Pith. sign in

REVIEW 4 major objections 5 minor 29 references

A few-shot deep-learning model predicts per-element phase and amplitude corrections for a 96-element transcranial focused ultrasound array from CT images, matching simulation-based focusing after ten-point fine-tuning at a 2,535× speedup.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 12:08 UTC pith:Q36LJWNU

load-bearing objection Solid few-shot surrogate for tFUS phase/amplitude correction, but all performance numbers are judged against the same simulator that generated the labels; the point-monopole element model is the main external-validity risk. the 4 major comments →

arxiv 2607.29182 v1 pith:Q36LJWNU submitted 2026-07-31 eess.IV cs.LG

Few-shot Deep Learning for Phase-Amplitude Aberration Correction in Transcranial Focused Ultrasound

Minju Seol , Minjee Seo , Seonaeng Cho , Kyungho Yoon This is my paper
classification eess.IV cs.LG
keywords transcranial focused ultrasoundaberration correctionfew-shot learningdeep learningphase predictionamplitude predictionCT imagingphased array transducer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper aims to show that a deep-learning surrogate can replace expensive time-reversal simulations for correcting skull-induced aberrations in transcranial focused ultrasound. The model takes CT-derived skull geometry along with transducer and target positions, and predicts the phase and amplitude each of the 96 elements should emit. After fine-tuning on only ten target points for a new skull, it reproduces focal fields that nearly match full simulation, with focal centroid error under 0.5 mm and dice overlap above 94%. If this holds, real-time patient-specific steering becomes practical, enabling iterative treatment planning and closed-loop therapy.

Core claim

The central claim is that a geometry-aware encoder—fed an ROI patch from CT, Fourier-embedded element and target coordinates, skull-layer distances, and time-of-flight—produces a shared feature that supports both a phase-classification head and an amplitude-regression head. Phase is handled as a 32-bin circular classification with circular-expectation decoding, avoiding wrap-around errors. Joint training on simulated time-reversal labels yields per-element corrections that, when applied, reconstruct focal fields with 94.4% Dice, 92.3% peak pressure ratio, and 0.467 mm centroid error relative to the simulation ground truth.

What carries the argument

The central mechanism is the geometry feature encoder that fuses a cropped 3D skull patch with Fourier-embedded positions and ray-derived segment distances and time-of-flight. This shared feature feeds two decoders: an amplitude regression head and a phase classification head over 32 circular bins, decoded by circular expectation. The circular soft-label and cosine loss directly penalize angular error, making phase periodicity a built-in inductive bias rather than an afterthought.

Load-bearing premise

All ground-truth phases, amplitudes, and evaluation fields come from time-reversal simulations that assign bone acoustic properties purely by CT intensity thresholds; if that simulation model is not faithful to real skull transmission, the learned corrections correct simulation artifacts rather than actual aberrations.

What would settle it

Measure the true per-element phase and amplitude with a hydrophone for a focused ultrasound beam transmitted through an ex vivo skull, compare the model's predicted corrections to those measurements for the same target, and reconstruct the resulting focus; a large mismatch or a degraded focal peak would show the simulation-based labels do not transfer to reality.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A full 96-element steering profile is produced in 0.029 seconds, compared with 73.5 seconds for simulation, so per-patient correction becomes a real-time operation.
  • Only ten target points are needed to adapt to an unseen skull, removing the need for a full patient-specific simulation before treatment.
  • Simultaneous phase and amplitude correction addresses both focal steering and intensity preservation, potentially improving safety and efficacy.
  • The speedup makes iterative treatment planning—adjusting targets or re-planning during a session—practicable.

Where Pith is reading between the lines

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

  • Because the model learns a mapping from skull geometry to correction parameters, the same architecture could, in principle, be retrained for other transducer layouts, frequencies, or multi-focus patterns, if simulation labels were available.
  • The stated few-shot behavior is essentially a domain-adaptation result; it suggests the learned skull-geometry features are transferable across subjects, which might be exploited in a meta-learning framework.
  • The real test is experimental: the model's accuracy is measured against the same simulation that generated its labels, so a phantom or ex vivo validation against hydrophone measurements would determine whether the learned corrections transfer to tissue.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a few-shot deep learning surrogate for transcranial focused ultrasound (tFUS) aberration correction. For each element of a 96-element 3D phased array, the model predicts the phase and amplitude corrections from CT-derived skull geometry features, after pretraining on 11 skulls and fine-tuning on 10 target points of the held-out skull. The training and ground-truth labels come from k-Wave PSTD time-reversal simulations with HU-threshold acoustic properties. Leave-one-out evaluation across 12 skulls reports phase CMAE 0.155 rad, amplitude rMAE 9.089%, focal centroid error 0.467 mm, Dice 94.422%, and peak pressure ratio 92.332%, with a claimed ~2,535x speedup over TR simulation.

