Detection of disconnected MR coils and identification of misplaced coils.

Abstract

A system and associated method for supporting MR imaging includes logic that receives measurements from RF sensors (SS1-8) positionable outside the bore of an MR imaging device, and processes the measurements to establish whether i) there is at least one surface RF coil that is not electrically coupled to the circuitry of the MR imaging device, and / or ii) there is at least one surface RF coil located on or at the patient table of the MR imaging device.

Description

[Technical Field]

[0001] The present invention relates to a system for supporting MR imaging, a patient table for MR imaging, a method for supporting X-ray images, a method for training a machine learning model for use in supporting MR imaging, a computer program element, and a computer-readable medium. [Background technology]

[0002] Magnetic resonance imaging (MRI) is a well-known imaging modality because it allows for the non-invasive acquisition of images of internal anatomical structures and / or pathophysiology. Unlike X-ray imaging modalities, MRI uses non-ionizing radiation.

[0003] The same MRI scanner includes a cylindrical bore in which the patient lies on a patient table during imaging. The space within the bore is narrow. The scanner includes magnets and coils built around the bore. Both are operated together to acquire data that can be processed into slice images to reveal details of internal anatomical structures, for example, for diagnostic purposes.

[0004] Some imaging protocols may require the use of additional mobile coils (also called surface coils) to achieve better image quality in a specific region of interest. One or more surface coils are placed close to the patient in the region of interest. The surface coils include lead wires with connectors that are electrically connected to the MR scanner system.

[0005] Occasionally, unconnected surface coils are observed to be inadvertently left with the patient on the patient bed. This can degrade image quality, interrupt the imaging workflow, cause damage to the surface coil, or even injure the patient. [Overview of the project] [Problems that the invention aims to solve]

[0006] Therefore, improvements to MRI may be necessary, especially to make MRI safer.

[0007] The object of the present invention is solved by the subject matter of the independent claims, and further embodiments are incorporated into the dependent claims. It should be noted that the embodiments of the present invention described below apply equally to patient tables for MR imaging, methods for supporting X-ray imaging, methods for training machine learning models for use in supporting MR imaging, computer program elements, and computer-readable media. [Means for solving the problem]

[0008] According to a first aspect of the present invention, a system for supporting MR imaging,

[0009] A system is provided comprising a logic unit configured to receive measurements from one or more RF (radio frequency) sensors that can be positioned outside the bore of an MR imaging device, and to process the measurements to i) determine whether there is at least one surface RF coil that is not electrically coupled to the circuitry of the MR imaging device, and / or ii) locate at least one surface RF coil on or within the patient table of the MR imaging device, wherein the measurement is based on a response signal provided by at least one surface coil. Processing of the measurement by the logic unit is preferably performed while the patient and sensors are outside (or at least partially or entirely outside) the bore. The measurement is preferably a (frequency) spectral measurement.

[0010] The electrical coupling of surface coils can be achieved through connections via leads, cables, etc. The system can detect whether or not the surface coils are connected in this manner.

[0011] In an embodiment, the logic unit is based on a machine learning ("ML") module that is pre-trained. Non-ML embodiments are also envisioned. In such non-ML embodiments, the logic unit may be configured to adapt a measured value to a reference signal ("signature") stored in advance. In both ML and non-ML embodiments, the measured value may be in the time domain or the frequency domain. The reference signal stored in advance may be that of a surface coil that is connected or not connected. In an embodiment, the logic unit is further configured to receive and process an image acquired by a camera of the patient table.

[0012] In an embodiment, the system is operable before the introduction of the patient into the bore of the MRI imaging device, and the patient is introduced only if it is established that at least one (preferably all) surface coils are i) connected and / or ii) positioned at a location that matches the intended imaging protocol. In an embodiment, the measured value is an S21 measured value.

[0013] In an embodiment, one or more RF sensors are communicatively coupled to the logic unit in an MRI-compliant manner.

[0014] In an embodiment, the coupling is arranged as an optical link.

[0015] In an embodiment, the system includes an output port indicating any one or more of i) a non-connected surface RF coil is detected, ii) the position of a (connected or non-connected) surface RF coil on or above the patient table, and iii) whether the position matches the intended imaging protocol.

[0016] In another aspect, an MRI image patient table is provided that includes one or more RF sensors that can be used in the above-described system according to any of the embodiments.

[0017] In some embodiments, one or more sensor RFs are integrated into the surface of the table to make it more comfortable for the patient or to facilitate cleaning the patient table.

[0018] In another embodiment, an imaging device is provided which includes a system as described in any of the embodiments or models described above, and i) a patient table and / or an MR imaging device.

[0019] In another embodiment, a method for supporting MR imaging is provided, and this method is The steps include receiving measurements from one or more RF sensors that can be positioned outside the bore of the MR imaging device, A method comprising the steps of processing measurements to determine whether i) there is at least one surface RF coil that is not electrically coupled to the circuitry of the MR imaging device, and / or ii) to locate at least one surface RF coil on or to the patient table of the MR imaging device, wherein the measurements are based on response signals provided by the at least one surface coil.

[0020] In another aspect, a method is provided for training a machine learning model based on training data in order to obtain a machine learning module to be pre-trained.

[0021] This specification proposes performing radio-RF measurements to test for unconnected coils on the patient table before the patient enters the bore. This is proposed as a standalone check, or preferably as a redundant check in addition to camera-based evaluation. The patient may proceed into the MRI bore only if both checks are negative, i.e., neither results in the detection of unconnected coils.

[0022] In addition, or alternatively, it is proposed to perform measurements using a connected coil, allowing for at least rough positioning of the coil on the patient table before the patient enters the bore. In this case as well, this is proposed as a standalone check, or preferably as a redundant check in addition to camera-based assessment. In the case of a misplaced coil, the patient is asked to reposition the coil, or, ultimately, the staff is warned to provide assistance.

[0023] The proposed system and method facilitate interrupted MR workflows and reduce surface coil damage and patient injury due to heat buildup in disconnected surface coils. The RF sensor, also referred to herein as the “Guardian Coil,” can be based on existing MR coil technology. No additional technologies such as RFID are required. The proposed system is not vendor-independent. Existing MR setups can be modified. After appropriate training, machine learning embodiments enable robust detection with a low error rate, further improving workflow, image throughput, and patient safety. To avoid workflow interruptions, detection of disconnected or misplaced surface coils may be beneficially used in medical imaging systems.

[0024] In another embodiment, a computer program element is provided that, when executed by at least one processing unit, is adapted to cause the processing unit to perform the method according to any one of the embodiments described above.

[0025] In yet another embodiment, a computer-readable medium is provided on which program elements are stored. definition

[0026] The term "user" refers to a person who operates the imaging device or monitors the imaging procedure, such as a healthcare professional or another person. In other words, the user is generally not the patient.

[0027] The term “patient” is used herein to refer to a person being imaged. The use of the term “human” does not exclude animals being imaged for veterinary purposes other than those being imaged. Furthermore, images of inanimate objects, such as baggage in security checks or products in non-destructive testing, are not excluded herein.

[0028] Generally, a “machine learning component” is a computerized configuration that implements an ML algorithm configured to perform a task. In an ML algorithm, the task's performance is measurably improved by the training experience. The training experience can include providing the instrument with more (newer and appropriately modified) training data. Performance can be objectively test-measured when supplying the system with test data. Performance can be defined in terms of a specific error rate that should be achieved with given test data. See, for example, TM Mitchell, “Machine Learning”, page 2, section 1.1, McGraw-Hill, 1997.

