Diffusion magnetic resonance imaging (dMRI), based on diffusion nuclear magnetic resonance (NMR), is a non-invasive imaging modality that provides information about the architecture of any physical structure, in which the spin-carrying atoms or molecules can diffuse over a certain time set by the NMR measurement. Such physical structures can include, but are not limited to, composite materials, continuous media, random media, porous media, porous rocks, and biological tissues. For example, biological tissue architecture restrics the random motion of water molecules, which makes dMRI of water protons an essential part of any clinical or research brain MRI protocol. Typical experimental settings probe such motion at a scale of micrometers or tens of micrometers, orders of magnitude below MRI imaging resolution (see, e.g., Kiselev, 2017; Novikov et al., 2019; and Alexander et al., 2019). Hence, tissue microstructure imaging with dMRI can become sensitive, and possibly specific, to developmental, aging and disease processes that originate at this scale and could provide biomarkers of said processes (see, e.g., Assaf, 2008; and Jelescu & Budde, 2017). Herein, both diffusion NMR and diffusion MRI can be referred to as dMRI, given that diffusion NMR can be seen as a particular case of dMRI when imaging is not performed, but diffusion encoding methodology is similar for the purposes of the present disclosure.
The information content of the dMRI signal depends on the level of coarse-graining (see, e.g., Novikov et al., 2019) over the diffusion length controlled by the diffusion time t. With sufficiently long t approximately 30-50 ms or greater, typically used in the clinic, each tissue compartment (e.g., intra-axonal space, extra-axonal space within each fiber fascicle, etc) can be asymptotically considered as fully coarse-grained, such that diffusion in it becomes approximately Gaussian. Assuming Gaussian diffusion in each compartment drastically simplifies biophysical modeling. Indeed, the signal Sαexp(−tr B Dα) (e.g., normalized to 1 in the absence of diffusion weighting) from any “Gaussian compartment” labeled by α and defined by its diffusion tensor Dα, becomes fully encoded by a 3×3 symmetric B-tensor (see, e.g., Westin et al., 2016; and Topgaard, 2017). The overall signal at clinically-relevant moderate diffusion weightings is then represented by the cumulant expansion (van Kampen, 1981; Kiselev, 2010).
where b=tr B is the dMRI b-value, see Materials and Methods for details. Here, Einstein's convention of summation over repeated Roman indices is assumed hereon, and
are components of the overall diffusion and covariance tensors D and C. Tensor D has 6 independent parameters, or “degrees of freedom” (dof), and C has 21 dof (see . . .
denote averages over the diffusion tensor distribution
(D) that characterizes the tissue—or, equivalently, over the compartment index α, and double brackets
. . .
denote the cumulants (see, e.g., van Kampen, 1981).
For a model-independent signal representation (1), which can be utilized even without assuming Gaussian diffusion in all compartments, a fundamental problem is to classify symmetries and define tensor invariants, i.e., combinations of tensor components that are independent of the choice of basis (such as, but not limited to, mean diffusivity). These rotational invariants facilitate the generation of basis-independent scalar maps. They can be the easiest to visualize (as opposed to multi-dimensional objects such as tensors), and to study as candidate markers of development, aging and disease.
Tensor invariants can serve as ideal hardware-independent “fingerprint” of a dMRI signal; their information content can underpin future classifiers of development, pathology and aging, for any tissue or organ, such as, but not limited to, brain. A practical problem is to relate the invariants to tissue properties, and to find ways of their fast and robust estimation.
Thus, it may be beneficial to provide exemplary systems, methods and computer-accessible medium can overcome at least some of the deficiencies described herein above.
To that end, it is possible to provide exemplary systems, methods and computer-accessible medium according to exemplary embodiments of the present disclosure, which can be referred to herein, as an example only, as exemplary systems, methods and computer-accessible medium utilizing and/or providing Rotational Invariants of the Cumulant Expansion (or “RICE” techniques/procedures, methods, systems and/or computer-accessible medium).
In these exemplary embodiments of the present disclosure, a full classification of rotational invariants of the cumulant expansion (RICE), i.e., invariants of diffusion tensor D and covariance tensor C (in total, 3+18 RICE invariants) can be provided, in terms of irreducible representations of the group of rotations, their geometric meaning can be elucidated, and the invariants can be related to the problem of addition of quantum-mechanical angular momenta.
In additional exemplary embodiments of the present disclosure, exemplary formulas can be provided and utilized for calculating of most or all RICE invariants, and connect them with tissue biophysics embodied by the distribution (D). It is possible to express conventional scalar contrasts—mean diffusivity (MD), fractional anisotropy (FA), mean kurtosis (MK), and microscopic fractional anisotropy (μFA)—in terms of some of the RICE invariants.
According to certain exemplary embodiments of the present disclosure, iRICE acquisitions can be provided based on the icosahedral directions with the smallest number of measurements required to determine, e.g., only the tensor elements that yield the above conventional contrasts in 1-2 minutes for the whole brain on a clinical scanner.
With the advent of precision medicine and quantitative imaging, tensor invariants can provide a parsimonious and hardware-independent “fingerprint” of a dMRI signal. Representing the signal's information content in terms of scalar invariant maps according to the underlying symmetries can improve and moreover underpin machine learning classifiers of brain pathology, development and aging, while fast iRICE protocols can enable translation of advanced dMRI into a clinical practice.
According to further exemplary embodiments of the present disclosure, exemplary systems, methods and computer-accessible medium can be provided for determining invariants associated with at least one physical structure. Using such exemplary embodiments, it is possible to receive at least one particular component which is a component of a diffusion tensor and/or a component of a covariance tensor. The particular component(s) can be associated with the physical structure(s). In such exemplary embodiments, it is also possible to generate the invariants of the diffusion tensor and/or the covariance tensor based on the particular component. For example, the physical structure can be, but not limited to, (i) a biological tissue, (ii) a composite material, (iii) a continuous medium, and/or (iv) a random medium, (v) porous media and/or (vi) porous rocks. The particular component can be based on diffusion magnetic resonance (dMR) image of the physical structure (which can be, but not limited to, at least one tissue). The invariants can associated with at least one parameters of such structure (e.g., the tissue).
In certain exemplary embodiments of the present disclosure, the diffusion tensor can be split into a scalar part of degree 0 and a symmetric trace-free (STF) part of degree 2. Alternatively or in addition, the covariance tensor can be split into a fully symmetric part and an asymmetric part. Alternatively or in addition, the fully symmetric part of the covariance tensor can be split into a first part of degree 0, a second part of degree 2, and a third part of degree 4. Alternatively or in addition, the asymmetric part can be split into a first part of degree 0 and a second part of degree 2.
