Method and apparatus for acquiring and reconstructing a diffusion weighted magnetic resonance image sequence of a coverage volume

Abstract

A method for creating a plurality of diffusion-weighted magnetic resonance (MR) image sequences of an object is described, wherein each of the MR image sequences represents a same series of consecutive cross-sectional slices covering a volume of the object. The method comprises: (a) providing a plurality of image raw data set sequences collected using at least one radio frequency receiver coil of a magnetic resonance imaging device, wherein each image raw data set comprises a plurality of data samples generated using a combined diffusion-weighted spin echo and single shot stimulated echo sequence with diffusion encoding gradients; (b) performing a regularized nonlinear inverse reconstruction procedure on the image raw data set sequences generated using diffusion encoding gradients of zero or lower intensity to provide coil sensitivity and MR image sequences without or with lower diffusion weighting; and (c) performing a regularized linear inverse reconstruction procedure on all image raw data set sequences with diffusion encoding gradients of zero or lower intensity and higher intensity to provide MR image sequences without or with lower diffusion weighting and MR image sequences with higher intensity, each of the MR images representing one of the cross-sectional slices and being created by using the sensitivity of the at least one receiver coil determined in step (b) for the same cross-sectional slice and from a current estimate of the image content according to a difference between the image content of the adjacent cross-sectional slices. Furthermore, an MRI device for creating a diffusion-weighted MR image sequence of an object is described.

Description

Technical Field

[0001] This invention relates to a method for creating (especially for acquiring and reconstructing) magnetic resonance (MR) image sequences with diffusion contrast, wherein the MR images cover volumes of directly adjacent cross-sections and provide insensitivity to magnetic field inhomogeneities and improve image quality. Furthermore, this invention relates to a magnetic resonance imaging (MRI) apparatus configured to implement said method. Applications of this invention cover the field of MR imaging, particularly medical MR imaging (e.g., brain imaging or prostate imaging) or non-medical research in the natural sciences (e.g., the study of artifacts). Background Technology

[0002] In this specification, reference is made to the prior art, particularly the art related to MR image acquisition and reconstruction, which provides the technical background of the invention:

[0003] [1] V. Baliyan et al., World Journal of Radiology, Vol. 8, pp. 785-798, 2016;

[0004] [2] CSSchouten et al., Quantitative Imaging Med Surg, Vol. 4, pp. 239-250, 2014;

[0005] [3] K. Hirata et al., Medicine, Vol. 97, No. 19, e0447, 2018;

[0006] [4] US 4 748 409 A (J. Frahm et al.);

[0007] [5] J. Frahm et al., Journal of Magnetic Resonance, Vol. 65, pp. 130-135, 1985;

[0008] [6] K. D. Erboldt et al., Magnetic Resonance Imaging in Medicine, Vol. 23, pp. 179-192, 1992;

[0009] [7] UGNolte et al., Magnetic Resonance Imaging in Medicine, Vol. 44, pp. 731-736, 2000;

[0010] [8] AAKhalil et al., PLoS ONE, Vol. 11, e0161416, 2016;

[0011] [9] ABBakushinsky and MYKokurin, Iterative Methods for Approximate Solution of Inverse Problems, Springer, Dordrecht, Netherlands, 2005;

[0012]

[10] A. Merrem et al., Invest Radiol, Vol. 52, pp. 428-433, 2017;

[0013]

[11] A. Merrem, Thesis “Undersampled Radial STEAM MRI Methodological Developments and Applications”, University of George Augustus Göttingen, 2018;

[0014]

[12] A. Merrem et al., NMR Biomed, Vol. 32, doi:10.1002 / nbm.4074, 2019;

[0015]

[13] US 8 384 383 B2 (J. Frahm et al.);

[0016]

[14] M. Uecker et al., *NMR Biomed*, Vol. 23, pp. 986-994, 2010; and

[0017]

[15] J. Frahm et al., The Open Med Imaging, Vol. 8, pp. 1-7, 2014.

[0018] Diffusion-weighted (DW) magnetic resonance imaging (MRI), also known as MRI, provides information about dynamic molecular processes, such as the self-diffusion and / or transport of water molecules, by altering the signal intensity of MR images. This technique is highly clinically relevant because the diffusion characteristics differ between normal and pathological tissues, enabling diagnostically relevant comparisons, such as between normal tissue and ischemic brain tissue (i.e., stroke) or between normal tissue and tumor tissue (i.e., cancer).

[0019] Diffusion encoding of MRI signals is typically accomplished using a DW spin echo sequence, which consists of two radiofrequency pulses and a set of self-compensating strong magnetic field gradients. The spin echo sequence typically includes an initial excitation radiofrequency pulse, a first interval of half the spin echo time TE / 2, a refocusing radiofrequency pulse, and another interval TE / 2 before the spin echo forms at time TE (relative to the center of the initial excitation pulse). For diffusion encoding, the spin echo sequence is typically supplemented by a pair of identical magnetic field gradients, one gradient applied in each of the two spin echo intervals. These gradients can be unipolar or bipolar, but can also be more complex gradient waveforms. In the absence of motion, i.e., without any spatial displacement of water molecules, the nuclear spin moments of water protons are completely refocused in the spin echo signal, thus providing maximum spin echo signal intensity, which decays only through T2 relaxation. However, in the presence of molecular diffusion or transport, the spin echo signal is further attenuated because the excited water protons move to locations with different magnetic field strengths—or more precisely, when the water protons move between the coded diffusion gradient applied in the first spin echo interval and the refocusing diffusion gradient applied in the second spin echo interval—preventing complete refocusing of the spin moment. The relevant “diffusion time” is the duration between these mutually compensating gradients (centers) in the first and second intervals of the spin echo sequence.

[0020] Because the positional displacement caused by molecular transport processes is at a microscopic scale and therefore very small, DW MRI typically employs very strong diffusion-encoding gradients to ensure sufficient contrast (i.e., signal intensity difference) between DW and non-DW images (i.e., images acquired without a diffusion-encoding gradient but under otherwise identical conditions). These images with different diffusion weights are used to calculate quantitative maps of relevant diffusion characteristics, such as quantitative maps of the apparent diffusion coefficient (ADC).

