Out-of-distribution testing for magnetic resonance imaging.

Abstract

A medical system 100, 300 is disclosed that includes a memory 110 that stores machine executable instructions 120. The medical system further includes a computing system 104. Execution of the machine executable instructions causes the computing system to receive (202) a test magnetic resonance image reconstructed or reconstructed from undersampled k-space data, receive (204) a test signal in response to inputting the test magnetic resonance image to an out-of-distribution test neural network, and provide (206) the test signal. The test neural network outputs a test signal in response to receiving the test magnetic resonance image. The test signal indicates whether the test magnetic resonance image is within a training distribution defined by a set of training data.

Description

[Technical field]

[0001] The present invention relates to magnetic resonance imaging, and in particular to compressed sensing in magnetic resonance imaging. [Background technology]

[0002] In a magnetic resonance imaging (MRI) scanner, a large static magnetic field is used to align the nuclear spins of atoms as part of a procedure to produce images of the patient's inside the body. This large static magnetic field is called the B0 field or main magnetic field. MRI can be used to spatially measure and image various quantities or properties of a subject.

[0003] Compressed sensing (CS) is one of the means to reduce the time required to acquire k-space data for magnetic resonance images. Medical images are typically compressed or may have a sparse representation. The idea behind compressed sensing is that because medical images can have a sparse representation, images can be acquired and reconstructed by sampling less k-space data than required by the Nyquist criterion (herein referred to as undersampled k-space data). To do this, the k-space data is sampled such that artifacts from undersampling appear in image space as random noise.

[0004] An iterative process is typically used to reconstruct an image. First, an image is reconstructed from the acquired or measured k-space data. In traditional CS, a filter module converts the image into a sparse representation, such as a wavelet representation. The converted image is then thresholded to remove noise and typically converted back to image space to produce a denoised image. A data consistency module is then used to refine the image, ensuring consistency with the measured k-space data. The data consistency module takes the denoised image and adjusts the k-space transformation of the image so that it is more consistent with the measured k-space data. The effect of this is that the noise due to undersampling is reduced. The image can then be improved by iteratively processing the image through the filter and data consistency modules.

[0005] US Patent Application US20170372155A discloses image quality scoring of images from a medical scanner, where deep machine learning can be used to create a generative model of the expected high quality image. Deviations from the generative model of the input image are used as input feature vectors for a discriminative model. The discriminative model may also operate on other input feature vectors derived from the input image. Based on these input feature vectors, the discriminative model outputs an image quality score. Summary of the Invention

[0006] The invention provides a medical system, a computer program and a method as set forth in the independent claims. Embodiments are set forth in the dependent claims.

[0007] There are various ways to implement CS algorithms. Various elements, or possibly the entire numerical CS scheme above, can be replaced by a trained neural network. The problem with using neural networks is that they provide very good results when given data within the training distribution (data similar to the data used to train the neural network). If data outside the training distribution is given to the neural network, the results produced by the neural network may be inaccurate. This can, for example, lead to inaccurate magnetic resonance images being reconstructed or the reconstruction algorithm failing.

[0008] Even with modern computers, CS algorithms can be time consuming for large amounts of k-space data. Embodiments may provide a means of assessing how accurate a reconstructed image is before it is reconstructed, or providing an estimate of how well a CS reconstruction will work before the CS reconstruction is performed. This may, for example, lead to increased confidence in the resulting images, and may be used to avoid reconstructing CS images when the algorithm is likely to fail or perform poorly.

[0009] To achieve this, the undersampled k-space data is first reconstructed into a test magnetic resonance image, for example by Fourier transforming the undersampled k-space data. As a result, an image with noise and artifacts is generated. However, the insight disclosed herein is that the noise and artifacts in the test image can be used to identify whether the undersampled k-space data belongs to a training distribution. A particular training distribution provides noise and image artifacts that can be used by an out-of-distribution test neural network (image classification neural network) to provide a test signal that represents whether the test magnetic resonance image is within the training distribution defined by a set of training data.

[0010] In one embodiment of this, the out-of-distribution testing neural network is trained as a classifier within a generative adaptive neural network. An out-of-distribution testing neural network trained in this manner is particularly useful for performing out-of-distribution (OOD) testing.

[0011] In one aspect, the present invention provides a medical system including a memory storing machine executable instructions. The memory may also store an out-of-distribution test neural network, or the out-of-distribution test neural network may be remote or virtual (cloud-based computing system). The out-of-distribution test neural network is a neural network. The out-of-distribution test neural network may be a classification network configured to receive an image and provide a classification of the image. The test neural network outputs a test signal in response to receiving a test magnetic resonance image. The test signal represents whether the test magnetic resonance image is within a training distribution defined by a set of training data. The set of training data may be data used to provide training, such as deep learning, of a neural network or a collection of neural networks.

[0012] The test signal may take different forms for different examples. In one example, the test signal is a binary classification or a representation of a particular classification. For example, a 1 may indicate that the test magnetic resonance image is within the training distribution. In the same example, a 0 may indicate that the test magnetic resonance image is outside the training distribution. In another example, the test signal may be a probability indicating that the test magnetic resonance image is within the training distribution. Either example may be used, for example, as an indication or level of confidence that the test magnetic resonance image is within the training distribution.

[0013] The medical system further includes a computing system. In this specification, the term "computing system" is intended to refer to one or more computing devices located at one or more locations. For example, some parts of the computing system may be located at different locations and / or may be available as a remote computing system or a cloud-based computing system. Some parts of the computing system may be available online on demand provided by a virtual machine.

[0014] Execution of the machine-executable instructions causes the computing system to optionally receive undersampled k-space data representative of a region of interest of a subject. The label "undersampled k-space data" is a designation for particular k-space data. K-space data is data that is sampled by a magnetic resonance imaging system when imaging a region of interest of a subject. Undersampled k-space data refers to k-space data that does not meet the Nyquist criterion. However, it is possible to reconstruct an image using image reconstruction techniques such as compressed sensing.

[0015] Execution of the machine executable instructions further causes the computing system to request or cause a test magnetic resonance image to be reconstructed from the undersampled k-space data. The reconstruction may be performed by the computing system or may be a reconstruction performed on a remote or cloud-based computing system. The reconstruction may, for example, use an analysis algorithm that performs a Fourier transform on the k-space data. When the undersampled k-space data is reconstructed using a Fourier transform, it may be noisy, especially if it is undersampled. Execution of the machine executable instructions further causes the computing system to receive a test signal in response to inputting the test magnetic resonance image as an input to the out-of-distribution test neural network. Execution of the machine executable instructions further causes the computing system to provide the test signal.