According to additional exemplary embodiments of the present disclosure, the first parts of the diffusion tensor, of the fully symmetric part of the covariance tensor, and of the asymmetric part of the covariance tensor, respectively, can be used to generate of the invariants which can be proportional to full traces thereof. Alternatively or in addition, the second part of the diffusion tensor can be used to generate at least two of the invariants (e.g., intrinsic variants) which can be based on traces of second and third powers of such second part. Alternatively or in addition, the second part of the fully symmetric part of the covariance tensor can be used to generate at least two of the invariants which are based on traces of second and third powers of such second part. Alternatively or in addition, the second part of the asymmetric part of the covariance tensor can be used to generate at least two of the invariants which can be based on traces of second and third powers of such second part. Alternatively or in addition, the third part of the fully symmetric part of the covariance tensor can be used to generate at least four of the invariants which can be based on traces of second, third, fourth and fifth powers of such third part. Alternatively or in addition, the third part of the fully symmetric part of the covariance tensor can be used to generate at least two invariants based on traces of cubic powers of at least two eigentensors, that are determined from an eigentensor decomposition of such third part.
In yet additional exemplary embodiments of the present disclosure, eigenbases of the second part of the fully symmetric part of the covariance tensor, the second part of the asymmetric part of the covariance tensor, and the third part of the fully symmetric part of the covariance tensor can be used to generate at least one of the invariants of the covariance tensor, based on relative orientations of the eigenbases. Further, a first set of the invariants of the covariance tensor can be given by parameters of a rotation of the eigenbasis of the second part of the asymmetric part of the covariance tensor relative to the eigenbasis of the second part of the fully symmetric part of the covariance tensor. Alternatively or in addition, a second set of the invariants of covariance tensor can be given by parameters of a rotation of the eigenbasis of the third part of the fully symmetric part of the covariance tensor relative to the eigenbasis of the second part of the fully symmetric part of the covariance tensor. The eigenbasis of the third part of the fully symmetric part of the covariance tensor can be generated based on the eigenbasis corresponding to a largest eigenvalue of an eigentensor decomposition thereof.
According to further exemplary embodiments of the present disclosure, a kurtosis tensor can be generated based on the fully symmetric part of the covariance tensor, and kurtosis invariants can be generated based on the invariants of the fully symmetric part of the covariance tensor. Alternatively or in addition, it is possible to utilize the invariants to determine contrasts, which includes (i) mean, axial or radial diffusivity, (ii) fractional anisotropy, (iii) mean, axial and/or radial kurtosis, and/or (iv) microscopic fractional anisotropy. Alternatively or in addition, it is possible to utilize the invariants to determine contrasts which includes (i) isotropic variance, and/or (ii) anisotropic variance.
In yet further exemplary embodiment of the present disclosure, it is possible to generate compartmental tensor covariances associated with the tissue parameters. The compartmental tensor covariances can include size-size covariance, shape-shape covariance, and size-shape covariance. It is additionally possible to generate a size-shape correlation.
According to further exemplary embodiments of the present disclosure, it is possible to utilize parameter maps (e.g., RICE maps) as inputs to machine-learning or artificial-intelligence classifiers, such as, but not limited to, neural networks, trained for an automatic pathology detection and staging. One of the advantages of such parameter maps as inputs for machine-learning classifiers can be in their independence of the hardware and measurement protocol. In such exemplary manner, it is possible to combine training data from multiple sites, scanner manufacturers, field strengths, and other differing tissue-independent characteristics, that could otherwise limit the effectiveness of training of such classifiers.
According to certain further exemplary embodiments of the present disclosure, exemplary systems, methods and computer-accessible medium can be provided for determining invariants associated with at least one physical structure. Using such exemplary embodiments, it is possible to receive information related to the at least one diffusion magnetic resonance (dMR) image of the at least one physical structure. It is then possible to generate the invariants of a diffusion tensor and/or covariance tensors using (i) a particular number of acquisitions and/or particular directions, and (ii) the information.
For example, in an exemplary embodiments, the particular number and the particular directions of the diffusion acquisitions can correspond to spherical designs. The spherical designs can correspond to at least one number of directions that can be configured to form a spherical design. The spherical designs can be provided by half of the octahedron vertices, half of icosahedron vertices, all octahedron vertices, or all icosahedron vertices. It is also possible to generate the mean diffusivity and the mean kurtosis based on an un-weighed diffusion image, 6 icosahedron vertices for a first b-shell, and/or 6 icosahedron vertices for a second b-shell. The first b-shell can be selectable as approximately b=1 ms/m{circumflex over ( )}2, and the second b-shell can be selectable as approximately b=2 ms/m{circumflex over ( )}2.
In yet another exemplary embodiment of the present disclosure, it is possible to generate the mean diffusivity, fractional anisotropy, the mean kurtosis, and/or microscopic fractional anisotropy based on an un-weighed diffusion image, 6 icosahedron vertices for the first b-shell, 6 icosahedron vertices for the second b-shell, and/or a spherical tensor encoding (STE) acquisition at the third b-value. The first b-shell can be selectable as approximately b=1 ms/m{circumflex over ( )}2. The second b-shell can be selectable as approximately b=2 ms/m{circumflex over ( )}2. The third b-value (STE) can be selectable as approximately b=1.5 ms/m{circumflex over ( )}2.
According to certain further exemplary embodiments of the present disclosure, exemplary systems, methods and computer-accessible medium can be provided for determining at least one component of at least one tensor associated with at least one physical structure. Using such exemplary embodiments, it is possible to receive first information related to at least one diffusion magnetic resonance (dMR) image of the physical structure(s), receive second information related to at least one constraint on the component(s) of at least one tensor. It is then possible to generating the component(s) which is/are a component of a diffusion tensor and/or a component of a covariance tensor based on the first information and the second information.
In a further exemplary embodiment of the present disclosure, the constraint(s) can be for the value of a part of the tensor(s) of degree 4. The value of the part of the tensor(s) of degree 4 can be zero. The component(s) of the covariance tensor can be estimated in a symmetric trace-free (STF) basis. Alternatively or in addition, the constraint(s) can be formulated for the STF basis coefficients of degree 4. Alternatively or in addition, the constraint(s) can be on a further part of the tensor(s) of degree 2 to be axially symmetric. Alternatively or in addition, the constraint(s) can be on the part and the further part of the tensor(s) can be aligned therebetween or with an eigenbasis of the diffusion tensor.
These and other objects, features and advantages of the exemplary embodiments of the present disclosure will become apparent upon reading the following detailed description of the exemplary embodiments of the present disclosure, when taken in conjunction with the accompanying claims.
