acceptodds
Under review as a conference paper at ICLR 2027

GaLA-SVR: Feed-Forward 3D Gaussian Framework with Physics-Aware Latent Aggregation for Slice-to-Volume Reconstruction

Abstract

Reconstructing coherent 3D volumes from sparse, motion-corrupted MRI slices requires accurate motion estimation and faithful integration of incomplete observations. Existing learning-based SVR methods often reconstruct an intermediate volume assembled from motion-corrected slice intensities for subsequent refinement, rather than leveraging semantically rich slice features across stacks. This can limit their ability to exploit complementary observations from different views. In sparsely observed regions, refinement needs to hallucinate missing structures from incomplete volume intensity information. To address this, we propose GaLA-SVR, a feed-forward 3D Gaussian framework with physics-aware Latent Aggregation for Slice-to-Volume Reconstruction. GaLA-SVR bypasses existing intermediate volume reconstruction, with its core consisting of a physics-aware latent aggregation scheme and a dense patch-wise correspondence-based motion estimator. First, we propose a novel Point Spread Function (PSF)-aware feature aggregation scheme to learn a comprehensive 3D latent representation from 2D slice features, maintaining structural completeness even in sparsely observed or unobserved regions, which is then decoded into anisotropic 3D Gaussians for high-fidelity volumetric rendering. To achieve the precise spatial alignment essential for feature aggregation, we reformulate slice motion estimation as a dense patch-wise correspondence problem, facilitated by a novel Intra- and Inter-Slice Attention (IISA) block. Experiments across three MRI datasets demonstrate that GaLA-SVR achieves SOTA volumetric reconstruction quality with accurate slice motion estimation. Code will be released upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D gaussian representation, slice-to-volume reconstruction, 3D latent representation, feed-forward 3D Gaussian

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NQ7tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method, GaLA-SVR, to reconstruct MRI scans into 3D volume measurements using 2D slices. The method corrects and synthesizes different 2D MRI slices from different angles by using a point-spread function to create more accurate and informed 3D reconstructions. This method is evaluated on the accuracy of different MRI reconstructions and how well the motion estimation is predicted.

Critical strengths
  1. This paper identified and explained the problem of 3D MRI reconstruction clearly, and the theory behind the method is well-defined with sound logic. It explains the limitations of current existing methods and proposes a logical solution.
  2. In the Experiment section, the results (Figure 3) and table (Table 1) match each other and it is clear where there are improvements on all benchmarks. The results are presented cleanly and make it easy to identify how the paper’s method is able to be close to the ground truth.
  3. In the Method section, the figure (Figure 2) displays the method relatively clearly, visually corresponding to each part of the method that is explained.
Critical weaknesses
  1. The datasets used for reconstruction of the 3D MRI volumes are under simulation, and not real-world patient motion data. This concerns if the results, no matter how well they performed, can perform equally as well in real patient scans.
  2. There is risk that the anatomy is not accurate to real-patient data and is instead derived from learned patterns from each stage. Even though there is noise reduction being done, it is possible to hallucinate structures that are not there and not accurate to human anatomy.
  3. Due to the limited amount of data, there might be too much of a generalization about the brain structure, as every patient’s brain will vary. If there is a case where a patient does not have the typical brain anatomy, it could result in more hallucinations about the entire structure.
  4. The comparison methods listed in the main results (Table 1 and Table 2) were mainly published more than two years ago; it is not possible to evaluate the performance compared with the SOTA methods in recent years.
Questions for authors

How well can the simulated MRI volume datasets be trusted for fair evaluation when part of the evaluation is for real motion accuracy from a patient’s movement?

6 comments
AuthorsOpenReview ↗

To reviewer NQ7t [Part I]

We appreciate the reviewer's thoughtful evaluation and helpful recommendations, which have strengthened our paper. The manuscript has been revised accordingly, and we respond to each comment in turn below.

W1 & Q1: Real motion-corrupted data concern

To assess generalization beyond simulated motion, we further evaluate all methods on real motion-corrupted neonatal scans from the dHCP dataset [Ref1]. We select 30 subjects from the originally acquired, motion-corrupted T2-weighted stacks of the dHCP neonatal cohort. Each subject consists of 2 to 4 image stacks with in-plane resolution of 0.8 mm, slice gap of 0.8 mm, and slice thickness of 1.6 mm. Each stack is skull-stripped with SynthStrip [Ref2] from FreeSurfer, resampled in-plane to 1.0 mm, center-cropped to 128×128128\times128, intensity-normalized, and reoriented to the slice convention of the simulated stacks. The same preprocessing is applied to all methods. Since no ground-truth slice motion is available, we report volume reconstruction metrics only, using the official dHCP reconstruction of the same session. All learning-based models are trained only on simulated stacks on the dHCP dataset and applied to the real motion-corrupted data in a zero-shot manner.

As shown in Table R1, GaLA-SVR achieves the best zero-shot performance among all learning-based methods, with a PSNR of 21.00 dB and an SSIM of 0.761, outperforming the strongest learning-based baseline cSVR (20.66 dB). It also surpasses all optimization-based methods, including the recent Gaussian-based methods GaussianSVR and GSVR, as well as SVRTK and NeSVoR. Note that these four methods need long per-subject optimization time. NeSVoR, GaussianSVR, and GSVR start from slice transformations provided by a pretrained SVoRT. In contrast, GaLA-SVR is end-to-end and reconstructs a subject directly from the input stacks in 0.56 s with a single forward pass, about 200× faster than NeSVoR (110.5 s). This indicates that the learned prior transfers well to real acquisitions. Qualitative comparison can be found in Figure 8 in the Appendix of the revised manuscript. We have also added these results and discussion in the revised manuscript in Section 4.2.

Table R1. Quantitative comparison of reconstruction performance on real scans on the dHCP dataset. PSNR, SSIM, and NRMSE are reported as mean ± standard deviation; inference time is the mean runtime per subject. Best results are in bold and second-best in italics.