[0029] Herein, exemplary embodiments of the present invention will be described with reference to the following drawings, which are not to a fixed scale unless otherwise specified. [Brief explanation of the drawing]

[0030] [Figure 1A] This shows an imaging device for magnetic resonance imaging (CT). [Figure 1B] This shows an imaging device for magnetic resonance imaging (CT). [Figure 2] An exemplary surface coil that can be used in MRI is shown. [Figure 3] The image shows a patient tabletop including a radio frequency sensor. [Figure 4] A block diagram of the SYS guardian system envisioned in this specification, according to one embodiment, is shown. [Figure 5A] This shows a spectrogram that can function as input for a parental control system. [Figure 5B] This shows a spectrogram that can function as input for a parental control system. [Figure 6] A flowchart illustrating methods for assisting with MR imaging is shown. [Figure 7] A block diagram of a machine learning model that may be used in the embodiments described herein is shown. [Figure 8] Figure 9 shows a training system for training a machine learning model, and a method for training a machine learning model. [Modes for carrying out the invention]

[0031] Referring first to Figure 1A, this shows a schematic block diagram of an imaging apparatus AR assumed herein in an embodiment. The imaging apparatus AR includes an MRI (magnetic resonance CT) apparatus IA configured to acquire MR medical images of a human or animal patient PAT. The MRI imaging apparatus may be combined with a hybrid imaging or therapeutic system. A hybrid imaging system includes an MRI PET (positron emission tomography) / SPECT (single-photon emission computed tomography) system or a CT (computed tomography) PET / SPECT system. A hybrid therapeutic imaging system further includes an MRI-LINAC (linear accelerator) or an MRI hyperthemia system, or other systems.

[0032] A magnetic resonance CT scanner IA includes a housing HS that defines a generally cylindrical scanner bore BR in which the associated imaging target PAT is placed. Inside the housing HS is a main magnetic field coil MG. The main magnetic field coil can generally be a solenoid configuration to generate a main magnetic field B0 oriented along the Z direction parallel to the central axis of the scanner bore BR. The main magnetic field coil is typically a superconducting coil located within a cryogenic shroud, but a resistive main magnet can also be used. Superconducting magnets are preferred, especially at higher magnetic field strengths, and therefore higher frequencies. Figure 1A shows a closed cylindrical design, which is not a requirement in this specification, as open-type MRI scanners with U-shaped magnets are also included herein. However, the principles predicted herein are particularly beneficial for closed-type MRI scanners because the patient space is restricted, as will become clear below.

[0033] The housing HS may also house or support a magnetic field gradient coil GC for selectively generating a magnetic field gradient along the Z direction, along in-plane directions transverse to the Z direction (such as along the Cartesian X and Y directions), or along other selected directions.

[0034] The housing HS also houses or supports a radio frequency head or body coil MC (referred to herein as the main coil) for selectively exciting and / or detecting magnetic resonances. While birdcage coils are common below 128 MHz, in addition to birdcage coils, other coils, such as transverse electromagnetic (TEM) coils, phased coil arrays, or any other type of radio frequency coil, can be used as volume transmit coils. The housing HS typically includes a decorative inner liner that defines the scanner bore BR.

[0035] The MRI apparatus may further include a patient bed or table PT. The patient PAT is expected to lie on the table PT during imaging. While the patient PAT lies on the table PT outside the bore BR of the imaging apparatus, it is prepared for imaging during the preparation phase. Once preparation is complete and imaging begins, the patient table PT is operated to introduce the patient PAT into the bore along the image axis Z, in the longitudinal axis direction of the cylindrical bore BR.

[0036] More specifically, the table PT includes a payload surface PS for receiving the patient. The payload surface has appropriate dimensions and shape. For example, the payload surface PS may be rectangular in shape with dimensions suitable for receiving an average-sized human patient, such as approximately 2 x 1 meter, with its longest side aligned with direction Z. The payload surface is usually flat, and there may be a more comfortable mattress component placed between the patient and the payload surface. The payload surface on which the patient is placed is supported from below by the undercarriage UC. The undercarriage may include rollers or be positioned on rails to facilitate moving the bed so that the patient moves onto and away from the imaging device AI. The undercarriage UC may house compartments, pockets, shelves, etc., for staff to place equipment.

[0037] A user interface (not shown) integrated into the table can enable the table's operation, or the table can be operated remotely by an image controller integrated circuit, a computerized device described in more detail below. The undercarriage UC can also house actuators that enable the table's operation and adjust the table's height relative to the ground. In addition, or alternatively, the actuators may enable the payload surface to be advanced horizontally along the Z-axis relative to the undercarriage with the patient on it, so that the patient can be introduced into the bore to initiate imaging.

[0038] Table adjustment, particularly the advancement of the payload surface, may be influenced by manual actuators such as a set of levers or handwheels. However, preferably, the table PT is motorized to include appropriate electromechanical actuators, such as stepping motors, which can be remotely controlled by the bed's local user interface or via a controller IC.

[0039] During use, after introducing the patient PAT into the bore BR, the main magnetic field coil MG preferably generates a main magnetic field B0 in the Z direction of at least 3.0T, more preferably greater than 3.0T, for example, 7.0T or more. Values ​​less than 3T, such as 1T or 1.5T, are also conceivable.

[0040] In general, the operation of an MRI includes an image control stage integrated circuit and an image reconstruction stage (RECON). Both stages may be managed by one or more appropriately configured computing devices. In the control stage, a configured computing device (referred to herein as the image controller), such as a general-purpose or dedicated computing device appropriately programmed and interfaced, coordinates the aforementioned reception and transmission of high-frequency signals through the main coil (MC), the control of the inclined coil (GC), and the activation of the magnet (MG), or other actions. The image reconstruction stage is managed by a computing device that executes the aforementioned image reconstruction software or algorithms to process the stored k-space values ​​into an MR image in the image region for display, storage, or further processing. The control stage and the reconstruction stage may be integrated into a single computing system or device.

[0041] First, looking at the image controller IC, it includes a magnetic resonance CT controller that operates a magnet controller to selectively energize the magnetic gradient coil GC, and operates a radio frequency transmitter coupled to one or more radio frequency coils MC to selectively energize the main, fixed, radio frequency coil or coil MC located inside the hole, causing it to emit a transmit signal and interact with B0 and tissue. By selectively operating the magnetic gradient coil GC and one or more radio frequency coils MC, magnetic resonances are generated and spatially encoded in at least a portion of the selected region of interest of the imaging target PAT. In this way, an image plane including at least a portion of the ROI can be selected.

[0042] The magnetic resonance CT controller operates a radio frequency receiver coupled to one or more radio frequency coils (MCs) to receive image data in the formation of magnetic resonance k-space data samples, which are stored in k-space memory. The image data can be stored in any suitable format.

[0043] The image reconstruction stage RECON can reconstruct k-space samples into a reconstructed image that includes at least a portion of the region of interest of the object being imaged, by applying an appropriate reconstruction algorithm such as a Fourier transform reconstruction algorithm.

[0044] The reconstructed images are stored in image memory, displayed on the operator console user interface, stored in non-volatile memory, transmitted via a local intranet or the internet, or otherwise viewed, stored, manipulated, etc. The operator console user interface can also enable radiologists, technicians, or other operators of the magnetic resonance CT scanner IA to communicate with the magnetic resonance CT controller IC to select, modify, and execute magnetic resonance CT sequences.

[0045] As shown in Figure 1A, the image reconstructor RECON and the image controller IC are coupled to the aforementioned hardware and / or software components of the imager IA via a suitable communication network CN. The communication network CN may be wired or wireless and enables the flow of data to and from the imaging device IA. Control signals may be passed from the image controller integrated circuit to the imaging device I. Received signals picked up by the main coil MC may be passed to the image controller integrated circuit via the network CN, and / or to a storage device in k-space memory, from where they may be passed to the image reconstructor RECON for image processing. It is also conceivable that data from the coil MC may be directly supplied to the image reconstructor RECON as a data feed for processing.

[0046] Some imaging protocols may require the use of an additional RF coil, sometimes called a surface coil (SC). Unlike the main RF coil (MC) or inclined coil (GC) described above, the surface coil (SC) has coil components located above and below it, and is separated from it. While coilworks like the main inclined coil (MC) or inclined coil (GC) are stationary components fixed and installed around the bore (BR) within the imager housing (IA), the surface coil (SC) is a movable object that can be removed from or placed within the bore along with the patient (PAT). More specifically, surface coils may be used to provide more detailed field imaging, improved intra-slice spatial resolution, and a higher signal-to-noise ratio for specific regions of interest to be imaged. Such regions of interest ("ROI") may include certain biological structures such as the head, chest, wrist, knee, or foot. Where required, surface coils (SC) may be used alongside, in addition to, or instead of, the main coil (MC).

[0047] Figure 2 shows examples of two such surface coils SC1,2. These are called surface coils SC because they are designed to be attached to the patient's body surface. A surface coil SC has a form factor corresponding to a main body portion that is attached and remains there during imaging. Attached via straps or other fastening means, the surface coil is positioned near the ROI being imaged.

