Autonomous magnetic resonance scan for a given medical test

Abstract

For autonomous MR scanning of a given medical test, a simplified MR scanner can be used without or with little input or control by a technical expert, e.g., by a physician, radiologist, or person trained in MR scanner operation. The MR scanner autonomously positions, scans, checks quality, analyzes, and / or outputs answers to diagnostic questions with or without MR images. Artificial intelligence-based scan analysis allows continuous or instant changes to the scan configuration to acquire the data desired to answer the diagnostic question. By using a simplified MR scanner, both the positioning of the patient relative to the MR scanner and the localization of the scan by the MR scanner are jointly solved. Sensors can sense the patient at the scan position, with reduced radio frequency requirements allowing a more open bore.

Description

[0001] This patent document claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 308,508, filed on February 10, 2022, with the filing date under 35 U.S.SC §119(e), which is incorporated herein by reference. Background Technology

[0002] Magnetic resonance (MR) imaging equipment is currently designed to accommodate manual workflows, where a scanner first reconstructs images and then presents these images to a radiologist for interpretation. This comes with several drawbacks. First, because reconstruction is a deterministic function of the acquired data, it involves suppression information. While some suppression may be beneficial (e.g., in the presence of noise), the criteria for deciding what to retain and what to suppress are subjective and loosely correlated with the clinical decisions being addressed. Second, much of the current internal workings of MR imaging are designed to simplify the reconstruction task rather than optimize the information rate of the scan. Finally, these simplifying assumptions also place high demands on MR hardware. The k-space sampling used for reconstruction requires linear gradients, and spatially constant image contrast requires uniform B0 and B1 fields. These hardware requirements result in more expensive MR imaging equipment. Summary of the Invention

[0003] By way of introduction, the preferred embodiments described below include methods, systems, instructions, and non-transitory computer-readable media for autonomous MR scanning for a given medical test. The simplified MR scanner can be used with little or no input or control from a technical expert (e.g., a physician, radiologist, or person trained in MR scanner operation). The MR scanner autonomously locates, scans, examines quality, analyzes, and / or outputs answers to diagnostic questions, with or without MR images. Artificial intelligence-based scan analysis allows for continuous or real-time changes to the scan configuration to obtain the desired data to answer diagnostic questions. By using the simplified MR scanner, both patient positioning relative to the MR scanner and the localization of the scan by the MR scanner are jointly addressed. Sensors can sense the patient at the scan location, where reduced radio frequency requirements allow for a more open aperture.

[0004] In a first aspect, a data analysis method for magnetic resonance (MR) scanning is provided. An MR scanner scans a patient during an MR examination using a first configuration based on medical tests, thereby generating first raw data. A first machine learning model analyzes the first raw data, resulting in a change in the first configuration. The MR scanner is controlled based on a second configuration derived from the change in the first configuration. The MR scanner then scans the patient again using the second configuration for the medical tests, as part of the same MR examination, thereby generating second raw data. Diagnostic outputs of the medical tests from the MR examination are output based on the first and second raw data.

[0005] In one embodiment, the MR scan has a non-uniform main magnetic field, a non-uniform first pulse, and / or a nonlinear gradient.

[0006] In another embodiment, the analysis includes determining a diagnostic value for the medical test and an uncertainty in the value using a first machine learning model. The control includes changing from the first configuration to a second configuration if the uncertainty exceeds a threshold. For example, backpropagation based on the first machine learning model specifies the next measurement. Based on missing information identified by backpropagation, parameters of the second configuration are set for the next measurement.

[0007] As another embodiment, the analysis includes determining that insufficient information has been collected by using an MR scan with a first configuration, and the control includes determining a second configuration to obtain information.

[0008] In yet another embodiment, the analysis includes inputting first raw data and a first configuration into a first machine learning model, and having the first machine learning model output a diagnostic output as the answer to a diagnostic question.

[0009] Further embodiments include: locating the patient relative to the MR scanner, and localizing the MR scan by solving a second machine learning model that solves for both localization and localization; and checking the quality of the first raw data using the second machine learning model, wherein the control includes performing controls to improve the quality without human input.

[0010] In yet another embodiment, control includes control via a second machine learning model, which is a reinforcement learning model. In some embodiments, the MR scan using the first configuration continues until the reinforcement learning model generates a change.

[0011] In a second aspect, a magnetic resonance (MR) system includes an MR scanner configured by controlled settings to scan an area of ​​a patient. The scan provides scan data. A patient support for the patient is movable relative to the MR scanner. Sensors are configured to sense the patient on the patient support. A processor is configured to jointly perform (1) positioning the patient by movement of the patient support and / or the MR scanner, and (2) localizing the scan of the area by the MR scanner.

[0012] In one embodiment, the processor includes a machine learning detector configured to detect the localization and localization based on the output of the sensor and the input of the scan data, and includes a machine learning actor configured to move the patient support and change the localization when the detected localization and localization are incorrect.

[0013] As another embodiment, the processor includes a machine learning model configured to jointly perform localization and localization as a single solution for both localization and localization.

[0014] In other embodiments, the processor is further configured to analyze scan data once the patient is located and the scan is localized. The analysis is performed using a machine learning model configured to output a diagnosis in response to input of the scan data. The processor is further configured to modify the settings based on the analysis.

[0015] In another embodiment, the processor is further configured to monitor artifacts in the scan data and change the settings based on the artifact level.

[0016] In yet another embodiment, the sensor is a camera, and the MR scanner is an open-aperture MR scanner, wherein the camera's field of view extends to the patient while being localized within the open aperture of the open-aperture MR scanner.

[0017] In a third aspect, a method for autonomous magnetic resonance (MR) scanning is provided. The patient is positioned within an MR scanner. An area of ​​the patient is localized by the MR scanner for scanning that area. The MR scanner scans the patient, thereby generating scan data. The scan data is analyzed, and a diagnostic answer is generated from the scan data. Positioning, localization, MR scanning, analysis, and generation are performed without human control.