[0016] This embodiment may be beneficial because the out-of-distribution test neural network tests images reconstructed from undersampled k-space data. Because the undersampled k-space data is undersampled, image artifacts are present in the test magnetic resonance image. The out-of-distribution test neural network may be configured to recognize artifacts in the test magnetic resonance image and determine from the artifacts in the image whether the undersampled k-space data is within or covered by the training distribution. The test signal may be used, for example, for other control elements in the program, or may be provided with further reconstruction of the undersampled k-space data, for example using compressed sensing reconstruction. The test signal may also be used before performing a numerically intense reconstruction of the undersampled k-space data to determine whether it is worth performing. Testing the test magnetic resonance image is very effective because it is in an image space where the out-of-distribution test neural network has not been trained to look at specific locations in k-space. For example, the neural network can be trained to directly evaluate the undersampled k-space data, but it needs to be trained for a specific k-space sampling pattern. In this example, because the undersampled k-space data is converted to an image and is already in image space, the out-of-distribution test neural network is inherently able to obtain an image of the correct size or format. This means that the out-of-distribution test neural network is highly independent of the particular sampling pattern used to obtain the undersampled k-space data. This may allow for flexible modification or adjustment of the k-space sampling pattern.

[0017] In another embodiment, execution of the machine-executable instructions further causes the computing system to receive a clinical magnetic resonance image reconstructed from the undersampled k-space data according to a compressed sensing magnetic resonance imaging reconstruction algorithm if the test signal indicates that the test magnetic resonance image is within the training distribution. This reconstruction of the clinical magnetic resonance image may be performed by the computing system, or the computing system may transmit the undersampled k-space data to a remote or cloud-based computing system to perform the reconstruction.

[0018] In this embodiment, the reconstruction of a clinical magnetic resonance image from undersampled k-space data depends on the test signal, which can be exploited to avoid lengthy reconstruction of a clinical magnetic resonance image when the data is problematic.

[0019] It may also enable operators of medical imaging systems, such as magnetic resonance imaging systems, to detect whether undersampled k-space data is likely to result in high quality clinical magnetic resonance images before lengthy image reconstructions are performed, which may allow, for example, to obtain more data or to reacquire the data while the subject is still in the clinic.

[0020] In another embodiment, the compressed sensing magnetic resonance imaging reconstruction algorithm uses an image processing neural network to generate a reconstruction of a clinical magnetic resonance image in an iterative manner. This may be performed by a computing system or may request the reconstruction from a remote or cloud-based computing system. The image processing neural network may take different forms for different examples. For example, the compressed sensing algorithms typically used for magnetic resonance image reconstruction are formulated as analysis algorithms that process data and perform data consistency iteratively. An intermediate image may be reconstructed at each iteration. Often, this image is processed using a filter, such as a denoising filter, before the data consistency step. The image processing neural network may be, for example, a denoising filter.

[0021] In another embodiment, the image processing neural network is trained using a set of training data. This embodiment is particularly beneficial because the test signal can be used to evaluate how effective the image processing neural network used in the compressed sensing magnetic resonance imaging reconstruction algorithm is. This can be used, for example, for a confidence score or to control whether the compressed sensing magnetic resonance imaging reconstruction algorithm should actually be used or whether additional data or additional algorithms should be used.

[0022] In another embodiment, the image processing neural network is configured as a denoising filter to remove noise from intermediate images between each iteration of a compressed sensing magnetic resonance imaging reconstruction algorithm.

[0023] In another embodiment, the image processing neural network serves as an image compression algorithm. As mentioned above, in compressed sensing, a sparsifying transformation may be used in combination with thresholding and desparsifying. These operations, when combined, formally serve as a compression algorithm and are the theoretical foundation of compressed sensing theory. The statement that the image processing neural network serves as an image compression algorithm means that the image processing neural network is used for one or more of the sparsifying transformation, thresholding, and desparsifying. For example, the neural network may be trained to perform all these tasks simultaneously. In another embodiment, the image processing algorithm is also trained using a set of training data. To train the image processing neural network, full sampled k-space data may be acquired. The full sampled k-space data may then be used to reconstruct training images. This provides a reference for the output of the neural network during training. The undersampled k-space data may be simulated by acquiring fully simulated k-space data and removing a portion of it so that it is undersampled. The undersampled k-space data may be combined with images reconstructed from the full sampled k-space data and used as training data in various situations for various types of neural networks.

[0024] In another embodiment, the compressed sensing magnetic resonance imaging reconstruction algorithm is a numerical image reconstruction algorithm that finds a solution to an undecidable linear system that represents the reconstruction of a clinical magnetic resonance image from an undersampled k-space image. In this embodiment, the compressed sensing magnetic resonance imaging reconstruction algorithm does not use a neural network, but is a conventional compressed sensing reconstruction algorithm. Although no training data was used to train parts of the compressed sensing reconstruction algorithm, this embodiment may be beneficial because it is still possible to use an out-of-distribution test neural network to detect insufficient undersampled k-space data. For example, it is still possible to construct a set of training data from the full sampled k-space data that is used to generate the training output image and to create the undersampled synthetic k-space data. This can then be used to train the out-of-distribution test neural network. If there is a problem with the undersampled k-space data after measurement, for example, if the subject moves or other events occur that impair the quality of the undersampled k-space data, the out-of-distribution test neural network will detect it.

[0025] In another embodiment, the compressed sensing magnetic resonance image reconstruction algorithm includes an image reconstruction neural network that reconstructs clinical magnetic resonance images from undersampled k-space data at each stage of the iterative compressed sensing algorithm. In this example, instead of using traditional reconstruction such as Fourier transform or algorithmic sparse transform, there is an image reconstruction neural network that performs this task. As above, the image reconstruction neural network can be trained by constructing training data from the fully sampled k-space data. The images used to train the image reconstruction neural network are images obtained from the reconstruction of the fully sampled k-space data using an analysis algorithm, after which the synthetic undersampled k-space data can be obtained by obtaining the fully sampled k-space data and removing some of the samples.

[0026] The image reconstruction neural network in this example may be an iterative or single pass reconstruction algorithm.

[0027] In another embodiment, the out-of-distribution test neural network is trained as a discriminative neural network in a generative adversarial network using training data. Within the generative adversarial network, there is a generative neural network and a discriminative neural network. The generator and the discriminator are trained together. The discriminative neural network is trained to generate artificial inputs to the generative adversarial network. The generative adversarial network is trained using a combination of fake data from the generator and correct or real data. The advantage of training the out-of-distribution test neural network with the generative adversarial network is that it becomes very good at detecting whether the test magnetic resonance images represent images for the training distribution.

[0028] In another embodiment, the generative adversarial network includes a generative neural network that generates simulated images in response to receiving a noise distribution. In this example, a vector or other noise is input to the generative neural network used to generate the simulated images. The simulated images and the training data are then used to train an out-of-distribution test neural network.

