Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Endo-NeRF++: Uncertainty-Aware Neural Rendering with Multi-Resolution Hash Encoding for Dynamic Surgical Scene Reconstruction

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 Endo-NeRF++ claims that uncertainty-guided adaptive sampling, multi-resolution hash encoding, and temporal feature blending improve dynamic surgical scene reconstruction, with the adaptive variant beating EndoNeRF by up to 1.22 dB PSNR.

desk verdict A plausible engineering extension of EndoNeRF with real but thin evidence; the adaptive-sampling ablation is internally inconsistent and needs fixing before the central claim can be trusted. read the letter →

arxiv 2607.27825 v2 pith:ADIER2TZ submitted 2026-07-30 eess.IV cs.CV

Gousia Habib , Laura Ruotsalainen This is my paper
classification eess.IVcs.CV
keywords NeuralradiancefieldsDynamicsurgicalscenereconstructionMulti-resolutionhashencodingUncertaintyquantificationAdaptivesamplingTemporalfeatureblendingEndoscopicvideo3D
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

Endo-NeRF++ is an extension of the EndoNeRF neural rendering pipeline for dynamic surgical scenes. It replaces the single MLP with multi-resolution hash-grid encoders, blends temporally sampled features, and reallocates training samples toward regions the model is unsure about. On robotic pulling and cutting tissue video, the adaptive variant reports PSNR 29.537/27.377 dB, SSIM 0.960/0.931, and LPIPS 0.048/0.074, outperforming EndoNeRF with up to 1.22 dB higher PSNR, 5.3% higher SSIM, and 55.1% lower LPIPS. If this holds, the same components could make deformable endoscopic reconstruction more accurate and more temporally consistent without changing the underlying EndoNeRF architecture.

What carries the argument

The mechanism that carries the argument is the per-sample uncertainty score $S_{u,j}^{(sample)} = \tilde{\sigma}^2_{\sigma,j} + \lambda \tilde{\sigma}^2_{c,j}$, where the two variances are min–max normalized along each ray and combine aleatoric variance (predicted log-variances) with epistemic variance (variance across $K$ Monte Carlo dropout passes). After a 5,000-iteration warm-up, 30% of rays are selected by a ray-level opacity-weighted aggregate of this score and re-sampled so that uncertain regions receive more points; multi-resolution hash grids $\{H_i\}$ with temporal blending weights $\beta_t^{(i)}$ supply the features that two MLP heads turn into predicted means and variances for color and density.

What would settle it

Compute per-pixel Pearson and Spearman correlations between predicted uncertainty and absolute color error on held-out frames of the pulling and cutting sequences; if the correlations are near zero or negative, or if reallocating samples with shuffled uncertainty scores produces the same PSNR/SSIM/LPIPS gains, the claim that uncertainty guidance drives the improvement fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the combination of three modifications to EndoNeRF—multi-resolution hash-grid encoding, temporal feature blending, and uncertainty-guided adaptive sampling—improves reconstruction of deformable surgical scenes. The authors show that each modification contributes: hash-grid encoding with temporal blending raises SSIM from 0.912 to 0.949 on the pulling sequence, adding uncertainty estimation preserves the gain while producing error-correlated confidence maps, and the full adaptive version reaches the best overall numbers on both tissue types. The reported margins against the EndoNeRF baseline are up to 1.22 dB PSNR, 5.3% SSIM, and 55.1% LPIPS, with the largest gains on the pulling sequence.

Load-bearing premise

The adaptive-sampling gain rests on the assumption that the per-sample uncertainty score computed from Monte Carlo dropout variances is a reliable guide to where reconstruction error is large; the paper's own correlation tables show only weak-to-moderate agreement (Pearson around 0.20–0.38, Spearman around 0.28), so if that signal does not track error, the reported gains would not be explained by uncertainty guidance.

