Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Super-temporal-resolution Photoacoustic Imaging with Dynamic Reconstruction through Implicit Neural Representation in Sparse-view

T0 review · 4 major / 6 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 An implicit neural representation, trained only on sparse sensor data, reconstructs dynamic photoacoustic images and interpolates unmeasured frames for 4x temporal super-resolution.

desk verdict A plausible first application of INR spatiotemporal reconstruction to dynamic PACT, but the quantitative claims rest on an inverse-crime-prone simulation and the temporal super-resolution is only qualitative. read the letter →

arxiv 2506.03175 v1 pith:YJPMGWXL submitted 2025-05-29 eess.IV cs.CV

Youshen Xiao , Yiling Shi , Ruixi Sun , Hongjiang Wei , Fei Gao , Yuyao Zhang This is my paper
classification eess.IVcs.CV
keywords photoacousticcomputedtomographyimplicitneuralrepresentationsparse-viewreconstructiontemporalsuper-resolutiondynamicimagingunsuperviseddeeplearninglow-rankregularizationtotalvariation
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

The paper claims that dynamic photoacoustic computed tomography can be reconstructed from sparse-view sensor data by treating the whole image sequence as a single continuous function of space and time, encoded in a small multilayer perceptron. The network is trained only on the measured sensor signals, with no external dataset or ground truth, using data consistency plus temporal total variation and low-rank penalties. On a simulated mouse-abdomen dataset the method reaches average PSNR of 39.87 dB with 128 sensors and 37.69 dB with 64 sensors, outperforming classical delay-and-sum and universal back-projection reconstructions. Because the learned function is continuous in time, the same trained network can be queried at denser time coordinates to produce interpolated frames, giving 4x temporal super-resolution from one quarter of the acquired frames. Qualitative results on live finger and fish data support the same conclusion.

What carries the argument

The central object is the implicit neural representation of the dynamic image: a small multilayer perceptron with three hidden layers of 256 ReLU neurons and a Sigmoid output that maps Fourier-encoded spatiotemporal coordinates to initial-pressure values. The Fourier feature mapping $\gamma(p) = [\cos(2\pi B p), \sin(2\pi B p)]$ with $L=256$ frequency bands is adopted so the network can represent high-frequency spatial and temporal details that raw-coordinate MLPs tend to miss. The forward photoacoustic model $A$ projects the network output into sensor space to define the data-consistency loss, while temporal total variation and nuclear-norm penalties on the image sequence provide explicit regularization. During inference the optimized function is evaluated at any chosen set of spatiotemporal coordinates, so both spatial and temporal sampling densities can be increased without changing the network.

What would settle it

Take a dynamic PACT phantom with a known ground-truth sequence containing a target moving at increasing speed or a pulsating source, train the network on one quarter of the frames with 64 sensors, and compare interpolated frames to ground truth. If the interpolated-frame PSNR falls below the simple linear-interpolation baseline or below the DAS reconstruction as speed increases, the temporal-continuity assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that a dynamic PACT image sequence is well represented as an implicit continuous mapping $f: (x,y,t) \mapsto I$, parameterized by an MLP with Fourier feature encoding, whose weights are optimized directly against the sparse sensor measurements $y$ via the loss $\|y - A x_\theta\|_2^2 + \lambda_D \|\mathrm{TV}_t(x_\theta)\|_1 + \lambda_L \|x_\theta\|_*$. With this loss, no training database or prior images are needed, and the continuity of the representation acts as an implicit regularizer alongside the explicit sparsity and low-rank terms. The paper demonstrates on simulated mouse abdominal data that this yields cleaner reconstructions than DAS and UBP at both 128 and 64 sensors, and in a separate experiment shows that a network trained on only five of twenty frames can interpolate the remaining fifteen frames with high fidelity, achieving 4x temporal super-resolution without retraining.

Load-bearing premise

The load-bearing premise is that the dynamic scene changes smoothly enough in time that a single continuous MLP trained on a subset of frames can accurately interpolate the unacquired frames; the paper notes the need to reduce large inter-frame motion differences but does not quantify the motion regime.

