pith. sign in

arxiv: 2606.26236 · v1 · pith:BCVX6NBHnew · submitted 2026-06-24 · 📡 eess.IV · cs.CV· eess.SP

Rendering Novel Views of MRI Using 3D Gaussian Splatting

Robin Y. Park , Mark C. Eid , Rhydian Windsor , Amir Jamaludin , Ana I.L. Namburete , Jo\~ao F. Henriques , Andrew Zisserman This is my paper

Pith reviewed 2026-06-26 01:04 UTC · model grok-4.3

classification 📡 eess.IV cs.CVeess.SP
keywords Gaussian SplattingMRI resamplingSpinal stenosisNovel view synthesisVolumetric reconstructionAnisotropic MRIRadiological grading
0
0 comments X

The pith

Adapting 3D Gaussian Splatting to sparse spinal MRIs produces resampled views that improve accuracy of stenosis severity gradings.

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

This paper adapts 3D Gaussian Splatting to reconstruct volumes from sparse anisotropic MRIs of the spine. It then renders new imaging planes aligned with the anatomy needed for evaluating stenosis. The resulting images lead to more accurate ordinal grades for stenosis conditions than either the original scans or those resampled by voxel interpolation. Clinicians could use this to extract better diagnostic information from existing low-resolution acquisitions without acquiring new scans.

Core claim

By starting from sparse anisotropic MRIs and using 3D Gaussian Splatting to create a volumetric model, the method allows sampling and rendering of novel view planes that are optimally aligned with the target spinal anatomy. When these resampled scans are used to predict stenosis grades, they yield higher accuracy across conditions than raw scans missing complete in-plane anatomy or scans resampled via inverse-distance weighted voxel interpolation.

What carries the argument

3D Gaussian Splatting adapted for volumetric reconstruction from sparse MRIs, representing the scene as a collection of 3D Gaussians that are rendered into novel aligned 2D views.

If this is right

  • Gaussian Splatting resampling produces higher stenosis grading accuracy than raw anisotropic scans.
  • It also outperforms voxel interpolation resampling across all tested stenosis conditions.
  • The novel views supply complete in-plane anatomy where the original slices do not.
  • Existing sparse MRI data can be repurposed for improved clinical grading without new acquisitions.

Where Pith is reading between the lines

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

  • The approach could be tested on other anisotropic modalities such as CT to check for similar grading gains.
  • If the Gaussians preserve fine intensity detail, the method might incidentally support limited super-resolution in the rendered planes.
  • Deployment would need checks that the splatting step does not systematically shift intensity distributions used in grading.

Load-bearing premise

The Gaussian Splatting reconstruction must produce views whose intensities and geometry match the true anatomy closely enough that grading gains come from better plane alignment rather than from reconstruction artifacts.

What would settle it

Acquire high-resolution isotropic reference scans in the aligned planes and compare stenosis gradings from the Gaussian Splatting renders against those references; if accuracy does not exceed that of the raw scans or voxel interpolation, the claim is false.

Figures

Figures reproduced from arXiv: 2606.26236 by Amir Jamaludin, Ana I.L. Namburete, Andrew Zisserman, Jo\~ao F. Henriques, Mark C. Eid, Rhydian Windsor, Robin Y. Park.

