Anomaly Detection Using Magnetic Resonance Fingerprinting

Abstract

The present invention provides a medical imaging system (100, 300, 500, 600), comprising: a memory (110) for storing machine executable instructions (120); and a processor (106) for controlling the medical instrument. Execution of the machine executable instructions causes the processor to: receive (200) MRF magnetic resonance data (122) of a region of interest (310) acquired according to an MRF magnetic resonance imaging protocol; reconstruct (202) an MRF vector (124) for each voxel in a set of voxels describing the region of interest using the MRF magnetic resonance data acquired according to the MRF magnetic resonance imaging protocol; calculate (204, 714) a pre-processed MRF vector (126) for each voxel in the set of voxels by applying a predetermined pre-processing routine to the MRF vector for each voxel, wherein the predetermined pre-processing routine includes normalizing the pre-processed MRF vector for each voxel; calculate (206, 716) an outlier map (130) for the set of voxels by assigning an outlier score (800) to the pre-processed MRF vector using a machine learning algorithm.

Description

Technical Field

[0001] The present invention relates to magnetic resonance imaging, and in particular to magnetic resonance fingerprinting. Background Art

[0002] Magnetic resonance fingerprinting (MRF) is a technique in which multiple RF pulses distributed in time are applied so that signals from different materials or tissues have unique contributions to the measured magnetic resonance (MR) signal. A limited dictionary of pre-calculated signal contributions from a set or fixed number of substances is compared with the measured MR signal. This can be used, for example, to determine intrinsic properties, such as T1, T2, and B1+ values. In other examples, the comparison between the dictionary of pre-calculated signals and the measured signal can be used to determine the material composition within a single voxel. For example, if a voxel is known to contain only water, fat, and muscle tissue, the contributions from these three materials only need to be considered and intra-voxel component matching can be used to accurately determine the composition of the voxel. Magnetic resonance fingerprinting technology is introduced in the journal article "Magnetic Resonance Fingerprinting" by Ma et al. (Nature, Vol. 495, pp. 187-193, doi:10.1038 / nature11971).

[0003] International patent application WO 2017 / 007696 A1 discloses a system during operation that iteratively captures MR signals of one or more types of nuclei in one or more parts of a biological life form based on scan instructions corresponding to a dynamic scan plan. The MR signals in a given iteration can be associated with voxels of an associated size at a three-dimensional (3D) position in at least the corresponding part of the biological life form. If the system detects a potential abnormality when analyzing the MR signals from a given iteration, the system dynamically modifies the scan plan based on the detected potential abnormality, medical history, and / or MR scan history. Subsequent measurements of the MR signals can be associated with the same or different items: type of nucleus, part of the biological life form, voxel size, and / or 3D position. Summary of the invention

[0004] The invention provides a medical imaging system, a computer program product and a method in the independent claims. Embodiments are given in the dependent claims.

[0005] As mentioned above, a dictionary of signals for a fixed number of substances is used in magnetic resonance fingerprinting. If a particular substance is not in the dictionary, but is present in the object as an image, the match to the dictionary can fail. This may lead to errors in the calculation of the content within the voxel. In this case, it is difficult to know whether there is an abnormality (such as tumor tissue) within the object or whether the preparation of the dictionary is insufficient.

[0006] Embodiments of the present invention can provide a module for identifying abnormal voxels without prior preparation of an MRF dictionary. To do this, a machine learning algorithm is prepared. The machine learning algorithm is trained or prepared using MRF magnetic resonance data collected from an unknown object including tissue containing abnormal or abnormal anatomical structures. The machine learning algorithm is then used to calculate an outlier map of the MRF signal (also referred to as an MRF vector) for each voxel. The machine learning algorithm assigns an outlier score to each voxel. The outlier score is a numerical value that is a measure of how much the tested MRF vector changes from the MRF vector used to train or configure the machine learning algorithm.

[0007] The outlier map can then be displayed or plotted (possibly superimposed on other medical images) to help healthcare professionals identify voxels that produce abnormal or anomalous MRF vectors. This can be used to check the effectiveness of conventional MRF images for monitoring cancer growth and / or for identifying areas of the subject for further imaging or study.

[0008] In one aspect, the present invention provides a medical imaging system comprising: a memory for storing machine executable instructions; and a processor configured to control the medical instrument. In different examples, the medical imaging system can take different forms. In some examples, the medical imaging system is a system for processing or modifying images or data related to images. In other examples, the medical imaging system may also include components for acquiring medical imaging data that is then processed or can be processed into data suitable for rendering.

[0009] In the case where the medical imaging system comprises only a memory and a processor and possibly other normal components of a computer or workstation, controlling the medical instrument encompasses receiving processing and computing data for generating data that is rendered or can be rendered into a medical image. In other cases, the medical imaging system may comprise other components. Accordingly, the processor may be configured to control these other components.

[0010] Execution of the machine executable instructions causes the processor to receive MRF magnetic resonance data of a region of interest acquired according to an MRF magnetic resonance imaging protocol.

[0011] Execution of the machine executable instructions also causes the processor to reconstruct an MRF vector for each voxel in a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol. In magnetic resonance fingerprinting, one or more changes in the pulse sequence are made and data in k-space is sampled. The sampled k-space data is then reconstructed into an image. Instead of reconstructing the acquired magnetic resonance data directly into an image, a series of images is generated instead. One or more values ​​for this sequence of images are then assembled into an MRF vector or a fingerprint for each voxel. The MRF vector may also be referred to as an MRF signal. In conventional magnetic resonance fingerprinting, the MRF vector will then be compared with a dictionary including MRF fingerprints or vectors for known substances or properties of the object. In this embodiment, however, the MRF vector that can be compared with the magnetic resonance fingerprinting dictionary is handled differently.

[0012] Execution of the machine executable instructions also causes the processor to calculate a preprocessed MRF vector for each voxel in the set of voxels by applying a predetermined preprocessing routine to the MRF vector for each voxel. The predetermined preprocessing routine includes normalizing the preprocessed MRF vector for each voxel. In this execution, the MRF vector for each voxel is calculated by the processor as a preprocessed MRF vector. The predetermined preprocessing routine is a routine for normalizing the MRF vector for each voxel. Execution of the machine executable instructions also causes the processor to calculate an outlier map for the set of voxels by assigning outlier scores to the preprocessed MRF vector using a machine learning algorithm.

[0013] After the MRF vector has been processed by a predetermined preprocessing routine, the resulting preprocessed MRF vector is then input into a machine learning algorithm. The output of the machine learning algorithm is an outlier map. As used herein, an outlier map covers assigning values ​​to each voxel, which indicates how close the preprocessed MRF vector is to a previous measurement that is considered normal. The outlier map is then a measure of how well the MRF vector matches the expected value for each voxel. This can have several advantages. First, the outlier map can be used to indicate which areas of the object should be evaluated or studied more strongly by a physician or other medical expert. This embodiment also has the advantage that it is able to identify abnormalities in a magnetic resonance fingerprinting protocol that may not be otherwise identified. In conventional magnetic resonance fingerprinting, it was mentioned earlier that the MRF vector is compared with a magnetic resonance fingerprinting dictionary. The construction of the dictionary requires that the relevant properties of the object (such as t1, t2) or other relaxation values ​​or concentrations of various compounds and substances are added to the dictionary. The use of a machine learning algorithm enables the identification of abnormal parts of an object without having to use a magnetic resonance fingerprinting dictionary.

[0014] The outlier map can, for example, be stored for later use or it can be plotted or displayed in different ways on a display. The outlier map can also be used in various ways. The outlier map can be used to plan additional imaging protocols on the object, which can, for example, be conventional magnetic resonance imaging or even further magnetic resonance fingerprinting studies. The outlier map can also be useful for identifying abnormal areas of the object, for example for cancer screening or tracking the growth or remission of cancer during treatment.

[0015] In another embodiment, the medical imaging system further comprises a magnetic resonance imaging system. The memory further comprises an MRF pulse sequence command, which is configured to control the magnetic resonance imaging system to acquire MRF magnetic resonance data from the region of interest. The MRF pulse sequence command is a pulse sequence command. As used herein, a pulse sequence command encompasses commands or data that can be converted into commands for controlling a magnetic resonance imaging system to acquire magnetic resonance data. In this example, it is used to acquire the MRF magnetic resonance data. The MRF magnetic resonance data is magnetic resonance data. The term MRF is used to specify specific magnetic resonance data.

[0016] The memory also includes an MRF pulse sequence command configured to control the magnetic resonance imaging system to acquire MRF magnetic resonance data from the region of interest. Execution of the machine executable instructions also causes the processor to control the magnetic resonance imaging system to acquire the MRF magnetic resonance data. In this embodiment, the medical imaging system includes a magnetic resonance imaging system that is used to acquire the MRF magnetic resonance data.

[0017] In another embodiment, execution of the machine executable instructions further causes the processor to receive a segmentation of the voxels. The segmentation identifies a voxel type for each of the set of voxels. For example, the segmentation may identify a tissue type or an organ type for a particular voxel. The outlier score is assigned at least in part by using the voxel type as an input to the machine learning algorithm. In this embodiment, the machine learning algorithm may be able to use the description of the voxels in the segmentation to further identify whether a particular voxel has what is considered a normal MRF vector. This may have the advantage of making the outlier map more accurate.

[0018] The segmentation of the set of voxels can take different forms of different paradigms. For example, the MRF magnetic resonance data can be used together with a magnetic resonance fingerprinting dictionary to generate one or more MRF magnetic resonance images. These MRF magnetic resonance images can, for example, be segmented or the classification of the voxels using the magnetic resonance fingerprinting dictionary can also be used as input to the machine learning algorithm. In other examples, a conventional magnetic resonance image acquired separately can be segmented and the segmentation can be used.