[0029] In another embodiment, the generative adversarial network includes a generative neural network that generates a simulated image in response to receiving a simulated test image. For example, instead of adding a noise vector to the generated simulated image, undersampled k-space data may be input to the generative neural network. This may have the advantage of making the out-of-distribution test neural network function more robust and accurate.

[0030] In another embodiment, the test magnetic resonance images are reconstructed from the undersampled k-space data using a single Fourier transform. This embodiment has the advantage that when the undersampled k-space data is reconstructed using a single Fourier transform, it contains image artifacts. These image artifacts are then detectable by the out-of-distribution test neural network and / or by one used to assess whether the undersampled k-space data is within the training distribution.

[0031] In another embodiment, the undersampled k-space data is parallel imaging k-space data collected for a set of reference coils. The computing system is configured to reconstruct coil images for each of the set of receive coil images using a single Fourier transform. The test magnetic resonance image is reconstructed by combining the coil images for each receive coil using the set of coil sensitivity maps. This embodiment is also beneficial because the individual images of each coil also contain artifacts. In this case, each out-of-distribution test neural network can simultaneously test the k-space data from all sets of receive coils by testing the test magnetic resonance image.

[0032] In another embodiment, the memory further includes pulse sequence commands configured to control the magnetic resonance imaging system to acquire undersampled k-space data from the region of interest. Execution of the machine-executable instructions further causes the computing system to acquire the undersampled k-space data by controlling the magnetic resonance imaging system with the pulse sequence commands.

[0033] In another embodiment, the medical system further includes a magnetic resonance imaging system.

[0034] In another embodiment, execution of the machine executable instructions further causes the computing system to provide a warning if the test signal indicates that the test magnetic resonance imaging results are outside of the training distribution, which may be a warning provided using the computer's user interface, or may be provided using other audio, visual, or tactile means.

[0035] In another embodiment, execution of the machine executable instructions further causes the computing system to request a reacquisition of the undersampled k-space data if the test signal indicates that the test magnetic resonance image is outside the training distribution. This portion may be implemented, for example, in the control of the magnetic resonance imaging system, which may allow an operator to immediately reacquire the undersampled k-space data if this is detected.

[0036] In another embodiment, if the test signal indicates that the test magnetic resonance image is outside the training distribution, the execution of the machine executable instructions further performs a reconstruction of the clinical magnetic resonance image using a purely numerical reconstruction algorithm. This reconstruction may be performed by the computing system, or may be performed by a remote or cloud-based computing system. For example, if a neural network is used for the compressed sensing reconstruction, the failure indicated by the test signal may prevent the use of the algorithm using the neural network. Thus, essentially, in this embodiment, the test signal is used to select an alternative reconstruction of the clinical magnetic resonance image.

[0037] In another embodiment, execution of the machine executable instructions further causes the computer system to control the magnetic resonance imaging system to continue acquiring undersampled k-space data and repeat the following steps: receiving undersampled k-space data representative of a region of interest of the subject, reconstructing a test magnetic resonance image from the undersampled k-space data, and receiving a test signal in response to inputting the test magnetic resonance image to an out-of-distribution test neural network if the test magnetic resonance image is outside the training distribution. In this example, the out-of-distribution test neural network is used to check whether enough k-space data has been acquired. For example, if the quality of the test magnetic resonance image has too many artifacts, the system may be controlled to acquire more k-space data. This may be done iteratively and the process may be stopped when the test signal indicates that the magnetic resonance image is within the training distribution.

[0038] In another embodiment, the training data includes artifact-free magnetic resonance images reconstructed from the fully sampled k-space data and simulated undersampled k-space data reconstructed from the fully sampled k-space data, which may be used, for example, in the reconstruction of medical images or utilized to train neural networks for use within compressed sensing algorithms.

[0039] In another embodiment, the training data for the out-of-distribution test neural network includes simulated undersampled k-space data constructed from the full sampled k-space data and a simulated test magnetic resonance image reconstructed from the simulated undersampled k-space data. The simulated undersampled k-space data and the simulated undersampled k-space data can be used, for example, to train the out-of-distribution test neural network directly using deep learning or as part of a generative adaptive network as described above. In this case, the image reconstructed from the full sampled k-space data is additionally used for training.

[0040] In another aspect, the present invention provides a method of operating a medical system.

[0041] The method optionally includes receiving undersampled k-space data representative of a region of interest of the subject. The method further includes receiving a test magnetic resonance image reconstructed from the undersampled k-space data. The method further includes receiving a test signal in response to inputting the test magnetic resonance image to an out-of-distribution test neural network. The test neural network outputs the test signal in response to receiving the test magnetic resonance image. The test signal indicates whether the test magnetic resonance image is within a training distribution defined by the set of training data. The method further includes providing the test signal.

[0042] In another embodiment, an out-of-distribution training neural network is trained as a discriminative neural network within a generative adversarial network using the training data.

[0043] In another aspect, the invention provides a computer program product comprising machine-executable instructions for execution by a computing system.

[0044] Execution of the machine-executable instructions causes the computing system to optionally further receive undersampled k-space data representative of a region of interest of the subject. Execution of the machine-executable instructions causes the computing system to further reconstruct or receive a test magnetic resonance image from the undersampled k-space data. Execution of the machine-executable instructions causes the computing system to further receive a test signal in response to inputting the test magnetic resonance image into the out-of-distribution test neural network. The out-of-distribution test neural network outputs a test signal in response to receiving the test magnetic resonance image. The test signal indicates whether the test magnetic resonance image is within a training distribution defined by the set of training data. Execution of the machine-executable instructions causes the computing system to further provide the test signal.

[0045] One or more of the above embodiments of the present invention may be combined, unless the embodiments to be combined are contradictory.

[0046] 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 may be referred to generally herein as "circuits," "modules," or "systems"). Additionally, aspects of the present invention may take the form of a computer program product embodied by one or more computer-readable medium(s) having computer-executable code embodied thereon.

[0047] Any combination of one or more computer readable media may be utilized. The computer readable media may be computer readable signal media or computer readable storage media. As used herein, "computer readable storage media" may encompass any tangible storage media capable of storing instructions executable by a processor or computing system of a computing device. The computer readable storage media may also be referred to as computer readable non-transitory storage media. The computer readable storage media may also be referred to as tangible computer readable media. In some embodiments, the computer readable storage media may be capable of storing data accessible by the 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 CD-ROM, CD-RW, CD-R, DVD-ROM, DVD-RW, or DVD-R disks, etc. The term computer-readable storage medium also refers to various types of recording media that a computing device can access via a network or communication link. For example, data may be retrieved via a modem, the Internet, or a local area network. Computer executable code embodied in a computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, fiber optic cable, RF, etc., or any suitable combination thereof.