[0021] However, the strong magnetic field gradient at the same time makes available DW MRI techniques highly sensitive to (macroscopic) motion. This problem, for example, in medical imaging, refers to involuntary body movements and motion caused by breathing, heartbeats, or peristalsis. Therefore, standard practice is to combine diffusion-encoded spin echo sequences with high-speed (preferably single-shot) readout MRI sequences to produce DW images that effectively “freeze” the actual conditions—images free from motion-induced artifacts. Several technical recommendations for this strategy have been proposed over the past few decades.

[0022] As described in a recent review [1], the preferred solution to overcome motion sensitivity in DW MRI is to use echo-plane imaging (EPI) readout. Currently, DW EPI is the only commercially available DW MRI method. Despite some serious problems, it is commonly (if not limited to) used in medical MR imaging and is widely accepted for clinical applications. The main drawbacks of DW EPI are as follows: As a gradient echo MRI technique, EPI relies on applying a single radiofrequency excitation pulse and acquiring multiple gradient echoes, which inevitably increases the echo time. Therefore, EPI is significantly sensitive to magnetic field inhomogeneities. For example, in the human body, this magnetic field distortion is caused by the unavoidable differences in the susceptibility of different tissues. This magnetic field distortion often results in severe geometric distortion and false positive or false negative signal changes in EPI images, reducing diagnostic accuracy. For example, inhomogeneity effects occur in the lower and anterior parts of the brain (i.e., near air-filled cavities or dental implants), and even more frequently throughout the body, such as in MRI of the prostate (i.e., near air-filled rectum). However, in clinical settings, DW-EPI is currently the only viable option because the images have a good signal-to-noise ratio (SNR).

[0023] Another solution for DW MRI is to use single-shot excitation spin echo MRI sequences (see [2] and [3]), which avoids sensitivity to magnetic field inhomogeneities and still enables high-speed acquisition. This technique uses multiple radiofrequency refocusing spin echoes (instead of gradient echoes from EPI) to generate multiple spin echo signals for image reconstruction and spatially encode them. However, a significant drawback of this approach is the requirement for a large number of high-power radiofrequency excitations (i.e., refocusing pulses with high flip angles), which often results in violations of the specific absorption rate (SAR) limit for energy deposition in the human body. Therefore, this single-shot excitation spin echo method has not been widely used in human applications.

[0024] Another suggestion for single-shot DW MRI is to combine DW spin echo sequences with readout sequences that include a series of radiofrequency refocused stimulated echoes. Single-shot STEAM (stimulated echo acquisition mode) MRI sequences were described as early as 1984 (see [4] and [5], etc.). In this case, multiple stimulated echoes are generated by radiofrequency pulses with low flip angles, thus avoiding any SAR problems. Recently, several versions of DW single-shot STEAM MRI have been developed that address the sensitivity issues of DW EPI, but their drawback is that the SNR is lower than that of DW EPI (see [6], [7], and [8]). Therefore, DW STEAM MRI has not yet gained clinical acceptance.

[0025] To overcome the insufficient SNR of DW STEAM MRI in specific brain applications, it has been proposed to combine this technique with an undersampled radial acquisition mode and use a two-step method to reconstruct the image. The two-step method first determines the coil sensitivity of the non-DW sequence in the original image dataset by performing a regularized nonlinear inversion of the coil sensitivity of adjacent cross sections. The joint reconstruction of the non-DW images and their coil sensitivity is achieved by iterative regularized Gauss-Newton method, as described in textbook [9]. In the second step, all non-DW and DW images are reconstructed by using a linear inversion of the coil sensitivity of the non-DW images without any further spatial regularization (

[10] ,

[11] ). Unfortunately, coil sensitivity provides very little spatial frequency information, which makes it impossible to significantly improve the above problems with weakly regularized reconstruction, and may even lead to numerical instability.

[0026] Another previous version of DW STEAM MRI, developed for a specific prostate application (

[12] ), employed a multiple acquisition method. In addition to again using weak regularization that only uses the sensitivity of adjacent coils (or even no spatial regularization), this method required more complex mathematical processing to achieve joint reconstruction of DW images from multiple acquisitions involving different macroscopic motion states.

[0027]

[13] and

[14] present an extremely fast method for acquiring and reconstructing dynamic non-DW MR image sequences using previous images described in [9] and regularization of the sensitivity of all coils. Acquisition times can be kept in the tens of milliseconds range due to the use of gradient echo MRI sequences with significant undersampling, non-Cartesian trajectories for spatial coding, and image reconstruction via regularized nonlinear inversion. Thus, temporal changes of the subject can be monitored in real time with high temporal fidelity, depending on the dynamic process to be studied (

[15] ). However, the techniques in

[13] and

[14] are primarily designed for collecting images of a single slice of the subject, thus neither gaining coverage of the volume of the subject nor diffusion weighting. Although images of different slices of the subject are also considered in

[13] and

[14] , this is limited to a few slices, for example, less than five slices. Furthermore, the corresponding applications are implemented as staggered multi-slice data acquisition, which results in a loss of temporal resolution and increased sensitivity to motion in the techniques in

[13] and

[14] .

[0028] Purpose of the invention

[0029] The object of this invention is to provide an improved method for creating, in particular acquiring raw image data and reconstructing a DW MR image sequence covering the volume of a subject under investigation. This method overcomes the shortcomings of conventional techniques and / or enables new applications of MR imaging. Specifically, the object of this invention is to provide an improved method for creating a cross-sectional DW MR image sequence that seamlessly covers the volume, while increasing acquisition speed, reducing (or even eliminating) sensitivity to motion, increasing SNR, and / or reducing (or even eliminating) sensitivity to magnetic field inhomogeneities. Specifically, in the context of medical imaging, the object is to provide a method for creating a DW MR image sequence that seamlessly covers the volume of the human body, while reducing (or even eliminating) sensitivity to motion and magnetic field inhomogeneities, thereby enabling medical MR imaging methods applicable to the entire human body or different parts of the body. Furthermore, the object of this invention is to provide an improved MRI device, particularly suitable for implementing a method for acquiring and reconstructing spatially continuous sequences of diffusion-weighted MR images of a volume. Summary of the Invention

[0030] The above-mentioned objective is achieved by an MR image creation method and / or MRI apparatus including the features of the independent claims. Advantageous embodiments of the invention are defined in the dependent claims.