[0019] In another embodiment, the voxel type is an anatomical location derived from an annotated anatomical atlas.

[0020] In another embodiment, the voxel type is a tissue type.

[0021] In another embodiment, the voxel type is an organ type.

[0022] In another embodiment, the voxel type can be considered as a so-called global voxel. In the case of global voxels, there is no difference in use. This can be combined with other results. For example, the outlier map can be calculated using the anatomical location, the tissue type, the organ type and / or the global voxel designation. These different combinations of outlier maps can be combined to produce a weighted or global outlier map or can be screened individually. For example, in some instances, an outlier map defined using the global voxel definition may not identify abnormal voxels, but when segmentation is used and a specific organ type is identified for each voxel, it can actually show that the voxel is abnormal for that specific type of organ, but is normal in the global aspect.

[0023] In another embodiment, the execution of the machine executable instructions further causes the processor to reconstruct the magnetic resonance image by reconstructing the magnetic resonance image according to the MRF vector for each voxel using the magnetic resonance fingerprinting dictionary. The execution of the machine executable instructions may also include optionally calculating the segmentation of the voxels according to the reconstruction of the magnetic resonance image. This embodiment may be beneficial because a magnetic resonance image generated according to magnetic resonance fingerprinting is generated and the outlier map is also provided. The outlier map can, for example, be superimposed or compared with the magnetic resonance imaging image. This can assist in the interpretation of the magnetic resonance image and / or guide a physician or other user of the magnetic resonance image to view specific areas identified as abnormal by the outlier map.

[0024] In another embodiment, execution of the machine executable instructions further causes the processor to reconstruct a magnetic resonance image by reconstructing the magnetic resonance image based on the imaging magnetic resonance data. The imaging magnetic resonance data describes the region of interest. For example, the memory may contain imaging magnetic resonance pulse sequence commands that are used to control the magnetic resonance imaging system to acquire the imaging magnetic resonance data. This embodiment may be beneficial because the outlier image may be, for example, thresholded and voxels above or below a specific value may be indicated on the magnetic resonance image. This may help identify abnormal anatomical regions of the object or regions that should be imaged or further studied.

[0025] As used herein, a magnetic resonance image encompasses an image generated by magnetic resonance data. In the above examples, the magnetic resonance image is either directly calculated from the MRF vector or from imaging magnetic resonance data. The magnetic resonance image can map a single or multiple parameters. For example, the magnetic resonance image can be a T1-weighted conventional magnetic resonance image, or it can also be a T1 value calculated from the magnetic resonance fingerprint recognition vector. In some examples, the magnetic resonance image can include more than one value or a multidimensional signal for each specific voxel. The magnetic resonance image can then be drawn using different colors or schemes for showing multidimensional signals.

[0026] In another embodiment, execution of the machine executable instructions further causes the processor to identify abnormal voxels by thresholding the outlier map. Execution of the machine executable instructions further causes the processor to render a medical image including the magnetic resonance image. The abnormal voxels are marked in the medical image.

[0027] In another embodiment, execution of the machine executable instructions further causes the processor to modify the machine learning algorithm by receiving training MRF magnetic resonance data acquired according to an MRF magnetic resonance imaging protocol for a training region of interest. Execution of the machine executable instructions further causes the processor to modify the machine learning algorithm by reconstructing a training MRF vector for each voxel in a set of training voxels describing the training region of interest using the training MRF magnetic resonance data acquired according to the MRF magnetic resonance imaging protocol.

[0028] Execution of the machine executable instructions further causes the processor to modify the machine learning algorithm by applying the predetermined preprocessing routine to the MRF vector for each voxel by calculating a trained preprocessed MF vector for each of a set of training voxels. Execution of the machine executable instructions further causes the processor to modify the machine learning algorithm by adjusting a variable parameter of the machine learning algorithm using the trained preprocessed MRF vector.

[0029] In another embodiment, the predetermined pre-processing routine includes reducing the dimension of the MRF vector for each voxel. The MRF vector includes components or values ​​for each image acquired in the MRF magnetic resonance imaging protocol. It may be possible to reduce the dimension of the MRF vector by the predetermined pre-processing routine to simplify the calculation of the outlier map by the machine learning algorithm. This may have the benefit of reducing the calculation time of the outlier map and / or improving the reliability of the outlier map.

[0030] In another embodiment, the reduction of the dimensionality of the MRF vector for each voxel comprises applying a Fourier transform to the MRF vector and truncating the Fourier transformed MRF vector above a predetermined frequency value. For example, the MRF vector may have hundreds of different entities. The MRF vector may be Fourier transformed using a discrete Fourier transform and then essentially a low pass filter is applied to it.

[0031] In another embodiment, the reduction of the dimension of the MRF vector includes compressing the MRF vector using a principal component analysis algorithm. This can be useful for eliminating components of the MRF vector that do not contain relevant information.

[0032] In another embodiment, the reduction of the dimensionality of the MRF vector may include calculating multiple relaxation times using the magnetic resonance fingerprinting dictionary. For example, different relaxation times selected from T1, T2, T2* or other specific combinations of relaxation times may be useful in identifying abnormal voxels.

[0033] In another embodiment, the predetermined pre-processing routine includes applying a mask to remove selected voxels from the set of voxels.For example, if a voxel is within a portion of the imaging volume that does not contain any tissue, it may not be useful or constructed to apply or calculate the outlier therein.

[0034] In another embodiment, the predetermined pre-processing routine includes deleting a voxel selected from the set of voxels if the MRF vector is below a predetermined measure or a predetermined magnitude. For example, if a particular voxel is on the boundary of an object and does not contain a sufficient amount of tissue or other material, it may not be useful for calculating the outlier map. In some embodiments, these above methods may be applied.

[0035] In another embodiment, the trained machine learning algorithm is an outlier detection algorithm. An outlier detection algorithm is an algorithm configured to detect how much a measurement deviates from other measurements in order to raise suspicion that it is generated by a different mechanism or a different physical process. This may also be referred to as anomaly detection. The outlier detection algorithm may be, for example, a density-based technique such as a k-nearest neighbor or a local outlier factor. Neural networks or various ensemble techniques or fuzzy logic may also be used.

[0036] In another embodiment, the machine learning algorithm is an isolation forest algorithm.

[0037] In another embodiment, the machine learning algorithm is a k-nearest neighbors algorithm.

[0038] In another embodiment, the machine learning algorithm is a one-class support vector machine algorithm.

[0039] In another embodiment, the predetermined pre-processing routine includes calculating a spatial average MRF vector for a predetermined area around each of the set of voxels. The predetermined pre-processing routine also includes appending the spatial average MRF vector to the pre-processed MRF vector. These actions are performed by the processor. The predetermined area can, for example, be a voxel within a predetermined distance of the voxel.

[0040] The predetermined area may also be a segmented area. For example, the segmentation may be used to divide the voxels into regions belonging to different organs or tissue types. The spatial average MRF vector may be calculated for a specific tissue type or for a specific organ. The outlier map may then be used to detect abnormal voxels within a specific tissue type of an organ. This may be useful for indicating abnormal anatomical structures or diseased tissues.

[0041] This embodiment may have the advantage that the outlier map is a combined outlier map of the MR fingerprint of each voxel and the information about the spatial environment around the voxel.

[0042] Adding the mean MRF vector to the preprocessed MRF vector enables detection of abnormal spatial variations as abnormalities. When also combined with voxel type, abnormal spatial variations within specific organs or other segmented regions can also be detected.

[0043] In another embodiment, the predetermined pre-processing routine includes calculating a spatial gradient MRF vector for each of the set of voxels. The predetermined pre-processing routine also includes appending the spatial gradient MRF vector to the pre-processed MRF vector before calculating the outlier map. Spatial changes in the MRF vector can also be used to detect abnormal voxels. This can be useful, for example, in detecting abnormal spatial changes in the MRF vector. As with the spatial average MRF vector, the voxel type can also be input into the machine learning algorithm. This can be useful in detecting abnormalities or sudden changes in the MRF vector.

[0044] In another aspect, the present invention provides a computer program product comprising machine executable instructions for execution by a processor. Execution of the machine executable instructions causes the processor to receive MRF magnetic resonance data of a region of interest acquired according to an MRF magnetic resonance imaging protocol. Execution of the machine executable instructions also causes the processor to reconstruct an MRF vector for each voxel in a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol.

[0045] Execution of the machine executable instructions further causes the processor to calculate a preprocessed MRF vector for each of the set of voxels by applying a predetermined preprocessing routine to the MRF vector for each voxel. The predetermined preprocessing routine includes normalizing the preprocessed MRF vector for each voxel. Execution of the machine executable instructions further causes the processor to calculate an outlier map for the set of voxels by assigning outlier scores to the preprocessed MRF vectors using a machine learning algorithm.

[0046] In another aspect, the present invention provides a method for operating a medical imaging system. The method includes receiving MRF magnetic resonance data of a region of interest collected according to an MRF magnetic resonance imaging protocol. The method also includes reconstructing an MRF vector for each voxel in a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol. The method also includes calculating a preprocessed MRF vector for each of the set of voxels by applying a predetermined preprocessing routine to the MRF vector for each voxel. The predetermined preprocessing routine includes normalizing the preprocessed MRF vector for each voxel. The method also includes calculating an outlier map of the set of voxels by assigning an outlier score to the preprocessed MRF vector using a machine learning algorithm.

[0047] It should be understood that one or more of the above-described embodiments of the present invention can be combined as long as the combined embodiments are not mutually exclusive.