Further objects, features and advantages of the present disclosure will become apparent from the following detailed description taken in conjunction with the accompanying Figures showing illustrative embodiments of the present disclosure, in which:
Throughout the drawings, the same reference numerals and characters, unless otherwise stated, are used to denote like features, elements, components or portions of the illustrated embodiments. Moreover, while the present disclosure will now be described in detail with reference to the figures, it is done so in connection with the illustrative embodiments and is not limited by the particular embodiments illustrated in the figures and the appended claims.
To obtain all 21 dof of C, the necessary (but not sufficient) condition can be rank B>1 (see, e.g., Westin et al., 2016). This should be contrasted with diffusion kurtosis imaging (DKI) (see Jensen et al., 2005) yielding only S, the fully symmetric part of C (see, e.g.,
with 15 dof, as it involves only linear tensor encoding (LTE, rank B=1) provided by PGSE (see, e.g., Stejskal and Tanner, 1965). In this exemplary case,
and similarly for all n! permutations of indices for n-th order tensors (see, e.g., Thorne, 1980).) Estimation of higher-order cumulants (see, e.g., Ning et al., 2021) is challenging due to a small convergence radius of the cumulant series (see, e.g., Kiselev & Il'yasov, 2007).
The number of nonzero eigenvalues of the B tensor generally reflects how many dimensions of the diffusion process are being probed simultaneously. The requirement rank B>1 for probing the full C tensor can mean that some of its dof are describing diffusion along more than one dimension. According to the exemplary embodiments of the present disclosure, without the loss of generality, we focus on axially symmetric B (see, e.g.,
parametrized by its trace b giving the overall scale; the unit vector ĝ along its symmetry axis; and the dimensionless shape parameter β (see, e.g., Eriksson et al., 2015). Compared to conventional linear tensor encoding (LTE, β=1), according to the exemplary embodiments of the present disclosure, varying the extra degree of freedom β changes the B-tensor shape, e.g., β=0 for spherical encoding (STE, isotropic B-tensor), and β=−1/2 for planar encoding (PTE, two equal nonzero eigenvalues) (see, e.g., Cory et al., 1990; Mitra, 1995; Mori & Van Zijl, 1995; Shemesh et al., 2015). Non-axially symmetric B-tensors are not necessary for accessing all (b2) information, and are typically not employed.
The complementary approach based on representations of the SO(3) group of rotations (see, e.g., Backus, 1970; Itin & Hehl, 2013; Itin & Hehl, 2015) has so far led to the following. The tensor C=S+A is decomposed (see, e.g., Backus, 1970) into a fully symmetric part S (Equation 3 above), and its asymmetric (not antisymmetric) complement A, as shown in
Tensor S (and kurtosis tensor, Equation (3)) further splits into irreducible representations of SO(3) with =0, 2, 4:
where S(0) is fully isotropic (=0) and defined by one dof (the full trace Siijj of S), whereas S(2) and S(4) are symmetric trace-free (STF) tensors that are parametrized by 5 and 9 dof, respectively, with the total dof count 15=1+(2·2+1)+(2·4+1) (see Materials and Methods).
Importantly, tensor A (6 dof=21−15) is equivalent to a symmetric 3×3 tensor (see, e.g., Itin & Hehl, 2013):
where δij is Kronecker's delta, and ϵijk is the fully antisymmetric Levi-Civita tensor. Therefore, A splits into irreducible representations with =0 and 2,
Hence, non-LTE B-tensor shapes (see, e.g., Westin et al., 2016) probe the A tensor. Specifically, STE is sensitive to the isotropic (=0) part A(0), while more general B-tensor shapes, such as PTE, also probe the
=2 part A(2), as shown in
The correspondence between symmetric trace-free (STF) tensors and spherical harmonics (SH) (see, e.g., Eq. (19) herein below) indicates that, if the tensor glyph S({circumflex over (n)})=Sijkl ninjnknl is drawn, where n is a unit vector, then the S(0) part gives the directional average of S({circumflex over (n)}) (a ball). The S(2) part is responsible for a “single-fiber” glyph parametrized by five Y2m({circumflex over (n)}) (turning the ball into an ellipsoid, much like D({circumflex over (n)})=Dijninj in diffusion tensor imaging, or DTI). The S(4) part, parametrized by nine Y4m({circumflex over (n)}), depicts multiple lobes coming from fiber crossings, as shown in
Among the 5 dof of S(2), 3 absolute angles can define the orientation of the glyph. The remaining 2 dof parametrize the three eigenvalues that sum up to zero, as tr S(2)=0. Thereby, the 3 eigenvalues of S(0)+S(2) (one coming from S(0) and two from S(2)) are the invariants that determine the semi-axes of an ellipsoid.
For S(4), the picture is more complex. The 3 absolute angles again determine the orientation of the object Sijkl(4)ninjnknl, while the remaining 6 dof are the invariants determining its shape (see
Similarly to S(0)+S(2) or DTI examples above, the A tensor can describe a “single-fiber” glyph, with its 1+2 invariants parametrizing the semi-axes of the corresponding ellipsoid, and the remaining 3 dof determining its orientation in space. In particular, for axially symmetric B-tensors, Equation (4), the contribution of the A tensor to the second term in the cumulant expansion (1),
can be expressed as a scalar ∝App=(3/2) Aiikk (full trace), plus an ellipsoid parametrized by the tensor (7a). As expected, Eq. (9) vanishes for LTE (β=1). For STE, only the first (scalar) term survives, whereas PTE is the cleanest encoding to probe the A tensor ellipsoid: the first term vanishes, and the second term yields (b2/12) Apqgpgq.
The representation theory (see, e.g., Backus, 1970; Itin & Hehl, 2013; and Itin & Hehl, 2015), while elegant, has so far not been tied to constructing the invariants. Here we reconstruct all invariants of C by combining the characteristic polynomial approach (see, e.g., Betten, 1987; and Basser & Pajevic, 2007) with the representation theory (see, e.g., Backus, 1970; and Itin & Hehl, 2013; Itin & Hehl, 2015). Namely, according to exemplary embodiments of the present disclosure, it is possible to construct the invariants for each representation of
separately (see
It is then possible to complement these intrinsic invariants with the mixed ones defining the relative orientations between irreducible components of C with >0.
For example, the total number of invariants for any tensor equals to its number of dof minus 3 absolute rotation angles (see, e.g., Ghosh et al., 2012), yielding 3 for D (DTI), 12 for S or W tensors (DKI), and 18 for C. Applying this argument to each irreducible representation, the number of intrinsic invariants is 1 for =0, and (2
+1)−3=2(
−1)=2, 6, . . . for
=2, 4, . . . . The relative angles between irreducible components likely do not change upon rotation of the basis (the tensor transforms as a whole). Therefore, the 21 dof of C are split between (1+2+6)+(1+2)=12 intrinsic invariants according to decomposition (10); 3+3=6 mixed invariants (relative angles between components A(2), S(2), and S(4)); and 3 absolute rotation angles, as shown in
Eq. (2) can be reviewed through the lens of representation theory. In particular, the right-hand side of (2) involves a direct product of two compartment tensors, which can symbolically denote D⊗D (dropping compartment index α for brevity). Each D=D(0)+D(2) can be split into =0 and 2 representations (e.g., by going into the STF tensor basis, as shown in
correspond to the addition of two angular momenta with =0 or
=2.