[0048] A computer-readable signal medium may include a propagated data signal that includes computer-executable code (e.g., 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 capable of communicating, propagating, or transmitting a program for use by or in connection with an instruction execution system, apparatus, or device.

[0049] "Computer memory" or "memory" is one example of a computer-readable storage medium. Computer memory is any memory directly accessible by a computing system. "Computer storage" or "storage" is another example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments, computer storage may also be computer memory, and vice versa.

[0050] 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 potentially including multiple computing systems or processing cores. A computing system may be, for example, a multi-core processor. A computing system may also refer to a collection of multiple processors aggregated in a single computing system or distributed among multiple computing systems. The term computing system should also be interpreted as meaning a collection or network of multiple computing devices, each of which includes one or more processors or computing systems. Machine-executable code or instructions may be executed by multiple computing systems or processors aggregated in the same computing device or distributed across multiple computing devices.

[0051] Machine executable instructions or computer executable code may include instructions or programs that cause a processor or other computing system to perform aspects of the invention. Computer executable code for carrying out operations of aspects of the invention may be written in any combination of one or more of object-oriented programming languages ​​such as Java, Smalltalk, C++, and traditional procedural programming languages ​​such as C, 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 precompiled form, and may be used with an interpreter that generates machine executable instructions on the fly. In other examples, the machine executable instructions or computer executable code may be in the form of programming for a programmable logic gate array.

[0052] 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 and partially on a remote computer, or entirely on a remote computer or server. In the latter case, 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 a connection may be established to an external computer (e.g., via the Internet using an Internet Service Provider).

[0053] 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 will be understood that each block or group of blocks of the flowcharts, illustrations, and / or block diagrams may be implemented by computer program instructions in the form of computer executable code, where applicable. It will also be understood that blocks in different flowcharts, illustrations, and / or block diagrams may be combined, where not mutually inconsistent. These computer program instructions may be provided to a computing system of a general purpose computer, special purpose computer, or other programmable data processing device to form a machine. The instructions executed via the computing system of the computer or other programmable data processing device create means for implementing the functions / acts identified in one or more blocks of the flowcharts and / or block diagrams.

[0054] These machine-executable instructions or computer program instructions may be stored on a computer-readable medium that can cause a computer, other programmable data processing apparatus, or other device to function in a particular manner. The instructions stored on the computer-readable medium create an article of manufacture that includes instructions that implement the functions / acts identified in one or more blocks of the flowcharts and / or block diagrams.

[0055] The machine-executable instructions or computer program instructions may be loaded onto a computer, other programmable data processing apparatus, or other device and may cause the computer, other programmable apparatus, or other device to execute a series of operational steps to produce a computer-implemented process. The instructions executing on the computer or other programmable apparatus provide a process for implementing the functions / operations identified in one or more blocks of the flowcharts and / or block diagrams.

[0056] 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" may also be referred to as a "human interface device." A user interface may provide information or data to an operator and / or receive information or data from an operator. A user interface may allow a computer to receive input from an operator and may provide output from the computer to a user. In other words, a user interface may allow an operator to control or manipulate a computer and an interface may allow a computer to display the effects of the operator's control or manipulation. 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, pedals, wired gloves, remote control, and accelerometer are all examples of user interface elements that allow receiving information or data from an operator.

[0057] 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 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 MIDI interface, an analog input interface, and a digital input interface.

[0058] As used herein, "display" or "display device" encompasses an output device or user interface adapted to display images or data. A display may output visual, auditory, or tactile data. Examples of displays include, but are not limited to, computer monitors, television screens, touch screens, tactile electronic displays, Braille screens,

[0059] These include cathode ray tubes (CRT), storage tubes, bi-stable displays, electronic paper, vector displays, flat panel displays, 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.

[0060] Medical imaging data is defined herein as a record of measurements made by a tomographic medical imaging system representing a subject. The medical imaging data may be reconstructed into a medical image. A medical image is defined herein as a reconstructed two- or three-dimensional visualization of the anatomical data contained within the medical imaging data. This visualization may be performed using a computer.

[0061] As used herein, k-space data is defined as the measurement recording of radio frequency signals emitted by atomic spins using the antenna of a magnetic resonance apparatus during a magnetic resonance imaging scan. Magnetic resonance data is an example of medical tomography image data.

[0062] Undersampled k-space data is defined as k-space data that contains less k-space data than is necessary to meet the Nyquist criterion.

[0063] As used herein, a magnetic resonance imaging (MRI) image or MR image is defined as a reconstructed two- or three-dimensional visualization of anatomical data contained within magnetic resonance imaging data, which visualization may be performed using a computer. [Brief description of the drawings]

[0064] Preferred embodiments of the present invention will now be described, by way of example only, with reference to the following drawings, in which:

[0065] [Figure 1] FIG. 1 shows an example of a medical system. [Diagram 2] FIG. 2 shows a flow chart illustrating a method of using the medical system of FIG. [Diagram 3] FIG. 3 shows a further example of a medical system. [Figure 4] FIG. 4 shows a flow chart illustrating a method of using the medical system of FIG. [Diagram 5]FIG. 5 shows an example of a generative adversarial network that can be used to train an out-of-distribution test neural network. [Figure 6] FIG. 6 illustrates the use of an out-of-distribution testing neural network. [Figure 7] FIG. 7 shows a further example of a generative adversarial network that can be used to train an out-of-distribution test neural network. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0066] Elements with like numbers in the figures are equivalent elements or perform the same function. An element described previously is not necessarily described in a later figure if the function is equivalent.

[0067] FIG. 1 illustrates an example of a medical system 100. The medical system is shown as including a computer 102. The computer 102 is intended to represent one or more computing devices. The computer 102 may be integrated into a magnetic resonance imaging system, for example, as part of a control system for the magnetic resonance imaging system. In other examples, the computer 102 may be a remote computer system used to reconstruct images remotely. For example, the computer 102 may be a server in a radiology department or a virtual computer system in a cloud computing system.

[0068] Computer 102 is further shown as including a computer system 104. Computing system 104 is intended to represent one or more processors or processing cores, or other computing systems, located in one or more locations. Computing system 104 is shown as being connected to an optional hardware interface 106. The optional hardware interface 106 may, for example, enable computing system 104 to control other components, such as a magnetic resonance imaging system.

[0069] The computing system 104 is further shown as being connected to an optional user interface 108, which may, for example, enable an operator to control and operate the medical system 100. The computing system 104 is further shown as being connected to a memory 110. The memory 110 is intended to represent various types of memory that may be connected to the computing system 104.

