Pith. sign in

REVIEW 4 major objections 5 minor 35 references

D2IP: Deep Dynamic Image Prior for 3D Time-sequence Pulmonary Impedance Imaging

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Deep Dynamic Image Prior (D2IP) claims that a single untrained 3D convolutional network, warm-started and propagated across frames, can reconstruct 3D time-sequence pulmonary electrical impedance tomography faster and more accurately than…

desk verdict Plausible acceleration of DIP for 3D time-sequence EIT, but the headline accuracy numbers rest on test-set hyperparameter tuning and need a held-out validation. read the letter →

arxiv 2507.14046 v1 pith:TNBVISW3 submitted 2025-07-18 eess.IV cs.CVcs.LG

Hao Fang , Hao Yu , Sihao Teng , Tao Zhang , Siyi Yuan , Huaiwu He , Zhe Liu , Yunjie Yang This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords deepimagepriorelectricalimpedancetomography3Dtime-sequenceimagingunsupervisedreconstructiontemporalparameterpropagation4Dtotalvariationpulmonarymonitoringinverseproblems
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 is trying to show that 3D time-sequence electrical impedance tomography can be solved without any labeled training data and without the usual thousands of per-frame network iterations. It proposes Deep Dynamic Image Prior (D2IP), in which the unknown 3D conductivity change of each frame is the output of one untrained 3D convolutional network. Two parameter-transfer strategies, reusing optimized weights from an earlier task and from the previous frame, shorten convergence to roughly 250 iterations for each frame after a warm start, while a lightweight network and 4D spatio-temporal total variation keep cost and noise down. On two simulated breathing sequences and two clinical participants, D2IP reports higher mean structural similarity (24.8% average MSSIM increase), lower error (8.1% ERR reduction), and 7.1 times less computation than the compared baselines. If correct, this would make continuous bedside lung monitoring with EIT practical in settings where training data are scarce or ethically difficult to collect.

What carries the argument

The load-bearing mechanism is the pair of weight-transfer rules, Unsupervised Parameter Warm-Start (UPWS) and Temporal Parameter Propagation (TPP), applied to an untrained 3D convolutional generator called 3D-FastResUNet. UPWS first optimizes the network on an arbitrary single-frame reconstruction task and transfers those final weights to initialize the first time-sequence frame; TPP then initializes each frame from the optimized weights of the previous frame, so the fixed random input stays constant and only the weights travel along the sequence. 3D-FastResUNet uses depthwise separable 3D convolutions, residual blocks, squeeze-and-excitation units, and an atrous spatial pyramid pooling module to keep the forward pass cheap, and a 4D total-variation regularizer explicitly couples spatial smoothness with temporally decaying differences to earlier frames. The iteration counts 1800:450:250 encode the design: a longer warm start, a moderate first-frame refinement, and short subsequent refinements.

What would settle it

Run the same reconstruction protocol on an experimental phantom with known conductivity contrasts, or on simulated data from a nonlinear forward solver, and compare D2IP's images with the known truth; if the error ratio or structural similarity gains over the baselines disappear in the high-contrast frames, the linearized forward model underlying the data-fidelity term is the failing assumption.

Watch

Extended reading notes

Core claim

The central claim is that the bottleneck of unsupervised deep-image-prior reconstruction is not the prior itself but the repeated cost of optimizing an untrained network from scratch for every frame; once the same random input and network are reused, each subsequent frame can be initialized from the previous optimized weights and converged in far fewer iterations without losing temporal coherence. D2IP represents each conductivity change as the output of 3D-FastResUNet applied to a fixed random input, and minimizes a linearized data-fidelity term plus a 4D total-variation term that couples the current frame to earlier frames through exponentially decaying weights. The reported iteration budget is 1800 for the upstream warm-start phase, 450 for the first frame, and 250 for every later frame, which yields the claimed quality and speed gains over quadratic regularization, total-variation regularization, and a shallow-image-prior baseline on simulated and clinical pulmonary data.

Load-bearing premise

The method assumes the linearized forward model V = JΣ (where J is the sensitivity matrix) is accurate enough for the 3D time-sequence task; if J fails for large conductivity contrasts or for voltage references that differ from the one used to build J, the data-fidelity loss is wrong and the reconstructed images are biased.