[0031] According to a first general aspect of the invention, the above-mentioned object is achieved by a method for creating a diffusion-weighted MR image sequence of a subject under investigation (in particular, acquiring raw image data and reconstructing the diffusion-weighted MR data), wherein each of the MR image sequences represents a series of consecutive cross-sectional slices covering the volume of the subject.

[0032] The method of the present invention includes the following steps: providing a plurality of raw image dataset sequences, said plurality of raw image dataset sequences being collected using at least one radio frequency receiver coil of a magnetic resonance imaging (MRI) device.

[0033] Each raw image dataset includes the image content of one image from the MR images to be reconstructed, and each raw image dataset refers to one of the cross-sectional slices; that is, each of the MR images to be reconstructed represents one of the cross-sectional slices. The cross-sectional slices provide a series of directly adjacent cross-sections covering the volume of the object. The position of each cross-sectional slice is shifted in a direction perpendicular to the imaging plane by slice shifting to cover the volume of the object under investigation. The slice shift is equal to a predetermined slice thickness of the cross-sectional slice. Providing raw image data with directly adjacent cross-sections means that the provided raw image data has no or negligible overlap of the image content of the adjacent cross-sections, and no or negligible gaps between the image content of the adjacent cross-sections. The spatial orientation of the imaging plane (e.g., the longitudinal z-direction relative to the main magnetic field of the MRI device) can be selected according to the imaging task, such as based on the anatomical location of the organ to be imaged in the human body. The spatial orientation of the imaging plane can be set by the direction of the spatially encoded magnetic field gradient in the MRI device.

[0034] Furthermore, each raw image dataset includes multiple data samples generated using a combined diffusion-weighted spin echo and single-excitation stimulated echo sequence with a diffusion-encoded gradient (hereinafter referred to as a DW single-excitation STEAM sequence). The DW single-excitation STEAM sequence spatially encodes the MRI signal (i.e., the raw image dataset) received through the at least one radio frequency receiver coil using a non-Cartesian k-space trajectory (preferably a radial k-space trajectory).

[0035] The diffusion coding gradients of the plurality of original image dataset sequences have at least two different intensities and at least three different directions. Therefore, each original image dataset sequence is distinguished by a predetermined diffusion coding gradient intensity (gradient magnitude, which can be zero) and a predetermined diffusion decoding gradient direction, wherein at least two different sequences of the original image datasets use different diffusion coding gradient intensities, and at least three different sequences of the original image datasets use different diffusion coding gradient directions.

[0036] At least one of the original dataset sequences was generated using a diffusion coding gradient of zero or lower intensity. This sequence can also be referred to as a "zero-intensity or lower-intensity sequence" which was generated without using a diffusion coding gradient, or whose generation used a diffusion coding gradient with an intensity lower than that of the diffusion coding gradients in the remaining original image dataset sequences. The at least one zero-intensity or lower-intensity sequence was collected using the lowest intensity of all applied diffusion coding gradients or without using a diffusion coding gradient.

[0037] Other original dataset sequences were generated using a higher-intensity diffusion coding gradient. These sequences can also be referred to as “higher-intensity sequences”, which were generated using a diffusion coding gradient with an intensity greater than zero and higher than that of the zero-intensity or lower-intensity sequences.

[0038] Optionally, the original image dataset sequence can be repeatedly acquired to advantageously reduce noise; that is, the original image dataset sequence may include other original image data sequences with the same intensity or orientation. Therefore, using this preferred variant, at least one of the zero-intensity or low-intensity sequences and / or the high-intensity sequences can be repeatedly collected.

[0039] Each image raw dataset comprises a uniformly distributed set of lines in k-space with equivalent spatial frequency content, wherein the lines of each image raw dataset pass through the center of k-space and cover a continuous range of spatial frequencies, and the number of lines in each image raw dataset is chosen such that each image raw dataset is undersampled below a sampling rate constraint defined according to the Nyquist-Shannon sampling theorem (also known as the Whitaker-Kachchenkov-Shannon sampling theorem). Spatial coding using undersampled radial trajectories advantageously provides image raw data collected at high acquisition rates. The positions of the lines in each image raw dataset differ from those of adjacent cross-sectional slices.

[0040] Furthermore, the method of the present invention includes the following steps (first reconstruction step): performing a regularized nonlinear inverse reconstruction process on a sequence of original image datasets generated using zero-intensity or low-intensity diffusion-encoded gradients (i.e., zero-intensity or low-intensity sequences) to provide a sequence of coil sensitivities and MR images with no diffusion weighting or with low diffusion weighting (compared to the remaining MR images). Each of the coil sensitivities and MR images represents one of the cross-sectional slices, and each of the coil sensitivities and MR images is created by simultaneously estimating the sensitivity of the at least one receiver coil and the image content and based on the difference between the current estimate of the sensitivity of the at least one receiver coil and the image content and the estimate of the sensitivity of the at least one receiver coil and the image content of the adjacent cross-sectional slice.

[0041] Furthermore, the method of the present invention includes the following steps (second reconstruction step): performing a regularized linear inverse reconstruction process on all original image dataset sequences (i.e., all collected original image dataset sequences) having diffusion-encoded gradients of zero or low intensity and high intensity, to provide at least one MR image sequence with no diffusion weighting or low diffusion weighting and at least one MR image sequence with high intensity, each of the MR images representing one of the cross-sectional slices, and created by using the sensitivity of the at least one receiver coil determined for the same cross-sectional slice in the first reconstruction step and based on the difference between the current estimate of the image content and the estimate of the image content of adjacent cross-sectional slices. Preferably, at least three MR image sequences with high intensity are provided to obtain DW images in at least three orthogonal directions in space.