[0018] According to one embodiment, the analysis includes checking the quality of the scanned data by a machine learning model. The processor modifies the configuration of the MR scanner based on the quality change. In one example, the check includes inspection by a machine learning model acting as a quality generator, which identifies deviations from the quality regardless of the type of artifact. In another example, the check includes inspection targeting the type of artifact, where the change includes modifying the configuration to reduce that type of artifact.

[0019] This invention is defined by the following claims, and nothing in this section should be construed as limiting those claims. Other aspects and advantages of the invention are discussed below in conjunction with preferred embodiments and may subsequently be claimed independently or in combination. Attached Figure Description

[0020] Figure 1 This is a block diagram of an embodiment of an MR system for autonomous medical imaging;

[0021] Figure 2 This is a flowchart illustrating one embodiment of a method for autonomous MR scanning;

[0022] Figure 3 This is a flowchart illustrating one embodiment of a method for joint localization and localization;

[0023] Figure 4 This is a flowchart illustration of one embodiment of a method for analyzing configuration changes during an MR inspection;

[0024] Figure 5 This is a flowchart illustrating one embodiment of a method for quality inspection of motion artifacts as part of analysis used to control scanning; and

[0025] Figure 6 This is a flowchart illustrating an embodiment of a method for quality inspection of radio frequency interference artifacts as part of analysis used to control scanning. Detailed Implementation

[0026] Data acquired by MR scanners is leveraged by enabling end-to-end data artificial intelligence (AI) analysis from raw signals to clinical findings. This end-to-end approach improves performance and standardizes it. The ability to directly analyze the acquired data allows for relaxed hardware requirements and / or enables shorter protocols. Relaxed hardware constraints reduce costs by lowering design requirements, allowing for more accessible MR devices. Devices with a relaxed design may not provide images or useful images, but can provide analysis as a way to understand the captured measurements. Image reconstruction steps can be skipped. Fully automated, low-cost MR devices can be provided where image reconstruction is not the goal. First, MR devices are not built to produce images. Focusing on scanning for data analysis also enables faster protocols on high-field (i.e., image-quality) scanners in autonomous operation.

[0027] In autonomous MR scanning, the scan can be adapted during a given patient examination. The control of the MR device is adapted to perform a given medical test, rather than following a scanning protocol used for imaging. The scanning protocol is adapted to allow the test to be performed with minimal supervision from a technical expert. AI analyzes the acquired data and reconfigures the MR scanner to extract the most useful information for answering diagnostic questions.

[0028] AI can be used to control patient localization and region of interest (ROI) localization to further assist autonomous scanning. Localization and localization can be addressed jointly rather than sequentially, for example, in cases where sensors other than MR can sense the patient even when the patient is inside an MR scanner.

[0029] AI can be used for quality monitoring during scanning to further assist autonomous scanning. Continuous quality monitoring occurs during scanning, allowing for analysis to control the scan without human input during inspection.

[0030] In one embodiment, for a simplified MR scanner, MR is automated, thereby reducing the technical burden on the operator and enabling widespread deployment in settings where skilled technicians are unavailable. Automation provides a self-service workflow where patients perform their own scans under guidance from the device. This automation is performed using protocol optimizations for the scan, with the aim of reducing the development costs of MR scanners. A more simplified MR scanner can be used by employing AI-based configuration during scanning. This is particularly important for low-cost hardware that breaks down standard MR assumptions and the resulting sequence development heuristics. Real-time control allows the scanner to adapt to the acquired data online. This potentially enables more efficient customized protocols than a "one-size-fits-all" approach.

[0031] Figure 1An embodiment of a system for performing MR scanning by an MR scanner is shown. The MR scanner 90 scans a given patient 140. The MR scanner 90 performs... Figure 2 The MR scanner 90 operates at least partially autonomously, including combined patient localization and positioning, adaptive configuration during patient scanning, and / or quality check analysis and adaptation. The MR scanner 90 may have reduced requirements, such as non-uniform fields and / or nonlinear gradients. The MR scanner 90 may be capable of image reconstruction, but may use faster or different scans for analysis directly derived from the scan data.

[0032] The MR scanner 90 includes a main field magnet 100, a gradient coil 110, a whole-body coil 120, a local coil 130, and a patient support (e.g., a bed) 150. The system includes the MR scanner 90, a processor 160, a memory 170, and a display 180. One or more sensors 190 may be provided separate from the coils 110, 120, and 130. Additional, different, or fewer components may be provided for the MR scanner 90 and / or the system. For example, the local coil 130 or the whole-body coil 120 may not be used. In another example, the processor 160, memory 170, and display 180 are provided without the coils 100-120 and the patient support 150, such as a workstation that operates on scan data stored in the memory 170. In yet another example, the processor 160, memory 170, and / or display 180 are part of the MR scanner 90.

[0033] The MR scanner 90 is configured, through controlled settings, to scan an area of ​​the patient 140. The scan provides scan data in the scan domain. The MR scanner 90 scans the patient 140 to provide raw measurements (measurements in a possible nonlinear frequency domain). In cases where hardware defects cause the spatial encoding to be non-Fourier, the measured response is referred to as raw data or scan data, rather than k-space data. In cases where the spatial encoding is Fourier, the scan or raw data can be k-space data. For the scan, the master field magnet 100 creates a static fundamental magnetic field B0 within the patient 140, positioned on the patient support 150. Gradient coil 110 generates a positioning-dependent magnetic field gradient superimposed on the static magnetic field. The gradient coil 110 generates positioning-dependent and shim-adjusted magnetic field gradients in three orthogonal directions and generates a sequence of magnetic field pulses. The whole-body coil 120 and / or local coil 130 receive radio frequency (RF) transmitted pulses, thereby generating magnetic field pulses (B1) that spin-rotate protons in the imaging area of ​​the patient 140.

[0034] In response to an applied RF pulse signal, as the whole-body coil 120 and / or the local coil 130 return to an equilibrium position established by the static and gradient magnetic fields, the whole-body coil 120 and / or the local coil 130 receive MR signals, i.e., signals from excited protons within the body. The MR signals are detected and processed by a detector, thereby providing an MR dataset of raw data. The raw storage array of the memory 170 stores the corresponding individual measurements that form the MR dataset.

