betaFeedback

WorksEngineering

Susceptibility source separation via Closed-form Alternating Newton-type Optimization

Abstract

Quantitative Susceptibility Mapping (QSM) provides a bulk magnetic susceptibility estimate per voxel, so contributions from paramagnetic and diamagnetic sources within the same voxel partially cancel. Susceptibility source separation methods identify these compartments by fitting a three-pool complex exponential model to multi-echo gradient-echo (GRE) data. However, conventional voxel-wise fitting with general-purpose constrained solvers requires several hours per whole-brain volume and often fails in high-susceptibility regions. We present CANON-QSM (Closed-form Alternating Newton-type Optimization for QSM), a solver that reformulates the subproblems of the three-pool model and solves them alternately via exact closed-form updates and Newton-type steps, yielding a substantial per-voxel computational speedup. Both in silico and in vivo evaluations demonstrate not only a massive reduction in reconstruction time, completing a whole-brain volume in under two minutes, but also improved accuracy and robustness against artifacts in high-susceptibility regions. Finally, an echo-time analysis of the three-pool model shows that source separation is driven by second-order nonlinearities in signal decay and phase evolution rather than first-order bulk estimates.

Introduction

Quantitative susceptibility mapping (QSM) estimates the magnetic susceptibility of tissue from the phase of gradient-echo (GRE) MRI [1, 2, 3] and is widely used as a biomarker of brain iron and myelin in aging and neurodegenerative diseases. Standard QSM calculates a single bulk susceptibility value per voxel under the assumption that the local phase evolves linearly with echo time (TE). In brain parenchyma, however, paramagnetic iron and diamagnetic myelin frequently coexist within the same voxel. Because their induced frequency shifts have opposite signs, they partially cancel, so the net bulk susceptibility can remain close to zero even when both sources are abundant, and a change in bulk susceptibility cannot be uniquely attributed to iron or myelin without additional information.

Susceptibility source separation resolves this ambiguity by exploiting the fact that both source types dephase spins and accelerate transverse relaxation, whereas their frequency shifts have opposite signs. For example, χ\chi-separation combines R2∗R_2^* and QSM with R2R_2 or R2′R_2' maps [4], requiring an auxiliary spin-echo acquisition that increases scan time.

Alternatively, DECOMPOSE-QSM [5] relies solely on multi-echo GRE data, modeling the complex signal of each voxel as the sum of paramagnetic, diamagnetic, and neutral pools with decay rates and frequency shifts derived from static dephasing theory [6].

A major practical obstacle to GRE-only source separation is its computational cost. DECOMPOSE-QSM fits each voxel by alternating over three parameter blocks, solving each block iteratively with a general-purpose constrained MATLAB solver [5]. This procedure requires several hours per brain [7], and fitting artifacts have been reported in large cerebral veins [8]. Although a deep-learning approximation has recently been proposed to accelerate inference [7], network predictions are not guaranteed to satisfy the physical forward model. In this work, we solve the three-pool model directly by reformulating each subproblem. Specifically, the pool-concentration and baseline-relaxation blocks admit exact, iteration-free closed-form solutions, while the susceptibility block is formulated using a phase residual that is insensitive to 2π2\pi wrapping, paired with an analytic Jacobian and closed-form 2×22 \times2 Newton-type (Levenberg-Marquardt) steps with a parallel multi-start strategy. We term this method Closed-form Alternating Newton-type Optimization for QSM (CANON-QSM). Because all subproblems reduce to elementary per-voxel operations, the solver executes voxel-parallel on a GPU, providing a massive computational speedup. We validate CANON-QSM on a realistic numerical brain phantom and in vivo, achieving whole-brain susceptibility separation in under two minutes.

Furthermore, we bridge a theoretical gap in the literature from a signal-processing perspective by expanding the three-pool complex signal in powers of echo time. We show that linear-phase QSM and mono-exponential R2∗R_2^* fitting capture only first-order echo-time quantities, whereas paramagnetic and diamagnetic sources are separated by second-order effects: the echo-time drift of the apparent susceptibility and the curvature of the log-magnitude decay. This analysis explains why per-echo QSM reconstructions are required as input.

Notation. All quantities refer to a single voxel unless stated otherwise. Let NeN_e denote the number of echoes, indexed by ee, and let t=(TE1,…,TENe)⊤\mathbf{t} = (\mathrm{TE}_1,\ldots,\mathrm{TE}_{N_e})^{\top} with 0<TE1<⋯<TENe0 < \mathrm{TE}_1 < \cdots< \mathrm{TE}_{N_e}. The symbols ⊙\odot and ⊘\oslash denote elementwise multiplication and division, and exp⁡\exp, log⁡\log, ∣⋅∣\lvert\cdot\rvert, and arg⁡\arg act elementwise, with arg⁡∈(−π,π]\arg\in(-\pi,\pi]. Here, vH\mathbf{v}^{H} denotes the conjugate transpose, vˉ\bar{\mathbf{v}} the complex conjugate, 1\mathbf{1} a vector of ones, and clip⁡[l,u](x)=min⁡{max⁡{x,l},u}\operatorname{clip}_{[l,u]}(x) = \min\{\max\{x,l\},u\}.

GRE signal and dipole deconvolution

Let Ω⊂R3\Omega\subset\mathbb{R}^{3} denote the imaging domain placed in a main static magnetic field B0=B0b^0\mathbf{B}_0 = B_0\hat{\mathbf{b}}_0 with unit direction b^0=(0,0,1)⊤\hat{\mathbf{b}}_0 = (0,0,1)^{\top} and gyromagnetic ratio γ\gamma. In standard QSM, a voxel at spatial position r∈Ω\mathbf{r} \in\Omega is characterized by a bulk magnetic susceptibility distribution χ(r)\chi(\mathbf{r}) (in ppm) and an effective transverse relaxation rate R2∗(r)≥0R_2^*(\mathbf{r}) \ge0. Under the linear mono-exponential model, the continuous-time complex GRE signal at echo time t≥0t \ge0 is given by [9, 10]

SGRE(r,t)=M0(r)e−R2∗(r)tei(ϕ0(r)+γB(r)t),(1)S_{\mathrm{GRE}}(\mathbf{r},t)=M_0(\mathbf{r})e^{-R_2^*(\mathbf{r})t}e^{i\left(\phi_0(\mathbf{r})+\gamma B(\mathbf{r})t\right)}, \tag*{(1)}

where M0(r)≥0M_0(\mathbf{r}) \ge0 is the equilibrium spin-density magnitude at t=0t=0, ϕ0(r)∈(−π,π]\phi_0(\mathbf{r}) \in(-\pi,\pi] is the transmit/receive radiofrequency phase offset, and B(r)B(\mathbf{r}) is the zz-component of the total magnetic field perturbation. Consequently, the unwrapped phase ϕ(r,t)\phi(\mathbf{r},t) of SGRE(r,t)S_{\mathrm{GRE}}(\mathbf{r},t) evolves linearly with time according to

ϕ(r,t)=ϕ0(r)+2πf(r)t,f(r)=γ2πB(r),(2)\phi(\mathbf{r},t)=\phi_0(\mathbf{r})+2\pi f(\mathbf{r})t,\qquad f(\mathbf{r})=\frac{\gamma}{2\pi}B(\mathbf{r}), \tag*{(2)}

where f(r)f(\mathbf{r}) (or, equivalently, the relative field shift ψtotal(r)=B(r)/B0\psi_{\mathrm{total}}(\mathbf{r})=B(\mathbf{r})/B_0) is the total frequency shift. This perturbation splits into a harmonic background field Bbg(r)B_{\mathrm{bg}}(\mathbf{r}), generated by susceptibility sources outside the region of interest Ω\Omega (such as air-tissue interfaces) that satisfies ΔBbg=0\Delta B_{\mathrm{bg}}=0 in Ω\Omega, and a local tissue field Bloc(r)B_{\mathrm{loc}}(\mathbf{r}) induced by tissue magnetization within Ω\Omega.

Under the Lorentzian sphere approximation, the relative local field perturbation ψ(r):=Bloc(r)/B0\psi(\mathbf{r}):=B_{\mathrm{loc}}(\mathbf{r})/B_0 is linked to the spatial susceptibility distribution χ(r)\chi(\mathbf{r}) via three-dimensional convolution with the unit magnetic dipole kernel d(r)d(\mathbf{r}) [1, 3]:

F(ψ)(k)=D(k)F(χ)(k),D(k)=13−(b^0⋅k)2∥k∥2(k≠0),(3)\mathcal{F}(\psi)(\mathbf{k})=D(\mathbf{k})\mathcal{F}(\chi)(\mathbf{k}),\qquad D(\mathbf{k})=\frac{1}{3}-\frac{\left(\hat{\mathbf{b}}_0\cdot\mathbf{k}\right)^2}{\lVert\mathbf{k}\rVert^2}\qquad(\mathbf{k}\ne0), \tag*{(3)}

with D(0)=0D(0)=0, where F\mathcal{F} denotes the continuous 3D Fourier transform. Standard QSM estimates a single scalar χ(r)\chi(\mathbf{r}) per voxel by unwrapping ϕ(r,t)\phi(\mathbf{r},t), removing ϕ0(r)\phi_0(\mathbf{r}) and Bbg(r)B_{\mathrm{bg}}(\mathbf{r}), and solving the ill-posed dipole deconvolution problem, whose Fourier kernel vanishes on the magic-angle cone.

The single-compartment model in (1) assumes a homogeneous sub-voxel frequency shift and a mono-exponential magnitude decay. In brain tissue, however, a single voxel typically contains water protons surrounding two distinct populations of sub-voxel magnetic perturbers: paramagnetic inclusions, dominated by iron stored in ferritin, and diamagnetic inclusions, dominated by myelin. Each inclusion distorts the microscopic magnetic field within the voxel. Protons precessing near an inclusion experience local field offsets and lose phase coherence, accelerating the transverse decay of the voxel signal, while the net microscopic field offset shifts its mean frequency. Because paramagnetic and diamagnetic inclusions induce frequency shifts of opposite signs that partially cancel in the net phase (2) while both accelerate R2∗R_2^* relaxation, the composite voxel signal departs from the mono-exponential model in (1).