MethodPSNR ↑SSIM ↑NRMSE ↓Inf. (s) ↓
Optimization-based
SVRTK20.52 ± 1.850.745 ± 0.0480.540 ± 0.130856
NeSVoR20.55 ± 1.780.745 ± 0.0490.538 ± 0.126110.5
GaussianSVR19.95 ± 1.690.735 ± 0.0480.574 ± 0.1271183.5
GSVR19.54 ± 1.100.684 ± 0.0440.595 ± 0.098142.3
Learning-based
SVRnet19.70 ± 1.260.715 ± 0.0260.528 ± 0.0490.18
PlaneInVol19.76 ± 1.290.723 ± 0.0240.520 ± 0.0540.20
SVoRT20.11 ± 1.230.726 ± 0.0310.530 ± 0.0841.17
FC-SVR14.63 ± 0.730.544 ± 0.0321.039 ± 0.1103.07
SVRMamba20.05 ± 1.120.704 ± 0.0570.558 ± 0.0740.21
cSVR20.66 ± 1.720.742 ± 0.0430.529 ± 0.1138.85
GaLA-SVR21.00 ± 1.660.761 ± 0.0410.507 ± 0.1060.56

[Ref1] Hughes, Emer J., et al. "A dedicated neonatal brain imaging system." Magnetic Resonance in Medicine 2017

[Ref2] Hoopes, Andrew, et al. "SynthStrip: Skull-stripping for any brain image." NeuroImage 2022

AuthorsOpenReview ↗

To reviewer NQ7t [Part II]

W2 & W3: Hallucination concern

For the concern that the anatomy may not be accurate on real-patient data, our response to W1 has shown that GaLA-SVR generalizes to real motion-corrupted neonatal scans, with the reconstructed anatomy consistent with the official dHCP [Ref1] reconstruction and no structure appearing that is absent from the reference (Table R1 and Figure 8 in the Appendix of the revised manuscript).

To further demonstrate that the reconstruction of GaLA-SVR is driven by the input slices rather than by learned patterns or anatomical priors alone, we compare it on the 76 dHCP test subjects with two prior-only predictions, which do not use the input slices at all. The first is the mean of the 675 training volumes, which is the same for every test subject. The second is, for each test subject, the training volume that is closest to its ground-truth volume, i.e., the best result that could be obtained by retrieving a brain from the training set. As shown in Table R2, they reach only 24.46 and 23.25 dB, whereas GaLA-SVR reaches 29.47 dB and is better on all 76 subjects. The performance therefore cannot be explained by learned patterns or anatomical priors alone. In addition, we conducted a perturbation experiment on the FeTA test set to verify that the reconstruction depends on the slice motion. We feed the same input slices to the network, but replace the slice motions estimated by the network with artificially perturbed ones before the PSF-aware aggregation. The perturbed motions are obtained by adding an independent random error to the ground-truth motion of each slice (a translation of LL mm plus a rotation of L/35.9L/35.9 rad about the slice center). As shown in Table R3, the PSNR decreases monotonically from 30.24 dB with the ground-truth motions to 29.32, 26.89, and 22.34 dB at L=2L=2, 4, and 8 mm, which would not happen if the reconstruction were generated from learned patterns alone.

Table R2. Comparison with predictors that do not use the input slices on the 76 dHCP test subjects. Values are mean ± standard deviation.

PredictionPSNR ↑SSIM ↑NRMSE ↓
Mean of the training volumes24.46 ± 1.020.810 ± 0.0110.345 ± 0.035
Most similar training volume23.25 ± 0.750.798 ± 0.0080.396 ± 0.031
GaLA-SVR29.47 ± 0.870.883 ± 0.0070.194 ± 0.022

Table R3. Reconstruction of GaLA-SVR on the FeTA test set when the slice transformations passed to the aggregation are replaced. In-brain disp. is the mean displacement of the pixels inside the brain. Values are mean ± standard deviation.

Slice transformationsED (mm)In-brain disp. (mm)PSNR ↑SSIM ↑
GT0.000.0030.24 ± 2.100.888 ± 0.007
GT + 2 mm3.54 ± 0.052.15 ± 0.0329.32 ± 2.190.884 ± 0.010
GT + 4 mm6.95 ± 0.204.32 ± 0.0626.89 ± 2.470.862 ± 0.020
GT + 8 mm14.17 ± 0.328.58 ± 0.1822.34 ± 2.270.814 ± 0.030
GaLA-SVR (estimated)6.10 ± 2.783.07 ± 1.2929.11 ± 2.080.881 ± 0.011

To examine how GaLA-SVR behaves when a subject does not have the typical brain anatomy, we tested it on structures that never appear in the training data. We inserted synthetic structures into the ground-truth volumes of dHCP test subjects (a hypointense sphere of radius 3 mm or 8 mm), and re-simulated the stacks with the original motion, slice thickness, and noise. Figure 4 in the Appendix of the revised manuscript shows that GaLA-SVR reconstructs the inserted structures at the correct location and with a similar extent, and that the error outside the structure remains at the same level as in the surrounding tissue. This indicates that the reconstruction follows the anatomy of the individual subject rather than a generalized brain structure, and that the presence of an atypical structure does not lead to hallucinations in the rest of the brain.

[Ref1] Hughes, Emer J., et al. "A dedicated neonatal brain imaging system." Magnetic Resonance in Medicine 2017

AuthorsOpenReview ↗

To reviewer NQ7t [Part III]

W4: Baseline evaluation concern

We would like to clarify that the original comparison already included the state-of-the-art learning-based SVR methods, among which FC-SVR (2024) and SVRMamba (2025) were published within the last two years. Since GaLA-SVR reconstructs a volume in an end-to-end fashion within a second, we therefore considered learning-based baselines with similar computational requirements to be a fairer comparison, with only one optimization-based method, SVRTK, as reference. Following the suggestion, we have added four recent methods to all comparisons: three optimization-based methods, NeSVoR (2023), GaussianSVR (2026), and GSVR (2026), and cSVR [Ref1] (2026), a recently published method that combines a feed-forward motion estimator with an optimization-based reconstruction. Note that the first three methods require extra initialization of slice transformations provided by an external pretrained SVoRT method. Their results are summarized in Table R4 and Table R5.