Editorial extensions

If this is right

  • 3D time-sequence EIT can be reconstructed without labeled training data or per-frame network re-initialization, which is what makes bedside pulmonary monitoring feasible.
  • Starting each frame from the previous frame's optimized weights enforces temporal coherence as a byproduct of fast convergence, rather than through an explicit motion model.
  • The reported quality gains (24.8% higher average MSSIM, 8.1% lower ERR) and the 7.1 times speedup are demonstrated on two simulated breathing scenarios and two clinical participants.
  • Reconstruction cost scales with the number of frames but not with the amount of training data, since no dataset is needed at all.

Reading between the lines

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

  • Not claimed in the paper: the same warm-start plus frame-to-frame weight propagation should transfer to other dynamic inverse problems with linear measurements, such as dynamic CT or PET, because the mechanism only assumes adjacent frames are close in solution space.
  • Not claimed in the paper: the exponential temporal decay in the 4D-TV term may smooth out brief ventilation transients (for example, a cough or rapid derecruitment), so testing on non-sinusoidal breathing waveforms would reveal the limits of the temporal prior.
  • Not claimed in the paper: the 1800-iteration warm start is a fixed overhead; reusing a warm start across longer sequences or across patients under the same electrode protocol could cut the per-sequence cost further, but the paper only demonstrates reuse within one protocol.
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 / 5 minor

Summary. The paper proposes D2IP, an unsupervised deep image prior framework for 3D time-sequence electrical impedance tomography (tsEIT). Three components are introduced: Unsupervised Parameter Warm-Start (UPWS), Temporal Parameter Propagation (TPP), and a lightweight 3D-FastResUNet backbone, together with 4D total variation regularization. The reconstruction is formulated as a per-frame optimization of network parameters against measured voltages and a spatio-temporal regularizer, with parameters warm-started and propagated across frames. Experiments on two simulated breathing sequences and two clinical patients show faster convergence and improved reconstruction quality relative to Tikhonov, TV, and R-SIP, including a reported 24.8% increase in average MSSIM and 8.1% reduction in ERR, and 7.1x faster reconstruction over 20 frames.

Significance. If the reported gains hold, D2IP would be a useful contribution to dynamic EIT reconstruction: it is unsupervised, does not require labeled training data, and addresses the computational bottleneck of standard DIP methods for 3D time-sequence data. The paper includes a clear architectural and algorithmic description, an ablation study of the acceleration strategies, convergence analysis under noise, and a clinical feasibility study with IRB approval and informed consent, which are strengths. The central claim, however, depends on simulation-based quantitative comparisons whose reliability is weakened by the hyperparameter selection procedure and the absence of statistical error bars.

