Adaptive reconstruction of magnetic resonance images.

Abstract

The present disclosure relates to a method that includes providing a trained machine learning model that reconstructs an image from input data. The method includes receiving (201) a multidimensional array containing M-dimensional acquired data, determining (205) a subset of at least one selected K-dimensional value of the array, and for each value of the subset, determining (207) an M-dimensional array containing the acquired data corresponding to the value, resulting in a set of M-1-dimensional arrays, inputting (209) the set of M-dimensional arrays to the trained machine learning model, and receiving a reconstructed image from the trained machine learning model.

Description

[Technical field]

[0001] The present invention relates to scanning imaging systems, and in particular to medical analysis systems for reconstructing magnetic resonance images. [Background technology]

[0002] Magnetic resonance imaging (MRI) scanners rely on a large static magnetic field (B0) to align the nuclear spins of atoms as part of the procedure to generate images of the patient's body. These images reflect various quantities or characteristics of the subject. Image reconstruction from undersampled k-space data plays an important role for MRI. Deep learning in particular has been shown to have the potential to significantly speed up MRI reconstruction with fewer measurements. In the ISMRM-2018 abstract (page 2796): "Integrating spatial and temporal correlation into a deep neural network for low-delay reconstruction of highly undersampled radial dynamic images", H. Takeshima discloses a dynamic reconstruction method using a deep neural network (DNN) where a single image is reconstructed from M consecutive frames and N adjacent slices. Summary of the Invention

[0003] Various embodiments provide a medical analysis system, a method and a computer program product for reconstructing magnetic resonance images as described by the subject matter of the independent claims. Advantageous embodiments are described in the dependent claims.

[0004] Embodiments of the present invention provide means for reconstructing an actual (center) line, slice, or slab. At this time, the correlation with adjacent lines, slices, or slabs is considered. The present invention can reduce the dimensionality of the optimization problem in reconstruction, which reduces the computational amount and time of MR image reconstruction. This reduces the dimensionality of the multi-dimensional array of dimension M of the MR image data by only one or more dimensions (dimensions with K≧1), resulting in a reduced dimension of M-K (M minus K), selects one or more lower parts of the reduced array (each lower part includes a correlated array), and uses the lower part as an input to a model trained to reconstruct an image from data having the reduced dimension M-K.

[0005] In one aspect, the present invention relates to a medical analysis system for reconstructing magnetic resonance images. The medical analysis system includes a processor and at least one memory storing machine-executable instructions. The processor controls the medical analysis system. The medical analysis system includes a trained machine learning model. The trained machine learning model reconstructs an MR image from input data. Execution of the machine-executable instructions causes the processor to receive a multi-dimensional array including acquired data of dimension M, determine values of a subset of at least one selected dimension of the M-dimensional array (K selected dimensions, K≧1 and K<M), and for each value of each subset of at least one subset, determine an M-K dimensional array including the acquired data corresponding to the value, such that a set of M-K dimensional arrays is obtained as a result, and input the set of M-K dimensional arrays to the trained machine learning model and receive an image reconstruction output from the trained machine learning model.

[0006] For example, if two dimensions are selected (i.e., K=2), two subsets of values ​​(one subset for each selected dimension) are determined. For each pair consisting of a first value of one of the two subsets and a second value of the other of the two subsets, an M-2 dimensional array is determined. For example, if the first subset has two values ​​and the second subset has three values, six possible pairs are defined from these two subsets, and therefore six M-2 arrays are determined for the six pairs. If three dimensions are selected (i.e., K=3), three subsets of values ​​(one subset for each selected dimension) are determined. For each triplet consisting of a first value of one of the three subsets and a second and third value of the other two subsets of the three subsets, an M-3 dimensional array is determined. For example, if one dimension is selected (K=1), one subset of values ​​is determined. For each value of the subset, an M-1 dimensional array is determined, and so on.

[0007] Trained models, such as deep learning models, perform better when the dimensionality of the problem is increased. For example, a 3D problem is solved by processing a 3D undersampled orthogonal scan with a pure 3D reconstruction. This results in better image quality than solving the 3D problem slice by slice (e.g., N times the 2D problem). However, solving a 2D problem N times is more practical than solving the entire 3D problem in one go. This is more practical in terms of computation time and memory during training and inference, which allows the application of more powerful deep learning solutions.