[0070] The memory is shown as including machine-executable instructions 120. The machine-executable instructions 120 enable the computing system 104 to perform tasks such as controlling other components and performing various data and image processing tasks. The memory 110 is further shown as including an out-of-distribution test neural network 122. The out-of-distribution test neural network 122 is configured to receive test magnetic resonance images 126 and provide test signals in response thereto. Alternatively, the out-of-distribution test neural network may be located on a remote or cloud-based computing system.

[0071] The memory 110 is further shown as including k-space data 124, which is k-space data 124 acquired using a magnetic resonance imaging system. The memory 110 is further shown as including a test magnetic resonance image 126 reconstructed from the k-space data 124, which can be done, for example, using a Fourier transform according to a magnetic resonance imaging protocol. The k-space data 124 is undersampled k-space data 124. Undersampled means that the k-space data does not meet the Nyquist criterion. When the undersampled k-space data 124 is reconstructed into a test magnetic resonance image 126, the test magnetic resonance image 126 includes image artifacts, such as distortion and noise, due to undersampling. The out-of-distribution test neural network 122 may be, for example, an image classification neural network. The out-of-distribution test neural network 122 may be trained to recognize whether the undersampled k-space data 124 is within a training distribution defined by a set of training data. As used herein, training data includes data that may be used to train a neural network. The training data thus provides input to the neural network and a training output or training test signal that can be compared to the output of the out-of-distribution test neural network 122 .

[0072] Memory 128 is further shown as including a test signal 128 received in response to inputting test magnetic resonance image 126 as an input to out-of-distribution test neural network 122. Test signal 128 may be used to indicate or provide a probability indicating whether or not test magnetic resonance image 126 is included in the training distribution.

[0073] The memory 110 is shown as including an optional compressed sensing magnetic resonance imaging reconstruction algorithm 130, which is an algorithm used to reconstruct a clinical magnetic resonance image 132 from the undersampled k-space data 124 according to a compressed sensing algorithm. The compressed sensing magnetic resonance imaging reconstruction algorithm 130 may be conventional without using a neural network, or may be various types of neural networks incorporated into the compressed sensing magnetic resonance imaging reconstruction algorithm. In some examples, the data used to train the neural network component of the compressed sensing magnetic resonance imaging reconstruction algorithm 130 is trained using a training distribution defined by the same set of training data used to train the out-of-distribution test neural network 122.

[0074] In some examples, the out-of-distribution test neural network 122 was trained with a GAN network. The out-of-distribution test neural network 122 was a discriminator in the GAN network. Using the discriminator of the GAN network as the out-of-distribution test neural network 122 may have the technical advantage of being much more robust and effective in detecting whether the undersampled k-space data 124 is included in the training distribution.

[0075] FIG. 2 shows a flow chart illustrating a method of operation of the medical system 100 of FIG. 1. The method and system show steps performed by the computing system 104. Alternatively, the image reconstruction and / or the use of the out-of-distribution test neural network may be performed on a remote or cloud-based computing system. First, in step 200, optionally undersampled k-space data 124 is received. The undersampled k-space data 124 represents a region of interest of a subject during a magnetic resonance imaging examination. Next, in step 202, a test magnetic resonance image 126 is reconstructed from the undersampled k-space data 124 or a pre-reconstructed test magnetic resonance image 126 is received. The reconstruction may be performed using a Fourier transform. Next, in step 204, a test signal 128 is received in response to inputting the test magnetic resonance image 126 as an input to the out-of-distribution test neural network 122. First, in step 206, a test signal 128 is provided. This may be used for various purposes. For example, in some examples, the test signal 128 may be a probability. In this case, the test signal can be used as an analogy to a confidence measure. Thus, the test signal may be added to data or metadata representing a subsequent clinical magnetic resonance image 132. In other examples, the test signal 128 can be used to control the behavior of the reconstruction and / or magnetic resonance imaging. For example, if the test signal 128 indicates that the undersampled k-space data 124 is not within the training distribution, it may be beneficial to use a different reconstruction algorithm or to reacquire or acquire more undersampled k-space data 124.

[0076] Figure 3 illustrates a further example of a medical system 300. The medical system 300 illustrated in Figure 3 is similar to the medical system 100 of Figure 1, except that it additionally includes a magnetic resonance imaging system 302 controlled by the computing system 104.

[0077] The magnetic resonance imaging system 302 has a magnet 304. The magnet 304 is a superconducting cylindrical magnet with a bore 306 passing through it. Different types of magnets can be used. For example, it is possible to use both split cylindrical magnets and so-called open magnets. Split cylindrical magnets are similar to standard cylindrical magnets, but differ in that the cryostat is split into two parts to allow access to the iso-plane of the magnet; such magnets can be used, for example, with charged particle beam therapy. Open magnets have two magnet parts, one located above the other to allow enough space to accommodate a subject between them, the arrangement of the two parts being similar to that of a Helmholtz coil. Open magnets are popular because the subject is less occluded. Inside the cryostat of the cylindrical magnet is a collection of superconducting coils.

[0078] Within the bore 306 of the cylindrical magnet 304 is an imaging zone 308 in which there is a magnetic field strong enough and uniform to perform magnetic resonance imaging. A region of interest 309 is shown within the imaging zone 308. Magnetic resonance data is typically acquired about the region of interest. A subject 318 is shown supported by a subject support 320 such that at least a portion of the subject 318 is within the imaging zone 308 and the region of interest 309.

[0079] Also within the magnet bore 306 are a set of magnetic field gradient coils 310 used for preliminary magnetic resonance data acquisition to spatially encode magnetic spins within an imaging zone 308 of the magnet 304. The magnetic field gradient coils 310 are connected to a magnetic field gradient coil power supply 312. It should be understood that the magnetic field gradient coils 310 are representative. Typically, the magnetic field gradient coils 310 include three separate coil sets for spatial encoding in three orthogonal spatial directions. The magnetic field gradient power supply supplies electrical current to the magnetic field gradient coils 310. The electrical current supplied to the magnetic field gradient coils 310 may be controlled, ramped or pulsed as a function of time.

[0080] Next to the imaging zone 308 is a radio frequency coil 314 for manipulating the orientation of the magnetic spins in the imaging zone 308 and for receiving radio signals from the spins in the imaging zone 308. The radio frequency antenna may include multiple coil elements. The radio frequency antenna may also be referred to as a channel or antenna. The radio frequency coil 314 is connected to a radio frequency transceiver 316. The radio frequency coil 314 and the radio frequency transceiver 316 may be replaced by separate transmit and receive coils and separate transmitters and receivers. It should be understood that the radio frequency coil 314 and the radio frequency transceiver 316 are representative. The radio frequency coil 314 is also intended to represent a dedicated transmit antenna and a dedicated receive antenna. Similarly, the transceiver 316 may also represent a separate transmitter and receiver. The radio frequency coil 314 may also have multiple receive / transmit elements, and the radio frequency transceiver 316 may have multiple receive / transmit channels. For example, if a parallel imaging technique such as SENSE is implemented, the radio frequency coil 314 has multiple coil elements.