major comments (4)
  1. [§IV.D and §V.A.2] The quantitative claim of 24.8% higher MSSIM and 8.1% lower ERR is not supported by a valid evaluation protocol. Section IV.D states that the D2IP parameters (λ, λs, λt, learning rate, iteration counts) are set based on convergence curve analysis, quantitative evaluation metrics, and trial-and-error, and the quantitative metrics used in that tuning (CC, PSNR, MSSIM, ERR) are exactly the metrics reported in Fig. 6b on the same simulation data. The Tikhonov baseline is tuned over only a narrow range (0.001–0.01) on the same test data. This asymmetric and test-set-based tuning can inflate the apparent advantage of D2IP. Please add a held-out validation protocol (e.g., tune on one simulated case or on a subset of frames and report on the remaining data), and report the hyperparameter values chosen before evaluating the final model.
  2. [§V.A.2, Fig. 6b] No error bars, confidence intervals, or repeated trials are reported for any of the quantitative metrics. DIP-based reconstruction is stochastic because of the random input tensor Z and random network initialization, so a single run could give results that are not representative. Please run the experiments multiple times with different random seeds and noise realizations, report mean ± standard deviation for CC, PSNR, MSSIM, and ERR, and perform a statistical significance test (e.g., paired test) for the claimed improvements before the headline numbers can be accepted.
  3. [§II, Eq. (1); §IV.A] The linearized forward model V = JΣ is used as the data fidelity term without any validation of its accuracy for the simulated scenarios. Case 1 involves a large relative conductivity change (from 0.20 S/m to 0.105 S/m, roughly a 47% reduction), where the linear approximation can become inaccurate, and the sensitivity matrix J is computed at a single reference state. Please quantify the linearization error by comparing the voltage predictions of J against the full COMSOL nonlinear forward model over the simulated breathing cycle, or alternatively reconstruct with a nonlinear forward model and check whether the reported qualitative and quantitative conclusions are preserved.
  4. [§V.B, Fig. 9] The clinical evaluation is purely qualitative, based on two patients, and the statement that D2IP is 'consistent with CT findings' is made through visual inspection without any quantitative measure. Given that the abstract claims accurate reconstruction on clinical pulmonary datasets, the current evidence is insufficient. Please either temper the clinical claim to 'qualitative feasibility in two patients' or provide quantitative comparison to CT (e.g., lung region overlap, ventilation-weighted intensity correlation, or a blinded radiologist scoring).
minor comments (5)
  1. [§II and §III.A] The symbol V is used for two different objects: in Eq. (1) it is the voltage measurement matrix of size M×T, while in Eq. (3) it is the reshape operator V(·). Please use separate notation, e.g., U(·) for the reshape operator, to avoid ambiguity.
  2. [§III.E, Eq. (13)] The normalization factor in Eq. (13) is typeset as '1/Pi−1 j=1 αj,i', which is unclear; it should be written as 1 / (∑_{j=1}^{i−1} α_{j,i}) or a reciprocal with proper parentheses.
  3. [§IV.A] The text says each breathing sequence is simulated over 20 frames, but Case 2 yields 19 frames because the first frame serves as the reference. Please state this explicitly in the data description to avoid confusion.
  4. [§III.E and §IV.D] In UPWS (Eq. (6)), the regularization term R is applied to a single frame, yet the definition of R4D-TV in Eq. (11) includes temporal TV across previous frames; clarify which regularization is actually used in the warm-start phase.
  5. [References] Reference [35] cites He et al. for atrous spatial pyramid pooling, but ASPP is more commonly attributed to Chen et al. (DeepLab). Please check and correct the citation to the appropriate source.

Circularity Check

1 steps flagged · score 4.0 of 10

The reported accuracy gains are partially selected by hyperparameter tuning on the same quantitative metrics used for evaluation, while the reconstruction objective itself is not circular.

  1. fitted input called prediction [Section IV.D "Parameter Settings"; reported in Section V.A.2 "Qualitative and Quantitative Reconstruction Results"]
    "The parameter settings for D2IP are determined based on convergence curve analysis, quantitative evaluation metrics, and empirical tuning through trial-and-error."

    The same quantitative metrics that define the paper's headline gains (CC, PSNR, MSSIM, ERR, as reported in Fig. 6b and averaged in V.A.2) are explicitly used to select D2IP's hyperparameters, including lambda, lambda_s, lambda_t, learning rate, and iteration counts. The subsequent claim of a 24.8% increase in average MSSIM and an 8.1% reduction in ERR is therefore not an independent evaluation of a fixed method; it is partly a selection result. The reconstruction objective itself still minimizes voltage fidelity plus 4D-TV rather than directly fitting ground truth, so the method is not entirely forced, but the quantitative superiority claim is statistically coupled to the tuning procedure.

full rationale

The derivation chain of D2IP is self-contained: the conductivity is represented as the output of an untrained 3D-FastResUNet, and Eq. (4) minimizes a voltage data-fidelity term plus 4D-TV regularization, without using ground-truth conductivity in training. This is not self-definitional, and the simulation ground truth provides an external benchmark for the reconstruction quality. The self-citations to prior work, such as R-SIP [25] and MAIP [23], are used as baselines or architectural starting points, not as load-bearing uniqueness theorems, so they do not by themselves constitute circularity. The one significant circular feature is the evaluation protocol: Section IV.D states that D2IP's parameter settings were determined using quantitative evaluation metrics, and Section V.A.2 uses those same metrics (CC, PSNR, MSSIM, ERR) to report the method's superiority. This makes the headline accuracy numbers in-sample and inflated, though the method retains independent content in its voltage-fidelity optimization and in the inference-time advantage. The clinical data are qualitative and do not support quantitative claims, so they do not add circularity. Overall, the central reconstruction derivation is not circular, but the reported quantitative performance claim is partially fitted to the evaluation metric.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a linear forward model, the empirical DIP property, and a set of five hyperparameters tuned on the evaluation data. No new physical entities are introduced.