Theory

Sub-voxel Susceptibility and Signal Model

Following [5], at each voxel we model the signal as the sum of three pools: paramagnetic (+), diamagnetic (−), and neutral (0). We aim to recover the concentrations c=(c+,c−,c0)⊤\mathbf{c} = (c^+, c^-, c^0)^\top, which represent the signal fractions at t=0t = 0 contributed by each pool, satisfying ck≥0c^k \ge0 and c++c−+c0=1c^+ + c^- + c^0 = 1, the intrinsic susceptibilities χ+≥0\chi^+ \ge0 and χ−≤0\chi^- \le0 (in ppm), collected in χ=(χ+,χ−)⊤\boldsymbol{\chi} = (\chi^+, \chi^-)^\top, and the baseline relaxation rate R≡R2,0∗R \equiv R_{2,0}^* of the neutral medium. In the static dephasing regime, where water diffusion during the echo train is negligible relative to the spatial scale of the field perturbations, a susceptibility source χ\chi adds a transverse decay rate a∣χ∣a\lvert\chi\rvert and a frequency shift pχp\chi to the signal of its pool [6].

The complex GRE signal of the voxel at time tt after excitation is therefore a sum of complex exponentials [5]:

S(t)=c+e−(aχ++R)t+ipχ+t+c−e−(−aχ−+R)t+ipχ−t+c0e−Rt,(4)S(t) = c^+ e^{-(a\chi^+ + R)t + ip\chi^+t} + c^- e^{-(-a\chi^- + R)t + ip\chi^-t} + c^0 e^{-Rt}, \tag*{(4)}

where

a=2π93γB0⋅10−6,p=23γB0⋅10−6.(5)a = \frac{2\pi}{9\sqrt{3}}\gamma B_0 \cdot10^{-6}, \qquad p = \frac{2}{3}\gamma B_0 \cdot10^{-6}. \tag*{(5)}

At 3 T, a≈323.5 s−1ppm−1a \approx323.5\ \mathrm{s}^{-1}\mathrm{ppm}^{-1} and p≈535.0 rad s−1ppm−1p \approx535.0\ \mathrm{rad}\ \mathrm{s}^{-1}\mathrm{ppm}^{-1}. Since χ+≥0≥χ−\chi^+ \ge0 \ge\chi^-, both aχ+a\chi^+ and −aχ−-a\chi^- equal a∣χ±∣a\lvert\chi^\pm\rvert, thus, the decay rate reports the magnitude of each perturbation and the frequency shift reports its sign. Both sources accelerate transverse decay, but they rotate the phase in opposite directions. The model has five independent real parameters, and each echo provides two real measurements, so at least three echoes are required.

We combine both effects of each source into the complex coefficients

u+=−a+ip,u−=a+ip,(6)u^+ = -a + ip, \qquad u^- = a + ip, \tag*{(6)}

so that u±χ±=−a∣χ±∣+ipχ±u^\pm\chi^\pm= -a\lvert\chi^\pm\rvert+ ip\chi^\pm, and (4) becomes

S(t)=e−Rt(c+eu+χ+t+c−eu−χ−t+c0).(7)S(t) = e^{-Rt}\left(c^+e^{u^+\chi^+t} + c^-e^{u^-\chi^-t} + c^0\right). \tag*{(7)}

The factor e−Rte^{-Rt} is shared by all pools, and the normalized model satisfies S(0)=1S(0) = 1.

The measured GRE phase also contains non-local phase contributions that (7) does not model, such as background fields from air-tissue interfaces outside the voxel and receiver phase offsets—so fitting the model directly to the raw phase would falsely attribute these terms to χ±\chi^\pm. As in [5], the measured phase is therefore replaced by the phase induced by the locally reconstructed susceptibility, yielding the synthesized local signal

y(t)=M(t)M0eipχQSM(t)t,(8)y(t) = \frac{M(t)}{M_0}e^{ip\chi^{\mathrm{QSM}}(t)t}, \tag*{(8)}

where M(t)M(t) is the measured GRE magnitude and χQSM(t)\chi^{\mathrm{QSM}}(t) is the QSM map reconstructed from the echo acquired at time tt. This procedure removes background fields, inverts the dipole convolution (3), and re-encodes the result via (8) using the same scaling constant pp, preserving the echo-time phase dependence exhibited by mixed voxels. The normalization factor M0M_0 is extrapolated per voxel so that ∣y(t)∣→1\lvert y(t)\rvert\to1 as t→0t \to0, matching S(0)=1S(0) = 1: M0=exp⁡(μ^+R^2∗TE)M_0 = \exp(\hat{\mu} + \hat{R}_2^* \mathrm{TE}) is the t=0t = 0 intercept of a log-linear fit of M(t)M(t), where μ^\hat{\mu} is the mean log-magnitude across echoes, TE is the mean echo time, and R^2∗\hat{R}_2^* (in s−1\mathrm{s}^{-1}) is the fitted mono-exponential decay rate. Note that y(t)y(t) is a derived observation rather than raw kk-space data, as its errors include those propagated from the QSM reconstruction, deviate from a complex Gaussian distribution, and may be spatially correlated across voxels.

Discrete problem

Sampling each pool of (4) at the acquired echo times gives the basis vectors

b±=[e(−R+u±χ±)TE1⋮e(−R+u±χ±)TENe]=exp⁡((−R+u±χ±)t),b0=exp⁡(−Rt).(9)\mathbf{b}^{\pm} = \begin{bmatrix} e^{(-R+u^{\pm}\chi^{\pm})\mathrm{TE}_1}\\ \vdots\\ e^{(-R+u^{\pm}\chi^{\pm})\mathrm{TE}_{N_e}} \end{bmatrix} = \exp\left((-R+u^{\pm}\chi^{\pm})t\right), \qquad \mathbf{b}^{0} = \exp(-Rt). \tag*{(9)}

Collecting the unknowns in θ=(c,χ,R)\boldsymbol{\theta} = (\mathbf{c}, \boldsymbol{\chi}, R), the discrete model signal vector s(θ)=[S(TEe)]e\mathbf{s}(\boldsymbol{\theta}) = [S(\mathrm{TE}_e)]_e is

s(θ)=c+b++c−b−+c0b0=Bc,B=[b+ b− b0]∈CNe×3,(10)\mathbf{s}(\boldsymbol{\theta}) = c^+\mathbf{b}^+ + c^-\mathbf{b}^- + c^0\mathbf{b}^0 = \mathbf{B}\mathbf{c}, \qquad \mathbf{B} = [\mathbf{b}^+\ \mathbf{b}^-\ \mathbf{b}^0] \in\mathbb{C}^{N_e \times3}, \tag*{(10)}

defined on the feasible parameter set

Θ={(c,χ,R):c≥0, 1Tc=1, 0≤χ+≤χmax⁡, −χmax⁡≤χ−≤0, 0≤R≤Rmax⁡}.(11)\Theta= \{(\mathbf{c}, \boldsymbol{\chi}, R) : \mathbf{c} \ge0,\ \mathbf{1}^{\mathsf{T}}\mathbf{c} = 1,\ 0 \le\chi^{+} \le\chi_{\max},\ -\chi_{\max} \le\chi^{-} \le0,\ 0 \le R \le R_{\max}\}. \tag*{(11)}

By (7), s=e−Rt⊙β\mathbf{s} = e^{-Rt} \odot\boldsymbol{\beta}, where β=c+eu+χ+t+c−eu−χ−t+c01\boldsymbol{\beta} = c^{+}e^{u^{+}\chi^{+}t} + c^{-}e^{u^{-}\chi^{-}t} + c^{0}\mathbf{1} is the signal mixture with the common relaxation decay factored out. Because the shared factor e−Rte^{-Rt} is real and positive, it affects only the magnitude, so all phase information is carried by β\boldsymbol{\beta}, and the observations are the samples of (8), y=[y(TEe)]e\mathbf{y} = [y(\mathrm{TE}_{e})]_{e}.

Model (10) is linear in c\mathbf{c} once (χ,R)(\boldsymbol{\chi}, R) are fixed, a separable structure well known in nonlinear least squares [11].

Geometrically, because c\mathbf{c} lies in the probability simplex, s\mathbf{s} lies in the triangle in CNe\mathbb{C}^{N_{e}} whose vertices are the three pool signals b+\mathbf{b}^{+}, b−\mathbf{b}^{-}, and b0\mathbf{b}^{0}. The susceptibilities and RR position the vertices, while the concentrations select a point inside the triangle.

Across a set of voxels VV, the observations form a map y:V→CNe\mathbf{y}: V \to\mathbb{C}^{N_{e}}. Since (10) contains no spatial coupling between voxels (which has already been resolved by the per-echo dipole inversion), the whole-brain inverse problem decouples into ∣V∣|V| independent voxel problems, enabling massively parallel voxel-wise computation.

Output maps

The products c±χ±c^{\pm}\chi^{\pm} alone do not capture the nonlinear phase and magnitude evolution of the multi-pool signal. Chen [5] therefore define effective component susceptibilities from the phase of the fitted sub-voxel signals. We first specify how the phase is tracked across echoes, and then define the output maps.

Because the principal argument arg⁡\arg is defined only modulo 2π2\pi, the wrapping operator maps any real phase to its principal value in (−π,π](-\pi,\pi],

wrap⁡(x)=x−2π⌊x−π2π⌋,(12)\operatorname{wrap}(x) = x - 2\pi\left\lfloor\frac{x-\pi}{2\pi} \right\rfloor, \tag*{(12)}

so that arg⁡z=wrap⁡(ϑ)\arg z = \operatorname{wrap}(\vartheta) for any phase ϑ\vartheta of z≠0z \ne0, and arg⁡(z1z‾2)=wrap⁡(arg⁡z1−arg⁡z2)\arg(z_{1}\overline{z}_{2}) = \operatorname{wrap}(\arg z_{1} - \arg z_{2}). For a discrete signal m=(m1,…,mNe)∈CNe\mathbf{m} = (m_{1}, \ldots, m_{N_{e}}) \in\mathbb{C}^{N_{e}} with nonzero entries, we define its unwrapped phase U(arg⁡m)∈RNeU(\arg\mathbf{m}) \in\mathbb{R}^{N_{e}} as the phase tracked continuously from t=0t=0: starting from phase 00 (m0=1m_{0}=1) at TE0=0\mathrm{TE}_{0}=0, it accumulates the wrapped phase increments between consecutive echoes [12]. The effective susceptibility of m\mathbf{m} is then defined as the ratio of the summed unwrapped phases to pp times the summed echo times:

Ψ(m)=∑e=1NeU(arg⁡m)ep∑e=1NeTEe.(13)\Psi(\mathbf{m}) = \frac{\sum_{e=1}^{N_{e}} U(\arg m)_{e}}{p\sum_{e=1}^{N_{e}} \mathrm{TE}_{e}}. \tag*{(13)}

For a single-pool signal with linear phase pχtp\chi t, Ψ(m)\Psi(\mathbf{m}) returns χ\chi exactly.