Figure 1
Figure 1. Figure 1: Paired axial scans: Slices from two T2w axial scans of the same patient, with moderate left subarticular stenosis at the L5-S1 intervertebral disc (IVD) level (see [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of stenosis: Axial slices showing severe stenosis at L5-S1. See [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Axial scans intersecting in 2D & 3D: The oblique plane in the 3D plot (right image) shows the mid-slice of the axial volume aligned with the disc (blue lines in left image). The yellow planes are non-aligned slices of a volume that does not follow the angle of the disc, but intersect the intervertebral disc region of interest (green box in left image). In this section, we describe how we apply 3D Gaussian … view at source ↗
Figure 4
Figure 4. Figure 4: Example of axial region of interest: Red lines indicate sagittal slices that intersect the axial slice. SpineNet is used to detect the intervertebral disc (IVD) pertaining to the level in the sagittal scans; the anterior and posterior intersection of the IVD region with the sagittal slices are marked in yellow. The midpoint of the posterior intersection, indicated by the blue dot, is used as a reference po… view at source ↗
Figure 5
Figure 5. Figure 5: 3D Gaussian Splatting-reconstructed example across training epochs: [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stenosis regions: The extrusion of the disc into each demarcated zone leads to compression of nerves and vascular structures in that region, causing stenosis. Image adapted from the Miami Neuroscience Center [28] using Nano Banana 2 [8]. The presence and severity of degenerative spinal conditions are characterised by categorical grading scales ordered by severity [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Stenosis severity gradings: Examples of right subarticular and neural foraminal stenosis gradings. In the RSNA dataset, normal and mild are combined into a single category. The red ellipse indicates the location of the stenosis in moderate and severe images. The subarticular and foraminal zones are described in [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Grading model: We employ the same network described in [32], where each slice in the volume is encoded using 2D Resnet18, then aggregated at volume-level using a global class (CLS) token. The output of the CLS token from the transformer encoder is passed to a linear layer that predicts the severity of each classification task. 5.1 Performance metrics Image quality metrics include PSNR, SSIM, and LPIPS (usi… view at source ↗
Figure 9
Figure 9. Figure 9: Gaussian weight mask: The red box shows the region relevant for stenosis grading. An anisotropic Gaussian weight mask focused on the region of interest is applied to get pixel-wise weights to compute image quality metrics. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Resampled images: A shows the originally aligned slice. B shows an image reconstructed using Voxel Interpolation (R-VI) while C shows a slice reconstructed using Gaussian Splatting (R-GS). The red box shows the region relevant for grading [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Examples of rendered slices: Ground truth images acquired from an aligned-view volume compared against views rendered from a 3D Gaussian Splatting reconstruction from a non￾aligned source volume. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
read the original abstract

The objective of this paper is to improve radiological gradings measured on MRIs of spines, by resampling scans so that the new view planes are better aligned with the target anatomy than the original sparse images. To this end, we adapt 3D Gaussian Splatting to form a volumetric reconstruction starting from sparse anisotropic MRIs, and imaging planes aligned with the anatomy relevant for clinical evaluation are then sampled and rendered. The novel view plane is optimal for diagnostic radiological grading of the target anatomy, whereas the original MRI is not. The resampled scans are then used to predict ordinal severity grades of localised stenosis conditions in spinal MRIs. We compare our method against Voxel Interpolation resampling, which takes the average of inverse-distance weighted nearest neighbour intensities for each target coordinate. Experiments show that across all stenosis conditions, resampled scans using Gaussian Splatting produce more accurate stenosis gradings compared to the raw scans which do not include the complete anatomy in-plane, as well as images resampled using Voxel Interpolation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper adapts 3D Gaussian Splatting to reconstruct a volumetric representation from sparse anisotropic spinal MRIs, then renders novel imaging planes aligned with the target anatomy for stenosis grading. It compares the resulting gradings against those from the original raw scans and from voxel-interpolation resampling, claiming superior accuracy for the Gaussian Splatting approach across all tested stenosis conditions.

Significance. If the empirical claim holds with proper controls and quantitative validation, the work could provide a practical route to obtaining anatomy-aligned views from existing anisotropic acquisitions without additional scanning, which would be relevant for spinal MRI diagnostics where plane misalignment is common.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'resampled scans using Gaussian Splatting produce more accurate stenosis gradings' is asserted without any reported accuracy numbers, dataset size, number of stenosis conditions, statistical tests, or implementation details; this absence makes the empirical result impossible to evaluate or reproduce.
  2. [Abstract] Abstract: no description is given of how stenosis grades were obtained (e.g., reader protocol, number of readers, ground-truth definition), so it is impossible to determine whether the reported improvement is attributable to better plane alignment or to other factors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these comments on the abstract. We agree that the abstract requires additional quantitative and methodological details to allow proper evaluation of the claims, and we will revise it in the next version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'resampled scans using Gaussian Splatting produce more accurate stenosis gradings' is asserted without any reported accuracy numbers, dataset size, number of stenosis conditions, statistical tests, or implementation details; this absence makes the empirical result impossible to evaluate or reproduce.

    Authors: We agree that the abstract should report key quantitative results to support the claim. In the revised manuscript we will expand the abstract to include the observed accuracy improvements across stenosis conditions, the dataset size, the number of conditions evaluated, and references to the statistical tests and implementation details already present in the methods and results sections. revision: yes

  2. Referee: [Abstract] Abstract: no description is given of how stenosis grades were obtained (e.g., reader protocol, number of readers, ground-truth definition), so it is impossible to determine whether the reported improvement is attributable to better plane alignment or to other factors.