[0008] The subject matter allows for decomposing an M-dimensional problem into multiple MK-dimensional problems. An example is fast orthogonal 3D scanning, where k-space is completely measured in the read direction. The latter allows for decomposing the 3D reconstruction problem into Nread_out times of 2D problems to allow for decomposition of the optimization problem along the read direction. By decomposing the problem into smaller problems, the computational and memory load is reduced since the trained machine learning model has to deal with less data. Computational load and memory consumption are important factors during inference (e.g., during image reconstruction on a scanner) and during training. This is particularly advantageous for deep neural networks, for the following reasons: The size of the network in a deep learning solution is limited by hardware constraints (e.g., availability of GPU memory of computing power) during training and on the scanner. Since performance in terms of output quality usually improves with increasing model size, it is beneficial to use as large a network as possible. Thus, solving smaller problems allows for the use of larger models and thus improves output quality. The subject matter further optimizes solutions to fewer multiple MK-dimensional problems. In particular, the present subject matter employs an "N and a half" D approach to solve an M-dimensional problem using correlations between N(MK) dimensional data (for M=3, the approach becomes a 2.5D approach). By using information from a few correlated slabs or slices (e.g., adjacent 2D slices are correlated slices) rather than the entire M-dimensional volume, the correlation between slices or slabs is used to improve the image quality of an actual single slice or slab. This "N and a half" D approach can be used in situations where the problem can be separated along one or more dimensions.

[0009] For example, an MRI scanner is used to scan or image a target volume (e.g., the brain) within a subject to acquire data. This results in M-dimensional acquired data. The acquired data includes k-space data (e.g., undersampled k-space data) or other image data, such as aliased image data. K-space data is defined herein as the recorded measurements of radio frequency signals emitted from atomic spins using the antenna of a magnetic resonance device during a magnetic resonance imaging scan.

[0010] The trained machine learning model provides an image reconstruction output in response to receiving a set of X-dimensional arrays of data. The image reconstruction output is a result of reconstructing image data of one X-dimensional (XD) array of a set of XD arrays and considering correlations between the one XD array and the remaining XD arrays of the set of XD arrays. The X dimensions are provided such that MK=X. The machine learning model is trained to directly learn the mapping between the input (e.g., undersampled k-space data or aliased image) and the output MR image. For example, the image reconstruction output is a magnetic resonance image, such as a reconstructed one-dimensional, two-dimensional, or three-dimensional visualization of the anatomical data contained in the k-space data. For example, the undersampled k-space data is used as an input to the trained machine learning model, and the desired image from the fully sampled k-space data is the reconstructed image output.

[0011] Each array in the set of MK arrays includes at least one dimension that represents spatial frequency information in one of three directions of the object. In one embodiment, K=1, and thus one dimension is selected and a subset of values ​​along the selected dimension is determined.

[0012] In one embodiment, execution of the machine-executable instructions further causes the processor to determine values ​​of the subset such that a correlation between an MK dimension array associated with a value (or a given value) of the subset and the remaining MK dimension arrays of the set is higher than a predefined threshold.

[0013] Correlations along dimensions can be very strong, especially in a local manner (eg, adjacent slices), and therefore can limit the output quality if not taken into account.

[0014] This embodiment allows solving partial multidimensional problems where not all dimensional data are fully considered, but only "local" data. That is, when optimizing a particular multidimensional problem of a (MK)-dimensional point, the data of the neighborhood of that point is also provided as input to the trained machine learning model. For example, instead of passing the data of one point in the reading direction (e.g., passing one ky / kz space), a range of data in the reading direction is fed to the trained machine learning model. The output of the trained machine learning model is still the solution of the original multidimensional problem, but the machine learning model can now take advantage of the (local) correlation that exists along the reading direction.

[0015] The correlation between MK dimensional arrays depends, for example, on the position. For example, 2D adjacent slices of a 3D acquisition are highly correlated due to the lack of gaps between the slices. In particular, adjacent 2D slices are very similar due to the slow spatial variation of the scanned object. For example, if M=3 and a 3D k-space k(x,y,z) array is provided with a selected dimension where x(x=x1,x2,x3,...), slices k(x1,y,z), k(x2,y,z) and k(x3,y,z) corresponding to successive values ​​x1, x2 and x3 are the correlated slices. Thus, the 2D slices k(x1,y,z), k(x2,y,z) and k(x3,y,z) are a set of 2D arrays. In another example, if M=4 and a 4D k-space k(x,y,z,t) array is provided with a selected dimension being time t (t=t1,t2,t3,...), then the 3D slabs k(x,y,z,t1) and k(x,y,z,t4) are the correlated sets of the 3D array. Correlation may be between slabs acquired at different time points, not necessarily consecutive time points. For example, in cardiac imaging, correlation is strongest between adjacent phases of the cardiac cycle that are not measured consecutively in time.