As it is known from quantum mechanics (see, e.g., Tinkham, 2003), the addition of angular momenta +
yields all possible momenta between |
1−
2| and
1+
2. Mathematically speaking (see, e.g., Hall, 2015), the tensor product of representations
1 and
2 is reducible, and splits into a direct sum of irreducible representations with
=|
1−
2|, . . . ,
1+
2.
The successive terms in Eq. (12) yield: 0⊗0=0; 0 ⊗2=2 ⊗0=2; 2⊗2=0⊕2⊕4. In the latter case, representations with =1 and 3 forbidden by parity were not included. Hence, the right-hand side of Eq. (2) is a direct sum of representations 0⊕2⊕0⊕2⊕4. This is precisely what we get from Eq. (10) representing the left-hand side of Eq. (2).
Thus, importantly, according to the exemplary embodiments of the present disclosure, the representations of SO(3) group can dictate the relations between distinct components of the C tensor and tissue properties. Furthermore, the coefficients with which the STF component covariances (see
For example, not all 3+18 RICE invariants are equally important. According to an exemplary embodiments of the present disclosure, it is possible to identify, e.g., 7 of the invariants (e.g., 2 from D and 5 from C) which are related to previously studied contrasts. Such invariants all correspond to L2-norms of the representations of C and D, Eqs. (10) and (11). These RICE invariants can be called the main RICE invariants, and a special notation for them can be introduced, as they play a unique and/or important role in synthesizing dMRI contrasts:
Interestingly, it is possible to express conventional scalar contrasts—mean diffusivity (MD), fractional anisotropy (FA), mean kurtosis (MK), and microscopic fractional anisotropy (μFA)—in terms of D0, D2, S0, and A0:
Further, axial and radial kurtosis for an axially-symmetric fiber tract can be expressed via D0, D2, S0, S2 and S4, without the need to rotate to the basis aligned with the tract. Further, it is possible to also identify a previously unexplored contrast coming from a combination of S(2) and A(2) elements, which can be used to define the size-shape correlation index of compartmental tensors.
Table 1 shows the exemplary minimal instant/icosahedral RICE (iRICE) protocols to obtain MD, FA, MK, based on measurements with only 12 B tensors; and MD, FA, MK, μFA based on measurements with only 15 B tensors, in accordance with certain exemplary embodiments of the present disclosure. These numbers can notably be fewer than 21=6+15 or 27=6+21 necessary to determine all tensor components. These acquisition schemes can involve fewer measurements than was previously anticipated.
As described herein, it is possible to use spherical designs (special directions on a sphere) which can, e.g., guarantee that relevant tensor components are estimated, whereas the inessential ones (that do not contribute to the above main invariants) are canceled.
A systematic derivation of all rotational invariants of the cumulant expansion (RICE) from an irreducible decomposition of the diffusion and covariance tensors can be provided, in accordance with the exemplary embodiments of the present disclosure, as shown in
For example, =0; 2 for
=2; and 6 for
=4). Together with these 1+2+1+2+6=12 intrinsic invariants, C has 9 basis-dependent absolute angles defining the orientations of its S(2), S(4), and A(2) components via the rotation matrices
S
S
A
S
S
A
With RICE, we uncover multiple diffusion contrasts that have been overlooked so far. We separate all rotational invariant information contained in D and C tensors, as shown in the exemplary maps of )2, e.g., are the only ones whose combinations have been studied in the literature, since these together are equivalent to standard DTI+DKI+μFA contrasts. Furthermore, from these it is also possible to define a novel size-shape correlation (SSC) invariant, in accordance with the exemplary embodiments of the present disclosure. Indeed,
describing mixed invariants are parametrized by Euler angles of the bases of S(4) and A(2) sectors relative to the eigenbasis of S(2). These angles correspond to intrinsic (active) rotations of the S(2) frame along z by {tilde over (α)}, then along new x′ by {tilde over (β)}, and along new z″ by {tilde over (γ)} (equivalent to the product of extrinsic active rotations
({tilde over (α)},{tilde over (β)},{tilde over (γ)})=X({tilde over (α)})Y({tilde over (β)})Z({tilde over (γ)}) in the fixed S(2) frame to obtain the S(4) and A(2) frames). Underlying tissue microstructure introduces correlations between invariants. For example, the near-zero relative angles {tilde over (β)} in the white matter tracts exemplify the alignment of the eigen frames from different representations with the local tract orientation.
=2 and
=4 invariants relative to
=0, together with histograms containing gray and white matter voxels. It is evident that the contribution of high-order information in D, W, and A decreases with the degree e for both tissues. This is more pronounced in gray matter and less so in crossing fibers or highly aligned WM regions, e.g., the corpus callosum. In the exemplary maps of
, since axially symmetric tensors satisfy |
|/
≡1 for all
. The above shows that D(2) and W(2) have a high axial symmetry, while the opposite holds for W(4) and A(2). For axially symmetric tensors these ratios equal 1. In the kurtosis tensor, the 9 elements associated with
=4 have a 5-10 times smaller intensity than the trace. This empirical observation about the smallness of S(4) component of the fully-symmetric part S of covariance tensor can be used for making the estimation of components of D, S and C tensors more robust, such as, but not limited to, imposing the constraint on the component S(4) to be small, as further discussed herein.
The bottom rows of
Minimal and fully sampled acquisitions were thoroughly compared. Exemplary maps for MD+FA+MK are shown in
The main exemplary RICE maps, ∝tr1/2()2, are related to conventional contrasts, Eqs. (36)-(39), without assumptions on the tissue diffusion tensor distribution
(D). An advantage of RICE maps over other signal representations can be that they belong to distinct irreducible representations of rotations, and thus, represent “orthogonal” (complementary) contrasts up to
(b2). RICE parameters thereby can be used to represent a hardware- and measurement-independent “fingerprint” of the tissue (or, more generally, of a physical structure under consideration) in a mutually complementary way, without any further assumptions. RICE maps (i.e., RICE parameters determined for a number of voxels in an organ or tissue) can be used as inputs to machine-learning or artificial-intelligence classifiers, such as, but not limited to, neural networks, trained for an automatic pathology detection and staging. The particular advantage of such RICE parameter maps as inputs for machine-learning classifiers is in their independence of the hardware and measurement protocol; this enables combining training data from multiple sites, scanner manufacturers, field strengths, and other differing tissue-independent characteristics, that could otherwise limit the effectiveness of training of said classifiers.