The paramagnetic and diamagnetic component susceptibilities (PCS and DCS) are defined as the effective susceptibilities of a voxel in which the opposing source is replaced by the neutral medium, and the composite susceptibility χcomp\chi_{\mathrm{comp}} is that of the full three-pool signal:

PCS=Ψ(c+b++(c0+c−)b0),(14)\mathrm{PCS} = \Psi(c^{+}\mathbf{b}^{+} + (c^{0}+c^{-})\mathbf{b}^{0}), \tag*{(14)}
DCS=Ψ(c−b−+(c0+c+)b0),(15)\mathrm{DCS} = \Psi(c^{-}\mathbf{b}^{-} + (c^{0}+c^{+})\mathbf{b}^{0}), \tag*{(15)}
χcomp=Ψ(s).(16)\chi_{\mathrm{comp}} = \Psi(\mathbf{s}). \tag*{(16)}

Each of these vectors has the form e−Rt⊙ge^{-Rt}\odot\mathbf{g} with g\mathbf{g} independent of RR, since a positive real multiplier does not alter the complex argument; PCS, DCS, and χcomp\chi_{\mathrm{comp}} are independent of RR. Expanding each signal at short echo times (Theorem 2 in Appendix § 9) shows that as TEe→0\mathrm{TE}_{e}\to0, where UU acts as the identity, PCS, DCS, and χcomp\chi_{\mathrm{comp}} converge to c+χ+c^{+}\chi^{+}, c−χ−c^{-}\chi^{-}, and the net susceptibility c+χ++c−χ−c^{+}\chi^{+}+c^{-}\chi^{-}, respectively. At practical echo times, however, they differ substantially because the active pool decays faster than the neutral pool. For example, with c=(0.3,0.3,0.4)\mathbf{c}=(0.3,0.3,0.4) and χ=(0.15,−0.10)\boldsymbol{\chi}=(0.15,-0.10) ppm at 3 T, PCS = 0.012 ppm whereas c+χ+=0.045c^{+}\chi^{+}=0.045 ppm. This difference reflects the nonlinear phase evolution that the component maps are designed to capture [5].

Geometrical View

The model in (4) admits an intuitive geometric interpretation. Each term in (4) represents a rotating phasor in the complex plane: the paramagnetic phasor rotates counter-clockwise, the diamagnetic phasor rotates clockwise, and the neutral phasor remains stationary on the real axis. In addition, the two source phasors shrink faster than the neutral phasor. The composite signal S(t)S(t) is the vector sum of these three arrows (Figure 1a).

Figure 1. Phasor view of the three-pool model [5] (Eq. (4)) at 3 T for c=(0.3,0.3,0.4)\mathbf{c} = (0.3, 0.3, 0.4), χ=(0.15,−0.10)\boldsymbol{\chi} = (0.15, -0.10) ppm, and R=20 s−1R = 20\,\mathrm{s}^{-1}. (a) At t=12t = 12 ms, the paramagnetic (red) and diamagnetic (blue) phasors rotate in opposite directions and nearly cancel in the sum SS (black), which remains close to the neutral phasor (gray). Dashed curves show the trajectories of the two source phasors, and dotted lines indicate vector addition. (b) Trajectory of S(t)S(t) for 0≤t≤400 \le t \le40 ms (black) compared with a single-pool signal sharing the same first-order terms (gray, dashed), and dots mark the five echoes of the in vivo protocol (TE=7.2\mathrm{TE} = 7.2–32.032.0 ms). (c) Phase of the two signals versus tt: both start with the same initial slope, determined by the net susceptibility, but the phase of the mixed voxel bends and changes sign as the faster-decaying paramagnetic pool attenuates, leaving the diamagnetic pool dominant at later echoes.

At early echoes, the two opposing rotations nearly cancel, so S(t)S(t) stays close to the neutral phasor and its phase remains near zero. At later echoes, the faster-decaying source phasor fades, breaking the cancellation and causing the phase trajectory of S(t)S(t) to bend (Figure 1b,Figure 1c). This curvature provides the information needed to separate the sources. The formal statements and proofs are given in Appendix § 9 (Theorem 2 and Corollary 1).

At short echo times, arg⁡S(t)\arg S(t) grows linearly with slope proportional to the net susceptibility c+χ++c−χ−c^{+}\chi^{+} + c^{-}\chi^{-}, while log⁡∣S(t)∣\log|S(t)| decays linearly at the rate R+a(c+χ++c−∣χ−∣)R + a\left(c^{+}\chi^{+} + c^{-}|\chi^{-}|\right). Standard linear-phase QSM and mono-exponential R2∗R_{2}^{*} fitting measure only these two first-order quantities, so a voxel containing high concentrations of both iron and myelin and a voxel containing neither can produce identical first-order measurements.

In a mixed voxel, however, the apparent susceptibility χapp(t)=arg⁡S(t)/(pt)\chi_{\mathrm{app}}(t) = \arg S(t)/(p t)—which is what a QSM map reconstructed from the echo at time tt estimates—is not constant. To first order in tt, it drifts linearly with echo time (Corollary 1):

χapp(t)≈c+χ++c−χ−−aκt,κ=c+(1−c+)(χ+)2−c−(1−c−)(χ−)2.(17)\chi_{\mathrm{app}}(t) \approx c^{+}\chi^{+} + c^{-}\chi^{-} - a\kappa t,\qquad\kappa= c^{+}(1-c^{+})(\chi^{+})^{2} - c^{-}(1-c^{-})(\chi^{-})^{2}. \tag*{(17)}

The sign of the slope −aκ-a\kappa indicates which source dominates the quadratic phase drift. For the mixed voxel in Figure 1, χapp(t)\chi_{\mathrm{app}}(t) shifts from 0.0085 ppm at 7.2 ms to −0.0123-0.0123 ppm at 32 ms, closely matching the values of 0.0089 ppm and −0.0122-0.0122 ppm predicted by (17). Whereas an echo-combined QSM averages out this temporal drift, reconstructing QSM per echo preserves it—explaining why the synthesized signal (8) must be constructed from per-echo QSM maps.

Methods

CANON-QSM

Following [5], the parameters are estimated by alternating between three blocks, each minimized with the other two fixed:

(A)minimize⁡c≥0, 1Tc=1 ∥Bc−y∥2,(B)minimize⁡0≤R≤Rmax⁡ ∥log⁡s(R)−log⁡y∥2,(C)minimize⁡0≤χ+≤χmax⁡ ∥log⁡s(χ)−log⁡y∥2,−χmax⁡≤χ−≤0.(18)\begin{aligned} \text{(A)}\quad&\underset{\mathbf{c}\ge0,\ \mathbf{1}^{\mathsf{T}}\mathbf{c}=1}{\operatorname{minimize}}\ \|\mathbf{B}\mathbf{c}-\mathbf{y}\|^{2},\\ \text{(B)}\quad&\underset{0\le R\le R_{\max}}{\operatorname{minimize}}\ \|\log\mathbf{s}(R)-\log\mathbf{y}\|^{2},\\ \text{(C)}\quad&\underset{0\le\chi^{+}\le\chi_{\max}}{\operatorname{minimize}}\ \|\log\mathbf{s}(\chi)-\log\mathbf{y}\|^{2},\qquad-\chi_{\max}\le\chi^{-}\le0. \tag*{(18)} \end{aligned}

Block (A) uses a least-squares cost on the signal, in which the model is linear in c\mathbf{c}. Blocks (B) and (C) use a cost on the complex logarithm, which separates log-magnitude and phase. The following subsections describe how CANON-QSM solves each block. The derivations are given in Appendix § 10.

Subproblem (A): Concentrations in Closed Form

To prevent a pool from vanishing early in the fit, we impose a lower bound (floor) cf∈[0,1/3)c_f \in[0,1/3) on every concentration by writing c=cf1+(1−3cf)q\mathbf{c}=c_f\mathbf{1}+(1-3c_f)\mathbf{q}, where q=(q+,q−,q0)T\mathbf{q}=(q^+,q^-,q^0)^\mathsf{T} is nonnegative and sums to one. With x=(q+,q−)T\mathbf{x}=(q^+,q^-)^\mathsf{T} and q0=1−q+−q−q^0=1-q^+-q^-, define the floor-corrected data y~=(y−cfB1)/(1−3cf)\tilde{\mathbf{y}}=(\mathbf{y}-c_f\mathbf{B1})/(1-3c_f), the source-versus-neutral contrasts d±=b±−b0\mathbf{d}^{\pm}=\mathbf{b}^{\pm}-\mathbf{b}^0, D=[d+ d−]\mathbf{D}=[\mathbf{d}^+\ \mathbf{d}^-], z=y~−b0\mathbf{z}=\tilde{\mathbf{y}}-\mathbf{b}^0, and

G=Re⁡(DHD)∈R2×2,h=Re⁡(DHz)∈R2.(19)\mathbf{G}=\operatorname{Re}(\mathbf{D}^{\mathsf{H}}\mathbf{D})\in\mathbb{R}^{2\times2},\qquad\mathbf{h}=\operatorname{Re}(\mathbf{D}^{\mathsf{H}}\mathbf{z})\in\mathbb{R}^2. \tag*{(19)}

Theorem 1 (Exact solution of subproblem (A)). Block (A) with the floor is equivalent to the two-variable convex quadratic program

minimize⁡x∈T 12xTGx−hTx,T={x≥0, x++x−≤1}.(20)\underset{\mathbf{x}\in T}{\operatorname{minimize}}\ \frac{1}{2}\mathbf{x}^{\mathsf{T}}\mathbf{G}\mathbf{x}-\mathbf{h}^{\mathsf{T}}\mathbf{x},\qquad T=\{\mathbf{x}\geq0,\ x^++x^-\leq1\}. \tag*{(20)}

Its global minimizer is the lowest-cost point among four candidates: the unconstrained minimizer G−1h\mathbf{G}^{-1}\mathbf{h} (if it lies inside TT), and the minimizer on each of the three edges of TT,

(0,clip⁡[0,1]h2G22),(clip⁡[0,1]h1G11,0),(τ,1−τ),τ=clip⁡[0,1]Re⁡(eHw)∥e∥2,(21)\left(0,\operatorname{clip}_{[0,1]}\frac{h_2}{G_{22}}\right),\qquad\left(\operatorname{clip}_{[0,1]}\frac{h_1}{G_{11}},0\right),\qquad\left(\tau,1-\tau\right),\quad\tau=\operatorname{clip}_{[0,1]}\frac{\operatorname{Re}(\mathbf{e}^{\mathsf{H}}\mathbf{w})}{\lVert\mathbf{e}\rVert^2}, \tag*{(21)}

with e=d+−d−\mathbf{e}=\mathbf{d}^+-\mathbf{d}^- and w=z−d−\mathbf{w}=\mathbf{z}-\mathbf{d}^-. This holds also when G\mathbf{G} is singular.

