Grouped Isochromat MR Simulation

• Grouped isochromat simulation is a method that consolidates Bloch-equation updates for isochromats sharing identical parameters.
• It partitions isochromats based on key parameters like T1, T2, and gradient-induced phase factors to reduce redundant computations.
• Evaluations demonstrate simulation speedups ranging from 3 to 72 times, significantly enhancing efficiency in large-scale MR simulations.

Searching arXiv for the primary paper and closely related MR simulation work on grouped/compact isochromat representations. Simulation using grouped isochromats is a Bloch-equation magnetic resonance simulation strategy that relaxes the conventional assumption that every isochromat must always be simulated independently. In the formulation introduced as a “combined update with grouped isochromats,” isochromats are partitioned into groups that share the same subsequence-dependent evolution operator, so that expensive transition calculations can be performed once per group rather than once per isochromat. The method is sequence-aware rather than universal: it is exact when the grouped isochromats truly share the same update matrix or the same analytic decay and phase factors over a subsequence, and it is especially targeted at with-RF and with-ADC subsequences having single-axis gradients or no gradients. Reported evaluations in VMRscan showed simulation-time reductions of 3 to 72 times, including large-scale fast spin echo (FSE) and echo-planar imaging (EPI) cases with 27.5 million isochromats (Takeshima, 5 Aug 2025).

1. Problem setting and conceptual basis

Conventional MR simulators typically discretize an object into many isochromats and numerically update each isochromat’s magnetization through the pulse sequence. The computational burden is substantial because realistic simulations often require millions of isochromats, and RF-containing intervals may themselves be discretized into microsecond-scale time steps. The method using grouped isochromats starts from the observation that many isochromats can evolve identically over a given subsequence, even though they remain distinct state instances. When that occurs, the expensive shared transition can be computed once and reused across all members of the group (Takeshima, 5 Aug 2025).

For the $k$-th isochromat, the Bloch model is written as

$M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$

with

$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$

where

$M_{static}(k)=(0,0,M_0(k))^T,$

$$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$

and

$B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$

The sequence contribution is decomposed into the RF term

$B_1(k,t)=B_x(k,t)+iB_y(k,t),$

and the longitudinal term

$B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$

For static isochromats,

$r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$

The simulator partitions pulse sequences into three subsequence classes: with-RF subsequences, with-ADC subsequences, and relaxation subsequences. Their stated costs in a straightforward simulator are:

• naïve with-RF: $O(N_{RF}K)$,
• with-ADC: $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$0,
• relaxation: $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$1.

Here $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$2 is the number of isochromats, $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$3 the number of RF update steps, $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$4 the number of ADC samples, and $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$5 the number of coils. This asymptotic structure explains why speed becomes a limiting factor for long RF pulse trains, large-scale phantoms, FSE and EPI sequences, and simulations using subvoxel refinement.

2. Group construction and validity conditions

Grouped simulation is subsequence-type dependent. The method does not claim that arbitrary isochromats can always be grouped. Grouping is valid only when the subsequence has a sufficiently simple gradient structure and the grouped isochromats share all quantities that determine the update operator $M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$6. In the theory section, this requirement is expressed as equality of

$M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$7

within a valid group (Takeshima, 5 Aug 2025).

The paper focuses on practical gradient classes: Gx-only, Gy-only, Gz-only, and no-grads. Under these cases, the grouping key simplifies because only the coordinate along the active gradient enters the gradient-induced phase term.

Gradient type	Group-common quantities
no-grads	$M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$8
Gx-only	$M(k,t)=(M_x(k,t),M_y(k,t),M_z(k,t))^T$9
Gy-only	$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$0
Gz-only	$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$1

For a Gx-only subsequence,

$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$2

so isochromats may differ in $$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$3 and $$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$4 and still be grouped. The same logic applies to Gy-only and Gz-only subsequences. If no gradient is active, then

$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$5

so only $$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$6 must match.

A recurring misconception is that grouping is a general-purpose aggregation of nearby spins. The stated method is narrower. If more than one gradient component is nonzero in a way that makes $$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$7 depend on multiple coordinates, grouping becomes much more restrictive because multiple spatial coordinates must match simultaneously. The discussion identifies this as the main remaining limitation and the reason the method currently targets single-axis gradient subsequences.

3. Mathematical formulation of grouped updates

The simulator uses the augmented state

$$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$8

For a small interval $$\frac{dM(k,t)}{dt}=\gamma \big(M(k,t) \times B(k,t)\big)+C(k)(M_{static}(k) - M(k,t)),$$9 with piecewise constant field,

$M_{static}(k)=(0,0,M_0(k))^T,$0

Over a with-RF subsequence broken into $M_{static}(k)=(0,0,M_0(k))^T,$1 small steps, the combined transition matrix is

$M_{static}(k)=(0,0,M_0(k))^T,$2

The conventional combined-transition approach still computes one such matrix per isochromat, so preparation remains $M_{static}(k)=(0,0,M_0(k))^T,$3. Grouping changes this preparation cost to $M_{static}(k)=(0,0,M_0(k))^T,$4, while the final state update remains $M_{static}(k)=(0,0,M_0(k))^T,$5 because each member still needs a matrix-vector multiplication (Takeshima, 5 Aug 2025).