Editorial extensions

If this is right

  • On the two EndoNeRF sequences, the adaptive variant improves PSNR by up to 1.22 dB, SSIM by up to 5.3%, and LPIPS by up to 55.1% relative to EndoNeRF.
  • Hash-grid encoding plus temporal blending alone accounts for part of the gain, e.g., pulling-sequence SSIM rises from 0.912 to 0.949 before adaptive sampling is added.
  • The uncertainty maps align with surgical tools, tissue borders, and strongly deformed areas, giving a confidence signal for downstream use in surgical scene understanding.
  • The same hyperparameters transfer to the StereoMIS dataset, where the adaptive variant beats EndoNeRF on both tested sequences.
  • The method keeps EndoNeRF's tool-aware canonicalization, depth cueing, and bidirectional mapping intact, so the reported gains come from the three new components rather than a different overall architecture.

Reading between the lines

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

  • A direct consequence the authors do not draw: the same uncertainty-guided sampling rule could be applied to other deformable neural rendering pipelines, concentrating computation on moving boundaries in any dynamic scene, not just endoscopy.
  • Because the paper's uncertainty-error correlations are weak, an alternative explanation for the adaptive-sampling gains is that non-uniform sample allocation alone, rather than the accuracy of the uncertainty estimate, is what helps; comparing against random or shuffled reallocation would separate these.
  • The evaluation is limited to four short sequences, so the reported margins should be read as evidence about those clips; testing on longer multi-view surgical recordings would show whether the gains generalize.
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 Endo-NeRF++, an extension of the EndoNeRF dynamic surgical scene reconstruction method, with three claimed contributions: multi-resolution hash-grid encoding, temporal feature blending, and uncertainty-guided adaptive sampling based on Monte-Carlo dropout variances. Experiments on two endoscopic sequences (pulling and cutting) report that the adaptive variant improves PSNR by up to 1.22 dB, SSIM by up to 5.3%, and LPIPS by up to 55.1% over EndoNeRF, with additional cross-dataset results on StereoMIS. The paper also reports uncertainty calibration metrics (Pearson/Spearman correlation, AUSE) and qualitative comparisons.

Significance. If the reported results are reliable, the proposed combination of efficient hash encoding and uncertainty-aware sampling would be a useful step for deformable surgical scene reconstruction, an application area where existing NeRF-based methods are computationally heavy and rarely quantify uncertainty. The paper is clearly written in terms of the overall architecture, and it makes a reasonable attempt to evaluate uncertainty quality rather than only photometric metrics. However, the experimental support is weakened by a very small proprietary dataset (two sequences, roughly 250 frames, single viewpoint), by apparent inconsistencies between Tables 2, 3, and 4, and by selection of the hash configuration on the test sequences. These issues directly affect the central claim that uncertainty-guided adaptive sampling produces the reported gains.