[0035] The MR scanner 90 is configured by the processor 160 to perform scanning. Any of various scanner controls can be set, such as k-space coordinates, TR, TE, flip angle, pulse envelope, carrier frequency, timing, duration, and / or raw transmitted pulses. With or without user input or changes, the protocol can establish settings at least initially for a particular scan. Any level of versatility can be provided for the settings, such as abstractions of the actual variables used for specific hardware. The memory 170 stores the configuration (e.g., a predetermined pulse sequence for the imaging protocol, and magnetic field gradient and intensity data, as well as data indicating the timing, orientation, and spatial volume of the gradient magnetic field to be applied during the scan) and the resulting raw data or measurements.

[0036] The MR scanner 90 can be an image aperture MR scanner 90, such as one with a uniform B0 field provided by a field intensity of 0.5T or higher. The imaging aperture B0 field has, for example, a VRMS of <0.5ppm over the volume of interest. The image aperture MR scanner 90 provides an imaging aperture linear gradient with, for example, <2% geometric distortion.

[0037] In other embodiments, the MR scanner 90 has fewer design constraints, such as being designed and constructed for analysis and / or imaging without reconstruction. For example, it may provide a non-uniform master magnetic field (e.g., a 10% variation in the patient's scan area), a non-uniform B0 field or emission pulse (e.g., >0.5 ppm), and / or a nonlinear gradient (e.g., >2% geometric distortion). As another example, the master magnet 100 is 0.1 T or less. Where such a less uniform magnetic field is required, the aperture in which the patient is located during scanning can be open. An open-aperture scanner allows sufficient space for the field of view of sensor 190 to sense (e.g., take a photograph or view) the patient within the aperture. The field of view of sensor 190 extends to the patient while the patient is within the open aperture of the open-aperture MR scanner 90. For example, the open aperture could be a chair or bed serving as a patient support 150 without a surrounding shell, or it could be a shell open at more than two ends along the longitudinal direction of patient 140, such as having open sides and open ends (i.e., the shell is above and below the patient but not on the top, bottom, or sides). As another example, the hole can be cylindrical, but with a diameter exceeding 4, 5, or 6 feet.

[0038] For less restrictive designs, the MR scanner 90 can be placed in a room without a Faraday cage. No radio frequency shielding is provided outside or around the MR scanner 90. In other embodiments, the MR scanner 90 is placed in a room formed as a Faraday cage.

[0039] The patient support 150 is a flat or contoured plate (e.g., a bed) on which the patient 140 lies or is supported. In an open aperture, if the aperture is large, the patient support 150 can be formed as a recliner or chair.

[0040] The patient support 150 is movable relative to the MR scanner 90 (i.e., the main field magnet 100, gradient coil 110, and whole-body coil 120). A motor with gears, pulleys, and / or other transmissions moves the patient support into and out of the aperture, such as moving it longitudinally along the aperture or along the patient support 150. Other movements are possible, such as raising and lowering the patient support 150, laterally moving the patient support 150 (orthogonal to the side where the patient 140 lies on their back), and / or rotating it in one, two, or three dimensions. One or more sensors can measure the position of the patient support 150 relative to the MR scanner 90.

[0041] The patient support 150, with patient 140, is moved to a more uniform or most uniform portion of the magnetic field created by the main field magnet. Using gradient coils 110, the MR scanner 90 can position the region of interest (ROI) or scan area at different locations within the field of view (FoV) of the MR scanner 90. The patient support 150 moves patient 140 such that the RIO is within the field of view of the MR scanner 90 to allow localization. For example, the prostate is to be scanned. The patient support 150 moves the lower abdomen of patient 140 to be centered and / or moved into the aperture. The MR scanner 90 then scans the prostate area instead of other scanable areas, with patient 140 positioned at the location established by the patient support 150. In an alternative embodiment, an open aperture allows the patient to move themselves within the aperture or the FoV of the MR scanner 90.

[0042] Positioning by the patient support 150 and then localization (i.e., identification and manipulation to scan the region of interest) by the MR scanner 90 can be processed sequentially. This could be in the case where the aperture obstructs the sensor 190 from sensing the patient near the region of interest. Alternatively, after the patient 140 is positioned by the patient support 150, a pilot or pre-MR scan is performed to locate the region of interest. In designs with smaller constraints, such as those with larger or open apertures, the sensor 190 can be used to locate the region of interest for manipulation or localization of the MR scan, with or without a pre-MR scan. This allows patient positioning and localization to be performed as a single, unified step. Positioning and localization are performed jointly, rather than sequentially. This allows control of mechanical actuators and external sensors to manipulate the MR field. This is a direct result of the open-aperture low-field hardware concept, where patient positioning can still change once in the FoV of the MR device, external sensors 190 can still approach the patient 140 in scan positioning, and / or the MR scanner 90 may have a smaller FoV and require mechanical steps (i.e., patient support positioning) to reach a given ROI.

[0043] Sensor 190 is one or more sensors. Sensor 190 is positioned outside the aperture or FoV of MR scanner 90, but may be inside the aperture. Sensor 190 is mounted to the housing of MR scanner 90, a robotic arm, a wall, a ceiling, or a sensor tree. Sensor 190 is positioned such that when patient 140 is inside the aperture, sensor FoV captures all or part of patient 140. For example, sensor FoV reaches the external portion of patient 140 through the region of interest, while patient 140 is positioned by patient support 150 where MR scanner 90 can be localized to scan the region of interest.

[0044] Sensor 190 can be an active or passive sensor. For example, sensor 190 may be a camera for acquiring optical images or a depth camera for acquiring optical images with depth. An infrared camera or a camera for video resolution may be used. Other types of passive sensors may be provided, such as a laser rangefinder, a radio frequency sensor, or a weight pad that senses weight at different locations on the patient support 150. In alternative or additional examples, sensor 190 may be an active sensor, such as an ultrasound scanner that tactilely emits acoustic energy and receives echoes, or a camera on a robotic arm that actively moves its camera.

