arxiv: 2604.14788 · v1 · submitted 2026-04-16 · 💻 cs.AI

Recognition: unknown

Sequence Search: Automated Sequence Design using Neural Architecture Search

Rokgi Hong , Hongjun An , Sooyeon Ji , Jongho Lee

Authors on Pith no claims yet

Pith reviewed 2026-05-10 10:41 UTC · model grok-4.3

classification 💻 cs.AI

keywords MRI sequence designneural architecture searchBloch simulatorpulse sequence optimizationautomated MR sequencegradient-based learningmagnetic resonance imaging

0 comments

The pith

A neural architecture search method can automatically generate MRI pulse sequences from tissue properties and design goals alone, recovering standard sequences and finding new variants.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Sequence Search, a framework that builds MRI pulse sequences step by step using neural architecture search. It starts with basic inputs on tissue properties, scan settings, and objectives such as contrast or signal efficiency, then iteratively creates and refines sequence candidates. Each candidate is evaluated and improved through gradient descent on a differentiable Bloch simulator that models the physics of spin interactions. The approach recovered classic sequences like spin-echo and inversion recovery without being given their structures in advance. It also produced alternative designs that use less radiofrequency energy while maintaining similar imaging behavior.

Core claim

Sequence Search establishes that neural architecture search, paired with gradient-based optimization against a differentiable Bloch simulator, can produce valid MR pulse sequences that meet specified objectives without any initial sequence template or large training data, successfully replicating conventional spin-echo, T2-weighted spin-echo, and inversion recovery sequences while also identifying non-standard three-RF spin-echo-like sequences with reduced RF energy and altered refocusing phases.

What carries the argument

Neural architecture search that iteratively generates candidate pulse sequences and optimizes them via gradient descent on loss functions computed by a differentiable Bloch simulator.

If this is right

Sequence design no longer requires an expert to supply a starting template or hand-tune parameters.
New sequences can be generated for arbitrary combinations of tissue parameters and imaging goals.
Sequences with lower RF energy deposition become discoverable without explicit prior constraints.
The same search process can be applied to other objective functions such as specific contrast types or scan speed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The framework could be extended to jointly optimize sequence and reconstruction parameters in a single loop.
Validation on multiple scanner vendors would be needed before clinical deployment of any novel sequence.
Similar search methods might apply to designing sequences for other modalities if accurate differentiable simulators exist.

Load-bearing premise

The Bloch simulator must model real MRI hardware and tissue behavior closely enough that sequences optimized in simulation remain effective and safe when run on actual scanners.

What would settle it

Running one of the discovered non-standard sequences on a physical MRI scanner and measuring whether the acquired images and signal intensities match the simulator predictions within acceptable error bounds.

Figures

Figures reproduced from arXiv: 2604.14788 by Hongjun An, Jongho Lee, Rokgi Hong, Sooyeon Ji.

**Figure 1.** Figure 1: Overview of Sequence Search. Sequence Search designs a sequence within the given search space based on the input properties and design objectives, without requiring prior knowledge of sequence design principles. Sequence Search consists of three main parts. The sequence scheduler recommends a candidate sequence. The Bloch simulator processes this sequence in conjunction with the given properties, generatin… view at source ↗

**Figure 3.** Figure 3: ) were log-transformed to compress their wide dynamic range. A small constant of 0.001 ms was added prior to logarithmic transformation to prevent numerical instability when idle times approached zero. Weight parameters of the sequence operations were optimized using SGD (learning rate 0.01), and architecture parameters using ADAM [22] (learning rate 0.001). C. Bloch simulator A Bloch simulator was implem… view at source ↗

**Figure 5.** Figure 5: Results of the GM–WM contrast enhancement experiment (contrast weight = 30). (a) Optimal two-RF spin-echo identified by grid search (88°–177°, TE = 37.0 ms). (b, c) Designed sequences: (b) two-RF Hahn-echo-like sequence with a refocusing phase near 180° discovered under the identical-T1 condition; (c) inversion-prepared spinecho-like sequence discovered under the different-T1 condition (TI = 452.9 ms). (d… view at source ↗