[0016] In one embodiment, execution of the machine executable instructions further causes the processor to determine values ​​of further subsets and repeat step d) of determining an MK dimensional array and input step e) for each further subset of the further subsets. Steps d) and e) are repeated to provide an image reconstruction output for each MK dimensional array of the M dimensional arrays. For example, when M=3 and K=1, each 2D slice along the selected dimension is input to the trained machine learning model along with one or more slices that are correlated with each 2D slice. For example, when the number of 2D slices along the selected dimension is 100, steps d) and e) are repeated 100 times to provide 100 reconstructed image outputs (2D slice images).

[0017] In one embodiment, the subsets are non-overlapping subsets. This allows further saving of processing resources while still providing accurate reconstruction results. For example, if M=3 and K=1, only one slice of the correlation slices of the subset of 2D slices is reconstructed to represent all slices of the subset. If the number of 2D slices along the selected dimension is 100, steps d) and e) are repeated a number of times less than 100 (which is the number of all subsets of the selected dimension values). For example, if the first selected subset has 3 slices, the next subset is selected from the remaining 100-3 slices, if the next subset contains 5 slices, the next subset is selected from the remaining 100-8 slices, and so on.

[0018] According to one embodiment, the M dimensions of acquired data are 3D k-space data (M=3), the selected dimension is the read direction (K=1), and a set of M-1 dimensional arrays represent a corresponding set of 2D slices.

[0019] For example, M=3 dimensions correspond to the spatial dimensions kx, ky, and kz of k-space. The 3D acquired data is k-space data.

[0020] This embodiment is advantageous for the following reasons: By using a 2D neural network model, the subject matter can be used with the limited memory of a GPU to get the most performance out of the 2D neural network, since using a 3D neural network model can quickly run into memory issues when passing a complete image or patch through the 3D neural network.

[0021] In one embodiment, the subset of values ​​includes a predefined odd number of values, which allows obtaining slices with high correlation, especially when the actual value is the median value of the determined subset. In another example, the subset of values ​​includes a randomly selected number of values ​​that is smaller than a predefined maximum number.

[0022] In one embodiment, determining the subset of values ​​includes identifying a center value and two or more surrounding values ​​of the center value along the selected dimension, where the subset includes the center value and the surrounding values. This allows for a more systematic way of selecting the subset, further improving performance of the subject matter. This is particularly advantageous since adjacent slices are highly correlated.

[0023] In one embodiment, the trained machine learning model is a deep neural network (DNN) model. For example, the input layer of the DNN includes nodes, and each of these nodes receives the value of an element of one set of sets of MK-dimensional arrays. In another example, the trained machine learning model is an ISTA-net model. This enables fast reconstruction of magnetic resonance images from sparsely sampled MR data in k-space using a deep learning neural network.

[0024] In another aspect, the invention relates to a method including providing a trained machine learning model for reconstructing an image from input k-space data; receiving a multi-dimensional array including M dimensions of acquired data; determining values ​​for a subset of K selected dimensions of the array, where K is an integer greater than or equal to 1 and less than M; and for each value of the subset, determining MK dimensional arrays including acquired data corresponding to the value, resulting in a set of MK dimensional arrays; inputting the set of MK arrays to the trained machine learning model; and receiving a reconstructed image from the trained machine learning model.

[0025] In another aspect, the invention relates to a computer program product comprising machine-executable instructions for execution by a processor, the execution of the machine-executable instructions causing the processor to perform the method of any of the preceding embodiments.

[0026] In one aspect, the present invention relates to a medical analysis system for reconstructing magnetic resonance images. The medical analysis system includes a processor and at least one memory having machine executable instructions stored therein. The processor controls the medical analysis system. The medical analysis system includes a trained machine learning model. The trained machine learning model reconstructs an image from input data. Execution of the machine executable instructions causes the processor to receive a dimensional array including M multi-dimensional acquired data, determine values ​​of a subset of selected dimensions of the M dimensional array, and for each value of the subset, determine an M-1 dimensional array including acquired data corresponding to the value, resulting in a set of M-1 dimensional arrays, input the set of M-1 dimensional arrays to the trained machine learning model, and receive an image reconstruction output from the trained machine learning model.

[0027] It will be understood that one or more of the above-described embodiments of the present invention can be combined, unless the combined embodiments are mutually exclusive. [Brief description of the drawings]

[0028] Preferred embodiments of the invention will now be described, by way of example only, with reference to the drawings in which:

[0029] [Figure 1] FIG. 1 is a schematic diagram of a control system in accordance with the present subject matter. [Diagram 2] FIG. 2 is a flow chart of a method for reconstructing a magnetic resonance image according to an example of the present subject matter. [Diagram 3] FIG. 3 shows a graph illustrating the expected gains in image quality (IQ), memory cost and increased computational load when using the present subject matter. [Figure 4] FIG. 4 shows a cross-sectional and functional diagram of the MRI system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0030] In the following, elements with the same number in the figures are similar elements or perform equivalent functions. Elements already described will not necessarily be described in subsequent figures if their functions are equivalent.

