Pith. sign in

REVIEW 2 major objections 5 minor 57 references

MoNetV2: Enhanced Motion Network for Freehand 3D Ultrasound Reconstruction

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MoNetV2 fuses IMU motion data with ultrasound images to cut freehand 3D reconstruction drift.

desk verdict Solid incremental extension of MoNet with strong experiments, but the IMU zero-mean preprocessing is an unvalidated assumption that the online losses cannot fix. read the letter →

arxiv 2506.15835 v1 pith:F2GN5SIS submitted 2025-06-16 eess.IV cs.AIcs.CV

Mingyuan Luo , Xin Yang , Zhongnuo Yan , Yan Cao , Yuanji Zhang , Xindi Hu , Jin Wang , Haoxuan Ding
show 3 more authors
Wei Han Litao Sun Dong Ni
This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords freehand3Dultrasoundreconstructioninertialmeasurementunitonlineself-supervisedlearningmulti-levelconsistencymulti-modalfusioncumulativedriftscanningvelocityandtacticsdeep
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Freehand 3D ultrasound reconstruction aims to turn a sweep of 2D ultrasound frames into a 3D volume without an external tracking arm. The paper's claim is that this becomes markedly more accurate and more transferable to new scanning styles when the network is given inertial motion data and is adapted online to each scan, rather than reconstructing from image content alone. The proposed MoNetV2 fuses image features with a velocity feature built from IMU acceleration and Euler angles, then uses three online consistency constraints and one multi-modal self-supervised loss to keep cumulative drift low even when the sonographer changes speed or scanning tactic. On arm, carotid, and thyroid datasets covering linear, curved, sector, and loop scans, the paper reports consistently lower drift and angle error against existing reconstruction methods, as well as better cross-dataset generalization. If true, the practical consequence is cheaper, tracker-free 3D ultrasound that stays reliable under realistic hand motion and across different organs.

What carries the argument

The load-bearing mechanism is the temporal and multi-branch structure (TMS) together with four online losses. The TMS constructs a velocity feature $f^v_i$ by adding the IMU acceleration feature to the preceding image feature, $f^v_i = f_{i-1} + f^a_i$, so that acceleration acts as increments of velocity between frames; a ConvLSTM then smooths this feature over the sequence, and IMU Euler angles are concatenated before the final transformation regression. The online losses are: scan-level velocity consistency (SVC), which samples subsequences at intervals $k$ and requires $\hat{\theta}^k$ to equal the composition of smaller-step estimates; path-level appearance consistency (PAC), which reorders subsequences of a loop tactic and requires the same appearance to produce the same positions; patch-level motion consistency (PMC), which uses the FILM frame-interpolation model to build an approximate geodesic between frames and matches normalized per-patch content differences to normalized per-patch 3D motion; and the multi-modal self-supervised strategy (MSS), which uses IMU acceleration and Euler angle as pseudo-labels. Each loss supervises the same inter-frame transformation estimator, so the network is continuously adapted to the statistics of the scan being reconstructed.

What would settle it

Take a scan in which the electromagnetic ground truth shows the probe moving at the first or last frame, feed that scan to MoNetV2, and compare its FDR, ADR, and MD with the current test-set results; if the rest assumption matters, drift should worsen noticeably on these non-rest scans. A simpler check: compute the mean of $(A_i - g_i)$ for such scans and show that it is not zero, which would directly invalidate the debiasing step in Eq. (2).

Watch

Extended reading notes

Core claim

MoNetV2 claims that the velocity of the probe is the right bridge between ultrasound images and inertial measurements, and that exploiting this bridge both in the network and in online losses removes most of the accumulated drift that plagues image-only reconstruction. The temporal and multi-branch structure computes a velocity feature by adding IMU-acceleration features to the previous image feature, feeds it through a ConvLSTM for temporal smoothing, and concatenates IMU Euler angles at the output; the image branch remains a ResNet18 backbone. During testing, three consistency losses supervise the estimated inter-frame transformations using only the scan itself: scan-level velocity consistency enforces that transformations estimated at different sampling intervals compose correctly; path-level appearance consistency reorders subsequences of a loop scan and demands that revisited anatomy yield the same positions; patch-level motion consistency matches per-patch image-content differences, measured along a frame-interpolated geodesic, to per-patch 3D motion distances. A multi-modal self-supervised loss then aligns the estimated acceleration and Euler angles with the IMU readings. The paper reports that the full system outperforms five prior methods on all six metrics on three datasets, and retains that advantage in cross-dataset tests.

