Detection of artifacts in magnetic resonance images using neural networks.

Abstract

Disclosed herein is a medical system comprising a memory storing machine executable instructions and an image processing module, the image processing module comprising an image processing neural network portion and an artificial structure prediction portion, the image processing module comprising an input configured to receive magnetic resonance data, the image processing neural network portion comprising a first output configured to output a corrected magnetic resonance image in response to receiving the magnetic resonance data at the input, and the artificial structure prediction portion comprising a second output configured to output artificial structure data describing the likelihood of an artificial structure in the corrected magnetic resonance image.

Description

[Technical field]

[0001] The present invention relates to magnetic resonance imaging, and more particularly to processing magnetic resonance images using neural networks. [Background technology]

[0002] Recently, neural networks have been used to reconstruct, filter or correct magnetic resonance images. The advantage of this approach is that neural networks are very fast and can be trained to perform complex image and reconstruction tasks. The disadvantage of using neural networks is that they can add structures to the image that are not real. This is called hallucinations in the literature. In this specification, artificial structures and hallucinations are used synonymously. Summary of the Invention [Problem to be solved by the invention]

[0003] The presence of artifacts or illusions in magnetic resonance images can have a negative effect since the physician examining the magnetic resonance images may see different anatomical features than in the actual subject.

[0004] US Patent Application Publication No. 2020 / 0249300 discloses reconstructing a magnetic resonance image from accelerated magnetic resonance imaging (MM) data. In one embodiment, a method for reconstructing a magnetic resonance (MR) image includes estimating a plurality of sets of coil sensitivity maps from undersampled k-space data, estimating undersampled k-space data acquired by a multi-coil radio frequency (RF) receiver array, reconstructing a plurality of initial images using the undersampled k-space data and the plurality of sets of estimated coil sensitivity maps, iteratively reconstructing the plurality of images using the initial images and the plurality of sets of coil sensitivity maps to generate a plurality of final images, each of the plurality of images corresponding to a different one of the plurality of sets of sensitivity maps, and combining the plurality of final images output from the trained deep neural network to generate an MR image. [Means for solving the problem]

[0005] The invention provides a medical system, a computer program product and a method in the independent claims. Embodiments are given in the dependent claims.

[0006] An embodiment may provide a means for detecting artifacts in magnetic resonance images by using an image processing module comprising an image processing neural network portion that receives magnetic resonance data and outputs a corrected magnetic resonance image in response. The image processing neural network portion is trained from a ground truth image set of selected (e.g. magnetic resonance) images. The image processing module also comprises an artifact detection portion that outputs artifact data in response to the magnetic resonance data being input to the image processing neural network. These artifact data represent one or more aspects of the relationship of the receiver magnetic resonance data for aggregating ground truth information. The receiver magnetic resonance data represent the corrupted image to be corrected. The correction is performed by the image processing neural network portion that is trained based on a specific ground truth data set of the selected images. The artifact prediction portion returns information that represents the likelihood of an artifact, such as an illusion. For this purpose, the artifact portion is trained based on a global ground truth that represents certain aspects of a more global set of images (compared to the set of known images that form the ground truth for the image processing neural network portion). Thus, the image processing network can return, on the one hand, a corrected magnetic resonance image at its first output matching the ground truth data on which the image processing neural network part was trained, and on the other hand, its second output also returns data representative of the aspect of the relationship of the received magnetic resonance data to the aggregate ground truth data. Thus, the image processing module can jointly balance the corrected images returned by the image processing neural network part and the artificial structure part between a close correspondence with the ground truth formed by the training data set of known images and an appropriate correction in some aspects with the aggregate ground truth. Furthermore, it is noted that the image processing neural network and the artificial structure prediction network can be jointly trained on the basis of the known images as well as the aggregate ground truth. Various ways of implementing the image processing module are described below.In particular, the aggregate ground truth may be formed by a reference magnetic resonance image, which may be a template or atlas image derived from a significantly more extensive image collection than the training dataset of known images. The relationship between aspects of the receiver magnetic resonance (input) data and aspects of the aggregate ground truth may relate to statistical measures of difference maps or to comparisons of volumetric proportion data between segmented anatomical structures. The ground truth of the image processing neural network part may be selected at the medical institution where the medical system is operated, or by a specific radiologist or team of radiologists referring to the patient to be examined for the magnetic resonance imaging examination.

[0007] In one aspect, the present invention provides a medical system comprising a memory storing machine executable instructions, which may be machine executable code or instructions, and an image processing module, the image processing module comprising an image processing neural network portion and an artificial structure prediction portion, the image processing module comprising an input configured to receive magnetic resonance data.

[0008] As used herein, magnetic resonance data encompasses either a magnetic resonance image or k-space data that can be used to reconstruct a magnetic resonance image. Thus, the image processing module is configured to receive as input k-space data or a magnetic resonance image. If the image processing module receives k-space data, the image processing module reconstructs a magnetic resonance image from the k-space data. If the magnetic resonance data is a magnetic resonance image, the image processing module is used to perform some image processing task on the magnetic resonance image. This can be, for example, noise removal, artifact removal, motion reduction, or other such similar tasks.

[0009] The image processing neural network comprises a first output configured to output a corrected magnetic resonance image in response to receiving magnetic resonance data at the input. The artificial structure prediction portion comprises a second output configured to output artificial structure data describing the likelihood of an artificial structure in the corrected magnetic resonance image. An artificial structure as used herein encompasses an anatomical structure or feature in the corrected magnetic resonance image that is artificially added by the image processing neural network portion. Artificial structures in magnetic resonance images are also commonly known in the literature as hallucinations. The use of neural networks to construct magnetic resonance images from k-space data or to perform various image processing tasks such as artifact or noise removal is well established. However, a common technical problem is the inability to examine neural networks and understand what limitations they have, what limitations they have, and what limitations are real or what limitations are modified and have so-called hallucinations.