major comments (4)
  1. [Section 5.5, Tables 2-4] The evidence for the central contribution, uncertainty-guided adaptive sampling, is internally inconsistent. Table 3, titled "Performance of uncertainty-guided adaptive sampling compared with the baseline," reports Pulling PSNR 29.256, SSIM 0.943, LPIPS 0.055 and Cutting PSNR 27.356, SSIM 0.931, LPIPS 0.074, which are identical to the Table 2 results for EndoNeRF++ with uncertainty estimation but without any stated adaptive sampling. In contrast, Table 4's "Unc. EndoNeRF++" row reports Cutting SSIM 0.922 and LPIPS 0.087, values that differ from both Table 2 and Table 3, while its Pulling row matches Tables 2 and 3 exactly. Table 3 also reports Pulling Spearman correlation 0.605 and AUSE 0.317, whereas Table 2 reports Spearman 0.280 and AUSE 0.698 for the same sequence. Because the adaptive-sampling ablation table duplicates or contradicts the other tables, the comparison between uniform and uncertainty-guided sampling is not verifiable from the reported data, and the claimed improvements cannot currently be attributed to the proposed mechanism.
  2. [Section 4.1 and Section 5.2] There is a mismatch between the method description and the implementation. Equations (13)-(14) and the surrounding text define a per-sample uncertainty score that reallocates additional points along each ray proportionally to normalized variance ("Samples with higher uncertainty are allocated additional points proportionally to their normalized uncertainty score"). However, Section 5.2 states that "adaptive ray sampling selects 30% uncertainty-guided rays and 70% uniformly sampled rays," which describes selecting entire rays rather than reallocating samples within rays. The paper therefore does not specify whether the reported results come from the per-sample reallocation derived in Section 4.1 or from a ray-selection heuristic. This ambiguity must be resolved before the adaptive-sampling results can be interpreted.
  3. [Table 2 and Section 5.5] The uncertainty signal used to drive adaptive sampling is weakly correlated with reconstruction error: Table 2 reports Pearson correlations of 0.201 and 0.380 and Spearman correlations of 0.280 and 0.284 for the Pulling and Cutting sequences, respectively. Even the best correlations over trials are modest (Pearson 0.302-0.511, Spearman 0.465-0.475). With such weak correlations, the claimed substantial gains from uncertainty-guided sampling (up to 1.22 dB PSNR, 5.3% SSIM, 55.1% LPIPS) are not mechanistically explained. The paper should report whether the sampling reallocation actually concentrates samples in regions with high error, and should include an ablation that directly compares uniform sampling with uncertainty-guided sampling under otherwise identical settings, rather than relying on the inconsistent tables.
  4. [Section 5.7, Table 5] The multi-resolution hash configuration is selected on the test sequences. Table 5 evaluates several hash configurations on the same Pulling and Cutting sequences, and the authors then state that the configuration "is thus used in all subsequent studies." This is a selection-on-the-test-set procedure: the reported results for the chosen configuration are optimistic, and the generalization claim to new surgical scenes is not supported by the experimental design. The paper should either use a validation split for configuration selection or report results across configurations without presenting the best one as the final model.
minor comments (5)
  1. [General] There are numerous typos and inconsistent naming throughout: "T able" in table captions, "EndoNerf" versus "EndoNeRF", "E-DSSR" versus "DSSR", and "Qualtative Results" in Section 5.8. These should be corrected.
  2. [Equation (18)] The negative log-likelihood loss in Eq. (18) sums over rays and over samples j, but the ground-truth color C_gt is defined per ray. The notation should clarify how the per-sample color loss relates to the final rendered pixel color, since the predictive mean color for each sample is compared with the ray's ground truth.
  3. [Section 5.2] The implementation paragraph lists "K = 4000" in the same sentence as the depth loss weight and exponential moving average coefficient, but K is defined earlier as the number of stochastic forward passes for uncertainty estimation (Eq. 7). Please clarify what K=4000 refers to, and if it is the depth refinement interval, use a distinct symbol.
  4. [References] Reference [31] is listed as "Unknown" and reference [26] appears to duplicate reference [27]; several related works discussed in the text (e.g., EndoSurf, NeRFscopy) are not included in the quantitative comparisons. Please complete the reference list and clearly distinguish the baselines used in experiments from those only discussed.
  5. [Section 6] The limitations paragraph appropriately mentions dataset scarcity and computational cost, but the paper does not discuss the consequences of using a single viewpoint and a proprietary dataset for the validity of the reported improvements. A sentence acknowledging this would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: rendering, uncertainty estimation, and adaptive sampling do not reduce to their own inputs by construction, though the reported tables are internally inconsistent.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs by construction. The rendered color and depth (Eqs. 15–16) depend only on the predictive means (Eq. 8), while the uncertainty scores (Eqs. 13–14, 17) are separate outputs used for sampling allocation; no predicted quantity is defined in terms of the target it is supposed to explain. The uncertainty branch is trained with an NLL loss (Eq. 18) that encourages predicted variances to match the model's own residuals, so a positive uncertainty-error correlation on training data would be partly loss-induced; however, Table 2 reports correlations on test frames, and the weak values (Pearson 0.20–0.38, Spearman about 0.28) show the signal is not forced. The central comparisons are against the external EndoNeRF and E-DSSR baselines (Tables 1 and 4), and the cited EndoNeRF works have no author overlap with the present paper, so no self-citation chain carries the argument. The serious problem in the manuscript is internal inconsistency rather than circularity: Table 3's reconstruction numbers duplicate Table 2 and disagree with the Adap. EndoNeRF++ row of Table 4, so the adaptive-sampling improvement is not verifiable from the reported data, and Section 5.2's implementation note describes ray selection rather than the per-sample reallocation in Eqs. (13)–(14). These are data-integrity and reporting issues, not circular reductions, and per the instructions they do not raise the circularity score.