[0048] Each surface coil SC1, 2 may include an RF loop LP made of conductive material, coupled to a circuit CT within the body of the surface coil. The RF loop and circuit CS are housed in a housing, such as a plastic casing. Optionally, the surface coils SC1, 2 may include a skin-friendly surface lining to make the wearing experience more comfortable for the patient. The surface coils SC1, 2 may be positioned and fixed, for example, on the patient's abdomen in a supine position, or positioned beneath the patient between the payload surface and the patient. The operation of the surface SC is substantially the same as that of the main coil MC, i.e., picking up resonant signals emitted by atomic nuclei in patent tissue. The surface coil SC circuit CT may be arranged on one or more circuit boards. The surface coil SC circuit CT may include a (coil-shaped) loop LP with a capacitor for tuning. The coil is coupled to a preamplifier via an impedance (inductor) and one or more diodes. Active detuning is performed by a diode in series with a (preferably variable) inductor, which, together with the capacitor, is capable of resonating at the Larmor frequency during use. The circuit CT may include active and / or passive circuit components to protect against high-current damage. The passive circuit components may include one or more fuses. The circuit structure CT described is illustrative, and other circuit designs are conceivable. A preamplifier provides an output signal representing the nuclear response picked up by the surface coil SC during use. The preamplifier output signal can then be digitized and stored in k-space memory for processing into an MR image. The circuit may further include a memory component, such as an EPROM or other, where the coil's ID can be stored for ID checking by the imaging device IA. The memory component may further store a sanity hardware check routine, program, which is executed at power-up when the surface coil is connected to the scanner IA's connection circuit SPC.

[0049] The surface coils SC1,2 may include leads L, e.g., electrical plugs, terminated at connectors CNT in wired embodiments. The surface coil connector CNT can be coupled to a connection circuit SPC, as shown in Figure 1A. The connection circuit SPC may include a socket SK into which the plug or connector CNT is received to establish a connection through which electrical signals pass. The circuit SPC, which communicates with the leads L and connector CNT, allows the surface coil SC loop LP and its internal circuit CT to be coupled to the signal chain of the communication network CN. Thus, signals picked up by the surface coil SC during MRI can be processed into an MR image by passing from the coil SC through a controller integrated circuit and / or k-space memory or image reconstructor RECON. The connection circuit SPC may be located as the front end of the imaging device adjacent to the entrance to the bore BR, or it may be integrated into the bed PT, chassis UC, or similar, from which data is transferred via cablework or other suitable data connections and coupled to the communication network CN.

[0050] A surface coil SC that is coupled to the signal chain by such a wired connection via a connecting circuit SPC and lead L / connector CNT device is referred to herein as “connected.” Otherwise, this state is referred to herein as “not connected / or disconnected.” Two or more surface coils SC1, 2 may be used together for a given patient PAT during imaging. When connected, the picked-up signal is transmitted to a communication network, and operating power is supplied to the coil to operate active circuit components and protect them from current damage.

[0051] Surface coils SC1 and SC2 are preferably arranged as receiving-only coils. Depending on the imaging protocol, surface coils SC may be used together with the main coil. Alternatively, the main coil MC may be used without surface coils SC1 and SC2, or only one or more surface coils without the main coil MC may be used to pick up the nuclear response signal.

[0052] Occasionally, one or more unconnected surface coils SC3 may be left on the surface PS of the table PT along with the patient PAT. This can lead to undesirable situations when imaging is initiated. Such a situation is schematically illustrated in the side view of Figure 1B, where a “bad” surface coil SC3 remains unconnected to the patient on the surface PS. The unconnected coil SC3 may not be immediately apparent to the patient or staff, as it may be obscured from view by the drape DR, or by other coverings worn by the patient or covering the patient during imaging. Some lead wires L1, L2 of properly connected coils SC1, SC2 may be visible, emerging from under the covering, and one may believe that everything is in good order according to protocol, but there may still be an unconnected coil remaining, perhaps tucked under the mattress, pillow, or drape, or otherwise escaping visual attention, along with its lead wire LS. There may be designated racks, such as the aforementioned compartments / pockets in the bed's PT undercarriage or similar, for arranging unused surface coils, but the daily practice in busy imaging departments does not always allow for such tensions. On-site staff with various MRI scanners may "borrow" a surface coil assigned to one scanner and use it with another. While this may have been intentional, borrowed coils are not always returned. Therefore, a quick visual inspection of the surface coil rack, even when all spaces are filled, does not necessarily mean that no surface coils remain somewhere on the bed.

[0053] The condition of a faulty coil is undesirable because it can induce uncontrolled high currents in the faulty (surface) coil SC3 during the operation of the magnet MG in the imaging procedure. The faulty coil SC3 can become hot, be damaged, or cause burns to the patient, or damage other equipment in contact with the hot, unconnected coil SC3. The main coil MC, unintentionally placed too close, can cause high currents, which in turn lead to undesirable overheating. Such overheating is not likely to occur under normal use because the coil SC is an active circuit component that can operate when connected, especially when connected. Passive components without active components may not be able to prevent overheating on their own.

[0054] To prevent unintended rogue surface coils SC from remaining with the patient during imaging, a guardian system SYS is envisioned herein that can operate to prevent the patient from being introduced into the bore BR if such rogue surface coils SC are actually present. In other words, the proposed guardian system is configured to detect the presence of one or more unconnected surface coils SC3 on the table PT in close proximity to the patient. Advantageously, this allows for the avoidance of image interruption. Often, such rogue coils SC3 can be found due to a specific signal signature that is eventually picked up in the signal chain, but only when the imaging procedure is already underway. Such image abortion may be necessary because the bore BR does not leave much room for the patient to rotate, move, or hold the rogue surface coil, for example, by passing it to a staff outside the bore. Also, the patient may be sedated or too alert to help. Consequently, in the absence of the guardian system SYS, if such rogue coils SC3 are left behind, the imaging procedure must be aborted, and the patient must be removed from the bore for the rogue coils SC3 to be recovered. It will be understood that this could have a serious impact on the workflow and reduce image throughput. Image throughput is a factor that needs to be considered, especially for busy imaging departments in large urban hospital facilities, and is already being pushed to its limits.

[0055] In addition to detecting the presence of unconnected, faulty coils SC, or alternatively, the Guardian System SYS envisioned herein may be configured to verify, even if connected, that surface coils are properly spatially positioned relative to the patient, particularly to the ROI. As previously mentioned, depending on the imaging protocol used, surface coils must be positioned in or around the ROI being imaged. Otherwise, image quality may be significantly degraded, requiring re-imaging, which also slows down the image throughput. This function is sometimes referred to as a positioning / locating function, in contrast to the faulty coil detection function described above.

[0056] Both functionalities, presence detection and localization, can be performed jointly as preferred, or one can perform the other without the other as needed. The user can choose to operate either one or both of the functions. Each of the above system functions may be performed on its own as a simple check, or both may be combined into a more elaborate check for localization and detection of the presence of a tampered coil.

[0057] The guardian system SYS may include one or more auxiliary sensor AUXs above the patient bed, such as an optical camera placed in the examination room in front of the imaging device IA, to image the patient / bed before introducing the patient into the bore. This is schematically shown in Figure 1A. Two or more such cameras may be used and do not necessarily have to be positioned for planar image capture. Some sensor AUXs may be configured to capture side views and / or front or rear images. While this specification primarily refers to optical cameras, this is not a requirement of this specification. The auxiliary sensor AUXs may not necessarily be optical cameras and may instead be depth-sensing cameras, near-infrared (IR) cameras, or any other suitable sensor type such as LIDAR, or even more. Two or more types of sensors, in particular some or all of the above types of sensors may be used in combination.

[0058] Location and / or detection of the presence of an irregular surface coil (hereinafter simply referred to as "presence detection") by the Guardian System SYS can be based on images acquired by the sensor AUXS. The images represent the patient on the outer diameter of the table PT.