[0031] Various structures, systems and devices are depicted diagrammatically in the figures for purposes of explanation only and so as to not obscure the present invention with details that are well known to those skilled in the art. However, the attached figures are included to illustrate illustrative examples of the disclosed subject matter.

[0032] 1 is a schematic diagram of a medical analysis system 100. The medical analysis system 100 includes a control system 111 that is connected to a scanning imaging system (or acquisition component) 101. The control system 111 includes a processor 103 and a memory 107, each capable of communicating with one or more components of the medical system 100. For example, the components of the control system 111 are coupled to a bidirectional system bus 109.

[0033] It will be understood that the methods described herein are at least partially non-interactive and automated by a computerized system. For example, the methods may also be implemented in software 121 (including firmware), hardware, or a combination thereof. In an exemplary embodiment, the methods described herein are implemented in software as executable programs and executed by a specialized or general-purpose digital computer, such as a personal computer, a workstation, a minicomputer, or a mainframe computer.

[0034] The processor 103 is a hardware device for executing software, particularly that stored in the memory 107. The processor 103 may be any custom made or commercially available processor, a central processing unit (CPU), a coprocessor among several processors associated with the control system 111, a semiconductor based microprocessor (in the form of a microchip or chipset), a microprocessor, or generally any device for executing software instructions. The processor 103 controls the operation of the scanning imaging system 101.

[0035] The memory 107 may include any one or combination of volatile memory elements (e.g., random access memory (RAM, e.g., DRAM, SRAM, ADRAM, etc.)) and non-volatile memory elements (e.g., ROM, erasable programmable read-only memory (EPROM), electronically erasable programmable read-only memory EEPROM, programmable read-only memory (PROM), etc.). It should be noted that the memory 107 may have a distributed architecture where various components are located remotely from each other but accessible by the processor 103. The memory 107 may store instructions or data related to at least one other component of the medical analysis system 100.

[0036] The control system 111 further includes a display device 125 for displaying, for example, text, images, etc., in a user interface 129. The display device 125 may be a touch screen display device.

[0037] The medical analysis system 100 further includes a power supply 108 for powering the medical analysis system 100. For example, the power supply 108 may be a battery or an external power source, such as power provided by a standard AC outlet.

[0038] The scanning imaging system 101 includes at least one of an MRI, CT, and PET-CT imager. The control system 111 and the scanning imaging system 101 may or may not be an integral part, i.e., the control system 111 may or may not be external to the scanning imaging system 101.

[0039] The scanning imaging system 101 includes components that are controlled by the processor 103 such that the scanning imaging system 101 provides image data to a control system 111. Settings of the scanning imaging system 101 enable operation of the scanning imaging system 101. The operation of the scanning imaging system 101 is, for example, automatic. Figure 4 illustrates an example of components of the scanning imaging system 101, which is an MRI system.

[0040] The connection between the control system 111 and the scanning imaging system 101 may include, for example, a BUS Ethernet connection, a WAN connection, an Internet connection, or the like.

[0041] In one example, the scanning imaging system 101 provides output data, such as images, corresponding to specified measurements. The control system 111 receives data, such as MR image data, from the scanning imaging system 101. For example, the processor 103 may receive information (automatically or on demand) from the scanning imaging system 101 in a compatible digital format so that such information can be displayed on the display device 125. Such information may include operating parameters, alert notifications, and other information regarding the use, operation, and functionality of the scanning imaging system 101.

[0042] The medical analysis system 100 communicates with other scanning imaging systems 131 and / or databases 133 via a network 130. The network 130 may include, for example, a wireless local area network (WLAN) connection, a wide area network (WAN) connection, a local area network (LAN) connection, or a combination thereof. The database 133 may include information about the patient, the scanning imaging system, the anatomy, the scanning geometry, the scanning parameters, the scan, etc. The database 133 may include, for example, an electronic medical record (EMR) database including the patient's EMR, a radiology information system database, a medical image database, a PACS, a hospital information system database, and / or other databases that compare data that may be used for planning the scanning geometry. The database 133 may include, for example, a training data set for training performed by the subject matter.