[0042] Advantageously, compared to

[10] and

[11] , all non-DW and DW images are obtained as solutions to a fully spatially regularized linear inverse problem. Specifically, in the first reconstruction step, the invention extends the regularization of the nonlinear inverse problem to directly adjacent images (with low and high spatial frequency components) and all their associated coil sensitivities. Furthermore, the linear inverse problem in the second reconstruction step involves regularization of directly adjacent images while using the coil sensitivities determined in the first reconstruction step. The resulting reconstructions can achieve a high degree of data undersampling, thereby improving acquisition speed. These reconstructions further achieve computational robustness and high spatial fidelity, and improve the quality of each cross-sectional image.

[0043] In fact, the numerical solutions in the first and second reconstruction steps formally employ a similar regularization originally developed for dynamic real-time MRI (

[13] and

[14] ), but now regularization is performed spatially rather than temporally. In other words, the present invention utilizes the similarity between consecutive directly adjacent images (first and second steps) and coil sensitivity (first step), thus achieving spatial fidelity of the similarity previously demonstrated for temporal fidelity when using temporal regularization of the preceding image and its coil sensitivity (

[15] ).

[0044] In other words, the nonlinear inverse reconstruction process is an iterative process in which a regularized linearization of the nonlinear MRI signal equation is solved in each iteration step, mapping the unknown spin density to be measured and its coil sensitivity to data acquired from the at least one receiver coil. Similarly, the linear inverse reconstruction process is an iterative process in which a regularized linear MRI signal equation is solved in each iteration step, mapping the unknown spin density to be measured to data acquired from the at least one receiver coil, while using the coil sensitivity determined in the nonlinear reconstruction process. In both cases, the iterative regularized Gauss-Newton method is preferably used to perform the nonlinear and linear inverse reconstructions. The inventors have discovered that inverse reconstruction processes employing the similarity of temporally continuous images of a given image plane as described in

[13] and

[14] can be used to reconstruct spatially continuous images of continuous cross-sectional slices, i.e., images of different imaging planes. This is a surprising result because it was not anticipated prior to this invention that adjacent cross-sectional slices would have sufficient similarity to successfully apply a regularized inverse reconstruction process, even if the object has a step change in the spin density. In contrast to

[13] and

[14] , the method of the present invention does not primarily provide a time-varying (dynamic) MR image sequence, but rather a spatially distributed (static) MR image sequence of the object, wherein the spatially distributed (static) MR images further have diffusion contrast.

[0045] Compared to

[11] and

[12] , the present invention relies on single-shot DW STEAM MRI acquisition combined with linear inverse reconstruction of all images through regularization of directly adjacent images. Therefore, the method according to the invention is suitable for improving spatial resolution because, compared to

[11] and

[12] , higher spatial frequencies are introduced into the nonlinearity and linear inversion in the first and second reconstruction steps, respectively. Advantageously, the present invention offers unique advantages for medical imaging and can be directly applied to various organ systems, particularly the human brain and prostate.

[0046] In summary, this invention provides a technique for volumetric DW single-excitation STEAM MRI that is free of sensitivity artifacts and offers higher image quality compared to previous methods. The invention acquires a series of non-DW and DW datasets of directly adjacent cross-sections that seamlessly cover the volume of the subject under investigation. It determines the optimal coil sensitivity by performing a nonlinear inverse reconstruction of the non-DW dataset in the first reconstruction step, while the primary reconstruction of all non-DW and DW images is accomplished as a solution to a linear inverse problem spatially regularized to adjacent images. This invention provides (i) volumetric DW MRI that is not sensitive to magnetic field inhomogeneities, thus avoiding geometric distortion, signal gaps and false positives, and negative signal variations opposite to DW EPI. Compared to previous DW STEAM MRI methods, this invention further provides (ii) a higher degree of data undersampling and thus reduced measurement time, (iii) higher computational stability, speed, and spatial fidelity, (iv) higher spatial resolution and SNR, (v) universal applicability to different organ systems, and (vi) clinically feasible scan times in the case of medical imaging.

[0047] According to a second general aspect of the invention, the above-mentioned object is achieved by an MRI device configured to create diffusion-weighted MR image sequences of subjects under investigation, and comprising an MRI scanner and a control device. The MRI scanner includes a main magnetic field device, at least one radio frequency excitation coil, three magnetic field gradient coils, and at least one radio frequency receiver coil. According to the invention, the control device, like a computer device, is adapted to control the MRI scanner to collect the plurality of raw image dataset sequences and reconstruct the MR image sequences using a method described in the first aspect of the invention or an embodiment thereof.

[0048] According to a preferred embodiment of the present invention, the method for creating a DW MR image sequence further includes the step of: performing a quantitative map calculation on the MR image sequence for the subject under investigation, including pixel-level calculation of at least one of the following: average diffusion-weighted MRI signal averaged in the gradient direction; representation of diffusion characteristics, especially the diffusion tensor; the trajectory of the diffusion tensor; the apparent diffusion coefficient; fractional anisotropy; and representation of the diffusion direction, especially the main diffusion direction. The calculation of the quantitative map has unique advantages for medical imaging because it provides a basis for individual sequential diagnosis.

[0049] According to another preferred embodiment of the invention, the reconstruction process of at least one of the first reconstruction step and the second reconstruction step includes a filtering process to suppress image artifacts. Advantageously, filtering can improve the image quality. In a particularly preferred variant, the filtering process includes applying a spatial filter, especially a nonlocal mean filter for each MR image.