Appendix A gives the implementation form

$M_{static}(k)=(0,0,M_0(k))^T,$6

The relaxation matrix is

$M_{static}(k)=(0,0,M_0(k))^T,$7

and the rotation matrix is

$M_{static}(k)=(0,0,M_0(k))^T,$8

where

$M_{static}(k)=(0,0,M_0(k))^T,$9

These expressions make the grouping criterion explicit: if two isochromats share the same $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$0, $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$1, and $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$2 at each time step in the subsequence, then they share $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$3, $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$4, $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$5, and therefore the same combined transition matrix.

For static isochromats and $$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$6, the analytic no-RF evolution is

$$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$7

$$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$8

The received signal is

$$C(k)=\operatorname{diag}\!\left(-\frac{1}{T_2(k)},-\frac{1}{T_2(k)},-\frac{1}{T_1(k)}\right),$$9

The most distinctive grouped derivation concerns ADC intervals. The total signal is rewritten as

$B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$0

with

$B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$1

For a group with shared $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$2 and shared $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$3,

$B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$4

This grouped-ADC formula implies that the simulator can evolve one complex block signal per group and coil during the ADC window, rather than updating every member isochromat at every ADC sample.

4. Algorithmic workflow and implementation

The method was implemented in the CPU-based simulator VMRscan. The workflow begins with preprocessing and grouping. For each isochromat, the simulator extracts the subsequence-dependent values that must be equal; the resulting tuple acts as a grouping key, and isochromats with the same key are collected into the same group. The paper assumes this grouping is performed once before simulation (Takeshima, 5 Aug 2025).

The sequence is then partitioned into with-RF, with-ADC, and relaxation subsequences. For with-RF and with-ADC subsequences, the simulator checks whether the active gradients fit one of the recognized types: Gx-only, Gy-only, Gz-only, or no-grads. If so, grouped simulation is used; if not, the simulator falls back to the conventional per-isochromat method from the author’s earlier work.

For with-RF subsequences that satisfy the grouping rule, the simulator computes one combined transition matrix $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$5 per group and applies it to each member’s magnetization vector. The acceleration is therefore concentrated in the preparation stage rather than in the final member-wise application.

For with-ADC subsequences that satisfy the grouping rule, the stated procedure has three steps:

1. At $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$6, compute

$B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$7

using the group sum.

1. Compute the $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$8 sample values using the grouped signal equations for the $B(k,t)=(B_x(k,t),B_y(k,t),B_z(k,t))^T.$9 group states.
2. After the ADC subsequence, update the actual magnetization states $B_1(k,t)=B_x(k,t)+iB_y(k,t),$0 for all $B_1(k,t)=B_x(k,t)+iB_y(k,t),$1 isochromats using the analytic no-RF formulas, because subsequent subsequences still require the individual states.

This yields the stated grouped-ADC complexity:

• Step 1: $B_1(k,t)=B_x(k,t)+iB_y(k,t),$2,
• Step 2: $B_1(k,t)=B_x(k,t)+iB_y(k,t),$3,
• Step 3: $B_1(k,t)=B_x(k,t)+iB_y(k,t),$4.

A second misconception addressed by the implementation is that grouped simulation eliminates persistent per-isochromat state storage. It does not. Signal formation during ADC is grouped, but individual magnetization states are still maintained and later updated for all isochromats. Relaxation subsequences are processed conventionally, because their cost is already low.

The method assumes static isochromats for the analytic no-RF formulas used in with-ADC and relaxation subsequences, recognizes only one active gradient axis or no gradients, and may use preprocessing or clustering to reduce the number of distinct $B_1(k,t)=B_x(k,t)+iB_y(k,t),$5 combinations.

5. Performance, scaling, and accuracy trade-offs

The evaluation in VMRscan used a CPU with 8 performance cores, 16 efficient cores, 32 processor threads, base frequencies 3.2 GHz and 2.4 GHz, and boost up to 6.0 GHz. Simulations were tested with and without SIMD instructions and always with multi-threading. Two phantoms were evaluated: a circles phantom designed to be favorable for grouping, and a brain phantom derived from measured multi-slice brain data and clustered using bisecting K-means in the space $B_1(k,t)=B_x(k,t)+iB_y(k,t),$6 (Takeshima, 5 Aug 2025).

