acceptodds
Under review as a conference paper at ICLR 2027

RIVER: Relational 4D Gaussian Reconstruction from Ultra-Sparse Dynamic Projections

Abstract

Ultra-sparse dynamic digital subtraction angiography (DSA) couples angular undersampling with temporal contrast variation. Representative radiative Gaussian models predict attenuation through per-primitive queries, leaving vascular neighborhood dependencies implicit. We introduce RIVER, a relational 4D Gaussian representation in which neighboring primitives explicitly condition the forward attenuation model. A coverage graph supports broad message propagation, while a confidence graph selects structurally supported relations for regularization. Directional-state stabilization and transport-inspired, segment, and tubular priors complement this relational representation without interpreting its latent states as calibrated hemodynamic quantities. Across ten private clinical acquisitions and two public acquisitions, RIVER improves projection fidelity most clearly under severe sparsity: at ten views, private-cohort PSNR increases from 29.76 to 31.40 dB over the strongest compared baseline. On the private cohort, registered 3D evaluation yields lower Hausdorff distances, while CD, clDice, and projection-domain perceptual gains are not uniformly improved. Capacity, self-context, neighbor-context, and priors-only controls distinguish the tested contributions of forward communication and auxiliary regularization. Mechanism diagnostics further show that broad message coverage coexists with selective structural participation. These results support explicit inter-primitive conditioning as a representation bias for ultra-sparse dynamic vascular reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Relational Learning, 4D Gaussian Reconstruction, Sparse-View Tomography, Dynamic Medical Imaging

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pJJ9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RIVER targets sparse-view dynamic digital subtraction angiography. Geometry (pi,Σi)(p_i,\Sigma_i) is time-invariant and only per-primitive attenuation ai(t)a_i(t) varies, rendered by a differentiable X-ray projection. Relative to a 4DRGS-style predictor that queries (H3(pˉi),H4(pˉi,t))(H_3(\bar p_i),H_4(\bar p_i,t)), RIVER appends a 32-d relation vector hih_i produced by one round of degree-normalized message passing over K=4K=4 nearest neighbors; hih_i carries no time input. Two graphs share the primitive support: a coverage graph for aggregation and a confidence graph (K=12K=12) whose gated edges define vessel segments.

Critical strengths

The contribution is decomposed rather than reported as a single full-versus-base comparison. Table 4 isolates four effects -- capacity, self-processing, neighbor communication, and priors -- and includes a capacity-matched independent control, so the part of the gain attributable to each design choice is visible; the text also states directly that neighbor context alone does not improve CD over 4DRGS. Metrics on which the method does not lead (LPIPS at 20 and 30 views, CD at 10 views, clDice against R$^2$-Gaussian at all four settings) are reported in the tables, and the normalization and Otsu protocols are given in Section 4 with their scope stated.

The 10-view operating point is a real result: PSNR 31.40 against 29.76 for the strongest baseline, SSIM best at all four view counts, and HD best at all four 3D settings.

Critical weaknesses

Only FDK is used as a classical baseline, although TV, SART and learned primal-dual reconstruction are cited in the related work and are the standard sparse-view alternatives. Together with per-image min--max normalization and per-case Otsu thresholding, the reported metrics measure normalized fidelity rather than absolute attenuation; the limitations section states this, but no metric for it is given.

The main gain comes from capacity rather than from the relational mechanism, by the authors' own control. In Table 4 the 10-view PSNR is 28.76 for 4DRGS, 30.12 for the capacity-matched independent predictor, and 31.18 with neighbor context: 1.36 of the 2.64 dB is capacity, and the part attributable to neighborhood conditioning is 1.06 dB. The margin quoted against 4DRGS therefore includes a component unrelated to the proposed mechanism.

Questions for authors

Would TV, SART and a learned primal-dual reconstruction be feasible at 10 views? If so, a comparison would place the method in the standard sparse-view setting. Since images are min--max normalized and volumes Otsu-thresholded, an error metric on absolute attenuation, or an explicit statement that calibrated recovery is out of scope, would also help.