[0081] The transceiver 316 and the gradient controller 312 are shown as being connected to the hardware interface 106 of the computer system 102 .

[0082] The memory 110 is further shown as including a plurality of pulse sequence commands 330. The pulse sequence commands 330 may be commands configured to control the magnetic resonance imaging system 302 to acquire undersampled k-space data 124 from the region of interest 309, or data convertible into such commands.

[0083] Figure 4 shows a flow chart illustrating a method of operation of the medical system 300 of Figure 3. Beginning at step 400, the magnetic resonance imaging system 302 is controlled using pulse sequence commands 330 to acquire undersampled k-space data 124. The method then proceeds to steps 200, 202, 204, 206, and 208 as shown in Figure 2.

[0084] In compressed sensing magnetic resonance imaging (CS-MRI), undersampling of k-space is performed to acquire images faster, which improves all KPIs of the quadruple aim. However, if the acceleration rate is too high, reconstruction of such undersampled data can result in poor quality and a large number of artifacts. Deep learning-based algorithms have proven effective in reducing reconstruction errors, allowing for higher undersampling rates.

[0085] However, the lack of control and the almost unpredictable behavior on never-before-seen data is one of the main drawbacks of AI compared to traditional algorithms. It has been frequently observed that deep learning solutions for medical imaging applications can produce realistic artifacts if the processed data deviates too far from the data used for training (out-of-distribution OOD). Users of AI-powered systems may therefore be unable to spot and correct mistakes or avoid using the system for difficult or ambiguous symptoms. This problem creates a large gap between experiments with neural networks and their practical application in real-world tasks.

[0086] The embodiment can provide a means to estimate whether a new input to a deep learning system is sufficiently similar to the training data using a possibly adversarially trained discriminative model (out-of-distribution test neural network 122). The well-known discriminator properties of generative adversarial networks (GANs) are exploited to learn the distribution of the training data, which is then applied to the OOD estimation. In use, the model provides a test signal 128, which may provide the probability that an object belongs to the training distribution, or directly provide a binary decision. Furthermore, knowing this makes it possible to determine the feasibility of reconstructing undersampled data without spending extra time and computational power on the reconstruction process. Thus, the present invention enables and facilitates the application of deep learning solutions in real-world environments.

[0087] In an example, a discriminative model may be provided to distinguish whether a zero filled image is in-distribution or out-of-distribution. To that end, we propose to utilize the well-known discriminator properties of generative adversarial networks (GANs) to learn the distribution of training data and apply it to OOD estimation.

[0088] Such an approach solves several problems. First, it is assumed that the reconstruction, or the features of the reconstruction network, have some relationship to the distribution of the undersampled objects, which is rarely the case in practice. Second, both methods require one or more forward passes for a reconstruction network such as Adaptive CS-Net, which can be computationally expensive. Working directly with zero-filled images frees these requirements, making OOD estimation both computationally feasible and more interpretable.

[0089] In the simplest terms, a GAN is used to generate realistic zero-filled images (fake test magnetic resonance images 508). As the OOD scoring mechanism, we employ the GAN's discriminator (out-of-distribution test neural network 122). In the next paragraphs, a simplified approach is presented to convey the general concept. In later paragraphs, we present a more advanced and powerful generation mechanism.

[0090] By default, GAN samples a noise vector z using a normal or uniform distribution and creates an image Gout using a deep neural network generator G. A discriminator D is added to distinguish whether the discriminator input is real or generated. Note that in this case, Gout is not a reconstructed image, but a zero-padded image. Hence, the output of the discriminator, the value p D estimates the probability that the input is a realistic representation of a zero-padded image present in the training distribution. During training time, D only recognizes the training data as real, and therefore implicitly learns to distinguish between objects from the training distribution, i.e., the zero-padded images used to train the reconstruction model, and objects from other distributions. Therefore, we propose a discriminator D and its output p as a method for computing the OOD score of an input. D I suggest using:

[0091] Thus, a basic system may include one or more of the following elements: Noise vector sampled from a z-normal or uniform distribution A generator model for creating realistic zero-filled images from Gz G out -The output of the reconstruction model G, i.e. a realistic zero-filled image p D - the probability that the input of the classifier network D is a real (not generated) object L d - Loss of the discriminative model D. It is also used as the loss term for the generative model G. Undersampled Image (UI) - Inverse Fast Fourier Transform (iFFT) of the result of element-wise multiplication of the fully sampled k-space with an initial binary undersampling mask

[0092] FIG. 5 illustrates one way to train the out-of-distribution test neural network 122. FIG. 5 illustrates a generative adversarial neural network 500 with a generative neural network 502 and a discriminative neural network 504. In this case, the discriminative neural network 504 is the out-of-distribution test neural network 122. The generative neural network 502 is configured to receive a noise vector 506 and output a fake test magnetic resonance image 508. The fake test magnetic resonance image 508 is input to the discriminator 504. During training, it is known whether the image 508 is real or not. If the discriminator 504 guesses wrong, the discriminator is trained. If the generative neural network 502 fails to fool the discriminator 122, the generative neural network is trained. During training, both the generator 502 and the discriminator 504 are improved together. Real data is also used to train the discriminator 504. In this case, the training test magnetic resonance images 510 are constructed from the full sampled k-space data 512. A mask 513 is used to construct simulated undersampled k-space data 514 by removing a portion of the full sampled k-space data 512. This is then Fourier transformed to generate the training test magnetic resonance images 510. During training, the classifier 504 may be tested using both the training test magnetic resonance images 510 and the fake test magnetic resonance images 508. The full sampled k-space data 512 is used to construct the simulated undersampled k-space data 514. The space represented by the resulting image 510 and the simulated undersampled k-space data 514 represent the training distribution defined by the set of training data. The training data in this case are pairs of the simulated undersampled k-space data 514 and the training test magnetic resonance images 510 developed therefrom. The fully sampled k-space data 512 may be used to construct training data for other neural networks, such as a neural network that may be incorporated into a compressed sensing reconstruction algorithm to generate clinical magnetic resonance images.