**Figure 6.** Figure 6: Results of the CSF nulling experiment (CSF nulling weight = 0.25). (a) Optimal inversion recovery sequence identified by grid search (177°–89°; TI = 3.75 s). (b, c) Designed sequences: (b) inversionprepared gradient-echo-like sequence (TI = 3.76 s); (c) inversionprepared spin-echo-like sequence (TI = 3.34 s). (d–f) Simulated transverse-magnetization evolution for GM, WM, and CSF corresponding to (a–c) a… view at source ↗

read the original abstract

Developing an MR sequence is challenging and remains largely constrained by human intuition. Recently, AI-driven approaches have been proposed; however, most require an initial sequence for parameter optimization or extensive training datasets, limiting their general applicability. In this study, we propose "Sequence Search," an automated sequence design framework based on neural architecture search. The method takes tissue properties, imaging parameters, and design objectives as inputs and generates pulse sequences satisfying the design objectives, without requiring prior knowledge of conventional sequence structures. Sequence Search iteratively generates candidate sequences through neural architecture search and optimizes them via a differentiable Bloch simulator and objective-specific loss functions using gradient-based learning. The framework successfully replicated conventional spin-echo, T2-weighted spin-echo, and inversion recovery sequences. Less intuitive solutions were also discovered, such as three-RF spin-echo-like sequences with reduced RF energy and refocusing phases deviating from the conventional Hahn-echo. This work establishes a generalizable framework for automated MR sequence design, highlighting the potential to explore configurations beyond conventional designs based on human intuition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a clean NAS-plus-differentiable-Bloch-sim framework that recovers standard MRI sequences and some variants without seed designs, but everything stays in simulation with no hardware checks or quantitative metrics.

read the letter

The main takeaway is that Sequence Search combines neural architecture search with a differentiable Bloch simulator to generate pulse sequences directly from tissue properties and objectives. It skips the need for initial human sequences or large datasets that limited earlier AI methods for this task. That combination is the actual step forward here. They demonstrate recovery of conventional spin-echo, T2-weighted spin-echo, and inversion recovery sequences, plus some three-RF variants that use less RF energy and non-standard refocusing phases. The optimization runs via gradient descent on simulator outputs and objective losses, which keeps the process grounded in the physics model rather than fitting to known designs. The approach looks internally consistent on paper and avoids obvious circularity. The soft spots are straightforward. All reported results are simulation-only, with no error tables, no comparison to real scanner data, and no mention of how post-hoc selection was handled. Standard Bloch models leave out B1 inhomogeneity, gradient nonlinearities, and hardware SAR limits, so the less-intuitive sequences could be exploiting those gaps. The abstract gives no numbers on fidelity or reproducibility steps, which leaves the central claims thin. This is for readers working on automated MRI design or NAS applications in physics simulation. Someone exploring new sequence spaces will see a workable template, but anyone needing practical sequences should treat the novel outputs as hypotheses only. I would send it to peer review. The core framework is clear enough and the claims are falsifiable, so referees can push on the validation gap without starting from scratch.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes 'Sequence Search', a neural architecture search (NAS) framework for automated MRI pulse sequence design. It accepts tissue properties, imaging parameters, and design objectives as inputs, iteratively generates candidate sequences via NAS, and optimizes them using a differentiable Bloch simulator combined with gradient-based learning and objective-specific loss functions. The central claims are that the method replicates conventional sequences (spin-echo, T2-weighted spin-echo, inversion recovery) without prior structural knowledge and discovers less intuitive variants, such as three-RF spin-echo-like sequences with reduced RF energy and non-Hahn refocusing phases.