On CamCAN and dHCP, where sufficient training subjects are available, GaLA-SVR remains the best on all three metrics. Its PSNR gains over the strongest added method on each dataset, 0.91 dB over cSVR on CamCAN and 0.45 dB over NeSVoR on dHCP, are statistically significant (P<0.01), while being orders of magnitude faster than per-subject optimization. On FeTA, where only 65 training subjects are available, all learning-based methods are limited by the data size, and NeSVoR and GaussianSVR, which benefit from per-subject optimization with an already good initialization, achieve the highest PSNR (31.85 and 31.31 dB). The reconstruction of cSVR is optimization-based, so it is likely less affected by the limited training data. We also added the results and discussion in Section 4.2 in the revised manuscript.

Table R4. Quantitative comparison of reconstruction performance on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. Values are mean ± standard deviation. Best results are in bold and second-best in italics.

MethodFeTA PSNR ↑FeTA SSIM ↑FeTA NRMSE ↓CamCAN PSNR ↑CamCAN SSIM ↑CamCAN NRMSE ↓dHCP PSNR ↑dHCP SSIM ↑dHCP NRMSE ↓
Optimization-based
NeSVoR31.85 ± 4.150.890 ± 0.0340.167 ± 0.11125.94 ± 1.360.864 ± 0.0090.183 ± 0.03429.02 ± 0.940.878 ± 0.0080.204 ± 0.028
GaussianSVR31.31 ± 2.830.889 ± 0.0090.166 ± 0.05325.02 ± 2.200.842 ± 0.0410.224 ± 0.06527.55 ± 1.560.859 ± 0.0150.245 ± 0.045
GSVR25.65 ± 2.260.716 ± 0.0750.317 ± 0.10023.40 ± 2.760.782 ± 0.0590.275 ± 0.09127.85 ± 0.730.787 ± 0.0810.226 ± 0.031
Learning-based
cSVR29.12 ± 2.430.879 ± 0.0130.206 ± 0.04026.27 ± 0.860.861 ± 0.0070.182 ± 0.02328.21 ± 0.750.872 ± 0.0070.225 ± 0.023
GaLA-SVR29.11 ± 2.080.881 ± 0.0110.198 ± 0.02127.18 ± 1.210.873 ± 0.0150.171 ± 0.02329.47 ± 0.870.883 ± 0.0070.194 ± 0.022

Table R5. Quantitative comparison of transformation errors and inference time on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. ED and GD are mean ± standard deviation; inference time is the mean runtime per subject. Best transformation results are in bold and second-best in italics.

MethodFeTA ED (mm) ↓FeTA GD (rad) ↓FeTA Inf. (s) ↓CamCAN ED (mm) ↓CamCAN GD (rad) ↓CamCAN Inf. (s) ↓dHCP ED (mm) ↓dHCP GD (rad) ↓dHCP Inf. (s) ↓
Optimization-based
NeSVoR5.96 ± 4.440.098 ± 0.082115.496.37 ± 5.910.077 ± 0.079146.803.98 ± 1.770.068 ± 0.038106.42
GaussianSVR3.35 ± 2.190.061 ± 0.043647.336.33 ± 5.530.079 ± 0.0741443.543.85 ± 2.040.073 ± 0.037805.60
GSVR11.38 ± 7.150.205 ± 0.13678.149.16 ± 6.310.118 ± 0.086318.684.35 ± 1.770.079 ± 0.029165.26
Learning-based
cSVR6.31 ± 2.920.113 ± 0.0584.444.56 ± 1.130.054 ± 0.0167.635.23 ± 4.010.084 ± 0.0815.30
GaLA-SVR6.10 ± 2.780.110 ± 0.0540.113.22 ± 1.040.041 ± 0.0140.273.79 ± 3.620.066 ± 0.0790.17

[Ref1] Xu, Junshen, et al. "cSVR: Convolutional Slice-to-Volume Reconstruction." MICCAI PIPPI Workshop 2026

Reviewer_TpyuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GaLA-SVR, a feed-forward 3D Gaussian framework for slice-to-volume reconstruction (SVR). The method consists of three main components: (1) slice motion estimation is reformulated as dense patch-wise canonical coordinate correspondence, using IISA blocks and a Planar Kabsch solver to analytically recover each slice’s rigid transformation; (2) a PSF-aware feature aggregation scheme actively queries and fuses 2D slice features from predefined 3D candidate cells to form a 3D latent representation; (3) a sparse Transformer decodes the latent into anisotropic 3D Gaussians, which are evaluated at voxel centers to obtain the reconstructed volume. Experiments are conducted on three T2w MRI datasets, FeTA, dHCP, and CamCAN, reporting reconstruction quality, motion estimation errors, inference time, and ablations.

Critical strengths

(1) The paper targets a representative bottleneck of existing learning-based SVR methods, namely assembling an intermediate intensity volume and then refining it. Instead, it aggregates semantic features in a 3D latent space, avoiding the representational bottleneck of intermediate volumes. The overall pipeline is end-to-end and requires a single forward pass at inference. (2) Changing SVR motion estimation from sparse regression of global parameters/anchor points to dense patch correspondence plus an analytical Planar Kabsch solver provides richer spatial supervision. The IISA block has a clear inductive bias, and ablations show that inter-slice attention is particularly important.

Critical weaknesses

(1) Quantitative comparison with the most relevant 2026 work is missing. The paper cites 3D Gaussian SVR works such as Wang et al., 2026b and Dannecker et al., 2026. However, no comparison was made with these methods. In addition, some feed-forward SVR methods should also be compared (cSVR: Convolutional Slice-to-Volume Reconstruction, 2026). This makes the “state-of-the-art” conclusion insufficiently supported. The authors should add comparisons with these methods. (2) The novelty of PSF-aware aggregation is limited and needs clearer differentiation from existing PSF modeling work. Dannecker et al. already exploit the closedness of Gaussians under convolution to derive an analytical solution for the PSF forward model. The contribution here is to use through-plane PSF weights for feed-forward feature fusion, not to invent PSF weighting itself.

Questions for authors

(1) Can the authors add quantitative comparisons with Wang et al., 2026b and Dannecker et al., 2026., and (cSVR: Convolutional Slice-to-Volume Reconstruction, 2026)? If direct comparison is infeasible due to unavailable code/data, can the authors provide qualitative or indirect comparisons under the same simulated slice protocol? (2) Dependence on the atlas and candidate cell selection is not analyzed. The method uses a fixed atlas to define 3D candidate cells and dilates the brain region by one cell. For subjects with anatomical abnormalities, segmentation failures, or incomplete fields of view, the atlas may miss valid regions. The authors should analyze the impact of atlas quality, dilation range, and the number of candidate cells on reconstruction and motion estimation.