[0048] As will be appreciated by those skilled in the art, various aspects of the present invention may be implemented as an apparatus, method, or computer program product. Accordingly, various aspects of the present invention may take the form of a fully hardware embodiment, a fully software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects (all of which may be generally referred to herein as "circuits," "modules," or "systems"). In addition, various aspects of the present invention may take the form of a computer program product implemented in one or more computer-readable media having computer executable code implemented thereon.

[0049] Any combination of one or more computer-readable media can be utilized. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. As used herein, "computer-readable storage medium" encompasses any tangible storage medium that can store instructions that can be executed by a processor of a computing device. A computer-readable storage medium can be referred to as a computer-readable non-transitory storage medium. A computer-readable storage medium can also be referred to as a tangible computer-readable medium. In some embodiments, a computer-readable storage medium can also be capable of storing data that can be accessed by a processor of a computing device. Examples of computer-readable storage media include, but are not limited to, floppy disks, magnetic hard disk drives, solid-state hard disks, flash memories, USB thumb drives, random access memories (RAM), read-only memories (ROM), optical disks, magneto-optical disks, and register files of processors. Examples of optical disks include compact disks (CDs) and digital versatile disks (DVDs), such as CD-ROMs, CD-RWs, CD-Rs, DVD-ROMs, DVD-RWs, or DVD-R disks. The term computer-readable storage medium also refers to various types of recording media that can be accessed by a computer device via a network or communication link. For example, data can be retrieved on a modem, the Internet, or a local area network. Computer executable code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0050] A computer readable signal medium may include a propagated data signal having computer executable code implemented 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 may be any computer readable medium that is not a computer readable storage medium and is capable of conveying, propagating, or transmitting a program used by or in conjunction with an instruction execution system, apparatus, or device.

[0051] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory directly accessible by a processor. "Computer storage" or "storage device" is another example of a computer-readable storage medium. A computer storage device is any non-volatile computer-readable storage medium. In some embodiments, a computer storage device may also be computer memory, or vice versa.

[0052] As used herein, "processor" encompasses an electronic component capable of executing a program or machine executable instructions or computer executable code. References to computing devices including "processors" should be interpreted as being capable of containing more than one processor or processing core. The processor may be, for example, a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed between multiple computer systems. The term computing device should also be interpreted as being capable of referring to a collection or network of computing devices each including one or more processors. Computer executable code may be executed by multiple processors that may be within the same computing device or may even be distributed between multiple computing devices.

[0053] The computer executable code may include machine executable instructions or programs that cause a processor to perform aspects of the present invention. The computer executable code for performing operations for aspects of the present invention may be written in any combination of one or more programming languages ​​and compiled into machine executable instructions, including object-oriented programming languages ​​such as Java, Smalltalk, C++, etc. and conventional procedural programming languages ​​such as "C" programming language or similar programming languages. In some instances, the computer executable code may take the form of a high-level language or take a precompiled form and be used in conjunction with a decoder that generates machine executable instructions at work.

[0054] The computer executable code may be executed 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 through any type of network including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0055] Aspects of the present invention are described with reference to the flow chart, diagram and / or block diagram of the method, device (system) and computer program product according to an embodiment of the present invention. It should be understood that, when applicable, each square frame or part of the square frame of the flow chart, diagram and / or block diagram can be implemented by computer program instructions in the form of computer executable code. It should also be understood that, when mutually non-exclusive, the combination of the square frames in different flow charts, diagrams and / or block diagrams can be combined. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device that produces the machine, so that the instructions executed by the processor of the computer or other programmable data processing device create a unit for implementing the function / action specified in the flow chart and / or one or more block diagram frames.

[0056] These computer program instructions may also be stored in a computer-readable medium, which can direct a computer, other programmable data processing apparatus, or other device to work in a specific manner so that the instructions stored in the computer-readable medium produce a product including instructions for implementing the functions / actions specified in the flowchart and / or one or more block diagrams.

[0057] The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operating steps to be performed on the computer, other programmable apparatus, or other device, thereby producing a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide a process for the functions / actions specified in the flowchart and / or one or more block diagram boxes.

[0058] As used herein, "user interface" is an interface that allows a user or operator to interact with a computer or computer system. "User interface" can also be referred to as "human-machine interface device". The user interface can provide information or data to the operator and / or receive information or data from the operator. The user interface can enable the input from the operator to be received by the computer and can provide output to the user from the computer. In other words, the user interface can allow the operator to control or manipulate the computer, and the interface can allow the computer to indicate the effect of the control or manipulation of the operator. The display of data or information on a display or graphical user interface is an example of providing information to the operator. The reception of data by keyboard, mouse, tracking ball, touch pad, pointing stick, graphic input board, joystick, game controller, webcam, earphone, pedal, wired gloves, remote controller and accelerometer is all examples of user interface components that realize the reception of information or data from the operator.

[0059] As used herein, "hardware interface" encompasses an interface that enables a processor of a computer system to interact with and / or control an external computing device and / or apparatus. A hardware interface may allow a processor to send control signals or instructions to an external computing device and / or apparatus. A hardware interface may also enable a processor to exchange data with an external computing device and / or apparatus. 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 LAN connection, a TCP / IP connection, an Ethernet connection, a control voltage interface, a MIDI interface, an analog input interface, and a digital input interface.

[0060] As used herein, a "display" or "display device" encompasses an output device or user interface suitable for displaying images or data. A display may 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 (CRTs), memory tubes, bi-stable displays, electronic paper, vector displays, flat panel displays, vacuum fluorescent displays (VFs), light emitting diode (LED) displays, electroluminescent displays (ELDs), plasma display panels (PDPs), liquid crystal displays (LCDs), organic light emitting diode displays (OLEDs), projectors, and head mounted displays.

[0061] Magnetic resonance (MR) data is defined herein as the measurement of radio frequency signals emitted by atomic spins recorded using the antenna of a magnetic resonance device during a magnetic resonance imaging scan. MRF magnetic resonance data is magnetic resonance data. Magnetic resonance data is an example of medical image data. A magnetic resonance imaging (MRI) image or MR image is defined herein as a reconstructed two-dimensional or three-dimensional visualization of anatomical data contained within the magnetic resonance imaging data. This visualization can be performed using a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Preferred embodiments of the invention will be described hereinafter by way of example only and with reference to the accompanying drawings, in which:

[0063] Figure 1 An example of a medical imaging system is illustrated;

[0064] Figure 2 The diagram shows the operation Figure 1 A flowchart of a method of a medical imaging system;

[0065] Figure 3 illustrates another example of a medical imaging system;

[0066] Figure 4 The diagram shows the operation Figure 3 A flowchart of a method of a medical imaging system;

[0067] Figure 5 illustrates another example of a medical imaging system;

[0068] Figure 6 illustrates another example of a medical imaging system;

[0069] Figure 7 A flow chart illustrating an example of a method is shown;

[0070] Figure 8 An example of an outlier plot is shown;

[0071] Fig. 9Contains figures that belong to the supplement and illustrate automated wavelet balancing;

[0072] Fig.10 Contains figures that belong to the supplement and illustrate the definitions of HH, LH, and HL;

[0073] Fig.11 contains figures belonging to the supplement and showing a flow chart CS-SENSE including the estimation of the Laplace scaling measure;

[0074] Fig.12 contains figures that are supplementary and illustrate the relationship between 1D and 2D scales; and

[0075] Fig.13 The figures are included as supplementary and show the estimated Laplace scaling for the wavelet transform.

[0076] Reference numerals list

[0077] 100 Medical Imaging Systems

[0078] 102 Computer

[0079] 104 Hardware or network interface

[0080] 106 processors

[0081] 108 User Interface

[0082] 110 Memory

[0083] 120 machine executable instructions

[0084] 122 MRF magnetic resonance data

[0085] 124 MRF vectors for multiple voxels

[0086] 126 Preprocessed MRF vectors for multiple voxels

[0087] 128 Machine Learning Algorithms

[0088] 130 Outlier Plot

[0089] 200 Receiving MRF magnetic resonance data of the region of interest acquired according to the MRF magnetic resonance imaging protocol

[0090] 202 Reconstructing an MRF vector for each voxel in a set of voxels describing a region of interest using MRF magnetic resonance data according to an MRF magnetic resonance imaging protocol

[0091] 204 Calculate a preprocessed MRF vector for each voxel in a set of voxels by applying a predetermined preprocessing routine to the MRF vector for each voxel

[0092] 206 Compute an outlier map for a set of voxels by assigning outlier scores to preprocessed MRF vectors using a machine learning algorithm

[0093] 300 Medical Imaging Systems

[0094] 302 Magnetic Resonance Imaging System

[0095] 304 Magnet

[0096] 306 Magnet bore

[0097] 308 Imaging Area

[0098] 309 Area of ​​Interest

[0099] 310 Magnetic Field Gradient Coil

[0100] 312 Magnetic Field Gradient Coil Power Supply

[0101] 314 RF Coil

[0102] 316 Transceiver

[0103] 318 objects

[0104] 320 Object Support

[0105] 340 MRF pulse sequence command

[0106] 400 Controlling a magnetic resonance imaging system to acquire MRF magnetic resonance data

[0107] 500 Medical Imaging Systems

[0108] 542 Magnetic Resonance Fingerprint Recognition Dictionary

[0109] 544 Magnetic resonance images

[0110] 544' Magnetic Resonance Image

[0111] 546 Image Segmentation

[0112] 546' Image Segmentation

[0113] 548 Identification of abnormal voxels

[0114] 548' Identification of abnormal voxels

[0115] 550 Medical Image Rendering

[0116] 550' Medical Image Rendering

[0117] 600 Medical Imaging Systems

[0118] 640 Imaging pulse sequence command

[0119] 642 Imaging MRI Data

[0120] 700 Start

[0121] 702 Collection of MRF Data from Healthy Subjects

[0122] 704 A) Masking data based on organizational segmentation

[0123] 706 B) Optionally reduce the dimensionality of the MRF signal

[0124] 708 C) Normalized signal

[0125] 710 Training an outlier detection algorithm using selected / reduced MRF signals

[0126] 712 Collecting MRF data from the test object

[0127] 714 Prepare data according to steps A, B, and C

[0128] 716 Determine the abnormality score for each voxel

[0129] 718 Visualization of abnormal graph

[0130] 720 More objects?