Significance. If the results transfer beyond simulation, the framework would be a practically useful fast surrogate for patient-specific TR-based aberration correction. The strengths are the leave-one-out protocol across 12 skulls, the component ablations (Table 3), and the public code release. However, the entire evaluation is in silico: labels, predictions, and focal metrics are all generated by the same PSTD model with a point-monopole element representation. The reported quantitative values are therefore properties of that simulator, not of physical tFUS through real skulls. The fine-tuning step also implicitly requires 10 full TR simulations per new skull, which materially affects the speedup claim. These caveats are not limitations that invalidate the internal logic of the surrogate, but they must be stated and the claims adjusted accordingly.

major comments (4)
  1. [§4.1, Real-time Applicability] The ~2,535× speedup is misleading because it compares only per-target inference time (0.029 s) against one TR simulation (73.52 s) and ignores the 10 TR simulations needed to generate fine-tuning labels for a new skull. Those 10 simulations take roughly 735 s, so the total one-time adaptation cost is ~752 s. For a single target, the proposed pipeline is then slower than running one TR simulation; even for 100 targets, the total time (~752 s + 2.9 s) is only ~9.7× faster than 100 TR simulations. The abstract and conclusion also quote the speedup without this caveat. The authors should report end-to-end timings including label generation, or explicitly define the speedup as per-target inference after a one-time adaptation cost, and state that amortization requires many targets.
  2. [§2.1, Phased Array Transducer Configuration] Each transducer element is modeled as a point monopole. At 250 kHz (wavelength ~6 mm in water), a real phased-array element has finite aperture, frequency-dependent directivity, and mutual coupling. The point-source idealization discards these effects, and the optimal phase/amplitude corrections could differ meaningfully for a physical 96-element array. The cited ex vivo validation [7,18] used a single-element transducer and a different configuration, so it does not establish the validity of the point-source model for this array. The authors should either (a) provide evidence that point-source corrections transfer to finite-aperture elements, e.g., a simulation comparison with finite-aperture source terms, or (b) state clearly that the reported focal metrics are for the point-source simulation model and are not yet shown to apply to a real device.
  3. [§3.4, Model Implementation] The fine-tuning protocol is underspecified. It is not stated how the 10 fine-tuning target points are selected for each held-out skull, and no variance is reported over different selections. Because only 10 labels are used, the choice could materially affect the results. The paper should describe the selection rule (e.g., random, grid, farthest-point) and, ideally, report results averaged over multiple random selections. In addition, the text should be explicit that fine-tuning requires 10 full TR simulations per new skull, which means the method does not avoid patient-specific simulation; rather, it reduces the number of simulated target points from 100 to 10. This clarification is needed in Section 3.4 and in the abstract/conclusion.
  4. [§1 and §4, Comparison to Prior Work] The introduction states that Zhang et al. [29] and Naftchi-Ardebili et al. [12] have limitations, and the paper claims the first joint phase-and-amplitude prediction for a 3D array. However, no quantitative comparison to these methods is provided. Without a table or clearly justified qualitative comparison on the same data, the claimed advantage over prior DL-based aberration correction is not supported. The authors should add a numerical comparison (or explain why a direct comparison is impossible, e.g., different transducer configurations) and also compare against a simple no-correction or single-frequency phase-correction baseline to contextualize the reported focal metrics.
minor comments (5)
  1. [Abstract and Conclusion] The phrase '~2,535× speedup' should be qualified as 'per-target inference speedup after a one-time adaptation cost' to avoid overstatement.
  2. [§3.3] The Gaussian kernel used for circular soft labels is not fully specified; the standard deviation of the Gaussian should be reported. This is a hyperparameter that affects the phase classification loss.
  3. [§3.5] The definition of PPR ('peak pressure ratio') is given only as 'peak-pressure preservation.' Please specify the ratio (predicted peak / TR peak) and whether it is expressed as a percentage, as in Table 2.
  4. [§4.1] The text mentions '12×90 test cases' after saying 100 target points are sampled. It should be stated explicitly that 10 points per skull are reserved for fine-tuning and the remaining 90 are used for evaluation.
  5. [Overall] The paper would benefit from a zero-shot (no fine-tuning) performance row to quantify the contribution of the few-shot adaptation step. This is not essential but would strengthen the 'few-shot' narrative.