3 comments
AuthorsOpenReview ↗

To reviewer Tpyu [Part I]

We are grateful to the reviewer for the positive assessment and the valuable suggestions. All of them have been taken into account in the revised paper, and our point-by-point responses are given below.

W1 & Q1: Baseline evaluation concern----------------------------------------------------------------------

We would like to first clarify why NeSVoR, GaussianSVR, and GSVR were not included originally. GaLA-SVR is a feed-forward method that estimates the slice transformations and reconstructs the volume in an end-to-end fashion within a second. In contrast, NeSVoR, GaussianSVR, and GSVR are optimized separately for each subject, take orders of magnitude longer, and require extra initializations of slice transformations provided by an external pretrained SVoRT method. We therefore considered learning-based baselines with similar computational requirements to be a fairer comparison, with only one optimization-based method, SVRTK, as reference. Nevertheless, following the suggestion, we added the comparison against these three methods along with the recent cSVR [Ref1], which combines a feed-forward motion estimator with an optimization-based reconstruction, to all comparisons. Their results are summarized in Table R1 and Table R2.

On CamCAN and dHCP [Ref5], where sufficient training subjects are available, GaLA-SVR remains the best on all three metrics. Its PSNR gains over the strongest added method on each dataset, 0.91 dB over cSVR on CamCAN and 0.45 dB over NeSVoR on dHCP, are statistically significant (P<0.01), while being orders of magnitude faster than per-subject optimization. On FeTA, because there are only 65 training subjects, the results are likely limited by the data size. All learning-based methods are trained from scratch on the same stacks and are affected in the same way. NeSVoR and GaussianSVR, which benefit from per-subject optimization with an already good initialization, achieve the highest PSNR (31.85 and 31.31 dB) on the FeTA dataset. The reconstruction of cSVR is optimization-based, so it is likely less affected by the limited training data. We have also incorporated the results and discussion in the revised manuscript.

Table R1. Quantitative comparison of reconstruction performance on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. Values are mean ± standard deviation. Best results are in bold and second-best in italics.

MethodFeTA PSNR ↑FeTA SSIM ↑FeTA NRMSE ↓CamCAN PSNR ↑CamCAN SSIM ↑CamCAN NRMSE ↓dHCP PSNR ↑dHCP SSIM ↑dHCP NRMSE ↓
Optimization-based
NeSVoR31.85 ± 4.150.890 ± 0.0340.167 ± 0.11125.94 ± 1.360.864 ± 0.0090.183 ± 0.03429.02 ± 0.940.878 ± 0.0080.204 ± 0.028
GaussianSVR31.31 ± 2.830.889 ± 0.0090.166 ± 0.05325.02 ± 2.200.842 ± 0.0410.224 ± 0.06527.55 ± 1.560.859 ± 0.0150.245 ± 0.045
GSVR25.65 ± 2.260.716 ± 0.0750.317 ± 0.10023.40 ± 2.760.782 ± 0.0590.275 ± 0.09127.85 ± 0.730.787 ± 0.0810.226 ± 0.031
Learning-based
cSVR29.12 ± 2.430.879 ± 0.0130.206 ± 0.04026.27 ± 0.860.861 ± 0.0070.182 ± 0.02328.21 ± 0.750.872 ± 0.0070.225 ± 0.023
GaLA-SVR29.11 ± 2.080.881 ± 0.0110.198 ± 0.02127.18 ± 1.210.873 ± 0.0150.171 ± 0.02329.47 ± 0.870.883 ± 0.0070.194 ± 0.022

Table R2. Quantitative comparison of transformation errors and inference time on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. ED and GD are mean ± standard deviation; inference time is the mean runtime per subject. Best transformation results are in bold and second-best in italics.

MethodFeTA ED (mm) ↓FeTA GD (rad) ↓FeTA Inf. (s) ↓CamCAN ED (mm) ↓CamCAN GD (rad) ↓CamCAN Inf. (s) ↓dHCP ED (mm) ↓dHCP GD (rad) ↓dHCP Inf. (s) ↓
Optimization-based
NeSVoR5.96 ± 4.440.098 ± 0.082115.496.37 ± 5.910.077 ± 0.079146.803.98 ± 1.770.068 ± 0.038106.42
GaussianSVR3.35 ± 2.190.061 ± 0.043647.336.33 ± 5.530.079 ± 0.0741443.543.85 ± 2.040.073 ± 0.037805.60
GSVR11.38 ± 7.150.205 ± 0.13678.149.16 ± 6.310.118 ± 0.086318.684.35 ± 1.770.079 ± 0.029165.26
Learning-based
cSVR6.31 ± 2.920.113 ± 0.0584.444.56 ± 1.130.054 ± 0.0167.635.23 ± 4.010.084 ± 0.0815.30
GaLA-SVR6.10 ± 2.780.110 ± 0.0540.113.22 ± 1.040.041 ± 0.0140.273.79 ± 3.620.066 ± 0.0790.17

[Ref1] Xu, Junshen, et al. "cSVR: Convolutional Slice-to-Volume Reconstruction." MICCAI PIPPI Workshop 2026

AuthorsOpenReview ↗

To reviewer Tpyu [Part II]

W2: PSF-aware feature aggregation clarification-------------------------------------------------------

We would like to clarify that the purpose of employing the PSF differs between GSVR and our approach. GSVR [Ref3] uses a Gaussian PSF to render slice intensities from the Gaussian primitives in closed form, which avoids costly Monte Carlo sampling and accelerates the per-subject optimization. In GaLA-SVR, the PSF is not used to render slice intensities, so there is no need to invent a new PSF weighting. Instead, we adopt the standard Gaussian PSF model and derive from it a through-plane weight for aggregating learned 2D slice features into a 3D latent representation in a single forward pass. The ablation in Table 4(a) in the manuscript confirms its benefit: replacing the PSF weights with uniform weights lowers the PSNR from 29.47 to 28.10 dB. We have clarified this distinction in the revised manuscript in Section 2.