Load-bearing premise

The preprocessing of the IMU acceleration assumes the probe is at rest at the first and last frames of every scan, so the mean of the gravity-corrected acceleration is zero and can be subtracted; if the probe is already moving at either end of the scan, the acceleration signal is biased and the online self-supervised loss can steer the estimated trajectory away from the true path.

Editorial extensions

If this is right

  • If MoNetV2's claims hold, freehand 3D ultrasound can achieve tracker-level accuracy with a cheap inertial sensor attached to the probe, without an external positioning system.
  • Scans with uneven velocity should no longer require uniform frame spacing or careful speed control, because scan-level velocity consistency explicitly trains the estimator to compose across intervals.
  • Loop scans and other maneuvers that revisit anatomy should accumulate less drift, because path-level appearance consistency anchors revisited regions.
  • Reconstruction should transfer to new organs and scanning tactics without retraining, as the cross-dataset arm/carotid/thyroid results indicate.
  • Reconstructed volumes become clinically usable for downstream measurements: the paper reports strong correlation with ground-truth vessel volume size and length in carotid scans.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A test the paper does not report: remove the image branch and feed only IMU data through the TMS. This would separate how much of the gain comes from the sensor fusion versus the consistency losses, and would expose how quickly performance degrades when the acceleration debiasing assumption is violated.
  • The scan-level velocity consistency loss is a generic compositionality constraint that does not require an IMU. Applied to any frame-to-frame pose estimator, it could provide a similar online adaptation signal, though the paper only evaluates it inside MoNetV2.
  • The equal weighting of the four online losses and the fixed 60-iteration schedule are tuned to these datasets; the paper's own thyroid experiments use shorter scans and tighter sampling intervals, so in very long or very short clinical sweeps the loss balance and iteration count would likely need adjusting.
  • The method assumes rigid probe motion, and the paper notes that tissue deforms under probe pressure; the online pseudo-labels would then pull the trajectory toward the sensor path rather than the deforming anatomy, so extending this to non-rigid tissue would require separating rigid trajectory from elastic deformation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes MoNetV2, a deep network for freehand 3D ultrasound reconstruction that fuses ultrasound images with IMU acceleration and orientation through a temporal multi-branch structure (TMS). At test time, it applies three online consistency losses (scan-level velocity consistency, path-level appearance consistency, patch-level motion consistency) and a multi-modal self-supervised loss (MSS) that correlates the estimated acceleration with IMU acceleration and the estimated Euler angles with IMU angles. The method is trained on EM-tracked ground truth and adapted per test scan without ground truth. Experiments on three datasets (arm, carotid, thyroid) with six metrics, ablations of inputs and components, cross-dataset generalization, and a downstream vessel segmentation evaluation claim consistent improvements over CNN-OF, ResNet18, DC2-Net, RecON, MoNet, and OSCNet.

Significance. If the claims hold, MoNetV2 is a meaningful advance for freehand 3D ultrasound: it demonstrates that a lightweight IMU can be combined with online consistency constraints to improve both accuracy and cross-dataset generalizability. The empirical study is unusually thorough, including three datasets, six metrics, component/input ablations, a four-way generalizability study, an online-loss-weight sensitivity analysis, and a vessel-segmentation downstream validation. The test-time losses are genuinely self-supervised in that they do not use ground truth. The main reservations are the unverified zero-mean acceleration assumption in Section III-A and the terse statistical reporting; these affect the strength of the central claim but are addressable within the manuscript's scope.