[0010] The incorporation of the artifact prediction portion provides a means of identifying possible artifacts or illusions in the corrected magnetic resonance image. This may be useful, for example, when a doctor or other medical professional is examining the corrected magnetic resonance image and wants to know which parts of the corrected magnetic resonance image may have or be erroneous. This can also be used as a control function. For example, if an image processing module is used to reconstruct an image from raw k-space data, a different neural network or even an algorithmic method may be used instead to reconstruct the corrected magnetic resonance image, thereby avoiding possible artifacts or illusions in the magnetic resonance image.

[0011] The medical system further comprises a computing system. Execution of the machine executable instructions causes the computing system to receive magnetic resonance data. Execution of the machine executable instructions further causes the computing system to receive a corrected magnetic resonance image at a first output and man-made structure data at a second output in response to input of the magnetic resonance data to the input of the image processing module. Execution of the machine executable instructions further causes the computing system to provide a warning signal in response to the man-made structure data meeting a predetermined criterion.

[0012] For example, if the artificial data structures meet a predetermined criteria, a warning signal can be provided. The warning signal can take different forms in different examples. In one example, the warning signal can be an audible or tactile warning. In another example, the warning signal can be a warning rendered by a display connected to the medical system. In yet another example, the warning signal can be a heat map or probability map indicating which locations in the corrected magnetic resonance image are likely to have artificial or illusory structures.

[0013] The medical system may be integrated into different types of systems. In one example, the medical system processes k-space data into a magnetic resonance image or may be a stand-alone system that processes magnetic resonance images. It may be located, for example, on a server or remotely located as a web service. In another example, the machine executable instructions and image processing modules are integrated into algorithms or image processing modules used to reconstruct the magnetic resonance image. In yet another example, the medical system may incorporate or be integrated into a medical imaging system, such as a magnetic resonance imaging system.

[0014] In another embodiment, the artificial structure data includes template matching parameters for matching the corrected magnetic resonance image to a reference magnetic resonance image. Execution of the machine executable instructions further causes the computing system to calculate an image difference map between the reference magnetic resonance image and the corrected magnetic resonance image using the template matching parameters. For example, the corrected magnetic resonance image may be transformed or transformed to match the reference magnetic resonance image or vice versa. The image difference map may then look at individual pixels or voxels of the two images and calculate the difference between the two of them. In some examples, the image difference map may be further processed.

[0015] For example, the image difference map may be thresholded so that only differences above a certain value or within a certain neighborhood are registered. Computing the image difference map may also involve some pre-processing of the corrected magnetic resonance image. For example, the contrast of a reference magnetic resonance image is determined, a corrected magnetic resonance image is also determined, and then one or the other image is adjusted so that both have the same contrast before the image difference map is calculated.

[0016] Execution of the machine executable instructions further causes the computational system to determine whether the image difference map algorithmically meets a predetermined criterion by determining whether the image difference map exceeds a predetermined statistical measure. A warning signal is provided if the image difference map meets the predetermined criterion. As previously described, the predetermined statistical measure may be a thresholding process to see if there are voxels that exceed a certain difference. Additional statistical criteria may also be applied. For example, the number of different voxels may also be one or more criteria used to determine whether a warning signal is provided. The predetermined criteria may also include looking at how clusters of voxels that exceed a predetermined threshold are grouped together. For example, if there are one or two isolated voxels, this may not trigger a warning signal, but if there is an area above a certain size that exceeds a threshold, this may be used to trigger a warning signal in some examples.

[0017] The template matching parameters may be different in different examples. For example, in one case it may be a three-dimensional deformation field that defines the pixel-by-pixel or voxel shift between the two images. The template matching parameters may also represent a lower dimensional shift based on identifying landmarks in both images, which may then be used to perform the transformation process.

[0018] This embodiment can be beneficial because it is often difficult to determine whether there are hallucinations in a magnetic resonance image, and therefore comparing the corrected magnetic resonance image to a reference magnetic resonance image can provide a very effective and objective means of evaluating the image to see if there are artificial structures or hallucinations.

[0019] In another embodiment, execution of the machine executable instructions further causes the computing system to receive subject metadata describing the subject, and to select a reference magnetic resonance image from a template database using the subject metadata. For example, age, sex, and other details may be used to select a particular reference magnetic resonance image. This may have the advantage that the reference magnetic resonance image is closer to the subject's anatomy.

[0020] In another embodiment, the memory further comprises an image generation neural network configured to generate a reference magnetic resonance image in response to receiving the magnetic resonance data. For example, both the artificial structure prediction part and the image processing neural network part are realized as neural networks. For example, an autoencoder or an image generator trained using an adversarial loss term (as in the generative adversarial network concept) can be used to generate the reference magnetic resonance image.

[0021] In one example, only subject metadata is used as input to an image generation neural network, where the image generation neural network can be trained using subjects with normal anatomy.

[0022] In another embodiment, the artificial structure prediction part is implemented as a template matching algorithm. This embodiment can be beneficial because the image processing neural network part and the artificial structure prediction part can be two separate executable groups of code or program. For example, the artificial structure prediction part can take a corrected magnetic resonance image as its input, and then use an algorithm to determine template matching parameters.

[0023] In another embodiment, the artificial structure data includes a three-dimensional segmentation mask defining a plurality of predefined anatomical structures. For example, the artificial structure prediction portion may segment the corrected magnetic resonance image to determine the three-dimensional segmentation mask. This may be done using a segmentation algorithm, such as a conventional neural network or a deformable shape algorithm. The machine executable instructions further cause the computing system to receive predefined volumetric ratio data describing one or more ratios between the plurality of predefined anatomical structures. Execution of the machine executable instructions further causes the computing system to calculate, from the artificial structure data, measured volumetric ratio data describing one or more ratios between the plurality of predefined anatomical structures. Execution of the machine executable instructions further causes the computing system to determine whether a predefined criterion is met by comparing the predefined volumetric ratio data with the measured volumetric ratio data.

[0024] In this embodiment, the corrected magnetic resonance image can be segmented to determine these different three-dimensional segmentation masks.Then, the various ratios between these different volumes can be calculated, and these ratios can be compared with predetermined volume ratio data.If there is an artificial structure in the corrected magnetic resonance image or illusion, this is very likely to be detected by the imbalance of the difference between the predetermined volume ratio data and the measured volume ratio data.