free parameters (5)
  • lambda_TV = 0.002
    Weight for the combined 4D-TV term in the loss (Eq. 14); tuned by trial-and-error on simulation data.
  • lambda_s = 1
    Spatial weighting factor in the 4D-TV decomposition (Eq. 11).
  • lambda_t = 0.1
    Temporal weighting factor in the 4D-TV decomposition (Eq. 11).
  • learning_rate = 5e-4 (simulation), 1e-4 (clinical)
    Adam learning rate; chosen from convergence curve analysis in Fig. 5a.
  • iteration_counts = 1800 (warm-start), 450 (first frame), 250 (subsequent frames)
    Set from convergence plateaus in Fig. 5b, with margin to avoid overfitting.
assumptions (3)
  • domain assumption Linearized EIT forward model V = JΣ
    The reconstruction (Eq. 1) treats EIT as linear with a fixed sensitivity matrix J; this is a standard approximation in difference EIT but is not validated here.
  • domain assumption Deep Image Prior implicit regularization
    The method relies on the empirical property that optimizing an untrained network with a data-fidelity loss yields plausible images; this is an observed phenomenon, not a proven theorem.
  • domain assumption Hyperparameter transfer from simulation to clinical data
    The same lambda values and learning rate schedule are used for patient data, assuming the simulation is representative of clinical conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of D2IP: Deep Dynamic Image Prior for 3D Time-sequence Pulmonary Impedance Imaging." pith.science (2026). https://pith.science/paper/TNBVISW3

@misc{pith2026250714046,
  author       = {Pith},
  title        = {Pith review of: D2IP: Deep Dynamic Image Prior for 3D Time-sequence Pulmonary Impedance Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TNBVISW3}},
  note         = {Machine review of arXiv:2507.14046}
}
read the original abstract

Unsupervised learning methods, such as Deep Image Prior (DIP), have shown great potential in tomographic imaging due to their training-data-free nature and high generalization capability. However, their reliance on numerous network parameter iterations results in high computational costs, limiting their practical application, particularly in complex 3D or time-sequence tomographic imaging tasks. To overcome these challenges, we propose Deep Dynamic Image Prior (D2IP), a novel framework for 3D time-sequence imaging. D2IP introduces three key strategies - Unsupervised Parameter Warm-Start (UPWS), Temporal Parameter Propagation (TPP), and a customized lightweight reconstruction backbone, 3D-FastResUNet - to accelerate convergence, enforce temporal coherence, and improve computational efficiency. Experimental results on both simulated and clinical pulmonary datasets demonstrate that D2IP enables fast and accurate 3D time-sequence Electrical Impedance Tomography (tsEIT) reconstruction. Compared to state-of-the-art baselines, D2IP delivers superior image quality, with a 24.8% increase in average MSSIM and an 8.1% reduction in ERR, alongside significantly reduced computational time (7.1x faster), highlighting its promise for clinical dynamic pulmonary imaging.

Figures

Figures reproduced from arXiv: 2507.14046 by the authors.

Figure 1
Figure 1. Geometric interpretation of different reconstruction strategies in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed D2 IP framework. (a) DIP-based op￾timization using the 3D-FastResUNet with depthwise separable convolutions [32]. (b) UPWS first initializes the 3D-FastResUNet parameters using Kaiming initialization, followed by an iterative optimization process on an arbitrary single-frame reconstruction task. (c) TPP propagates learned parameters across sequential frames, ensuring temporal consistency… view at source ↗
Figure 3
Figure 3. The architecture of the proposed 3D-FastResUNet. C. Temporal Parameter Propagation To further accelerate convergence and enforce temporal con￾sistency across sequential frames, we introduce the Temporal Parameter Propagation (TPP) strategy. Instead of optimizing each frame independently, TPP transfers the optimized param￾eters of the previous frame as initialization for the next frame. This strategy leverages tempor… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of the two simulated lung expansion scenarios: (a) Lung [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Convergence analysis of D2 IP under different settings. (a) Effect of learning rate on D2 IP convergence. (b) Ablation study of acceleration strategies including UPWS and TPP. (c) Convergence performance of D2 IP under different levels of input noise (10–60 dB). Each b…
Figure 6
Figure 6. Figure 6: Qualitative and quantitative comparison of different reconstruction methods on Case 1. (a) Visual comparison of reconstructed frames (frames 1, 10, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of reconstructed frames (frames 1, 10) using [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of accumulated inference time across 20 frames for [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Real-world clinical experiment setup and reconstruction results. (a) Axial CT images of two participants (viewed from the feet upwards). (b) Top: the [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages

  1. [1]

    Pulmonary functional imaging: Part 1—state-of-the-art technical and physiologic underpinnings,

    Y . Ohno, J. B. Seo, G. Parraga, K. S. Lee, W. B. Gefter, S. B. Fain, M. L. Schiebler, and H. Hatabu, “Pulmonary functional imaging: Part 1—state-of-the-art technical and physiologic underpinnings,” Radiology, vol. 299, no. 3, pp. 508–523, 2021

    work page 2021

  2. [2]

    Pulmonary functional imaging: part 2—state-of-the- art clinical applications and opportunities for improved patient care,

    W. B. Gefter, K. S. Lee, M. L. Schiebler, G. Parraga, J. B. Seo, Y . Ohno, and H. Hatabu, “Pulmonary functional imaging: part 2—state-of-the- art clinical applications and opportunities for improved patient care,” Radiology, vol. 299, no. 3, pp. 524–538, 2021

    work page 2021

  3. [3]

    Deep learning to quantify pulmonary edema in chest radiographs,

    S. Horng, R. Liao, X. Wang, S. Dalal, P. Golland, and S. J. Berkowitz, “Deep learning to quantify pulmonary edema in chest radiographs,” Radiology: Artificial Intelligence , vol. 3, no. 2, p. e190228, 2021

    work page 2021

  4. [4]

    Imaging the acute respiratory distress syndrome: past, present and future,

    L. Bitker, D. Talmor, and J.-C. Richard, “Imaging the acute respiratory distress syndrome: past, present and future,” Intensive Care Medicine , vol. 48, no. 8, pp. 995–1008, 2022

    work page 2022

  5. [5]

    Imaging advances in chronic obstructive pulmonary disease. insights from the genetic epidemiology of chronic obstructive pulmonary disease (copdgene) study,

    S. P. Bhatt, G. R. Washko, E. A. Hoffman, J. D. Newell Jr, S. Bodduluri, A. A. Diaz, C. J. Galban, E. K. Silverman, R. San Jos´e Est´epar, and D. A. Lynch, “Imaging advances in chronic obstructive pulmonary disease. insights from the genetic epidemiology of chronic obstructive pulmonary disease (copdgene) study,” American journal of respiratory and critic...

    work page 2019

  6. [6]

    Computer-aided diagnosis of pulmonary fibrosis using deep learning and ct images,

    A. Christe, A. A. Peters, D. Drakopoulos, J. T. Heverhagen, T. Geiser, T. Stathopoulou, S. Christodoulidis, M. Anthimopoulos, S. G. Mougiakakou, and L. Ebner, “Computer-aided diagnosis of pulmonary fibrosis using deep learning and ct images,” Investigative radiology , vol. 54, no. 10, pp. 627–632, 2019

    work page 2019

  7. [7]

    Expanding applications of pulmonary mri in the clinical evaluation of lung disor- ders: Fleischner society position paper,

    H. Hatabu, Y . Ohno, W. B. Gefter, G. Parraga, B. Madore, K. S. Lee, T. A. Altes, D. A. Lynch, J. R. Mayo, J. B. Seo et al. , “Expanding applications of pulmonary mri in the clinical evaluation of lung disor- ders: Fleischner society position paper,” Radiology, vol. 297, no. 2, pp. 286–301, 2020

    work page 2020

  8. [8]

    Overview of mri for pulmonary functional imaging,

    Y . Ohno, S. Hanamatsu, Y . Obama, T. Ueda, H. Ikeda, H. Hattori, K. Murayama, and H. Toyama, “Overview of mri for pulmonary functional imaging,” The British Journal of Radiology, vol. 95, no. 1132, p. 20201053, 2022

    work page 2022

Show all 35 references
  1. [9]