AuthorsOpenReview ↗

To reviewer Tpyu [Part III]

Q2: Sensitivity analysis on dilation range of atlas and latent grid resolution

We would like to first clarify how the atlas is used. It is a fixed, coarse support built from the union of 20 training masks followed by a one-cell dilation, and it only defines which latent cells are queried. No segmentation of the test subject is required at inference, so a segmentation failure on a test subject cannot affect the atlas. To address the concern that a fixed atlas may miss valid regions when a subject's anatomy lies outside its support, we added two analyses on the dHCP dataset [Ref2], as suggested. The results are also included in the revised manuscript (Appendix C, Tables 6–8).

(1) Impact of atlas quality and dilation range. We measure how well the atlas covers each test subject by its recall, i.e., the fraction of the brain cells of the subject that lie inside the atlas. As shown in Table R3 and Table R4, with the default one-cell dilation, the recall is at least 0.9989 on all test subjects. To study the effect of an atlas that misses valid regions, we trained a model without dilation, for which the recall drops to 0.9585 in the worst case. Since cells outside the atlas are never queried, the uncovered regions cannot be reconstructed, and the PSNR decreases from 29.47 to 29.07 dB. An incomplete atlas therefore degrades the reconstruction, but only moderately. Enlarging the dilation from one to two cells does not improve the coverage further and brings no gain in reconstruction (29.40 dB), while the number of candidate cells increases by 27%. The reconstruction is thus not sensitive to an oversized atlas: cells outside the brain project onto the background of the input slices and contribute little. This also means that, for abnormal subjects whose anatomy may exceed the atlas, a larger dilation can be used as a safeguard at a moderate computational cost and without loss of accuracy. The choice of dilation has little effect on motion estimation either (ED between 3.78 and 3.79 mm across the three settings), because the atlas is only used in the feature aggregation, which takes place after the slice motion has been estimated.

Table R3. Effect of the dilation range of the atlas on the dHCP dataset. Values are mean ± standard deviation. Best results are in bold.

DilationPSNR ↑SSIM ↑NRMSE ↓ED (mm) ↓GD (rad) ↓TED (mm) ↓
0 cell29.07 ± 0.960.880 ± 0.0080.201 ± 0.0253.789 ± 3.5980.065 ± 0.0772.178 ± 2.491
1 cell29.47 ± 0.870.883 ± 0.0070.194 ± 0.0223.785 ± 3.6230.066 ± 0.0792.193 ± 2.563
2 cells29.40 ± 0.920.881 ± 0.0070.196 ± 0.0233.776 ± 3.5610.068 ± 0.0762.214 ± 2.488

Table R4. Coverage of the atlas with different dilation ranges on the dHCP dataset. Recall and precision are mean ± standard deviation over test subjects.

DilationActive cellsRecall ↑Min. recall ↑Precision ↑
0 cell18,7320.9964 ± 0.00700.95850.903 ± 0.014
1 cell24,4131.0000 ± 0.00010.99890.695 ± 0.015
2 cells30,9161.0000 ± 0.00001.00000.549 ± 0.012

(2) Impact of number of candidate cells. Besides the dilation, the number of candidate cells is also determined by the resolution of the latent grid (Table R5). Reducing the grid from 46×46×3446\times46\times34 to 23×23×1723\times23\times17 lowers the PSNR by 1.94 dB, because each cell must represent an eight-times larger region. Increasing it to 70×70×5170\times70\times51 yields a smaller gain of 0.39 dB, while the number of cells roughly triples and training takes about 2.2 times longer. The default grid offers a favorable trade-off between accuracy and efficiency. As with the dilation range, the resolution of the latent grid has little effect on motion estimation (ED between 3.78 and 3.80 mm across the three grids), because the latent grid is only used in the feature aggregation, which takes place after the slice motion has been estimated.

Table R5. Effect of the resolution of the 3D latent grid on the dHCP dataset. Values are mean ± standard deviation. Best results are in bold.

Latent gridPSNR ↑SSIM ↑NRMSE ↓ED (mm) ↓GD (rad) ↓TED (mm) ↓
23×23×1723\times23\times1727.53 ± 0.800.859 ± 0.0080.237 ± 0.0263.793 ± 3.750.066 ± 0.0802.168 ± 2.563
46×46×3446\times46\times3429.47 ± 0.870.883 ± 0.0070.194 ± 0.0223.785 ± 3.6230.066 ± 0.0792.193 ± 2.563
70×70×5170\times70\times5129.86 ± 1.000.887 ± 0.0070.191 ± 0.0243.803 ± 3.640.067 ± 0.0802.184 ± 2.471

[Ref2] Hughes, Emer J., et al. "A dedicated neonatal brain imaging system." Magnetic Resonance in Medicine 2017

Reviewer_Cu9aOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GaLA-SVR is a feed-forward MRI slice-to-volume reconstruction network that jointly estimates rigid slice motion and a high-resolution 3D volume from corrupted 2D stacks. A shared CNN encodes slices, while IISA attention blocks model slice relationships to predict 3D patch coordinates. An intensity-weighted Kabsch solver converts these correspondences into rigid transforms. Using the estimated motion, slice features are back-projected to a fixed atlas grid and fused with Gaussian PSF weighting. A sparse transformer then decodes active cells into anisotropic Gaussians rendered onto the voxel grid. The network is trained end-to-end with coordinate, motion, L1, and D-SSIM losses. On simulated data, it achieves the best reconstruction metrics, improving PSNR by 1.9–2.5 dB over the strongest baseline, while also providing strong motion accuracy and 0.1–0.3 s inference per subject.

Critical strengths

The paper is well motivated. Rather than assembling motion-corrected intensities into an intermediate volume, GaLA-SVR fuses learned slice features directly in 3D before decoding. The motion head is also principled: dense canonical patch coordinates plus a weighted planar Procrustes solver produce valid, differentiable rotations and richer supervision than direct anchor regression, as supported by the dHCP ablation. Querying a fixed grid with PSF-weighted slice features is similarly well justified and consistently outperforms simpler alternatives by about 1–1.7 dB. IISA efficiently models slice-level motion using per-slice tokens. Experiments span fetal, neonatal, and adult cohorts, with large, consistent gains and sub-second inference.