[0050] Another advantage of the present invention is that the raw image data can be selected by a high degree of undersampling, i.e., relative to a fully sampled reference (e.g., for using a rotating straight line and radial encoding according to the sampling theorem), which is derived from the product of n / 2 and the number of data samples per line. The degree of undersampling can be at least 5 times, especially at least 10 times, thereby accelerating data acquisition in the same manner as real-time MRI (e.g., see

[14] ). Thus, the number of lines in each raw image dataset can be reduced. Specifically, for medical imaging, it has been found that a number of lines equal to or less than 30, especially equal to or less than 20, in each raw image dataset is sufficient to obtain high-quality MR image sequences.

[0051] According to another preferred embodiment of the invention, the lines in each original image dataset can be selected such that the lines in consecutive original image datasets are rotated relative to each other by a predetermined angular displacement. As an advantage, this rotation improves the regularization effect in the image reconstruction.

[0052] According to another preferred embodiment of the invention, the intensity of the diffusion coding gradient with zero or lower intensity is equal to or less than 100 s mm. -2 The b-value, especially zero. When the diffusion coding gradient strength is below this limit, the coil sensitivity calculated using the first reconstruction step is advantageously used in the second reconstruction step, improving signal quality. As a further advantage, the resulting images (sequences) without diffusion weighting or with low diffusion weighting provide a better reference for combination with corresponding images (sequences) with high diffusion weighting to calculate quantitative maps (sequences) of various diffusion features, such as the apparent diffusion coefficient, which is crucial for diagnostic purposes, for example, in medical MR imaging.

[0053] In contrast to the lower-intensity diffusion coding gradient, the higher-intensity diffusion coding gradient preferably has an intensity equal to or greater than b = 200 s mm. -2 The value of b, especially equal to or greater than b = 800s mm -2 The value of b, and equal to or less than b = 3000s mm -2 The b-value. For example, up to 1000 mm. -2 The b value is typically used in standard neuroimaging applications [1].

[0054] In another preferred embodiment of the invention, a radio frequency pulse and magnetic field gradient module for frequency-selective saturation is used before collecting the raw dataset for each image. Advantageously, the frequency-selective saturation can achieve a specific contrast according to the imaging task.

[0055] Advantageously, the method of the present invention for reconstructing MR image sequences can be implemented during and / or immediately after the collection of raw image data using at least one radio frequency receiver coil of the MRI device. In this case, providing the series of raw image datasets includes the following steps: arranging the object in the MRI device including the at least one receiver coil, applying the DW single-excitation STEAM sequence to the object, and collecting the series of raw image dataset sequences using the at least one receiver coil. Reconstruction of the MR image sequences can be rapidly completed after the raw image data collection is completed.

[0056] According to an alternative embodiment, the method of the present invention for reconstructing the MR image sequence can be performed independently of collecting the raw image data under predetermined measurement conditions. In this case, for example, the raw image dataset can be received from a data storage device (such as a data cloud storage device) and / or by transmitting data from a remote MRI device. Attached Figure Description

[0057] Further details and advantages of the invention will be described with reference to the accompanying drawings, which schematically illustrate:

[0058] Figure 1 : A combined diffusion-weighted spin echo and single-excitation stimulated echo sequence with diffusion-coded gradient for raw data acquisition in a preferred embodiment of the MR imaging method according to the present invention;

[0059] Figure 2 Preferred embodiments of the MRI device according to the present invention;

[0060] Figure 3 : A flowchart illustrating the features of a preferred embodiment of the MR imaging method according to the present invention;

[0061] Figure 4 A flowchart of the first reconstruction step of the MR imaging method according to the present invention is shown;

[0062] Figure 5 A flowchart of the second reconstruction step of the MR imaging method according to the present invention is shown; and

[0063] Figure 6 and Figure 7Examples of clinical applications of the DW MRL of the present invention.

[0064] Preferred Implementation

[0065] Preferred embodiments of the present invention will now be described with particular reference to the data stream of the original data acquisition and reconstruction process, the basic components of the MRI device of the present invention, and practical application examples.

[0066] As disclosed in

[13] or

[14] , design details of the k-space trajectory and mathematical formulas and implementations for inverse reconstruction are preferably provided. Specifically, as disclosed in

[13] , a regularized nonlinear inverse reconstruction process is preferably implemented to reconstruct the time series of MR images of the subject under investigation. Therefore, the entire contents of

[13] are incorporated herein by reference, especially all details concerning the image reconstruction of the cross-sectional MR image sequence of the subject under investigation in the first reconstruction step. All process steps applied to the time series of the original dataset and the MR image sequence in

[13] can be applied in the same manner to the slice sequence of the original dataset generated using zero-intensity or low-intensity sequences. The linear inverse reconstruction process implemented in the second reconstruction step follows the nonlinear inverse reconstruction process but uses the coil sensitivity determined in the first reconstruction step. In both cases, numerical optimization is accomplished by an iterative regularized Gauss-Newton method with spatial regularization of adjacent cross-sections. The linear inverse reconstruction process is not limited to applying the regularized Gauss-Newton method. Alternatively, other known numerical methods can be used to solve the inverse problem.

[0067] Further details of the MRI apparatus, the numerical implementation of the mathematical formulas using available software tools, and optional additional image processing or computational steps for providing quantitative maps are not described to the extent known in conventional MRI techniques. Furthermore, the following description exemplarily refers to parallel MR imaging, where the raw image data includes MRI signals received using multiple radio frequency receiver coils. It should be emphasized that the application of the invention is not limited to parallel MR imaging, but can even be performed using a single receiver coil.

[0068] Figure 1 The use of schematically illustrates Figure 2The illustrated embodiment of the MRI device 100 applies and collects a combined diffusion-weighted spin echo and single-excitation stimulated echo sequence with diffusion-coded gradients. The MRI device 100 includes an MRI scanner 10 having a main magnetic field device 11, at least one radio frequency excitation coil 12, three magnetic field gradient coils 13, and a radio frequency receiver coil 14. The main magnetic field device 11 generates a static main magnetic field parallel to the longitudinal axis (e.g., z-axis) of the MRI scanner aperture. The object to be investigated 1 is housed in the MRI device 100. Furthermore, the MRI device 100 includes a control device 20 adapted to control the MRI scanner 10 to collect a sequence of raw image datasets and utilize... Figures 3 to 5 The illustrated embodiment's step S1 reconstructs the MR image sequence. The control device 20 includes at least one GPU 21, which is preferably used to implement... Figure 4 and Figure 5 The nonlinear and linear inversion algorithms shown are illustrated.