    Authors: We acknowledge the absence of this information from the abstract. We will revise the abstract to include a concise summary of the grading process (reader protocol, number of readers, and ground-truth definition) as described in the methods section, so readers can assess whether the improvement is attributable to plane alignment. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an empirical adaptation of 3D Gaussian Splatting for MRI resampling followed by comparative experiments on stenosis grading accuracy. No derivation chain, self-referential equations, fitted parameters presented as predictions, or load-bearing self-citations are present. The central claim rests on observable grading improvements versus raw scans and voxel interpolation, which is externally falsifiable and does not reduce to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no equations, parameters, or modeling assumptions; free_parameters, axioms and invented_entities cannot be identified.

pith-pipeline@v0.9.1-grok · 5735 in / 1050 out tokens · 29406 ms · 2026-06-26T01:04:37.189335+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

44 extracted references · 1 linked inside Pith

  1. [1]

    Y. Bao, T. Ding, J. Huo, Y. Liu, Y. Li, W. Li, Y. Gao, and J. Luo. 3D Gaussian Splatting: Survey, Technologies, Challenges, and Opportunities.IEEE Transactions on Circuits and Systems for Video Technology, 35(7):6832–6852, July 2025

    2025

  2. [2]

    Blondiaux and C

    E. Blondiaux and C. Garel. Fetal cerebral imaging – ultrasound vs. MRI: An update. Acta Radiologica, 54(9):1046–1054, Nov. 2013

    2013

  3. [3]

    Bonilla, S

    S. Bonilla, S. Zhang, D. Psychogyios, D. Stoyanov, F. Vasconcelos, and S. Bano. Gaussian Pancakes: Geometrically-Regularized 3D Gaussian Splatting for Realistic Endoscopic Reconstruction. In M. G. Linguraru, Q. Dou, A. Feragen, S. Gian- narou, B. Glocker, K. Lekadir, and J. A. Schnabel, editors,Medical Image Computing and Computer Assisted Intervention – ...

    2024

  4. [4]

    Y. Cai, Y. Liang, J. Wang, A. Wang, Y. Zhang, X. Yang, Z. Zhou, and A. Yuille. Radiative gaussian splatting for efficient X-ray novel view synthesis. InECCV, pages 283–299. 2024

    2024

  5. [5]

    L. S. de Vries, M. J. N. L. Benders, and F. Groenendaal. Imaging the premature brain: Ultrasound or MRI?Neuroradiology, 55(2):13–22, Sept. 2013

    2013

  6. [6]

    M. C. Eid, A. I. Namburete, and J. F. Henriques. UltraGauss: Ultrafast gaussian reconstruction of 3D ultrasound volumes. InICLR, 2026

    2026

  7. [7]

    B. Fei, J. Xu, R. Zhang, Q. Zhou, W. Yang, and Y. He. 3D Gaussian Splatting as a New Era: A Survey.IEEE Transactions on Visualization and Computer Graphics, 31(8):4429–4449, Aug. 2025

    2025

  8. [8]

    Nano banana 2 generative AI image model

    Google DeepMind. Nano banana 2 generative AI image model. Image generated using the Nano Banana 2 model in Gemini. Accessed: 27 May 2026]

    2026

  9. [9]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks. InICML, pages 1321–1330. PMLR, 2017

    2017

  10. [10]

    J. Guo, J. Wang, D. Kang, W. Dong, W. Wang, and Y.-h. Liu. Free-surgs: Sfm-free 3D gaussian splatting for surgical scene reconstruction. InMICCAI, pages 350–360, 2024

    2024

  11. [11]

    K.-T. Hong, Y. Cho, C. H. Kang, K.-S. Ahn, H. Lee, J. Kim, S. J. Hong, B. H. Kim, and E. Shim. Lumbar spine computed tomography to magnetic resonance imaging synthesis using generative adversarial network: Visual turing test.Diagnostics, 12(2):530, 2022

    2022

  12. [12]

    J. P. Hornak.The basics of MRI. Rochester Institute of Technology, 1996

    1996

  13. [13]