[0059] In another preferred embodiment, localization and / or presence detection is achieved by evaluating RE signals received from one, preferably multiple, sets of RF sensors SS. These sensors are structurally similar to surface coils in that they include RF circuits coupled to RF coils, as described above in Figure 2. However, as part of a guardian system SYS, these RF sensors SS are not used for medical imaging but are used in addition to one or more surface coils SC (used for imaging purposes) and the coilwork of the imaging device, main coil MC and inclined coil GC, within the imaging device housing around the bore BR. The RF sensors SS may be configured as transceivers, but this is not necessary as some or all may be receive-only and some may be transmit-only. The RF sensors SS may be used in addition to the auxiliary sensors AUXS described above in a two-stage redundant inspection setup, or they may be used in place of such auxiliary sensors AUXS in a simpler one-stage inspection setup.

[0060] These additional sets of one or more RF coils, referred to herein as guardian coils SS, are positioned outside the bore BR when in use. They are configured to pick up search signals (as described more fully below) that can be processed by the system SYS to determine whether correct localization is achieved and / or whether there are any faulty coils that are not actually connected on the table surface, before the patient is moved into the bore BR.

[0061] Therefore, presence detection and localization can be based on the analysis of optical images acquired by an additional camera AUX and / or RF search signals received by an RF sensor SS. The patient can only be introduced into the bore if both analyses reveal that the correct configuration is present and / or that no unconnected coils exist. If at least one of the above analyses fails, the guardian system SYS can intercept and block requests, for example, remotely issued by the image controller, to introduce the patient into the bore. Thus, a redundant inspection system is established by the system to ensure proper localization and / or absence of misplaced coils with a low error rate. In particular, false negatives or inverted positives can be avoided. The two functions of localization and misplaced coil detection are described more fully below in Figure 4. The guardian system SYS does not necessarily have to actively intervene in the control signal chain; instead, it may simply issue an all-clear or warning signal.

[0062] Next, referring to Figure 3, this illustrates one embodiment of how additional guardian coils SS, SS1 to SS8, may be positioned near the patient. In one embodiment, the guardian coils SS are positioned on the surface PS of a table. The coils may be arranged in a regular layout along direction Y in Figure 4, as shown in the plan view. In one embodiment, a 2x4 layout of eight such guardian sensors SS1 to SS8 is used, but any other layout, regular or irregular, of any number of such coils is also assumed herein.

[0063] It will be understood that if more additional sensors SS1 to SS8 are placed per unit area of ​​the surface PS and properly densely packed, position detection can operate more accurately.

[0064] The Guardian SS coils may be placed on the bare payload surface and may be fixed by adhesive or other means. In one example, the mattress may be placed on top, sandwiching the coil SS between the surface PS and any mattress. For even greater comfort, the payload surface PF may include appropriate recesses into which the surface coils are integrated so as to be essentially coplanar with the upper surface of the payload surface PS. The Guardian coils SS1 to 8 are partially or completely integrated with the payload surface PS. Typically, since a completely RF-transparent material is used for the payload surface PS, the Guardian coils SS1 to 8 may be completely embedded and submerged to the depth of the surface so that their presence cannot be revealed by visual inspection. Instead of placing them in or on the surface PS, the Guardian coils SS1 to 8 can be integrated within the undercarriage UC, provided that sufficient proximity to the patient is ensured. Sufficient proximity to the signal frequencies assumed herein is, for example, in the range of up to 10 cm. It may also be possible to place the Guardian coils SS raised above the surface, on a framework around a table, or within a raised edge, or mounted on any of the same.

[0065] In the example, the guardian coils SS are circular with a diameter of approximately 2 cm. They are preferably placed in the transmit chain of the MR scanner IA as transceivers that can be coupled to the communication network CN. Some or each guardian coil may have its own transmit / receive channel. Alternatively, fewer such channels (e.g., a single channel) may be used, i.e., switched by a multiplexer to connect each of the transmit coils in sequence, while the other channels are switched to receive mode. In transmit mode, the RF search signal described above is transmitted. The RF search signal interacts with one or more surface coils, including the rogue coil, to produce a response signal, which is then detected in receive mode by one or more guardian coils. The response RE signal, the presence of the rogue coil and / or the correct placement (whether the coils are connected) of the coils may be processed as described in Figure 4 below. Preferably, rogue coil detection is performed first. Once the rogue coil is removed, the correct placement of the remaining (connected) coils is queryed. Alternatively, both presence detection and placement checks are performed at the same time.

[0066] To avoid interference with MR signal reception at the surface coil SC for MR imaging, a local battery-powered, electrically isolated, but still wirelessly linked and software-controlled, protective coil SS may also be used. This is because the RE signal transmitted and received by the guardian coil SS for detecting the misaligned surface coil SC3 does not need to be phase-locked to the MR receiving chain, as it is required for RF transmission for MR imaging. Full B1 clarity can be provided to the patient table PT using an optical or wireless link operating at a much higher frequency than the RF used for imaging.

[0067] Refer to the block diagram in Figure 4 to explain the operation of the parental control system SYS in more detail.

[0068] The input signal is received at input port IN. The input signal is either an optical or other image acquired by the patient's camera system AUX on table PT in front of the RE signal bore, provided by a guardian sensor positioned on the patient's table as described above in Figure 3.

[0069] In response to the two functions described above, the system may include a position component LC and / or a presence detection component PC.

[0070] The position component LC processes the signals received from the guardian sensor and / or the images received from the camera AUX into the probability of the correct position of connected or unconnected coils. The correct position is the position defined by the imaging protocol used. For example, in a wrist image, the ROI is the wrist, which is where at least one surface coil SC should be positioned.

[0071] In the embodiments, optical images acquired by the auxiliary camera system AUXS are image-processed by the position component LC to determine whether the surface coil CS position is correct according to the imaging protocol. Alternatively or additionally, the RE signal received by the guardian coil, a typical signal fingerprint of the surface coil SC, is processed by the position component LC. A machine learning module MLM' may also be used. These embodiments are described in more detail below.

[0072] In the redundant inspection embodiment, the position component LC processes two input data types, namely the optical image signal and the high-frequency signal from the guardian sensor, separately and using different techniques, to reach their respective probabilities p1 and p2. Each probability represents the probability of the exact position of the surface coil based on processing each data type. Both probabilities can be thresholded by threshold TH against their respective minimum error probability thresholds. The logic AND coupler Σ1 checks whether both conditions by the two thresholds are met, and only if they are met does a confirmation output signal generate at the output port OUT. Otherwise, a warning signal is issued, as will be described in more detail below.

[0073] The presence detection component PC for establishing the presence of the forgotten or incorrect surface coil SC3 can similarly act on two data types, or only one of them. The presence detection component of the logic unit CL processes either the patient's optical image on the outer bore of the table, or the RE signal received by the guardian coil SS, or both data types in the redundant examination setup described above for the position component LC.

[0074] Here again, preferably, redundancy checks are performed, and each data type is processed separately by different techniques to reveal two respective probabilities p1, p2 regarding the presence of a faulty coil. Both probabilities are thresholded by their respective thresholds TH for their respective error probability thresholds. For the positional component LC, the error probability thresholds may or may not be different. The result of the thresholding by thresholds TH is coupled by a logical "AND" operator Σ1 to an output signal indicating, as previously described, whether the patient's PAT is introduced into the lumen BR via the bed PT and imaged. In addition, or instead, the output signal may be used to actively control the imaging procedure in cooperation with an image controller, such as in an autonomous imaging setup.

[0075] If only a simple non-redundancy check is required for either the presence component PC or the location component LC, based on only one of two data types, then only one threshold TH is needed, and the logical AND coupler Σ1 is not required. The output signal provided at output OUT is based on this single thresholding, as shown by the dashed line in Figure 4.

[0076] Location and presence checks may also be used as standalone components, with one without the other. Furthermore, while redundancy checks by each component LC may be useful, PCs are not required in all embodiments, and each component may act not as a redundancy checker, but as a simple checker, on only one data type, i.e., one of the images acquired by the RE signal camera AUX, received by the protective coil.

[0077] However, redundant testing based on both data types is advantageous because relying solely on the optical image provided by the sensor AUX is prone to errors. A direct line of sight to identify the faulty coil SC3 is not always present. As mentioned above in Figure 1B, the faulty coil SC3 may be obstructed by a strap, cover, the patient themselves, etc. The additional redundant channel provided by the response RE signal received by the guardian coil SS complements the optical image-based processing because a line of sight is not required. Furthermore, relying solely on RF signal processing can lead to boundary predictions in time, in which case the additional check provided by the camera's AUX image can help reach a final decision. Redundant testing can reduce primary and secondary error rates (false positives or negatives) compared to single-channel testing.