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

The framework inherits EndoNeRF's deformation and depth assumptions; the paper adds several hand-tuned hyperparameters (hash-grid configuration, uncertainty loss weights, sampling ratio, warm-up length) that are selected on the same evaluation sequences. No new physical entities are postulated.

free parameters (4)
  • Hash grid configuration = resolutions {8,16,32}, scales {1.4,1.7,1.9}, 16 levels, 8 channels
    Selected via Table 5 on the same two test sequences; the chosen config is the best of three tried, so final numbers partly reflect search on the evaluation set.
  • Uncertainty loss weight lambda_unc = 0.02 (cutting), 0.03 (pulling)
    Tuned per sequence; no sensitivity analysis is reported.
  • Adaptive ray sampling fraction = 30% uncertainty-guided, 70% uniform
    Set by hand after warm-up; no ablation across ratios is provided.
  • Warm-up iterations = 5k
    Chosen to let the deformation network stabilize before uncertainty sampling begins; no sensitivity analysis is given.
assumptions (4)
  • domain assumption The deformation network G_phi can map deformed-space points to a consistent canonical space using tool motion encoding w_t from a single viewpoint.
    Inherited from EndoNeRF and central to all rendering and sampling. If canonicalization fails under large tissue deformation, hash-grid features and uncertainty scores lose meaning.
  • domain assumption Monte Carlo dropout variances approximate epistemic uncertainty in this dynamic setting.
    The method relies on dropout at inference to produce uncertainty scores; the paper does not calibrate or validate dropout against a Bayesian alternative.
  • ad hoc to paper Min-max normalization of variances along each ray preserves the relative uncertainty ranking needed for sampling.
    Equation 13 normalizes variances per ray without justification that this ranking is stable across frames or views.
  • domain assumption The STTR-light stereo depth maps and manually annotated tool masks are accurate enough to guide training.
    Depth cues and tool masks are inherited from EndoNeRF; noisy depth or masks would corrupt the deformation field and the uncertainty signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Endo-NeRF++: Uncertainty-Aware Neural Rendering with Multi-Resolution Hash Encoding for Dynamic Surgical Scene Reconstruction." pith.science (2026). https://pith.science/paper/ADIER2TZ

@misc{pith2026260727825,
  author       = {Pith},
  title        = {Pith review of: Endo-NeRF++: Uncertainty-Aware Neural Rendering with Multi-Resolution Hash Encoding for Dynamic Surgical Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADIER2TZ}},
  note         = {Machine review of arXiv:2607.27825}
}
read the original abstract