[0131] 722 End

[0132] 800 Outlier score DETAILED DESCRIPTION

[0133] Like numbered elements in these figures are either equivalent elements or perform the same function. If the function is equivalent, an element that has been discussed previously will not necessarily be discussed in a later figure.

[0134] Figure 1 An example of a medical imaging system 100 is illustrated. Figure 1The medical imaging system 100 includes a computer 102. The computer 102 includes a hardware interface or network interface 104, which is shown as being connected to a processor 106. The hardware interface or network interface 104 can be used, for example, to exchange data with other computer systems or other components of the medical imaging system 100. For example, if the medical imaging system 100 will include a magnetic resonance imaging system, the hardware interface 104 can be used to send commands to control the magnetic resonance imaging system. The processor 106 is also shown as being connected to a user interface 108 and a memory 110. The memory 110 can be any combination of memory accessible to the processor 106. This can include things such as main memory, cache memory, and non-volatile memory (such as flash RAM, a hard drive, or other storage device). In some examples, the memory 110 can be considered a non-transitory computer readable medium.

[0135] The memory 110 is shown as containing machine executable instructions 120. The machine executable instructions 120 are commands that enable the processor 106 to perform various functions (such as control of other components of the medical imaging system 100) and perform various numerical and data processing tasks. The memory 110 is also shown as containing MRF magnetic resonance data 122. The MRF magnetic resonance data 122 is acquired according to a magnetic resonance fingerprinting or MRF magnetic resonance imaging protocol and describes a region of interest. The memory 110 is also shown as containing MRF vectors for a plurality of voxels 124 calculated using the MRF magnetic resonance data 122. The memory 110 is also shown as containing pre-processed MRF vectors for a plurality of voxels 126. The pre-processed MRF vectors for a plurality of voxels 126 are calculated using the MRF vectors 124 for a plurality of voxels. The memory 110 is also shown as containing a machine learning algorithm 128. The machine learning algorithm 128 can take the pre-processed MRF vectors for a plurality of voxels 126 as input and is configured to output an outlier map 130. An outlier map 130 is also shown as being stored in memory 110 .

[0136] Figure 2 The diagram shows the operation Figure 1Flowchart of a method of a medical imaging system 100 of the present invention. First, in step 200, MRF magnetic resonance data 122 is received. The MRF magnetic resonance data can be received, for example, via a network interface 104, via a storage medium connected to a computer 102, or it can also be acquired by a magnetic resonance imaging system controlled by the medical imaging system 100. Next, in step 202, MRF vectors for a plurality of voxels 124 are reconstructed using the MRF magnetic resonance data 122 according to an MRF magnetic resonance imaging protocol. Next, in step 204, preprocessed MRF vectors 126 for a plurality of voxels are calculated for each of a group of voxels by applying a predetermined preprocessing routine to the MRF vector 124 for each voxel. The predetermined preprocessing routine includes at least normalizing the preprocessed MRF vector for each voxel. Finally, in step 206, the method includes calculating an outlier map 130 for a group of voxels by assigning an outlier score to the preprocessed MRF vector 126 using a machine learning method 128.

[0137] Figure 3 Another example of a medical imaging system 300 is illustrated. In this example, the medical imaging system also includes a magnetic resonance imaging system 302. The magnetic resonance imaging system 302 includes a magnet 304. The magnet 304 is a superconducting cylindrical magnet with a bore 306 passing through it. The use of different types of magnets is also possible: for example, it is also possible to use both a split cylindrical magnet and a so-called open magnet. The split cylindrical magnet is similar to a standard cylindrical magnet except that the cryostat has been split into two parts to allow access to the isoplane of the magnet, and such a magnet can be used, for example, in conjunction with charged particle beam therapy. The open magnet has two magnet parts, one above the other, with enough space to receive the object: an arrangement of two partial areas similar to the arrangement of the Helmholtz coil. Open magnets are popular because the object is less restricted. In the cryostat of the cylindrical magnet, there is a collection of superconducting coils. Within the bore 306 of the cylindrical magnet 304, there is an imaging zone 308 where the magnetic field is strong and consistent enough to perform magnetic resonance imaging. A region of interest 309 is shown within the imaging zone 308. An object 318 is shown supported by an object support 320 so that at least a portion of the object 318 is within the imaging zone 308 and the region of interest 309.

[0138] In the bore 306 of the magnetic field, there is also a set of magnetic field gradient coils 310, which are used to acquire preliminary magnetic resonance data to spatially encode magnetic spins in the imaging zone 308 of the magnet 304. The magnetic field gradient coils 310 are connected to a magnetic field gradient coil power supply 312. The magnetic field gradient coils 310 are intended to be representative. Typically, the magnetic field gradient coils 310 include three separate coil groups for spatially encoding in three orthogonal spatial directions. The magnetic field gradient power supply supplies current to the magnetic field gradient coils. The current supplied to the magnetic field gradient coils 310 is controlled as a function of time and can be ramped or pulsed.

[0139] Adjacent to the imaging zone 308 is a radio frequency coil 314 for manipulating the orientation of magnetic spins within the imaging zone 308 and for receiving radio transmissions from spins also within the imaging zone 308. The radio frequency antenna may contain 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 intended to also 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 performed, the radio frequency coil 314 will have multiple coil elements.

[0140] In this example, the subject 318 is positioned so that the head region of the subject is within the region of interest 309. In other examples, other parts of the body of the subject 318 may be positioned in the region of interest 309.

[0141] The transceiver 316 and the gradient controller 312 are shown connected to the hardware interface 104 of the computer system 102. The memory 110 is further shown as containing MRF pulse sequence commands 340. The MRF pulse sequence commands 340 enable the processor 106 to control the magnetic resonance imaging system 302 to acquire the MRF magnetic resonance data 122.

[0142] Figure 4 The diagram shows the use of Figure 3 Flowchart of a method for a medical imaging system 300 of the present invention. The method starts at step 400. In step 400, the magnetic resonance imaging system 302 is controlled to acquire MRF magnetic resonance data 122 using the MRF pulse sequence command 340. The method then proceeds to Figure 2 Step 200 is illustrated in the flowchart of FIG. Figure 4 Then follow the method Figure 2 The method shown.

[0143] Figure 5 Another example of a medical imaging system 500 is shown. The medical imaging system 500 is similar to Figure 3 3. The medical imaging system 300 in FIG. 3. The memory 110 of the medical imaging system 500 is shown as containing several additional items. The memory 110 is shown as containing a magnetic resonance fingerprinting dictionary 542. The memory 110 is also shown as containing a magnetic resonance image 544 generated using the magnetic resonance fingerprinting dictionary 542 and the MRF magnetic resonance data 122 according to a conventional magnetic resonance fingerprinting imaging protocol.

[0144] Thus, the magnetic resonance image 544 in the memory 110 can be considered as a magnetic resonance fingerprinting magnetic resonance image. The memory 110 is also shown as containing an image segmentation 546 generated from the magnetic resonance image 544. The image segmentation 546 can, for example, assign a voxel type to each of the MRF vectors for the plurality of voxels 124. The same is true for the pre-processed MRF vectors 126. The machine learning algorithm 128 can then take these voxel type identifications as additional inputs for determining the outlier map 130. The memory 110 is also shown as containing an identification of an abnormal voxel 548 calculated by thresholding the outlier map 130. The memory 110 is also shown as containing a rendering of a medical image 550 as an identification of the abnormal voxel 548 and the magnetic resonance image 544.

[0145] Figure 6 A further example of a medical imaging system 600 is illustrated. Figure 6 The medical imaging system 600 in Figure 3 1. The medical imaging system 300 in FIG. 1 has several additional items in its memory 110. The memory 110 is also shown as containing imaging pulse sequence commands 640. The imaging pulse sequence commands are instructions that enable the processor 106 to control the magnetic resonance imaging system to acquire imaging magnetic resonance data 642. The imaging pulse sequence commands 640 are in accordance with a conventional magnetic resonance imaging protocol (such as T1, T2) or a proton density weighted magnetic resonance imaging protocol. Therefore, the imaging magnetic resonance data 642 can be considered as magnetic resonance data acquired in accordance with a conventional magnetic resonance imaging protocol.

[0146] The memory 110 is also shown as containing a magnetic resonance image 544' reconstructed from the imaging magnetic resonance data 642. The memory 110 is also shown as containing an image segmentation 546' of the magnetic resonance image 544'. The memory 110 is also shown as containing an identification of abnormal voxels 548' accomplished by thresholding the outlier map 130. The image segmentation 546' can, for example, have been used as an input to the machine learning algorithm 128 to generate the outlier map 130. The memory 110 is shown as also including a rendering of a medical image 550' combining the identification of abnormal voxels 548' and the magnetic resonance image 544'.

[0147] Multiparametric MR images yield additional information about tissue compared to single MR contrast. MR fingerprinting (MRF) can be considered a multidimensional imaging technique because each fingerprint includes a large number of data points. These are partially correlated, but still span a parameter space that can be larger than the two or three parameter dimensions known from standard multiparametric imaging, depending on the number of parameters encoded in the MRF sequence.

[0148] Paradigms can overcome one or more of the following disadvantages:

[0149] - Screening applications usually require knowledge about a specific clinical problem so that a protocol can be chosen and the radiologist can be asked to search for specific patterns. Therefore, screening applications will always be specific and require many resources.