[0078] In any of the above embodiments, whether both position check LC and presence check C are performed, or only one of the two is performed without the other, a control signal is issued directly to an electromagnetic actuator in the patient table TB via the output port OUT, and once the checks are passed, the patient can be automatically introduced into the bore. Such direct automatic control is useful in autonomous imaging setups, where it is the imaging system and / or the patient itself, rather than staff, performing the support tasks. Such tasks may include preparing the patient and surface coil SC for imaging, or introducing / removing the patient into or from the bore, and positioning / removing the coil SC from the system IA / bed PT. Alternatively, in some non-autonomous setups, there is no such automatic introduction; instead, an all-clear signal is issued via the output interface OU. The all-clear signal drives a transducer, such as a loudspeaker, to sound with the corresponding sound signal, or a lamp to emit a light signal, or a display controller to control a display device DD in the control room to inform the user. It is then the user who initiates the introduction of the patient into the bore. If the check fails, an alert signal is issued. The request to introduce the patient into the bore may be blocked. Such all clear signals to provide auditory, visual, or tactile confirmation / warning may also be used in autonomous settings to inform the patient.

[0079] In embodiments where both location check LC and presence check PC are performed, these can be further combined by a second-stage logical "and" operator (∧)Σ2, and the patient can only be introduced into the bore if there are no unconnected surface coils and the position check indicates correct placement. Otherwise, a warning signal is issued and / or, in an automated setting, the introduction is stopped / prohibited or permitted to proceed.

[0080] The control logic unit CL may be implemented by a processing unit PU. The processing unit may be integrated into the computing system at the imaging control integrated circuit stage.

[0081] To describe in more detail the operation of the presence detection component PC based on the RE signal received by the guardian coil SS, it is preferably (but not necessarily) implemented by a machine learning module MLM. The machine learning module includes a pre-trained learning model M. The machine learning model is pre-trained on the training data, as will be described in more detail below.

[0082] Figures 5A and 5B show an example of such an RF input signal in the form of a spectrogram that can be obtained by Fourier analysis of the RE signal received and sampled in the guardian coil SS in the time domain. Specifically, the f-domain data shows frequency along the horizontal axis versus intensity along the vertical axis. The spectrogram is shown for a baseline BL with spikes caused by the resonance response of the unconnected coil SC3.

[0083] To detect unconnected coils, the guardian coils may operate as follows: Each or some of the guardian coils SS are then switched to transmit using a sweep over the frequency range where resonance of the unconnected MR RF receiving coil is expected. Typically, this range is the Larmor frequency + / - 10 MHz. Then, S21 measurements are performed. Specifically, while one of the guardian coils SS is transmitting (in transmit mode), at least one, preferably all, of the other guardian coils SS are held in receive mode to acquire spectra, and at least some of them will show resonance if any unconnected coil is present, as shown in Figures A, B. Additional such spectral measurements may be acquired, here by switching some or all guardians to transmitters one by one, with the earlier transmitting coils SS switching back to receive mode, etc. Figures 5A, B show such S21 measurements using two transceivers, one in transmit mode and one in receive mode. A typical resonance spike indicates the presence of an unconnected MR coil. Figure 5A shows the spectrum of a FlexM coil, and Figure 5B shows the spectral response of a forward coil. Such spectra are typical of MR coils, and consequently, the presence of a false coil can be detected by software-based processing as proposed herein. It has been found that the spectral signature changes in a characteristic manner depending on the coil type, and therefore, the coil type can also be arbitrarily established herein. For comparison, the flat baseline trace BL in Figures 5A and 5B is obtained without inventing any (unconnected) coils.

[0084] Some or all of the spectra measured in the S21 procedure may be processed by a machine learning module (MLM) previously trained on the data to be labeled. Training data may be collected in a training session, where one or more different such unconnected surface test coils are placed at different positions on a patient table, and a spectrogram is recorded by manipulating a protective coil and switching between transmit and receive modes to pick up responses induced from the unconnected test coils on the table. Similar spectrograms can be recorded using connected test surface coils. A human expert can easily determine the presence or absence of a surface coil from the experimental setup or from the spectrogram itself. Thus, the training data can be labeled and then provided to the machine learning module as training input. The parameters of the machine learning model are adapted in an optimization procedure, as will be described in more detail below.

[0085] The task of classifying the spectrograms of coil responses, such as those recorded in one or more S21 measurements as shown in Figure 5, is well-suited to machine learning-based processing because predicting the presence of a false coil based on simple thresholding appears difficult. While not excluded herein in non-ML embodiments, "classical" analytical signal processing methods tend to be more error-prone due to noise and other factors. Unlike analytical processing, machine learning approaches do not require specific assumptions about dependencies and can therefore be trained more robustly on previously acquired training data. In an ML optimization procedure, an initial set of parameters for the ML model is adapted in light of the training data. These adapted parameters can be thought of as "encoding" the relationship between the spectrogram and, on the other hand, the presence of a false coil.

[0086] The architecture of ML model M can be selected as an artificial neural network having one or more fully connected hidden layers, with a combinatorial layer (such as a softmax layer) as the output layer. The output layer couples feature maps from previous layers to a probability distribution regarding the presence of rogue coils. When processing a spectrogram as shown in Figure 5, a hybrid architecture including both convolutional and fully convolutional layers can be used. Alternatively, a purely fully connected neural network is used because these types of NNs have been found to perform better when processing non-image data such as the spectrogram in Figure 5. Here again, the coupling layer in the terminal output layer couples the previous feature maps to a desired probability distribution across two labels lb1, lb2 (lb1 = "rogue coil present", lb2 = "rogue coil not present").

[0087] Moving on to processing the optical images, a machine learning module based on a neural network, for example, can be used. The machine learning module can process the optical images received from the camera to determine whether or not an unconnected coil SC3 is present. To facilitate this classification task, surface coils can be equipped with a specific marker CM that allows for easier identification of the presence of an unconnected coil. Coils on lead cables may also be considered for coil recognition. While presence detection based on optical images taken from above the table PT is preferred, as shown in Figure 1A, further images of the patient on the table can be obtained using one or more additional cameras positioned on the sides around the table. When optical image data is processed by machine learning, a neural network architecture with convolutional layers is preferably used, and this type of NN model has been found to yield good results for processing image type data for the task of recognizing whether or not an object is present in an image.

[0088] Instead of using an ML approach, the existing component PC can be processed using classical image processing techniques such as segmentation, shape detection, and intensity thresholding. For example, using segmentation-based image processing, the rogue coil SC3 can be detected as a characteristic "tadpole" shape formed by the main body and the attached lead wires. The leads terminate within the connector CNT. Therefore, if unconnected, this condition can be found by segmenting the "tadpole" structure. If the tails (leads) terminate in a flared structure, the connector is visible, and therefore the surface coil is not connected. Model-based segmentation (MBS) can be used, for example, prioritizing shape and using the tadpole-shaped planar image installation area.

[0089] In an image-based embodiment of the positioning component LC, optical images from the camera AUXS can be used to derive the patient's position on the table and to check whether the surface coil SC is approximately positioned correctly on the patient for the planned scan.

[0090] The processing of high-frequency signals and / or optical images by the presence detection component PC may be performed by a separate machine learning component MLM, or it may be performed integrally by a single machine learning module.

[0091] Now, moving on to the implementation of the localization component LC, this can be done with or without machine learning. In the ML embodiment, the same architecture as described above can be used for the presence component PC. The ML embodiment is described in more detail in Figures 7 to 9 below.

[0092] As described above in the S21 measurement protocol for the present component PC, the response spectrum readings from the distributed guardian sensor SS on the table are obtained by sequentially switching the guardian coil SS between transmission and reception.

[0093] In non-ML-based embodiments, the intensity of spectral readings from some or each of the sensor SS can be compared. Sensor SS closer to the surface coil SC can respond with higher intensity readings. Readings collected from some or all of the SS coils can be compared. The guardian coil with the highest intensity reading is then positioned closest to the surface coil. Since the positions of the sensor SS are known, the positions of the surface coils within the error margin corresponding to the distance of the guardian coils in each layout are also known. Since the imaging protocol is known and therefore an ROI, the correct positions of the coils can be correlated with the guardian sensor layout. Based on this correlation, the precise surface coil SC placement can be established. Signal intensity or intensity at a given target frequency, such as the Larmor frequency, can be used, but any other frequency in the spectrum can also be used.