[0069] according to Figure 3 A preferred embodiment of the method of the present invention includes a raw data acquisition step S1 of providing multiple raw image dataset sequences collected using a DW single-excitation STEAM sequence, a first reconstruction step S2 of performing regularized nonlinear inverse reconstruction on zero-intensity or low-intensity sequences, a second reconstruction step S3 of performing regularized linear inverse reconstruction on all non-DW and DW image sequences using at least one coil sensitivity of zero-intensity or low-intensity sequences with the same cross-sectional slice, and a calculation step S4 of calculating a quantitative map.

[0070] Therefore, a portion of the method of the present invention includes reconstructing a cross-sectional STEAM MR image sequence from multiple slices that have no diffusion coding gradient or have a low diffusion coding gradient, wherein the STEAM MR image and associated coil sensitivity are obtained through a nonlinear inverse reconstruction process that regularizes adjacent images and their coil sensitivities (first reconstruction step S2). Another portion of the method of the present invention includes performing linear inverse reconstruction on all non-DW and DW STEAM MR images of the same slice by regularizing adjacent images (second reconstruction step S3). This step can be applied in conjunction with as many different diffusion coding gradients as possible, with zero, low, or high intensity as required by the specific application (e.g., gradients with different intensities or orientations, or gradients with similar intensities for improving SNR).

[0071] Raw data collection

[0072] To provide raw data (especially for collecting raw data), multiple directly adjacent slices 2 with constant slice thickness were defined within the volume of the investigated object 1, such as an organ of a patient. Figure 2The illustrations are schematically shown. The number and thickness of slice 2 are selected based on the size of the organ, the slice thickness used, the measurement time, and / or the imaging resolution to be acquired. As a practical example of human brain imaging, approximately 50 slices with a slice thickness of 3 mm were defined. For each slice, a different DW single-shot STEAM sequence was applied, as described below.

[0073] In the raw data acquisition step S1 of the method of the present invention, multiple raw image dataset sequences are provided. These sequences are collected together with the DW single-shot STEAM sequences, such as... Figure 1 As shown. Each image in the original dataset represents one slice. For object 1 (e.g., a patient's tissue or organ), use... Figure 1 The DW single-shot excitation STEAM sequence shown encodes MRI signals received using radio frequency receiver coil 14. The portion in parentheses is repeated for each line in k-space. The DW single-shot excitation STEAM sequence is constructed to collect data samples along a non-Cartesian k-space trajectory. A slice shift A equal to the slice thickness is achieved by varying the radio frequency excitation pulse.

[0074] Figure 1 Each DW single-shot STEAM MRI sequence combines a leading spin echo sequence (with or without a diffusion-coded gradient) with a single-shot STEAM MRI readout sequence. Figure 1 The image shown is used to obtain Figure 6 and Figure 7 The clinically relevant version shown refers to a preferred example having a pair of unipolar diffuse-coded gradients (shaded lines) in a dual-pulse spin echo sequence (SE = spin echo) and a single-shot stimulated STEAM MRI sequence (STE = stimulated echo, G...). s =Slice selection gradient, RF = radio frequency pulse, dashed line = spoiler gradient) radial encoded gradient (G) x G y Typically, spread-coded spin echo modules may have different variations, such as including more complex gradient waveforms or employing two or more refocusing pulses.

[0075] In the actual implementation of the present invention Figure 1The described DW single-excitation STEAM sequence is applied n times. For example, one of the original dataset sequences (zero-intensity or low-intensity sequence) is not generated using the diffusion-coded gradient from the leading spin echo sequence, while other original dataset sequences are generated using higher-intensity diffusion-coded gradients. An implementation with a minimum number of repetitions (n=4) represents an application using only one zero-intensity or low-intensity diffusion-coded gradient, as well as an application using three diffusion-coded gradients with higher intensities and different directions. However, the number n can also correspond to hundreds of repetitions, for example, when using 128 gradient directions and multiple averages to achieve advanced post-processing applications such as fiber tracing imaging.

[0076] To avoid ambiguity regarding directional differences in diffusion characteristics, clinical applications of DW MRI typically use three or six diffusion-encoding gradients with different orientations to measure DW images in at least three directions (but preferably six). Such protocols are used to calculate the average DW image across the diffusion-encoding directions to obtain an image with "isotropic" (i.e., direction-independent) diffusion weighting.

[0077] Furthermore, the clinical application of DW MRI typically relies on two or more different gradient intensities (e.g., 0 and 1000 s mm). -2 The b-value is a feature. This type of information, along with DW images with multiple gradient directions, is used to compute orientation-independent average DW images, diffusion tensors, corresponding mappings of ADCs, and other diffusion features (e.g., fractional anisotropic mappings). For specific applications of DW MRI, more than two gradient intensities can be applied, for example, up to 10.

[0078] To improve SNR, zero-intensity or low-intensity raw data sequences and / or high-intensity raw data sequences of the original dataset can be repeatedly collected. The number of repetitions can be selected based on the application conditions of the MRI method.

[0079] Image Reconstruction

[0080] First, through a regularized nonlinear inverse reconstruction process ( Figure 3 The first reconstruction step S2) in the process obtains a sequence of MR images with zero or low diffusion weighting. This process jointly estimates each image and its associated coil sensitivity while taking advantage of the spatial similarity with directly adjacent images and their coil sensitivities (e.g., using an iteratively regularized Gauss-Newton method).

[0081] Unlike

[13] , each original dataset of images represents another in a continuous cross-sectional slice 2, such as Figure 2 The schematic illustration is shown in the figure.