Circularity Check

0 steps flagged

No significant circularity; core training/evaluation loop is a standard held-out surrogate benchmark.

full rationale

The paper's derivation chain is a supervised learning loop: inputs are CT-derived RoI patches, element/target coordinates, intersection points, and ToF features (Sec. 3.1); labels are phase/amplitude pairs extracted from PSTD time-reversal simulations (Sec. 2.2); and the reported metrics are computed on held-out skulls after 10-point fine-tuning with the remaining 90 targets per skull evaluated. No equation defines a predicted phase/amplitude in terms of the label itself, and no fitted parameter is renamed as a prediction; the amplitude and phase heads output scalars/distributions from geometry features, with evaluation against TR ground truth consuming those outputs. The closest candidate for circularity is the Sec. 2.2 reliance on the authors' prior work [7,18] to assert that the HU-threshold acoustic property table matches ex vivo measurements. That is a self-citation, but it points to externally obtained experimental validation rather than to the present paper's own fitted values, so under the review rules it is independent support and does not make the core surrogate derivation circular. The same-simulation evaluation is a benchmark-consistency property (the model learns to reproduce the simulator), which limits external validity but is not a circular reduction; the paper itself defers phantom/clinical validation to future work in Sec. 5. No load-bearing step reduces Eq. X to Eq. Y by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The only invented objects are neural network components (geometry encoder, phase/amplitude heads), which are not claimed to be physical. The central claim rests on the fidelity of the HU-based simulation model, on the ray-tracing feature extraction, and on the representativeness of the target grid; all are domain assumptions carried from prior work.

free parameters (5)
  • phase_bin_count = 32
    Discretization of [-pi, pi) into 32 uniform bins for phase classification; sets a hard floor on achievable phase error (about 0.196 rad) and is not ablated.
  • loss_weight_ratio = 0.3 KL / 0.7 circular cosine
    Hand-set weighting of the two terms in the phase loss; no sensitivity analysis is reported.
  • fine_tune_target_count = 10
    Number of TR simulations used per held-out skull for few-shot adaptation; chosen as 'few-shot' but the cost and sensitivity are not analyzed, and the cost is excluded from the speedup claim.
  • fourier_feature_dimension = 16
    Frequency count in the Fourier feature embedding for coordinates; the ablation only tests presence versus absence, not the scale of the embedding.
  • roi_patch_size = 40x40x80
    Hand-fixed crop between skull boundaries; determines the skull geometry seen by the encoder. No ablation.
axioms (4)
  • domain assumption The HU-threshold acoustic property model (Table 1) faithfully represents skull-induced phase and amplitude aberrations
    Invoked in Sec. 2.2 for all ground-truth generation; values are cited from prior studies and the group's ex vivo validation [7,18], but not independently verified here. If wrong, the surrogate learns to reproduce simulation errors.
  • domain assumption PSTD time-reversal simulation with point-monopole elements gives the correct correction profile
    Section 2.2 uses k-Wave PSTD with a 96-element point-monopole source model; no experimental validation of this array model is provided in this paper.
  • domain assumption Ray-casting with four HU-threshold intersections captures the acoustic path through the skull
    Section 2.1 defines the RoI patch and intersections using 'adaptive and fixed HU thresholds'; the thresholds are not specified, and the four-intersection model assumes a single ray path, omitting multipath and skull heterogeneities.
  • domain assumption The 100 target points and the 10-point fine-tune set are representative of clinical targeting
    Target points lie in a 10x10x10 mm cube near the array center; the fine-tune point selection procedure is not described, so generalization to other brain targets is assumed.

pith-pipeline@v1.3.0-daily-deepseek · 8187 in / 14432 out tokens · 133130 ms · 2026-08-03T12:08:33.517822+00:00 · methodology