For the circles phantom, the reported dimensions were $B_1(k,t)=B_x(k,t)+iB_y(k,t),$7, spatial size $B_1(k,t)=B_x(k,t)+iB_y(k,t),$8, and $B_1(k,t)=B_x(k,t)+iB_y(k,t),$9 isochromats, with 10 regions of constant $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$0, $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$1, and $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$2. The numbers of groups were 10 for no-grads, 1128 for Gx-only, 1124 for Gy-only, and 100 for Gz-only. Pulseq implementations of T1W SE, T2W SE, EPI, and FSE were evaluated. Table 2 reports 512 with-RF and 256 with-ADC subsequences for T1W/T2W SE, 2 with-RF and 96 with-ADC for EPI, and 312 with-RF and 260 with-ADC for FSE.

Selected timings illustrate how strongly the speedup depends on sequence structure. For the circles phantom, without SIMD, T1W was 55.04 s ungrouped vs 2.72 s grouped, EPI was 17.99 s vs 0.25 s, and FSE was 58.09 s vs 2.54 s. With SIMD, the corresponding values were 10.10 s vs 1.41 s for T1W, 3.09 s vs 0.15 s for EPI, and 10.55 s vs 1.32 s for FSE. For the brain phantom with clustering and 3.5 million isochromats, the ungrouped SIMD times were 28.51 s for T1W, 29.07 s for T2W, 8.88 s for EPI, and 29.35 s for FSE; grouped with 256 clusters and SIMD, the times were 6.18 s, 7.69 s, 1.01 s, and 5.60 s, respectively. The paper summarizes these results as overall speedups of 4 to 72 times in simulation set #1.

A second scaling study varied subvoxeling from $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$3 to $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$4, giving isochromat counts from $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$5 to $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$6. For the largest case, using SIMD and multi-threading, the conventional method simulated EPI in 66.38 s and FSE in 208.44 s, while grouped simulation with brain-256 required 7.11 s for EPI and 38.11 s for FSE, corresponding to about 9.34\times and 5.47\times. Across these larger-scale tests, the reported speedup was 3 to 14 times.

The paper states that runtime remained approximately proportional to the number of isochromats even with grouping. This suggests that the residual $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$7 member-wise state update remains significant after the shared $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$8 and $B_z(k,t)=G(t)r(k,t)+\Delta B_0(k).$9 components have been reduced to dependence on $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$0.

Accuracy requires separating exactness of the grouped update from approximation introduced by phantom preprocessing. Under the grouping conditions, the grouped method itself is exact. Approximation enters when the phantom is clustered to reduce the number of distinct parameter combinations. In the brain phantom experiments, 256 clusters gave visually close results to the reference in EPI, FSE, T1W, and T2W, errors were larger for long-TR sequences such as T2W and EPI, and errors decreased as the number of clusters increased. The stated trade-off is therefore between speed and faithfulness to the original parameter maps, rather than between grouped computation and its own exact grouped model.

6. Relation to adjacent compact isochromat representations

Grouped-isochromat simulation belongs to a broader family of strategies that replace many explicit sub-isochromat updates with a more compact representation of shared dynamics. A closely related but distinct example is the simulation of effective transverse relaxation $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$1 proposed by Hidenori Takeshima, where the unresolved Lorentzian frequency ensemble is represented by one spatial isochromat augmented with derivative states rather than by grouping discrete isochromats with identical keys (Takeshima, 27 Jan 2026).

That work is explicitly described as not a standard grouped-isochromat method. Instead, it uses a compact continuous ensemble representation of a Lorentzian frequency distribution through a linear phase model and the propagated partial derivatives

$r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$2

Its central relation is

$r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$3

and its practical motivation is that conventional simulators may require 100+ isochromats per point in the frequency dimension to represent a Lorentzian realistically. By propagating the augmented state

$r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$4

and applying analytic solutions and combined transitions, the method reported that simulations with $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$5 were only 2.0 to 2.7 times longer than corresponding simulations without $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$6, while analytic solutions accelerated 19 times and combined transitions accelerated up to 17 times.

The connection between the two papers is conceptual rather than algorithmic. Both reduce explicit ensemble size by exploiting shared structure, and both preserve persistent state information required for subsequent sequence evolution. The grouped-isochromat method shares transition preparation and grouped ADC evolution across discrete spatial isochromats that have identical subsequence dynamics. The $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$7 method instead replaces a many-frequency discretization by one representative object with extra derivative information. A plausible implication is that “simulation using grouped isochromats” is best understood as one member of a larger class of compact-state MR simulation methods, in which the principal design question is how much internal ensemble structure can be preserved without retaining every explicit sub-isochromat.

Under that broader view, the grouped-isochromat method is characterized by exact reuse under strict equality conditions and by sequence-aware grouping rules tied to single-axis gradients. Its practical value is highest when $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$8, the phantom contains repeated or clustered $r(k,t)=r(k)=(r_x(k),r_y(k),r_z(k))^T.$9 values, and the sequence spends substantial time in with-RF or with-ADC subsequences of recognized gradient types. Its main limitation remains the rapid growth in group count when multiple gradient components are active simultaneously, which is why extending the method to more general gradient types is described as “not easy to accelerate simulations efficiently with such groups” (Takeshima, 5 Aug 2025).