There is certain previously-unexplored information, complementary to μFA, in the non-symmetric part of C. It is related to the size-shape correlation of compartmental diffusion tensors. The complementary information needed for such contrast is present in the anisotropic part A(2) of A. This is inaccessible by LTE+STE acquisitions, although can be measured combining LTE and PTE, as shown in the diagram of
For example, not all RICE invariants may be equally important. Some can reflect more prominent signal features and thus show higher SNR maps. Furthermore, although exemplary RICE invariants can be mathematically independent, it was determined, according to the exemplary embodiments of the present disclosure, that some correlations exist between them. This highlights that brain tissue (b2) signals may not have independent 18 rotationally invariant dof. For example, in white matter, the mixed invariants provided in the exemplary maps of
The STF basis allowed the analysis of the different degrees of anisotropy in D, W (or S), and A. It was observed in both gray and white matter that the relative contributions from degrees =0,2,4 are not equally important. As
becomes higher, tensor elements become smaller (as shown in
=4 elements in kurtosis while it is practically absent in the diffusion tensor or in
=2 kurtosis elements. This facilitates a fast approximate computation of axial and radial diffusion/kurtosis without the need to project D and W onto the main eigenvector of D.
A comparison between fiber-basis projection maps (axial and radial diffusion/kurtosis) and axial symmetry approximations shows high agreement (see, e.g.,
Furthermore, the empirical observation of smallness of S4 (equivalently, of W4) as compared to =0 and 2 components, can lead to improvement of precision, accuracy and/or robustness of estimating the tensor components, by constraining all elements of S4m (equivalently, of W4m) to be small; and/or setting those elements to zero during estimation. For example, such smallness constraint can be imposed by using a corresponding training set with small or zero S4m, when learning the mapping between the measurements and the tensor components of D and C tensors.
Throughout the disclosure of the exemplary embodiments, MK can be referred to as, e.g., the trace of the kurtosis tensor, as defined in Eq. (37) (see, e.g., Hansen et al., 2013; Jespersen, 2018). Such definition has two main advantages over the definition using the average of the directional apparent kurtosis: First, it is a tensor derived quantity, the rotational invariant . Second, estimating it from noisy measurements is a better-posed inverse problem presenting fewer outliers (“black voxels”).
According to certain exemplary embodiments of the present disclosure, two acquisition protocols minimizing the number of directions for MD, FA, MK, and μFA can be provided. For example, using STF decompositions of the diffusion, kurtosis, and covariance tensors, we separate the contribution of each tensor element to the parametric maps of interest. By canceling unnecessary contributions, e.g., it is possible to reduce the number of free parameters, allowing the estimation of different contrasts combinations in a faster scan. The exemplary minimal acquisitions, in accordance with the certain exemplary embodiments of the present disclosure, can include, e.g., 6×LTEb=1+6×LTEb=2=12 DWIs for MD+FA+MK in a 1-minute scan, and 6×LTEb=1+3×STEb=1.5+6×LTEb=2=15 DWIs for MD+FA+MK+μFA in a 2-minute scan, see Table 2. For example, as discussed herein, DWI can refer to diffusion-weighted MR image, or measurement. The feasibility of exemplary minimal protocols is provided on a normal volunteer. The smallest spherical designs facilitating a joint fit are used but the robustness and precision of both minimal protocols can be enhanced by using spherical designs that have more directions, e.g., full icosahedrons for each shell.
Table 1 shows a comparison between conventional fast protocols and the minimal fast protocols, in accordance with the exemplary embodiments of the present disclosure. One publication (see, e.g., Hansen et al., 2013) suggested a fast DKI approach of 3×LTEb=1+9×LTEb=2=12 DWI, relying on a two-step fitting rather than a joint one. This is suboptimal due to nonzero contributions of W elements at b=1 ms/μm2, which bias MD and MK estimation. Nonetheless, the same or similar estimation approach can be applied to the framework, in accordance with the exemplary embodiments of the present disclosure, with only 1×STEb1+6×LTEb=2=7 DWI or 3×LTEb=1+6×LTEb=2=9 DWI. Similarly, another publication (see, e.g., Nilsson et al., 2020) proposed a 3-minute protocol for extracting MD, MK, and μFA. In such conventional approach, thicker slices and more measurements are used, but the same direction sets to compute spherical means. These were used to fit a moments-based approximation involving diffusion isotropic and anisotropic variances from which MK and μFA can be computed. However, this may not facilitate a simultaneous estimation of FA since such expression only holds for spherical means. Recently, a further publication (see, e.g., Kerkela et al., 2021) used Monte Carlo simulations to show that computing μFA from the estimation of the cumulants D and C has a minimum finite-b bias. Due to this, protocols, in accordance with the exemplary embodiments of the present disclosure, can facilitate a joint fit of the cumulants from directional signals rather than from spherical means.
In general, the dMRI signal depends on all sequence timings and the corresponding Larmor frequency gradient values g(t); mathematically, the signal is a functional of g(t), or, equivalently, of the encoding function q(t)=∫0tg(t′)dt′ (the anti-derivative of g(t)): S=S[q(t)]. Even for a conventional pulsed-gradient diffusion sequence (missing citation), one obtains a multi-dimensional phase diagram in the space of sequence parameters (see, e.g., Kiselev, 2021; and Novikov, 2021).
However, when each compartment is fully coarse-grained, the measurement is completely determined by the B-tensor with elements
calculated based on q(t). The distribution (D) of compartment tensors in a given voxel gives rise to the overall signal
where Dα are compartment diffusion tensors (see, e.g., Basser & Pajevic, 2003; Jian et al., 2007; and Glenn et al., 2015). Normalization ∫dD(D)=1 implies that the fractions add up to unity, Σαƒα=1.
Equation (16) can be the most general form of a signal from multiple Gaussian compartments. It is valid when the transient processes have played out, such that tensors Dα have all become time-independent, and thereby higher-order cumulants in each compartment are negligible (see, e.g., Novikov et al., 2019). In this case, the signal (16) is a function of the B-tensor: S[q(t)]→S(B), while tissue is fully represented by the distribution (D). This long-t picture of multiple Gaussian compartments (anisotropic and non-exchanging) underpins a large number of dMRI modeling approaches, in particular, the Standard Model (SM) of diffusion (see, e.g., Novikov et al., 2019) and its variants (see, e.g., Jespersen et al., 2007; Jespersen et al., 2010; Fieremans et al., 2011; Zhang et al., 2012; Sotiropoulos et al., 2012; and Jensen et al., 2016). Furthermore, this picture contains the SM extension onto different fiber populations in a voxel, lifting the key SM assumption of a single-fascicle “kernel” (response).