[0025] The segmentation can be performed in various ways. It may be, for example, a single neural network with Y-Net that can generate both the corrected magnetic resonance image and the artificial structure data. The image processing module may be a separate neural network and the artificial structure prediction part may be a separate program or executable code. For example, the artificial structure prediction part may be a segmentation algorithm or may be implemented as a separate neural network. This embodiment also has the advantage that the corrected magnetic resonance image is compared with real or raw reference data for the anatomical structures. This may be very effective in detecting artificial structures or illusions.

[0026] In another embodiment, execution of the machine executable instructions causes the computing system to receive subject metadata. Execution of the machine executable instructions further causes the computing system to select predetermined volumetric ratio data from a volumetric ratio database using the subject metadata. This can be particularly useful since different volumetric ratios can be fine-tuned to different parameters describing the subject, such as age and sex.

[0027] In another embodiment, the artificial structure prediction portion is implemented as an image segmentation algorithm, which may be, for example, a conventional neural network trained to perform image segmentation, or it may be, for example, an algorithmic segmentation algorithm, such as a deformable model.

[0028] In another embodiment, the artificial structure prediction part is implemented as a neural network. The artificial structure prediction part takes the corrected magnetic resonance image as input. This is applicable to both the template matching embodiment and the embodiment using the 3D segmentation mask. The neural network can be implemented as a ResNet neural network or a U-net neural network, for example, as some examples. The artificial structure prediction part can be trained, for example, by acquiring a set of magnetic resonance images and then manually labeling the template matching parameters and the segmentation. This training data can then be used using deep learning techniques to train the artificial structure prediction part, for example, when the artificial structure prediction part is a convolutional neural network.

[0029] In another embodiment, the artificial structure prediction portion is implemented as a neural network. The output artificial structure data is a spatially dependent probability map describing the likelihood of an artificial structure in the corrected magnetic resonance image. In this embodiment, this spatially dependent probability map is output directly. There are several different ways to implement this. In one case, the artificial structure prediction portion and the image processing neural network portion can be implemented together as a single neural network, for example, as a Y-net neural network, as described below. In another example, the artificial structure prediction portion and the image processing neural network portion are implemented as two separate neural networks. For example, these two neural networks can be two separate U-net neural networks.

[0030] In another embodiment, the image processing neural network is a Y-net neural network. The Y-net neural network is formed from a U-net neural network structure configured to output a corrected magnetic resonance image at a first output in response to receiving the magnetic resonance data. The Y-net further comprises a decoding branch configured to output artificial structure data describing artificial structures in the corrected magnetic resonance image at a second output in response to receiving the magnetic resonance data. The decoding branch is connected to the U-net neural network structure. The U-net neural network structure includes an image processing neural network portion. The decoding branch includes an artificial structure prediction portion.

[0031] As used herein, the Y-net neural network encompasses the U-net neural network with an additional decoding branch. This embodiment may be beneficial because the addition of an additional decoding branch to the U-net may be an effective way to detect hallucinations generated by the U-net neural network structure. The use of the Y-net neural network may be used in many of the above-mentioned embodiments. For example, the decoding branch may be used to output template matching parameters. In this case, the U-net may be trained by using data with a known input of magnetic resonance data, either image or k-space data, and then the ground truth data may be the corrected magnetic resonance image and appropriate template matching parameters.

[0032] Y-net may also be useful in embodiments with a 3D segmentation mask. In this case, the decoding branch performs segmentation or predicts the segmentation of the corrected magnetic resonance image. In this case, the training data may be known magnetic resonance data in either image or k-space paired with the corrected magnetic resonance image and ground truth data with a reference segmentation mask, and deep learning may be used to train the Y-net neural network. Similarly, Y-net may be used to very effectively directly generate a space-dependent probability map using the decoding branch. For example, this may be done by acquiring magnetic resonance data again in k-space or image space and using the U-net part to generate a corrected magnetic resonance image. The template data or the human can then pass through and identify the regions of the image that contain hallucinations, which can be used as ground truth data. The Y-net may then be trained again using deep learning.

[0033] In another embodiment, the magnetic resonance data describes the subject's brain. This embodiment may work particularly well because the structure or normal structure of neuroanatomical structures is fairly consistent between individuals. For example, the use of reference magnetic resonance images or volume data may therefore be very accurate in predicting the hallucinated structures in the corrected magnetic resonance images.

[0034] In another embodiment, the U-net neural network structure includes a lowest resolution convolutional layer. The decoding branch is connected to the lowest resolution convolutional layer. This embodiment may be particularly useful because the lowest resolution convolutional layer usually has a lower resolution than the corrected magnetic resonance image, but nevertheless contains data that can accurately describe where the hallucinated structures may be. In this case, it may be useful to predict template matching parameters, volume fractions, or even directly output a spatially dependent probability map.

[0035] In another embodiment, the magnetic resonance data is image data. In this case, the magnetic resonance data is an already reconstructed magnetic resonance image. The corrected magnetic resonance image is a correction of the magnetic resonance data.

[0036] In another embodiment, the image processing module is incorporated into a magnetic resonance imaging reconstruction algorithm configured to reconstruct a clinical magnetic resonance image in response to receiving the k-space data. The magnetic resonance data is an intermediate magnetic resonance image calculated from the k-space data during reconstruction of the clinical magnetic resonance image. This embodiment may be particularly beneficial as it may enable a means to provide higher reliability when a neural network is incorporated into a conventional magnetic resonance imaging reconstruction algorithm.

[0037] An example of this is the use of the image processing module as a denoising filter used in a compressed sensing algorithm. In the compressed sensing algorithm, there is a denoising filter used before the data consistency step. The warning signal can then be used to re-trigger data acquisition or to modify the operation of the magnetic resonance imaging reconstruction algorithm. For example, if there is a high possibility of a large number of hallucinations, the algorithm can switch to using a traditional algorithmic noise filter instead of a neural network-based noise filtering. This can help to increase the confidence that the clinical magnetic resonance image is more accurate and that it is less likely to contain unrealistic or hallucinatory artifacts.