    Huang, L

    Y. Huang, L. Bai, B. Cui, K. Yuan, G. Wang, M. I. Hoque, N. Padoy, N. Navab, and H. Ren. SurgTPGS: Semantic 3D surgical scene understanding with text promptable gaussian splatting. InMICCAI, pages 584–594, 2025

    2025

  14. [14]

    Iddrisu, S

    K. Iddrisu, S. Malec, and A. Crimi. 3d reconstructions of brain from mri scans using neural radiance fields. In L. Rutkowski, R. Scherer, M. Korytkowski, W. Pedrycz, R. Tadeusiewicz, and J. M. Zurada, editors,Artificial Intelligence and Soft Computing, pages 207–218, Cham, 2023. Springer Nature Switzerland

    2023

  15. [15]

    Jamaludin, T

    A. Jamaludin, T. Kadir, and A. Zisserman. SpineNet: automated classification and evidence visualization in spinal MRIs.Medical image analysis, 41:63–73, 2017

    2017

  16. [16]

    Y. Jia, A. Gholipour, Z. He, and S. K. Warfield. A new sparse representation framework for reconstruction of an isotropic high spatial resolution mr volume from orthogonal anisotropic resolution scans.IEEE transactions on medical imaging, 36(5):1182–1193, 2017

    2017

  17. [17]

    Kastryulin, J

    S. Kastryulin, J. Zakirov, N. Pezzotti, and D. V. Dylov. Image quality assessment for magnetic resonance imaging.IEEE Access, 11:14154–14168, 2023. 16 Rendering Novel Views of MRI

    2023

  18. [18]

    J. N. Katz, Z. E. Zimmerman, H. Mass, and M. C. Makhni. Diagnosis and management of lumbar spinal stenosis.JAMA, 327(17):1688–1699, 2022

    2022

  19. [19]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  20. [20]

    K. H. Kim, E.-C. Lee, Y. D. Yoon, D.-W. Shin, H.-W. Koo, and B.-J. Lee. Translation of computed tomography images to T2-weighted magnetic resonance images of lum- bar spine using generative adversarial networks on sagittal images.Scientific Reports, 15(1):18385, 2025

    2025

  21. [21]

    Landauer and K

    F. Landauer and K. Trieb. Diagnostic limitations and aspects of the lumbosacral transitional vertebrae (LSTV).Applied Sciences, 12(21):10830, 2022

    2022

  22. [22]

    S. Lee, J. W. Lee, J. S. Yeom, K.-J. Kim, H.-J. Kim, S. K. Chung, and H. S. Kang. A practical mri grading system for lumbar foraminal stenosis.American Journal of Roentgenology, 194(4):1095–1098, 2010

    2010

  23. [23]

    Lin, R.-M

    S.-I. Lin, R.-M. Lin, and L.-W. Huang. Disability in patients with degenerative lumbar spinal stenosis.Archives of physical medicine and rehabilitation, 87(9):1250–1256, 2006

    2006

  24. [24]

    S. Liu, M. Yang, T. Xing, and R. Yang. A survey of 3D reconstruction: the evolution from multi-view geometry to NeRF and 3DGS.Sensors, 25(18):5748, 2025

    2025

  25. [25]

    J.-T. Lu, S. Pedemonte, B. Bizzo, S. Doyle, K. P. Andriole, M. H. Michalski, R. G. Gonzalez, and S. R. Pomerantz. Deep spine: automated lumbar vertebral segmentation, disc-level designation, and spinal stenosis grading using deep learning. InMLHC, pages 403–419. PMLR, 2018

    2018

  26. [26]

    Marzol, I

    K. Marzol, I. Kolton, W. Smolak-DyĹL’ewska, J. Kaleta, M. Mazur, P. Spurek, et al. MedGS: Gaussian splatting for multi-modal 3D medical imaging.arXiv preprint arXiv:2509.16806, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    J. L. Melancia, A. F. Francisco, and J. L. Antunes. Spinal stenosis.Handbook of clinical neurology, 119:541–549, 2014

    2014

  28. [28]

    Spinal stenosis symptoms, diagnosis & treatment, 2019

    Miami Neuroscience Center. Spinal stenosis symptoms, diagnosis & treatment, 2019. Published: 18 October 2019; last modified: 8 July 2020

    2019

  29. [29]

    Mildenhall, P

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

    2020

  30. [30]