Geometrically, block (A) finds the point of the triangle with vertices b+\mathbf{b}^+, b−\mathbf{b}^-, b0\mathbf{b}^0 that is closest to the data. G\mathbf{G} is the real Gram matrix of the two triangle edges that leave b0\mathbf{b}^0. The closest point is either the orthogonal projection onto the plane of the triangle, if it falls inside, or lies on one of the three edges. The edges correspond to one pool being absent (c+c^+, c−c^- or c0c^0 at the floor), and the vertices of TT (a single pool) appear when a ratio is clipped. The update is closed-form and needs no iterations.

Subproblem (B): Baseline Relaxation in Closed Form

Since s=e−Rt⊙β\mathbf{s}=e^{-R\mathbf{t}}\odot\boldsymbol{\beta} and e−Rte^{-R\mathbf{t}} is real and positive, RR affects only the log-magnitude of the model and not its phase.

Proposition 1 (Exact solution of block (B)). Let ρ=log⁡∣y∣−log⁡∣β∣\boldsymbol{\rho}=\log|\mathbf{y}|-\log|\boldsymbol{\beta}|. The cost in (19) equals ∥Rt+ρ∥2\lVert R\mathbf{t}+\boldsymbol{\rho}\rVert^2 plus a term that does not depend on RR, and its minimizer over [0,Rmax⁡][0,R_{\max}] is

R⋆=clip⁡[0,Rmax⁡](−tTρtTt).(22)R^\star=\operatorname{clip}_{[0,R_{\max}]}\left(-\frac{\mathbf{t}^{\mathsf{T}}\boldsymbol{\rho}}{\mathbf{t}^{\mathsf{T}}\mathbf{t}}\right). \tag*{(22)}

This is a log-linear R2⋆R_2^\star fit, through the origin, of ∣y∣/∣β∣|\mathbf{y}|/|\boldsymbol{\beta}|: the data with the decay explained by the two sources divided out.

Subproblem (C): Susceptibilities by Levenberg-Marquardt

The phase part of (20), arg⁡s−arg⁡y\arg\mathbf{s}-\arg\mathbf{y}, jumps by 2π2\pi whenever one of the two phases wraps and the other does not. This occurs for large venous susceptibilities, at high field and at late echoes, and it adds spurious penalties of order (2π)2(2\pi)^2 to the cost. We therefore use the residual

r(χ)=[log⁡∣s∣−log⁡∣y∣arg⁡(s⊙yˉ)]∈R2Ne,(23)\mathbf{r}(\boldsymbol{\chi})= \begin{bmatrix} \log|\mathbf{s}|-\log|\mathbf{y}|\\ \arg(\mathbf{s}\odot\bar{\mathbf{y}}) \end{bmatrix}\in\mathbb{R}^{2N_e}, \tag*{(23)}

whose phase part is the phase difference wrapped to (−π,π](-\pi,\pi], arg⁡(s⊙yˉ)=wrap⁡(arg⁡s−arg⁡y)\arg(\mathbf{s}\odot\bar{\mathbf{y}})=\operatorname{wrap}(\arg\mathbf{s}-\arg\mathbf{y}) with wrap as in (12). In other words, ∥r∥2\lVert\mathbf{r}\rVert^2 is the original cost of (20) with every spurious 2π2\pi jump removed, and it coincides with the original cost whenever no echo has such a jump. Block (C) thus becomes a bounded nonlinear least-squares problem in two variables.

With s±=c±b±\mathbf{s}^{\pm}=c^{\pm}\mathbf{b}^{\pm}, the Jacobian of r\mathbf{r} has columns

j±=[Re⁡v±Im⁡v±],v±=u±(s±⊙s)⊙t.(24)\mathbf{j}^{\pm}= \begin{bmatrix} \operatorname{Re}\mathbf{v}^{\pm}\\ \operatorname{Im}\mathbf{v}^{\pm} \end{bmatrix}, \qquad \mathbf{v}^{\pm}=u^{\pm}(\mathbf{s}^{\pm}\odot\mathbf{s})\odot\mathbf{t}. \tag*{(24)}

The sensitivity to χ±\chi^{\pm} is thus the fraction of the signal coming from pool ±\pm, multiplied by its complex coefficient and by TE: late echoes and abundant pools are the most informative. Each Levenberg-Marquardt (LM) step [13, 14, 15] solves

(JTJ+λσˉI)δ=−JTr,σˉ=12tr⁡(JTJ),(25)\left(\mathbf{J}^{\mathsf{T}}\mathbf{J}+\lambda\bar{\sigma}\mathbf{I}\right)\boldsymbol{\delta}=-\mathbf{J}^{\mathsf{T}}\mathbf{r},\qquad\bar{\sigma}=\frac{1}{2}\operatorname{tr}(\mathbf{J}^{\mathsf{T}}\mathbf{J}), \tag*{(25)}

a 2×22\times2 system solved in closed form by Cramer’s rule. The step interpolates between a Gauss-Newton step (λ→0\lambda\to0) and a short gradient-descent step (λ→∞\lambda\to\infty). Scaling the damping by σˉ\bar{\sigma} makes λ\lambda independent of the units of χ\chi. If a concentration reaches zero, then s±=0\mathbf{s}^{\pm}=0, the corresponding Jacobian column vanishes, and χ±\chi^{\pm} can no longer be updated. The floor cfc_f prevents this:

Proposition 2. If c±≥cf>0c^{\pm} \ge c_f > 0, TEe>0\mathrm{TE}_e > 0 for all ee, and ss has nonzero entries, then j±≠0j^{\pm} \ne0 and the matrix in (25) is positive definite for every λ>0\lambda> 0. Whenever J⊤r≠0\mathbf{J}^{\top}\mathbf{r} \ne0, the LM step is a strict descent direction for ∥r∥2\lVert\mathbf{r}\rVert^2.

We therefore start with a large floor and decrease it geometrically over the iterations, so that both sources remain observable while the fit is far from the solution, and the floor no longer biases the final estimate.

Exact unwrapping of the output maps requires ∣pχ∣max⁡e(TEe−TEe−1)<π|p\chi| \max_e(\mathrm{TE}_e-\mathrm{TE}_{e-1}) < \pi with TE0=0\mathrm{TE}_0 = 0 [12], which for a single pool means ∣χ∣<0.82|\chi| < 0.82 ppm at 3 T and ∣χ∣<0.31|\chi| < 0.31 ppm at 7 T for our protocols. The bound χmax⁡\chi_{\max} should therefore be of the same order as these values. Chen et al. used χmax⁡=0.5\chi_{\max} = 0.5 ppm [5]. In our in silico experiments we set χmax⁡=max⁡(0.2,1.5max⁡∣χGT∣)=0.48\chi_{\max} = \max(0.2, 1.5 \max|\chi_{\mathrm{GT}}|) = 0.48 ppm, where χGT\chi_{\mathrm{GT}} is the ground-truth susceptibility of the phantom. For in vivo data χmax⁡\chi_{\max} is chosen per dataset, since no ground truth is available.

Multi-start strategy and output selection

Block (C) is non-convex, so each voxel runs KK initializations (restarts) in parallel. The deterministic restart starts from χ=(0.1χmax⁡,−0.1χmax⁡)\chi= (0.1\chi_{\max}, -0.1\chi_{\max}) and from the two-point estimate R=−log⁡(∣yNe∣/∣y1∣)/(TENe−TE1)R = -\log(|y_{N_e}|/|y_1|)/(\mathrm{TE}_{N_e}-\mathrm{TE}_1). The other K−1K-1 restarts draw χ+/χmax⁡\chi^+/\chi_{\max} and −χ−/χmax⁡-\chi^-/\chi_{\max} uniformly from [0.01,0.99][0.01, 0.99] and RR uniformly from [0,max⁡{60,2Rˉ0}] s−1[0,\max\{60,2\bar{R}_0\}]\,\mathrm{s}^{-1}, where Rˉ0\bar{R}_0 is the mean two-point estimate. Block (A) is solved first, so the initial concentrations play no role.

Blocks (B) and (C) reduce the log-domain cost, whereas block (A) reduces the signal-domain cost, so no single cost decreases monotonically over the alternation. After each outer iteration, we therefore select the restart with the lowest score

S=∥s−y∥2∥y∥2+∥r∥2Ne,(26)\mathcal{S} = \frac{\lVert\mathbf{s}-\mathbf{y}\rVert^2}{\lVert\mathbf{y}\rVert^2} + \frac{\lVert\mathbf{r}\rVert^2}{N_e}, \tag*{(26)}

which weights both costs, and the output is the best state over all iterations.

Implementation

Algorithm 1 summarizes the solver. Because all voxels and restarts are mutually independent, we process batches of voxels ×K\times K restarts in parallel on a GPU in PyTorch [16], using spatial chunks to bound memory consumption. The computational complexity per volume is O(TK(nLM)NeNvox)O(TK(n_{\mathrm{LM}})N_eN_{\mathrm{vox}}), which scales linearly with the number of voxels and echoes.