    Regional lung perfusion as determined by electrical impedance tomography in comparison with electron beam ct imaging,

    I. Frerichs, J. Hinz, P. Herrmann, G. Weisser, G. Hahn, M. Quintel, and G. Hellige, “Regional lung perfusion as determined by electrical impedance tomography in comparison with electron beam ct imaging,” IEEE transactions on medical imaging , vol. 21, no. 6, pp. 646–652, 2002

    2002

  2. [10]

    Eichler, J

    L. Eichler, J. Mueller, J. Grensemann, I. Frerichs, C. Z ¨ollner, and S. Kluge, “Lung aeration and ventilation after percutaneous tracheotomy measured by electrical impedance tomography in non-hypoxemic criti- cally ill patients: a prospective observational study,” Annals of i...

    2018

  3. [11]

    Supervised descent learning for thoracic electrical impedance tomography,

    K. Zhang, R. Guo, M. Li, F. Yang, S. Xu, and A. Abubakar, “Supervised descent learning for thoracic electrical impedance tomography,” IEEE Transactions on Biomedical Engineering, vol. 68, no. 4, pp. 1360–1369, 2020

    2020

  4. [12]

    Early screening of lung function by electrical impedance tomography in people with normal spirometry reveals unrecognized pathological features,

    S. Qu, E. Feng, D. Dong, L. Yang, M. Dai, I. Frerichs, S. Liu, Y . Gao, J. Zheng, L. Song et al. , “Early screening of lung function by electrical impedance tomography in people with normal spirometry reveals unrecognized pathological features,” Nature Communications , vol. 16...

    2025

  5. [13]

    Multi-modal eit imaging using lensfree and flexible impedance sensor,

    H. Fang, R. B. Liu, J. Sun, P. O. Bagnaninchi, Z. Liu, and Y . Yang, “Multi-modal eit imaging using lensfree and flexible impedance sensor,” in 2024 IEEE SENSORS . IEEE, 2024, pp. 1–4

    2024

  6. [14]

    Three- dimensional electrical impedance tomography,

    P. Metherall, D. C. Barber, R. H. Smallwood, and B. H. Brown, “Three- dimensional electrical impedance tomography,” Nature, vol. 380, no. 6574, pp. 509–512, 1996

    1996

  7. [15]

    Tikhonov regularization and prior information in electrical impedance tomography,

    M. Vauhkonen, D. Vad ´asz, P. A. Karjalainen, E. Somersalo, and J. P. Kaipio, “Tikhonov regularization and prior information in electrical impedance tomography,”IEEE transactions on medical imaging, vol. 17, no. 2, pp. 285–293, 1998

    1998

  8. [16]

    Adaptive l {p} regu- larization for electrical impedance tomography,

    J. Li, S. Yue, M. Ding, Z. Cui, and H. Wang, “Adaptive l {p} regu- larization for electrical impedance tomography,” IEEE Sensors Journal, vol. 19, no. 24, pp. 12 297–12 305, 2019

    2019

  9. [17]

    In vivo impedance imaging with total variation regularization,

    A. Borsic, B. M. Graham, A. Adler, and W. R. Lionheart, “In vivo impedance imaging with total variation regularization,” IEEE transac- tions on medical imaging , vol. 29, no. 1, pp. 44–54, 2009

    2009

  10. [18]

    Deep d-bar: Real-time electrical impedance tomography imaging with deep neural networks,

    S. J. Hamilton and A. Hauptmann, “Deep d-bar: Real-time electrical impedance tomography imaging with deep neural networks,” IEEE transactions on medical imaging , vol. 37, no. 10, pp. 2367–2377, 2018

    2018

  11. [19]

    Image reconstruction based on convolutional neural network for electrical resistance tomography,

    C. Tan, S. Lv, F. Dong, and M. Takei, “Image reconstruction based on convolutional neural network for electrical resistance tomography,” IEEE Sensors Journal , vol. 19, no. 1, pp. 196–204, 2018

    2018

  12. [20]

    Electrical resistance tomography image reconstruction with densely connected convolutional neural network,

    F. Li, C. Tan, and F. Dong, “Electrical resistance tomography image reconstruction with densely connected convolutional neural network,” IEEE Transactions on Instrumentation and Measurement , vol. 70, pp. 1–11, 2020