[0082] Figure 4The data stream is summarized after the raw data acquisition step S1 (providing the measured raw data) and the optional preprocessing step S1.1 (preprocessing the measurement data), which includes the first reconstruction step S2 as described in

[13] .

[0083] Using the preprocessing step S1.1, gradient delay correction S1.1A is performed on the original image data (where the correction k-space line may have an unexpected shift from the k-space center), and interpolation step S1.1B is performed (where non-Cartesian data is interpolated onto a Cartesian grid). Steps S1.1A and S1.1B can be implemented as disclosed in

[13] .

[0084] Subsequently, in the first reconstruction step S2, the coil sensitivity and MR image sequence of object 1 are reconstructed through the regularized nonlinear inverse reconstruction process described in

[13] . Starting from the initial inference of the MR image and coil sensitivity of the first cross-sectional slice S21, each MR image is created by iteratively estimating the sensitivity of the receiver coil and the image content S22. Step S22 includes nonlinear inverse reconstruction using the iterative regularized Gauss-Newton method, which includes a convolution-based conjugate gradient algorithm S23. The number of iterations (Newton steps) is selected according to the image quality requirements of the specific imaging task. Finally, the reconstructed coil sensitivity sequence is output for further processing in the second reconstruction step (S25).

[0085] As a result of step S2, a sequence of cross-sectional STEAM MR images and associated coil sensitivities are obtained, each representing one of slices 2, wherein the STEAM MR images and associated coil sensitivities are generated without using a diffusion coding gradient or with a lower diffusion coding gradient. Preferably, for further processing in step S3, only the coil sensitivities are used, while the MR images obtained in step S2 are discarded.

[0086] Subsequently, as Figure 5 As shown, execute Figure 3 The second reconstruction step S3 is shown. As a result of the first reconstruction step S2, the composite coil sensitivity of each cross-sectional image is provided in step S31. Figure 5 As shown in step S32, the coil sensitivity is used to compute all STEAMMR image sequences with arbitrary diffusion weights using spatially regularized linear inversion. The linear inverse reconstruction also employs an iteratively regularized Gauss-Newton method, with the number of iterations (Newton steps) selected based on the image quality requirements of the specific imaging task.

[0087] In step S32, a regularized linear inverse reconstruction process is performed on each original image dataset sequence with a diffusion coding gradient of zero or low intensity and on each original image dataset sequence with a diffusion decoding gradient of high intensity. Accordingly, processing at least one original image dataset sequence with a diffusion coding gradient of zero or low intensity results in at least one MR image sequence having no diffusion weighting or low diffusion weighting. Furthermore, processing at least one original image dataset sequence with a diffusion coding gradient of high intensity results in at least one MR image sequence having high intensity.

[0088] In other words, regardless of its diffusion weighting, all MR image sequences are calculated through linear inversion and regularization of adjacent images (the coil sensitivity sequence is fixed, such as...). Figure 3 (as shown in step S2). Each MR image in each sequence represents one of the cross-sectional slices. Each MR image is created by using the sensitivity of at least one receiver coil for the same cross-sectional slice (linear inversion) and based on the difference between the current estimate of the image content (i.e., the current raw image dataset) and the estimate of the image content of the adjacent cross-sectional slice (i.e., the previous raw image dataset (regularized)). As a result, the MR images to be acquired are provided in step S33 for further processing.

[0089] The nonlinear and linear inverse reconstructions of the first few images, performed in the first reconstruction step S2 and the second reconstruction step S3 respectively, do not fully benefit from spatial regularization. Specifically, for the inverse reconstruction of the first image in each sequence, there is no prior information, i.e., no images from adjacent slices. This problem can be circumvented in several ways. First, the output and display of the first few (e.g., 5) images can be discarded. Second, the data acquisition can be extended to include several additional slices beyond the target volume to be covered. Or third, at the cost of reconstruction time, the reconstruction can be initialized using the last image from the initial reconstruction and its coil sensitivity, and the reconstruction can be repeated in reverse order.

[0090] In step S1, zero-intensity or low-intensity sequences and high-intensity sequences from the original dataset can be repeatedly collected to improve the SNR. Multiple MR images with the same diffusion weighting can be averaged for further processing. Similarly, other steps for the conventional processing, storage, display, or recording of image data can be followed. Using zero-, low-, or high-diffusion weighting to compute additional MR images has advantages for calculating the map in step S4 (see below).

[0091] For medical imaging, the MR image sequence obtained through reconstruction step S3 is used to calculate the quantitative map of the subject. Figure 3Step S4 in the process. The calculation of the quantitative map involves pixel-level calculations of various diffusion properties, such as the average diffusion-weighted MRI signal and the apparent diffusion coefficient averaged along the gradient direction. Step S4 is performed in a manner known from conventional techniques.

[0092] Experimental Example

[0093] The experimental examples of the invention are described below with particular reference to applications in the field of medical imaging. All examples involve studies on healthy human subjects.

[0094] Figure 6 An example representing a clinically relevant application of the present invention's DW MRI in the human brain is shown. This example summarizes three selected cross sections (from top to bottom) of a normal human brain DW single-emission STEAM MRI study, where a total of 51 diffusion-weighted and diffusion-weighted images (b=0 and 1000 s mm) of directly adjacent cross sections were acquired within a 2 minute 30 second measurement time. -2 (6 diffusion directions). The results show the non-DW image (left), the average DW image, and the calculated ADC map (right). Regardless of whether the sensitivity differences of the transverse brain slices represented by the images are small (top row) or strong (bottom row), none of the images show the inhomogeneity artifacts (i.e., geometric distortion and signal variation) commonly found in DWEPI measurements.

[0095] The same findings also apply to abdominal scans. Figure 7 Selected results from a single-shot DW-excitation STEAM MRI study of the prostate in normal subjects are presented, a study corresponding to another area of ​​high clinical importance. In this case, non-DW images, mean DW images (b=0 and 600s mm) are included. -2 ADC maps (with 6 gradient directions) and three cross-sections (and) Figure 6 The tissues in the samples were identical and were selected from a total of 31 directly adjacent cross sections collected during a measurement period of 4 minutes and 30 seconds.