Experimental Design

We evaluated CANON-QSM in the following experiments:

In silico: We used the realistic susceptibility-separation brain phantom [17], which provides ground-truth isotropic χ+\chi^+ and χ−\chi^- maps (matrix size 256×320×320256 \times320 \times320, 0.64 mm isotropic resolution, FOV =164×205×205= 164 \times205 \times205 mm3^3). Five echoes (TE =4= 4–36 ms, Δ\DeltaTE =8= 8 ms) were simulated at 7 T. Complex Gaussian noise with standard deviation σ=mean⁡(M1)/SNR\sigma= \operatorname{mean}(M_1)/\mathrm{SNR}, where M1M_1 is the first-echo magnitude, was added directly to the synthesized per-echo input signal to isolate solver performance from upstream preprocessing errors. We also evaluated ROI-specific reconstruction error and multi-start convergence reliability at 3 T and 7 T.

In vivo: One healthy volunteer was scanned at 3 T using a 3D multi-echo GRE sequence with institutional ethics approval (voxel size 0.60×0.60×1.00.60 \times0.60 \times1.0 mm3^3, TR/TE1_1/Δ\DeltaTE =43.4/7.2/6.2= 43.4/7.2/6.2 ms, 5 echoes, bandwidth 550.5 Hz/pixel, matrix size 352×352×160352 \times352 \times160, flip angle 17∘17^\circ). Preprocessing comprised ROMEO phase unwrapping [18], PDF background field removal [19], WH-QSM harmonic remnant estimation [20], and per-echo dipole inversion via DEMIX-QSM [21].

In both experiments, CANON-QSM was compared against DECOMPOSE-QSM run for 10 iterations as recommended in [5], using the reference implementation in QSM-CI [22]. We report composite, PCS, and DCS maps, and evaluate reconstruction accuracy using NRMSE and XSIM [23].

Results

In silico: Figure 2 compares the noiseless reconstructions. Both methods separate the three pools across the composite, PCS, and ∣DCS∣|\mathrm{DCS}| maps, with CANON-QSM preserving sharper tissue contrast and exhibiting smaller residual errors relative to the ground truth. Across the whole-brain volume, CANON-QSM (T=20T = 20, K=20K = 20) achieved a lower composite NRMSE than DECOMPOSE-QSM (10 iterations) both in the noiseless case (1.8% vs. 10.5%) and at SNR =100= 100 (6.0% vs. 13.8%), as well as consistently higher XSIM across all noise levels (Figure 3). In the noiseless ROI analysis (Figure 4), CANON-QSM yielded lower NRMSE in 8 of 11 segmented structures, with the largest improvements in white matter (0.04% vs. 14.3%), caudate (0.01% vs. 11.8%), thalamus (0.20% vs. 15.1%), and putamen (0.15% vs. 10.2%). DECOMPOSE-QSM yielded slightly lower NRMSE in venous blood (0.24% vs. 0.74%) and the dentate nucleus (0.88% vs. 1.6%), while both methods achieved negligible error in the globus pallidus.

Figure 2. In silico 7 T phantom comparison (noiseless) in sagittal, coronal, and axial views. Both CANON-QSM (T=20T = 20, K=20K = 20) and DECOMPOSE-QSM (10 iterations) resolve the three-pool model across the Composite, paramagnetic component susceptibility (PCS), and magnitude diamagnetic component susceptibility (∣|DCS∣|) maps. CANON-QSM yields sharper tissue contrast, smaller inter-method differences, and lower composite reconstruction error (normalized root-mean-square error [NRMSE]: 1.81% vs. 10.48%, XSIM: 0.9996 vs. 0.9749).

Figure 3. Noise robustness on the in silico 7 T brain phantom. CANON-QSM (T=20T = 20, K=20K = 20) consistently achieves lower composite normalized root-mean-square error (NRMSE) and higher structural similarity (XSIM) than DECOMPOSE-QSM (10 iterations) across all signal-to-noise ratio (SNR) levels (SNR = 50-300 and noiseless [clean]). Although reconstruction accuracy for both methods degrades with increasing noise, CANON-QSM maintains superior performance across the entire SNR range.

Figure 4. Composite normalized root-mean-square error (NRMSE) in 11 segmented structures of the noiseless in silico 7 T phantom. DECOMPOSE-QSM was run for 10 iterations and CANON-QSM with T=20T = 20 and K=20K = 20. CANON-QSM achieves lower error in 8 of 11 structures, whereas DECOMPOSE-QSM yields lower error in venous blood and the dentate nucleus (DN), and both methods achieve negligible error in the globus pallidus (GP). Abbreviations: CSF, cerebrospinal fluid, GM, gray matter, RN, red nucleus, SNandSTN, substantia nigra and subthalamic nucleus, WM, white matter.

Figure 5 shows the percentage of voxels in which the selected solution converges to within 0.005 ppm of the 512-start reference solution as a function of the number of restarts KK. At 7 T, every tissue reached the reference solution in at least 99.9% of voxels for K≥16K \ge16. At 3 T, convergence was slower: with K=20K = 20, the convergence rate reached 79.5% in cortical gray matter, 94% in the caudate, 95% in the putamen, 97% in veins, and nearly 100% in white matter and the globus pallidus, increasing to K=64K = 64 raised convergence to at least 96% across all tissues. Thus, K=20K = 20 is sufficient at 7 T and in most structures at 3 T, whereas cortical gray matter at 3 T benefits from additional restarts.

Figure 5. Reliability of the multi-start strategy on the in silico phantom at SNR = 100 across six tissue classes. Curves show the percentage of voxels in which the selected solution recovers paramagnetic and diamagnetic component susceptibilities (PCS and DCS) within 0.005 ppm of the reference solution (defined as the best-scoring solution among 512 random starts plus a start at the ground-truth parameters) as a function of the number of restarts KK at 3 T (left) and 7 T (right). The vertical dashed line marks K=20K = 20 (the default value used in this work), and the horizontal dotted line marks the 95% convergence threshold. Abbreviations: GM, gray matter, GP, globus pallidus, WM, white matter.

In vivo: Figure 6 displays the in vivo reconstructions. DECOMPOSE-QSM produced severe fitting artifacts in the great vein of Galen (sagittal view) in both the composite and PCS maps—erroneously assigning paramagnetic venous voxels to the diamagnetic pool, consistent with [8]—and exhibited reduced tissue contrast at 10 iterations. CANON-QSM eliminated these venous artifacts and yielded clear delineation of deep gray matter nuclei and white matter tracts. For the 352×352×160352 \times352 \times160 volume (Table 1), CANON-QSM completed in 113 s using 1.68 GB of peak GPU memory, compared with approximately 5 h for the MATLAB reference implementation of DECOMPOSE-QSM, corresponding to a ≈ 160× speedup.

MethodHardwareTimePeak memory
DECOMPOSE-QSM (Matlab Reference)CPU, 18 threads≈ 5 h–
CANON-QSM (T=20T = 20, K=20K = 20)GPU (RTX 5090)113 s1.68 GB (GPU)

Table 1. Whole-brain runtime (in vivo, 352×352×160352 \times352 \times160, 5 echoes).

Figure 6. In vivo 3 T whole-brain results in coronal, sagittal, and axial views. CANON-QSM (T=20T = 20, K=20K = 20) yields consistent susceptibility separation across the Composite, paramagnetic component susceptibility (PCS), and magnitude diamagnetic component susceptibility (∣DCS∣\lvert\mathrm{DCS}\rvert) maps. In contrast, DECOMPOSE-QSM (at the recommended 10 iterations) displays severe fitting artifacts in the great vein of Galen (sagittal view) in both the composite and PCS maps, falsely assigning paramagnetic venous voxels to the diamagnetic pool (consistent with [8]), alongside reduced tissue contrast.

Discussion

Susceptibility source separation is a rapidly growing area of QSM, motivated by the fact that the bulk susceptibility value can remain near zero even when paramagnetic and diamagnetic sources are both abundant within a voxel. In this work, we introduced CANON-QSM, a solver for the three-pool model of DECOMPOSE-QSM [5] in which two of the three alternating blocks are solved in closed form and the third is formulated as a two-variable Levenberg-Marquardt problem with a residual that is insensitive to phase wrapping. We also analyzed the physics of the echo train. Expanding the signal in powers of echo time shows that linear-phase QSM and mono-exponential R2∗R^{*}_{2} fitting measure only the net susceptibility and the total decay, whereas the sources are separated by the second-order drift of the apparent susceptibility with echo time. On the in silico phantom, CANON-QSM reached a lower composite NRMSE than DECOMPOSE-QSM (with 10 iterations) without noise (1.8% vs. 10.5%), and it processed a whole-brain volume in under two minutes on a GPU. In in vivo data, the whole-brain reconstruction took 113 s, compared with approximately 5 h for the reference implementation, and the venous fitting artifacts of the reference [8] were eliminated.

The substantial speedup stems from replacing iterative inner-loop solvers with exact block updates, which removes inner iterations and makes parallel restarts per voxel computationally affordable. Combined with the decaying concentration floor, which prevents pool gradients from vanishing early in the fit, and the wrapped phase residual, this multi-start strategy also explains the improved accuracy and the absence of 2π2\pi wrapping artifacts in large veins. Nevertheless, because CANON-QSM shares the physical forward model of DECOMPOSE-QSM, it inherits the assumptions of the static dephasing regime [6], isotropic susceptibility, and error propagation from the per-echo QSM inputs. Finally, while our in vivo proof of concept demonstrates feasibility at 3 T, a larger cohort and histological or clinical dataset will be necessary to establish robustness and generalizability across diverse acquisition protocols and pathological scenarios.

Conclusions

We introduced CANON-QSM, a solver for the three-pool susceptibility source separation model in which two of the three alternating blocks are solved in closed form and the third is formulated as a two-variable Newton-type (Levenberg-Marquardt) problem with a residual that is robust to phase wrapping. An expansion of the signal in powers of echo time explains why the model requires per-echo QSM reconstructions: linear-phase QSM and mono-exponential R2∗R_2^* fitting capture only the net susceptibility and the total decay, whereas the sources are separated by the second-order drift of the apparent susceptibility with echo time, at the cost of information that vanishes rapidly for short echo trains. Because every voxel is independent and every update reduces to elementwise operations, the solver runs voxel-parallel on a GPU, achieving whole-brain reconstructions in under two minutes without venous phase-wrapping artifacts.