Critical weaknesses

My main concern is that the evaluation is entirely simulation-based under fixed acquisition geometry. Training and test data use the same SVoRT-style simulator, with no intra-slice motion, bias field, or random variation in stack number or orientation, and there are no real motion-corrupted results. Given that prior methods such as SVoRT and NeSVoR evaluate on real fetal data, this is a substantial limitation.

A second concern is that performance may reflect a strong learned anatomical prior rather than faithful use of the input. GaLA-SVR decodes within a fixed population atlas from canonically aligned targets, and reconstruction quality remains high even when motion accuracy degrades. For example, on FeTA the motion error is 6.10 mm despite 2.5–3.5 mm slice thickness, while GaLA-SVR still leads by about 1.9 dB. Likewise, coarsening the patch grid noticeably worsens motion but reduces PSNR by only 0.32 dB. It seems plausible that the decoder may hallucinate plausible anatomy.

The baseline evaluation also needs strengthening. NeSVoR and two recent Gaussian-splatting SVR methods are cited but not compared, even though reconstruction quality is the paper’s main claim. Finally, all learned baselines are retrained from scratch, while FeTA has only 65 training subjects, raising concerns that some results may reflect data-size sensitivity rather than method quality.

Questions for authors
  1. How do you make sure that none of the details in the final reconstruction are hallucinated?
  2. How does your method performs against recent methods such as NeSVoR or CineVol?
4 comments
AuthorsOpenReview ↗

To reviewer Cu9a [Part I]

We sincerely appreciate the reviewer's positive and constructive feedback. It is very helpful in improving our manuscript. We have carefully incorporated it into the revised paper. Please see our responses below, which we hope will effectively address your concerns.

W1: Real motion-corrupted data concern----------------------------------------------------------------------

To assess generalization beyond simulated motion, we further evaluate all methods on real motion-corrupted neonatal scans from the dHCP dataset [Ref1]. We select 30 subjects from the originally acquired, motion-corrupted T2-weighted stacks of the dHCP neonatal cohort. Each subject consists of 2 to 4 image stacks with in-plane resolution of 0.8 mm, slice gap of 0.8 mm, and slice thickness of 1.6 mm. Each stack is skull-stripped with SynthStrip [Ref2] from FreeSurfer, resampled in-plane to 1.0 mm, center-cropped to 128×128128\times128, intensity-normalized, and reoriented to the slice convention of the simulated stacks. The same preprocessing is applied to all methods. Since no ground-truth slice motion is available, we report volume reconstruction metrics only, using the official dHCP reconstruction of the same session. All learning-based models are trained only on simulated stacks on the dHCP dataset and applied to the real motion-corrupted data in a zero-shot manner.

As shown in Table R1, GaLA-SVR achieves the best zero-shot performance among all learning-based methods, with a PSNR of 21.00 dB and an SSIM of 0.761, outperforming the strongest learning-based baseline cSVR (20.66 dB). It also surpasses all optimization-based methods, including the recent Gaussian-based methods GaussianSVR and GSVR, as well as SVRTK and NeSVoR. Note that these four methods need long per-subject optimization time. NeSVoR, GaussianSVR, and GSVR start from slice transformations provided by a pretrained SVoRT. In contrast, GaLA-SVR is end-to-end and reconstructs a subject directly from the input stacks in 0.56 s with a single forward pass, about 200× faster than NeSVoR (110.5 s). This indicates that the learned prior transfers well to real acquisitions. Qualitative comparison can be found in Figure 8 in the Appendix of the revised manuscript.

Table R1. Quantitative comparison of reconstruction performance on real scans on the dHCP dataset. PSNR, SSIM, and NRMSE are reported as mean ± standard deviation; inference time is the mean runtime per subject. Best results are in bold and second-best in italics.

MethodPSNR ↑SSIM ↑NRMSE ↓Inf. (s) ↓
Optimization-based
SVRTK20.52 ± 1.850.745 ± 0.0480.540 ± 0.130856
NeSVoR20.55 ± 1.780.745 ± 0.0490.538 ± 0.126110.5
GaussianSVR19.95 ± 1.690.735 ± 0.0480.574 ± 0.1271183.5
GSVR19.54 ± 1.100.684 ± 0.0440.595 ± 0.098142.3
Learning-based
SVRnet19.70 ± 1.260.715 ± 0.0260.528 ± 0.0490.18
PlaneInVol19.76 ± 1.290.723 ± 0.0240.520 ± 0.0540.20
SVoRT20.11 ± 1.230.726 ± 0.0310.530 ± 0.0841.17
FC-SVR14.63 ± 0.730.544 ± 0.0321.039 ± 0.1103.07
SVRMamba20.05 ± 1.120.704 ± 0.0570.558 ± 0.0740.21
cSVR20.66 ± 1.720.742 ± 0.0430.529 ± 0.1138.85
GaLA-SVR21.00 ± 1.660.761 ± 0.0410.507 ± 0.1060.56

[Ref1] Hughes, Emer J., et al. "A dedicated neonatal brain imaging system." Magnetic Resonance in Medicine 2017

[Ref2] Hoopes, Andrew, et al. "SynthStrip: Skull-stripping for any brain image." NeuroImage 2022

AuthorsOpenReview ↗

To reviewer Cu9a [Part II]

W2: Learned anatomical prior concern----------------------------------------------------------------------------------

To address this concern, we conducted additional experiments that test whether the reconstruction of GaLA-SVR relies on the input slices or on a learned anatomical prior. They are summarized below.

(1) Relation between reconstruction quality and motion accuracy. We would like to first clarify that the motion estimation is not "noticeably worsened" when the patch grid is coarsened. From 8×88\times8 to 2×22\times2, the ED on dHCP increases from 3.79 to 4.26 mm, i.e., by only 0.47 mm, which is less than half of the in-plane pixel size (1 mm) and far below the slice thickness (2.5–3.5 mm). The small PSNR drop of 0.32 dB therefore corresponds to a small change in the estimated motion, and does not indicate that the reconstruction is insensitive to it.