[0043] The memory 107 further includes an artificial intelligence (AI) component 150. The AI ​​component 150 may or may not be part of the software component 121. The AI ​​component 150 includes, for example, a trained machine learning model 160. The trained machine learning model 160 receives a set of X-dimensional arrays and provides an image reconstruction output representing one of the set of arrays. For example, the set of X-dimensional arrays includes three arrays Arr1, Arr2, and Arr3, and the trained model reconstructs the image output of one of the arrays, array Arr1, while the other arrays Arr2 and Arr3 are used to consider their correlation with Arr1 in the reconstruction. Thus, the trained model is input with three arrays to receive the image reconstruction output of array Arr1. The image reconstruction output may be a subset or part of the final output. For example, the image reconstruction output is various images from the entire volume. There is a mapping of a subrange of the M-dimensional input to a subrange of the P-dimensional output.

[0044] The machine learning model 160 is trained, for example, by providing multiple sets of XD arrays and associating each set of X-dimensional arrays with a fully sampled k-space reconstructed image that corresponds to an array (Arr1) in the set of X-dimensional arrays.

[0045] 2 is a flow chart of a method for reconstructing a magnetic resonance image of a target volume, such as a subject's brain. An MRI system (such as that shown in FIG. 4) acquires data by scanning or imaging the target volume of the subject.

[0046] In step 201, an M-dimensional array is received including acquired data (by an MRI system). The acquired data may be, for example, k-space data or other image data, such as aliased image data. The acquired data includes at least one dimension, for example representing one of three directions of a target volume. The M dimensions may include further dimensions, such as a time dimension. In one example, M≧2.

[0047] For ease of explanation, we assume M=3, which corresponds to spatial dimensions that are, say, kx, ky, and kz. In this case, an M-dimensional array is defined, say, as mr(x,y,z), where x has some Sx values, y has Sy values, and z has Sz values, i.e. the array mr has a size of Sx*Sy*Sz.

[0048] In step 203, at least one dimension of the M dimensions is selected or identified (i.e., K dimensions are selected (K≧1)). In this example, K=1, but is not limited thereto. For example, the selected dimension may be a dimension selected by a user. For example, a user input is received. The user input indicates the selected dimension. In another example, an automatic selection of a dimension may be performed. For example, a random selection is performed. For ease of explanation, the selected dimension is a dimension x of the array mr. The selected dimension has a corresponding set of values ​​in the M-dimensional array. The set of values ​​includes Sx values, where Sx is the number of phase encodings, for example, if the selected dimension is a dimension corresponding to a phase encoding.

[0049] The dimensions are selected such that the M-1 dimensional array has dimension M-1=X, where X is the dimension of the set of arrays used as input to the trained machine learning model 160. Each array in the set of M-1 dimensional arrays includes at least one dimension that represents spatial frequency information in one of three directions of the target volume.

[0050] In step 205, the values ​​of a subset of the set of values ​​are determined. Following the above example, the values ​​of the subset are, for example, the Sx values ​​of the subset. The values ​​of the subset are determined by selecting the first value of the subset (or the current value) from the Sx values ​​of the set and further selecting the remaining values ​​of the subset. The remaining values ​​and the current value may or may not form a range (the values ​​are consecutive values ​​of an array). In one example, the remaining values ​​are selected such that the current value is the center value. For example, assume that a set of values ​​is provided in the order v1, v2, v3, ..., vSx. If the current value is v4, the remaining values ​​may be v3 and v5, or v2, v3, v5, and v6. v4 is the center value of the subset defined by v3, v4, and v5, or the center value of the subset defined by v2, v3, v4, v5, and v6. In one example, if the current value being processed is the first value v1 (or the last value vSx), the determined subset includes v1 and values ​​beyond v1, such as v2 and v3.

[0051] Each value vx of the set of Sx values ​​defines a 2D array mrx(y,z) associated with the value vx. The 2D array mrx(y,z) represents a 2D slice at the value vx. For example, value v4 is associated with 2D array mr4(y,z), v3 is associated with 2D array mr3(y,z), v2 is associated with 2D array mr2(y,z), and so on. In one example, the remaining values ​​associated with each of the current values ​​are determined by determining the correlation of the 2D arrays of each of the remaining values ​​with the 2D array of the current value. Following the above example, if the current center value is v4, the method determines, for each of the values ​​v1 to vSx (except v4), a correlation value between the 2D array associated with each value and the 2D array mr4(y,z). Only the values ​​vx associated with a correlation value higher than a predefined threshold are selected. The selected value and the current value form a subset.

[0052] In one example, the subset of values ​​comprises a predefined number of values, for example 3, i.e. the subset comprises the current central value and two surrounding values. This is advantageous as it allows an optimal selection method taking into account the correlation of the remaining values ​​with the central value.

[0053] In step 207, for each value of the subset, an M-1 dimensional array is determined that contains the retrieved data corresponding to the value. This results in a set of M-1 dimensional arrays. Following the above example, if the subset determined in step 205 is [v3, v4, v5], then the three arrays mr3(y,z), mr4(y,z), and mr5(y,z) are the determined M-1 dimensional arrays of the three values ​​of the subset [v3, v4, v5], respectively.