[0093] First, a generative convolutional neural network G receives a noise vector z from a normal or uniform distribution. The goal of G is to generate realistic zero-filled images to fool a discriminative convolutional neural network D122. G generates Gout 508, which is sent to D122 along with the UI. The goal of D122 is to accurately distinguish the real UI 510 from the fake 508 from G. D is a p D This outputs G out can be interpreted as the probability that p is genuine. D may be converted to a binary output. D122 learns to predict the correct output by minimizing LD, which can be any classification loss, such as binary cross-entropy. G also gathers information from LD, but in contrast to D122, its goal is to maximize LD. Thus, G502 and D122 are alternately optimized to solve the adversarial minimax problem.

number

[0094] FIG. 6 illustrates the use of an out-of-distribution test neural network 122. In this figure, the partial k-space is undersampled k-space data 124. This is then Fourier transformed to generate a test magnetic resonance image 126. This is input to the out-of-distribution test neural network 122, which is trained as a discriminator in the GAN 500 of FIG. 5. This outputs a test signal 128. In this example, the test signal 128 is a discrete 0 or 1. In other cases, it may be a generated probability. FIG. 6 illustrates how a discriminative model can be used for OOD estimation of previously unseen data.

[0095] The partially sampled k-space is converted to an image using iFFT and sent to D. In one embodiment, pD can be used as an estimate of the probability that the input data is in the distribution. In another embodiment, pD can be converted to a binary output using a threshold function to directly answer whether the data belongs to the training distribution or not. Furthermore, these values ​​can be used to reject the AI-based solution and revert to a traditional solution, or to inform the user that the reconstruction may be inaccurate and suggest reacquiring the data.

[0096] Although the proposed system solves the problem, in practice it may suffer from problems that characterize the training of GANs: mode collapse and non-convergence. In the first case, this system, without any supervision at all, may easily reach a local minimum where G produces a limited variety of samples. This prevents the discriminator from generalizing well. In the second case, the model parameters oscillate, become unstable and never converge. To eliminate these potential problems, one form of self-supervision during training is described below. The following components are added to the proposed basic approach: Missing Information Image (MII) - iFFT of the element-wise multiplication of the fully sampled k-space with the inverse initial binary undersampling mask Lg - The loss function of the reconstructed model G. It is calculated as the distance between Gout and UI.

[0097] FIG. 7 shows an alternative generative adversarial neural network 700 that can be used to train the out-of-distribution test neural network 122. This example differs in that no random or noise vectors are input to the generator 502, but instead a missing information image 702 is used. This is an image constructed from the full sampled k-space data 512 using an alternative mask 703 that can change the sampling to a different undersampling pattern or miss some data so that the data is incomplete. This missing information image 702 is then input to the generative neural network and used to generate a fake test magnetic resonance image 508. A loss function 704 is then constructed from the fake test magnetic resonance image 508 and the undersampled image 706. The undersampled image 706 is created from the undersampled k-space data 512 and the same mask 513 used to construct the test magnetic resonance image 510. The full sampled k-space data 512 and the mask 513 are used to provide simulated undersampled k-space data 514. This is then Fourier transformed to produce the undersampled image 706 .

[0098] This example can be beneficial because it generates realistic fake test magnetic resonance images 508 very quickly during training, allowing the out-of-distribution test neural network 122 to be trained very effectively.

[0099] Figure 7 outlines the modified training procedure. First, a generative convolutional neural network G502 receives MII702 as input. The goal is still to generate realistic zero-padded images 508 to fool D122. But now, L d By maximizing the L between the UI and the output of the model g It learns to do so by also minimizing L. In fact, Lg is 1It can be any loss for image-to-image tasks, such as , SSIM, MS-SSIM, or a combination of these. Gout is then sent to D122 along with the UI, whose goal is to accurately distinguish real UI from fake ones from G502. Note that this is very similar to the basic approach above, where the model samples from a noise vector z. In this case, the generator is replaced by a stronger preconditioner.

[0100] Note that although this is a powerful and viable solution, other generators may be employed.

[0101] Networks G502 and D122 are trained using (2) as an adversarial system, but only D is used during evaluation time to estimate the OOD score. This system may appear to solve the same problem as Adaptive CS-Net, with only the discriminative network D122 added. In fact it does the opposite. Standard image reconstruction models attempt to recover the undersampled image UI. Our network G recovers the missing information image, which is essential to train a sufficient discriminative model.

number

[0102] The examples may provide one or more of the following advantages.

[0103] The method may be used directly on the input data before it passes through the reconstruction chain, potentially improving response time and reducing computational load.

[0104] The quality of the proposed system increases with increasing acceleration rate. By increasing the degree of undersampling, the amount of information contained in the MII increases, which makes the adversarial system easier to train and more stable during evaluation.

[0105] It can be shown formally that a GAN discriminative model learns the distribution of the training data, and the result of that work is therefore a direct estimation of the test samples belonging to the training distribution.

[0106] An example is where an adversarially trained discriminative model may be used to estimate whether a new input to a deep learning system is sufficiently similar to the training data. The well-known discriminator properties of generative adversarial networks (GANs) are exploited to learn the distribution of the training data, which is then applied to the OOD estimation. This property can be used to determine the feasibility of reconstructing undersampled data without spending extra time and computational power on the reconstruction process.

[0107] The architectures of G502 and D122 are not limited, except that they are convolutional neural networks capable of solving image-to-image and binary classification tasks, respectively. However, the use of state-of-the-art models for reconstruction of undersampled MR data, such as Adaptive-CS-Net, for G502 could significantly improve the overall quality and reliability of the system.

[0108] In the example, the out-of-distribution test neural network (G502) may be implemented using various types of neural networks, for example ResNet, DenseNet, or a basic CNN configured for image classification may be effectively used.

[0109] The generative neural network (D122) may be implemented using various types of neural networks. In general, any neural network configured for image processing may be effectively used. For example, an image processing neural network such as a U-net may be used.

[0110] The output from D122 can be used in a variety of ways, depending on the business needs and the overall design of the reconstruction system. In one embodiment, p is used as a direct estimate of the probability that the input data is in the distribution. D ∈[0,1] can be used. This may be provided to the user "as is" to decide whether to reject the AI-based solution and revert to a traditional solution or to continue with the default deep learning model. In another embodiment, the value of pd may be used by another smart system that performs the decision. In another embodiment, the binary output of D can be used immediately. In this case, the threshold that the model learned during training is applied.

[0111] Some examples may be applied to validate AI / deep learning based models for accelerated MR scan reconstruction, in which case the approach is as follows:

[0112] Training time: Collect training data (fully sampled raw MR data set (fully sampled k-space data)). The inverse undersampling mask along with the iFFT is applied to the initial data to generate input data (MII). Apply an undersampling mask to the initial data (UI) along with the iFFT Train networks G and D to solve (1) using MII and UI. Deploy D to production. System Usage Time: The patient is scanned with a desired level of undersampling to obtain input data. Calculate pd and the binary output of D. Using pd and the binary output of D, the output of the AI ​​(OOD or non-OOD) is verified. If the input data (test magnetic resonance image 126) is classified as not OOD (by test signal 128), do nothing. Otherwise, choose one or more of the following steps: We return to other algorithms that do not suffer from the OOD problem (Compressed SENSE). Notify the user or manufacturer of the problem. Corrective measures such as rescanning to obtain MR will be taken.