The confidence graph is gated by vv, and vv is itself optimized under the smoothness, floor and residual terms. How much of the segment structure would remain if vv were held at its initialization, or if edges were taken directly from the images?

This topic lies outside my primary area of expertise, so my reading of the reconstruction literature may be incomplete. I will weigh the response together with the other reviews before finalizing my rating.

5 comments
AuthorsOpenReview ↗

Rebuttals (1/3)

Q1: Would TV, SART and a learned primal–dual reconstruction be feasible at 10 views? If so, a comparison would place the method in the standard sparse-view setting. Since images are min–max normalized and volumes Otsu-thresholded, an error metric on absolute attenuation, or an explicit statement that calibrated recovery is out of scope, would also help.

Response.

Thanks for the suggestion.

  • We add SART and TV-iTV as classical sparse-view baselines and retain 4DRGS as a dynamic DSA baseline. All methods use the same ten input projections from a single case and are evaluated against the same 234 held-out measured projections. SART and TV-iTV reconstruct static volumes, while 4DRGS and RIVER model time-varying attenuation. As shown in Table 13, RIVER achieves the highest PSNR and SSIM and the lowest raw RMSE and MAE among the compared methods. These results demonstrate improved held-out projection fidelity over both the evaluated classical methods and 4DRGS in this ten-view example.

Table 13. Ten-view reconstruction comparison on a single case, evaluated against 234 held-out measured DSA projections.

MethodPSNR↑\uparrowSSIM↑\uparrowRaw RMSE↓\downarrowRaw MAE↓\downarrow
SART22.2330.27480.141800.09141
TV-iTV22.2340.27490.141780.09140
4DRGS27.7520.83980.078160.02529
RIVER29.2470.86230.066250.02244

Learned primal–dual reconstruction is feasible in principle, but its original supervised formulation requires paired training data. We currently lack an independent training set matched to dynamic DSA, and the synthetic benchmark is reserved for evaluation. SART, TV-iTV, and 4DRGS can operate on the available acquisition without a separate supervised training dataset.

  • As for the physical calibration and intensity-scale evaluation, physically calibrated attenuation recovery is outside the scope of the current evaluation. The available data do not provide an independently calibrated attenuation reference, so we cannot report a validated reconstruction error in physical units such as mm−1\mathrm{mm}^{-1}.

To assess intensity agreement without per-image normalization, we additionally report raw RMSE and MAE in Table 13. RIVER improves both metrics, showing that its projection-domain advantage also holds on the dataset’s original logarithmic DSA intensity scale. These measurements assess intensity fidelity on that scale.

AuthorsOpenReview ↗

Rebuttals (2/3)

Q2: The confidence graph is gated by the directional states vi\mathbf v_i, which are themselves optimized under the smoothness, floor and residual terms. How much of the segment structure remains if vi\mathbf v_i is held at initialization, or if edges are taken directly from the images?

Response.

Thanks for the reviewer’s suggestions. The clinical dataset does not provide independently annotated branch identities and junction labels for evaluating graph connectivity. We therefore construct the synthetic vascular benchmark described in our response to Reviewer e48r, Question 1. It contains two explicitly disconnected vascular trees with ten branches and four labelled junctions, providing known references for assessing erroneous connections and structural preservation. These labels are used only for evaluation. We evaluate two controls corresponding to the alternatives raised by the reviewer. The frozen-state control fixes vi\mathbf v_i at its PCA initialization, including magnitude, to assess how much structure remains without learning the directional states. The image-derived affinity control uses image support to weight candidate edges while keeping the directional states trainable, testing whether image evidence can replace the learned affinity. All variants share the same candidate-edge construction rule, input projections, reconstruction budget, and evaluation protocol.