To validate that the reconstruction of GaLA-SVR does depend on the slice motion, we conducted a perturbation experiment in which the motion error is varied over a much larger range. On the FeTA test set, we feed the same input slices to the network, but replace the slice motions estimated by the network with artificially perturbed ones before the PSF-aware aggregation. The perturbed motions are obtained by adding an independent random error to the ground-truth motion of each slice (a translation of LL mm plus a rotation of L/35.9L/35.9 rad about the slice center). As shown in Table R2, the PSNR decreases monotonically from 30.24 dB with the ground-truth motions to 29.32, 26.89, and 22.34 dB at L=2L=2, 4, and 8 mm. If the decoder mainly generated plausible anatomy from a learned prior, the PSNR would remain nearly unchanged. Instead, an increase of the ED from 3.54 to 6.95 mm costs 2.43 dB, i.e., about 0.7 dB per mm, which is consistent with the 0.32 dB drop observed for the 0.47 mm increase in the patch-grid ablation. This indicates that GaLA-SVR did not use anatomical priors to hallucinate the reconstructed volume.

Table R2. Reconstruction of GaLA-SVR on the FeTA test set when the slice transformations passed to the aggregation are replaced. In-brain disp. is the mean displacement of the pixels inside the brain. Values are mean ± standard deviation.

Slice transformationsED (mm)In-brain disp. (mm)PSNR ↑SSIM ↑
GT0.000.0030.24 ± 2.100.888 ± 0.007
GT + 2 mm3.54 ± 0.052.15 ± 0.0329.32 ± 2.190.884 ± 0.010
GT + 4 mm6.95 ± 0.204.32 ± 0.0626.89 ± 2.470.862 ± 0.020
GT + 8 mm14.17 ± 0.328.58 ± 0.1822.34 ± 2.270.814 ± 0.030
GaLA-SVR (estimated)6.10 ± 2.783.07 ± 1.2929.11 ± 2.080.881 ± 0.011

(2) The motion error on FeTA. We want to clarify that ED is not the displacement of brain tissue. ED is the mean distance between three anchor points per slice under the estimated and ground-truth transformations: the slice center and the two corners of one edge of the 128×128128\times128 mm slice, same as SVoRT [Ref 3]. The anchor points at the edge are about 90 mm away from the slice center and far outside the fetal brain. A small rotation error therefore already produces a large ED. If we instead average the displacement over the pixels inside the brain, the error is 3.07 mm, which is comparable to the slice thickness of 2.5–3.5 mm. Therefore, the lead of the PSNR is reasonable, not because of potential hallucination.

(3) Anatomical prior. To demonstrate that GaLA-SVR makes faithful use of the input rather than relying on a learned anatomical prior, we compare it on the 76 dHCP test subjects with two prior-only predictions, which do not use the input slices at all. The first is the mean of the 675 training volumes, which is the same for every test subject. The second is, for each test subject, the training volume that is closest to its ground-truth volume, i.e., the best result that could be obtained by retrieving a brain from the training set. As shown in Table R3, they reach only 24.46 and 23.25 dB, whereas GaLA-SVR reaches 29.47 dB and is better on all 76 subjects. The performance therefore cannot be explained by a learned anatomical prior alone.

Table R3. Comparison with predictors that do not use the input slices on the 76 dHCP test subjects. Values are mean ± standard deviation.

PredictionPSNR ↑SSIM ↑NRMSE ↓
Mean of the training volumes24.46 ± 1.020.810 ± 0.0110.345 ± 0.035
Most similar training volume23.25 ± 0.750.798 ± 0.0080.396 ± 0.031
GaLA-SVR29.47 ± 0.870.883 ± 0.0070.194 ± 0.022

[Ref3] Xu, Junshen, et al. "SVoRT: Iterative transformer for slice-to-volume registration in fetal brain MRI." MICCAI 2022

AuthorsOpenReview ↗

To reviewer Cu9a [Part III]

W3: Baseline evaluation concern

We would like to first clarify why NeSVoR, GaussianSVR, and GSVR were not included originally. GaLA-SVR is a feed-forward method that estimates the slice transformations and reconstructs the volume in an end-to-end fashion within a second. In contrast, NeSVoR, GaussianSVR, and GSVR are optimized separately for each subject, take orders of magnitude longer, and require extra initializations of slice transformations provided by an external pretrained SVoRT method. We therefore considered learning-based baselines with similar computational requirements to be a fairer comparison, with only one optimization-based method, SVRTK, as reference. Nevertheless, following the suggestion, we added these three methods and the recent cSVR [Ref1], which combines a feed-forward motion estimator with an optimization-based reconstruction, to all comparisons. Their results are summarized in Table R4 and Table R5.

On CamCAN and dHCP, where sufficient training subjects are available, GaLA-SVR remains the best on all three metrics. Its PSNR gains over the strongest added method on each dataset, 0.91 dB over cSVR on CamCAN and 0.45 dB over NeSVoR on dHCP, are statistically significant (P<0.01), while being orders of magnitude faster than per-subject optimization. On FeTA, because there are only 65 training subjects, the results are likely limited by the data size. All learning-based methods are trained from scratch on the same stacks and are affected in the same way. NeSVoR and GaussianSVR, which benefit from per-subject optimization with an already good initialization, achieve the highest PSNR (31.85 and 31.31 dB) on the FeTA dataset. The reconstruction of cSVR is optimization-based, so it is likely less affected by the limited training data. We have also incorporated these results and discussion in the revised manuscript.

Table R4. Quantitative comparison of reconstruction performance on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. Values are mean ± standard deviation. Best results are in bold and second-best in italics.

MethodFeTA PSNR ↑FeTA SSIM ↑FeTA NRMSE ↓CamCAN PSNR ↑CamCAN SSIM ↑CamCAN NRMSE ↓dHCP PSNR ↑dHCP SSIM ↑dHCP NRMSE ↓
Optimization-based
NeSVoR31.85 ± 4.150.890 ± 0.0340.167 ± 0.11125.94 ± 1.360.864 ± 0.0090.183 ± 0.03429.02 ± 0.940.878 ± 0.0080.204 ± 0.028
GaussianSVR31.31 ± 2.830.889 ± 0.0090.166 ± 0.05325.02 ± 2.200.842 ± 0.0410.224 ± 0.06527.55 ± 1.560.859 ± 0.0150.245 ± 0.045
GSVR25.65 ± 2.260.716 ± 0.0750.317 ± 0.10023.40 ± 2.760.782 ± 0.0590.275 ± 0.09127.85 ± 0.730.787 ± 0.0810.226 ± 0.031
Learning-based
cSVR29.12 ± 2.430.879 ± 0.0130.206 ± 0.04026.27 ± 0.860.861 ± 0.0070.182 ± 0.02328.21 ± 0.750.872 ± 0.0070.225 ± 0.023
GaLA-SVR29.11 ± 2.080.881 ± 0.0110.198 ± 0.02127.18 ± 1.210.873 ± 0.0150.171 ± 0.02329.47 ± 0.870.883 ± 0.0070.194 ± 0.022

Table R5. Quantitative comparison of transformation errors and inference time on FeTA, CamCAN, and dHCP for the added methods and GaLA-SVR. ED and GD are mean ± standard deviation; inference time is the mean runtime per subject. Best transformation results are in bold and second-best in italics.