Editorial extensions

If this is right

  • Frame rate of a dynamic PACT system can be increased fourfold from the same raw acquisition by training the network on one quarter of the frames and querying the other time points.
  • Sparse sensor arrays with as few as 64 elements become practical for dynamic imaging, since the method's PSNR on simulated data stays above 37 dB while DAS and UBP degrade sharply.
  • The method transfers across transducer geometries and anatomical sites without retraining on external data, because the network learns only the current scan's sensor signals.
  • Temporal interpolation is achieved without extra hardware or a separate registration step, which would lower deployment cost for high-frame-rate photoacoustic imaging.
  • The explicit low-rank and temporal-TV penalties plus implicit continuity may also suppress background noise in high-noise in vivo acquisitions, as shown on the fish data.

Reading between the lines

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

  • Beyond the paper: if the smooth-temporal-variation assumption holds, the same coordinate-to-intensity framework could be applied to other dynamic inverse problems with a known forward model, such as contrast-enhanced ultrasound or dynamic fluorescence tomography, where temporal continuity is similarly strong.
  • Beyond the paper: the paper does not characterize the range of inter-frame motion over which interpolation stays accurate; a systematic study varying frame spacing and target speed on a phantom would map the method's failure envelope.
  • Beyond the paper: the iterative per-scan optimization is not real-time, so a natural extension is meta-learning an initialization or amortizing the INR across similar anatomies to preserve the continuity prior while cutting reconstruction time.
  • Beyond the paper: ablating the temporal TV and low-rank terms separately would identify which regularizer carries the interpolation capability, since the implicit INR continuity alone might be sufficient or insufficient.
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

4 major / 6 minor

Summary. The paper proposes an unsupervised implicit neural representation (INR) method for dynamic photoacoustic computed tomography (PACT) reconstruction from sparse-view sensor data. The method trains a multilayer perceptron with Fourier feature encoding to map spatiotemporal coordinates to image intensity, using a data-consistency loss against the measured sensor data plus explicit temporal total variation and low-rank regularizers. The authors claim substantial improvements over classical DAS and UBP reconstruction on a simulated mouse abdominal PACT dataset (e.g., PSNR 37.69 dB with 64 sensors), and demonstrate 4x temporal super-resolution by training on one quarter of the frames and querying denser time coordinates. Real-data experiments on finger and fish are presented as qualitative validation of generalization.

Significance. If the quantitative claims are established, the method would be a valuable unsupervised alternative to supervised deep-learning reconstruction, with the notable advantage of temporal interpolation without retraining and no need for external training data. The paper ships a code repository, which is a concrete strength for reproducibility, and the idea of using INR continuity as an implicit regularizer in dynamic PACT is timely given recent work in dynamic MRI and cone-beam CT. However, the central quantitative evidence currently rests on a simulation that may suffer from inverse crime, a non-standard PSNR metric, and a missing comparison against direct optimization of the same regularized objective; the temporal super-resolution claim is not quantitatively evaluated. These gaps prevent the paper from establishing its headline improvements, though they are addressable with additional experiments.