We evaluate the internal graphs from two perspectives: cross-component connection errors and preservation of the known junctions. Primitive centers are assigned to the nearest labelled reference branch within a tolerance of 0.50.5 mm. Among retained edges whose endpoints are both mapped, connections between the two disconnected reference trees are counted as false edges. The false-edge rate is their number divided by the total number of mapped retained edges. A labelled junction is counted as preserved when an eligible mapped connection exists within its 33 mm junction window. Table 14 reports these measurements at three edge-weight thresholds.

Table 14. Confidence-graph comparison on the known-topology synthetic vascular benchmark.

qqConfigurationMapped retained edgesFalse edgesFalse-edge rate (%)Junctions preserved
0.10RIVER25,970700.26954/4
0.10Frozen states33,7461980.58674/4
0.10Image-derived graph47,9223420.71374/4
0.25RIVER25,970700.26954/4
0.25Frozen states32,1381880.58504/4
0.25Image-derived graph9,95800.00002/4
0.50RIVER25,970700.26954/4
0.50Frozen states26,1561460.55824/4
0.50Image-derived graph00–0/4

At the operating threshold q=0.50q=0.50, the frozen-state control retains 26,156 mapped strong edges and all four junctions, showing that the measured junction structure remains preserved with the initial directional states. Learning these states improves connection selectivity: RIVER retains all four junctions while reducing false cross-component connections from 146 to 70 and the false-edge rate from 0.5582% to 0.2695%. The image-derived graph preserves all four junctions at q=0.10q=0.10, but only two at q=0.25q=0.25 and none at q=0.50q=0.50. Thus, RIVER maintains junction preservation across the evaluated thresholds while suppressing erroneous connections.

We additionally evaluate the final reconstructed vessel masks using clDice. RIVER achieves 0.3209, compared with 0.2968 for frozen states and 0.2975 for the image-derived graph. This provides complementary evidence that the learned configuration improves centerline agreement in the final reconstruction, alongside improved selectivity of the internal confidence graph.

AuthorsOpenReview ↗

Rebuttals (3/3)

Q3: The main gain comes from capacity rather than from the relational mechanism: 4DRGS gives 28.76 dB, the capacity-matched independent predictor 30.12 dB, and neighbor context 31.18 dB. The margin quoted against 4DRGS therefore includes a component unrelated to the proposed mechanism.

Response.

We agree that increased capacity contributes substantially to the overall improvement. The capacity-matched independent predictor improves PSNR by 1.36 dB over 4DRGS, so the full 2.64 dB RIVER–4DRGS margin cannot be attributed entirely to neighborhood conditioning. We assess the relational contribution through the controlled comparisons in Table 15.

Table 15. Ten-view cohort ablation separating model capacity, self-processing, neighborhood conditioning, and auxiliary priors.

VariantPSNR↑\uparrowSSIM↑\uparrowCD (mm)↓\downarrowclDice↑\uparrow
4DRGS28.760.834.030.339
Capacity-matched independent30.120.854.480.386
Self context30.670.864.310.401
Neighbor context, no priors31.180.864.260.409
Priors only, no neighbor context30.940.864.190.421
Full RIVER31.400.874.070.433

At matched capacity, neighbor context improves PSNR from 30.12 to 31.18 dB, demonstrating an additional benefit beyond increasing predictor capacity. The comparison with self context further distinguishes neighbor communication from processing each primitive’s own information: introducing neighboring information raises PSNR from 30.67 to 31.18 dB. Finally, Full RIVER improves over the priors-only configuration from 30.94 to 31.40 dB, showing that neighborhood conditioning remains useful when the auxiliary priors are already present. These comparisons support an incremental relational contribution without assigning the entire baseline margin to that mechanism.

The structural results also distinguish the contribution of neighborhood conditioning from that of the full design. Neighbor context alone improves clDice over 4DRGS but has a higher CD. Full RIVER achieves the highest clDice in the comparison, increasing it from 0.339 to 0.433, while its CD of 4.07 mm is close to the 4DRGS value of 4.03 mm. These results support improved centerline agreement from the combined configuration.