[0045] Processor 160 configures MR scanner 90 and / or determines one or more analytical values ​​from scanned data. Processor 160 is a general-purpose processor, digital signal processor, graphics processing unit, application-specific integrated circuit, field-programmable gate array, artificial intelligence processor, tensor processor, digital circuit, analog circuit, a combination thereof, or other known or later-developed device for manipulating raw data and / or applying artificial intelligence. Processor 160 can be a single device, multiple devices, or a network. For more than one device, parallel or sequential processing partitioning can be used. Different devices constituting the image processor can perform different functions, such as configuring MR scanner 90 for scanning by one device and determining analysis by another device based on the raw data. In one embodiment, processor 160 is a control processor or another processor of MR scanner 90. Other processors, either internal to MR scanner 90 or external to MR scanner 90, can be used.

[0046] Processor 160 is configured by software, firmware, and / or hardware to perform its tasks. Processor 160 operates according to instructions stored on a non-transitory medium (e.g., memory 170) to perform the various actions described herein.

[0047] Processor 160 is configured to jointly (1) locate patient 140 by movement of patient support 150 and / or MR scanner 90, and (2) localize the scan of that area by MR scanner 90. Using sensor 190 and / or MR scanning, processor 160 resolves patient support positioning and localization of the scan area (e.g., region of interest) within the MR scanner FoV. The solution is joint, such as resolving both positioning and localization together. The solution can be iterative, such as moving patient support 150 and setting localization in a first solution, and then refining the solution.

[0048] In one embodiment, processor 160 applies machine learning model 175, which is configured to jointly perform localization and positioning as a single solution for both. The machine learning model may be a neural network that accepts sensor outputs as input (e.g., from sensor 190, any patient support positioning sensors, and / or MR scan data) and outputs the localization and region of interest of the patient support 150 to be placed there, relative to the MR scanner FoV, for localization. The machine learning model is trained to control or act on the patient support 150 and the MR scanner 90 to achieve the solution without user input or control. For example, the machine learning model is trained to be a machine learning detector configured to detect current localization and positioning from the output of sensor 190 and scan data, and is a machine learning actor configured to move the patient support 150 and / or change the localization if the detected localization and / or positioning is incorrect (not optimal or infeasible, e.g., localization outside the MR scanner FoV).

[0049] In alternative or additional embodiments, processor 160 is configured to analyze scan data once the patient is located and the scan is localized. This analysis is performed by machine learning model 175. Machine learning model 175 may be different from or the same as the model used for localization. Machine learning model 175 is trained and configured to output a diagnosis in response to input scan data.

[0050] Diagnostic questions such as cancer or cancer staging are answered by machine learning model 175 in response to input scan data. For example, processor 160 is configured to determine an analysis specific to the patient. This analysis may be a global analysis, such as representing characteristics of the patient in general, rather than a positional representation as in imaging. For example, a global analysis is a clinical finding such as (1) the patient does not have or may have cancer, or (2) further scanning or imaging is unnecessary or required. Processor 160 is configured to determine the value of the analysis by applying machine learning model 175 to the raw data without reconstruction from the raw data. The image reconstruction step is skipped or not provided. AI analysis is performed directly from the raw data to the clinical finding as a diagnosis. Processor 160 uses machine learning model 175 to determine the analysis without reconstruction.

[0051] In another example, processor 160 is configured to change settings based on analysis. In addition to detecting diagnostics, processor 160 also configures MR scanner 90. The initial configuration may be based on the diagnostic question or why the patient is being scanned. A default configuration is used. Once the scan begins, processor 160 can change the configuration. Based on the acquired scan data, the configuration is changed to acquire more relevant or more definitive data. In one embodiment, machine learning model 175 includes a machine learning actor that determines the configuration. Analysis of the scan data can be used by the actor, such as identifying desired data that is not yet sufficiently available. For example, backpropagation of machine learning model 175 used to answer diagnostic questions is used to identify information (scan data) that leads to greater uncertainty in the answer. The actor then configures MR scanner 90 to reacquire that information or to acquire missing information.

[0052] Processor-driven changes enable reconfiguration during patient scanning in a given or single examination. Instead of following protocols designed to reconstruct specific types of images, a combination of scanning and analysis of scan data used for diagnosis allows for modifications to the scan to acquire scan data more quickly and / or more thoroughly, enabling more confident answers to diagnostic questions.

[0053] In yet another additional or alternative embodiment, processor 160 is configured to monitor for artifacts in the scan data and modify the scan (e.g., settings or data to be maintained) based on the level of artifacts. For example, machine learning model 175 includes a machine learning detector to detect general artifacts or specific types of artifacts, and includes a machine learning actor to modify the scan configuration or workflow to resolve any detected artifacts.

[0054] Machine learning model 175 is one or more models. It can be arranged hierarchically, sequentially, or otherwise. Machine learning model 175 is formed from one or more networks and / or other machine learning architectures (e.g., support vector machines). For example, as used herein, the machine learning network is a deep learning neural network. In another example, the machine learning network is a neural network of sequences of transformers and / or attention layers. In one embodiment, machine learning model 175 includes a model for detection. Different events (e.g., localization and localization, diagnostic values, and / or artifacts) are detected by different machine learning models or a common multi-task model trained to detect different events. Different actions (e.g., moving a patient support and / or configuring the MR scanner 90 for localization, artifact reduction, and / or information gain to obtain information to answer diagnostic questions) are performed by a common machine learning model (e.g., a reinforced deep learning network or a machine-trained actor) or different machine learning models for different actions.

[0055] Machine learning model 175 is trained using training data with or without a benchmark. A loss based on the trained model's output compared to the objective function or the benchmark is used to optimize the training. Any optimization, such as Adam, can be used. Any loss, such as cross-entropy, L1 loss, or L2 loss, can be used. Pre-training, cross-training, and / or continuous training can be used. Training data is collected from a database of examples executed under expert control. The benchmark can be curated or created through expert review. Alternatively or additionally, training data can be created by modeling or synthetically created using a model from an MR scan.