[0150] - Abnormalities in MRI are almost exclusively based on contrast changes in qualitative MR images and cannot detect global changes in tissue properties.

[0151] - Multi-parameter imaging, such as magnetic resonance fingerprinting (MRF), can offer some more flexibility in generating contrast, but typically still requires a priori information about what contrast to generate, and in the case of MRF, what signal to expect in order to build a pre-computed dictionary.

[0152] The paradigm can overcome these problems by proposing a way to detect irregularities and abnormal features from multidimensional (MRF) image data. The evaluation can be performed in either multidimensional parameter space (e.g., T1, T2, diffusion) or multidimensional feature space, which can be obtained by compressing the high-dimensional MRF signal without explicitly extracting physical tissue parameters before processing. Without the need for a model that accurately includes all physical effects, the method allows abnormal patterns in the image to be found and displayed. This can be used as a computer-aided diagnostic tool that can be applied in daily diagnostic imaging and will benefit screening applications as well as follow-up examinations, so that radiologists can be made aware of areas in the image that may need closer investigation.

[0153] Examples may include one or more of the following features:

[0154] - Setup for measuring MRF signals;

[0155] - Algorithms that learn the features of MRF signals and compute anomaly scores in multidimensional feature space;

[0156] - Method for calculating anomaly maps based on measured MRF data;

[0157] - Methods for displaying anomaly maps optionally in combination with other parameter maps.

[0158] As mentioned previously, MRF is typically performed by matching the measured fingerprint signal to a precomputed dictionary. This requires that the model underlying the dictionary computation is accurate and includes physical effects associated with the imaging system and the object. Once the tissue parameters are estimated, the proposed analysis can be performed on this multi-parameter data. (This is in fact possible if the multi-parameter mapping is not an MRF.) A limitation of the approach mentioned above is that it requires the dictionary to be complete with respect to possible material / tissue types, since in other cases unknown tissues will be mapped to (wrong) known tissues. Detection of irregular or abnormal features of the investigated tissue is only possible when all these requirements are met. Potential errors in the estimated tissue parameters may compromise the anomaly map.

[0159] Signal anomaly assessment (computation of outlier maps) can also be performed directly on the MRF time signal or any other multidimensional feature vector without explicitly applying a signal model to extract tissue parameters. The approach is at the core of some examples:

[0160] Machine learning algorithms are trained on healthy volunteer data to be able to measure how likely it is that the probe signal is normal or has some abnormal irregularity. Without any knowledge of the underlying MR sequence or clinical question, the resulting map shows a measure of how "normal" or "abnormal" the tissue is.

[0161] Figure 7 A flow chart illustrating an example of the method is shown. The method begins at step 700 to start the method. Next, in step 702, magnetic resonance fingerprinting data is collected from a healthy subject. Step 704 is an optional step. In step 704, the data is masked based on tissue segmentation and / or signal intensity. The data in step 704 is the MRF vector 124. Step 702 is equivalent to Figure 4 Step 706 is optional. In step 706, the dimension of the magnetic resonance fingerprint signal is reduced. Next, in step 708, the MRF vector 124 is normalized. Steps 704, 706 and 708 are collectively equivalent to Figure 2Next, the method proceeds to step 710. In step 710, a machine learning algorithm, which may also be referred to as an outlier detection algorithm, is trained using the preprocessed MRF vectors 126 resulting from the application of steps 704, 706, and 708. Steps 702, 704, 706, 708, and 710 are collectively equivalent to a training or learning step for a machine learning algorithm. The steps after step 710 are equivalent to Figure 2 and 4 The method shown.

[0162] After step 710, the method proceeds to step 712. In step 712, magnetic resonance fingerprint recognition data is collected from the test object. This is equivalent to Figure 4 The difference between step 712 and step 702 is that in step 702, the magnetic resonance fingerprinting data is collected from what is referred to as a healthy or normal subject. It is expected that the data collected in step 702 may contain no abnormalities, very few abnormalities, or a reduced number of abnormalities. In step 712, it is unknown whether the magnetic resonance fingerprinting data was collected from a subject having any abnormalities in its anatomy. After step 712, the method proceeds to step 714. Step 714 is equivalent to Figure 2 and 4 204 in step 714. The steps performed in step 714 are equivalent and identical to the previously performed steps 704, 706, 708. Next in 716, an anomaly score is calculated for each voxel 716. Step 716 is equivalent to step 206. The anomaly score is equivalent to the outlier score. The method then proceeds to step 718. In step 718, an anomaly map, which may also be referred to as an outlier map, is visualized. Step 718 is optional. The outlier map or anomaly map may also be stored in memory for later use. The method then proceeds to step 720. Step 720 is a decision box. The question is whether there are more objects to process. If the answer is "no", then the method proceeds to step 722 and Figure 7 If the answer is "yes", the method proceeds back to step 712 and steps 712, 714, 716 and 718 are performed again.

[0163] Preparation of training MRI data:

[0164] First, MRF data / multi-parameter MR data are acquired from several healthy subjects. The voxels to be included in the training data set are selected by signal amplitude (low amplitude signals, e.g. outside the subject, are rejected) and optionally by tissue type (this can be achieved by computing a standard MRF classification map, computing a standard MRF parameter map and selecting a parameter range, manual segmentation, or automated model-based segmentation).

[0165] The signal selected as the training data set can optionally be compressed by dimensionality reduction. For example, in the case of spiral sampling (using a periodically rotating spiral), the Fourier transform of the subsequently selected signal with low frequency components can be used to eliminate sampling artifacts while maintaining most of the encoded tissue information.

[0166] The reduced signal vector is still multidimensional and can span from ten to hundreds of dimensions.

[0167] The signal is then normalized to allow a distance metric that is independent of the bulk material in the voxel. Optionally, the vector elements can be converted to absolute numbers to match the requirements of some machine learning algorithms. Alternatively, the training data can be estimated tissue parameters, T1, T2, proton density, diffusion and / or other parameters.

[0168] Training of machine learning algorithms:

[0169] Once the dataset is prepared, train or modify a machine learning algorithm. Some example algorithms that can be used are:

[0170] -Isolation Forest Algorithm:

[0171] This algorithm is well suited for multidimensional clustering and outlier analysis. Once trained, it computes a score for each test vector specifying its consistency with the training data.

[0172] -k-NN (k nearest neighbor) algorithm:

[0173] For each test vector, the algorithm returns the distance of the k nearest neighbors from the training data set. In the context of some examples, the average distance of the k nearest neighbors (where, for example, k=3) will be used as a measure of abnormality for the test vector.

[0174] -Single Class Support Vector Machine Algorithm:

[0175] The algorithm uses a set of training examples to define the boundary between inliers and outliers. Test data points are then classified as belonging to one of two categories. The algorithm is an example of a non-probabilistic binary classifier that uses a model that assigns new examples to one category or the other.

[0176] Analysis of test dataset:

[0177] The signals measured from the test subjects (patients to be screened) are prepared in the same way as the training signals (voxel selection, dimensionality reduction, normalization).

[0178] The signals are then tested using a trained machine learning algorithm that produces an abnormality score for each voxel.

[0179] Additionally, the location of the voxels (e.g., frontal lobe, hippocampus, etc.) can be considered for assessing the abnormality score of a given signal. This can be achieved by warping the acquired data to an annotated atlas, for example combining model-based segmentation with grid-based image warping.

[0180] Figure 8 An example of an outlier map 130, which may also be referred to as an anomaly map, is illustrated. The map 130 is a grayscale image that assigns an outlier score 800 to each voxel in the slice shown. The outlier score 800 is shown by the grayscale bar on the right side of the image 130. The outlier score 800 may also be referred to by the term anomaly score. Figure 8 The outliers or anomaly map of is calculated using a 3-NN algorithm: voxels containing fluid or very low signal have been excluded. The value indicates the average distance of the MRF signal from the three closest nearest neighbors selected from the training data set. In this case, this is the Euclidean distance of the 30-dimensional normalized feature vector. Since this is a healthy test subject, no significant anomalies are visible. This visualization can be superimposed (possibly in color) on a standard contrast or parameter map.

[0181] In another example, the abnormality map is analyzed statistically or via image analysis to generate a recommendation for the radiologist. The recommendation can then be displayed as a labeled image or text, such as "high probability for irregularity in the frontal lobe region" or "please check for the presence of a lesion in region __".

[0182] 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.

[0183] By studying the drawings, the disclosure and the claims, those skilled in the art can understand and implement other variations of the disclosed embodiments when practicing the claimed invention. In the claims, the word "comprising" does not exclude other elements or steps, and the words "one" or "an" do not exclude multiple. A single processor or other unit can perform the functions of several items recorded in the claims. Although specific measures are recorded in mutually different dependent claims, this does not indicate that a combination of these measures cannot be used advantageously. The computer program can be stored / distributed on a suitable medium, such as an optical storage medium or solid-state medium provided with other hardware or as part of other hardware, but the computer program can also be distributed in other forms, such as distributed via the Internet or other wired or wireless telecommunications systems. Any figure mark in the claims should not be interpreted as limiting the scope.

[0184] Replenish

[0185] Supplement Title: Automated Wavelet Equilibrium Supplement Authors: Miha Fuderer, Adri Duijndam, and Elwin deWeerdt

[0186] The authors of this supplement are different from the rest of the document. The authors of the supplement are not the authors of the rest of the document. The text between the header supplement above and the claims below belongs to the supplement.

[0187] summary:

[0188] Compressed sensing MRI is a valuable technique for speeding up acquisition time; however, it requires a factor that balances the impact of wavelet constraints relative to data fidelity constraints. This factor is often set "manually". The paradigm suggests automatically estimating this factor. The optimal balancing factor is calculated based on the statistics of a set of individual wavelet scales.