Beyond a faster solver for this three-pool model, a new theoretical formulation is needed to accurately account for the presence of multiple paramagnetic and diamagnetic species within a voxel without fixing their number a priori, as the exact number of sub-voxel compartments is generally unknown.

Acknowledgments

We thank ANID: FONDECYT de Iniciación 11240279, Fondecyt 1231535 and 1270460.

Data Availability Statement

The source code from this paper is available on GitHub at ignacontreras/CANON-QSM.

Theorems and proofs for Section 2.4

We define the following summary quantities:

χnet=c+χ++c−χ−(27)\chi_{\mathrm{net}} = c^{+}\chi^{+} + c^{-}\chi^{-} \tag*{(27)}

(net susceptibility),

χtot=c+χ++c−∣χ−∣(28)\chi_{\mathrm{tot}} = c^{+}\chi^{+} + c^{-}\lvert\chi^{-}\rvert \tag*{(28)}

(total susceptibility),

q=c+(χ+)2+c−(χ−)2,(29)q = c^{+}(\chi^{+})^{2} + c^{-}(\chi^{-})^{2}, \tag*{(29)}
vnet=q−χnet2,vtot=q−χtot2,(30)v_{\mathrm{net}} = q - \chi_{\mathrm{net}}^{2}, \qquad v_{\mathrm{tot}} = q - \chi_{\mathrm{tot}}^{2}, \tag*{(30)}
κ=c+(1−c+)(χ+)2−c−(1−c−)(χ−)2.(31)\kappa= c^{+}(1-c^{+})(\chi^{+})^{2} - c^{-}(1-c^{-})(\chi^{-})^{2}. \tag*{(31)}

Here, χnet\chi_{\mathrm{net}} represents the bulk susceptibility measured by traditional QSM, whereas χtot\chi_{\mathrm{tot}} governs the overall signal decay rate. The terms vnetv_{\mathrm{net}} and vtotv_{\mathrm{tot}} quantify the dispersion of the two susceptibilities around these values, and κ\kappa measures the asymmetry between positive and negative source distributions. Note that κ=c+(χ+)2−c−(χ−)2−χtotχnet\kappa= c^{+}(\chi^{+})^{2} - c^{-}(\chi^{-})^{2} - \chi_{\mathrm{tot}}\chi_{\mathrm{net}}, since χtotχnet=(c+)2(χ+)2−(c−)2(χ−)2\chi_{\mathrm{tot}}\chi_{\mathrm{net}} = (c^{+})^{2}(\chi^{+})^{2} - (c^{-})^{2}(\chi^{-})^{2} (using ∣χ−∣=−χ−\lvert\chi^{-}\rvert= -\chi^{-}). For a single source mixed with the neutral pool (c−=0c^{-}=0, 0<c+<10<c^{+}<1, χ+>0\chi^{+}>0), we have κ=c+(1−c+)(χ+)2>0\kappa= c^{+}(1-c^{+})(\chi^{+})^{2}>0.

Theorem 2 (Echo-time expansion). Let arg⁡S(t)\arg S(t) be the continuous unwrapped phase of S(t)S(t), with arg⁡S(0)=0\arg S(0)=0. For tt in a neighborhood of 00,

log⁡∣S(t)∣=−(R+aχtot)t+12(a2vtot−p2vnet)t2+O(t3),(32)\log\lvert S(t)\rvert= -(R+a\chi_{\mathrm{tot}})t + \frac{1}{2}(a^{2}v_{\mathrm{tot}}-p^{2}v_{\mathrm{net}})t^{2} + O(t^{3}), \tag*{(32)}
arg⁡S(t)=pχnett−apκt2+O(t3).(33)\arg S(t) = p\chi_{\mathrm{net}}t - ap\kappa t^{2} + O(t^{3}). \tag*{(33)}

Corollary 1 (Apparent susceptibility). The apparent susceptibility, defined as the phase divided by ptpt,

χapp(t)=arg⁡S(t)pt,\chi_{\mathrm{app}}(t) = \frac{\arg S(t)}{pt},

which corresponds to the QSM value reconstructed from the echo acquired at time tt, satisfies

χapp(t)=χnet−aκt+O(t2).(34)\chi_{\mathrm{app}}(t) = \chi_{\mathrm{net}} - a\kappa t + O(t^{2}). \tag*{(34)}

In terms of the original parameters,

χapp(t)≈(c+χ++c−χ−)−a(c+(1−c+)(χ+)2−c−(1−c−)(χ−)2)t.\chi_{\mathrm{app}}(t) \approx(c^{+}\chi^{+}+c^{-}\chi^{-}) - a\left(c^{+}(1-c^{+})(\chi^{+})^{2}-c^{-}(1-c^{-})(\chi^{-})^{2}\right)t.

Proof of Theorem 2. Express the signal as S(t)=e−Rtβ(t)S(t)=e^{-Rt}\beta(t) with

β(t)=c+eω+t+c−eω−t+c0,ω+=(−a+ip)χ+,ω−=(a+ip)χ−,\beta(t)=c^{+}e^{\omega^{+}t}+c^{-}e^{\omega^{-}t}+c^{0}, \qquad\omega^{+}=(-a+ip)\chi^{+}, \qquad\omega^{-}=(a+ip)\chi^{-},

where we used −a∣χ−∣=aχ−-a\lvert\chi^{-}\rvert=a\chi^{-}. Then log⁡S(t)=−Rt+log⁡β(t)\log S(t)=-Rt+\log\beta(t).

Expanding each exponential to second order and using c++c−+c0=1c^{+}+c^{-}+c^{0}=1 gives

β(t)=1+μ1t+12μ2t2+O(t3),μ1=c+ω++c−ω−,μ2=c+(ω+)2+c−(ω−)2.\beta(t)=1+\mu_{1}t+\frac{1}{2}\mu_{2}t^{2}+O(t^{3}), \qquad\mu_{1}=c^{+}\omega^{+}+c^{-}\omega^{-}, \qquad\mu_{2}=c^{+}(\omega^{+})^{2}+c^{-}(\omega^{-})^{2}.

Since β(0)=1\beta(0)=1, log⁡β(t)\log\beta(t) is analytic near t=0t=0 with log⁡β(0)=0\log\beta(0)=0, applying log⁡(1+z)=z−12z2+O(z3)\log(1+z)=z-\frac{1}{2}z^{2}+O(z^{3}) yields

log⁡β(t)=μ1t+12(μ2−μ12)t2+O(t3).\log\beta(t)=\mu_{1}t+\frac{1}{2}(\mu_{2}-\mu_{1}^{2})t^{2}+O(t^{3}).

Evaluating the coefficients, we obtain

μ1=c+(−a+ip)χ++c−(a+ip)χ−=−aχtot+ipχnet.\mu_{1}=c^{+}(-a+ip)\chi^{+}+c^{-}(a+ip)\chi^{-}=-a\chi_{\mathrm{tot}}+ip\chi_{\mathrm{net}}.

Next, (ω+)2=(a2−p2−2iap)(χ+)2(\omega^{+})^{2}=(a^{2}-p^{2}-2iap)(\chi^{+})^{2} and (ω−)2=(a2−p2+2iap)(χ−)2(\omega^{-})^{2}=(a^{2}-p^{2}+2iap)(\chi^{-})^{2}, so with qq from (31) and r=c+(χ+)2−c−(χ−)2r=c^{+}(\chi^{+})^{2}-c^{-}(\chi^{-})^{2},

μ2=(a2−p2)q−2iapr,μ12=a2χtot2−p2χnet2−2iapχtotχnet.\mu_{2}=(a^{2}-p^{2})q-2iapr, \qquad\mu_{1}^{2}=a^{2}\chi_{\mathrm{tot}}^{2}-p^{2}\chi_{\mathrm{net}}^{2}-2iap\chi_{\mathrm{tot}}\chi_{\mathrm{net}}.

Therefore,

μ2−μ12=a2(q−χtot2)−p2(q−χnet2)−2iap(r−χtotχnet)=a2vtot−p2vnet−2iapκ,\mu_{2}-\mu_{1}^{2}=a^{2}(q-\chi_{\mathrm{tot}}^{2})-p^{2}(q-\chi_{\mathrm{net}}^{2})-2iap(r-\chi_{\mathrm{tot}}\chi_{\mathrm{net}})=a^{2}v_{\mathrm{tot}}-p^{2}v_{\mathrm{net}}-2iap\kappa,

where we used the identity κ=r−χtotχnet\kappa=r-\chi_{\mathrm{tot}}\chi_{\mathrm{net}}.

Combining these expressions,

log⁡S(t)=−(R+aχtot)t+ipχnett+12(a2vtot−p2vnet−2iapκ)t2+O(t3).\log S(t)=-(R+a\chi_{\mathrm{tot}})t+ip\chi_{\mathrm{net}}t+\frac{1}{2}(a^{2}v_{\mathrm{tot}}-p^{2}v_{\mathrm{net}}-2iap\kappa)t^{2}+O(t^{3}).

Taking the real part yields log⁡∣S(t)∣\log\lvert S(t)\rvert in (34), and the imaginary part yields the unwrapped phase arg⁡S(t)\arg S(t) in (35). Dividing (35) by ptpt gives (36).