[0038] In another embodiment, the magnetic resonance data is k-space data. In this case, the image processing module takes this k-space data as input and outputs a corrected magnetic resonance image. This structure can take a variety of different forms. In one case, the image processing neural network part can perform the entire reconstruction, and the artificial structure prediction part can be completely separate. For example, the output image is used for template matching or segmentation is applied to it. In another case, a Y-net structure can be trained to receive the k-space data and then output both the corrected magnetic resonance image and the artificial structure prediction part.

[0039] In another embodiment, the medical imaging system further comprises a magnetic resonance imaging system. The memory further comprises pulse sequence commands configured to control the magnetic resonance imaging system to acquire magnetic resonance data to describe the imaging zone according to a magnetic resonance imaging protocol. In this example, acquiring magnetic resonance data can have two meanings. In one case, it can simply mean acquiring k-space data. This is the case when the image processing module reconstructs a corrected magnetic resonance image directly from the k-space data. Another interpretation that the magnetic resonance commands can take is that the magnetic resonance commands are magnetic resonance images, and the pulse sequence controls the magnetic resonance imaging system to acquire k-space commands, which are then reconstructed by the computing system into magnetic resonance commands. Execution of the machine executable instructions further causes the computing system to control the magnetic resonance imaging system to acquire magnetic resonance data. As previously mentioned, this can also include image reconstruction.

[0040] In another embodiment, the image processing neural network portion is configured to perform noise removal.

[0041] In another embodiment, the neural network portion is configured to perform artifact correction.

[0042] In another embodiment, the neural network portion is configured to perform motion compensation.

[0043] In another embodiment, the neural network portion is configured to perform super-resolution.

[0044] In another embodiment, the image processing neural network portion is configured to perform the deblurring.

[0045] In another embodiment, the image processing neural network portion is configured to perform a combination of the above-mentioned image processing techniques.

[0046] In another aspect, the present invention provides a computer program product comprising machine executable instructions and an image processing module, which may be machine executable instructions for execution by a computing system. The image processing module comprises an image processing neural network portion and an artificial structure prediction portion. The image processing module comprises an input configured to receive magnetic resonance data. The image processing neural network portion comprises a first output configured to output a corrected magnetic resonance image in response to receiving the magnetic resonance data at the input. The artificial structure prediction portion comprises a second output configured to output artificial structure data describing the likelihood of an artificial structure in the corrected magnetic resonance image.

[0047] Execution of the machine executable instructions causes the computing system to receive magnetic resonance data. Execution of the machine executable instructions further causes the computing system to receive both a corrected magnetic resonance image at a first output and man-made structure data at a second output in response to input of the magnetic resonance data to an input of the image processing module. Execution of the machine executable instructions further causes the computing system to provide a warning signal in response to the man-made structure data meeting predetermined criteria.

[0048] In another aspect, the present invention provides a method of medical imaging, the method comprising receiving magnetic resonance data, which again may be k-space data or a magnetic resonance image, depending on the particular example. The method further comprises receiving the corrected magnetic resonance image at a first output of the image processing module, and receiving artifact data at a second output of the image processing module in response to inputting the magnetic resonance data at the input of the image processing module. The artifact data describes the likelihood of artifacts in the corrected magnetic resonance image. The image processing module comprises an image processing neural network portion and an artifact prediction portion. The image processing module comprises an input configured to receive the magnetic resonance data. The image processing neural network comprises a first output configured to output the corrected magnetic resonance image in response to receiving the magnetic resonance data at the input. The artifact prediction portion comprises a second output that outputs the artifact data. The method further comprises providing a warning signal in response to the artifact data meeting a predetermined criterion.

[0049] It will be understood that one or more of the above-described embodiments of the present invention may be combined, unless the combined embodiments are mutually exclusive.

[0050] As will be appreciated by one of ordinary skill 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 may be referred to generally herein as a "circuit," "module," or "system," and further, 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 thereon.

[0051] 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 or a computing system of a computing device. A computer readable storage medium may 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 be capable of storing data that can be accessed by a computing system 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 register files of a computing system. Examples of optical disks include compact disks (CDs) and digital versatile disks (DVDs), such as CDROM, CDRW, CDR, DVDROM, DVDRW, or DVDR 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 retrieved over a modem, over the Internet, or over a local area network. Computer executable code embodied on a computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wireline, fiber optic cable, RF, etc., or any suitable combination of the foregoing.

[0052] A computer-readable signal medium may include a propagated data signal having computer-executable code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated signal may take any of 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 transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0053] "Computer memory" or "memory" is one example of a computer-readable storage medium. Computer memory is any memory directly accessible to a computing system. "Computer storage" or "storage" is a further example of a computer-readable storage medium. Computer storage is any non-volatile memory computer-readable storage medium. In some embodiments, computer storage may be computer memory, and vice versa.

[0054] As used herein, a "computing system" encompasses electronic components capable of executing programs, machine-executable instructions, or computer-executable code. References to a computing system, including examples of "computing system," should be interpreted as including more than one computing system or processing core, as the case may be. A computing system may be, for example, a multi-core processor. A computing system may also refer to a collection of computing systems within a single computer system or distributed among multiple computer systems. The term computing system should also be interpreted as referring to a collection or network of computing devices, each possibly comprising a processor or computing system. Machine-executable code or instructions may be executed by multiple computing systems or processors, which may be within the same computing device or may be distributed across multiple computing devices.

[0055] Machine-executable instructions or computer-executable code may comprise instructions or programs that cause a processor or other computing system to perform aspects of the present invention. Computer-executable code for performing operations for 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 traditional 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 the form of a high-level language or in a pre-compiled form and may be used in conjunction with an interpreter that generates the machine-executable instructions on the fly. In other cases, the machine-executable instructions or computer-executable code may form a program for a programmable logic gate array.

[0056] The computer executable code may run entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer, 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), or the connection may be to an external computer (e.g., via the Internet using an Internet Service Provider).

[0057] 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 present invention. It should be understood that each block or portion of a block in the flowcharts, diagrams, and / or block diagrams may be implemented by computer program instructions in the form of computer executable code, where applicable. Furthermore, it should be noted that combinations of blocks in different flowcharts, diagrams, and / or block diagrams may be combined, if not mutually exclusive. These computer program instructions may be provided to a general purpose computer, special purpose computer, or other programmable data processing device computing system to generate a machine such that the instructions, executed via the computer or other programmable data processing device computing system, create means for performing the functions / operations specified in the flowchart and / or block diagram blocks or blocks.