Reviewer_ZjZ2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RIVER, a framework for reconstructing 3D vessels from sparse-view DSA coupling time and angles. The proposed method incorporates neighboring query information into the 4DRGS method. The proposed methods were evaluated on private and public datasets in sparse-view settings. The authors reported view synthesis and reconstruction performance.

Critical strengths
  • The proposed method is sophisticated and adheres to the motivation of incorporating neighboring information.
  • The method achieves improvements over previous methods under all sparse-view settings.
Critical weaknesses
  • The proposed method incorporates four auxiliary loss functions and over nine hyperparmeters but yet results in marginal improvments over existing methods. The complexity of the method and marginal gains limits the impact of the proposed method.
  • The SSIM loss has already penalized local windows. The advantage of the proposed method over the SSIM loss is unjustified.
Questions for authors
  1. Neighboring information helps determine topological structure. But neighboring gaussian kernels have different attenuation under the same time point. This can be a factor that limits and conflates the methodological designs as time-varying attenuation is one of the key challenges for DSA reconstruction. The authors should provide clarifications on this point.

  2. The proposed method incorporates a photo loss and four auxiliary losses for neighboring constraints. However, the improvements are within or around the standard deviation of the second-best methods under 10, 15, 20, and 30 input views, as shown in Table 1. This suggests that neighboring information might not be a key limiting factor of existing methods. The SSIM loss penalizes local grids or neighboring structure. What is the advantage of the proposed method over the SSIM penalization. The authors should provide discussion on the above points.

  3. The proposed method consists of at least 9 hyperparameters (section C.2) to build affinity. The tradeoff between the methodological complexity and the marginal improvements is worth discussing. The complexity of hyperparameter selection can limit the reproductivity of the proposed method. The authors should provide sensitivity analysis or suggestions to ease hyperparameter selection.

In sum, the proposed method is motivated by neighboring relation. However, neighboring information might not be a critical limiting factor of existing methods, or it is already penalized by the SSIM loss. The authors should provide evidence or discussion to justify their innovation angle.

5 comments
AuthorsOpenReview ↗

Rebuttals (1/3)

Q1: Neighboring information helps determine topological structure, but neighboring Gaussian kernels can have different attenuation at the same time point. Could the neighborhood design conflate structural context with time-varying attenuation?

Response.

We agree that neighboring Gaussians can have different attenuation at the same time point. RIVER’s neighborhood design accommodates this difference because structural information sharing does not require attenuation agreement. Specifically, the message from primitive jj to primitive ii contains the sender position, directional state, and relative displacement. It does not average or copy the sender’s attenuation amplitude. Neighboring primitives can therefore exchange structural context while retaining different attenuation values.

Each primitive predicts its attenuation through its own spatial and temporal query:

a_i(\tau)=\sigma\left(f_\theta\left[H_3(\bar{\mathbf p}_i),H_4(\bar{\mathbf p}_i,\tau),\mathbf h_i\right]\right).\tag{2}

Although the predictor fθf_\theta is shared, its spatial encoding, spatiotemporal encoding, and neighborhood context are evaluated for each primitive. Thus, two adjacent Gaussians queried at the same τ\tau can produce different attenuation values and temporal curves. The neighborhood feature hi\mathbf h_i conditions this prediction without replacing the primitive-specific query, and no hard constraint imposes ai(τ)=aj(τ)a_i(\tau)=a_j(\tau).

Structural context can nevertheless influence the learned attenuation through joint optimization. The architectural distinction therefore explains how heterogeneous attenuation is accommodated, while its recovery quality is evaluated separately. We use the known-dynamics synthetic branch evaluation reported in our response to Reviewer e48r, Question 1, because the clinical sequences lack an independent 4D attenuation reference.