major comments (2)
  1. [Section III-A, Eq. (2)] The zero-mean correction of IMU acceleration is justified by the statement 'Assuming that V1 and VN are both equal to zero,' but the physical condition for the mean of (A_i - g_i) to vanish is V1 = VN, not necessarily zero. The data-collection protocol does not state that the probe is at rest at the first and last frames of every scan, and no endpoint-velocity analysis is reported. If V1 differs from VN, Eq. (2) removes the net acceleration and introduces a systematic bias into A, which is both an input to the TMS and a pseudo-label in the MSS loss. Because L_MSS (Eq. 10) uses Pearson correlation (invariant to additive offsets and scale), and because L_SVC (Eq. 5), L_PAC (Eq. 6), and L_PMC (Eq. 8, with z-score normalization) are also invariant to global rescaling of the estimated translations, the online losses cannot detect or correct a global trajectory bias induced by this preprocessing. The authors should quantify V1 and VN from the EM tracker for the collected scans or modify the preprocessing to avoid relying on this assumption.
  2. [Sections V-A and V-B, Tables I and II] The paper states 'MoNetV2 significantly outperformed all other methods across all datasets and metrics (p < 0.05, t-test)' and a similar claim in Section V-B, but no details of the test are given: it is not stated whether the t-test is paired or unpaired, one- or two-sided, whether the unit of analysis is scan or volunteer, or how multiple comparisons across six metrics and several methods are handled. Given that the headline claim is a universal superiority claim, the statistical reporting needs to be specific and, if applicable, include a multiple-comparison correction.
minor comments (5)
  1. [Section III-A] There is a typo: 'the IN’s velocity VN' should read 'the IMU’s velocity VN.'
  2. [Section V-A] The sentence reporting reductions of '32.14%/26.02%/19.05%' for FDR and '30.71%/32.96%/24.02%' for MEA does not state the baseline method to which these reductions are relative; it should be explicit.
  3. [Section III-C3 / Table VI] The choice of 63 interpolated images in PMC is justified only by monotonic improvement in Table VI; a brief sentence on the computational cost and on whether FILM, trained on natural images, is reliable for ultrasound speckle patterns would improve the presentation.
  4. [Section V-D] The vessel segmentation experiment uses semi-automatic labeling with the Pair software, but no information is given about inter- or intra-operator variability; a short statement would clarify the reliability of the reported vessel metrics.
  5. [Appendix G, Table IX] The Discussion says the average inference time is 'approximately 1.5 minutes,' while Table IX reports 87.83 seconds; these are consistent, but the Discussion should cite the table for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MoNetV2's supervised training uses independent EM ground truth and its online losses are self-consistency/multi-modal agreement terms, not fits to the evaluation target.

full rationale

The paper's central derivation chain is not circular. MoNetV2 is trained with ground-truth transformation parameters obtained from an electromagnetic (EM) positioning system, and the supervised loss in Eq. (4) directly compares estimated parameters to that independent ground truth. The reported evaluation metrics (FDR, ADR, MD, SD, HD, MEA) are also computed against EM-based ground truth. The online learning losses are consistency constraints: SVC (Eq. 5) enforces self-consistency between estimated transformations of subsequences, PAC (Eq. 6) enforces appearance-based consistency under reordering, PMC (Eq. 8) enforces agreement between normalized content differences and normalized estimated patch distances, and MSS (Eq. 10) enforces agreement between estimated acceleration/Euler angle and IMU measurements. None of these losses is fitted to the EM ground truth or to the evaluation metrics, so none constitutes a prediction that is equivalent to its input by construction. The use of IMU acceleration and Euler angle as both network inputs and pseudo-labels in MSS is a transparent multi-modal self-supervision strategy, not a hidden reuse of the target quantity. The preprocessing assumption in Eq. (2) that the probe velocity is zero at the first and last frames is an empirical assumption that may affect robustness, but it is not a circular derivation; it does not define the reconstruction target in terms of the output. The paper's self-citations to the authors' prior MoNet, OSCNet, and RecON works are used as baselines and as statements of incremental contribution, not as load-bearing justification for the method's correctness. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling-via-citation appears. The stated limitations (elastic tissue deformation, online inference time) are real limitations but are unrelated to circularity. Therefore the derivation is self-contained and no circular step is exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. The free parameters are standard network hyperparameters and loss weights chosen by ablation. The most important assumptions are the rest-state condition for acceleration preprocessing and the rigid-body assumption, both acknowledged in the paper.