[0058] These machine-executable instructions or computer program instructions may 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 function / act specified in a block or blocks of the flowcharts and / or block diagrams.

[0059] The machine-executable instructions or computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps executed on the computer, other programmable apparatus, or other device to produce a computer-implemented process, such that the instructions executing on the computer or other programmable apparatus provide a process for implementing the function / operation specified in the flowchart and / or block diagram block or blocks.

[0060] 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 sometimes referred to as a "human interface device" and can provide information or data to an operator and / or receive information or data from an operator. A user interface can allow input from an operator to be received by a computer and can provide output from the computer to a user. In other words, a user interface can allow an operator to control or manipulate a computer and an interface can allow a computer to show the effect of the operator's control or manipulation. The display of 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, pedals, wired gloves, remote control, and accelerometer are all examples of user interface components that allow for the reception of information or data from an operator.

[0061] As used herein, a "hardware interface" encompasses an interface that allows a computing system of a computer system to interact with and / or control external computing devices and / or devices. A hardware interface may allow a computing system to send control signals or instructions to external computing devices and / or devices. A hardware interface may also allow a computing system 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 RS232 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 MIDI interface, an analog input interface, and a digital input interface.

[0062] As used herein, a "display" or "display device" encompasses an output device or user interface adapted to display images or data. A display can output visual, audio, and / 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 (liquid), memory tubes, bi-stable displays, electronic paper, vector displays, flat panel displays, vacuum fluorescent displays (VF), light emitting diode (LED) displays, electroluminescent displays (ELD), plasma display panels (PDP), liquid crystal displays (LCD), organic light emitting diode displays (OLED), projectors, and head mounted displays.

[0063] K-space data is defined herein as being the recorded measurements of radio frequency signals emitted by atomic spins using the antenna of a magnetic resonance apparatus during a magnetic resonance imaging scan.

[0064] A magnetic resonance imaging (MRI) image or MR image is defined herein as a reconstructed two- or three-dimensional visualization of the anatomical data contained within the k-space data. This visualization can be performed using a computer.

[0065] In the following, preferred embodiments of the invention will be described, by way of example only, with reference to the drawings in which: [Brief description of the drawings]

[0066] [Figure 1] 1 shows an example of a medical system. [Diagram 2] 2 shows a flow chart illustrating a method of using the medical system of FIG. [Diagram 3] 1 illustrates an exemplary image processing module. [Figure 4] 4 shows a further example of an image processing module. [Diagram 5] 1 illustrates a further example of a medical system. [Figure 6] An example of how a template matching neural network can be trained is given below. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0067] Like numbered elements in these figures are equivalent elements or perform the same function. An element described above is not necessarily discussed in a subsequent figure if there is functional equivalence.

[0068] 1 illustrates an example of a medical system. In this example, the medical system 100 is shown to include a computer 102 having a computing system 104. The computing system 104 is shown to be in communication with an optional hardware interface 106 and an optional user interface 108. The optional hardware interface 106 allows the computing system 104 to communicate and exchange data with other components to control their operation and functionality. The user interface 108 may allow an operator or user to control the operation and functionality of the medical system 100.

[0069] The computing system 104 is further shown to be in communication with a memory 110. The memory 110 is intended to represent various types of memory and storage devices that may be in communication with the computing system 104. The memory 110 is shown to include machine-executable instructions 120. The machine-executable instructions 120 enable the computing system 104 to perform various computational and control tasks. For example, the machine-executable instructions 120 may include instructions that enable the computing system 104 to perform image processing or even reconstruct magnetic resonance images from k-space data. The memory 110 is further shown to include an image processing module 122.

[0070] The image processing module has one input and two outputs. The input is configured to receive magnetic resonance data, which in various examples can be either k-space data or a magnetic resonance image. A first output of the image processing module 122 is configured to output a corrected magnetic resonance image when magnetic resonance data 124 is input. A second output of the image processing module 122 is configured to output artifact data. The artifact data describes the likelihood of artifacts in the corrected magnetic resonance image. This can take different forms in different examples, for example, in some examples it can be a heat diagram or a plot showing the probability as a function of location in the corrected magnetic resonance image. In other cases it can be a composite score of the entire corrected magnetic resonance image or various parts of the corrected magnetic resonance image.

[0071] The memory 110 is further shown as including the magnetic resonance data 124. The memory is further shown as including the corrected magnetic resonance image 126 and artifact data 128 received from the image processing module 122 in response to the input of the magnetic resonance data 124. The memory 110 is further shown as including a warning signal 130. The warning signal 130 can take different forms in different examples. For example, in some examples it can cause a display or interaction with the user interface 108. In other examples, the warning signal 130 may itself be to display or store values ​​of various probabilities that the corrected magnetic resonance image 126 has artifacts or hallucinations therein.

[0072] Figure 2 shows a flow chart illustrating a method of operating the medical system 100 of Figure 1. First, in step 200, magnetic resonance data 124 is received. Next, in step 202, a corrected magnetic resonance image 126 is received at a first output of the output section and artificial structure data 128 is received at a second output in response to input of the magnetic resonance data 124 to the input of the image processing module 122. Finally, in step 204, a warning signal 130 is provided depending on whether the artificial data structure meets a predetermined criterion.

[0073] 3 shows one architecture of the image processing module 122. The image processing module is shown as having an input 300 and a first output 302 with a second output 304. Within the image processing module 122, the image processing neural network portion 306 and the artificial structure prediction portion 308 are separate. The image processing neural network portion 306 is an independent image processing neural network. It receives the magnetic resonance data 124 at its input and outputs the corrected magnetic resonance image 126.

[0074] The corrected magnetic resonance image 126 is directly input to the artificial structure prediction section 308, which then outputs the artificial structure data 128 at the second output 304. The artificial structure prediction section 308 can be implemented, for example, algorithmically or as a neural network. In one example, the artificial structure prediction section learns template matching between the corrected magnetic resonance image 126 and a reference magnetic resonance image. An image difference map can then be calculated between the reference magnetic resonance image and the corrected magnetic resonance image 126. This is beneficial because artificial or illusory structures in the corrected magnetic resonance image 126 can be detected using means independent of the neural network.