[0094] Optionally, instead of using the guardian coil spectral response, or in addition to it, the spectral responses of one or more connected surface coils are analyzed for signal intensity. Typically (but not necessarily), up to 30, for example, up to 32, or more or less, multiple receiving coil elements i are present in the surface MR coil SC. Each coil element generates a received signal with some intensity si. The element closest to the currently transmitting guardian coil SS provides the strongest signal. Considering the signal intensity at the target frequency received by some or all of the guardian coils allows for at least rough localization of the surface MR coil, as is known in the art.

[0095] Optionally, as an additional check, the coil load of some or all connected surface coils is measured, as is known in the art. Coils that are sufficiently close to the patient experience a higher load, i.e., a wider resonance width, while coils located further away from the patient experience a lower load or no load at all, and therefore a very small resonance width. The resonance width is determined by S11 measurement using coil connections. This information allows for further verification that the surface coils required for a particular scan are actually positioned on the patient.

[0096] A non-ML embodiment of the presence detection component PC based on RF measurements is described in more detail here. The guardian coil SS is known to be able to record the clear and distinct signal signatures of the connected surface coil SC. The individual RF functions F1-N of some or each surface coil SC can be stored in the database MEM during the preparation phase. However, if an unconnected surface coil SC3 is coupled with a coil present in the system, or if no such surface coil is present, the signal function is measured to be different from the baseline reading.

[0097] In one embodiment, an unconnected faulty coil SC3 can be detected using a guardian coil. The system begins to sequentially switch all or all of the connected coils SC on and off, and the guardian coil measures the individual signal functions F, for example, in the time domain.

[0098] The measured function registered in the guardian coil is compared to the signal function signature stored in the database MEM. The amplitude parameter (magnitude) A can be adjusted to fit the measured function F to some or each of the stored functions F = A1*F1. If the fitting result is within the error margin (e.g., <10%), the logic unit CL concludes that there are no unconnected coils in the system. Otherwise, if there is no probability of achieving a fit, there may be an unconnected faulty coil SC3 in the system. The user or autonomous system may be notified to stop the workflow and is advised to remove the unnecessary coil or connect it to the interface SPC.

[0099] Another telltale code for the unconnected coil SC3 shows their typical frequency response spectrum. This is TIFF0007865327000001.tif53. In another embodiment, these spectra are stored in the database MEM. The guardian coil SS is used for detecting the unconnected coil SC3. Used to measure TIFF0007865327000002.tif75. The frequency response to be measured. TIFF0007865327000003.tif65 is, Intensity parameters such as TIFF0007865327000004.tif617 By adjusting TIFF0007865327000005.tif64, it is fitted to a pre-stored reference spectrum. If two or more unconnected coils are present on the patient table, the individual coil parameters... A curve fitting is performed to match TIFF0007865327000006.tif779. Furthermore, combinations of unconnected coil spectra can be directly stored in the database.

[0100] Here, we refer to Figure 6, which shows a flowchart supporting MR imaging. It will be understood that the methods and steps described can be used to implement the Guardian System SYS described above. However, the methods described herein are not necessarily tied to the system architecture disclosed above and may be understood as teachings in themselves.

[0101] In step S610, preferably, a dual redundant check is initiated for the position and / or arrangement of unconnected surface coils in the MR setup.

[0102] In step S620, an optional check is performed in the connection circuit SPC to determine whether or not the surface coil is connected to the system.

[0103] In step S630, input data is collected.

[0104] Specifically, in step S630a, an optical image is acquired by a camera on a table where the patient is placed outside the bore. The image is preferably a plan view from above the table.

[0105] In addition to or instead of step S630a, in step S630b, a guardian sensor in the form of a radio frequency coil as described above is activated to transmit a signal and receive a response signal from an unnoticed, unconnected surface coil left on the patient table.

[0106] Next, the signals obtained in steps S630 B and / or C are processed in step S640 by a preferably pre-trained machine learning module, and in step S650, the probability of the presence of an unconnected coil and / or its correct position relative to the patient is output.

[0107] The probabilities may be thresholded in the redundancy check. If both probabilities exceed their respective error thresholds, an all-clear output signal is issued in step S670. The output all-clear signal can initiate or control the patient introduction into the imager bore, or it may simply indicate that the user has received the result of the check, and the user initiates the patient introduction.

[0108] Instead, if at least one of the threshold tests fails, it is flagged as a warning signal in step S660. The warning signal may be indicated to the user or patient by sound, light, display on a display device, etc. The warning signal may be transmitted to the user by pager, text message, email, etc. In autonomous imaging settings, the table movement can be actively controlled using alert signals. For example, the request for table movement to introduce the patient into the bore is actively interrupted by interference with the MR imager's control circuit / system.

[0109] The process flow then returns to optional step S620 and / or step S630, where new input data is collected after a certain period of time for the patient or staff to make corrections.

[0110] Once the response time has elapsed, the process flow continues and can be repeated as described above until both threshold tests are met and the patient can finally be introduced into the bore. The process flow can loop so that there may be two or more unnoticed faulty coils on the table.

[0111] As described above, a redundancy check step is preferred, but each of the above checks concerning the optical image and / or radio frequency reading from the guardian coil may, if necessary, be performed separately and without each other in a simpler non-redundancy check. The RE signal picked up by the guardian coil can be used to establish the presence and correct placement of the faulty coil.

[0112] Herein, we refer to Figure 7, which shows a schematic diagram of a neural network type ML model that may be used in the embodiment. As previously stated, the model may include one or more fully connected layers (as opposed to layers in a convolutional network). The neural network is configured for classification tasks. Consistent with what has been stated above, the classification tasks assumed herein include classifying a spectrogram obtained by a guardian coil as to whether or not it represents the presence of an unconnected surface coil (referred herein to as the “location classification task”). Another classification task is whether or not, given an image from an optical camera, it represents the presence of an unconnected surface coil. Yet another classification task is whether or not a surface coil is correctly positioned based on RF readings from a guardian coil or based on an image acquired by a camera (hereinafter referred to as the “presence detection classification task”).

[0113] For any of the classification tasks described above, model M can be trained by a computerized training system TS, as will be explained more fully below in Figure 8. During training, the training system TS adapts an initial set of (model) parameters θ of model M. In the context of neural network models, parameters are sometimes referred to as weights herein. Training data is collected in the data collection phase. Instead of a data collection phase, training data may be generated by simulation. Thus, after the data collection phase, two further processing phases can be defined with respect to the machine learning model M: a training phase and a deployment (or inference) phase.

[0114] In the training phase, prior to the deployment phase, the model is trained by adapting its parameters based on the training data. Once trained, the model can be used in the deployment phase to perform classification tasks. Training may be a one-off operation or it may be repeated after new training data becomes available.

[0115] The machine learning model M to be trained may be stored in one (or more) computer memories MEM. The model M to be trained may be deployed as a machine learning component MLM on a computing device PU, preferably an integrated image control IC computing system of an imager IA. Preferably, to achieve good throughput, the computing device PU includes one or more processors (CPUs) that support parallel computing, such as a multi-core design. In one embodiment, a GPU (or GPU) (graphical processing unit) is used.

[0116] Continuing to refer to Figure 7, this shows a neural network M in a feedforward architecture. Network M comprises multiple computing nodes arranged cascaded in a hierarchy L1–LN, with data flow from left to right, and therefore from hierarchy to hierarchy. Recurrent networks are not excluded herein.

[0117] In the deployment, input data x is applied to the input layer IL, where "x" is the spectrogram RF reading of the guardian coil or the patient bed / patient camera image. The input data x then propagates through a series of hidden layers L1–LN (only two are shown, but there may be simply one or more) and then appears in the output layer OL as the estimated output M(x). For this purpose, the output M(x) is the classification result with respect to the probability for any of the classification tasks described above. A similar information flow is applied to the training data when it is applied to the model M during training. There may be separate models M' (not shown) for localization classification tasks and presence detection classification tasks, each trained separately on their respective training data. Alternatively, there may be a single model for both tasks.