[0189] Detailed Description

[0190] Background of the Example

[0191] Compressed sensing (CS) is a technique with promising applications in MRI. It allows sampling a reduced set of data points in k-space while still reconstructing a good resulting image. This possibility is based on the knowledge that the resulting image is compressible. CS algorithms apply nonlinear operations in a domain where the image representation is assumed to be sparse. Typically, the domain is the wavelet transform of the image, and the nonlinear operations include "thresholding", which consists in placing those wavelet elements below a certain threshold to zero (and by that, in effect, removing more noise and artifacts than the relevant signal).

[0192] Typically, mathematical problems are expressed as minimization of the following function:

[0193]

[0194] Here, p is the image vector to be reconstructed; m is the measured data, A=MFS is the encoding matrix containing, where S contains the coil sensitivities, F denotes the Fourier transform and M denotes the matrix describing the (irregular) subsampling. In fact, the matrix A also contains regularization information, see the more detailed description below. W denotes the wavelet transform, and ||.||1 should be read as "L1 norm" (sum of absolute values). The factor "λ" will be discussed further. Note that regularization is not taken into account in this formulation, but can be easily entered without losing any essential elements, see the appendix accompanying it.

[0195] In practice, the first term tells that we want to find p that fits the data; the second term tells that p should be compressible. And λ weights one against the other. Needless to say that the resulting image will strongly depend on the choice of λ. In current practice, the factor λ is set ad hoc, i.e., manually tuned for a specific application. The problem or disadvantage is overcome by some embodiments.

[0196] We aim to derive a favourable choice of λ from the image or data properties and do so in a wide range of cases (compression factor, resolution, contrast, signal-to-noise conditions, etc.) without requiring the user to set the value.

[0197] A paradigm can contain the following elements, which can be applied independently of each other:

[0198] The factor λ is set depending on the histogram statistics of the wavelet components. More specifically, the histogram statistics of the finest wavelet components are derived from the statistics of the coarser scales.

[0199] • To preserve resolution, we limit the factor λ to a value that ensures that the signal energy in the finest wavelet scale will not decrease more than a fixed fraction (eg, to 90%) of the signal energy we would obtain if no sparsity term was applied.

[0200] Detailed descriptions of how to build and use some examples

[0201] Specific Example The following processing steps apply to each slice and each iteration of the Nesterov / Fista optimization scheme. See the Appendix for a detailed description of the computations required within the wavelet thresholding step.

[0202] - The Laplace scale of the image is calculated for each wavelet scale by analysis of the histogram of the values ​​of each wavelet scale. The Laplace scale is here called "b". In this step, a correction factor for the amount of background is taken into account.

[0203] - Estimate the parameters describing the exponential relationship between the Laplace scaling and the wavelet scaling. A combined estimation procedure is used here, where the so-called HH, LH and HL wavelet domain parts are separated in the exponential model but estimated in an integrated manner. The anisotropic nature of the voxel size is taken into account.

[0204] - Derive the lowest meaningful Laplace scale from this exponential relation. See the Appendix for details. These three steps together lead to a refined (and stable) estimate of b that belongs to the minimum wavelet scale.

[0205] -pass Derive the wavelet threshold, where σ is the standard deviation of the noise in the input data, and c is the matrix A H The largest eigenvalue of A (which is known).

[0206] - Calculate the permissible upper limit of the threshold as Here, f is the system tunability factor, for example 0.1, and R is the acceleration factor (the ratio of the sample density to the nominal required sampling density).

[0207] - Set the threshold to the maximum of the two previous values, i.e.,

[0208] - Apply wavelet thresholding using this threshold t.

[0209] Possible alternatives or refinements

[0210] The factor λ is set differently depending on the wavelet scale.

[0211] The thresholding function within a wavelet step is a function between hard thresholding and soft thresholding. More specifically,

[0212] We calculate the "subtraction value" (Here, is the wavelet component before thresholding). The value of p is preferably chosen to be 0.7; the essence of the idea is that it is chosen to be non-zero, but less than 1. The result of thresholding is calculated as

[0213]

[0214] Detectability

[0215] The detectability of details is an issue. However, as can be appreciated from the user interface, the system does not require manual setting of the wavelet parameters.

[0216] Some example applications

[0217] MRI

[0218] Supplementary Appendix - Background of the "Detailed Description" of the Supplementary Formula for the Inverse Problem Let the forward model for the measured data be given by,

[0219] m=MFSp+n(0.2)

[0220] Using the following definitions:

[0221] -m is a vector of measured data enhanced with zero values ​​at unmeasured k-space locations.

[0222] - p is a vector containing the pixels of the image,

[0223] -S is the matrix containing the coil sensitivities,

[0224] - F is the matrix representing the (full) Fourier transform.

[0225] - M is a block diagonal matrix containing values ​​of ones on the diagonal for measured k-space locations and zeros elsewhere (these weights can be generalized).

[0226] -n is the vector with noise.

[0227] The matrices S, F and M can be combined in the encoding matrix E = MFS. The noise is assumed to be a Gaussian distribution with zero mean. In the description, we will take the noise to be fully normalized, so the noise covariance matrix is ​​equal to the identity matrix.

[0228] The "model" of the familiar principle of regularization has a similar but simpler mathematical structure:

[0229] 0=Ip+n r (0.3)

[0230] The expected value of a given (complex-valued) image pixel is zero. The uncertainty is assumed to be Gaussian and its

[0231] The quality is specified in the diagonal covariance matrix R. This matrix will have low values ​​on the diagonal for voxels in the background, which effectively stabilizes the inverse problem and leads to a higher SNR.

[0232] The wavelet transform of an image p is given by the operation “Wp”. The sparse prior information with respect to the wavelet transform is formulated by the following probability density function:

[0233] p(p)∝exp(-‖B -1 Wp‖1) (0.4)

[0234] Here the uncertainty is assumed to have a Laplace distribution, which represents the assumption of a sparse distribution of the wavelet coefficients. The uncertainty is given by the diagonal matrix B.

[0235] Given these elements and using a Bayesian framework, the reconstruction problem boils down to minimizing the following function:

[0236]

[0237] Although the Laplace scaling is different from the wavelet scaling, in the optimization procedure (actually, in the included soft thresholding step), a constant value is used, using the form B = λ -1 I.

[0238] Solving this optimization problem requires a specialized algorithm described subsequently.

[0239] Optimization Algorithm

[0240] The optimization problem formulated above can be written in a more compact notation:

[0241]

[0242] Here, the data vector is redefined as:

[0243]

[0244] And, accordingly,

[0245]

[0246] The optimization problem (0.6) can be solved using an iterative soft thresholding algorithm, see the description of FISTA (Fast Iterative Soft Thresholding Algorithm)

[0247] For i = 1, 2, ..., imax, do:

[0248]

[0249] where S is the element-wise soft thresholding operator

[0250]

[0251] And c should be less than or equal to the matrix A H The largest eigenvalue of A.

[0252] Wavelet Thresholding

[0253] First, some basic properties of the wavelet transform of real-world images are introduced. Then, the estimation of some of the relevant properties is discussed. The application of wavelet thresholding is treated as follows.

[0254] An explanation of the wavelet transform is beyond the scope of this document. The reader should refer to https: / / en.wikipedia.org / wiki / Wavelet_transform and https: / / en.wikipedia.org / wiki / Discrete_wavelet_transform and references there.

[0255] Choice of wavelet transform

[0256] In principle, the examples can be applied to any type of wavelet transform used. The Daubechie family of wavelets and in particular the D4 wavelet is a well-known and popular choice. See also https: / / en.wikipedia.org / wiki / Discrete_wavelet_transform .

[0257] One property of this form of wavelet transform is that it is not shift invariant and is directionally selective. Adverse effects related to (the lack of) these properties have been seen on real data sets. One option to overcome these limitations is the use of a dual-tree complex wavelet transform.

[0258] Wavelet transform properties To introduce this topic, let us consider Fig. 9 The images and their corresponding wavelet transforms are shown (top left and middle). Two things can be observed: (1) for each wavelet scale, many coefficients have low absolute values; and (2) the amplitudes of the coefficients increase for coarser scales.

[0259] Fig. 9 The absolute values ​​of an image (top left) and a wavelet transform (top right) are shown. The top right shows the definition of the wavelet scale. The bottom left shows a histogram of a set of coefficients for one of the wavelet scales. When the width of the distribution is plotted as a function of the wavelet scale, the bottom right plot results. Note that the vertical scale is logarithmic. Let's look at these two observations in more detail in the following sections.

[0260] Laplace distribution of wavelet coefficients

[0261] In contrast to the first observation above, if we base the coefficients on a particular wavelet scale (see Fig. 9 If we generate a histogram from the graph below, we see that it roughly follows the Laplace distribution. For a single variable x, this distribution is given by:

[0262]

[0263] Where μ is the location parameter and b is the scale parameter. (To avoid confusion with the "wavelet scale", the scale parameter will mainly be referred to as the "Laplace scale"). Using this distribution, we can model the first observation above: for each wavelet scale, many coefficients have relatively low values. Significant coefficients are "sparse". This observation is used in the sparsity constraint as formulated in the minimization problem and is one of the most basic elements in the CS-SENSE algorithm. In addition to computational considerations, the wavelet transform was chosen for this reason. The image is sparse in the wavelet domain, and this information is used to constrain the output space of the SENSE scheme. Of course, in principle, any transform with a sparse representation of the image can be used.

[0264] Now from the inverse problem formulation, it is clear that we need to know the parameters of the Laplace distribution for each wavelet scale (in practice μ can be neglected, so it mainly belongs to the Laplace scale b). The estimation of these parameters is discussed below. It turns out that it is very helpful to use the relationship between the wavelet scales for this purpose.