[0075] In another example, the artificial structure prediction portion 308 includes a segmentation algorithm. The segmentation algorithm segments the image processing module 122. Again, this can be implemented as a neural network that performs the segmentation. However, this can be implemented algorithmically in a simple manner using a variety of different algorithms such as shape deformable models. These segmentations are equivalent to three-dimensional segmentation masks that define a number of predefined anatomical structures in the corrected magnetic resonance image 126. Various ratios between these volumes can be calculated and then compared to a number of predefined anatomical structure ratios. If the ratio between the measured data and the reference data changes more than a predefined amount, this can be used to trigger a warning that there may be artificial or hallucinated structures in the corrected magnetic resonance image. In some cases, this warning signal may be for the entire image or may be localized to where the various volumes that detected the discrepancy are located. In some cases, it may provide information about the location, and in other cases, it may just provide a warning for the entire corrected magnetic resonance image.

[0076] 4 shows a further structure of the image processing module 122. In this example, there is a single neural network 400 comprising an image processing neural network part 306 and an artificial structure prediction part 308. In this case, the artificial structure prediction part 308 is part of the neural network, as is the image processing neural network part 306. The neural network 400 may be, for example, a Y-Net. In this case, the single neural network 400 may directly output the spatially dependent probability map to the second output part 304.

[0077] 5 illustrates a further example of a medical system 500. The example 500 illustrated in FIG. 5 is similar to the example 100 of FIG. 1, except that it further comprises a magnetic resonance imaging system controlled by the computing system 104.

[0078] The magnetic resonance imaging system 502 comprises a magnet 504. The magnet 504 is a superconducting cylindrical magnet with a bore 506 through it. Different types of magnets can be used, for example both split cylindrical magnets and so-called open magnets. Split cylindrical magnets are similar to standard cylindrical magnets, except that the cryostat is split into two sections to allow access to the magnet's isoplane, such magnets can be used, for example, in conjunction with charged particle beam therapy. Open magnets have two magnet sections, one above the other, with a space between them large enough to accommodate the subject, i.e. an arrangement of the two section areas similar to the area of ​​a Helmholtz coil. Open magnets are popular because they are less confining to the subject. Inside the cryostat of the cylindrical magnet is a collection of superconducting coils.

[0079] Within the bore 506 of the cylindrical magnet 504 is an imaging zone 508 where the magnetic field is strong and sufficiently uniform to perform magnetic resonance imaging. A field of view 509 is shown within the imaging zone 508. Acquired magnetic resonance data is typically acquired for the field of view 509. A region of interest may be identical to the field of view 509 or may be a sub-volume of the field of view 509. A subject 518 is shown supported by a subject support 520 such that at least a portion of the subject 518 is within the imaging zone 508 and the field of view 509.

[0080] Also within the magnet bore 506 are a set of magnetic field gradient coils 510 used for preliminary magnetic resonance data acquisition for spatially encoding magnetic spins within the imaging zone 508 of the magnet 504. The gradient coils 510 are connected to a gradient coil power supply 512. The magnetic field gradient coils 510 are intended to be representative. Typically, the magnetic field gradient coils 510 include three separate coil sets for spatially encoding in three orthogonal spatial directions. The gradient power supply supplies current to the magnetic field gradient coils 510. The current supplied to the magnetic field gradient coils 510 is controlled as a function of time and may be ramped or pulsed.

[0081] Adjacent to the imaging zone 508 is a radio frequency coil 514 for manipulating the orientation of magnetic spins in the imaging zone 508 and for receiving radio transmissions from the spins in the imaging zone 508. A radio frequency antenna may include multiple coil elements. A radio frequency antenna may also be referred to as a channel or antenna. The radio frequency coil 514 is connected to a radio frequency transceiver 516. The radio frequency coil 514 and the radio frequency transceiver 516 may be replaced by separate transmit and receive coils, as well as separate transmitters and receivers. It is understood that the radio frequency coil 514 and the radio frequency transceiver 516 are representative. The radio frequency coil 514 is also intended to represent a dedicated transmit antenna and a dedicated receive antenna. Similarly, the transceiver 516 may represent a separate transmitter and receiver. The radio frequency coil 514 may also have multiple receive / transmit elements, and the radio frequency transceiver 516 may have multiple receive / transmit channels. The transceiver 516 and tilt controller 512 are shown as being connected to the hardware interface 106 of the computer system 102 .

[0082] The memory 110 is further shown as including pulse sequence commands 530 that may enable the computing system 104 to control the magnetic resonance imaging system 502 to acquire k-space data. In some examples, the magnetic resonance data 124 is k-space data. In other examples, the magnetic resonance data 124 is reconstructed from the acquired k-space data. The memory 110 is further shown as including an optional reference magnetic resonance image 532. The reference magnetic resonance image 532 may be, for example, from an anatomical atlas or may be a template image believed to represent the normal anatomy of the subject.

[0083] When the artificial structure prediction portion 308 determines the template matching parameters, the artificial structure data 128 is, for example, a difference image between the reference magnetic resonance image 532 and the corrected magnetic resonance image 126 calculated using the template matching parameters to match the two images. The memory 110 is further shown as including optional subject metadata 534. The optional subject metadata 534 may include data describing the subject, such as age, weight, and / or sex, which may be used to select the reference magnetic resonance image 532 from an optional template database 536.

[0084] Alternatively, memory 110 is also shown as including an optional image generation neural network 538, which can, for example, acquire subject metadata 534 and has been trained to generate reference magnetic resonance images 532. For example, image generation neural network 538 can be part of a trained GAN neural network.

[0085] The memory 110 is further shown as including predetermined volume ratio data 540, which may be used by embodiments in which the artificial structure prediction portion 308 or the artificial structure data includes a three-dimensional segmentation mask defining a plurality of predetermined anatomical structures. The memory 110 is also shown as including an optional volume ratio database 542, which may be used, for example, to provide the predetermined volume ratio data 540 in response to providing the subject metadata 534.