free parameters (5)
  • Maximum sampling interval K for SVC/PMC = 11 (arm/carotid), 4 (thyroid)
    Chosen by hand based on dataset sequence length; affects the range of subsequences used in scan-level and patch-level consistency losses.
  • Number of interpolated images in PMC = 63
    Selected via ablation (Table VI); higher numbers give slightly better metrics but increase computation.
  • Patch grid size for PMC (P x Q) = 32 x 32
    Selected via ablation (Table VII) as the best balance between capturing fine-grained motion and avoiding speckle noise.
  • Online learning iterations and learning rate = 60 iterations, learning rate 2e-6
    Chosen by hand; the paper shows metric curves converge by 60 iterations but does not justify the specific learning rate.
  • Online loss weights = 1, 1, 1, 1
    Selected via ablation (Table VIII) as equal weighting gives the best results; deviations cause small performance drops.
assumptions (5)
  • domain assumption The probe is at rest at the first and last frames of each scan (V1 = VN = 0), so the mean of (Ai - gi) is zero and can be subtracted.
    Invoked in Eq. (2) to preprocess IMU acceleration; if false, the acceleration signal is biased.
  • domain assumption The transformation between adjacent US frames is rigid.
    Stated in the problem formulation (Sec III-A); the method regresses 6-DOF rigid transformations. The discussion acknowledges elastic tissue deformation as a limitation.
  • domain assumption The EM positioning system provides accurate ground truth for transformation parameters.
    Used for training and evaluation (Sec IV-A); EM accuracy is 1.4 mm position and 0.5 degree orientation, which introduces noise into the ground truth.
  • ad hoc to paper Image patch content difference is monotonically related to local 3D motion distance.
    Assumed in the patch-level motion consistency loss (Sec III-C3, Eq. 8). This heuristic may fail in homogeneous or shadowed regions.
  • domain assumption IMU and US data are temporally aligned at 30 fps.
    Required to fuse IMU signals with image frames (Sec IV-A); any desync would degrade the motion branch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoNetV2: Enhanced Motion Network for Freehand 3D Ultrasound Reconstruction." pith.science (2026). https://pith.science/paper/F2GN5SIS

@misc{pith2026250615835,
  author       = {Pith},
  title        = {Pith review of: MoNetV2: Enhanced Motion Network for Freehand 3D Ultrasound Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2GN5SIS}},
  note         = {Machine review of arXiv:2506.15835}
}
read the original abstract

Three-dimensional (3D) ultrasound (US) aims to provide sonographers with the spatial relationships of anatomical structures, playing a crucial role in clinical diagnosis. Recently, deep-learning-based freehand 3D US has made significant advancements. It reconstructs volumes by estimating transformations between images without external tracking. However, image-only reconstruction poses difficulties in reducing cumulative drift and further improving reconstruction accuracy, particularly in scenarios involving complex motion trajectories. In this context, we propose an enhanced motion network (MoNetV2) to enhance the accuracy and generalizability of reconstruction under diverse scanning velocities and tactics. First, we propose a sensor-based temporal and multi-branch structure that fuses image and motion information from a velocity perspective to improve image-only reconstruction accuracy. Second, we devise an online multi-level consistency constraint that exploits the inherent consistency of scans to handle various scanning velocities and tactics. This constraint exploits both scan-level velocity consistency, path-level appearance consistency, and patch-level motion consistency to supervise inter-frame transformation estimation. Third, we distill an online multi-modal self-supervised strategy that leverages the correlation between network estimation and motion information to further reduce cumulative errors. Extensive experiments clearly demonstrate that MoNetV2 surpasses existing methods in both reconstruction quality and generalizability performance across three large datasets.

Figures

Figures reproduced from arXiv: 2506.15835 by the authors.

Figure 1
Figure 1. Pipeline of freehand 3D US reconstruction with an [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of the proposed enhanced motion network (Mo [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of online scan-level velocity consistency [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Overview of online path-level appearance consisten [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Overview of online patch-level motion consistency [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of IMU data (φ and A, blue line) and EM positioning data (r and A, red line). The abscissa of each subfigure indicates the image index. B. Evaluation Metrics For an N-length scan I, we calculate the actual spatial position P = {Pi |i = 1, 2, · · · , N} based…
Figure 7
Figure 7. Figure 7: The histogram and kernel density estimation (KDE) of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Typical reconstruction cases on arm (Row I), carotid [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Metric curves (with 95% confidence interval) de [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Typical drift curves for MoNetV2 and RecON in the [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Three cases of vessel segmentation (brown) from the [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Voxel-to-voxel distance models between vessel seg [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 14
Figure 14. Figure 14: Boxplot of the percentage of reconstructed over rea [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Precision-K, Recall-K, and F1-score-K curves on ar [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Error boxplots for the in-plane displacement, out- [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: Correlation of the sequence lengths and inference [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 56 canonical work pages