[0096] The application of this invention is not limited to medical imaging (as described in the examples above), but can also be used to image other objects such as workpieces or other technical objects.

[0097] The features of the invention disclosed in the foregoing specification, drawings and claims are of individual, combined or sub-combination importance for implementing the invention in different embodiments.

Claims

1. A method for creating multiple diffusion-weighted magnetic resonance (MR) image sequences of a subject under investigation, wherein each of the MR image sequences represents a series of consecutive cross-sectional slices covering a volume of the subject under investigation, the method comprising the steps of: (a) Provides a plurality of raw image dataset sequences, said plurality of raw image dataset sequences being collected using at least one radio frequency receiver coil of a magnetic resonance imaging (MRI) device, wherein - Each raw dataset of images includes the image content of one of the magnetic resonance images to be reconstructed; - Each image's original dataset refers to one of the cross-sectional slices; - Each image raw dataset includes multiple data samples generated using a combined diffusion-weighted spin echo and single-excitation stimulated echo sequence with a diffusion-coded gradient, the combined sequence spatially encoding the magnetic resonance imaging signal received through the at least one radio frequency receiver coil using a non-Cartesian k-space trajectory. - The diffusion coding gradients of the plurality of original image dataset sequences have at least two different intensities and at least three different directions, wherein one of the original dataset sequences is generated using a diffusion coding gradient of zero or low intensity, and the other original dataset sequences are generated using a diffusion coding gradient of higher intensity; - Each image's original dataset consists of a set of uniformly distributed lines in k-space with equivalent spatial frequency content; - The line in each image's original dataset passes through the center of the k-space and covers a continuous range of spatial frequencies; - Select the number of lines in the original dataset for each image such that the original dataset for each image is undersampled under the sampling rate limit defined by the Nyquist-Shannon sampling theorem; as well as - The position of the line in each image raw dataset is different from that of the adjacent cross-sectional slices in the image raw dataset; (b) Perform a regularized nonlinear inverse reconstruction process on the sequence of raw image datasets generated using diffusion-encoded gradients of zero or low intensity to provide a sequence of coil sensitivity and magnetic resonance images with or without diffusion weighting, each of the images representing one of the cross-sectional slices, and create the image by simultaneously estimating the sensitivity of the at least one receiver coil and the image content based on the difference between the current estimate of the sensitivity of the at least one receiver coil and the image content and the estimate of the sensitivity of the at least one receiver coil and the image content of the adjacent cross-sectional slice; as well as (c) Perform a regularized linear inverse reconstruction process on all sequences of original image datasets with zero or low intensity and high intensity diffusion coding gradients to provide sequences of magnetic resonance images with no diffusion weighting or low diffusion weighting and sequences of magnetic resonance images with high intensity, each of the magnetic resonance images representing one of the cross-sectional slices, and created by using the sensitivity of the at least one receiver coil determined in step (b) for the same cross-sectional slice and based on the difference between the current original image dataset and the previous original image dataset of the adjacent cross-sectional slice.

2. The method according to claim 1, further comprising the following step: (d) Performing quantitative mapping calculations on the magnetic resonance image sequence for the investigated subject, including pixel-level calculations of at least one of the following: - Average diffusion-weighted magnetic resonance imaging signal averaged along the gradient direction; - Representation of diffusion properties; - The trajectory of the diffusion tensor; - Apparent diffusion coefficient; - Fractional anisotropy; as well as - Indication of diffusion direction.

3. The method according to any one of the preceding claims, wherein - The reconstruction process in at least one of steps (b) and (c) includes a filtering process to suppress image artifacts.

4. The method according to claim 3, wherein The filtering process includes applying a spatial filter to each magnetic resonance image.

5. The method according to claim 1, wherein - Select the number of lines in the original dataset for each image such that the resulting undersampling level is at least 5 times.

6. The method according to claim 1, wherein - The number of lines in each image's original dataset is at most 30.

7. The method according to claim 1, wherein - Select the number of lines in each image raw dataset such that the lines of the image raw dataset of adjacent cross-sectional slices are rotated relative to each other by a predetermined angular displacement.

8. The method according to claim 1, wherein - The intensity of the diffusion-encoded gradient with zero or lower intensity is equal to or less than 100 s / mm. -2 The value of b.

9. The method according to claim 1, wherein - Before collecting the raw dataset for each image, use the radio frequency pulse and magnetic field gradient modules for frequency-selective saturation.

10. The method of claim 1, wherein the original image dataset sequence is provided by at least one of the following: - The object is arranged in the magnetic resonance imaging apparatus including the at least one receiver coil, the combined diffusion-weighted spin echo and single-excitation stimulated echo sequence is applied to the object, and the raw image dataset sequence is collected using the at least one receiver coil; and - Receive the sequence of raw image datasets by transmitting data collected from a remote magnetic resonance imaging device.

11. The method of claim 2, wherein The diffusion property is represented by the diffusion tensor.

12. The method according to claim 2, wherein The diffusion direction is represented by the main diffusion direction.

13. The method of claim 3, wherein The filtering process includes applying a nonlocal mean filter to each magnetic resonance image.

14. The method of claim 1, wherein - Select the number of lines in the original dataset for each image such that the resulting undersampling level is at least 10 times.

15. The method of claim 1, wherein - The number of lines in each image's original dataset is at most 20.

16. The method of claim 1, wherein - The intensity of the diffusion-encoded gradient with zero or lower intensity is zero.

17. A magnetic resonance imaging apparatus configured to create a diffusion-weighted magnetic resonance image sequence of a subject under investigation, comprising: - A magnetic resonance imaging scanner, comprising a main magnetic field device, at least one radio frequency excitation coil, three magnetic field gradient coils, and at least one radio frequency receiver coil, and - A control device adapted to control the magnetic resonance imaging scanner to collect the plurality of raw image dataset sequences and reconstruct the magnetic resonance image sequences using the method of any one of claims 1 to 16.

Images