0 comments
read the original abstract

Transcranial focused ultrasound (tFUS) is a non-invasive technique that delivers focused acoustic energy through the skull for neuromodulation and therapeutic applications. However, the heterogeneous structure of the skull induces complex, patient-specific phase and amplitude aberrations that distort the acoustic focus and deviate it from the intended target, compromising therapeutic efficacy and safety. Conventional time-reversal (TR) simulations can correct these aberrations but rely on computationally expensive full-wave solvers, making them impractical for real-time use and iterative treatment planning. We propose a few-shot deep surrogate framework that predicts per-element phase and amplitude corrections for a 96-element 3D phased-array transducer from patient CT images. A geometry-aware encoder extracts skull-path features shared across dedicated phase classification and amplitude regression branches, where phase periodicity is handled via circular expectation decoding. The framework is pretrained on diverse skull geometries and fine-tuned with only ten target points, enabling rapid adaptation to unseen patients without full patient-specific simulation. Evaluated via leave-one-out cross-validation across 12 skulls, it achieves a mean phase CMAE of 0.155 rad and amplitude rMAE of 9.089%, a focal centroid error of 0.467 mm, Dice score of 94.422%, and peak pressure ratio of 92.332%, with an approximately 2,535 times speedup over TR simulation. The code is available at https://github.com/Minju-Seol/fewshot-tfus-correction.

Figures

Figures reproduced from arXiv: 2607.29182 by Kyungho Yoon, Minjee Seo, Minju Seol, Seonaeng Cho.

Figure 1
Figure 1. Figure 1: The illustration of Intersection Extraction. The ray from a transducer element to the target defines four skull-interface intersection points that delimit the Region of Interest (RoI) 3D patch. Acoustic Path Feature Extraction For each transducer element, we cast a ray toward the target in CT voxel space and identify four skull-interface inter- [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed network. From the input geometry (RoI patch, target and transducer positions, and skull-path intersections), the geometry fea￾ture encoder produces a geometry feature that is fed to the amplitude and phase blocks to predict the per-element amplitude and phase. 3 Model Architecture The proposed framework predicts the phase and amplitude that each transducer element should emit to ac… view at source ↗
Figure 3
Figure 3. Figure 3: Pressure fields for a representative case (Skull 9, Target Point 66) in the zx- and yx-planes. GT (top) and Pred (bottom) with −6 dB FWHM contour (white) and target position (red cross). Focusing Performance The proposed framework achieved a mean FCE of 0.467 mm, Dice score of 94.422%, and PPR of 92.332%, as summarized in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages

  1. [1]

    The Journal of the Acoustical Society of America152(2), 1003–1019 (2022)

    Aubry, J.F., Bates, O., Boehm, C., Butts Pauly, K., Christensen, D., Cueto, C., Gélat, P., Guasch, L., Jaros, J., Jing, Y., et al.: Benchmark problems for transcra- nial ultrasound simulation: Intercomparison of compressional wave models. The Journal of the Acoustical Society of America152(2), 1003–1019 (2022)

    2022

  2. [2]

    Courant, R., Friedrichs, K., Lewy, H.: On the partial difference equations of math- ematical physics. IBM J. Res. Dev.11(2), 215–234 (1967).https://doi.org/10 .1147/rd.112.0215

    1967

  3. [3]

    Di Biase, L., Falato, E., Di Lazzaro, V.: Transcranial focused ultrasound (tfus) and transcranial unfocused ultrasound (tus) neuromodulation: from theoretical principles to stimulation practices. Front. Neurol.10, 549 (2019).https://doi. org/10.3389/fneur.2019.00549

    arXiv 2019

  4. [4]

    Elias, W.J., Huss, D., Voss, T., Loomba, J., Khaled, M., Zadicario, E., Frysinger, R.C., Sperling, S.A., Wylie, S., Monteith, S.J., Druzgal, J., Shah, B.B., Harrison, M., Wintermark, M.: A pilot study of focused ultrasound thalamotomy for essential tremor. N. Engl. J. Med.369(7), 640–648 (2013).https://doi.org/10.1056/NE JMoa1300962

    work page doi:10.1056/ne 2013

  5. [5]

    Fink, M.: Time reversal of ultrasonic fields. I. Basic principles. IEEE Trans. Ultra- son. Ferroelectr. Freq. Control39(5), 555–566 (2002).https://doi.org/10.110 9/58.156174 10 M. Seol et al

    2002

  6. [6]

    IEEE Trans

    Gâteau, J., Marsac, L., Pernot, M., Aubry, J.F., Tanter, M., Fink, M.: Transcranial ultrasonic therapy based on time reversal of acoustically induced cavitation bubble signature. IEEE Trans. Biomed. Eng.57(1), 134–144 (2009).https://doi.org/ 10.1109/TBME.2009.2031816

    arXiv 2009

  7. [7]