In that evaluation, RIVER obtains slightly lower branch and inter-branch difference errors than the internal controls. However, its sampled peak ratio of 0.1281 and inter-branch difference NRMSE of 0.9364 show that amplitude and heterogeneity recovery remain incomplete. These results characterize the remaining reconstruction errors; they do not isolate their cause as neighborhood conditioning. Our design therefore provides structural context while preserving primitive-specific attenuation prediction, without assuming that neighboring Gaussians must share the same attenuation.

AuthorsOpenReview ↗

Rebuttals (2/3)

Q2: What is the advantage of the proposed neighborhood design over SSIM penalization, and are the observed gains larger than the variability of the second-best method?

Response.

Thanks for the reviewer’s suggestion.

  • In SSIM penalization, SSIM constrains local agreement in rendered 2D projections, whereas RIVER conditions each primitive’s attenuation prediction on explicit 3D neighborhood information. Primitives contributing to the same detector window can change with viewpoint, and anatomically disconnected vessels can overlap in projection. RIVER instead aggregates neighboring primitives’ spatial and directional information in 3D before rendering. This provides structural context that is not explicitly encoded by the SSIM loss.

Table 7 evaluates this additional contribution under the same photometric objective, including SSIM. Neighbor context improves PSNR over both the capacity-matched independent predictor and the self-context variant, by 1.06 and 0.51 dB, respectively. Full RIVER also improves PSNR by 0.46 dB over the priors-only variant. These comparisons show that neighborhood information provides an incremental benefit when SSIM is already used, beyond increased capacity, processing of each primitive’s own features, and the auxiliary priors.

Table 7. Matched ten-case, ten-view ablation of photometric and relational components. The structural quantities used elsewhere are registration-sensitive proxies; PSNR and SSIM are the primary image-domain comparison here.

VariantPSNR$\uparrow$SSIM$\uparrow$
4DRGS28.760.83
Capacity-matched independent30.120.85
Self context30.670.86
Neighbor context, no priors31.180.86
Priors only, no neighbor context30.940.86
Full RIVER31.400.87

Table 8 directly compares neighborhood conditioning with and without the SSIM loss. At SSIM weights of 0 and 0.2, enabling neighborhood conditioning improves PSNR by 1.18 and 0.46 dB, respectively, accompanied by higher SSIM and lower raw RMSE and temporal-change NRMSE. The reported improvements therefore occur both without SSIM and in addition to SSIM. This supports the value of explicit 3D relational context beyond the local image agreement encouraged by SSIM.

Table 8. Provisional factorial control separating relational context from the SSIM coefficient.

VariantGraphλDSSIM\lambda_{\rm DSSIM}PSNR$\uparrow$SSIM$\uparrow$Raw RMSE$\downarrow$Δ\Delta-NRMSE$\downarrow$
Priors only, no graphoff0.030.30 ±\pm 0.430.85 ±\pm 0.040.079 ±\pm 0.0070.054 ±\pm 0.010
Priors only, no graphoff0.230.94 ±\pm 0.730.86 ±\pm 0.030.081 ±\pm 0.0070.057 ±\pm 0.010
Neighbor graph, no SSIMon0.031.48 ±\pm 0.390.86 ±\pm 0.040.068 ±\pm 0.0060.046 ±\pm 0.008
Full RIVERon0.231.40 ±\pm 0.620.87 ±\pm 0.020.070 ±\pm 0.0060.048 ±\pm 0.007
  • Regarding variability, Table 9 shows that the mean PSNR gain exceeds the strongest baseline’s between-case standard deviation at 10 views. At 10 views, RIVER improves PSNR by 1.644 dB, compared with the baseline standard deviation of 0.505 dB. At 15, 20, and 30 views, the gains are 0.571, 0.353, and 0.301 dB, respectively, compared with standard deviations of 0.601, 0.849, and 0.902 dB. RIVER therefore achieves higher cohort-mean PSNR across all four settings, with the largest improvement relative to baseline variability at the sparsest, ten-view setting.