    H. J. Park, S. Kim, S. Lee, N. Park, E. Chung, M. Rho, H. Kwon, and S. Kook. A practical MRI grading system for cervical foraminal stenosis based on oblique sagittal images.The British journal of radiology, 86(1025), 2013

    2013

  31. [31]

    M. S. Park, S.-H. Moon, H.-M. Lee, T.-H. Kim, J. K. Oh, S. Y. Lee, J. B. Oh, and K. D. Riew. Diagnostic value of oblique magnetic resonance images for evaluating cervical foraminal stenosis.The Spine Journal, 15(4):607–611, 2015

    2015

  32. [32]

    R. Y. Park, R. Windsor, A. Jamaludin, and A. Zisserman. Multi-view and multimodal radiological grading using spinal MRIs. InML-CDS Workshop, MICCAI 2025, 2025

    2025

  33. [33]

    T. J. Richards, A. E. Flanders, E. Colak, L. M. Prevedello, R. L. Ball, F. Kitamura, J. Mongan, M. Vazirabad, H.-M. Lin, A. Kendell, et al. The RSNA lumbar degenerative imaging spine classification (LumbarDISC) dataset.Radiology: Artificial Intelligence, page e250480, 2026

    2026

  34. [34]

    Sąsiadek and J

    M. Sąsiadek and J. Jacków-Nowicka. Degenerative disease of the spine: How to re- late clinical symptoms to radiological findings.Advances in Clinical and Experimental Medicine, 33(1):91–98, 2024

    2024

  35. [35]

    Sobański, R

    D. Sobański, R. Staszkiewicz, M. Stachura, M. Gadzieliński, and B. O. Grabarek. Pre- sentation, diagnosis, and management of lower back pain associated with spinal steno- sis: a narrative review.Medical Science Monitor: International Medical Journal of Experimental and Clinical Research, 29:e939237–1, 2023. 17 R. Park, M. Eid et al

    2023

  36. [36]

    Takahashi, A

    K. Takahashi, A. K. Yadav, K. Hashimoto, T. Tsubakino, T. Aizawa, and Y. Tanaka. Foraminal stenosis at L5–S1 as an overlooked pathology of bilateral radiculopathy: A case series.Journal of Orthopaedic Case Reports, 12(6):13, 2022

    2022

  37. [37]

    N. A. Tawfik, A. T. Ahmed, T. E. El-Shafei, and M. R. Habba. Diagnostic value of spinal ultrasound compared to MRI for diagnosis of spinal anomalies in pediatrics. Egyptian Journal of Radiology and Nuclear Medicine, 51(1):18, Jan. 2020

    2020

  38. [38]

    Van Reeth, I

    E. Van Reeth, I. W. Tham, C. H. Tan, and C. L. Poh. Super-resolution in magnetic resonance imaging: a review.Concepts in Magnetic Resonance Part A, 40(6):306–325, 2012

    2012

  39. [39]

    J. C. van Rijn, N. Klemetsö, J. B. Reitsma, C. B. Majoie, F. J. Hulsmans, W. C. Peul, J. Stam, P. M. Bossuyt, and G. J. den Heeten. Observer variation in MRI evaluation of patients suspected of lumbar disk herniation.American Journal of Roentgenology, 184(1):299–303, 2005

    2005

  40. [40]

    Verheijen, T

    E. Verheijen, T. Kapogiannis, D. Munteh, J. Chabros, M. Staring, T. Smith, and C. Vleggeert-Lankamp. Artificial intelligence for segmentation and classification in lumbar spinal stenosis: an overview of current methods.European Spine Journal, 34(3):1146–1155, 2025

    2025

  41. [41]

    Windsor, A

    R. Windsor, A. Jamaludin, T. Kadir, and A. Zisserman. Context-aware transformers for spinal cancer detection and radiological grading. InMICCAI, pages 271–281, 2022

    2022

  42. [42]

    Windsor, A

    R. Windsor, A. Jamaludin, T. Kadir, and A. Zisserman. Automated detection, labelling and radiological grading of clinical spinal MRIs.Scientific Reports, 14(1):14993, 2024

    2024

  43. [43]

    R. Zha, T. J. Lin, Y. Cai, J. Cao, Y. Zhang, and H. Li. R^2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. InNeurIPS, volume 37, pages 44907–44934, 2024

    2024

  44. [44]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, pages 586–595, 2018. 18

    2018