Significance. If the simulator-based results translate to hardware, the work would be significant for providing a generalizable, intuition-independent method to explore MRI sequence space beyond conventional designs. The combination of NAS with a differentiable Bloch simulator enables gradient-driven optimization without pre-defined targets or large training datasets, which is a methodological strength. However, the absence of quantitative metrics, error analysis, and experimental validation currently limits the assessed impact and generalizability.

major comments (2)

[Abstract] Abstract: The replication of conventional sequences and discovery of novel variants are stated without any quantitative metrics (e.g., contrast values, signal fidelity, sequence duration, or comparison to reference implementations), error bars, or ablation studies. This directly under-supports the central claim that the framework 'successfully replicated' and 'discovered' sequences.
[Abstract] Abstract (and implied Methods): The claim that less intuitive solutions (three-RF spin-echo-like sequences with reduced RF energy and non-Hahn phases) are valid rests on the differentiable Bloch simulator producing physically accurate gradients and signals. No details are given on simulator extensions for B1 inhomogeneity, gradient nonlinearity, RF pulse distortions, or SAR/hardware constraints, nor any cross-validation against real scanner data. This is load-bearing for the novelty claim, as standard Bloch models can be exploited by optimization.

minor comments (2)

[Abstract] The abstract could more precisely define the NAS search space and sequence representation (e.g., how RF pulses and gradients are encoded as architectures).
Consider adding a table or figure in the results that lists optimized sequence parameters, simulated signals, and objective losses for both conventional and novel designs to improve clarity and reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the presentation of results and clarify methodological assumptions.

read point-by-point responses

Referee: [Abstract] Abstract: The replication of conventional sequences and discovery of novel variants are stated without any quantitative metrics (e.g., contrast values, signal fidelity, sequence duration, or comparison to reference implementations), error bars, or ablation studies. This directly under-supports the central claim that the framework 'successfully replicated' and 'discovered' sequences.

Authors: We agree that the abstract would benefit from explicit quantitative support. The full manuscript includes signal evolution plots and objective function values demonstrating replication of spin-echo, T2-weighted spin-echo, and inversion recovery sequences, along with comparisons of RF energy and echo times for the discovered variants. In revision we will add specific metrics (e.g., contrast ratios, normalized signal fidelity, sequence duration) with error bars from repeated NAS runs, plus a brief ablation on the architecture search components. These will be summarized in the abstract and expanded in Results. revision: yes
Referee: [Abstract] Abstract (and implied Methods): The claim that less intuitive solutions (three-RF spin-echo-like sequences with reduced RF energy and non-Hahn phases) are valid rests on the differentiable Bloch simulator producing physically accurate gradients and signals. No details are given on simulator extensions for B1 inhomogeneity, gradient nonlinearity, RF pulse distortions, or SAR/hardware constraints, nor any cross-validation against real scanner data. This is load-bearing for the novelty claim, as standard Bloch models can be exploited by optimization.

Authors: The simulator is a standard differentiable Bloch implementation that computes exact gradients under the rotating-frame approximation for ideal rectangular RF pulses and linear gradients. We will add an explicit Methods subsection detailing these assumptions and the absence of B1 inhomogeneity, gradient nonlinearity, or SAR modeling. The discovered three-RF sequences achieve the target objective (e.g., spin-echo contrast with lower total RF energy) strictly within the simulated physics; the non-Hahn refocusing phases emerge because the optimizer is free to adjust flip angles and phases without a Hahn-echo constraint. While we acknowledge that real-hardware effects could alter performance, the optimization is not exploiting numerical artifacts but satisfying the Bloch equations under the stated model. Full scanner validation lies outside the current simulation-focused scope. revision: partial

standing simulated objections not resolved

Experimental validation on physical MRI hardware or in vivo subjects is absent; the work is limited to simulation results.

Circularity Check

0 steps flagged

No circularity: outputs derived via external simulator optimization

full rationale

The paper describes a NAS-based search that generates candidate sequences and optimizes them end-to-end using a differentiable Bloch simulator plus objective-specific losses. The resulting sequences (both conventional replications and novel variants) are produced by gradient descent on the simulator's forward model, not by re-expressing the inputs or by any self-referential definition. No equations or steps reduce the claimed discoveries to fitted parameters or prior self-citations; the simulator and NAS procedure are treated as independent computational engines. Replication of known sequences functions as external validation rather than a tautological outcome. This satisfies the default expectation of a non-circular derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard MRI physics modeling and ML optimization techniques rather than new physical postulates.

free parameters (1)

NAS hyperparameters and loss weights
The search process and objective-specific losses require multiple tunable parameters whose values are not detailed in the abstract.