[0054] In step 209, the set of M-1 dimensional arrays is input to the trained machine learning model 160 to receive a reconstructed image output from the trained machine learning model. Following the above example, a 2D image of the 2D slice mr4(y,z) associated with the current value v4 is reconstructed.

[0055] In one example, steps 205-209 are repeated for a further subset of values ​​of the selected dimension. Following the example above, a first run of the method is performed for value v1 of the set of values ​​v1 through vSx. Steps 205-209 are repeated for each value in the set of values ​​v2 through Sx. This results in Sx reconstructed 2D slices.

[0056] Figure 3 shows a block diagram illustrating the expected gain in memory cost 301 and increasing computational load 303 versus the input data size provided to a trained machine learning model. Figure 3 shows three different input sizes: one 2D slice, five 2D slices, and a full 3D volume. Figure 3 shows the expected gain, "cost" in terms of memory and computational load versus the input data size.

[0057] FIG. 4 shows a magnetic resonance imaging system 700 as an example of the medical system 100. The magnetic resonance imaging system 700 includes a magnet 704. The magnet 704 is a superconducting cylindrical magnet with a bore 706 inside. Different types of magnets can be used, for example a segmented cylindrical magnet and both so-called open or closed magnets. A segmented cylindrical magnet is similar to a standard cylindrical magnet, except that the cryostat is split into two sections to allow access to the isoplane of the magnet. Such magnets can be used, for example, in combination with charged particle beam therapy. An open magnet has two magnet sections, one above the other, with enough space between them to accommodate the subject 718 to be imaged. The arrangement of the two sections is similar to that of a Helmholtz coil. Inside the cryostat of the cylindrical magnet is a set of superconducting coils. Within the bore 706 of the cylindrical magnet 704 is an imaging zone or volume or anatomical structure 708 in which the magnetic field is strong and uniform enough to perform magnetic resonance imaging.

[0058] Also within the magnet bore 706 are a set of magnetic field gradient coils 710 that are used to acquire magnetic resonance data to spatially encode magnetic spins of a target volume within the imaging or examination volume 708 of the magnet 704. The magnetic field gradient coils 710 are connected to a magnetic field gradient coil power supply 712. The magnetic field gradient coils 710 are intended to be representative. Typically, the magnetic field gradient coils 710 will include three separate coil sets for encoding in three orthogonal spatial directions. The magnetic field gradient power supply supplies electrical current to the magnetic field gradient coils 710. The electrical current supplied to the magnetic field gradient coils 710 is controlled as a function of time and may be ramped or pulsed.

[0059] The MRI system 700 further includes an RF coil 714 for generating RF excitation pulses at the location of the subject 718 and adjacent to the examination volume 708. The RF coil 714 may include, for example, a surface coil or other specialized RF coil set. The RF coil 714 is alternately used for transmitting RF pulses and receiving magnetic resonance signals. For example, the RF coil 714 may be implemented as a transmit array coil including multiple RF transmit coils. The RF coil 714 is connected to one or more RF amplifiers 715.

[0060] The power supply 712 for the magnetic field gradient coils and the RF amplifier 715 are connected to a hardware interface of the control system 111. The memory 107 of the control system 111 includes, for example, a control module. The control module includes computer executable code that enables the processor 103 to control the operation and functionality of the magnetic resonance imaging system 700. It also enables basic operations of the magnetic resonance imaging system 700, such as the acquisition of magnetic resonance data.

[0061] As will be appreciated by those skilled in the art, aspects of the present invention may be embodied as an apparatus, a method, or a computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, all of which are generally referred to as a "circuit," "module," or "system." Additionally, aspects of the present invention may take the form of a computer program product embodied in one or more computer-readable medium(s) having computer-executable code embodied therein.

[0062] Any combination of one or more computer readable media may be utilized. A computer readable medium may be a computer readable signal medium or a computer readable storage medium. As used herein, a "computer readable storage medium" encompasses any tangible storage medium capable of storing instructions executable by a processor of a computing device. A computer readable storage medium may also be referred to as a computer readable non-transitory storage medium. A computer readable storage medium may also be referred to as a tangible computer readable medium. In some embodiments, a computer readable storage medium may also store data accessible by a processor of a computing device. Examples of computer readable storage media include, but are not limited to, floppy disks, magnetic hard disk drives, solid state hard disks, flash memory, USB thumb drives, random access memory (RAM), read only memory (ROM), optical disks, magneto-optical disks, and processor register files. Examples of optical disks include compact disks (CDs) and digital versatile disks (DVDs), such as CD-ROM, CD-RW, CD-R, DVD-ROM, DVD-RW, or DVD-R disks. The term computer-readable storage medium also refers to various types of recording media that can be accessed by a computer device over a network or communication link. For example, data can be obtained via a modem, the Internet, or a local area network. Computer executable code embodied in a computer-readable medium can be transmitted using any medium, including but not limited to wireless, wireline, fiber optic cable, RF, etc., or any suitable combination thereof.