  1. [1]

    S. Liu, Y . Wang, X. Y ang, B. Lei, L. Liu, S. X. Li, D. Ni, T. Wa ng, Deep learning in medical ultrasound analysis: a review, Eng ineering 5 (2) (2019) 261–275

    work page 2019

  2. [2]

    Demeulenaere, Z

    O. Demeulenaere, Z. Sandoval, P . Mateo, A. Dizeux, O. Vil lemain, R. Gallet, B. Ghaleh, T. Deffieux, C. Dem´ en´ e, M. Tanter, et al., Coronary flow assessment using 3-dimensional ultrafast ultrasound l ocalization microscopy, Cardiovascular Imaging 15 (7) (2022) 1193–120 8

    work page 2022

  3. [3]

    C. Chen, X. Y ang, Y . Huang, W. Shi, Y . Cao, M. Luo, X. Hu, L. Z hu, L. Y u, K. Y ue, et al., Fetusmapv2: Enhanced fetal pose estima tion in 3d ultrasound, Medical Image Analysis (2023) 103013

    work page 2023

  4. [4]

    Zhang, X

    J. Zhang, X. Tao, Y . Jiang, X. Wu, D. Y an, W. Xue, S. Zhuang, L. Chen, L. Luo, D. Ni, Application of convolution neural network alg orithm based on multicenter abus images in breast lesion detection , Frontiers in Oncology 12 (2022) 938413

    work page 2022

  5. [5]

    S. Xing, J. C. Romero, D. W. Cool, A. Mujoomdar, E. C. Chen, T. M. Peters, A. Fenster, 3d us-based evaluation and optimiz ation of tumor coverage for us-guided percutaneous liver thermal ab lation, IEEE Transactions on Medical Imaging 41 (11) (2022) 3344–3356

    work page 2022

  6. [6]

    Esposito, C

    M. Esposito, C. Hennersperger, R. G¨ obl, L. Demaret, M. S torath, N. Navab, M. Baust, A. Weinmann, Total variation regularization of pose signals with an application to 3d freehand ultrasound, IEEE Transactions on Medical Imaging 38 (10) (2019) 2245–2258

    work page 2019

  7. [7]

    H. Guo, H. Chao, S. Xu, B. J. Wood, J. Wang, P . Y an, Ultrasou nd volume reconstruction from freehand scans without tracking, IEEE Transactions on Biomedical Engineering 70 (3) (2023) 970–979

    work page 2023

  8. [8]

    M. Luo, X. Y ang, H. Wang, H. Dou, X. Hu, Y . Huang, N. Ravikum ar, S. Xu, Y . Zhang, Y . Xiong, W. Xue, A. F. Frangi, D. Ni, L. Sun, Re con: Online learning for sensorless freehand 3d ultrasound reco nstruction, Medical Image Analysis 87 (2023) 102810

    work page 2023

Show all 57 references
  1. [9]

    Q. Li, Z. Shen, Q. Li, D. C. Barratt, T. Dowrick, M. J. Clark son, T. V ercauteren, Y . Hu, Long-term dependency for 3d reconstr uction of freehand ultrasound without external tracker, IEEE Tran sactions on Biomedical Engineering (2023)

    2023

  2. [10]

    Y eung, L

    P .-H. Y eung, L. S. Hesse, M. Aliasi, M. C. Haak, W. Xie, A. I. Nambu- rete, I. 21st Consortium, et al., Sensorless volumetric rec onstruction of fetal brain freehand ultrasound scans with deep implicit re presentation, Medical Image Analysis (2024) 103147

    2024

  3. [11]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, Nature 52 1 (7553) (2015) 436–444

    2015

  4. [12]

    A. Bini, M. Bhat, Despeckling low snr, low contrast ultr asound images via anisotropic level set diffusion, Multidimensional Sys tems and Signal Processing 25 (2014) 41–65