major comments (4)
  1. [IV-A1, III-C] The simulation in §IV-A1 generates sinograms with a forward model from the mouse ground truth, and the loss L_DC in Eq. (14) uses the same forward operator A. The paper does not state that the simulation used a different discretization, a different numerical solver (e.g., k-Wave), or added noise. Under these conditions, iterative model-based and INR methods enjoy an 'inverse-crime' advantage over DAS/UBP, so the reported PSNR margins (18.20-18.76 dB, §IV-C1) may be largely a simulation artifact. Please rerun the simulation with an independent forward solver and/or with measurement noise, or explicitly justify why the inverse crime is not affecting the comparison.
  2. [Eq. (15), §IV-B] Equation (15) defines PSNR = 10 log10(1/||y - yhat||_2^2), without normalizing by the signal dynamic range or the number of pixels. Standard PSNR is 10 log10(MAX^2 / MSE). As written, the reported values (e.g., 37.69 dB) are not comparable to standard PSNR values from the literature, and the absolute dB improvements over DAS/UBP are difficult to interpret. Please report standard PSNR with the normalization used, and also state the normalization of GT and reconstructed images.
  3. [IV-D, Fig. 6] The temporal super-resolution experiment is only visually assessed: no PSNR/SSIM is reported for the held-out frames (e.g., frames 6-8 of the 20-frame mouse sequence) that were not used in training. Since the core claim of 4x temporal super-resolution rests on interpolating unacquired frames, quantitative evaluation against the GT of those held-out frames is required. Additionally, the smoothness assumption underlying temporal interpolation is not characterized; the paper notes in §IV-D that it chose one-quarter of the frames 'to reduce the impact of large inter-frame motion differences,' but no motion analysis (e.g., frame-to-frame displacement or residual analysis) is provided to define the regime of validity.
  4. [III-C, IV-B] The regularizers in Eq. (14) are the same temporal TV and low-rank terms as in the classical formulation Eq. (9), yet no comparison is made against directly solving Eq. (9) with a non-INR iterative optimizer (e.g., ADMM or FISTA). Without this baseline, the claimed superiority over DAS/UBP cannot be attributed to the INR representation itself, rather than to the explicit regularization terms. Please add such a baseline and, if possible, an ablation that removes the explicit regularizers to isolate the contribution of INR.
minor comments (6)
  1. [Eq. (7)] The dimensions of y are inconsistent: with A in R^{(S×F)×(N×N)} and x in R^{(N×N)×T}, y should be in R^{(S×F)×T}, not in R^{((S×F)×T)×T} as written.
  2. [IV-C2, IV-B] The text in §IV-C2 mentions TR as a baseline, but §IV-B lists only DAS and UBP; Fig. 4 includes TR. Clarify whether TR was compared and with which settings.
  3. [Fig. 4 and Fig. 5 captions] The Fig. 4 caption refers to an enlarged view of the 'heart region' in finger data; this appears to be a vessel region. Also, error maps cannot be computed for real data without GT; clarify what is actually displayed at the bottom of Fig. 4 and Fig. 5.
  4. [IV-A2, Fig. 4] §IV-A2 states the finger data were uniformly downsampled to simulate a sparse-view configuration with 64 sensors, but Fig. 4 shows results for 16, 32, 64, and 128 sensors; clarify the actual configurations used.
  5. [Eq. (16)] SSIM constants are given as c1=0.01^2 and c2=0.032; standard SSIM uses c2=(0.03 L)^2. Specify the value of L (dynamic range) and the intended constants, or correct the typo.
  6. [General] There are minor typos, including 'availble' in the code link and inconsistent capitalization of 'Fourier feature mapping' in §III-B.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reconstruction and temporal-interpolation claims are evaluated against external ground truth and held-out frames, with only a non-load-bearing self-citation.

full rationale

The paper's derivation chain is self-contained. The INR reconstruction is obtained by minimizing Eq. (14), L_total = ||y - A x_theta||_2^2 + lambda_D ||TV_t(x_theta)||_1 + lambda_L ||x_theta||_*, with x_theta generated by the MLP (Eqs. 10-12); this is a fitting procedure, not a derivation of the ground truth from the loss. The central quantitative claims are evaluated against the public mouse PACT ground-truth images (Sec. IV-A1, IV-C1) with PSNR/SSIM computed against GT (Eqs. 15-16), and the temporal-super-resolution claim is tested on held-out frames (Fig. 6: frames 6-8 between trained frames 5 and 9), so it is not equivalent to its training input. The self-citation of Feng et al. [21] (overlapping authors) is used only as prior INR-in-MRI inspiration; the PACT forward model, losses, and experiments are independent, so it is not load-bearing. The concern that the simulated data may be generated with the same forward operator A used in the data-consistency term (Sec. IV-A1 vs. Eq. 14) is a benchmark-validity/inverse-crime issue that could inflate measured gains, but it does not make the reported prediction reduce to its inputs by construction; the GT comparison remains external. No circular step was found.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method is an empirical fitting procedure with no new physical entities. It rests on three categories of assumptions: the PACT forward model, the INR capacity and continuity hypothesis, and the transfer of TV+low-rank priors from MRI to PACT. Key hyperparameters (λ_D, λ_L, σ) are tuned on the test dataset and not reported, so the quantitative results are partly a function of free parameters.