Given the forward model (16), an inverse problem is to restore ?(D) from measurements with different W. This problem is a matrix version of the inverse Laplace transform and is therefore ill-conditioned. Since in clinical settings, typical encodings are moderate (tr BDα˜1), the inverse problem can be formulated term-by-term for the cumulant expansion (1) of the signal (16).
The higher-order signal terms in Eq. (1) couple to successive cumulants of ?(D). The inverse problem maps onto finding the cumulants Di
(tensors of even order 2n) from a set of measurements. This becomes obvious by noting the analogy B→iλ with the standard cumulant series (see, e.g., van Kampen, 1981) ln
for the characteristic function p(λ)=∫dx e−iλxp(x) of a probability distribution p(x), such that the n-th term in Eq. (1) is
Hence, the B-tensor lowers the order by half: The 2n-th diffusion-displacement cumulant (the 2n-th order term from expanding ln S[q(t)] in q(t)) maps onto the n-th cumulant Di
of
(D) corresponding to the n-th order of expansion of InS(B) in B. The number of dof for this cumulant equals to that for a fully symmetric order-n tensor of dimension d=6, which is a number of assignments of n indistinguishable objects into d distinguishable bins: (n+d−1)!/n!(d−1)!=(n+5)!/n!5!=6, 21, 56, 126, . . . for n=1, 2, 3, 4, . . . .
The exemplary method of analysis, in accordance with the exemplary embodiments of the present disclosure, can utilize the representation theory of SO(3), the group of rotations in 3 dimensions (see, e.g., Tinkham, 2003; and Hall, 2015). A d-dimensional representation of a group is a mapping of each element (rotation) onto a d×d matrix that acts on a d-dimensional vector space. Representation theory provides a way to split a complex object (such as tensor D or C) into a set of independent simpler ones with certain symmetries, on which a group acts. In particular, the elements of an irreducible representation transform among themselves, and hence can be reviewed separately.
All irreducible representations of SO(3) are labeled by integers =0, 1, 2, . . . , and have dimension 2
+1. Each representation is equivalent to a set of 2
+1 symmetric trace-free (STF) tensors of order
(with
=0 corresponding to a scalar), and also to a set of 2
+1 spherical harmonics (SH)
(ĥ) with m=−
, . . . ,
(Thorne, 1980). Hence, there is a 1-to-1 correspondence between SH and STF tensors.
All in dMRI context are even due to time-translation invariance of the Brownian motion dictating even parity
(−{circumflex over (n)})=
({circumflex over (n)}). Hence, each cumulant or moment tensor, as in Eq. (1), can be split into a direct sum of irreducible representations with even e, connecting it with the orientation dispersion in the SH basis (see, e.g., Novikov et al., 2018; and Pozo et al., 2019). The STF-SH equivalence can be a useful tool to construct and provide geometric meaning to the invariants.
The tensors D, S, and A can be separated into irreducible parts, cf. Eqs. (11), (6) and (8) correspondingly. In components,
Thus, generate SH and form the standard STF tensor basis (see, e.g., Thorne, 1980). For L>
, it is possible to provide the basis tensors by symmetrization with kronecker symbols,
which generate standard spherical harmonics
Inverting Eq. (17), the STF components are related to their Cartesian counterparts via
where
and * denotes complex conjugation.
The above STF/Cartesian correspondence can be generalized to any order-L fully symmetric tensor S, which can be decomposed into a linear combination of its trace and STF tensors:
where in the last equation, the sum over l and m is implied, and =
is the degree-
component. The general mapping from Cartesian,
, to STF basis,
, can be computed from (Thorne, 1980).
Two intrinsic invariants for =2 can be obtained from the characteristic equation
since the coefficients of this cubic polynomial are rotationally invariant. It is also possible to express these invariants from tr (S(2))n for n=2,3 (note that tr S(2)=0):
which have a 1-to-1 mapping to the eigenvalues of S(2)
Intrinsic invariants for fourth-order tensors are more intricate since the characteristic equation can be written with more dof than in the second-order case. For the covariance or elasticity tensor this can be written as (Betten, 1987)
To solve this problem, certain publications (see, e.g., Betten, 1987; Basser & Pajevic, 2007) and others proposed to map the elasticity fourth-order 3D tensor to a second-order 6D tensor. For C→C6×6, it is also possible to use the mapping from (see, e.g., Basser & Pajevic, 2007):
The characteristic polynomial in Eq. (26) has degree 6 in λ and degree 3 in μ (Betten, 1987). From these coefficients, it is possible to extract invariants but these mix different irreducible representations. In the exemplary approaches utilized according to the exemplary embodiments of the present disclosure, in contrast to previous approaches, we use Eq. (26) to solve a more constrained problem: finding the intrinsic invariants of S(4), which has only 9 independent parameters. One of the advantages of the exemplary approaches can be that the obtained invariants all correspond to a particular symmetry, i.e., to the irreducible representation of SO(3) group of degree =4, which makes these invariants complementary to those coming from other irreducible representations. The corresponding characteristic polynomial is sixth-order in λ and first-order in μ. Although all these coefficients are rotationally invariant, e.g., only 4 are algebraically independent. These can be found setting μ=0 and they have a 1-to-1 mapping with the eigenvalues of S(4)6×6, or equivalently, tr (S(4))n, n=1, . . . , 6. It is noted that
Initially, it is possible to think that the characteristic equation of S(4) provides its 6 intrinsic invariants: tr (S(4))n, n=1, . . . , 6. This is not the case. Since each tensor element satisfies the characteristic equation, it is possible to prove that tr (S(4))2, tr (S(4))3, tr (S(4))4, tr (S(4))5 determine all traces of higher powers of S(4).
The remaining 2 intrinsic invariants of S(4) can be obtained from its eigentensor decomposition (see, e.g., Basser & Pajevic, 2007)
where λa and Eij(a) are the eigenvalues and eigentensors of S(4). This can be analogous to an eigenvalue decomposition of S(4)6×6 akin to Eq. (26) with μ=0. In the present case, e.g., λa are the eigenvalues and v(a) [vxx(a), vyy(a), vaa(a), vxy(a), vxz(a), vyz(a)] are the normalized eigenvectors, from which the following can be build
Eigentensors satisfy Eij(a)Eij(b)=Sab. From direct inspection of S(4)6×6 (as described herein) it can be seen that the eigentensor associated with a zero eigenvalue, λa
To obtain the remaining two invariants of S(4), the eigentensors should be utilized. The following can be defined:
Traces of powers of E and E are rotationally invariant. As for any second-order tensor, only traces of linear, quadratic and cubic powers of these matrices are algebraically independent. Further, by construction, tr E=tr {tilde over (E)}=0. One can also check that tr E2=5 and tr =tr (S(4))2, which is given by one of the previously found invariants. Therefore, it is possible to identify the two remaining independent intrinsic invariants of S(4) with tr E3 and tr
, without the loss of generality.