[0086] As mentioned above, neural network-based correction of artifacts can lead to artificial structures (hallucinations) in the resulting images. These synthetic structures can have a realistic appearance, which can complicate image interpretation and even lead to misdiagnosis.

[0087] The disclosed invention overcomes this problem by recasting the image correction setup into a multitasking framework that in some instances allows for automatic detection of network hallucinations. A dedicated multitasking network architecture can be used, where the primary task is artifact correction. Secondary tasks are designed to allow for detection of artificial structures that may occur in the artifact-corrected images. Three examples of this framework are described in detail.

[0088] Image artifacts are a common and persistent problem in clinical applications of magnetic resonance imaging (MRI). Because many artifact types (e.g., motion, Gibbs ringing, radial streaks) usually have a characteristic appearance, deep learning-based post-processing methods have been shown to enable substantial artifact reduction.

[0089] In certain cases, network-based image correction can result in artificial structures ("hallucinations") in the resulting images. In contrast to "classical" MR artifacts that can usually be easily identified by an experienced operator, these synthetic structures can have a realistic appearance. This complicates the interpretation of the images and can even lead to misdiagnosis if the hallucinations are not accurately identified.

[0090] Some examples can overcome these issues by refactoring the image correction settings into a multitasking framework, which allows for automatic detection of network hallucinations.

[0091] In this framework, a dedicated multitask network architecture can be used, where the primary task is artifact correction, and secondary tasks are designed to enable the detection of possible artificial structures in the artifact-corrected images.

[0092] Three implementations of this framework are described below as illustrative examples. 1) Template Matching A schematic diagram of this embodiment is shown in Figure 6. Figure 6 shows an example of how a template matching neural network can be trained. There is a single neural network 400, which is a Y-net. It has a U-net structure 602 and an additional decoding branch 604. In this example, the decoding branch 604 has several additional convolutional layers. In other examples, the decoding branch 604 can have a structure similar to the decoding branch of the U-net structure 602, completed with an additional skip connection. The decoding branch 604 is connected to the lowest resolution convolutional layer 606. The training data provided are the magnetic resonance data 124 paired with the ground truth data 600 and the template or atlas 532. During training, the corrupted image 124 is input to the input unit 300. In response, the corrected image 126 is output by the first output unit 302. Also, at the second output unit 304, the template matching parameters 607 are output. The corrected output 126 is compared to the ground truth data 600 using a reconstruction loss function 608. This reconstruction loss function 608 is used to provide loss data used to train the Y-Net 400 starting from the first output 302. The corrected output 126, the template matching parameters 607, and the reference magnetic resonance image 532 are input to a matching loss function 610 that provides a matching loss used to train the decoding branch 604 starting from the second output 304. For example, the problem of motion artifact correction in MR brain scans uses a tailored Y-Net architecture to simultaneously reduce artifacts and match the artifact-corrected output to a reference brain atlas, such as Talairach or MNI. To perform template matching, the second decoding branch of Y-Net is trained to provide the template matching parameters required by the specific matching algorithm. In the most common case, these parameters are 3D deformation fields that define the per-pixel shift across the volume. Assuming a certain degree of deformation field compressibility, a lower dimensional representation can also be chosen to simplify the associated learning problem. During training, the weights of the Y-Net are optimized using a standard reconstruction loss function, such as the mean squared error, in the artifact correction branch, and a matching loss function in the template matching branch. The latter is designed to achieve high sensitivity with respect to small anatomical differences, for example by using an edge-based loss function previously proposed for super-resolution tasks. During inference, potential hallucinations in the network-corrected images are detected by large deviations between the morphed artifact-corrected images and the template brain. 2) Volume In another embodiment, the secondary decoding pathway of Y-Net is trained to provide 3D segmentation masks for volumetric assessment of predefined intracranial compartments and brain structures. In this implementation, the reference masks are obtained by manual or automatic segmentation of ground truth data, and the template matching loss is replaced by, for example, a cross-entropy loss function commonly used for image segmentation tasks. During inference, detection of potential illusions in the artifact-corrected data is achieved by comparison of volumetric measurements to standard reference values. 3) Direct prediction of hallucination risk In another embodiment, a secondary decoding path of Y-Net is trained to provide hallucination risk directly, either as a global risk value or as a local risk map, the latter of which can be used to guide image interpretation highlighting "unreliable" image regions with a high risk of containing artificial anatomical structures.

[0093] Other examples may also be considered.

[0094] Patient or subject metadata is used as additional input data, thereby taking into account the diversity of brain morphology. In template matching embodiments, a dictionary of template brains with corresponding metadata is stored and appropriate dictionary entries are used during inference. In volumetric embodiments, the reference values ​​are adjusted for specific patient metadata.

[0095] Given the morphed artifact-corrected images and the template brain as input, an additional dedicated network for hallucination detection is used, which is trained to ignore normal anatomical variations but detect atypical differences (= hallucinations).

[0096] Both Y-Net branches are trained separately: we freeze the weights of one branch and update the weights of the other branch.

[0097] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered as illustrative or exemplary and not restrictive, and the invention is not limited to the disclosed embodiments.

[0098] Other variations to 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 "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" 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 measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. A computer program can be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium provided together with or as part of other hardware, but also in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be interpreted as limiting the scope. [Explanation of symbols]

[0099] 100 Medical Systems 102 Computer 104 Computing Systems 106 Hardware Interface 108 User Interface 110 Memory 120 Machine Executable Instructions 122 Image Processing Module 124 Magnetic Resonance Data 126 Corrected Magnetic Resonance Images 128 Artificial Structure Data 130 Warning Signal 300 Input section 302 First output section 304 Second output section 306 Image processing neural network part 308 Artificial structure prediction part 400 Single Neural Network 500 Medical Systems 502 Magnetic Resonance Imaging Diagnostic Equipment 504 Magnet 506 Magnet Bore 508 Imaging Zone 509 Field of view 510 Magnetic Gradient Coil 512 Gradient coil power supply 514 High Frequency Coil 516 Transceiver 518 Subject 520 Subject support part 530 Pulse sequence command (control command) 532 Reference Magnetic Resonance Imaging 534 Subject Metadata 536 Template Database 538 Image Generation Neural Network 540 Prescribed volume ratio data 542 Volume Ratio Database 600 ground truth data 602 U-net structure 604 Decode Branch 606 lowest resolution convolutional layers 607 Template Matching Parameters 608 Reconstruction Loss Function 610 Matching Loss Function