free parameters (4)
  • λ_D (temporal TV weight) = not reported
    Weights the temporal total variation penalty in Eq. 14; stated in Sec. III-D to be chosen empirically and fine-tuned on experimental results. No value is given, and it controls the smoothness of the reconstructed sequence.
  • λ_L (low-rank/nuclear norm weight) = not reported
    Weights the nuclear norm penalty in Eq. 14; also chosen empirically. No value given, affects the rank-reduction strength.
  • σ (Fourier feature scale) = not reported
    Defines the Gaussian distribution N(0, σ^2) from which the Fourier feature matrix B is sampled (Sec. III-B). Not specified in the paper; it sets the frequency bands of the encoding and thus the resolution of details.
  • L (number of Fourier frequency bands) = 256
    Fixed to 256 in Sec. III-D; this is an architectural choice that affects the frequency coverage of the encoding and was chosen empirically.
assumptions (4)
  • domain assumption The PACT forward model in Eqs. (1)-(6) accurately describes the acoustic wave propagation and sensor measurements for both simulation and in vivo data.
    The data consistency loss in Eq. 14 relies on A from Eq. 7; any mismatch between the model and the real system would bias the reconstructed images. Used throughout the method.
  • ad hoc to paper The dynamic PACT image sequence can be represented as a continuous mapping from spatiotemporal coordinates to intensity (Eq. 10).
    This is the core INR hypothesis; it presumes the MLP has sufficient capacity and that the continuity prior is appropriate for the image sequence.
  • domain assumption Temporal total variation and low-rank structure are valid priors for dynamic PACT images (Eq. 9, 14).
    These priors are transferred from dynamic MRI works [24,25] without validation on PACT data; if the image sequence is not temporally sparse or low-rank, the regularizers may distort the reconstruction.
  • ad hoc to paper A single MLP with Fourier features can be optimized via gradient descent to fit the image sequence from sparse sensor data.
    The paper relies on empirical optimization success; no convergence guarantee or analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Super-temporal-resolution Photoacoustic Imaging with Dynamic Reconstruction through Implicit Neural Representation in Sparse-view." pith.science (2026). https://pith.science/paper/YJPMGWXL

@misc{pith2026250603175,
  author       = {Pith},
  title        = {Pith review of: Super-temporal-resolution Photoacoustic Imaging with Dynamic Reconstruction through Implicit Neural Representation in Sparse-view},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJPMGWXL}},
  note         = {Machine review of arXiv:2506.03175}
}
read the original abstract

Dynamic Photoacoustic Computed Tomography (PACT) is an important imaging technique for monitoring physiological processes, capable of providing high-contrast images of optical absorption at much greater depths than traditional optical imaging methods. However, practical instrumentation and geometric constraints limit the number of acoustic sensors available around the imaging target, leading to sparsity in sensor data. Traditional photoacoustic (PA) image reconstruction methods, when directly applied to sparse PA data, produce severe artifacts. Additionally, these traditional methods do not consider the inter-frame relationships in dynamic imaging. Temporal resolution is crucial for dynamic photoacoustic imaging, which is fundamentally limited by the low repetition rate (e.g., 20 Hz) and high cost of high-power laser technology. Recently, Implicit Neural Representation (INR) has emerged as a powerful deep learning tool for solving inverse problems with sparse data, by characterizing signal properties as continuous functions of their coordinates in an unsupervised manner. In this work, we propose an INR-based method to improve dynamic photoacoustic image reconstruction from sparse-views and enhance temporal resolution, using only spatiotemporal coordinates as input. Specifically, the proposed INR represents dynamic photoacoustic images as implicit functions and encodes them into a neural network. The weights of the network are learned solely from the acquired sparse sensor data, without the need for external training datasets or prior images. Benefiting from the strong implicit continuity regularization provided by INR, as well as explicit regularization for low-rank and sparsity, our proposed method outperforms traditional reconstruction methods under two different sparsity conditions, effectively suppressing artifacts and ensuring image quality.