Within the present approach in accordance with the exemplary embodiments of the present disclosure, it becomes straightforward to assign meaning to tr (S(4))n, since these can be seen as shape tensor metrics of increasing order (n=2, . . . , 5). For tr E3 and tr this is less intuitive. Thus, one exemplary alternative can be, e.g., to take any pair of eigentensors, say Eij(6) and Eij(5) (where λ6 is the largest and λ5 is the second largest eigenvalue), and compute the 3-dimensional rotation matrix between their bases,
5,6, as shown in
5,6 is parametrized by two degrees of freedom φ and ψ (rotation angles). These exemplary (e.g., two) angles have a nontrivial 1-to-1 correspondence to the two invariants constructed above, tr E3 and tr
.
An alternative exemplary way to compute the main exemplary RICE invariants can be the following scalars for =0, and the L2-norms over m for l>0:
It is possible to present compartmental diffusion tensors as, e.g.:
where the STF basis allows a more intuitive separation of the degrees of freedom. The =0 component represents the trace of the tensor (size) and the
=2 components represent the shape in a given reference frame, see
Using Eqs. (5), (7), and (20), it is possible to compute the STF decomposition of D, S, and A, and group them according to their degree l:
where ij2m′
jk2n′
ki2m′ and
ij2m′
kl2n′
ijkl4m′ are proportional to the Clebsch-Gordan coefficients <2, m′, 2, n′|2, m> and <2, m′, 2, n′|4, m>, respectively. For example, the relation between the C tensor and the covariances of compartment diffusion tensors maps onto the addition of angular momenta (as discussed herein). The full exemplary (non-limiting) system can have, e.g., 27 independent linear equations with 27 unknowns, although all of them are not needed to generate some of the typical diffusion contrasts.
Conventional Contrasts from RICE
Interestingly, the first four equations in Eq. (35) are decoupled from the others. With D0 and D2, it is possible to compute:
where Vλ(D)=(1/2)D22 is the variance of the eigenvalues of D. This latter relation becomes evident if it is realized that in the eigenbasis D(2)=diag(λ1−
(see, e.g., Novikov et al., 2018).
For mean kurtosis, it is possible to work with exemplary tensor derived metrics rather than average apparent kurtosis (see, e.g., Lu et al., 2006; Hansen et al., 2013; and Jespersen et al., 2017):
Further, it is possible to extract the variance of the eigenvalues of Dα averaged over the diffusion tensor distribution. For this, it is possible to repeat the process outlined above for each compartment α and then take the average:
and use it to compute μFA (see, e.g., Westin et al., 2016; and Szczepankiewicz et al., 2016) without the need to assume axial symmetry in Dijα or assuming a functional form for (D):
Inverting the complete system in Eq. (35), together with Eq. (20), facilitates a determination for the covariances of STF components starting from the conventional Cartesian ones, Eq. (2), Cijkl→the Cartesian expressions of C (see, e.g., Basser & Pajevic, 2007; Westin et al., 2016; and Magdoom et al., 2021), the STF basis can facilitate a deeper understanding of the rotationally invariant information.
Often, a main fiber population in a voxel is assumed, thus, D and W can be projected to the main axis v1, yielding axial and radial projections:
If D and W possess axial symmetry around this main fiber population, e.g., it is possible to compute the above maps without projecting onto v1 and directly from RICE maps:
According to certain exemplary embodiments of the present disclosure, all contrasts above are independent of A2m elements, which contain previously unexplored information. Following the =2 part of the system in Eq. (35), it is possible to determine the 1×5 covariances <<D00αD2mα>>:
and define a novel invariant with size-shape correlation (SSC) information, complementary to DTI-DKI-μFA:
This contrast contains information about the correlation of the sizes and shapes of the microscopic compartments in a voxel. The normalization is chosen akin to correlation coefficients, such that SSC∈[0, 1], where SSC=0 for independent shapes and sizes and SSC=1 for a linear relationship. Due to the norm taken over m on the numerator, SSC can only take only positive values.
Spherical designs (see, e.g., Seymour & Zaslavsky, 1984) are sets of N points {{circumflex over (n)}i}i=1N∈2 on the unit sphere ∥{circumflex over (n)}∥=1 that for any rotation of the points scheme satisfy
where ƒ00/√{square root over (4π)} is the spherical mean of ƒL({circumflex over (n)}), which can be any function with finite degree L when expanded in spherical harmonics:
The smallest spherical designs for L=2 and L=4 are provided by tetrahedron and icosahedron vertices, N=4 and N=12, respectively. For functions with ƒL({circumflex over (n)})=ƒL(−{circumflex over (n)}), it is possible to further reduce N to half of the octahedron vertices for L=2 (the N=3 cyclic permutations of {circumflex over (n)}=(1,0,0)), and half of the icosahedron vertices for L=4 (the N=6 cyclic permutations of {circumflex over (n)}=1/√{square root over (1+φ2)} (1,±φ,0), where φ=(1+√{square root over (5)})/2).
the exemplary number of measurements of the minimal spherical designs can be much smaller than the total number of degrees of freedom in ƒL({circumflex over (n)}). ƒ4({circumflex over (n)}) has 15 degrees of freedom but only 6 measurements suffice for an unbiased computation of their spherical average. Thus, in accordance with the exemplary embodiments of the present disclosure, it is possible to use spherical designs as a minimal way to measure the isotropic part of ƒL({circumflex over (n)}), ƒ00, since they cancel contributions for 0<
≤L.
When all b-shells are acquired using spherical designs, it is possible to represent the dependence on Bij of the spherical mean signal using only the traces of the cumulant tensors: D00, S00, and A00, Eq. (17). If the interest is only in measuring MD and MK, then two LTE shells with 4-designs (N=6) will suffice. This also provides sufficient measurements to fit D2m, elements, enabling the computation of FA. The signal expression to be fit for such a protocol would be:
which only has 8 free parameters and can be robustly estimated from one b0 and two N=6 distinct b-shells, totaling 12 DWI. From these, we can compute D0, W0, and D2, and obtain MD, FA, and MK following Eqs. 36-37.