    2014

  5. [13]

    Ahmad, R

    N. Ahmad, R. A. R. Ghazilla, N. M. Khairi, V . Kasi, Review s on var- ious inertial measurement unit (imu) sensor applications, International Journal of Signal Processing Systems 1 (2) (2013) 256–262

    2013

  6. [14]

    D. B. Downey, A. Fenster, V ascular imaging with a three- dimensional power doppler system., AJR. American journal of roentgenol ogy 165 (3) (1995) 665–668

    1995

  7. [15]

    J.-F. Chen, J. B. Fowlkes, P . L. Carson, J. M. Rubin, Dete rmination of scan-plane motion using speckle decorrelation: Theoret ical consid- erations and initial test, International Journal of Imagin g Systems and Technology 8 (1) (1997) 38–44

    1997

  8. [16]

    T. A. Tuthill, J. Kr¨ ucker, J. B. Fowlkes, P . L. Carson, A utomated three- dimensional us frame positioning computed from elevationa l speckle decorrelation., Radiology 209 (2) (1998) 575–582

    1998

  9. [17]

    A. H. Gee, R. James Housden, P . Hassenpflug, G. M. Treece, R. W. Prager, Sensorless freehand 3d ultrasound in real tissue: S peckle decor- relation without fully developed speckle, Medical Image An alysis 10 (2) (2006) 137–149

    2006

  10. [18]

    R. W. Prager, A. H. Gee, G. M. Treece, C. J. Cash, L. H. Berm an, Sensorless freehand 3-d ultrasound using regression of the echo intensity, Ultrasound in medicine & biology 29 (3) (2003) 437–446

    2003

  11. [19]

    Tetrel, H

    L. Tetrel, H. Chebrek, C. Laporte, Learning for graph-b ased sensorless freehand 3d ultrasound, in: Machine Learning in Medical Ima ging, 2016, pp. 205–212

    2016

  12. [20]

    Hirano, A

    H. Hirano, A. Minagi, K. Takemoto, Universal adversari al attacks on deep neural networks for medical image classification, BMC m edical imaging 21 (2021) 1–13

    2021

  13. [21]

    Zhang, H

    S. Zhang, H. Fu, Y . Y an, Y . Zhang, Q. Wu, M. Y ang, M. Tan, Y . Xu, Attention guided network for retinal image segmentat ion, in: International Conference on Medical Image Computing and Co mputer- Assisted Intervention, Springer, 2019, pp. 797–805

    2019

  14. [22]

    X. Ruan, M. Dai, Z. Chen, Z. Y ou, Y . Zhang, Y . Li, Z. Dou, M. Tan, Temporal micro-action localization for videofluoroscopic swallowing study, IEEE Journal of Biomedical and Health Informatics (2 023)

  15. [23]

    Zhang, Y

    Y . Zhang, Y . Wei, Q. Wu, P . Zhao, S. Niu, J. Huang, M. Tan, C ollabora- tive unsupervised domain adaptation for medical image diag nosis, IEEE Transactions on Image Processing 29 (2020) 7834–7844

    2020

  16. [24]

    L. Sun, K. Y u, K. Batmanghelich, Context matters: Graph -based self- supervised representation learning for medical images, in : Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 35, 2 021, pp. 4874–4882

  17. [25]

    Huang, A

    C. Huang, A. Jiang, J. Feng, Y . Zhang, X. Wang, Y . Wang, Ad apting visual-language models for generalizable anomaly detecti on in medical images, in: IEEE/CVF Computer Vision and Pattern Recogniti on Con- ference (CVPR), 2024, pp. 11375–11385

    2024

  18. [26]

    Prevost, M

    R. Prevost, M. Salehi, J. Sprung, A. Ladikos, R. Bauer, W . Wein, Deep learning for sensorless 3d freehand ultrasound imaging, in : International Conference on Medical Image Computing and Computer-Assist ed Inter- vention, Springer, 2017, pp. 628–636

    2017

  19. [27]

    Z. Li, F. Liu, W. Y ang, S. Peng, J. Zhou, A survey of convol utional neural networks: analysis, applications, and prospects, IEEE Tra nsactions on Neural Networks and Learning Systems 33 (12) (2021) 6999–70 19