Reconstructing dynamic surgical scenes is crucial for robot-assisted minimally invasive surgery; however, it continues to be difficult because of tissue deformation, occlusions, specular reflections, and restricted viewpoints. In this study, we introduce Endo-NeRF++, a neural rendering framework that accounts for uncertainty in the reconstruction of dynamic surgical scenes. Expanding on EndoNeRF, the suggested approach incorporates multi-resolution hash-grid encoding, temporal feature merging, and uncertainty-informed adaptive sampling to enhance reconstruction accuracy and temporal coherence in deformable endoscopic scenes.The multi-resolution hash-grid representation within the framework effectively captures both coarse and fine anatomical details, while temporal feature blending ensures stable reconstruction during tissue deformation and surgical tool occlusions. Additionally, uncertainty-driven adaptive sampling assigns more samples to uncertain areas to enhance rendering quality and geometric coherence. Experiments on robotic surgical video sequences demonstrate that the proposed uncertainty-guided adaptive sampling improves PSNR by up to 1.22dB (4.3%), increases SSIM by up to 5.3%, and reduces LPIPS by up to 55.1% compared with the EndoNeRF baseline.

Figures

Figures reproduced from arXiv: 2607.27825 by the authors.

Figure 1
Figure 1. Proposed Framework 4 Method 4.1 Adaptive uncertainty-aware sampling and Multi resolution hash grids We adopt an uncertainty-aware adaptive sampling strategy tailored for deformable surgical scenes. A ray is parameterized as r(s) = o + sd, (1) [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of reconstruction results on representative frames from the tissue pulling and tissue cutting sequences [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of reconstruction results on representative frames from the tissue pulling and tissue cutting sequences. Qualitative Results For both the tissue pulling and tissue cutting sequences, qualitative comparisons are shown in [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Qualitative comparison of EndoNeRF and EndoNeRF++ on representative frames. Uncertainty maps reflect reconstruction errors These findings show that places with higher reconstruction error are con￾sistently highlighted by the anticipated uncertainty, yielding significan…
Figure 3
Figure 3. Figure 3: Qualitative comparison of EndoNeRF and EndoNeRF++ on representative frames. Uncertainty maps reflect reconstruction errors These findings show that places with higher reconstruction error are con￾sistently highlighted by the anticipated uncertainty, yielding significan…
Figure 4
Figure 4. Figure 4: Qualitative comparison of reconstruction results on representative frames from the tissue pulling and tissue cutting sequences. Considering both the quantitative and qualitative results, the findings demon￾strate that uncertainty-guided adaptive sampling efficiently ta…
Figure 4
Figure 4. Figure 4: Qualitative comparison of reconstruction results on representative frames from the tissue pulling and tissue cutting sequences. Considering both the quantitative and qualitative results, the findings demon￾strate that uncertainty-guided adaptive sampling efficiently ta…
Figure 5
Figure 5. Figure 5: Qualitative comparison on the StereoMIS dataset. Compared with EndoNeRF, the proposed EndoNeRF++ produces sharper tissue boundaries, preserves fine anatom￾ical details, and reduces reconstruction artifacts on both P22 and P26 sequences. Vinci Xi system, in order to ass…
Figure 5
Figure 5. Figure 5: Qualitative comparison on the StereoMIS dataset. Compared with EndoNeRF, the proposed EndoNeRF++ produces sharper tissue boundaries, preserves fine anatom￾ical details, and reduces reconstruction artifacts on both P22 and P26 sequences. Vinci Xi system, in order to ass…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages

  1. [1]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Chang, J.R., Chen, Y.S.: Pyramid stereo matching network. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    work page 2018

  2. [3]

    arXiv preprint arXiv:2602.23574 (2026)

    Duan, R., Wong, A.: Evidential neural radiance fields. arXiv preprint arXiv:2602.23574 (2026)

    arXiv 2026

  3. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2018)

    Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (2018)

    work page 2018

  4. [5]

    In: CVPR (2024)

    Goli, L., et al.: Bayes’ rays: Uncertainty quantification for neural radiance fields. In: CVPR (2024)

    work page 2024

  5. [6]

    International Journal of Computer Assisted Radiology and Surgery (2023)

    Hayoz, M., Hahne, C., Gallardo, M., Candinas, D., Kurmann, T., Allan, M., Sznitman, R.: Learning how to robustly estimate camera pose in endoscopic videos. International Journal of Computer Assisted Radiology and Surgery (2023). https://doi.org/10.1007/s11548-023-02919-w

    doi:10.1007/s11548-023-02919-w 2023

  6. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence30(2), 328– 341 (2008)