[0118] The model network M can be said to have a deep architecture because it has multiple hidden layers. In a feedforward network, "depth" is the number of hidden layers between the input layer IL and the output layer OL, while in a time network, depth is the number of hidden layers multiplied by the number of paths.

[0119] The layers of a network, specifically the input x and output M(x), as well as the inputs and outputs between hidden layers (referred to herein as feature maps), can be represented as matrices ("tensors") of two or more dimensions for computational and memory allocation efficiency. The network output M(x) can be represented as a two-dimensional vector. The vector can be normalized, so that two entries are added to 1 to represent probabilities as a result of the classification task described above. The output layer OL can combine or combine the previous feature maps with the vector to represent probabilities. For example, the output layer OL can be configured as a soft max layer to provide values ​​within unit intervals to represent probabilities. The data applied as input x to the input layer can be presented as a vector, such as a spectrogram, or as a matrix when an image is used.

[0120] The hidden layer L1–LN may contain one or more convolutional layers, but is preferably exclusively a fully connected layer. The number of layers is at least one, for example, 2 to 5, or any other number. The number may be a two-digit number.

[0121] Fully connected layers (FCs) are beneficial in non-image regression tasks, as is the case herein for processing spectrograms. However, in embodiments processing camera AUX images, one or more convolutional layers, or even just such convolutional layers, can still be used as hidden layers. Fully connected layers are distinguished from convolutional layers in that entries in the output feature map of the fully connected layer are combinations of all nodes received by that layer as input from the previous layer. In other words, for each feature map entry, the convolutional layer applies only to a subset of the inputs received from the previous layer.

[0122] It should be understood that the Model M shown in Figure 7 is merely one embodiment and does not limit the present disclosure. Other neural network architectures are also assumed herein, having more, fewer, or different functions than those described herein, such as pooling layers or dropout layers. Furthermore, the ML model M assumed herein is not necessarily of the neural network type. Support vector machines can be used as an alternative, for example. Other statistical classification models and methodologies based on sampling from training data are also assumed herein, such as alternative embodiments such as Bayesian networks, or random fields such as Markov random fields, and others such as decision trees and random forests.

[0123] Next, refer to Figure 8, which shows the training system TS for training a machine learning mod. The training system TS is configured for the learning parameters. The trainable parameters can include weights such as a machine learning model M in a neural network as discussed in Figure 7, or other neural network type models, or actually non-neural network type models, ML, or statistics.

[0124] The training data consists of k pairs of data (x k , y k ) includes the training data for each pair, and the training input data x k And related target y k This includes the following. Therefore, training data is organized in k pairs, particularly for supervised learning plans as assumed herein. However, it should be noted that unsupervised learning plans are not excluded herein.

[0125] Training input data x k Each pair may include a spectrogram (see, for example, Figure 5 above) or a camera image. The associated target y kOr "Grand Truth" is a label given by human experts or simply follows an experimental setup for training data collection. The label indicates the type of training example, such as whether the training input data represents the signature of a surface coil not connected or indicates the correct placement. The pair (x k , y k ) is recorded in the training data memory during the training data collection stage described above. Again, each element x k , y k can be presented as a vector or matrix. The training data, as described above, can be obtained by experiment (e.g., by placing a sample surface coil and acquiring test RF readings / spectrograms or test images) or by simulation.

[0126] This model is of the classifier type and attempts to classify x k into the label y k as will be described in more detail. In the training phase, an initial set of weights is pre-input into the architecture of a machine learning model M, such as the artificial neural network ("NN") shown in FIG. 7. The weights θ of the model M represent the parameterized M θ , and the goal of the training system TS is to optimize and thus adapt the parameter θ based on the training data (x k , y k ) pairs in the training batch. In other words, learning can be mathematically formulated as an optimization scheme where the cost function F is minimized, although a dual formulation that maximizes a utility function can be used instead.

[0127] Here, assuming the paradigm of the cost function F, this measures the error that occurs between the aggregated residuals, i.e., the data estimated by the neural network model NN and the target for some or all of the training data pairs k. TIFF0007865327000007.tif1191TIFF0007865327000008.tif11103

[0128] In equation (1a) and below, the function M() represents the result of the model NN applied to the input x. During training, the training input data x of the training pair k This is propagated through the network M that is initialized. Specifically, the k-th pair of training inputs x k The data is received at input IL, passes through the model, and then at output OL, the output training data M θ It will be output as (x).

[0129] The cost function F is the actual training output M generated by the model M. θ (x k ) and the desired goal y k The difference between (also known herein as the residual) is based on an appropriate similarity measure. The symbol ||·,· || in (1a) specifies a general operator for the cost function to measure the distance in a space of appropriate choice between the label and the prediction, given a current set of parameters θ (symbolized by an extremely long stroke to the right) evaluated over some or all training pairs indexed by "i". For binary classification, which is primarily assumed herein, the cost function F measure can be formulated as cross-entropy or Kalbach-Leibler divergence (KLD) distance / loss, or any other function appropriate for the classification task. For classification tasks such as binary classification, which is primarily assumed herein, the space is a space of probability distributions whose distance can be measured by cross-entropy or KLD or other measures. In this setting, the results provided by the model are considered to be samples drawn from an unknown probability distribution. The prediction is thought of as the probability for each label, and the cross-entropy minimization (1b) is the maximum likelihood estimate. Maximum likelihood estimates are also assumed herein.

[0130] In some cases, equation (1a), such as for the cross-entropy function or KLD, may be written in additive form as shown in equation (1b), and thus the evaluation "·|k" over all pairs of training data pairs "k" in (1a) is the sum.

[0131] Output training data M(x k ) is the target y related to the applied input training image data x. k This is an estimate of the output M(x). Generally speaking, this output M(x k ) and the associated target y of the kth pair currently being considered k An error exists between the two. Then, an optimization method such as inverse / forward propagation or other gradient-based methods is used to consider the pair (x k , y k Alternatively, the parameters θ of the model M can be adapted to reduce the residuals of a subset of training pairs from the complete training dataset.

[0132] The model parameter θ is the current (x k ,y k After one or more iterations in the first inner loop, which are updated by the updater UP for pairs in the training batch, the training system TS enters the second outer loop, where the next training data pair x k+1 , y k+1 Alternatively, the next batch of training pairs is processed accordingly. The structure of the updater UP depends on the optimization scheme used. For example, an internal loop managed by the updater UP may be implemented by one or more forward and backward passes in a forward / backpropagation algorithm. To improve the objective function while adapting the parameters, aggregated, e.g., summed residuals of all training pairs are considered up to the current pair. Aggregated residues can be formed by constructing the objective function F as the sum of squared residues, such as equation (2), of some or all of the residues considered for the pair. Other algebraic combinations are also possible instead of the sum of squares.

[0133] Optionally, one or more batch normalization operators ("BN", not shown) can be used. Batch normalization operators may be integrated into the model M, for example, by being coupled to one or more convolution operators within a layer. BN operators allow for the mitigation of slope effects, which are experienced during the learning phase of a slope-based learning algorithm in the learning phase of the model M, with a gradual decrease in the magnitude of the slope over repeated forward and reverse passes. Batch normalization operators BN can be used in training, but can also be used in deployment.

[0134] The training system shown in Figure 8 can be considered for all learning plans, particularly supervised plans. Unsupervised learning plans may also be assumed herein in alternative embodiments. A GPU may be used to implement the training system TS.

[0135] A fully trained machine learning module M can be stored in one or more memory locations (MEM) or a database. Training is performed per user to obtain a different pre-trained model for each user, as described above.

[0136] Figure 9 is a flowchart showing a method for training a machine learning model as described in any one of the embodiments described above.

[0137] Appropriate training data must be collected during the monitoring phase as described above. Preferably, a supervised learning plan is assumed herein, but this is not required as an unsupervised learning setup is also assumed herein.

[0138] In supervised learning, training data consists of pairs of appropriate data items, each pair containing training input data and its associated target training output data. Specifically, training data pairs (x k , y k As described above, this data is collected during the monitoring phase and stored in training data memory (not shown).

[0139] Continuing to refer to Figure 9, in step S910, the training data is paired (x k ,y k It is received in the form of ). Each pair is defined in Figure 8 as the training input x k and related target y k , x k Includes. In step S920, training input x k This is applied to the initialized machine learning model NN to generate the training output.