    2020

  13. [21]

    A learning- based method for solving ill-posed nonlinear inverse problems: A simulation study of lung eit,

    J. K. Seo, K. C. Kim, A. Jargal, K. Lee, and B. Harrach, “A learning- based method for solving ill-posed nonlinear inverse problems: A simulation study of lung eit,”SIAM journal on Imaging Sciences, vol. 12, no. 3, pp. 1275–1295, 2019

    2019

  14. [22]

    Electrical resistance tomography with conditional generative adversarial networks,

    Y . Chen, K. Li, and Y . Han, “Electrical resistance tomography with conditional generative adversarial networks,” Measurement Science and Technology, vol. 31, no. 5, p. 055401, 2020

    2020

  15. [23]

    Multi- frequency electrical impedance tomography reconstruction with multi- branch attention image prior,

    H. Fang, Z. Liu, Y . Feng, Z. Qiu, P. Bagnaninchi, and Y . Yang, “Multi- frequency electrical impedance tomography reconstruction with multi- branch attention image prior,” arXiv preprint arXiv:2409.10794 , 2024

    2024 arXiv

  16. [24]

    Deepeit: Deep image prior enabled electrical impedance tomography,

    D. Liu, J. Wang, Q. Shan, D. Smyl, J. Deng, and J. Du, “Deepeit: Deep image prior enabled electrical impedance tomography,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 8, pp. 9627–9638, 2023

    2023

  17. [25]

    Regular- ized shallow image prior for electrical impedance tomography,

    Z. Liu, Z. Chen, H. Fang, Q. Wang, S. Zhang, and Y . Yang, “Regular- ized shallow image prior for electrical impedance tomography,” IEEE Transactions on Instrumentation and Measurement , 2025

    2025

  18. [26]

    A multi-frequency electrical impedance tomog- raphy system for real-time 2d and 3d imaging,

    Y . Yang and J. Jia, “A multi-frequency electrical impedance tomog- raphy system for real-time 2d and 3d imaging,” Review of Scientific Instruments, vol. 88, no. 8, 2017

    2017

  19. [27]

    Time sequence learning for electrical impedance tomography using bayesian spatiotemporal priors,

    S. Liu, R. Cao, Y . Huang, T. Ouypornkochagorn, and J. Jia, “Time sequence learning for electrical impedance tomography using bayesian spatiotemporal priors,” IEEE Transactions on Instrumentation and Mea- surement, vol. 69, no. 9, pp. 6045–6057, 2020

    2020

  20. [28]

    Mmv-net: A multi- ple measurement vector network for multifrequency electrical impedance tomography,

    Z. Chen, J. Xiang, P.-O. Bagnaninchi, and Y . Yang, “Mmv-net: A multi- ple measurement vector network for multifrequency electrical impedance tomography,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 11, pp. 8938–8949, 2022

    2022

  21. [29]

    Multi-path fusion in sfcf-net for enhanced multi-frequency electrical impedance tomography,

    X. Tian, T. Zhang, L. Zhang, X. Liu, F. Fu, X. Shi, C. Xu et al. , “Multi-path fusion in sfcf-net for enhanced multi-frequency electrical impedance tomography,” IEEE Transactions on Medical Imaging, 2024

    2024

  22. [30]

    Dual-modal image reconstruction for electrical impedance tomography with over- lapping group lasso and laplacian regularization,

    Z. Liu, H. Gu, Z. Chen, P. Bagnaninchi, and Y . Yang, “Dual-modal image reconstruction for electrical impedance tomography with over- lapping group lasso and laplacian regularization,” IEEE Transactions on Biomedical Engineering, vol. 70, no. 8, pp. 2362–2373, 2023

    2023

  23. [31]

    Adam: A method for stochastic optimization,

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

    2014 arXiv

  24. [32]

    Xception: Deep learning with depthwise separable convolu- tions,

    F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1251–1258

    2017

  25. [33]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 1026–1034

    2015

  26. [34]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  27. [35]

    Spatial pyramid pooling in deep convolutional networks for visual recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep convolutional networks for visual recognition,” IEEE transactions on pattern analysis and machine intelligence , vol. 37, no. 9, pp. 1904– 1916, 2015

    1904

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.