Figures

Figures reproduced from arXiv: 2506.03175 by the authors.

Figure 1
Figure 1. Overview of the proposed method is as follows: All spatiotemporal coordinates are first input into a Fourier encoding and then fed into a MLP to [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The reconstruction results on the simulated dynamic abdominal PACT data of mouse for DAS, UBP, and the proposed method (from left to right) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The reconstruction results on the simulated dynamic abdominal PACT data of mouse for DAS, UBP, and the proposed method (from left to right) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The reconstruction results on in vivo finger data for DAS, UBP, TR, and the proposed method (from left to right) are shown with 16, 32, 64, and 128 sensors(projections). The enlarged view of the heart region is highlighted by yellow boxes, and it is noted that the prop…
Figure 5
Figure 5. Figure 5: The reconstruction results on in vivo fish data for DAS, UBP, and the proposed method (from left to right) are shown under sparse-view conditions with 128 and 64 sensors. The enlarged view of the fish head region is highlighted by orange boxes, and it is clearly observ…
Figure 6
Figure 6. Figure 6: (a) Temporal super-resolution pipeline for dynamic PACT recon [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

  1. [1]

    Yong Zhou, Junjie Yao, and Lihong V . Wang. Tutorial on photoacoustic tomography.Journal of Biomedical Optics, 21, 2016

    work page 2016

  2. [2]

    Lihong V . Wang. Tutorial on photoacoustic microscopy and computed tomography.IEEE Journal of Selected Topics in Quantum Electronics, 14:171–179, 2008

    work page 2008

  3. [3]

    Lihong V . Wang. Multiscale photoacoustic microscopy and computed tomography.Nature photonics, 3 9:503–509, 2009

    work page 2009

  4. [4]

    The delay multiply and sum beamforming algorithm in ultrasound b-mode medical imaging.IEEE Transactions on Medical Imaging, 34:940–949, 2015

    Giulia Matrone, Alessandro Stuart Savoia, Giosu `e Caliano, and Giovanni Magenes. The delay multiply and sum beamforming algorithm in ultrasound b-mode medical imaging.IEEE Transactions on Medical Imaging, 34:940–949, 2015

    work page 2015

  5. [5]

    Photoacoustic tomography in absorbing acoustic media using time reversal.Inverse Problems, 26(11):115003, sep 2010

    Bradley E Treeby, Edward Z Zhang, and B T Cox. Photoacoustic tomography in absorbing acoustic media using time reversal.Inverse Problems, 26(11):115003, sep 2010

    work page 2010

  6. [6]

    Minghua Xu and Lihong V . Wang. Universal back-projection algorithm for photoacoustic computed tomography. InSPIE BiOS, 2005

    work page 2005

  7. [7]

    Kagadis, George K

    George C. Kagadis, George K. Loudos, Konstantinos Katsanos, Steve G. Langer, and George Nikiforidis. In vivo small animal imaging: current status and future prospects.Medical physics, 37 12:6421–42, 2010

    work page 2010

  8. [8]

    Franc, Paul D

    Benjamin L. Franc, Paul D. Acton, Carina Mar ´ı, and Bruce H. Hasegawa. Small-animal spect and spect/ct: Important tools for pre- clinical investigation*.Journal of Nuclear Medicine, 49:1651 – 1663, 2008

    work page 2008

Show all 37 references
  1. [9]

    Real-time ultrasound doppler tracking and autonomous navigation of a miniature helical robot for accelerating thrombolysis in dynamic blood flow.ACS nano, 2022

    Qianqian Wang, Xingzhou Du, Dongdong Jin, and Li Zhang. Real-time ultrasound doppler tracking and autonomous navigation of a miniature helical robot for accelerating thrombolysis in dynamic blood flow.ACS nano, 2022

    2022

  2. [10]

    Real-time two-dimensional blood flow imaging using ultrasound doppler.Japanese Journal of Applied Physics, 26:9, 1987

    Chihiro Kasai and Koroku Namekawa. Real-time two-dimensional blood flow imaging using ultrasound doppler.Japanese Journal of Applied Physics, 26:9, 1987

    1987

  3. [11]

    Roberta Fusco, Vincenza Granata, Mauro Mattace Raso, Paolo Vallone, Alessandro Pasquale De Rosa, Claudio Siani, Maurizio di Bonito, An- tonella Petrillo, and Mario Sansone. Blood oxygenation level dependent magnetic resonance imaging (mri), dynamic contrast enhanced mri, and d...