[0265] Multi-scale relations

[0266] Let us define Fig. 9The wavelet scale s in (top right). It is defined as zero at the finest scale and increases by 1 for each coarser scale. The intermediate scales 0.5, 1.5, etc. are the horizontal or vertical components of the wavelet transform. It can be seen that the Laplace scale is larger for coarser wavelet scales. The correlation with respect to wavelet scale can be formulated as follows:

[0267] b s =b0q s ,s=0,0.5,1,…(0.12)

[0268] This relation (where q≈3) has been found to describe the data very well, except for the finest scales, see Fig. 9 The example is used to illustrate this point.

[0269] Since this piece of prior knowledge is essential to the approach for wavelet thresholding, its theoretical background will be discussed in more detail in the next section.

[0270] Background on the relationship between wavelet domain and k-space and multi-scale relationship

[0271] In order to provide a rational number for (0.12), it is important to understand the relationship between the wavelet scale and the position in k-space. A good way to illustrate this relationship is to study the Fourier transform of an image built from just one wavelet coefficient.

[0272] First, we need to realize that the wavelet coefficients are located in different parts of the wavelet domain HH, LH and HL, see below Fig.10 . In the indication of these sections, the first character corresponds to the first axis, and the second character corresponds to the second axis. The first axis is considered horizontal, and the second axis is considered vertical. "L" stands for low pass filtering, and "H" stands for high pass filtering. Of course the different sections repeated for different wavelet scales correspond to different "wavelet types". In the sequel to this document, we will often refer to these different sections as "types" which we have in three different flavors as mentioned above.

[0273] Fig.10 The definitions of the HH, LH and HL parts of the wavelet domain (repeated for each wavelet scale) are shown.

[0274] Now consider a disk in image space. The Fourier transform of such a disk with radius r is given by

[0275]

[0276] Thus, reducing the wavelet scale by one corresponds to increasing the position in k-space by a factor of two, which in turn corresponds to 2 1.5The magnitude of the reduction is a factor of ≈2.83. Deviations in this factor actually originate from different anatomical shapes and finite slice thickness. The effects of zero filling in k-space and the effects of anisotropic resolution are discussed in the next section.

[0277] Estimation of Laplace scaling

[0278] The Laplace scale is not known before the reconstruction of the CS-SENSE scan, as it is object dependent. There is no pre-scan or general information available to derive these values ​​up front, so these properties must be estimated as an integral part of the CS-SENSE algorithm. And they must be estimated from the (partially) reconstructed images. There are two problems that hinder the estimation of these properties.

[0279] - At the very beginning of an iterative scheme, or during early iterations, we have only an incomplete reconstruction. In addition to noise, the image will still contain aliasing artifacts.

[0280] - Even at later and final iterations, the image will still be contaminated by noise and residual artifacts (possibly incomplete unwrapping, but also artifacts due to imperfect coil sensitivity, motion and other common causes of artifacts). These will particularly affect the finest wavelet scales, so those cannot be estimated reliably.

[0281] Fig.11 The following scheme shown should overcome these problems. Wavelet properties are estimated during iterations and for each iteration. The estimate will be inaccurate for early iterations and become more accurate as the iterations proceed. Laplace scales are estimated for all wavelet scales. Subsequently, a multi-scale relationship will be used in order to accurately estimate the Laplace scale for the finest wavelet scale(s) from the coarser wavelet scales. This last aspect reduces the effects from random noise and the artifacts mentioned above. Fig.12 A flow chart CS-SENSE including the estimation of the Laplace scaling quantity is shown.

[0282] The details of these steps are discussed below.

[0283] Estimation of individual Laplace scaling

[0284] Basic Estimates

[0285] Given a one-dimensional Laplace distribution and a set of N independent and identically distributed samples x1…x N , the maximum likelihood estimate is the sample median, and the maximum likelihood estimate of b is

[0286]

[0287] For complex-valued images as occurs in MRI, the sample median is assumed to be close enough to zero that it can be neglected. The estimate for the Laplace scale is then simply Alternative methods have also been proposed that use the 90% percentage of classified samples for the estimate of b, but we have found that the maximum likelihood estimate is more accurate. Furthermore, it does not require classification of the absolute values ​​of the samples and is therefore more efficient.

[0288] Correction for background values

[0289] It is desirable to make the estimate of the Laplace scale independent of the amount of zero padding done in the image space. After all, the nature of the image content is not changed simply by adding an edge with zero values ​​around it. Such zero padding actions occur in order to stretch the image to a matrix size suitable for the wavelet transform.

[0290] Now, if the fraction of zero-valued samples is denoted by the symbol f, then the mean of the wavelet coefficients is reduced by a factor of (1-f) compared to the case without zero padding. Accordingly, in order to arrive at a "zero-padding independent" Laplace scale, the mean from the previous subsegment needs to be divided by (1-f).

[0291] Similar reasoning applies to all samples in the background (not part of the object). Therefore, more generally, the raw estimated Laplace scale needs to be divided by (1-f), where f is the ratio of samples in the background to the total number of samples.

[0292] Estimation of the relationship between the levels

[0293] For the CS-SENSE scheme, q and b0 are estimated by first estimating the Laplace scale using equation (0.14) for all wavelet scales (and the three parts HH, LH and HL), including background mask correction. From these Laplace scales and the relationship ((0.12), q and b0 are estimated. There are several aspects that need to be considered in order to arrive at a "picture" of the wavelet scale-dependent Laplace scale that depends only on the intrinsic object properties and is independent of acquisition or processing related details. These latter aspects cover:

[0294] - Zero filling (interpolation) in k-space

[0295] - Anisotropic resolution

[0296] Zero filling in k-space

[0297] As an example, consider zero padding with a factor of two along a particular direction. Since this corresponds to interpolation by a factor of two, it is clear that the contents of the image in the wavelet domain are shifted up one scale compared to the case without interpolation. See Section 0. This needs to be incorporated into the definition of wavelet scaling. Let's add more detail to the specification of wavelet scaling, and in particular the relationship between one-dimensional and two-dimensional wavelet scaling.

[0298] Fig.12 shows the relationship between 1D and 2D wavelet scaling. 1D scaling is plotted along two axes. Let us define the wavelet scaling of a 1D wavelet transform for either x or y scaling as

[0299] s x,y =0,1,2,…(0.15).

[0300] We can now define the wavelet scale for the 2D transform as

[0301]

[0302] s x,y =0.5,1,1.5,…

[0303] To increase zero-filling in k-space (or interpolation in image space), it is clear that the wavelet scale resulting from the wavelet transform corresponds to a reduced "physical size". The "true finest scale" of the wavelet transform as used in the determination of the energy conservation threshold should correspond to the acquisition voxel size. Given a fixed field of view, since the wavelet scale has a logarithmic relationship with the matrix size (and thus the voxel size), the "true scale" along one direction (taking x as an example) can be defined as:

[0304] s x,true =s x -log2 f x,zp

[0305] s x =0,1,2,…

[0306] f x,zp = zero filling factor in k-space, along Using the definition of the 2D scale (0.16), we get

[0307]

[0308] s=0,0.5,1,…(0.18)

[0309] This implies that, for example, with a zero-filling factor of 2 for both dimensions, the "true scale" is reduced by one compared to the "numerical scale." This ensures a consistent estimate of the Laplace scale that is more or less independent of the amount of zero-filling.

[0310] Anisotropic resolution and multi-scale estimation

[0311] Let us now turn to the subject of anisotropic resolution and results on wavelet transforms. And more specifically on which wavelet scale (which "box" in the wavelet domain) we will likely base the threshold on. Consider a synthetic example of anisotropic resolution compared to anisotropic resolution for a disk. Fig.13 The corresponding estimated Laplace scales are shown. It should be clear that due to the anisotropic resolution, the Laplace scales for the three different types (HH, LH and HL) are significantly different. Fig.13 The estimated Laplace scaling for the wavelet transform is shown. Fig.12 ).

[0312] To handle this correctly, the multiscaling model (0.12) is used for each "wavelet type" separately. The base of the exponential function q is assumed to be the same for the types. Therefore, we use the model

[0313]

[0314] where, for the s values, the values ​​after correction for zero padding are used. The unknown parameters can be estimated by taking the logarithm on both sides of these equations:

[0315]

[0316] in,

[0317] a HH =ln b HH

[0318] a LH =ln b LH

[0319] a HL =ln b HL

[0320]

[0321] and solve (0.20) in a weighted least squares sense.

[0322] Estimation Details

[0323] Several practical aspects of the estimation process need to be covered:

[0324] - Number of wavelet levels

[0325] - Horizontal weighting

[0326] - Least Squares Alternatives

[0327] All wavelet levels are used in the estimation process. However, weights are applied defined as:

[0328]

[0329] Effectively, therefore, the finest scale (s=0) is not used.

[0330] L1 fitting can be used as an alternative to least squares fitting.

[0331] Threshold selection

[0332] Two aspects need to be considered when determining the exact form of the sparse term of the inverse problem to be solved. The first is whether the threshold should vary according to the wavelet type or whether a constant level is required. The second is the term "wavelet balance", the level of thresholding (relative to the data fitting and regularization terms). These topics are described in the next section.