axioms (1)

domain assumption The Bloch equations provide a sufficiently accurate model of spin dynamics for sequence optimization.
The differentiable simulator is built on this standard physics description.

pith-pipeline@v0.9.0 · 5476 in / 1347 out tokens · 42799 ms · 2026-05-10T10:41:23.597353+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages · 1 internal anchor

[1]

Self-play reinforcement learning guides protein engi- neering,

Y . Wanget al., “Self-play reinforcement learning guides protein engi- neering,”Nat. Mach. Intell., vol. 5, no. 8, pp. 845–860, Jul. 2023, doi: 10.1038/s42256-023-00691-9

work page doi:10.1038/s42256-023-00691-9 2023
[2]

Inverse molecular design using machine learning: Generative models for matter engineering,

B. Sanchez-Lengeling and A. Aspuru-Guzik, “Inverse molecular design using machine learning: Generative models for matter engineering,” Science, vol. 361, no. 6400, pp. 360–365, Jul. 2018, doi: 10.1126/sci- ence.aat2663

work page doi:10.1126/sci- 2018
[3]

Deep reinforcement learning-designed radiofrequency waveform in MRI,

D. Shinet al., “Deep reinforcement learning-designed radiofrequency waveform in MRI,”Nat. Mach. Intell., vol. 3, no. 11, pp. 985–994, Nov. 2021, doi: 10.1038/s42256-021-00411-1

work page doi:10.1038/s42256-021-00411-1 2021
[4]

Deep reinforce- ment learning designed Shinnar-Le Roux RF pulse using root-flipping: DeepRFSLR,

D. Shin, S. Ji, D. Lee, J. Lee, S.-H. Oh, and J. Lee, “Deep reinforce- ment learning designed Shinnar-Le Roux RF pulse using root-flipping: DeepRFSLR,”IEEE Trans. Med. Imag., vol. 39, no. 12, pp. 4391–4400, Dec. 2020, doi: 10.1109/TMI.2020.3018508

work page doi:10.1109/tmi.2020.3018508 2020
[5]

DeepControl: 2DRF pulses facilitating B 1+ inhomogeneity and B 0 off-resonance compensation in vivo at 7 T,

M. S. Vinding, C. S. Aigner, S. Schmitter, and T. E. Lund, “DeepControl: 2DRF pulses facilitating B 1+ inhomogeneity and B 0 off-resonance compensation in vivo at 7 T,”Magn. Reson. Med., vol. 85, no. 6, pp. 3308–3317, Jun. 2021, doi: 10.1002/mrm.28667

work page doi:10.1002/mrm.28667 2021
[6]

Joint design of RF and gradient waveforms via auto-differentiation for 3D tailored excitation in MRI,

T. Luo, D. C. Noll, J. A. Fessler, and J.-F. Nielsen, “Joint design of RF and gradient waveforms via auto-differentiation for 3D tailored excitation in MRI,”IEEE Trans. Med. Imag., vol. 40, no. 12, pp. 3305– 3314, Dec. 2021, doi: 10.1109/TMI.2021.3083104

work page doi:10.1109/tmi.2021.3083104 2021
[7]

Automated pulse sequence generation (AUTOSEQ) using Bayesian reinforcement learning in an MRI physics simulation environment,

B. Zhu, J. Liu, N. Koonjoo, B. R. Rosen, and M. S. Rosen, “Automated pulse sequence generation (AUTOSEQ) using Bayesian reinforcement learning in an MRI physics simulation environment,” inProc. Joint Annual Meeting ISMRM-ESMRMB, 2018, p. 0438

2018
[8]

Multi-task convolutional neural network-based design of radio frequency pulse and the accompanying gradients for magnetic resonance imaging,

Y . Zhanget al., “Multi-task convolutional neural network-based design of radio frequency pulse and the accompanying gradients for magnetic resonance imaging,”NMR Biomed., vol. 34, no. 2, p. e4443, Feb. 2021, doi: 10.1002/nbm.4443

work page doi:10.1002/nbm.4443 2021
[9]

Learning optimal k-space acquisition and reconstruction using physics-informed neural networks,