    2021

  20. [28]

    H. Guo, S. Xu, B. Wood, P . Y an, Sensorless freehand 3d ult rasound reconstruction via deep contextual learning, in: Internat ional Conference on Medical Image Computing and Computer-Assisted Interven tion, Springer, 2020, pp. 463–472

    2020

  21. [29]

    Y eung, L

    P .-H. Y eung, L. Hesse, M. Aliasi, M. Haak, W. Xie, A. I. Na mburete, et al., Implicitvol: Sensorless 3d ultrasound reconstruct ion with deep implicit representation, arXiv preprint arXiv:2109.1210 8 (2021)

    2021

  22. [30]

    Prevost, M

    R. Prevost, M. Salehi, S. Jagoda, N. Kumar, J. Sprung, A. Ladikos, R. Bauer, O. Zettinig, W. Wein, 3d freehand ultrasound witho ut external JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 19 tracking using deep learning, Medical image analysis 48 (20 18) 187– 202

    2021

  23. [31]

    Miura, K

    K. Miura, K. Ito, T. Aoki, J. Ohmiya, S. Kondo, Localizin g 2d ultrasound probe from ultrasound image sequences using deep learning f or volume reconstruction, in: Medical Ultrasound, and Preterm, Peri natal and Paediatric Image Analysis, 2020, pp. 97–105

    2020

  24. [32]

    Kaderdina, M

    A. Kaderdina, M. J. B. Suarez, R. Garbi, E. Schaeffer, K. Mulpuri, A. Hodgson, Using deep learning to estimate frame-to-frame angle displacements in 2d ultrasound image sequences of an infant hip, CMBES Proceedings 45 (2023)

    2023

  25. [33]

    El Hadramy, J

    S. El Hadramy, J. V erde, K.-P . Beaudet, N. Padoy, S. Coti n, Tracker- less volume reconstruction from intraoperative ultrasoun d images, in: International Conference on Medical Image Computing and Co mputer- Assisted Intervention, Springer, 2023, pp. 303–312

    2023

  26. [34]

    Jordan, Serial order: a parallel distributed proces sing approach

    M. Jordan, Serial order: a parallel distributed proces sing approach. technical report, june 1985-march 1986, Tech. rep., Califo rnia Univ., San Diego, La Jolla (USA). Inst. for Cognitive Science (1986 )

    1985

  27. [35]

    V aswani, N

    A. V aswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jone s, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need, in: I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Syste...

    2017

  28. [36]

    Y . Liu, Y . Zhang, Y . Wang, F. Hou, J. Y uan, J. Tian, Y . Zhan g, Z. Shi, J. Fan, Z. He, A survey of visual transformers, IEEE Transact ions on Neural Networks and Learning Systems (2023)

    2023

  29. [37]

    M. Luo, X. Y ang, X. Huang, Y . Huang, Y . Zou, X. Hu, N. Ravik umar, A. F. Frangi, D. Ni, Self context and shape prior for sensorle ss freehand 3d ultrasound reconstruction, in: International Conferen ce on Medical Image Computing and Computer-Assisted Intervention, Spri nger...

    2021

  30. [38]

    G. Ning, H. Liang, L. Zhou, X. Zhang, H. Liao, Spatial pos ition estimation method for 3d ultrasound reconstruction based o n hybrid transfomers, in: 2022 IEEE 19th International Symposium on Biomed- ical Imaging (ISBI), 2022, pp. 1–5

    2022

  31. [39]

    Q. Li, Z. Shen, Q. Li, D. C. Barratt, T. Dowrick, M. J. Clar kson, T. V ercauteren, Y . Hu, Trackerless freehand ultrasound wit h sequence modelling and auxiliary transformation over past and futur e frames, in: 2023 IEEE 20th International Symposium on Biomedical Im aging (IS...