A similar exemplary procedure can be provided if it is desired to measure μFA. Here, a single STE measurement sensitive to (b2) must be added to the previous protocol to provide simultaneous sensitivity to A00 and insensitivity to A2m. Hence, the signal becomes:
which has 9 free parameters that can be estimated from one b0 and 13 DWI. Thus, it is possible to access D0, D2, S0, and A0, which provide us with MD, FA, MK, and μFA. Due to potential spurious time dependence of STE, e.g., 3 orthogonal rotations can be acquired.
For example, in both scenarios, it is possible to have more measurements than free parameters. However, being insensitive to high contributions greatly reduces the number of parameters affecting the signal (which have to be estimated), thereby pushing down the limit of minimum directions needed. Here, no assumptions are made on the shapes of D, W, or C. Table 1 contrasts theoretically minimal spherical designs against previous literature and our proposed protocols.
After providing inform consent, three healthy volunteers (23 year old female, 25 year old female, 33 year old male) underwent MRI in a whole body 3T-system (e.g., Siemens Healthcare, Prisma) using a 32-channel head coil. Maxwell-compensated free gradient diffusion waveforms were used to yield linear, planar, and spherical B-tensor encoding using a prototype spin echo sequence with EPI readout (Szczepankiewicz et al., 2019). Four diffusion datasets were acquired according to Table 2. Imaging parameters: voxel size=2×2×2 mm3, TR-4.2 s, TE=90 ms, bandwidth=1818 Hz/Px, RGRAPPA=2, partial Fourier=6/8, multiband=2. Total scan time was approximately 15 minutes per subject for all protocols.
All four protocols were processed identically and independently for each subject. Magnitude and phase data were reconstructed. Then, a phase estimation and unwinding step preceded the denoising of the complex images (see, e.g., Lemberskiy et al., 2019). Denoising was performed using the Marchenko-Pastur principal component analysis method (see, e.g., Veraart et al., 2016) on the real part of the phase-unwinded data. An advantage of denoising before taking the magnitude of the data is that Rician bias is reduced significantly. This data was also processed considering only magnitude DWI were acquired. Here, magnitude denoising and Rician bias correction were applied, these results can be found and described herein. Data was subsequently processed with the DESIGNER pipeline (see, e.g., Ades-Aron et al., 2018). Denoised images were corrected for Gibbs ringing artifacts accounting for the partial Fourier acquisition (see, e.g., Lee et al., 2021), based on re-sampling the image using local sub-voxel shifts. These images were rigidly aligned and then corrected for eddy current distortions and subject motion simultaneously (see, e.g., Smith et al., 2004). A b=0 image with reverse phase encoding was included for correction of EPI-induced distortions (see, e.g., Andersson et al., 2003). Further, DWI were locally smoothed based on similar spatial locations and signal intensities akin the method proposed by (see, e.g., Wiest-Daessle et al., 2007).
Four different variants of the cumulant expansion were fit to all four datasets described in Table 2. This depended on which parameters each protocol was sensitive to. The full DKI protocol was fit with a regular DKI expression. The minimal DKI using Eq. (48). The full RICE protocol with Eq. (1), and the minimal RICE protocol with Eq. (49). Weighted linear least squares were used for fitting (see, e.g., Veraart et al., 2013). All code for RICE parameter estimation were implemented in MATLAB (R2021a, MathWorks, Natick, Massachusetts).
Exemplary spherical designs can fulfill Eq. (45). One can check that for L=2, {n}={(1,0,0), (0,1,0), (0,0,1)}satisfies
For L=4, {{circumflex over (n)}}=1/√{square root over (1+<φ2)} {(1,φ,0), (0,1,φ), (φ,0,1), (1,−φ,0), (0,1,−φ), (−φ,0,1)}, where φ=(1+√{square root over (5)})/2:
Direct inspection of S(4)6×6 makes evident that: (i) trS(4)=0, (ii) One of the eigenvalues is zero (λa0=0) and its associated eigenvector is v(a0)=(1, 1, 1, 0, 0, 0)t.
With respect to the exemplary Full RICE method, in procedure 605, it is possible to acquire DWI using densely sampled general multishell q-space trajectory acquisition (such as, but not limited to, a combination of linear tensor encoding and planar tensor encoding). Then, in procedure 610, parameter estimation of O(b2) cumulant expansion can be performed. Further, in procedure 615, D and C can be transformed to SFT basis using symmetrization. Then, in procedure 620, D, W and A rotational invariants can be computed or otherwise determined. In procedure 625, contrasts can be computed or otherwise determined from invariants.
Turning to the exemplary FAST DKI/RICE method, in procedure 655, it is possible to acquire DWI using spherical designs (e.g., fast DKI or fast RICE). Then, in procedure 660, parameter estimation in STF basis omitting high order contributions can be performed. Thereafter, in procedure 665, rotational invariants for each order involved can be computed or otherwise determined. Further, in procedure 670, contrasts can be computed or otherwise determined from invariants.
As shown in
Further, the exemplary processing arrangement 705 can be provided with or include an input/output ports 735, which can include, for example a wired network, a wireless network, the internet, an intranet, a data collection probe, a sensor, etc. As shown in
The foregoing merely illustrates the principles of the disclosure. Various modifications and alterations to the described embodiments will be apparent to those skilled in the art in view of the teachings herein. It will thus be appreciated that those skilled in the art will be able to devise numerous systems, arrangements, and procedures which, although not explicitly shown or described herein, embody the principles of the disclosure and can be thus within the spirit and scope of the disclosure. Various different exemplary embodiments can be used together with one another, as well as interchangeably therewith, as should be understood by those having ordinary skill in the art. In addition, certain terms used in the present disclosure, including the specification, drawings and claims thereof, can be used synonymously in certain instances, including, but not limited to, for example, data and information. It should be understood that, while these words, and/or other words that can be synonymous to one another, can be used synonymously herein, that there can be instances when such words can be intended to not be used synonymously. Further, to the extent that the prior art knowledge has not been explicitly incorporated by reference herein above, it is explicitly incorporated herein in its entirety. All publications referenced are incorporated herein by reference in their entireties.
The following references are hereby incorporated by reference, in their entireties:
This application relates to and claims priority from International Application no. PCT/US2023/019622 filed on Apr. 24, 2023 and relates to and claims priority from U.S. Patent Application No. 63/333,856, filed on Apr. 22, 2022, the entire disclosure of which is incorporated herein by reference.
This invention was made with government support under Grant Nos. R01 NS088040 (NINDS), R01 EB027075 (NIBIB), awarded by the NIH, and the Center of Advanced Imaging Innovation and Research, a NIBIB Biomedical Technology Resource Center: P41 EB017183. The government has certain rights in the invention.
Number | Date | Country | |
---|---|---|---|
63333856 | Apr 2022 | US |
Number | Date | Country | |
---|---|---|---|
Parent | PCT/US2023/019622 | Apr 2023 | WO |
Child | 18923344 | US |