MethodFeTA ED (mm) ↓FeTA GD (rad) ↓FeTA Inf. (s) ↓CamCAN ED (mm) ↓CamCAN GD (rad) ↓CamCAN Inf. (s) ↓dHCP ED (mm) ↓dHCP GD (rad) ↓dHCP Inf. (s) ↓
Optimization-based
NeSVoR5.96 ± 4.440.098 ± 0.082115.496.37 ± 5.910.077 ± 0.079146.803.98 ± 1.770.068 ± 0.038106.42
GaussianSVR3.35 ± 2.190.061 ± 0.043647.336.33 ± 5.530.079 ± 0.0741443.543.85 ± 2.040.073 ± 0.037805.60
GSVR11.38 ± 7.150.205 ± 0.13678.149.16 ± 6.310.118 ± 0.086318.684.35 ± 1.770.079 ± 0.029165.26
Learning-based
cSVR6.31 ± 2.920.113 ± 0.0584.444.56 ± 1.130.054 ± 0.0167.635.23 ± 4.010.084 ± 0.0815.30
GaLA-SVR6.10 ± 2.780.110 ± 0.0540.113.22 ± 1.040.041 ± 0.0140.273.79 ± 3.620.066 ± 0.0790.17

[Ref1] Xu, Junshen, et al. "cSVR: Convolutional Slice-to-Volume Reconstruction." MICCAI PIPPI Workshop 2026

AuthorsOpenReview ↗

To reviewer Cu9a [Part IV]

Q1: Hallucination concern

Our response to W1 has shown that GaLA-SVR generalizes to real motion-corrupted neonatal scans, with the reconstructed anatomy consistent with the official dHCP reconstruction and no structure appearing that is absent from the reference (Table R1 and Figure 8 in the Appendix of the revised manuscript). To examine this more directly, we further tested whether GaLA-SVR reconstructs details that are present in the input slices but never appear in the training data. We inserted synthetic structures into the ground-truth volumes of dHCP test subjects (a hypointense sphere of radius 3 mm or 8 mm) and re-simulated the stacks with the original motion, slice thickness, and noise. Figure 4 in the Appendix of the revised manuscript shows that GaLA-SVR reconstructs the inserted structures at the correct location and with a similar extent, and that the error outside the structure remains at the same level as in the surrounding tissue. This indicates that the reconstruction follows the anatomy of the individual subject rather than a generalized brain structure, and that the presence of an atypical structure does not lead to hallucinations in the rest of the brain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to All Reviewers

We thank the Area Chair and reviewers for their careful assessment and constructive feedback. We are encouraged that the reviewers recognize the clear motivation of fusing learned slice features directly in a 3D latent space instead of assembling an intermediate intensity volume (Cu9a, Tpyu), the principled motion estimation design built on dense canonical patch coordinates, IISA blocks, and the analytical Planar Kabsch solver (Cu9a, Tpyu), and the end-to-end single-pass pipeline with clear and consistent improvements across fetal, neonatal, and adult cohorts (Cu9a, Tpyu, NQ7t). We have addressed their main concerns through additional experiments, methodological clarifications, and manuscript revisions:

  • Evaluation on real motion-corrupted data: We add a zero-shot evaluation on real motion-corrupted neonatal scans from 30 dHCP subjects, with all learning-based models trained only on simulated stacks. GaLA-SVR achieves the best PSNR, SSIM, and NRMSE among all learning-based and optimization-based methods (21.00 dB, 0.761, 0.507), while reconstructing a subject in 0.56 s (Reviewers Cu9a and NQ7t).
  • Faithfulness to the input and hallucination: We provide three analyses showing that the reconstruction is driven by the input slices rather than a learned anatomical prior: a motion-perturbation experiment in which PSNR decreases monotonically with the injected motion error, a comparison with prior-only predictors that GaLA-SVR outperforms by more than 5 dB on all 76 dHCP test subjects, and a synthetic-structure experiment in which unseen structures are reconstructed at the correct location without introducing errors elsewhere. We also clarify that ED is measured at anchor points far outside the brain; the in-brain displacement on FeTA is 3.07 mm, comparable to the slice thickness (Reviewers Cu9a and NQ7t).
  • Comparison with recent methods: We add NeSVoR, GaussianSVR, GSVR, and cSVR to all comparisons. GaLA-SVR remains the best on all reconstruction metrics on CamCAN and dHCP, with statistically significant PSNR gains (P<0.01), while being orders of magnitude faster than per-subject optimization. On FeTA, where only 65 training subjects are available, NeSVoR and GaussianSVR achieve higher PSNR, which we report and discuss in the revised manuscript (Reviewers Cu9a, Tpyu, and NQ7t).
  • Design choices and component contributions: We clarify how our PSF-aware feature aggregation differs from the PSF-based slice rendering in GSVR and add sensitivity analyses on the atlas dilation range and the latent grid resolution. We further add controlled ablations showing that feature-level aggregation outperforms three intensity-level variants by 1.01–1.98 dB and that the Gaussian decoder improves over a voxel decoder under identical settings (Reviewers Tpyu and NQ7t).

We appreciate Reviewer NQ7t's follow-up engagement during the discussion, which prompted the two additional ablations on feature-level aggregation and the Gaussian representation. All new results and discussions have been incorporated into the revised manuscript (Section 4.2 and Appendix C). We hope these additions resolve the reviewers' concerns, and we remain happy to provide any further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.