[0140] Related target y k The deviation or residual of the training output M(x) from is quantified by the cost function F. One or more parameters of the model are adapted in step S930, and in one or more iterations within the inner loop, to improve the cost function. For example, model parameters are fitted to reduce the residual measured by the cost function. The parameters include, in particular, the weights W of the neural network type mode, convolutional or fully connected, or partially both.

[0141] Next, the training method returns to step S910 in the outer loop, where the next pair of training data is supplied. In step S920, the model parameters are fitted so that the aggregated residues of all pairs considered in the batch are reduced, in particular, to be minimized. The cost function quantifies the aggregated residues. Reverse propagation or a similar gradient-based technique may be used in the inner loop.

[0142] More generally, the parameters of the model M are adjusted to improve an objective function F, which is either a cost function or a utility function. In embodiments, the cost function is configured to measure aggregated residues. In embodiments, this is done by summing over all or some residues for all pairs of residues where aggregated residues are considered. The method can be implemented on one or more general-purpose processing units TS, preferably having a processor capable of parallel processing to speed up training.

[0143] The components of the training system TS may be implemented as one or more software modules running on one or more processing computing systems, preferably having high-performance processors such as GPUs. Training is preferably not performed by the imager IA's processing unit PU, but may be performed externally by one or more more powerful computing systems such as servers. Once trained, the model may be loaded into the data storage MEM of one or more computing device PUs implementing the guardian system SYS.

[0144] While this specification has primarily referred to imaging of human patients in the medical field, the principles disclosed herein may be used instead to image animals. In addition, imaging of objects, baggage inspection or non-destructive material testing, and surface coils are also conceivable where necessary.

[0145] The system SYS method or components may run as one or more software modules on one or more general-purpose processing unit PUs, such as workstations associated with imagers IA, or on server computers associated with a group of imagers.

[0146] Alternatively, the System SYS method or some or all of its components may be located in hardware such as a field-programmable gate array (FPGA) or hardwired IC chip, application-specific integrated circuit (ASIC), or a properly programmed microcontroller or microprocessor, which is integrated into the imaging device IA. In further embodiments, the System SYS may be implemented partially in software and partially in hardware.

[0147] The methods or steps of different components of the system SYS may be implemented on a single data processing unit (PU). Alternatively, some or all components or method steps may be implemented on different processing units (PUs), possibly located remotely in a distributed architecture, and connected via an appropriate communication network such as a cloud setup or client-server setup.

[0148] One or more features described herein may be configured or implemented as circuits encoded in a computer-readable medium, or using and / or in combination thereof. Circuits may include discrete and / or integrated circuits, systems on a chip (SOC), and combinations thereof, machines, computer systems, processors and memory, and computer programs.

[0149] In another exemplary embodiment of the present invention, a computer program or computer program element is provided, characterized in that it is adapted to perform the method steps of a method according to one of the above-described embodiments on a suitable system.

[0150] Accordingly, the computer program elements may be stored in a computer unit, which may be part of an embodiment of the present invention. This computing unit may be adapted to perform or trigger the execution of the steps of the method described above. Furthermore, it may be adapted to operate the components of the apparatus described above. The computing unit may be adapted to operate automatically and / or to execute user sequences. The computer program may be loaded into the working memory of a data processor. Accordingly, the data processor may be equipped to perform the method of the present invention.

[0151] This exemplary embodiment of the present invention encompasses both computer programs that use the present invention from the outset and computer programs that, by means of updating, transform an existing program into a program that uses the present invention.

[0152] Furthermore, computer program elements can provide all the steps necessary to satisfy the procedure of the exemplary embodiment of the procedure described above.

[0153] According to a further exemplary embodiment of the present invention, a computer-readable medium such as a CD-ROM is presented, having computer program elements stored thereon, which are described in the previous section.

[0154] Computer programs may be stored and / or distributed on suitable media (in particular, but not necessarily, non-temporary media) such as optical storage media or solid-state media supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0155] However, computer programs may also be presented over a network such as the World Wide Web and downloaded from such a network into the working memory of a data processor. According to a further exemplary embodiment of the present invention, a medium is provided for making a computer program element available for download, and this computer program element is configured to perform the method according to one of the aforementioned embodiments of the present invention.

[0156] It should be noted that embodiments of the present invention are described with reference to different subject matter. In particular, some embodiments are described with reference to method-type claims, and other embodiments are described with reference to apparatus-type claims. However, unless otherwise notified, those skilled in the art will find that any combination of features belonging to one type of subject matter, as well as any combination of features relating to different subject matter, are gathered from the above and below descriptions and are deemed to be disclosed in this application. However, all features can be combined to provide a synergistic effect greater than the simple sum of the features.

[0157] Although the present invention is illustrated and described in detail in the drawings and the foregoing description, such illustrations and descriptions should be considered illustrative or exemplary and not limiting. The present invention is not limited to the embodiments disclosed. Other variations of the disclosed embodiments can be understood and practiced by those skilled in the art in carrying out the claimed invention, based on a study of the drawings, disclosure and dependent claims.

[0158] In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite articles “a” or “an” do not exclude plurals. A single processor or other unit may satisfy the functions of several items enumerated in the claims. The mere fact that certain means are referenced in different dependent claims does not imply that combinations of these means cannot be used advantageously. No reference numeral in the claims should be construed as limiting in scope.

Claims

1. A system to support MR imaging, A logic unit configured to receive measurements from a plurality of RF sensors located outside the bore of an MR imaging device during use of the system, wherein the logic unit processes the measurements to determine whether i) there is at least one surface RF coil that is not electrically coupled to the circuitry of the MR imaging device, and / or ii) to locate at least one surface RF coil on or in the patient table of the MR imaging device, wherein the measurements are based on a response signal provided by an RF loop forming the at least one surface RF coil, the RF loop being operable in the imaging operation of the MR imaging device, and the response signal being provided in response to a transmit signal that can be transmitted by at least one transmitter, the logic unit It has, Of the aforementioned RF sensors, one RF sensor operates in transmit mode, and the other RF sensors operate in receive mode. The aforementioned measurement is the S21 measurement. system.

2. The system according to claim 1, wherein the logic unit is based on a machine learning module to be trained.

3. The system according to claim 1, wherein the logic unit is based on fitting the measured values ​​to one or more reference signals that are stored in advance.

4. The system according to any one of claims 1 to 3, wherein the logic unit is further configured to receive and process images acquired by the camera on the patient table.

5. The system according to any one of claims 1 to 4, wherein the patient is introduced only when it is established that the at least one surface RF coil is i) connected to and / or ii) located in a position consistent with the intended imaging protocol, and the bore of the MR imaging apparatus is operational before patient introduction via the or one patient table.

6. The system according to any one of claims 1 to 5, comprising: i) detection of an unconnected surface RF coil; ii) the location of one or more of the surface RF coils on or thereof of the patient table; and iii) an output port indicating whether or not the location matches an intended imaging protocol.

7. The system according to any one of claims 1 to 6, wherein the plurality of sensors are integrated into the surface of the patient table or one patient table.

8. An imaging apparatus comprising the system according to any one of claims 1 to 7 and the MR imaging apparatus.

9. A method for supporting MR imaging, When the above method is performed, the steps include receiving measurement values ​​from a plurality of RF sensors located outside the bore of the MR imaging device, i) determining whether there is at least one surface RF coil that is not electrically coupled to the circuit of the MR imaging device, and / or (ii) processing the measurement to locate at least one surface RF coil on or in the patient table of the MR imaging device, wherein the measurement is based on a response signal provided by an RF loop forming the at least one surface RF coil, the RF loop being operable in the imaging operation of the MR imaging device, and the response signal being provided in response to a transmit signal that can be transmitted by at least one transmitter. It has, Of the aforementioned RF sensors, one RF sensor operates in transmit mode, and the other RF sensors operate in receive mode. The aforementioned measurement is the S21 measurement. method.

10. A method for training a machine learning model using training data to obtain a trained machine learning module according to any one of claims 2, 4 to 7.

11. A computer program element adapted to cause at least one processing unit to perform the method described in claim 9 or 10 when executed by the processing unit.

12. At least one computer-readable medium storing the program element described in claim 11, or the machine learning module described in claim 2.

Images