Proofs for Section 3.1

Proof of Theorem 1. Substituting c=cf1+(1−3cf)q\mathbf{c}=c_f\mathbf{1}+(1-3c_f)\mathbf{q} into the objective gives Bc−y=(1−3cf)(Bq−y~)\mathbf{B}\mathbf{c}-\mathbf{y}=(1-3c_f)(\mathbf{B}\mathbf{q}-\tilde{\mathbf{y}}). Since q0=1−x+−x−\mathbf{q}^{0}=\mathbf{1}-\mathbf{x}^{+}-\mathbf{x}^{-} (with x=(x+,x−)T\mathbf{x}=(\mathbf{x}^{+},\mathbf{x}^{-})^{\mathsf{T}}), we have Bq=b0+Dx\mathbf{B}\mathbf{q}=\mathbf{b}^{0}+\mathbf{D}\mathbf{x}, so Bq−y~=Dx−z\mathbf{B}\mathbf{q}-\tilde{\mathbf{y}}=\mathbf{D}\mathbf{x}-\mathbf{z}. For real x\mathbf{x},

∥Dx−z∥2=xTGx−2hTx+∥z∥2,\lVert\mathbf{D}\mathbf{x}-\mathbf{z}\rVert^{2}=\mathbf{x}^{\mathsf{T}}\mathbf{G}\mathbf{x}-2\mathbf{h}^{\mathsf{T}}\mathbf{x}+\lVert\mathbf{z}\rVert^{2},

because the imaginary part of the Hermitian matrix DHD\mathbf{D}^{\mathsf{H}}\mathbf{D} is antisymmetric and vanishes in quadratic forms. Hence, minimizing the objective is equivalent to (20), and G\mathbf{G} is positive semidefinite because xTGx=∥Dx∥2\mathbf{x}^{\mathsf{T}}\mathbf{G}\mathbf{x}=\lVert\mathbf{D}\mathbf{x}\rVert^{2}.

On each boundary edge of T\mathcal{T}, the cost reduces to a one-dimensional quadratic function constrained to [0,1][0,1]. The minimizers over the edges and the unconstrained interior minimizer G−1h\mathbf{G}^{-1}\mathbf{h} (when G\mathbf{G} is invertible and the point lies in T\mathcal{T}) form the candidate set C\mathcal{C}. When G\mathbf{G} is singular, let n≠0\mathbf{n}\ne0 satisfy Gn=0\mathbf{G}\mathbf{n}=0, then hTn=0\mathbf{h}^{\mathsf{T}}\mathbf{n}=0 because Dn=0\mathbf{D}\mathbf{n}=0 implies hTn=Re⁡((Dn)Hz)=0\mathbf{h}^{\mathsf{T}}\mathbf{n}=\operatorname{Re}((\mathbf{D}\mathbf{n})^{\mathsf{H}}\mathbf{z})=0. The objective is thus constant along lines parallel to n\mathbf{n}, which intersect the boundary at an edge candidate achieving the same minimum value. Taking the best among all candidates yields the global minimum.

Proof of Proposition 1. Since s=e−Rt⊙β\mathbf{s}=e^{-\mathbf{R}t}\odot\boldsymbol{\beta} and e−Rte^{-\mathbf{R}t} is real and positive, log⁡s=−Rt+log⁡∣β∣+iarg⁡β\log\mathbf{s}=-\mathbf{R}t+\log|\boldsymbol{\beta}|+i\arg\boldsymbol{\beta}. The real part of log⁡s−log⁡y\log\mathbf{s}-\log\mathbf{y} is −(Rt+ρ)-(\mathbf{R}t+\boldsymbol{\rho}), while the imaginary part (arg⁡β−arg⁡y)(\arg\boldsymbol{\beta}-\arg\mathbf{y}) is independent of R\mathbf{R}. The cost function is thus a convex parabola in R\mathbf{R}, (tTt)R2+2(tTρ)R+const(t^{\mathsf{T}}t)\mathbf{R}^{2}+2(t^{\mathsf{T}}\boldsymbol{\rho})\mathbf{R}+\mathrm{const}, whose unconstrained minimizer is −tTρ/tTt-t^{\mathsf{T}}\boldsymbol{\rho}/t^{\mathsf{T}}t. Projecting onto [0,Rmax⁡][0,R_{\max}] yields (22).

Proof of Proposition 2. Away from branch cuts where the wrapped phase residual satisfies ∣arg⁡(seyˉe)∣<π|\arg(s_{\mathrm{e}}\bar{y}_{\mathrm{e}})|<\pi, the partial derivative of r\mathbf{r} with respect to χ±\chi^{\pm} is Im⁡(∂log⁡s/∂χ±)\operatorname{Im}(\partial\log\mathbf{s}/\partial\chi^{\pm}), even if arg⁡se\arg s_{\mathrm{e}} crosses ±π\pm\pi. From (9), ∂b±/∂χ±=u±t⊙b±\partial\mathbf{b}^{\pm}/\partial\chi^{\pm}=\mathbf{u}^{\pm}t\odot\mathbf{b}^{\pm}, yielding ∂log⁡s/∂χ±=u±(s±⊘s)⊙t=v±\partial\log\mathbf{s}/\partial\chi^{\pm}=\mathbf{u}^{\pm}(\mathbf{s}^{\pm}\oslash\mathbf{s})\odot t=\mathbf{v}^{\pm} and leading to the Jacobian columns in (24). If cf±≥cf>0c_f^{\pm}\ge c_f>0 and TEe>0\mathrm{TE}_{\mathrm{e}}>0, every entry of v±\mathbf{v}^{\pm} is nonzero, ensuring σˉ>0\bar{\sigma}>0. The matrix M=JTJ+λσˉI\mathbf{M}=\mathbf{J}^{\mathsf{T}}\mathbf{J}+\lambda\bar{\sigma}\mathbf{I} is therefore positive definite for all λ>0\lambda>0, guaranteeing that JTr≠0\mathbf{J}^{\mathsf{T}}\mathbf{r}\ne0 defines a strict descent direction for ∥r∥2\lVert\mathbf{r}\rVert^{2}.

Algorithm

Algorithm 1 CANON-QSM

Input: yy, tt, u±u^{\pm}, pp, χmax⁡\chi_{\max}, Rmax⁡R_{\max}, TT, KK, nLMn_{\mathrm{LM}}, floor schedule {cf(ℓ)}ℓ=1T\{c_f^{(\ell)}\}_{\ell=1}^{T}, λ0\lambda_0, ν\nu

Output: PCS, DCS, χcomp\chi_{\mathrm{comp}}, and (c,R,χ)(c, R, \chi)

  1. Initialize KK restarts (χ,R)(\chi, R) (1 deterministic, K−1K-1 random), λ←λ0\lambda\leftarrow\lambda_0

  2. for ℓ=1,…,T\ell=1,\ldots,T do

    (A) Concentrations

    1. cf←cf(ℓ)c_f \leftarrow c_f^{(\ell)}, b±←exp⁡((−R+u±χ±)t)\mathbf{b}^{\pm} \leftarrow\exp((-R+u^{\pm}\chi^{\pm})t), b0←exp⁡(−Rt)\mathbf{b}^{0} \leftarrow\exp(-Rt), d±←b±−b0\mathbf{d}^{\pm} \leftarrow\mathbf{b}^{\pm}-\mathbf{b}^{0}

    1. z←(y−cfB1)/(1−3cf)−b0\mathbf{z} \leftarrow(\mathbf{y}-c_f\mathbf{B1})/(1-3c_f)-\mathbf{b}^{0}, G←Re⁡(DHD)\mathbf{G} \leftarrow\operatorname{Re}(\mathbf{D}^{\mathrm{H}}\mathbf{D}), h←Re⁡(DHz)\mathbf{h} \leftarrow\operatorname{Re}(\mathbf{D}^{\mathrm{H}}\mathbf{z})

    1. τ←clip⁡[0,1](Re⁡(eHw)/max⁡(∥e∥2,ϵ))\tau\leftarrow\operatorname{clip}_{[0,1]}(\operatorname{Re}(\mathbf{e}^{\mathrm{H}}\mathbf{w})/\max(\lVert\mathbf{e}\rVert^{2},\epsilon)) with e=d+−d−\mathbf{e}=\mathbf{d}^{+}-\mathbf{d}^{-}, w=z−d−\mathbf{w}=\mathbf{z}-\mathbf{d}^{-}

    1. τ1←clip⁡[0,1](h1/G11)\tau_1 \leftarrow\operatorname{clip}_{[0,1]}(h_1/G_{11}) if G11>0G_{11}>0 else 00, τ2←clip⁡[0,1](h2/G22)\tau_2 \leftarrow\operatorname{clip}_{[0,1]}(h_2/G_{22}) if G22>0G_{22}>0 else 00

    1. C←{(0,τ2),(τ1,0),(τ,1−τ)}\mathcal{C}\leftarrow\{(0,\tau_2),(\tau_1,0),(\tau,1-\tau)\}, if det⁡G>0\det\mathbf{G}>0 and G−1h∈T\mathbf{G}^{-1}\mathbf{h}\in\mathcal{T} then C←C∪{G−1h}\mathcal{C}\leftarrow\mathcal{C}\cup\{\mathbf{G}^{-1}\mathbf{h}\}

    1. x⋆←arg⁡min⁡x∈C∥Dx−z∥2\mathbf{x}^{\star}\leftarrow\arg\min_{\mathbf{x}\in\mathcal{C}}\lVert\mathbf{D}\mathbf{x}-\mathbf{z}\rVert^{2}, c←cf1+(1−3cf)(x1⋆,x2⋆,1−x1⋆−x2⋆)T\mathbf{c}\leftarrow c_f\mathbf{1}+(1-3c_f)(x_1^{\star},x_2^{\star},1-x_1^{\star}-x_2^{\star})^{\mathsf{T}}

    (B) Baseline relaxation

    1. β←c+eu+χ+t+c−eu−χ−t+c01\boldsymbol{\beta}\leftarrow c_{+}e^{u_{+}\chi^{+}t}+c_{-}e^{u_{-}\chi^{-}t}+c_{0}\mathbf{1}, ρ←log⁡∣y∣−log⁡∣β∣\rho\leftarrow\log|\mathbf{y}|-\log|\boldsymbol{\beta}|

    1. R←clip⁡[0,Rmax⁡](−tTρ/tTt)R\leftarrow\operatorname{clip}_{[0,R_{\max}]}(-\mathbf{t}^{\mathsf{T}}\rho/\mathbf{t}^{\mathsf{T}}\mathbf{t})

    (C) Susceptibilities

    1. for n=1,…,nLMn=1,\ldots,n_{\mathrm{LM}} do

    1. s±←c±exp⁡((−R+u±χ±)t)s^{\pm}\leftarrow c_{\pm}\exp((-R+u^{\pm}\chi^{\pm})t), s←s++s−+c0e−Rts\leftarrow s^{+}+s^{-}+c_{0}e^{-Rt}

