Helical MR Imaging with Off-Resonance Artifact Correction

Abstract

The present invention relates to a method for performing MR imaging of an object (10) located in an examination volume of an MR device (1). The object of the present invention is to achieve efficient and high-quality non-Cartesian MR imaging even in the case of strong B0 inhomogeneity. According to the present invention, the method comprises: subjecting the object to an imaging sequence that includes at least one RF excitation pulse and a modulated magnetic field gradient; acquiring MR signals along at least one non-Cartesian k-space trajectory; reconstructing an MR image based on the acquired MR signals; and using a deep learning network to detect one or more undersampling artifacts in the MR image caused by undersampling of k-space due to B0 inhomogeneity. Furthermore, the present invention relates to an MR device (1) and a computer program.

Description

Field of the Invention

[0001] The present invention relates to the field of magnetic resonance (MR) imaging. The present invention relates to a method for MR imaging of an object. The present invention also relates to an MR device and a computer program to be run on the MR device. Background Art

[0002] Today, image formation MR methods that utilize the interaction between a magnetic field and nuclear spins to form two-dimensional or three-dimensional images have been widely used, especially in the field of medical diagnosis, because they have advantages in many aspects for soft tissue imaging compared to other imaging methods, they do not require ionizing radiation and are generally non-invasive.

[0003] Generally, according to the MR method, an object (e.g., the body of a patient to be examined) is arranged in a strong and uniform magnetic field (B0), the direction of which simultaneously defines the axis (usually the z-axis) of the coordinate system on which the measurement is based. The magnetic field generates different energy levels depending on the magnetic field strength for individual nuclear spins, and this magnetic field strength can be excited (spin resonance) by applying an electromagnetic alternating field (RF field) that defines a frequency (the so-called Larmor frequency or MR frequency). From a macroscopic perspective, the distribution of individual nuclear spins generates an overall magnetization, and the overall magnetization can be deflected from the equilibrium state by applying an electromagnetic pulse (RF pulse) of an appropriate frequency, such that the magnetization precesses around the z-axis. The precessional motion describes the surface of a cone, the cone angle of which is called the flip angle. The amplitude of the flip angle depends on the strength and duration of the applied electromagnetic pulse. In the case of a so-called 90° pulse, the spin is deflected from the z-axis to the transverse plane (flip angle 90°).

[0004] After the termination of the RF pulse, the magnetization relaxes back to the original equilibrium state, in which the magnetization in the z-direction is re-established with a first time constant T1 (spin-lattice or longitudinal relaxation time), and the magnetization in the direction perpendicular to the z-direction relaxes with a second time constant T2 (spin-spin or transverse relaxation time). The change in magnetization can be detected by means of a receiving RF coil, which is arranged and oriented in the examination space of the MR device in such a way that it measures the change in magnetization in a direction perpendicular to the z-axis. After applying, for example, a 90° pulse, the decay of the transverse magnetization is accompanied by a transition of the nuclear spins (caused by local magnetic field inhomogeneities) from an ordered state with the same phase to a state with a uniform distribution of all phase angles (dephasing). The dephasing can be compensated, for example, by means of a refocusing pulse (e.g., a 180° pulse). This generates an echo signal (spin echo) in the receiving coil.

[0005] To achieve spatial resolution in vivo, a constant magnetic field gradient extending along three principal axes is superimposed on a homogeneous magnetic field, such that the spin resonance frequency has a linear spatial dependence. Then, the signal picked up in the receiving coil contains components of different frequencies, which can be associated with different positions in the body. The signal data obtained via the receiving coil corresponds to a spatial frequency domain called k-space. k-space data typically consists of multiple lines acquired with different phase encodings. Each line is digitized by collecting a large number of samples. The k-space data set is converted into an MR image with the aid of an image reconstruction algorithm.

[0006] The off-resonance effect due to B0 inhomogeneity and tissue-induced susceptibility changes is a major source of very common artifacts in MR images. Th. Kuestner et al. mentioned in the ISMRM-2019 abstract "Simultaneous detection and identification of MR artefact types in whole body imaging" of ISMRM-2018 (abstract 430) the CNN-based detection of artifacts originating from motion or field inhomogeneities in magnetic resonance images acquired using T1-weighted or T2-weighted FSE sequences. Using MR signal acquisition along a Cartesian k-space trajectory, it is well known that the inhomogeneity of B0 causes geometric distortion, which usually does not have a great impact on image quality. However, for non-Cartesian k-space trajectories (such as in spiral imaging), off-resonance usually manifests as blurring of the ideal image, which can seriously affect the diagnostic validity of the image. The degree of artifacts caused by B0 inhomogeneity is proportional to the main magnetic field strength.

[0007] Spiral imaging is a fast MR imaging technique that benefits from efficient k-space coverage and low sensitivity to flow artifacts. However, it is particularly vulnerable to B0 inhomogeneity when using long acquisition intervals (such as in single-shot spiral imaging). Such long acquisition intervals combined with parallel imaging techniques are crucial for obtaining the highest possible spatial resolution. This is meaningful in functional MR imaging (fMRI) and diffusion-weighted imaging (DWI) for alleviating problems of physiological and patient motion.

[0008] Deblurring methods for spiral MR imaging are known in the art. For example, it is known to acquire a B0 map and correct MR signal data for B0 inhomogeneity effects based on the B0 map (see, e.g., Ahunbay et al., "Rapid method for de-blurring spiral MR images" (Magn.Reson.Med. 2000, Vol. 44, pp. 491-494); Sutton et al., "Fast, iterative image reconstruction for MRI in the presence of field inhomogeneities" (IEEE Trans.Med.Imaging. 2003, Vol. 22, pp. 178-188); Nayak et al., "Efficient off-resonance correction for spiral imaging" (Magn.Reson.Med. 2001, Vol. 45, pp. 521-524).

[0009] However, even after applying deblurring methods of the above type, artifacts often remain in image regions with very strong susceptibility-induced magnetic field gradients. In the case of a spiral k-space trajectory, such artifacts appear as characteristic ring artifacts in the reconstructed and deblurred MR image and can overlap or cover anatomical details of interest. The reason for such remaining artifacts is that in the presence of strong (usually susceptibility-induced) local magnetic field gradients, the shape of the spiral k-space trajectory deviates to a corresponding large extent from the theoretical spiral shape for the corresponding voxel. This situation is illustrated in the Figure 4 two-dimensional k-space curve. Figure 4 a shows an "ideal" spiral k-space trajectory, which is obtained by applying sinusoidally modulated magnetic field gradients in the k x direction and the k y direction in the presence of a completely uniform main magnetic field B0. However, in Figure 4 b, B0 is inhomogeneous, with a strong gradient in the x direction, such that the corresponding voxel position "sees" a skewed k-space trajectory that deviates significantly from the ideal spiral shape. The consequence of this effect is under-sampling of k-space and violation of the Nyquist criterion in some parts of k-space. In Figure 4 b, insufficient signal data is sampled from important parts of k-space. These effects can lead to undesired remaining artifacts (hereinafter referred to as undersampling artifacts) caused by undersampling of k-space due to local gradients at strong local off-resonance positions, and such undersampling artifacts are not easily corrected. Summary of the Invention

[0010] From the above, it is readily appreciated that there is a need to improve MR imaging techniques. The object of the present invention is to solve the above limitations and enable efficient and high-quality non-Cartesian MR imaging even in the case of strong B0 inhomogeneity / strong B0 gradients.

[0011] According to the present invention, a method for MR imaging of an object positioned in an examination volume of an MR device is disclosed. The method comprises the steps of:

[0012] subjecting the object to an imaging sequence, the imaging sequence comprising at least one RF excitation pulse and a modulated magnetic field gradient;

[0013] acquiring MR signals along at least one non-Cartesian k-space trajectory;

[0014] reconstructing an MR image from the acquired MR signals; and

[0015] using a deep learning network to detect one or more undersampling artifacts caused by undersampling in k-space due to B0 inhomogeneity.