[0113] 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.

[0114] From the drawings, the disclosure, and the appended claims, other variations of the disclosed embodiments can be understood and realized by those skilled in the art practicing the claimed invention. In the claims, the terms "comprise" and "include" do not exclude other elements or steps, and singular elements do 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 several means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used advantageously. The computer program may be stored and / or distributed on a suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, or distributed 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 thereof. [Explanation of symbols]

[0115] 100 Medical Systems 102 Computer 104 Computing Systems 106 Optional Hardware Interface 108 Optional User Interface 110 Memory 120 Machine Executable Instructions 122 Out-of-Distribution Testing Neural Networks 124 Undersampled k-space data 126 Test Magnetic Resonance Imaging 128 test signals 130 Compressed Sensing Magnetic Resonance Imaging Reconstruction Algorithm 132 Clinical Magnetic Resonance Imaging 200 Receive undersampled k-space data representing a region of interest of a subject. 202 Reconstructing Test Magnetic Resonance Images from Undersampled k-Space Data 204 receiving a test signal in response to inputting a test magnetic resonance image to the out-of-distribution test neural network 206 Provide a test signal 208 Reconstructing clinical magnetic resonance images from undersampled k-space data using a compressed sensing magnetic resonance imaging reconstruction algorithm if the test signal indicates that the test magnetic resonance image is within the training distribution. 300 Medical Systems 302 Magnetic Resonance Imaging System 304 Magnet 306 Magnet Bore 308 Imaging Zone 309 Areas of Interest 310 Magnetic field gradient coil 312 Magnetic field gradient coil power supply 314 Radio Frequency Coil Transmitter / Receiver 318 Subjects 320 Subject support platform 330 Pulse Sequence Commands 400 A magnetic resonance imaging system is controlled with pulse sequence commands to acquire undersampled k-space data. 500 Generative Adversarial Neural Networks 502 Generative Neural Networks 504 Classification Neural Network 506 Noise Vector 508 Fake Test Magnetic Resonance Imaging 510 Training Test Magnetic Resonance Imaging 512 fully sampled k-space data 513 Mask 514 Simulated undersampled k-space data 700 Generative Adversarial Neural Networks 702 Missing Information Images 703 Alternative Mask 704 Loss Function 706 Undersampled Image

Claims

1. A medical system comprising a memory storing machine-executable instructions and a computing system, wherein upon execution of the machine-executable instructions, the computing system receives a test magnetic resonance image reconstructed from undersampled k-space data, and receives a test signal in response to causing an input of the test magnetic resonance image to an out-of-distribution test neural network, wherein the out-of-distribution test neural network outputs the test signal in response to receiving the test magnetic resonance image, and the test signal represents whether the test magnetic resonance image is within a training distribution defined by a set of training data, and provides the test signal.

2. Upon execution of the machine-executable instructions, the computing system further reconstructs a clinical magnetic resonance image from the undersampled k-space data using a compressed sensing magnetic resonance imaging reconstruction algorithm when the test signal indicates that the test magnetic resonance image is within the training distribution, according to the medical system of claim 1.

3. The compressed sensing magnetic resonance imaging reconstruction algorithm uses an image processing neural network to iteratively reconstruct the clinical magnetic resonance image, according to the medical system of claim 2.

4. The image processing neural network comprises a noise removal filter for removing noise from intermediate images between iterations, and is configured as any one of image compression algorithms, according to the medical system of claim 3.

5. The compressed sensing magnetic resonance imaging reconstruction algorithm is a numerical image reconstruction algorithm that finds a solution to an ill-conditioned linear system representing the reconstruction of the clinical magnetic resonance image from the undersampled k-space image, according to any one of claims 2 to 4 of the medical system.

6. The compressed sensing magnetic resonance imaging reconstruction algorithm includes an image reconstruction neural network that reconstructs the clinical magnetic resonance image from the undersampled k-space data at each stage of an iterative compressed sensing algorithm, according to any one of claims 2 to 4 of the medical system.

7. The out-of-distribution test neural network is trained as an identification neural network in the adversarial generation network using the training data, for the medical system according to any one of claims 1 to 6.

8. The adversarial generation network includes a generative neural network that generates a simulated image in response to receiving a noise distribution, for the medical system according to claim 7.

9. The adversarial generation network includes a generative neural network that generates a simulated image in response to receiving a simulated test image, for the medical system according to claim 7.

10. The test magnetic resonance image is reconstructed from the undersampled k-space data using a Fourier transform, for the medical system according to any one of claims 1 to 9.

11. The memory further includes a pulse sequence command for controlling a magnetic resonance imaging system to acquire the undersampled k-space data from a region of interest, and by executing the machine-executable instructions, the computing system further acquires the undersampled k-space data by controlling the magnetic resonance imaging system using the pulse sequence command, for the medical system according to any one of claims 1 to 10.

12. When the test signal indicates that the test magnetic resonance image is outside the training distribution, by executing the machine-executable instructions, the computing system further provides a warning signal; requests reacquisition of the undersampled k-space data; requests reconstruction of a clinical magnetic resonance image using a purely numerical reconstruction algorithm; controls the magnetic resonance imaging system to continue acquiring the undersampled k-space data; performs any one of the following steps: reconstructing the test magnetic resonance image from the undersampled k-space data representing the region of interest of the subject, and repeating receiving the test signal in response to input of the test magnetic resonance image to the out-of-distribution test neural network, for the medical system according to any one of claims 1 to 11.

13. The training data for the out-of-distribution test neural network includes simulated undersampled k-space data constructed from fully sampled k-space data and simulated test magnetic resonance images reconstructed from the simulated undersampled k-space data, the medical system according to any one of claims 1 to 12.

14. A method of operating a medical system, the method comprising: receiving a test magnetic resonance image reconstructed from undersampled k-space data; receiving a test signal in response to inputting the test magnetic resonance image into an out-of-distribution test neural network, the out-of-distribution test neural network outputting the test signal in response to receiving the test magnetic resonance image, the test signal indicating whether the test magnetic resonance image is within a training distribution defined by a set of training data; providing the test signal.

15. A computer program comprising machine-executable instructions executable by a computing system, the execution of the machine-executable instructions causing the computing system to: receive a test magnetic resonance image reconstructed from the undersampled k-space data; receive a test signal in response to inputting the test magnetic resonance image into an out-of-distribution test neural network, the out-of-distribution test neural network outputting the test signal in response to receiving the test magnetic resonance image, the test signal indicating whether the test magnetic resonance image is within a training distribution defined by a set of training data; provide the test signal.

Images