    2021

  4. [12]

    Tofts, Gunnar Brix, David L

    Paul S. Tofts, Gunnar Brix, David L. Buckley, Jeffrey L. Evelhoch, Elizabeth Henderson, Michael V . Knopp, Henrik B. W. Larsson, Ting- Yim Lee, Nina A. Mayr, Geoffrey J. M. Parker, Ruediger E. Port, June S. Taylor, and Robert M. Weisskoff. Estimating kinetic parameters from dy...

    1999

  5. [13]

    Madhavan, Jared T

    Ajay A. Madhavan, Jared T. Verdoorn, Darya P. Shlapak, M. C. Pope, John C. Benson, Felix E. Diehn, Greta B. Liebo, Vance T. Lehman, Dong Kun Kim, Michael P. Oien, Jeremy K. Cutsforth-Gregory, and Carrie M. Carr. Lateral decubitus dynamic ct myelography for fast cerebrospinal f...

    1903

  6. [14]

    Shah, Robin Schepers, Patrik Caluori, Vladimir Y

    Hasan Sari, Clemens Mingels, Ian Leigh Alberts, Jicun Hu, Dorothee Buesser, Vijay P. Shah, Robin Schepers, Patrik Caluori, Vladimir Y . Panin, Maurizio Conti, Ali Afshar-Oromieh, Kuangyu Shi, Lars E. Eriksson, Axel Rominger, and Paul Cumming. First results on kinetic modelling...

    2022

  7. [15]

    Ermilov, Mark A

    Refik Mert C ¸ am, Chao Wang, Weylan Thompson, Sergey A. Ermilov, Mark A. Anastasio, and Umberto Villa. Spatiotemporal image recon- struction to enable high-frame-rate dynamic photoacoustic tomography with rotating-gantry volumetric imagers.Journal of Biomedical Optics, 29, 2023

    2023

  8. [16]

    Current and future trends in photoacoustic breast imaging.Photoacoustics, 16, 2019

    Srirang Manohar and Maura Dantuma. Current and future trends in photoacoustic breast imaging.Photoacoustics, 16, 2019

    2019

  9. [17]

    Dynamic in vivo imaging of small animal brain using pulsed laser diode-based photoacoustic tomography system.Journal of biomedical optics, 22 9:1–4, 2017

    Paul Kumar Upputuri and Manojit Pramanik. Dynamic in vivo imaging of small animal brain using pulsed laser diode-based photoacoustic tomography system.Journal of biomedical optics, 22 9:1–4, 2017

    2017

  10. [18]

    Appleton, Konstantin Maslov, Lei Li, Ruiying Zhang, and Lihong V

    Li Lin, Peng Hu, Junhui Shi, Catherine M. Appleton, Konstantin Maslov, Lei Li, Ruiying Zhang, and Lihong V . Wang. Single-breath-hold pho- toacoustic computed tomography of the breast.Nature Communications, 9, 2018

    2018

  11. [19]

    Deep learning enhances multiparametric dynamic volumetric photoacoustic computed tomography in vivo (dl-pact).Advanced Science, 10, 2022

    Seongwook Choi, Jinge Yang, Soo Young Lee, Jiwoong Kim, Jihye Lee, Won Jong Kim, Seungchul Lee, and Chulhong Kim. Deep learning enhances multiparametric dynamic volumetric photoacoustic computed tomography in vivo (dl-pact).Advanced Science, 10, 2022. JOURNAL OF LATEX CLASS FI...