Claims

1. 1. A medical system, comprising: a memory storing machine-executable instructions; and an image processing module, the image processing module having an image processing neural network portion and an artificial structure prediction portion, the image processing module having an input configured to receive magnetic resonance data, the image processing neural network portion having a first output configured to output a corrected magnetic resonance image in response to receiving the magnetic resonance data at the input, and the artificial structure prediction portion having a second output configured to output artificial structure data describing the likelihood of artificial structures in the corrected magnetic resonance image; The image structure prediction neural network portion is trained from a ground truth image collection of magnetic resonance images. the artificial structure prediction portion is trained from a set of ground truth datasets representing global image aspects; The medical system further comprises: a computing system, wherein execution of the machine-executable instructions causes the computing system to: receiving the magnetic resonance data; receiving the corrected magnetic resonance image at the first output and the artificial structure data at the second output in response to inputting the magnetic resonance data at the input of the image processing module; providing a warning signal dependent on said artificial structure data meeting predetermined criteria; A computing system that executes A medical system having:

2. The artificial structure data includes template matching parameters for matching the corrected magnetic resonance image to a reference magnetic resonance image, and execution of the machine-executable instructions further causes the computing system to: calculating an image difference map between the reference magnetic resonance image template and the corrected magnetic resonance image using the template matching parameters; determining whether the image difference map algorithmically meets a predetermined criterion by determining whether the image difference map exceeds a predetermined statistical measure, and if the image difference map meets the predetermined criterion, the warning signal is provided; The medical system according to claim 1, wherein the medical system executes the following.

3. The medical system according to claim 2 , wherein the artificial structure prediction part is implemented as a template matching algorithm.

4. The artificial structure data includes a three-dimensional segmentation mask defining a plurality of predefined anatomical structures, and the machine-executable instructions further include: receiving predetermined volumetric ratio data describing one or more ratios between the plurality of predefined anatomical structures; calculating, from the artificial structure data, measured volume ratio data describing one or more ratios between the plurality of pre-defined anatomical structures; determining whether the predetermined criteria are met by comparing the predetermined volumetric ratio data with the measured volumetric ratio data; The medical system according to claim 1, wherein the medical system executes the following.

5. The medical system of claim 4 , wherein the artificial structure prediction portion is implemented as an image segmentation algorithm.

6. The medical system of claim 1 , wherein the artificial structure prediction portion is implemented as a neural network, and the artificial structure prediction portion is configured to receive the corrected magnetic resonance image as an input.

7. 2. The medical system of claim 1, wherein the artificial structure prediction portion is implemented as a neural network, and the output artificial structure data is a spatially dependent probability map describing the likelihood of artificial structures in the corrected magnetic resonance image.

8. 2. The medical system of claim 1, wherein the image processing module is a Y-net neural network, the Y-net neural network being formed from a U-net neural network structure configured to output the corrected magnetic resonance image at the first output in response to receiving the magnetic resonance data at the input, the Y-net further having a decoding branch configured to output artificial structure data describing artificial structures in the corrected magnetic resonance image at the second output in response to receiving the magnetic resonance data, the decoding branch being connected to the U-net neural network structure, the U-net neural network having the image processing neural network portion, and the decoding branch having the artificial structure prediction portion.

9. The medical system of claim 1 , wherein the magnetic resonance data is image data.

10. 10. The medical system of claim 9, wherein the image processing module is incorporated into a magnetic resonance imaging reconstruction algorithm configured to reconstruct a clinical magnetic resonance image in response to receiving k-space data, the magnetic resonance data being an intermediate magnetic resonance image calculated from the k-space data during the reconstruction of the clinical magnetic resonance image.

11. The medical system of claim 1 , wherein the magnetic resonance data is k-space data.

12. 12. The medical system of claim 1, further comprising a magnetic resonance imaging system, wherein the memory further comprises pulse sequence commands configured to control the magnetic resonance imaging system to acquire the magnetic resonance data from an imaging zone according to a magnetic resonance imaging protocol, and execution of the machine-executable instructions causes the computing system to control the magnetic resonance imaging system to acquire the magnetic resonance data.

13. 10. The medical system of claim 1, wherein the image processing neural network portion is configured for any one of noise removal, artifact correction, motion correction, deblurring, and combinations thereof.

14. 1. A computer program product having machine-executable instructions and an image processing module for execution by a computer system, the image processing module having an image processing neural network portion and an artificial structure prediction portion, the image processing module having an input configured to receive magnetic resonance data, the image processing neural network portion having a first output configured to output a corrected magnetic resonance image in response to receiving the magnetic resonance data at the input, and the artificial structure prediction portion having a second output configured to output artificial structure data describing the likelihood of artificial structures in the corrected magnetic resonance image, wherein execution of the machine-executable instructions causes the computer system to: receiving the magnetic resonance data; receiving the corrected magnetic resonance image at the first output and the artificial structure data at the second output in response to inputting the magnetic resonance data to the input of the image processing module; providing a warning signal dependent on said artificial structure data meeting predetermined criteria; A computer program product that causes the

15. 1. A method of medical imaging, said method comprising: receiving magnetic resonance data; receiving a corrected magnetic resonance image at a first output of the image processing module and artificial structure data at a second output of the image processing module in response to inputting the magnetic resonance data to an input of the image processing module, the artificial structure data describing the likelihood of artificial structures in the corrected magnetic resonance image, the image processing module having an image processing neural network portion and an artificial structure prediction portion, the image processing module having the input configured to receive the magnetic resonance data, the image processing neural network portion having the first output configured to output the corrected magnetic resonance image in response to receiving the magnetic resonance data at the input, and the artificial structure prediction portion having the second output configured to output the artificial structure data; providing a warning signal dependent on said artificial structure data meeting predetermined criteria; A method comprising:

Images