[0063] A computer-readable signal medium may include a propagated data signal in which computer-executable code is embodied, such as, for example, in baseband or as part of a carrier wave. Such a propagated signal may take a variety of forms, including, but not limited to, electromagnetic, optical, or any suitable combination thereof. A computer-readable signal medium is not a computer-readable storage medium, but may be any computer-readable medium that can communicate, propagate, or transfer a program for use in or in connection with an instruction execution system, apparatus, or device.

[0064] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory that is directly accessible by a processor. "Computer storage" or "storage" is an example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments, computer storage is also computer memory and vice versa.

[0065] As used herein, a "processor" encompasses an electronic component capable of executing a program, machine-executable instructions, or computer-executable code. References to a computing device that include a "processor" should be interpreted as potentially including multiple processors or processing cores. For example, a processor may be a multi-core processor. A processor may also refer to a collection of processors within a single computer system or a collection of processors distributed across multiple computer systems. The term computing device should also be interpreted as potentially referring to a collection or network of computing devices each including one or more processors. A computer-executable code may be executed by multiple processors within the same computing device or may be distributed across multiple computing devices.

[0066] Computer executable code may include machine executable instructions, or programs, that cause a processor to perform aspects of the present invention. Computer executable code for carrying out operations of aspects of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages, such as the "C" programming language or similar programming languages, and compiled into machine executable instructions. In some cases, the computer executable code may be in a high-level language or a compiled form and used in conjunction with an interpreter that generates the machine executable instructions on the fly.

[0067] The computer executable code may execute entirely on the user's computer, partially on the user's computer, as stand-alone software, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN). Alternatively, a connection to an external computer may be established (e.g., via the Internet using an Internet Service Provider).

[0068] Aspects of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block or part of the blocks of the flowcharts, diagrams, and / or block diagrams, where applicable, can be implemented in the form of computer executable code by computer program instructions. It will further be understood that combinations of blocks in different flowcharts, diagrams, and / or block diagrams can be combined if not mutually exclusive. Such computer program instructions are provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the instructions are executed by the processor of the computer or other programmable data processing apparatus to generate a machine that creates means for implementing the functions / acts specified in one or more blocks of the flowcharts and / or block diagrams.

[0069] These computer program instructions may also be stored on a computer-readable medium that can instruct a computer, other programmable data processing apparatus, or other device to function in a particular manner, such that the instructions stored on the computer-readable medium produce an article of manufacture that includes instructions that implement the functions / acts specified in one or more blocks of the flowcharts and / or block diagrams.

[0070] Computer program instructions may also be loaded into a computer, other programmable data processing apparatus, or other device to cause the computer, other programmable apparatus, or other device to perform a series of operational steps to generate a computer-implemented process, whereby the instructions executing on the computer or other programmable apparatus provide a process for implementing the functions / acts specified in one or more blocks of the flowcharts and / or block diagrams.

[0071] A "user interface" as used herein is an interface that allows a user or operator to interact with a computer or computer system. A "user interface" is also referred to as a "human interface device." A user interface can provide information or data to an operator and receive information or data from an operator. A user interface can be used to allow a computer to receive input from an operator and provide output from the computer to a user. That is, a user interface can be used to allow an operator to control or operate a computer, or an interface can be used to allow a computer to show the effects of an operator's control or operation. Displaying data or information on a display or graphical user interface is an example of providing information to an operator. Receiving data via a keyboard, mouse, trackball, touchpad, pointing stick, graphics tablet, joystick, gamepad, webcam, headset, shift lever, steering wheel, pedals, wired gloves, dance pad, remote control, and accelerometer are all examples of user interface components that allow for the reception of information or data from an operator.

[0072] As used herein, a "hardware interface" encompasses an interface that allows a processor of a computer system to interact with and control external computing devices and / or devices. A hardware interface allows a processor to send control signals and instructions to external computing devices and / or devices. A hardware interface also allows a processor to exchange data with external computing devices and / or devices. Examples of hardware interfaces include, but are not limited to, a universal serial bus, an IEEE 1394 port, a parallel port, an IEEE 1284 port, a serial port, an RS-232 port, an IEEE-488 port, a Bluetooth® connection, a wireless local area network connection, a TCP / IP connection, an Ethernet connection, a control voltage interface, a MID interface, an analog input interface, and a digital input interface.