    2023

  32. [40]

    X. Chen, H. Chen, Y . Peng, L. Liu, C. Huang, A freehand 3d u ltrasound reconstruction method based on deep learning, Electronics 12 (7) (2023) 1527

    2023

  33. [41]

    Housden, A

    R. Housden, A. H. Gee, R. W. Prager, G. M. Treece, Rotatio nal motion in sensorless freehand three-dimensional ultrasound, Ult rasonics 48 (5) (2008) 412–422

    2008

  34. [42]

    K. Ito, K. Y odokawa, T. Aoki, J. Ohmiya, S. Kondo, A probe -camera system for 3d ultrasound image reconstruction, in: Imaging for Patient- Customized Simulations and Systems for Point-of-Care Ultr asound, 2017, pp. 129–137

    2017

  35. [43]

    M. Luo, X. Y ang, H. Wang, L. Du, D. Ni, Deep motion network for freehand 3d ultrasound reconstruction, in: International Conference on Medical Image Computing and Computer-Assisted Interventi on, 2022, pp. 290–299

    2022

  36. [44]

    M. Luo, X. Y ang, Z. Y an, J. Li, Y . Zhang, J. Chen, X. Hu, J. Q ian, J. Chen, D. Ni, Multi-imu with online self-consistency for f reehand 3d ultrasound reconstruction, in: International Conferen ce on Medical Image Computing and Computer-Assisted Intervention, 2023

    2023

  37. [45]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning fo r image recognition, in: IEEE/CVF Computer Vision and Pattern Reco gnition Conference (CVPR), 2016, pp. 770–778

    2016

  38. [46]

    X. Shi, Z. Chen, H. Wang, D.-Y . Y eung, W.-K. Wong, W.-c. W oo, Convolutional lstm network: A machine learning approach fo r precipi- tation nowcasting, Advances in neural information process ing systems 28 (2015)

    2015

  39. [47]

    F. He, T. Liu, D. Tao, Why resnet works? residuals genera lize, IEEE Transactions on Neural Networks and Learning Systems 31 (12 ) (2020) 5349–5362

    2020

  40. [48]

    Greff, R

    K. Greff, R. K. Srivastava, J. Koutn´ ık, B. R. Steunebri nk, J. Schmid- huber, Lstm: A search space odyssey, IEEE Transactions on Ne ural Networks and Learning Systems 28 (10) (2017) 2222–2232

    2017

  41. [49]

    F. Reda, J. Kontkanen, E. Tabellion, D. Sun, C. Pantofar u, B. Curless, Film: Frame interpolation for large motion, in: European Co nference on Computer Vision (ECCV), 2022

    2022

  42. [50]

    Levenberg, A method for the solution of certain non-l inear problems in least squares, Quarterly of applied mathematics 2 (2) (19 44) 164–168

    K. Levenberg, A method for the solution of certain non-l inear problems in least squares, Quarterly of applied mathematics 2 (2) (19 44) 164–168

  43. [51]

    D. P . Kingma, J. Ba, Adam: A method for stochastic optimi zation, in: International Conference on Learning Representations (IC LR), 2015

    2015

  44. [52]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. C hanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch : An imperative style, high-performance deep learning library , Advances in Neural Information Processing System 32 (2019) 8026–8037

    2019

  45. [53]

    H. Song, M. Kim, D. Park, Y . Shin, J.-G. Lee, Learning fro m noisy labels with deep neural networks: A survey, IEEE Transactions on Ne ural Networks and Learning Systems (2022)

    2022

  46. [54]

    Liang, X

    J. Liang, X. Y ang, Y . Huang, H. Li, S. He, X. Hu, Z. Chen, W. Xue, J. Cheng, D. Ni, Sketch guided and progressive growing gan fo r realistic and editable ultrasound image synthesis, Medical Image Ana lysis 79 (2022) 102461

    2022

  47. [55]

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A . Xiao, C. Xu, Y . Xu, et al., A survey on vision transformer, IEEE tran sactions on pattern analysis and machine intelligence 45 (1) (2022) 8 7–110

    2022

  48. [56]

    S. Xie, R. Girshick, P . Doll´ ar, Z. Tu, K. He, Aggregated residual transformations for deep neural networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 201 7, pp. 1492– 1500

  49. [57]

    D. Qin, C. Leichner, M. Delakis, M. Fornoni, S. Luo, F. Y a ng, W. Wang, C. Banbury, C. Y e, B. Akin, et al., Mobilenetv4-universal mo dels for the mobile ecosystem, arXiv preprint arXiv:2404.10518 (20 24)

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.