    1. r←[log⁡∣s∣−log⁡∣y∣,arg⁡(s⊙yˉ)]\mathbf{r}\leftarrow[\log|s|-\log|\mathbf{y}|,\arg(s\odot\bar{\mathbf{y}})] ▷ wrapped residual

    1. v±←u±(s±⊘s)⊙t\mathbf{v}^{\pm}\leftarrow u^{\pm}(s^{\pm}\oslash s)\odot\mathbf{t}, J←[Re⁡v+Re⁡v−Im⁡v+Im⁡v−]\mathbf{J}\leftarrow\begin{bmatrix}\operatorname{Re}\mathbf{v}^{+}&\operatorname{Re}\mathbf{v}^{-}\\\operatorname{Im}\mathbf{v}^{+}&\operatorname{Im}\mathbf{v}^{-}\end{bmatrix}, σˉ←12tr⁡(JTJ)\bar{\sigma}\leftarrow\frac{1}{2}\operatorname{tr}(\mathbf{J}^{\mathsf{T}}\mathbf{J})

    1. δ←−(JTJ+λσˉI)−1JTr\boldsymbol{\delta}\leftarrow-(\mathbf{J}^{\mathsf{T}}\mathbf{J}+\lambda\bar{\sigma}\mathbf{I})^{-1}\mathbf{J}^{\mathsf{T}}\mathbf{r} ▷ 2×22\times2, Cramer’s rule

    1. χtrial←(clip⁡[0,χmax⁡](χ++δ+),clip⁡[−χmax⁡,0](χ−+δ−))T\chi_{\mathrm{trial}}\leftarrow(\operatorname{clip}_{[0,\chi_{\max}]}(\chi^{+}+\delta^{+}),\operatorname{clip}_{[-\chi_{\max},0]}(\chi^{-}+\delta^{-}))^{\mathsf{T}}

    1. if ∥r(χtrial)∥2<∥r(χ)∥2\lVert\mathbf{r}(\chi_{\mathrm{trial}})\rVert^{2}<\lVert\mathbf{r}(\chi)\rVert^{2} then

    1. χ←χtrial\chi\leftarrow\chi_{\mathrm{trial}}, λ←max⁡{λ/ν,λmin⁡}\lambda\leftarrow\max\{\lambda/\nu,\lambda_{\min}\}

    1. else

    1. λ←min⁡{νλ,λmax⁡}\lambda\leftarrow\min\{\nu\lambda,\lambda_{\max}\}

    1. end if

    1. end for

    Selection

    1. Among the KK restarts, track the state minimizing S=∥s−y∥2/∥y∥2+∥r∥2/NeS=\lVert\mathbf{s}-\mathbf{y}\rVert^{2}/\lVert\mathbf{y}\rVert^{2}+\lVert\mathbf{r}\rVert^{2}/N_e

    1. end for

    1. With the final stored state and Ψ(m)=1TU(arg⁡m)/(p1Tt)\Psi(\mathbf{m})=\mathbf{1}^{\mathsf{T}}\mathbf{U}(\arg\mathbf{m})/(p\mathbf{1}^{\mathsf{T}}\mathbf{t}):

    PCS←Ψ(c+b++(c0+c−)b0)\mathrm{PCS}\leftarrow\Psi(c^{+}\mathbf{b}^{+}+(c^{0}+c^{-})\mathbf{b}^{0}), DCS←Ψ(c−b−+(c0+c+)b0)\mathrm{DCS}\leftarrow\Psi(c^{-}\mathbf{b}^{-}+(c^{0}+c^{+})\mathbf{b}^{0})

    χcomp←Ψ(Bc)\chi_{\mathrm{comp}}\leftarrow\Psi(\mathbf{B}\mathbf{c})

Besides quantities TT and KK, Algorithm 1 uses a small number of fixed constants, listed in Table 2. With the exception of χmax⁡\chi_{\max}, which is chosen per dataset (Sec. 3.1.3), they were set once and kept identical in every experiment of this work, at both field strengths and for both the in silico and in vivo data:

SymbolRoleValue
TTouter alternating iterations20
KKrestarts per voxel20
nLMn_{\mathrm{LM}}LM steps per outer iteration10
λ0\lambda_0initial LM damping10−210^{-2}
ν\nudamping growth/shrink factor2
λmin⁡,λmax⁡\lambda_{\min},\lambda_{\max}damping limits10−7,10710^{-7}, 10^{7}
cf(1)→cf(T)c_f^{(1)}\rightarrow c_f^{(T)}concentration floor2×10−2→10−42\times10^{-2}\rightarrow10^{-4}
Rmax⁡R_{\max}upper bound of the baseline relaxation rate2000 s−12000\ \mathrm{s}^{-1}
χmax⁡\chi_{\max}susceptibility boundper dataset (0.48 ppm in silico)

Table 2. Fixed constants of CANON-QSM (Algorithm 1).

References

  1. [1]Robert W Brown, Y-C Norman Cheng, E Mark Haacke, Michael R Thompson, and Ramesh Venkatesan. Magnetic resonance imaging: physical principles and sequence design. John Wiley & Sons, 2014.DOI
  2. [2]Berkin Bilgic, Mauro Costagli, Kwok-Shing Chan, Jeff Duyn, Christian Langkammer, Jongho Lee, Xu Li, Chunlei Liu, José P. Marques, et al. Recommended implementation of quantitative susceptibility mapping for clinical research in the brain: a consensus of the ismrM electro-magnetic tissue properties study group. Magnetic Resonance in Medicine, 91(5):1834–1862, 2024.
  3. [3]Jae Kyu Choi, Hyoung Suk Park, Shuai Wang, Yi Wang, and Jin Keun Seo. Inverse problem in quantitative susceptibility mapping. SIAM Journal on Imaging Sciences, 7(3):1669–1689, 2014.DOI
  4. [4]H.-G. Shin et al. χ-separation: Magnetic susceptibility source separation toward iron and myelin mapping in the brain. NeuroImage, 240:118371, 2021.DOI
  5. [5]J. Chen, N.-J. Gong, K. T. Chaim, M. C. García Otaduy, and C. Liu. Decompose quantitative susceptibility mapping (QSM) to sub-voxel diamagnetic and paramagnetic components based on gradient-echo MRI data. NeuroImage, 242:118477, 2021.DOI
  6. [6]D. A. Yablonskiy and E. M. Haacke. Theory of NMR signal behavior in magnetically inhomogeneous tissues: The static dephasing regime. Magnetic Resonance in Medicine, 32(6):749–763, 1994.
  7. [7]F. Otsuka, G. Debiasi, C. Liu, C. Salmon, and M. C. García Otaduy. A deep learning approach to solve the DECOMPOSE-QSM model for source separation of magnetic susceptibility. In Proc. ISMRM Annual Meeting, 2026.
  8. [8]P. Fuchs, J. Chen, C. Liu, and K. Shmueli. Practical considerations for DECOMPOSE-QSM in the human brain. In Proc. ISMRM Annual Meeting, 2024.DOI
  9. [9]Rares Salomir, Baudouin Denis De Senneville, and Chrit TW Moonen. A fast calculation method for magnetic field inhomogeneity due to an arbitrary distribution of bulk susceptibility. Concepts in Magnetic Resonance Part B: Magnetic Resonance Engineering: An Educational Journal, 19(1):26–34, 2003.DOI
  10. [10]José P Marques and Richard Bowtell. Application of a Fourier-based method for rapid calculation of field inhomogeneity due to spatial variation of magnetic susceptibility. Concepts in Magnetic Resonance Part B: Magnetic Resonance Engineering: An Educational Journal, 25(1):65–78, 2005.
  11. [11]G. H. Golub and V. Pereyra. The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate. SIAM Journal on Numerical Analysis, 10(2):413–432, 1973.DOI
  12. [12]K. Itoh. Analysis of the phase unwrapping algorithm. Applied Optics, 21(14):2470, 1982.DOI
  13. [13]K. Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 2(2):164–168, 1944.DOI
  14. [14]D. W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial and Applied Mathematics, 11(2):431–441, 1963.DOI
  15. [15]J. J. Moré. The Levenberg–Marquardt algorithm: Implementation and theory. In G. A. Watson, editor, Numerical Analysis, pages 105–116. Springer, 1978.DOI
  16. [16]A. Paszke et al. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), pages 8024–8035, 2019.arxiv.org/abs/1912.01703
  17. [17]D. Ridani, B. De Leener, and E. Alonso-Ortiz. A realistic in silico brain phantom for quantifying susceptibility anisotropy-induced error in susceptibility separation. Magnetic Resonance in Medicine, 2026.DOI
  18. [18]B. Dymerska et al. Phase unwrapping with a rapid opensource minimum spanning tree algorithm (ROMEO). Magnetic Resonance in Medicine, 85(4):2294–2308, 2021.DOI
  19. [19]Tian Liu, Ildar Khalidov, Ludovic de Rochefort, Pascal Spincemaille, Jing Liu, A John Tsiouris, and Yi Wang. A novel background field removal method for MRI using projection onto dipole fields. NMR in Biomedicine, 24(9):1129–1136, 2011.DOI
  20. [20]Carlos Milovic, Berkin Bilgic, Bo Zhao, Christian Langkammer, Cristian Tejos, and Julio Acosta-Cabronero. Weak-harmonic regularization for quantitative susceptibility mapping. Magnetic resonance in medicine, 81(2):1399–1411, 2019.DOI
  21. [21]I. Contreras-Zúñiga, C. Tejos, and C. Milovic. Demixing streaking for quantitative susceptibility mapping. In IEEE 23rd International Symposium on Biomedical Imaging (ISBI), 2026.
  22. [22]A. W. Stewart, K. Eckstein, T. T. Dao, and S. Bollmann. QSM-CI: An automated continuous QSM challenge. In Proc. ISMRM Annual Meeting, 2024.DOI
  23. [23]Carlos Milovic, Cristian Tejos, Javier Silva, Karin Shmueli, and Pablo Irarrazaval. XSIM: A structural similarity index measure optimized for MRI QSM. Magnetic resonance in medicine, 93(1):411–421, 2025.

Paper details

vlnc:W_gfezpjjj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents