
| 202210025762.1 | Jan 2022 | CN | national |
| 2023-001205 | Jan 2023 | JP | national |