    2022

  12. [20]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions.ArXiv, abs/2006.09661, 2020

    2006 arXiv

  13. [21]

    Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction.ArXiv, abs/2301.00127, 2022

    Jie Feng, Rui jun Feng, Qing Wu, Zhiyong Zhang, Yuyao Zhang, and Hongjiang Wei. Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction.ArXiv, abs/2301.00127, 2022

    2022

  14. [22]

    Dynamic cone-beam ct reconstruction using spatial and temporal implicit neural representation learning (stinr).Physics in Medicine and Biology, 68, 2022

    You Zhang, Hua-Chieh Shao, Tinsu Pan, and Tielige Mengke. Dynamic cone-beam ct reconstruction using spatial and temporal implicit neural representation learning (stinr).Physics in Medicine and Biology, 68, 2022

    2022

  15. [23]

    Accel- eration of optoacoustic model-based reconstruction using angular image discretization.IEEE Transactions on Medical Imaging, 31:1154–1162, 2012

    Xos ´e Lu´ıs De´an-Ben, Vasilis Ntziachristos, and Daniel Razansky. Accel- eration of optoacoustic model-based reconstruction using angular image discretization.IEEE Transactions on Medical Imaging, 31:1154–1162, 2012

    2012

  16. [24]

    Sajan Goud Lingala, Yue Hu, Edward V . R. Di Bella, and Mathews Jacob. Accelerated dynamic mri exploiting sparsity and low-rank structure: k-t slr.IEEE Transactions on Medical Imaging, 30:1042– 1054, 2011

    2011

  17. [25]

    Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints.IEEE transactions on medical imaging, 31(9):1809–1820, 2012

    Bo Zhao, Justin P Haldar, Anthony G Christodoulou, and Zhi-Pei Liang. Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints.IEEE transactions on medical imaging, 31(9):1809–1820, 2012

    2012

  18. [26]

    Cand `es, and Daniel K

    Ricardo Otazo, Emmanuel J. Cand `es, and Daniel K. Sodickson. Low- rank plus sparse matrix decomposition for accelerated dynamic mri with separation of background and dynamic components.Magnetic Resonance in Medicine, 73, 2015

    2015

  19. [27]

    Deep low-rank plus sparse network for dynamic mr imaging.Medical image analysis, 73:102190, 2020

    Wenqi Huang, Ziwen Ke, Zhuoxu Cui, Jing Cheng, Zhilang Qiu, Sen Jia, Yanjie Zhu, and Dong Liang. Deep low-rank plus sparse network for dynamic mr imaging.Medical image analysis, 73:102190, 2020

    2020

  20. [28]

    Liyue Shen, John Pauly, and Lei Xing. Nerp: implicit neural repre- sentation learning with prior embedding for sparsely sampled image reconstruction.IEEE Transactions on Neural Networks and Learning Systems, 35(1):770–782, 2022

    2022

  21. [29]

    Naf: neural attenuation fields for sparse-view cbct reconstruction

    Ruyi Zha, Yanhao Zhang, and Hongdong Li. Naf: neural attenuation fields for sparse-view cbct reconstruction. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 442–452. Springer, 2022

    2022

  22. [30]

    Learning deep intensity field for extremely sparse-view cbct reconstruction

    Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li. Learning deep intensity field for extremely sparse-view cbct reconstruction. In International Conference on Medical Image Computing and Computer- Assisted Intervention, pages 13–23. Springer, 2023

    2023

  23. [31]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  24. [32]

    Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information proce...

    2020

  25. [33]

    Instant neural graphics primitives with a multiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41:1 – 15, 2022

    2022

  26. [34]

    Hamprecht, Yoshua Bengio, and Aaron C

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dr ¨axler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. InInternational Conference on Machine Learning, 2018

    2018

  27. [35]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2014

    2014 arXiv

  28. [36]

    Zhichao Liang, Shuangyang Zhang, Zongxin Mo, Xiaoming Zhang, Anqi Wei, Wufan Chen, and Li Qi. Organ-level instance segmen- tation enables continuous time-space-spectrum analysis of pre-clinical abdominal photoacoustic tomography images.Medical Image Analysis, 101:103402, 2025

    2025

  29. [37]

    Sheikh, and Eero P

    Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13:600–612, 2004

    2004

Pith tools

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