[0056] Memory 170 is a cache, buffer, RAM, removable media, hard disk, or other computer-readable storage media. Computer-readable storage media include various types of volatile and non-volatile storage media.

[0057] Memory 170 stores raw data, settings for control, information derived from the settings, machine learning model 175, and values ​​for localization, artifact levels, and / or analysis. Memory 170 may alternatively or additionally store instructions for processor 160. In response to one or more sets of instructions stored in or on a non-transitory computer-readable storage medium in memory 170, processor 160 performs the functions, actions, or tasks illustrated in the figures or described herein. The functions, actions, or tasks are independent of a particular type of instruction set, storage medium, processor, or processing strategy, and may be performed by software, hardware, integrated circuits, firmware, microcode, and the like, operating individually or in combination.

[0058] In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored at a remote location for transmission over a computer network. In yet another embodiment, the instructions are stored within a given computer, CPU, GPU, or system. Because some of the system components and method steps described in the accompanying drawings can be implemented in software, the actual connections between system components (or process steps) can vary depending on how this embodiment is programmed.

[0059] Display 180 is a CRT, LCD, plasma, projector, printer, or other display device. Display 180 is configured by loading an image onto a display plane or buffer. Display 180 is configured to display values ​​from analyses, such as one or more global analyses (e.g., clinical findings). The display may utilize or not utilize reconstructed images of patient 140. For example, the display may be part of a patient's report, text results, or electronic health record. Display 180 displays the analyzed values ​​to assist medical professionals in decision-making. For example, the value may indicate the detection of possible prostate cancer and / or the need for further scanning or imaging. In cases where MR scanner 90 is a low-cost scanner with no or poor imaging capabilities, the scan may be used as a first test using cheaper equipment to determine whether more expensive MR or other types of medical imaging should be performed.

[0060] Figure 2 This is a flowchart illustrating one embodiment of a method for autonomous MR scanning. Autonomous scanning is provided for a non-imaging or poorly imaging MR scanner 90, thereby allowing a low-cost, low-tech expertise system to assist in diagnosis and / or decision-making regarding whether to perform a high-cost imaging scan. The data analysis method provides automation of localization, MR scanning, analysis, and output generation. These actions are performed without human control. Some human control may be provided to initiate actions (e.g., automatic button press scanning) and / or initial protocol selection. In other embodiments, human control is provided for one or more actions.

[0061] The analysis of localization 200, quality control 230, and diagnostic scan 220 can be performed autonomously using one or more machine learning models. This method is... Figure 1 The system or another system performs the operation. The MR scanner 90 scans the patient. The processor performs detection and / or actions based on information from sensors, MR scans, and / or health records (e.g., the patient's scan sequence and / or clinical data). The display shows one or more clinical findings, utilizing or not utilizing the reconstructed images of the patient. Other components, such as a remote server or workstation, may be used to perform the generation and / or display.

[0062] During the application of a machine learning model to one or more different patients and corresponding different scan data, the same learning weights or values ​​of the machine learning model are used. At least for a given time period (e.g., weeks, months, or years) or a given number of uses (e.g., tens or hundreds of times), the model and values ​​of the learnable parameters do not change from one patient to the next. These fixed values ​​and the corresponding fixed model are applied sequentially and / or by different processors to scan data for different patients. The model can be updated, such as being retrained or replaced, but new values ​​are not learned as part of the application for a given patient. In other embodiments, continuous learning is used.

[0063] The method is performed in the indicated order (top to bottom or numerically) or other sequences. For example, actions 200 and 210 are performed together to obtain a solution (i.e., a combined solution for patient support positioning and MR scanner localization), which can be implemented sequentially (moving the patient support, then configuring the scanner). Additional, different, or fewer actions may be provided. For example, the scan is configured using preset, default, or user-input settings prior to action 220. As another example, the value (disease score 250) is stored in memory (e.g., a computerized patient medical record), transmitted over a computer network, and / or displayed on a monitor.

[0064] In action 200, the processor positions the patient relative to the MR scanner. The patient support is moved together with the patient on it to position a portion of the patient within the FoV of the MR scanner. In action 210, the processor localizes the MR scan. The MR scan is manipulated or controlled to scan less than the full Scan FoV, such as scanning the region of interest. The MR scan is localized by setting the area to be scanned relative to the MR scanner.

[0065] The positioning of action 200 and the localization of action 210 refer to actions taken to ensure that the diagnostic scan covers the clinical region of interest (e.g., the prostate). Figure 3 An example embodiment is shown. The patient is placed within an area (FoV) covered by the MR device itself (MR sensor 320) and / or other sensors such as a camera, laser rangefinder, or RF sensors (passive sensor 300 and / or active sensor 310). Sensors 300, 310, and 320 acquire measurements 305, 315, and 325. In action 330, measurements 305, 315, and 325 are provided to a processor to determine where the MR scanner 320 is focused on the target area of ​​interest.

[0066] Depending on the MR scanner, positioning 350 can be achieved either by using a mechanical actuator such as an electric motor or by providing feedback to the patient or technician to help them reach the correct positioning. In action 360, the parameters of the MR device (e.g., gradient intensity, pulse carrier frequency, multi-coil RF transmit and receive focus) are locally updated to focus the scan on the clinical region of interest (ROI). In action 340, the localization parameters are output to the MR scanner. The localization parameters can be the spatial coordinates of the target ROI in the room coordinate system or direct electrical control applied to the MR device to reach the ROI. The localization parameters can be provided to the active sensor 310 to manipulate the active sensor to examine or sense in the vicinity of the localized ROI.

[0067] In high-field scanners, a uniform field and linear gradient mean that the scanner can select cubes within a volume using a combination of slice-selective excitation, frequency, and phase readout encoding. In low-field scanners, non-uniform fields and / or nonlinear gradients can limit the FoV and the corresponding number of cubes. A more open aperture can allow for further patient support positioning in conjunction with localization to place the region of interest for scanning within the limitations of a given scanner FoV.

[0068] In one embodiment, the localization of action 200 and the localization of action 210 are achieved through a machine learning model. The processor applies measurements 305, 315, and / or 325 to the machine learning model to jointly solve for mechanical localization 350 and localization 360 / 340. The machine learning model includes a machine learning detector to determine where the region of interest is located relative to the MR scanner and / or whether the MR scanner is correctly focused on the target region of interest. The machine learning model includes a machine learning actor that alters the mechanical degrees of freedom of the device (e.g., localization 350) and the configuration of its sensors (including the MR sensor itself) (e.g., updating localization 360). The machine learning actor responds to all changes in the input solving location, such as patient localization and MR scanner localization.

[0069] exist Figure 2 In the example, positioning action 200 and localization action 210 are performed to locate the patient's prostate 215 for MR scanning. Other organs or regions may be targets.

[0070] In action 220, the MR scanner scans the patient using its configuration. The scan is guided by a protocol that establishes the values ​​for the scan's settings or controls. The scan generates measurements. A pulse sequence (i.e., multiple pulses from one or more coils) is created based on the MR scanner's configuration (e.g., the selected imaging protocol). The pulse sequence is transmitted from the coils into the patient's body. The resulting response is measured by receiving radio frequency signals at the same or different coils. The scan produces raw measurements as scan data.

[0071] This protocol is used for medical testing. It is designed to provide scan data that can be used to arrive at clinical findings. The scan data can be used to diagnose or answer diagnostic questions, such as whether a more detailed scan is needed, whether cancer is present in an organ, or the stage of cancer. Patients undergo MR scans to obtain clinical findings. For a given MR examination, the patient is positioned, the scan is localized, and then the patient is scanned to find clinical findings. MR scans are performed continuously in intervals of seconds or minutes to acquire scan data, thereby answering diagnostic questions.

[0072] MR scanning can be performed using an image aperture MR scanner. In other embodiments, scanning is performed using an MR scanner with a non-uniform master magnetic field, non-uniform pulses, and / or nonlinear gradients.

[0073] The configuration established by the protocol can be updated during the scan for a given examination. In order to arrive at an answer with sufficient confidence (e.g., a clinical finding), the configuration of the MR scanner may be changed or updated during the examination.

[0074] Figure 4 An example is shown. MR control parameters 450 are initially established by a selected protocol (e.g., prostate examination). In action 220, scan data (MR measurements) 400 are acquired via scanning. In action 410, a processor (e.g., a computer, workstation, server, or scanner processor) analyzes the scan data 400. This analysis may include analysis of the MR control parameters 450. The scan data 400 and control parameters 450 are analyzed to determine whether the configuration (control parameters 450) should be changed in action 420 and / or how to make the change in action 440. The change is to acquire different measurements, therefore parameter 450 is modified to obtain the desired measurements. During the MR examination, MR scan 220 is modified to obtain scan data optimized for diagnosis.

[0075] In one embodiment, the processor uses a machine learning model in data analysis action 410. Scan data 400 and MR scan parameters 450 are input into the machine learning model. Scan parameters 450 may be the type of scanner and / or settings of scanner controls (e.g., settings of the selected protocol). Scan parameters 450 may be (1) specific scanner controls, such as k-space coordinates, repetition time, echo time and / or flip angle, (2) the pulse envelope, carrier frequency, timing and duration of the pulses used in the MR scan of action 220, or (3) the original transmitted pulse itself or a representation of the transmitted pulse. Other formats of scan parameters may also be used.

[0076] In response to the input, the machine learning model determines a diagnostic value for the medical test and the uncertainty of that value. It outputs an answer to the diagnostic question and a confidence level in that answer. Neural networks, regressors, or other machine learning models can be trained to output clinical findings or analytical values ​​and the uncertainty of those values.

[0077] The machine learning model generates values ​​for each of one or more analyses of a patient. The analysis is a parameter representing the patient. This parameter is not a pixel-by-pixel or voxel-by-voxel imaging or representation, but instead represents more general information, such as clinical findings. The analysis or parameter is a global representation. When representing an organ, anatomical structure, and / or lesion, this representation is general to the region, rather than specific to a particular part of that region. For example, the analysis is whether the patient has an indication of cancer. As another example, the analysis is whether the tumor is benign or malignant. In yet another example, the analysis is cancer staging. One of the stages could be no cancer. As yet another example, the analysis is whether further scanning is recommended. Further scanning could be in the sense of further testing, whether by image scanning or by laboratory screening. Probabilities or non-binary values, such as the likelihood of cancer or other clinical findings, can be generated.

[0078] In action 420, the processor uses the configuration of parameter 450 to determine whether sufficient (yes) or insufficient (no) information has been collected by the MR scan of action 220. During diagnostic scanning, the MR scanner continuously acquires and analyzes measurements until the analysis confidently answers the clinical questions for which it scheduled the scan. The acquired data 400 is continuously analyzed during acquisition, allowing the system to determine whether it has collected enough information to make a decision. A machine learning model calculates a measure of the uncertainty of the analysis. If the analysis is a categorical variable (e.g., the presence or absence of a disease), the uncertainty can be measured as the entropy of the posterior distribution of all possible values ​​of the analysis. If the analysis has numerical values ​​(e.g., the volume of an organ or lesion), the uncertainty can be measured using a measure of the dispersion of the posterior distribution of the analysis (such as its standard deviation). The uncertainty output by the machine learning model is compared to a confidence threshold level. If the uncertainty is below the threshold (e.g., the confidence is above the threshold), then in action 430, the scan data 400 provides sufficient information to output a result. The value of the analysis is output.

[0079] If the confidence level of the analyzed value is low, the system then specifies which measurements to acquire next in action 440 to maximize the information gain about the previously acquired data 400. The processor controls the MR scanner using a different configuration based on the change from the previous configuration. Changing the configuration and corresponding MR control parameters 450 to acquire further measurements (MR data 400) allows for more confident answers to diagnostic questions. Changing the configuration also obtains information useful for machine learning models when detecting analyzed values.

[0080] In one embodiment, in action 440, a machine learning model is used to specify the next measurement. Backpropagation of the machine learning model used to determine the values ​​of the analysis identifies the information that contributes most to the uncertainty. This identifies missing or incorrect MR scan data 400, thereby identifying what measurement to perform next. Action 440 specifies the measurement that provides the missing scan data or replaces the incorrect MR data. Alternatively or additionally, the same or different machine learning models output control parameters 450 and / or the desired measurement in response to inputs of: (1) the identification of missing information or (2) the scan data 400 and parameters 450.

[0081] For data analysis in action 410, one or more machine learning models may be used. For example, a machine learning model (e.g., a machine learning detector) outputs a medical answer and its uncertainty in response to inputs of MR data 400 and control parameters 450. Another machine learning model is a machine learning actor used to specify measurements, set control parameters, or otherwise configure subsequent scans 220 for a given examination.

[0082] For machine learning actors, reinforcement learning models can be used. Since the system contains actor components that influence future data collection, deep reinforcement learning is used to train the system for machine learning. The system learns policies for making decisions. Training data can be synthesized using physical simulations of patient populations generated from clinical MRI datasets. The reward function in reinforcement learning is a measure of the amount of relevant information acquired by the MR scanner, such as information gain about the label (i.e., analysis or disease) or the reduction of uncertainty in a label predictor based on the acquired data.

[0083] The system can be broken down into parts and components used for pre-training. Localization and localization is one part. Diagnostic analysis is another part. Quality control is yet another part. The different parts can be pre-trained separately. Supervised training is used to pre-train the detector components (i.e., machine learning detectors such as localization detection, diagnostic data analysis, and quality control anomaly detection).

[0084] Once the detector is trained, the reinforcement learning of the actors is essentially a self-supervised process based on predicting future rewards, so the actors can continuously improve on the deployed devices. While allowing continuous learning to update device behavior online may be unacceptable, heterogeneous policy sampling can be used to perform learning without affecting the model's behavior. That is, using released actors to select actions, using those actions to train continuous actors, and then triggering a release process that includes network quality checks once the continuous actors are considered to have sufficient divergence from the released actors.

[0085] Machine learning actors through Figure 2 The entire process learns to make decisions or exercise control. Alternatively, separate machine learning actors can be provided for different components.

[0086] Machine learning models acting as detectors and / or actors can be any of a variety of architectures. For example, a machine learning model can be a neural network, such as a fully connected neural network (FCN) or a convolutional neural network (CNN). Any architecture or layer structure used for machine learning can be used. The architecture defines the structure, the learnable parameters, and the relationships between the parameters. In one embodiment, a convolutional or other neural network is used. Any number of layers and nodes within layers can be used. DenseNet, encoders, autoencoders, CNNs, FCNs, and / or other networks can be used. Some networks may include dense blocks (i.e., multiple layers whose outputs sequentially go to the next layer in the dense block and the final layer). Any known or later-developed neural network can be used. Any number of hidden layers can be provided between the input and output layers. The model is trained on the architecture.

[0087] Optimal control and reinforcement learning systems are synchronous systems that wait for an actor to decide on the next action. However, for MR scanners, real-time control may be impractical due to computational constraints, especially during the diagnostic phase, where the total delay between successive measurements may be on the order of milliseconds. To enable optimal control without slowing down the scanner, in action 440, the actor is implemented asynchronously, where the scanner maintains and follows a data-independent scan plan (configuration) until the actor requests a change to the plan. In this framework, the output of the scan is the scan plan, an open-ended generator of all future MR controls, not just the control for the next measurement. In this case, the output of the scan actor is the parameters of the scan plan model. MR scan 220 continues with the current configuration until the reinforcement learning model is generated and changed.

[0088] Refer again Figure 2Scanning action 220 can respond to action 230 for quality control. As part of the scan in action 220, signal quality 240 is determined and used to control the scan in action 440 and / or the analysis in action 410. The processor checks the quality of the raw data 400 from scan 220. According to one embodiment of quality control in action 230, the quality of the scan data is checked by a machine learning model. Then, in action 440, the processor can modify the configuration of the MR scanner based on quality and / or missing information. Quality checks can be controlled by removing information, thus affecting the missing information used to specify further measurements. Quality checks can be controlled by changing parameters to avoid artifacts, thus affecting the scan parameters 450 used in the modified configuration. Quality control periodically or continuously monitors the acquired data to check whether the MR scanner, patient, and / or environment are performing as expected, and to specify recovery steps in case of anomalies.

[0089] In one embodiment, a machine learning model is used to determine quality. Quality is determined in a general sense (not specific to a given type of artifact) or for each of one or more different types of artifacts. The machine learning model is a machine learning detector used to detect the level of artifacts. The machine learning model can be a machine learning actor used to control the scan to improve quality without human input. The actor causes reconfiguration to improve quality without user triggering and / or selecting settings to be used for quality improvement. The actor can be separate from the machine learning actor to change configurations for information gain, or it can be a common actor used to make decisions related to both purposes.

[0090] In one embodiment, the quality check of action 230 checks for one or more specific artifacts. Configuration changes are made to reduce this type of artifact. Identified error sources, such as motion and RF interference, are monitored individually.

[0091] Figure 5An example of monitoring and control to improve quality related to motion artifacts is shown. In action 500, MR data 400 is analyzed to determine if the patient has moved. Sensors may be used to monitor motion instead of or in addition to MR data 400. If there is no movement or the movement is below a threshold, the scan continues in action 505. If there is movement exceeding the threshold, the processor determines whether MR data 400 is still interpretable. If so, motion information is passed to data analysis in action 410. If not, localization may be adjusted in action 525 if the area is still within the MR scanner's field of view, as determined at action 520. If not, the patient support or patient positioning is adjusted in action 530. In the event of patient movement during scan 220, remedial steps may include adjustments to positioning and localization, adjustments to data analysis to account for motion (changing the imaging type to diffusion and / or adjusting data to account for motion), and providing feedback to the patient and / or operator.

[0092] Figure 6 An example of monitoring and control to improve quality related to radio frequency interference is shown. In action 600, MR data 400 is checked for any interference. If none is found, the scan continues in action 605. If some interference is present, in action 610, it is checked whether the interference is localized near the patient, such as scan data reflecting the patient having a phone or a metal object on them. If localized interference is present, in action 615, guidance for removing the source of interference is provided, such as issuing instructions to the patient. Scan 220 is repeated, and the checks for interference are repeated in action 620. If the interference persists, scan data 400 is discarded, and the scan is aborted in action 630. If there is no longer any interference, in action 625, MR data showing interference is discarded, and in action 410, MR data with no or minimal interference is retained and used for data analysis. Remedial actions for radio frequency interference may include: providing feedback to the patient and / or operator if the source of interference is identified as being close to and under their control; and attempting to repeat the measurement if the source of interference is environmental but intermittent.

[0093] Quality control for Action 230 can be achieved through the application of a processor to a machine learning model. The machine learning model trained to output values ​​for analysis can be a multi-task model also trained to output artifact detection. The machine learning model trained to output values ​​for analysis can use quality indicators as input to assess uncertainty. The machine learning model trained to output configuration or control scans can receive artifact detection information to output controls for scanning in cases of fewer artifacts and / or for rescanning.

[0094] In one embodiment, the inspection for artifact detection targets artifacts in general, rather than those of a specific type. For example, the machine learning model includes a machine learning generator for generating known good images or scan data. This generator identifies deviations in quality from the scan data, regardless of the type of artifact. This more general anomaly detection is a class of detections that can also be performed to capture unidentified sources of error. Unidentified artifacts do not have associated remedial steps; however, their detection can be considered by the data analysis in action 410 to assess uncertainty. Because a class of detections requires no supervision, this component can be continuously trained after the device is deployed, and the resulting machine learning generator can be used as pre-training for other trainable components. Anomalous data can also be mined and sent back to a central development hub for further root cause analysis.

[0095] After reconfiguration based on data analysis of analytical values ​​and / or quality control, the MR scanner performs a scan in action 220 to obtain the next measurement. Information with better certainty for estimating diagnostic information using machine learning models is obtained by targeting scans that yield missing or incorrect (less artifact) information. This scan is part of the same MR examination, so the patient does not leave the patient support and may not feel any delay or interruption in this one or a single scheduled MR scan. The resulting scans or raw data can then be used for analysis and / or to generate analytical output values. Repeat Figure 4 The method is to continue until the data analysis of Action 410 indicates that the clinical findings have sufficient confidence or low uncertainty.

[0096] In action 245, generate diagnostic output. Generate the output of medical tests from the MR examination. This output is the analytical value or answer (clinical finding) for the diagnostic question. Images of patient tissue may or may not be output.

[0097] Output is generated from multiple iterations of scan data. The scan data is accumulated, allowing sufficient information with sufficiently low artifacts to be collected for confident estimation using a machine learning model that generates diagnostic information. In an alternative embodiment, earlier scan data is analyzed and discarded, contributing to or being used to form the output, resulting in the reconfiguration and acquisition of better scan data that is used as input to generate the output.

[0098] A display (screen or device) shows the output (e.g., the analyzed values). The display shows values ​​representing patient parameters, such as values ​​from clinical findings. This display can be part of a patient report, a pop-up, a laboratory result, or part of the patient's electronic health record. The value is displayed as alphanumeric text, a graph, or a line graph or chart. The value can be displayed without images of the patient's anatomy. In other embodiments, the value is displayed along with images of the patient's anatomy, such as images acquired using non-MR modalities (e.g., X-rays) or images reconstructed from MR scans.

[0099] The display shows values ​​for use by users, radiologists, physicians, clinicians, and / or patients. These values ​​aid in diagnosis.

[0100] Although the subject matter has been described with reference to exemplary embodiments, the subject matter is not limited thereto. Rather, the appended claims should be interpreted broadly to include other variations and embodiments that can be made by those skilled in the art.

Claims

1. A method for data analysis of magnetic resonance (MR) scans, the method comprising: During an MR examination, a first configuration of a medical test-based MR scanner is used to perform an MR scan on the patient and generate first raw data. The first raw data is analyzed using a first machine learning model, and the analysis results in a change to the first configuration, wherein the analysis includes: The first raw data and the first configuration are input into the first machine learning model, and the first machine learning model outputs diagnostic output as the answer to the diagnostic question. The first machine learning model determines the diagnostic value for the medical test and the uncertainty of the value; Based on the change from the first configuration, the MR scanner is controlled using a second configuration, wherein the control includes changing from the first configuration to the second configuration if the uncertainty is higher than a threshold; The patient is subjected to an MR scan using the second configuration of the MR scanner used for the medical test, as part of the same MR examination, and second raw data is generated; and The diagnostic output of the medical test of the MR examination is generated from the first raw data and the second raw data.

2. The method of claim 1, wherein the MR scan using the first configuration and the second configuration comprises an MR scan utilizing a non-uniform main magnetic field, a non-uniform first pulse, and / or a nonlinear gradient.

3. The method of claim 1, wherein the control further comprises specifying the next measurement based on backpropagation of the first machine learning model, and setting parameters of the second configuration for the next measurement based on missing information identified by the backpropagation.

4. The method of claim 1, wherein the analysis includes determining that insufficient information has been collected by using the MR scan with the first configuration, and wherein the control includes determining the second configuration to obtain the information.

5. The method of claim 1, further comprising locating the patient relative to the MR scanner and localizing the MR scan by solving a second machine learning model for localization and localization.

6. The method of claim 1, further comprising checking the quality of the first raw data by a second machine learning model, wherein the control includes controlling to improve the quality without human input.

7. The method of claim 1, wherein control includes control by a second machine learning model, the second machine learning model including a reinforcement learning model.

8. The method of claim 7, wherein the MR scan using the first configuration continues until the reinforcement learning model generates the change.

Images