[0016] The gist of the present invention is to use a deep learning network to automatically identify the source of remaining undersampling artifacts in the reconstructed non-Cartesian (especially spiral) MR images. Once the undersampling artifacts are detected, they can be corrected in a targeted manner by a suitable algorithm.

[0017] In a preferred embodiment, before the step of detecting remaining undersampling artifacts (e.g., integrated in the step of reconstructing the MR image), the reconstructed MR image is deblurred based on a B0 map. By combining conventional B0-map-based deblurring with deep learning to detect residual undersampling artifacts, MR images of particularly high quality can be obtained even in the case of strong susceptibility effects and steep local magnetic field gradients.

[0018] In a further preferred embodiment, the deep learning network is trained to derive an artifact map from the MR image, the artifact map being a pictorial representation of only the detected aliasing artifacts. To achieve this, the deep learning network is preferably trained using: a set of modeled artifact maps at the output (last network layer) of the deep learning network; and a superposition of the training MR image and the corresponding modeled artifact maps at the input (first network layer) of the deep learning network. For example, the modeled artifact maps can include point spread functions of mono - or multi - voxel off - resonance at different off - resonance frequencies. In other words, the method of the present invention simulates the simplest form of aliasing artifacts as point spread functions (corresponding to the non - Cartesian k - space sampling scheme used) in the presence of strong local off - resonance and uses these point spread functions to train the deep learning network. To obtain a more realistic appearance, the aliasing artifacts can be modeled as a small convolutional local aggregation of different point spread functions, each point spread function at a slightly different off - resonance frequency but close in space. To further refine the artifact model, small geometric distortions can be added to the point spread functions to account for more complex local aggregations of off - resonance voxels. Then, the deep learning network is trained using: the modeled artifact maps (at the output of the network) as images showing only the predicted aliasing artifacts; and the superposition of the corresponding aliasing artifacts and the (artifact - free) training MR images (at the input of the network). This method employs the concept of "residual learning". It can be applied to complex and large amounts of MR image data.

[0019] Once the artifact map can be used as the output of the deep learning network, the detected aliasing artifacts can be easily corrected, for example, by subtracting the artifact map from the reconstructed (and optionally de - blurred) MR image to correct the detected aliasing artifacts.

[0020] In a further preferred embodiment, the correction is limited to predefined image regions and / or image regions in which the B0 map indicates that the inhomogeneity or degree of local variation (gradient) of the main magnetic field exceeds a given threshold. To optimize the image quality, the information in the B0 map can be used to appropriately limit or confine the artifact subtraction process to image regions that are clearly potential sources of artifacts. In this way, situations where anatomical structures that resemble undersampling artifact structures are removed (which would potentially destroy important diagnostic information in the MR image) are automatically avoided. In one possible embodiment, the B0 map information is directly used in the process of training a deep learning network in order to guide artifact detection. Instead of using only MR images with residual undersampling artifacts as the input to the network, the B0 map is also fed into the network at the input of the network. In this way, the deep learning network automatically takes into account the B0 map information, thus avoiding the situation where useful image structures are misidentified as artifacts.

[0021] In a further embodiment, an anatomical atlas can be fitted to the imaging data (preferably volumetric data, e.g., 3D or multi-slice data), thereby allowing the identification of predefined image regions that are crucial for the occurrence of undersampling artifacts (e.g., in the head: regions near the visual prefrontal cortex; or the inner ear cavity, etc.). This can be used to limit the detection of one or more undersampling artifacts, for example, in cases where the B0 map is not available. In such cases, the correction of the identified artifacts is also limited to these predefined regions.

[0022] In yet another preferred embodiment, the deep learning network is a convolutional network. The so-called F-Net network architecture can be used (see "Foveal Fully Convolutional Nets for Multi-Organ Segmentation" by Brosch T, Saalbach A. (Proceedings of SPIE, Volume 10574, 2018)), which relies on multiple resolution levels to extract high-level and low-level features. Alternatively, other network architectures can be adopted. For example, the so-called U-Net architecture can also be used (see "U-net: Convolutional networks for biomedical image segmentation" by Ronneberger, O., Fischer, P., Brox, T. (International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234-241, 20153)). If the B0 map and the input MR image have very different resolutions, a customized Y-Net architecture (see "Y-Net: A deep Convolutional Neural Network for Polyp Detection" by Mohammed, A. et al. (arXiv preprint arXiv:1806.01907, 2018)) may be appropriate. In such an architecture, fewer layers and / or channels can be used to implement the encoding path of the B0 map, which often has a smaller resolution, thereby reducing the overall size of the network.

[0023] The method of the present invention described so far can be performed by means of an MR device, which includes: at least one main magnet coil for generating a uniform static magnetic field within an examination volume; a plurality of gradient coils for generating switched magnetic field gradients in different spatial directions within the examination volume; at least one RF coil for generating RF pulses and / or receiving MR signals from an object positioned within the examination volume; a control unit for controlling the time succession of the RF pulses and the switched magnetic field gradients; and a reconstruction unit for reconstructing an MR image based on the received MR signals. The method of the present invention can be implemented, for example, by correspondingly programming the reconstruction unit and / or the control unit of the MR device.

[0024] The method of the present invention can be advantageously performed in most MR devices currently used clinically. For this purpose, it is only necessary to control the MR device with a computer program such that the MR device performs the above method steps of the present invention. The computer program can be present on a data carrier or in a data network for being downloaded and installed in the control unit of the MR device.

[0025] In addition, the sampling artifact detection and correction method of the present invention can be implemented as a retrospective artifact removal software tool to be installed on a diagnostic workstation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings disclose preferred embodiments of the present invention. However, it should be understood that the drawings are only designed for illustrative purposes and not for limiting the present invention. In the drawings:

[0027] Figure 1 An MR device for performing the method of the present invention is shown;

[0028] Figure 2 A B0 map, a reconstructed and deblurred MR head image, and a set of point spread functions modeled for different local off-resonances are shown;

[0029] Figure 3 A method for sampling artifact correction in a spiral MR head image of the present invention is illustrated;

[0030] Figure 4 A k-space diagram showing the source of sampling artifacts in a spiral MR imaging example is shown. DETAILED DESCRIPTION

[0031] Reference Figure 1 , schematically shows an MR device 1. The device includes a superconducting or resistive main magnet coil 2 such that a substantially uniform and temporally constant main magnetic field is created along the z-axis passing through the examination volume.

[0032] A magnetic resonance generation and manipulation system applies a series of RF pulses and switched magnetic field gradients to invert or excite nuclear magnetic spins, cause magnetic resonance, refocus magnetic resonance, manipulate magnetic resonance, spatially encode magnetic resonance and otherwise encode magnetic resonance, saturate spins, etc. to perform MR imaging.

[0033] More specifically, a gradient pulse amplifier 3 applies current pulses to a selected one of the whole-body gradient coils 4, 5, and 6 along the x-axis, y-axis, and z-axis of the examination volume. A digital RF frequency transmitter 7 emits RF pulses or pulse packets to the whole-body volume RF coil 9 via a transmit / receive switch 8 to emit RF pulses into the examination volume. A typical MR imaging sequence includes a packet of RF pulse segments of short duration, which are acquired together with each other, and any applied magnetic field gradients achieve a selected manipulation of nuclear magnetic resonance. The RF pulses are used to saturate resonance, excite resonance, invert magnetization, refocus resonance or manipulate resonance and select a part of the body 10 positioned in the examination volume. The MR signal is also picked up by the whole-body volume RF coil 9.

[0034] To generate an MR image of a limited region of the body 10, a set of local array RF coils 11, 12, 13 is placed adjacent to the region selected for imaging. The array coils 11, 12, 13 are capable of receiving MR signals induced by body coil RF transmission.

[0035] The resulting MR signals are picked up by the whole body volume RF coil 9 and / or the array RF coils 11, 12, 13 and demodulated by a receiver 14 preferably including a preamplifier (not shown). The receiver 14 is connected to the RF coils 9, 11, 12 and 13 via a transmit / receive switch 8.

[0036] The host computer 15 controls the gradient pulse amplifier 3 and the transmitter 7 to generate any one of a plurality of MR imaging sequences (e.g., echo planar imaging (EPI), echo volume imaging, gradient and spin echo imaging, fast spin echo (TSE) imaging, etc.) to acquire MR signals along a spiral k-space trajectory according to the present invention. For a selected sequence, the receiver 14 rapidly and successively receives single or multiple MR data along the corresponding k-space trajectory after each RF excitation pulse. The data acquisition system 16 performs analog-to-digital conversion on the received signals and converts each MR signal into a digital format for further processing. In modern MR devices, the data acquisition system 16 is a separate computer dedicated to acquiring raw image data.

[0037] Finally, the digital raw image data is reconstructed into an image representation by a reconstruction processor 17 which applies a Fourier transform or other suitable reconstruction algorithm. The MR image can represent a planar slice through the patient, an array of parallel planar slices, a three-dimensional volume, etc. The image is then stored in an image memory where the image can be accessed to convert, for example via a video monitor 18, a slice, a projection or other part of the image representation into a suitable format for visualization, and the video monitor 18 provides a human-readable display of the resulting MR image.

[0038] (e.g., by appropriately programming the host computer 15 and the reconstruction processor 17) the MR device 1 is arranged to perform the imaging method of the present invention as described above and below.

[0039] Continuing reference Figure 1 And further reference Figure 2 and Figure 3 to explain embodiments of the imaging method of the present invention.

[0040] In an exemplary embodiment, spin-echo T2-weighted spiral imaging is performed. During acquisition with an echo time of 60 milliseconds and an acquisition window of 50 milliseconds, only a single spiral trajectory is acquired. Parallel imaging is used, where the reduction factor R = 3. Through the use of iterative spiral SENSE image reconstruction and appropriate B0-map-based deblurring, a spiral spin-echo (SE) MR image is obtained (see Figure 2 the upper right image in). The (separately acquired) B0 field map for deblurring is also shown in Figure 2 (the upper left image). Due to the long acquisition window selected for optimizing scan efficiency and the high and strong varying local field inhomogeneities visible in the B0 map, after B0-map-based off-resonance correction (deblurring), uncorrected undersampling artifacts remain in the reconstructed MR image. The remaining undersampling artifacts appear as annular structures indicated by the arrows. In Figure 2 's bottom row, the point spread functions (single-voxel signal representations) calculated for different off-resonance frequencies given in Hertz and for a given spiral k-space trajectory used in MR signal acquisition are shown. For the position near the inner ear in the upper right MR image, the B0 map prediction deviation is approximately 250 Hertz. The corresponding inner ring sizes of the point spread functions are superimposed on the image. The point spread function simulates the signal contribution of a voxel or voxel aggregation centered at the marked position (small circle), which is distorted to an unrecognizable extent according to the existing strong local gradients and can no longer be corrected based on the B0 map. Undersampling artifacts may not be generated by just one voxel. It may be generated by small aggregations of several voxels that are close to each other, have different intensities, and experience different degrees of actual local inhomogeneities, resulting in a more complex distorted non-rotating or skewed artifact pattern. As described above, the calculated point spread function is used as a model artifact map to train a deep learning network. In the shown embodiment, the F-Net architecture that depends on multiple resolution levels to extract high-level and low-level features is used. Three different resolution levels are adopted, each with two convolutional layers.

[0041] As Figure 3 shown, the trained deep learning network analyzes the spiral MR image (upper left image) including undersampling artifacts. The network derives an artifact map from the MR image, which is the estimated result of the undersampling artifacts in the MR image. The detected artifacts are corrected based on the artifact map by subtracting the artifact map from the MR image. The corrected MR image is Figure 3The bottom image in. The arrows in the artifact image indicate undersampling artifacts that are misdetected based on misinterpretation of the deep learning network. By using the B0 map, regions with strong inhomogeneity and steep gradients can be identified. It can be predicted that the remaining undersampling artifacts will only appear in these regions of the MR image. Therefore, correction of the MR image based on the derived artifact map is only allowed in these regions. The information from the B0 map can be used to derive a corresponding subtraction mask that is "1" at the positions of strong inhomogeneity and steep gradients and smoothly drops to "0" in non-suspect regions, i.e., where the local inhomogeneity variation is below a critical threshold in this region, such that artifacts due to undersampling caused by off-resonance gradients cannot be predicted.

[0042] In another embodiment, an anatomical atlas can be fitted to the imaging data (preferably volumetric data, e.g., 3D or multi-slice data), allowing identification of regions with strong local susceptibility gradients that are crucial for the appearance of such artifacts (in the head, regions near the visual prefrontal cortex, or inner ear cavities, etc.), thereby supporting such weighted artifact subtraction in cases where field map information is not available.

[0043] In this way, situations where anatomical structures similar to the artifacts are erroneously removed and affect the clinical value of the image are automatically avoided.

Claims

1. A method for performing MR imaging on an object (10) positioned in an examination volume of an MR device (1), the method comprising: Subject the object (10) to an imaging sequence that includes at least one RF excitation pulse and a modulated magnetic field gradient; Acquire MR signals along at least one non-Cartesian k-space trajectory; Reconstruct an MR image from the acquired MR signals; and Use a deep learning network to detect one or more undersampling artifacts in the MR image caused by k-space undersampling due to inhomogeneity, wherein the deep learning network is trained to derive an artifact map from the MR image, the artifact map being a pictorial representation of only the detected at least one undersampling artifact.

2. The method according to claim 1, wherein, The non-Cartesian k-space trajectory is a spiral k-space trajectory.

3. The method according to claim 1 or 2, wherein, Before the step of detecting remaining undersampling artifacts, the reconstructed MR image is deblurred based on a B0 map.

4. The method according to any one of claims 1 - 3, wherein, The deep learning network is trained using: a set of modeled artifact maps at the output of the deep learning network; and a superposition of training MR images and corresponding modeled artifact maps at the input of the deep learning network.

5. The method according to claim 4, wherein, The modeled artifact maps include point spread functions of mono- or multi-voxel off-resonance calculated for the imaging sequence used.

6. The method according to claim 4 or 5, wherein, Correct the detected undersampling artifacts based on the artifact map derived by the deep learning network from the reconstructed MR image.

7. The method according to any one of claims 1 - 6, wherein, Detecting one or more undersampling artifacts is limited to a predefined image region and / or an image region where the B0 map indicates that the inhomogeneity or degree of local variation of the main magnetic field exceeds a given threshold.

8. The method according to any one of claims 1 - 7, wherein, During the detection of the undersampling artifacts, the B0 map is used as an additional input to the deep learning network.

9. The method according to any one of claims 1 - 8, wherein, The deep learning network is a convolutional network.

10. An MR device, comprising: At least one main magnet coil (2) for generating a uniform static magnetic field within the examination volume; A plurality of gradient coils (4, 5, 6) for generating switched magnetic field gradients in different spatial directions within the examination volume; at least one RF coil (9) for generating RF pulses within the examination volume and / or receiving MR signals from an object (10) positioned within the examination volume; a control unit (15) for controlling the temporal succession of the RF pulses and the switched magnetic field gradients; and a reconstruction unit (17) for reconstructing an MR image from the received MR signals, wherein the MR device (1) is arranged to perform the following steps: Subject the object (10) to an imaging sequence that includes at least one RF excitation pulse and a modulated magnetic field gradient; Acquire MR signals along at least one non-Cartesian k-space trajectory; Reconstruct an MR image from the acquired MR signals; and Use a deep learning network to detect one or more undersampling artifacts in the MR image caused by k-space undersampling due to inhomogeneity, wherein the deep learning network is trained to derive an artifact map from the MR image, the artifact map being a pictorial representation of only the detected at least one undersampling artifact.

11. A computer program comprising instructions for: Reconstructing an MR image from MR signals acquired using non - Cartesian k - space sampling; and Detecting, using a deep learning network, one or more undersampling artifacts in the MR image caused by undersampling of k - space due to inhomogeneity, wherein, The deep learning network is trained to derive an artifact map from the MR image, the artifact map being a pictorial representation of only the detected at least one undersampling artifact.

Images