[0073] As used herein, a "display" or "display device" encompasses an output device or user interface adapted for displaying images or data. A display can output visual, audio, or tactile data. Examples of displays include, but are not limited to, computer monitors, television screens, touch screens, tactile electronic displays, Braille screens, cathode ray tubes (CRTs), storage tubes, bi-stable displays, electronic paper, vector displays, flat panel displays, vacuum fluorescent displays (VFs), light emitting diode (LED) displays, emissive displays (ELDs), plasma display panels (PDPs), liquid crystal displays (LCDs), organic light emitting diode displays (OLEDs), projectors, and head mounted displays.

[0074] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments.

[0075] Other variations of the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprises" does not exclude other elements or steps, and the singular does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. The computer program can be stored / distributed on any suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, but can also be distributed in other forms, such as via the Internet or other wired or wireless communication systems. Any reference signs in the claims should not be interpreted as limiting the scope. [Explanation of symbols]

[0076] 100 Medical Systems 101 Scanning Imaging System 103 Processors 107 Memory 108 Power supply 109 Bus 111 Control System 121 Software 125 Display 129 User Interface 133 Database 150 AI Components 160 Machine Learning Models 201~209 Method steps 301 Memory Cost 303 Computational load 700 Magnetic Resonance Imaging System 704 Magnet 706 Magnet Bore 708 Imaging Zone 710 Magnetic Field Gradient Coil 712 Magnetic field gradient coil power supply 714 Radio Frequency Coil 715 RF Amplifier 718 Subjects

Claims

1. 1. A medical analysis system for reconstructing magnetic resonance images, comprising a processor and at least one memory storing machine executable instructions, the processor controlling the medical analysis system, execution of the machine executable instructions causing the processor to: providing a trained machine learning model that reconstructs an image from input data; receiving an M-dimensional multidimensional array including M dimensions of acquired data, where M≦4; determining at least a subset of values ​​for at least one selected dimension, K dimensions, K≧1 and K<M, of the M-dimensional multidimensional array; determining, for each value in each of said at least one subset, an M-K dimensional array containing said acquired data corresponding to said value, resulting in a set of M-K dimensional arrays; inputting the set of M-K dimensional arrays into the trained machine learning model and receiving an image reconstruction output from the trained machine learning model; Let them do so, Execution of the machine-executable instructions further causes the processor to determine values ​​of the subset such that a correlation between the M-K dimensional array associated with a median value of the subset and the remaining M-K dimensional arrays of the set is higher than a predefined threshold.

2. 2. The medical analysis system of claim 1, wherein execution of the machine executable instructions further causes the processor to determine additional subsets of values ​​and, for each additional subset of the additional subsets, repeat determining the M-K dimensional array and inputting the set of M-K dimensional arrays.

3. A medical analysis system as described in claim 2, wherein the subset and each of the further subsets do not overlap with each other.

4. 4. The medical analysis system of claim 1, wherein the M dimensional acquired data is 3D k-space data (M=3), the selected dimension (K=1) is the readout direction, and the set of M-1 dimensional arrays represents a corresponding set of 2D slices.

5. The medical analysis system of claim 1 , wherein the subset of values ​​comprises a predefined odd number of values.

6. 6. The medical analysis system of claim 1 , wherein determining the subset of values ​​comprises identifying a central value and two or more surrounding values ​​of the central value along the selected dimension, the subset including the central value and the surrounding values.

7. The medical analysis system of claim 1 , wherein the trained machine learning model is a deep neural network.

8. providing a trained machine learning model for reconstructing an image from input k-space data; receiving an M-dimensional multidimensional array containing M dimensions of acquired data, where M≦4; determining values ​​of at least a subset of at least one selected dimension, K dimensions, K≧1 and K<M, of the M-dimensional multidimensional array; determining, for each value of each of said at least one subset, an M-K dimensional array containing said acquired data corresponding to said value, resulting in a set of M-K dimensional arrays; inputting the set of M-K dimensional arrays into the trained machine learning model and receiving a reconstructed image from the trained machine learning model; determining values ​​of the subset such that a correlation between the M-K dimensional array associated with a central value of the subset and the remaining M-K dimensional arrays of the set is higher than a predefined threshold; A method comprising:

9. A computer program product comprising machine executable instructions for execution by a processor, the execution of the machine executable instructions causing the processor to perform at least a portion of the method of claim 8.

10. 8. A magnetic resonance imaging system including a medical analysis system according to any one of claims 1 to 7, said magnetic resonance imaging system being configured to acquire M-dimensional data.

Images