W. Peng, L. Feng, G. Zhao, and F. Liu, “Learning optimal k-space acquisition and reconstruction using physics-informed neural networks,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 20794–20803

2022
[10]

B-spline parameterized joint optimization of reconstruction and k-space trajecto- ries (BJORK) for accelerated 2D MRI,

G. Wang, T. Luo, J.-F. Nielsen, D. C. Noll, and J. A. Fessler, “B-spline parameterized joint optimization of reconstruction and k-space trajecto- ries (BJORK) for accelerated 2D MRI,”IEEE Trans. Med. Imag., vol. 41, no. 9, pp. 2318–2330, Sep. 2022, doi: 10.1109/TMI.2022.3161875

work page doi:10.1109/tmi.2022.3161875 2022
[11]

Deep- learning-based optimization of the under-sampling pattern in MRI,

C. D. Bahadir, A. Q. Wang, A. V . Dalca, and M. R. Sabuncu, “Deep- learning-based optimization of the under-sampling pattern in MRI,” IEEE Trans. Comput. Imag., vol. 6, pp. 1139–1152, Jul. 2020, doi: 10.1109/TCI.2020.3006727

work page doi:10.1109/tci.2020.3006727 2020
[12]

Active MR k-space sampling with reinforcement learning,

L. Pineda, S. Basu, A. Romero, R. Calandra, and M. Drozdzal, “Active MR k-space sampling with reinforcement learning,” inProc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent. (MICCAI), 2020, pp. 23–33

2020
[13]

Stochastic optimization of three-dimensional non-cartesian sampling trajectory,

G. Wang, J.-F. Nielsen, J. A. Fessler, and D. C. Noll, “Stochastic optimization of three-dimensional non-cartesian sampling trajectory,” Magn. Reson. Med., vol. 90, no. 2, pp. 417–431, Apr. 2023, doi: 10.1002/mrm.29645

work page doi:10.1002/mrm.29645 2023
[14]

Automated design of pulse sequences for magnetic resonance fingerprinting using physics-inspired optimization,

S. P. Jordanet al., “Automated design of pulse sequences for magnetic resonance fingerprinting using physics-inspired optimization,”Proc. Natl. Acad. Sci. U.S.A., vol. 118, no. 40, p. e2020516118, Oct. 2021, doi: 10.1073/pnas.2020516118

work page doi:10.1073/pnas.2020516118 2021
[15]

MRzero – Automated discovery of MRI sequences using supervised learning,

A. Loktyushinet al., “MRzero – Automated discovery of MRI sequences using supervised learning,”Magn. Reson. Med., vol. 86, no. 2, pp. 709– 724, Jun. 2021, doi: 10.1002/mrm.28727

work page doi:10.1002/mrm.28727 2021
[16]

An end- to-end AI-based framework for automated discovery of rapid CEST/MT MRI acquisition protocols and molecular parameter quantification (Au- toCEST),

O. Perlman, B. Zhu, M. Zaiss, M. S. Rosen, and C. T. Farrar, “An end- to-end AI-based framework for automated discovery of rapid CEST/MT MRI acquisition protocols and molecular parameter quantification (Au- toCEST),”Magn. Reson. Med., vol. 87, no. 6, pp. 2792–2810, Jun. 2022, doi: 10.1002/mrm.29173

work page doi:10.1002/mrm.29173 2022
[17]

Using deep reinforcement learning to actively, adaptively and autonomously control a simulated MRI scanner,

S. Walker-Samuel, “Using deep reinforcement learning to actively, adaptively and autonomously control a simulated MRI scanner,” inProc. ISMRM 27th Annual Meeting & Exhibition, 2019, p. 0473

2019
[18]

Hochreiter, S

X. He, K. Zhao, and X. Chu, “AutoML: A survey of the state- of-the-art,”Knowl.-Based Syst., vol. 212, p. 106622, Jan. 2021, doi: 10.1016/j.knosys.2020.106622

work page doi:10.1016/j.knosys.2020.106622 2021
[19]