    Jang, M., Choi, M., Jeong, I., Yoo, S.S., Yoon, K., Noh, G.: Deep learning-based real-time estimation of transcranial focused ultrasound acoustic field. Eng. Appl. Artif. Intell.156, 111157 (2025).https://doi.org/10.1016/j.engappai.2025. 111157

    doi:10.1016/j.engappai.2025 2025

  8. [8]

    BMC Biomed

    Jin, C., Moore, D., Snell, J., Paeng, D.G.: An open-source phase correction toolkit for transcranial focused ultrasound. BMC Biomed. Eng.2(1), 9 (2020).https: //doi.org/10.1186/s42490-020-00043-3

    work page doi:10.1186/s42490-020-00043-3 2020

  9. [9]

    Jing, Y., Meral, F.C., Clement, G.T.: Time-reversal transcranial ultrasound beam focusing using a k-space method. Phys. Med. Biol.57(4), 901 (2012).https: //doi.org/10.1088/0031-9155/57/4/901

    work page doi:10.1088/0031-9155/57/4/901 2012

  10. [10]

    Legon, W., Sato, T.F., Opitz, A., Mueller, J., Barbour, A., Williams, A., Tyler, W.J.: Transcranial focused ultrasound modulates the activity of primary so- matosensory cortex in humans. Nat. Neurosci.17(2), 322–329 (2014).https: //doi.org/10.1038/nn.3620

    work page doi:10.1038/nn.3620 2014

  11. [11]

    Magara, A., Bühler, R., Moser, D., Kowalski, M., Pourtehrani, P., Jeanmonod, D.: First experience with mr-guided focused ultrasound in the treatment of parkinson’s disease. J. Ther. Ultrasound2(1), 11 (5 2014).https://doi.org/10.1186/2050 -5736-2-11

    work page doi:10.1186/2050 2014

  12. [12]

    Naftchi-Ardebili, K., Singh, K., Popelka, G.R., Pauly, K.B.: A deep-learning model for one-shot transcranial ultrasound simulation and phase aberration correction. Med. Phys.53(1), e70259 (2026).https://doi.org/10.1002/mp.70259

    work page doi:10.1002/mp.70259 2026

  13. [13]

    Advances in neural information processing sys- tems32(2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems32(2019)

    2019

  14. [14]

    Pernot, M., Aubry, J.F., Tanter, M., Boch, A.L., Marquet, F., Kujas, M., Seilhean, D., Fink, M.: In vivo transcranial brain surgery with an ultrasonic time reversal mirror. J. Neurosurg.106(6), 1061–1066 (2007).https://doi.org/10.3171/jns. 2007.106.6.1061

    work page doi:10.3171/jns 2007

  15. [15]

    In: Inter- national Workshop on Digital Twin for Healthcare

    Seo, M., Shin, M., Noh, G., Yoo, S.S., Yoon, K.: Acoustic simulation with deep learning for low-intensity transcranial focused ultrasound digital twins. In: Inter- national Workshop on Digital Twin for Healthcare. pp. 58–68. Springer (2025). https://doi.org/10.1007/978-3-032-07694-6_6

    work page doi:10.1007/978-3-032-07694-6_6 2025

  16. [16]

    Shin, M., Peng, Z., Kim, H.J., Yoo, S.S., Yoon, K.: Multivariable-incorporating super-resolution residual network for transcranial focused ultrasound simulation. Comput. Meth. Programs Biomed.237, 107591 (2023).https://doi.org/10.101 6/j.cmpb.2023.107591

    arXiv 2023

  17. [17]

    Engineering Applications of Artificial Intelligence 138, 109349 (2024).https://doi.org/10.1016/j.engappai.2024.109349

    Shin, M., Seo, M., Cho, S., Park, J., Kwon, J.H., Lee, D., Yoon, K.: Physrfanet: Physics-guided neural network for real-time prediction of thermal effect during ra- diofrequency ablation treatment. Engineering Applications of Artificial Intelligence 138, 109349 (2024).https://doi.org/10.1016/j.engappai.2024.109349

    arXiv 2024

  18. [18]

    Shin, M., Seo, M., Yoo, S.S., Yoon, K.: tfusformer: Physics-guided super-resolution transformer for simulation of transcranial focused ultrasound propagation in brain stimulation. IEEE J. Biomed. Health Inform.28(7), 4024–4035 (2024).https: //doi.org/10.1109/JBHI.2024.3389708 Few-shot Deep Learning for Phase-Amplitude Aberration Correction 11

    arXiv 2024

  19. [19]

    In: NeurIPS

    Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Sing- hal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. vol. 33, pp. 7537–7547 (2020).https://doi.org/10.48550/arXiv.2006.10739

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.10739 2020

  20. [20]

    Treeby, B.E., Cox, B.T.: k-wave: Matlab toolbox for the simulation and reconstruc- tion of photoacoustic wave fields. J. Biomed. Opt.15(2), 021314–021314 (2010). https://doi.org/10.1117/1.3360308

    doi:10.1117/1.3360308 2010

  21. [21]

    Tufail, Y., Yoshihiro, A., Pati, S., Li, M.M., Tyler, W.J.: Ultrasonic neuromodu- lation by brain stimulation with transcranial ultrasound. Nat. Protoc.6(9), 1453– 1470 (2011).https://doi.org/10.1038/nprot.2011.371

    work page doi:10.1038/nprot.2011.371 2011

  22. [22]

    Ultrasonics132, 107026 (2023).https://doi.org/https://doi.org/10.1016/j.ultras.2023.107026

    Wang, L., Wang, H., Liang, L., Li, J., Zeng, Z., Liu, Y.: Physics-informed neural networks for transcranial ultrasound wave propagation. Ultrasonics132, 107026 (2023).https://doi.org/https://doi.org/10.1016/j.ultras.2023.107026

    arXiv 2023

  23. [23]

    Iradiology3(1), 26–46 (2025).ht tps://doi.org/10.1002/ird3.112

    Wang, M., Xu, Z., Cheng, B.: Systematic review of phase aberration correction algorithms for transcranial focused ultrasound. Iradiology3(1), 26–46 (2025).ht tps://doi.org/10.1002/ird3.112

    work page doi:10.1002/ird3.112 2025

  24. [24]

    IEEE Trans

    White, J., Clement, G.T., Hynynen, K.: Transcranial ultrasound focus reconstruc- tion with phase and amplitude correction. IEEE Trans. Ultrason. Ferroelectr. Freq. Control52(9), 1518–1522 (2005).https://doi.org/10.1109/TUFFC.2005.15160 24

    arXiv 2005

  25. [25]

    Wu, F., Thomas, J.L., Fink, M.: Time reversal of ultrasonic fields. il. experimen- tal results. IEEE transactions on ultrasonics, ferroelectrics, and frequency control 39(5), 567–578 (1992).https://doi.org/10.1109/58.156175

    work page doi:10.1109/58.156175 1992

  26. [26]

    Ultrasound Med

    Xu, L., Lee, W., Rotenberg, A., Böhlke, M., Yoon, K., Yoo, S.S.: Localized disrup- tion of blood albumin–phenytoin binding using transcranial focused ultrasound. Ultrasound Med. Biol.46(8), 1986–1997 (2020).https://doi.org/10.1016/j.ul trasmedbio.2020.04.011

    work page doi:10.1016/j.ul 1986

  27. [27]

    Yang, D., Fu, S., Zhao, M., Shi, Y.: The promise of transcranial focused ultrasound in disorders of consciousness: a narrative review. Crit. Care29(1), 1–6 (2025). https://doi.org/10.1186/s13054-025-05338-2

    work page doi:10.1186/s13054-025-05338-2 2025

  28. [28]

    Sensors21(17), 5962 (2021).https://doi.org/10.3390/s21175962

    Zhang, H., Zhang, Y., Xu, M., Song, X., Chen, S., Jian, X., Ming, D.: The effects of the structural and acoustic parameters of the skull model on transcranial focused ultrasound. Sensors21(17), 5962 (2021).https://doi.org/10.3390/s21175962

    work page doi:10.3390/s21175962 2021

  29. [29]

    Ul- trasonics152, 107641 (2025).https://doi.org/10.1016/j.ultras.2025.107641

    Zhang, Q., Sun, W., Deng, J., Qi, T., Wan, M., Lu, M.: Transcranial adaptive aberration correction using deep learning for phased-array ultrasound therapy. Ul- trasonics152, 107641 (2025).https://doi.org/10.1016/j.ultras.2025.107641

    arXiv 2025