    Hirschmuller, H.: Stereo processing by semiglobal matching and mutual informa- tion. IEEE Transactions on Pattern Analysis and Machine Intelligence30(2), 328– 341 (2008)

    work page 2008

  7. [8]

    IEEE Robotics and Automation Letters (2024)

    Kim, J.S., Kim, M., Kim, G., Oh, T.H., Kim, J.H.: Factorized multi-resolution hashgrid for efficient neural radiance fields: Execution on edge-devices. IEEE Robotics and Automation Letters (2024)

    work page 2024

  8. [9]

    arXiv preprint arXiv:2305.03027 (2023) Endo-NeRF++19

    Kirschstein, T., Qian, S., Giebenhain, S., Walter, T., Nießner, M.: Nersem- ble: Multi-view radiance field reconstruction of human heads. arXiv preprint arXiv:2305.03027 (2023) Endo-NeRF++19

    arXiv 2023

Show all 35 references
  1. [10]

    In: International Conference on 3D Vision (3DV) (2021)

    Lipson, L., Teed, Z., Deng, J.: Raft-stereo: Multilevel recurrent field transforms for stereo matching. In: International Conference on 3D Vision (3DV) (2021)

    2021

  2. [11]

    In: Medical Image Computing and Computer-Assisted Interven- tion (MICCAI)

    Long, Y., Li, Z., Yee, C.H., Ng, C.F., Taylor, R.H., Unberath, M., Dou, Q.: E-dssr: Efficientdynamicsurgicalscenereconstructionwithtransformer-basedstereoscopic depth perception. In: Medical Image Computing and Computer-Assisted Interven- tion (MICCAI). pp. 415–425. Springer (2021)

    2021

  3. [12]

    arXiv preprint arXiv:2505.03042 (2025)

    Luo, S.T.S.: A new perspective to understanding multi-resolution hash encoding for neural fields. arXiv preprint arXiv:2505.03042 (2025)

    2025 arXiv

  4. [13]

    Na- ture Biomedical Engineering1, 691–696 (2017)

    Maier-Hein, L., et al.: Surgical data science for next-generation interventions. Na- ture Biomedical Engineering1, 691–696 (2017)

    2017

  5. [14]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    2021

  6. [15]

    In: ECCV (2020)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)

    2020

  7. [16]

    IEEE Transactions on Medical Imaging29(11), 1876–1888 (2010)

    Mountney, P., Stoyanov, D., Davison, A.J., Yang, G.Z.: Three-dimensional tissue deformation recovery and tracking. IEEE Transactions on Medical Imaging29(11), 1876–1888 (2010)

    2010

  8. [17]

    ACM Transactions on Graphics (SIGGRAPH) (2022)

    Müller,T.,Evans,A.,Schied,C.,Keller,A.:Instantneuralgraphicsprimitiveswith a multiresolution hash encoding. ACM Transactions on Graphics (SIGGRAPH) (2022)

    2022

  9. [18]

    In: ISMAR (2011)

    Newcombe, R.A., et al.: Kinectfusion: Real-time dense surface mapping and track- ing. In: ISMAR (2011)

    2011

  10. [19]

    Surgical Oncology20(3), 189–201 (2011)

    Nicolau, S., Soler, L., Mutter, D., Marescaux, J.: Augmented reality in laparoscopic surgical oncology. Surgical Oncology20(3), 189–201 (2011)

    2011

  11. [20]

    Cureus17(3) (2025)

    Osman, E.I.A., Ismail, M.M.E.M., Mukhtar, M.A.H., Ahmed, A.U.B., Mohamed, N.A.A.E., Ibrahim, A.A.A., Mohamed, N.A.A., Ibrahim, A.A.: Artificial intelli- gence and robotics in minimally invasive and complex surgical procedures: a sys- tematic review. Cureus17(3) (2025)