Neural architecture search,

T. Elsken, J. H. Metzen, and F. Hutter, “Neural architecture search,” inAutomated Machine Learning: Methods, Systems, Challenges. Cham, Switzerland: Springer, 2019, pp. 63–77, doi: 10.1007/978-3-030-05318- 5 3

work page doi:10.1007/978-3-030-05318- 2019
[20]

Neural architecture search: A survey,

T. Elsken, J. H. Metzen, and F. Hutter, “Neural architecture search: A survey,”J. Mach. Learn. Res., vol. 20, no. 1, pp. 1997–2017, 2019

1997
[21]

ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

H. Cai, L. Zhu, and S. Han, “ProxylessNAS: Direct neural architecture search on target task and hardware,” inProc. Int. Conf. Learn. Represent. (ICLR), 2019, doi: 10.48550/arXiv.1812.00332

work page Pith review doi:10.48550/arxiv.1812.00332 2019
[22]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014, doi: 10.48550/arXiv.1412.6980

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2014
[23]

Theory of NMR signal behavior in magnetically inhomogeneous tissues: The static dephasing regime,

D. A. Yablonskiy and E. M. Haacke, “Theory of NMR signal behavior in magnetically inhomogeneous tissues: The static dephasing regime,” Magn. Reson. Med., vol. 32, no. 6, pp. 749–763, Dec. 1994, doi: 10.1002/mrm.1910320610

work page doi:10.1002/mrm.1910320610 1994
[24]

Postprocessing technique to correct for background gradients in image-based R 2* measurements,

M. A. Fernandez-Seara and F. W. Wehrli, “Postprocessing technique to correct for background gradients in image-based R 2* measurements,” Magn. Reson. Med., vol. 44, no. 3, pp. 358–366, Sep. 2000, doi: 10.1002/1522-2594(200009)44:3<358::aid-mrm3>3.0.co;2-i

work page doi:10.1002/1522-2594(200009)44:3 2000
[25]

NMR relaxation times in the human brain at 3.0 tesla,

J. P. Wansapura, S. K. Holland, R. S. Dunn, and W. S. Ball, Jr., “NMR relaxation times in the human brain at 3.0 tesla,”J. Magn. Reson. Imag., vol. 9, no. 4, pp. 531–538, Apr. 1999, doi: 10.1002/(sici)1522- 2586(199904)9:4<531::aid-jmri4>3.0.co;2-l

work page doi:10.1002/(sici)1522- 1999
[26]

Fast and high-resolution quantitative mapping of tissue water content with full brain coverage for clinically- driven studies,

M. Sabati and A. A. Maudsley, “Fast and high-resolution quantitative mapping of tissue water content with full brain coverage for clinically- driven studies,”Magn. Reson. Imag., vol. 31, no. 10, pp. 1752–1759, Dec. 2013, doi: 10.1016/j.mri.2013.08.001

work page doi:10.1016/j.mri.2013.08.001 2013
[27]

T 2* measurements in human brain at 1.5, 3 and 7 T,

A. M. Peterset al., “T 2* measurements in human brain at 1.5, 3 and 7 T,”Magn. Reson. Imag., vol. 25, no. 6, pp. 748–753, Jul. 2007, doi: 10.1016/j.mri.2007.02.014

work page doi:10.1016/j.mri.2007.02.014 2007
[28]

Chapter 8 – Nuclear magnetism of tissue,

F. A. Duck, “Chapter 8 – Nuclear magnetism of tissue,” inPhysical Properties of Tissue. London, U.K.: Academic Press, 1990, pp. 279– 317

1990
[29]

PyTorch: An imperative style, high-performance deep learning library,

A. Paszkeet al., “PyTorch: An imperative style, high-performance deep learning library,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 32, 2019, pp. 8024–8035

2019
[30]

MR-double-zero – Proof-of-concept for a framework to autonomously discover MRI contrasts,

F. Glang, S. Mueller, K. Herz, A. Loktyushin, K. Scheffler, and M. Zaiss, “MR-double-zero – Proof-of-concept for a framework to autonomously discover MRI contrasts,”J. Magn. Reson., vol. 341, p. 107237, Aug. 2022, doi: 10.1016/j.jmr.2022.107237

work page doi:10.1016/j.jmr.2022.107237 2022