[0333] Single or multi-scale thresholding

[0334] Given the above material and the general formulation of the inverse problem, one would expect that the best reconstruction would be one that would allow the threshold 1∝1 / b to decrease for increasing wavelet scale (which shows increasing b). While this is true from a purely quantitative point of view, the effect of level-dependent thresholding is that the noise becomes colored, which makes the image look artificial. They are unpleasant to the eye. In this case, the anatomy, coil sensitivity, and sampling pattern were taken from a neural scan. The normal data has been replaced with pure random white noise. Three different ways of thresholding have been applied, and the resulting images and corresponding wavelet transforms are shown (right). The color indicates the phase and can be ignored. The three methods are:

[0335] - Constant Threshold (top right),

[0336] - "Laplacian", horizontal correlation threshold, threshold from normal objects,

[0337] - "Laplace, energy preserving". This is as "Laplace", but the levels have been finite so that only a fixed percentage of the signal is removed according to the wavelet scaling.

[0338] It can be observed that while the noise for the constant threshold looks natural, the other results show colored noise (see also the corresponding wavelet transform).Based on these results and many real scans, it has been decided to apply only a constant threshold.

[0339] Threshold level / wavelet balance

[0340] Let's first list two of the main elements of some examples:

[0341] - "The application of the sparsity constraint in CS-SENSE does not require any user interaction". Or more precisely: the user does not have to specify the thresholding level.

[0342] - The application of the sparsity constraint does not result in a visible loss of resolution.

[0343] The first requirement implicitly requires an automatic scheme for determining threshold levels independent of any protocol parameters or other aspects of the scan.

[0344] Looking at the relationship between wavelet scale and type on the one hand and k-space (part 0) on the other hand, and the relationship between Laplace scale and wavelet scale and type, it is clear that the threshold should be based on the finest wavelet scale + type corresponding to the measured data. Since the HH type on the finest wavelet scale corresponds more or less to the non-measured (or filtered) part of k-space, it should not play a role in the determination of the threshold. Then, given the fact that the anisotropic resolution and possible anisotropic aspects of the object itself may cause significant amplitude differences between the LH and HL types, it has of course been decided to base the threshold on the minimum Laplace scale of these two types at the finest wavelet scale. In the formula, the Laplace scale is used:

[0345]

[0346] Now, the threshold can follow from this expression. But we also have the limitation that the threshold level should be chosen so that no visible resolution loss occurs due to thresholding. After some educated guesses and supporting experiments, the level has been chosen so that no more than 5% or 10% of the energy of the relevant wavelet scale / type is lost. Now let's see how the actual threshold can be derived from this level and the Laplace scale. In fact, a detailed description is beyond the scope of this document. An overview of the derivation is as follows. Given (one or more) Laplace scales, the inverse problem can be solved using the threshold using an iterative scheme (0.9):

[0347]

[0348] where the subscript "i" stands for "iterative scheme". Now, it would be extremely difficult to derive a bound on the threshold from the iterative scheme itself. However, it is possible to derive a bound for just a single thresholding step. Now, the derivation is based on the fact that, in practice, the CS-SENSE algorithm can be exactly split into a SENSE algorithm followed by a thresholding step, which uses the covariance matrix of the noise of the output image of the SENSE step. However, for practical results, the covariance matrix must be approximated. It can be shown that, using this approximation, the same problem can be solved in one iteration using the threshold,

[0349]

[0350] Here, the subscript "c" stands for "combined solution". Therefore, the two thresholds are related to each other via

[0351]

[0352] The next step comes from analyzing the threshold t c The bound on , so that part f of the energy of the signal is lost in the thresholding operation (for the combination scheme). This can be derived given the following condition:

[0353]

[0354] Combining this with the relationship (0.26) yields a bound on the threshold used in the iterative scheme:

[0355]

[0356] This value is used in the iteration scheme.

Claims

1. A medical imaging system, comprising: - a memory (110) for storing machine executable instructions (120); - a processor (106) for controlling a medical instrument, wherein execution of the machine executable instructions causes the processor to: - receiving (200) MRF magnetic resonance data (122) of a region of interest (310) acquired according to an MRF magnetic resonance imaging protocol; - reconstructing (202) an MRF vector (124) for each voxel of a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol; - calculating a pre-processed MRF vector for each voxel in the set of voxels (126) by applying a predetermined pre-processing routine to the MRF vector for each voxel, wherein the predetermined pre-processing routine includes normalizing the pre-processed MRF vector for each voxel; - calculating an outlier map (130) for the set of voxels by assigning outlier scores (800) to the pre-processed MRF vectors using a machine learning algorithm; and - receiving a segmentation of the set of voxels, wherein the segmentation identifies a voxel type for each voxel in the set of voxels, and wherein the outlier score is assigned at least in part by using the voxel type as an input to the machine learning algorithm.

2. The medical imaging system according to claim 1, wherein: The medical imaging system also includes a magnetic resonance imaging system, wherein the memory also includes MRF pulse sequence commands (340), the MRF pulse sequence commands being configured to control the magnetic resonance imaging system to acquire the MRF magnetic resonance data from the region of interest, wherein execution of the machine executable instructions causes the processor to control the magnetic resonance imaging system to acquire the MRF magnetic resonance data.

3. The medical imaging system according to claim 1 or 2, wherein: The voxel type is any of: an anatomical location derived from an annotated anatomical atlas, a tissue type, an organ type, a global voxel, and a combination thereof.

4. The medical imaging system according to claim 1 or 2, wherein: Execution of the machine executable instructions further causes the processor to reconstruct a magnetic resonance image according to any of: - reconstructing the magnetic resonance image (544) from the MRF vector for each voxel using a magnetic resonance fingerprinting dictionary (542); and - reconstructing the magnetic resonance image from imaging magnetic resonance data (642), wherein the imaging magnetic resonance data describes the region of interest.

5. The medical imaging system according to claim 4, wherein: Execution of the machine executable instructions further causes the processor to: - identifying outlier voxels by thresholding the outlier map; and - rendering a medical image comprising the magnetic resonance image, and wherein the abnormal voxels are marked in the medical image.

6. The medical imaging system according to claim 5, wherein: Execution of the machine executable instructions further causes the processor to train the machine learning algorithm by: - receiving (702) training MRF magnetic resonance data of a training region of interest acquired according to an MRF magnetic resonance imaging protocol; - reconstructing a training MRF vector for each voxel in a set of training voxels describing the training region of interest using the training MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol; - calculating a trained pre-processed MRF vector for each voxel in said set of training voxels by applying said predetermined pre-processing routine to said MRF vector for each voxel; and - Training (710) the machine learning algorithm using the trained preprocessed MRF vectors.

7. The medical imaging system according to claim 1 or 2, wherein: The predetermined pre-processing routine includes reducing the dimensionality of the MRF vector for each voxel.

8. The medical imaging system according to claim 7, wherein: Reducing the dimensionality of the MRF vector for each voxel includes any of the following: - applying a Fourier transform to the MRF vector and truncating the Fourier transformed MRF vector above a predetermined frequency value; - compressing the MRF vector using a principal component analysis algorithm; and - Use the MRF dictionary to compute multiple relaxation times.

9. The medical imaging system according to claim 1 or 2, wherein: The predetermined pre-processing routine includes any one of the following: - applying a mask to remove selected voxels from the set of voxels; - deleting said selected voxels from said set of voxels if said MRF vector is below a predetermined magnitude or a predetermined measure; and - combinations thereof.

10. The medical imaging system according to claim 1 or 2, wherein: The trained machine learning algorithm is an outlier detection algorithm.

11. The medical imaging system according to claim 1 or 2, wherein: The machine learning algorithm is any one of the following: an isolation forest algorithm, a k-nearest neighbor algorithm, and a one-class support vector machine algorithm.

12. The medical imaging system according to claim 1 or 2, wherein: Predetermined preprocessing routines include: - calculating a spatially averaged MRF vector for a predetermined region around each voxel in the set of voxels; - appending the spatially averaged MRF vector to the preprocessed MRF vector before computing the outlier map.

13. The medical imaging system according to claim 1 or 2, wherein: Predetermined preprocessing routines include: - calculating a spatial gradient MRF vector for each voxel in the set of voxels; and - appending the spatial gradient MRF vector to the preprocessed MRF vector before computing the outlier map.

14. A computer program product comprising machine executable instructions (120) for execution by a processor (106), wherein: Execution of the machine executable instructions causes the processor to: - receiving (200) MRF magnetic resonance data (122) of a region of interest (310) acquired according to an MRF magnetic resonance imaging protocol; - reconstructing (202) an MRF vector (124) for each voxel of a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol; - calculating a pre-processed MRF vector for each voxel in the set of voxels (126) by applying a predetermined pre-processing routine to the MRF vector for each voxel, wherein the predetermined pre-processing routine includes normalizing the pre-processed MRF vector for each voxel; - calculating an outlier map (130) for the set of voxels by assigning outlier scores (800) to the pre-processed MRF vectors using a machine learning algorithm; and - receiving a segmentation of the set of voxels, wherein the segmentation identifies a voxel type for each voxel in the set of voxels, and wherein the outlier score is assigned at least in part by using the voxel type as an input to the machine learning algorithm.

15. A method of operating a medical imaging system, wherein: The method comprises: - receiving (200) MRF magnetic resonance data (122) of a region of interest (310) acquired according to an MRF magnetic resonance imaging protocol; - reconstructing (202) an MRF vector (124) for each voxel of a set of voxels describing the region of interest using the MRF magnetic resonance data according to the MRF magnetic resonance imaging protocol; - calculating a pre-processed MRF vector for each voxel in the set of voxels (126) by applying a predetermined pre-processing routine to the MRF vector for each voxel, wherein the predetermined pre-processing routine includes normalizing the pre-processed MRF vector for each voxel; - calculating an outlier map (130) for the set of voxels by assigning outlier scores (800) to the pre-processed MRF vectors using a machine learning algorithm; and - receiving a segmentation of the set of voxels, wherein the segmentation identifies a voxel type for each voxel in the set of voxels, and wherein the outlier score is assigned at least in part by using the voxel type as an input to the machine learning algorithm.

Images