    2025

  12. [21]

    Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin- Brualla,R.:Nerfies:Deformableneuralradiancefields.In:InternationalConference on Computer Vision (ICCV) (2021)

    2021

  13. [22]

    Journal of Plastic, Reconstructive & Aesthetic Surgery71(3), 456–462 (2018)

    Pratt, P., Ives, M., Lawton, G., Simmons, J., Radev, N., Spyropoulou, L., Amiras, D.: Through the hololens looking glass: augmented reality for extremity recon- struction surgery. Journal of Plastic, Reconstructive & Aesthetic Surgery71(3), 456–462 (2018)

    2018

  14. [23]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural ra- diance fields for dynamic scenes. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

    2021

  15. [24]

    Journal of Robotic Surgery20(1), 186 (2026)

    Raptis,S.P.,Theocharopoulos,A.,Theocharopoulos,C.,Papadakos,S.P.,Levantis, G., Kontis, E., Vrahatis, A.G.: Artificial intelligence analysis of minimally invasive surgery data. Journal of Robotic Surgery20(1), 186 (2026)

    2026

  16. [25]

    In: WACV (2026)

    Safadoust, S., et al.: Warprf: Multi-view consistency for training-free uncertainty quantification. In: WACV (2026)

    2026

  17. [27]

    arXiv preprint arXiv:2602.15775 (2026) 20 Habib et al

    Salort-Benejam, L., Agudo, A.: Nerfscopy: Neural radiance fields for in-vivo time- varying tissues from endoscopy. arXiv preprint arXiv:2602.15775 (2026) 20 Habib et al

    2026

  18. [28]

    International Journal of Computer Vision47(1), 7–42 (2002)

    Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense stereo correspon- dence algorithms. International Journal of Computer Vision47(1), 7–42 (2002)

    2002

  19. [29]

    In: Proceedings of CVPR (2016)

    Schönberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of CVPR (2016)

    2016

  20. [30]

    arXiv preprint arXiv:2209.08718 (2022)

    Sünderhauf, N., Abou-Chakra, J., Miller, D.: Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields. arXiv preprint arXiv:2209.08718 (2022)

    2022 arXiv

  21. [31]

    arXiv preprint (2025)

    Unknown: Endo2dgs: Endoscopic scene reconstruction with high-fidelity geometry. arXiv preprint (2025)

    2025

  22. [32]

    In: Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Wang, Y., Long, Y., Fan, S., Dou, Q.: Neural rendering for stereo 3d reconstruc- tion of deformable tissues in robotic surgery. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 431–441. Springer (2022)

    2022

  23. [33]

    arXiv preprint arXiv:2404.15339 (2024)

    Wang, Y., Gong, B., Long, Y., Fan, S.H., Dou, Q.: Efficient endonerf recon- struction and its application for data-driven surgical simulation. arXiv preprint arXiv:2404.15339 (2024)

    2024 arXiv

  24. [34]

    In: RSS (2015)

    Whelan, T., et al.: Elasticfusion: Dense slam without a pose graph. In: RSS (2015)

    2015

  25. [35]

    In: Conference on Com- puter Vision and Pattern Recognition (CVPR) (2023)

    Yu, A., et al.: Uncertainty-aware neural radiance fields. In: Conference on Com- puter Vision and Pattern Recognition (CVPR) (2023)

    2023

  26. [36]

    Zha, R., Cheng, X., Li, H., Harandi, M., Ge, Z.: Endosurf: Neural surface re- construction of deformable tissues with stereo endoscope videos (2023).https: //doi.org/10.48550/arXiv.2307.11307,https://arxiv.org/abs/2307.11307

    2023 doi arXiv

  27. [37]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    Zhang, F., Prisacariu, V., Yang, R., Torr, P.H.: Ga-net: Guided aggregation net for end-to-end stereo matching. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    2019

Pith tools

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