Table 9. Reported PSNR (dB; mean ±\pm standard deviation across the ten cases). The strongest baseline is selected by its cohort-mean PSNR at each view count. Δ\Delta denotes the difference between the reported RIVER and baseline means. Boldface marks the higher mean PSNR in each row. Standard deviations describe between-case variation.

ViewsStrongest baselineBaseline PSNRRIVER PSNRΔ\Delta PSNR
10TOGS29.756±0.50529.756\pm0.50531.400±0.620\mathbf{31.400}\pm0.620+1.644+1.644
15TOGS32.639±0.60132.639\pm0.60133.210±0.710\mathbf{33.210}\pm0.710+0.571+0.571
20TOGS33.957±0.84933.957\pm0.84934.310±0.770\mathbf{34.310}\pm0.770+0.353+0.353
304DRGS35.405±0.90235.405\pm0.90235.706±0.857\mathbf{35.706}\pm0.857+0.301+0.301
AuthorsOpenReview ↗

Rebuttals (3/3)

Q3: The proposed method consists of at least 9 hyperparameters to build affinity. The tradeoff between methodological complexity and marginal improvements is worth discussing. The complexity of hyperparameter selection can limit reproducibility. Could the authors provide sensitivity analysis or suggestions to ease hyperparameter selection?

Response.

Thank you for the reviewer’s suggestion. We address the accuracy–complexity trade-off through computational-cost comparisons and component ablations, and provide radius sensitivity results together with practical guidance for selecting the remaining parameters.

  • For the methodological complexity and reconstruction gains, the benefit of the full design extends beyond projection quality to centerline agreement. As shown in Table 10, Full RIVER increases clDice from 0.449 for 4DRGS to 0.567 on a single case. Compared with neighbor context alone, clDice increases from 0.313 to 0.567. These substantial clDice gains demonstrate the contribution of the full configuration to centerline overlap with the registered reference. The additional accuracy comes with increased computational cost: Full RIVER takes 104.91 minutes, compared with 15.62 minutes for 4DRGS and 40.26 minutes for neighbor context alone. This comparison quantifies the accuracy–cost trade-off of adding the auxiliary priors to the neighborhood-based model. The matched ablations in Table 7, discussed in our response to Question 2, further demonstrate the incremental contributions of neighborhood conditioning and auxiliary priors under the same photometric objective.

Table 10. Accuracy and computational cost on a single case with 10 input views.

VariantPSNR↑\uparrowRaw RMSE↓\downarrowclDice↑\uparrowTime(min)
4DRGS27.7520.078160.44915.62
Neighbor context, no priors27.5880.079170.31340.26
Full RIVER29.2470.066250.567104.91
  • For the parameter sensitivity, we first examine the neighborhood radius, which determines the spatial extent of candidate connections. Table 11 compares radii of 2.0 and 3.0 mm on the case. The two settings yield similar projection and temporal-change errors: 2.0 mm provides slightly better projection fidelity, whereas the reference radius of 3.0 mm achieves higher clDice. This comparison illustrates the balance between local projection accuracy and centerline agreement when selecting the neighborhood extent.

Table 11. Neighborhood-radius sensitivity on a single case with 10 input views.

rr (mm)PSNR↑\uparrowRaw RMSE↓\downarrowΔ\Delta-NRMSE↓\downarrowclDice↑\uparrow
2.029.4320.065070.041350.545
3.029.2470.066250.041830.567
  • For the remaining parameters, Table 12 gives users a practical starting point and a clear order for parameter adjustment.

Table 12. Reference parameters and practical selection guidance.

GroupReference valuesSelection guidance
Distancer=3.0r=3.0 mm, sr=0.5s_r=0.5Start with the reference setting and relate the neighborhood extent to primitive spacing.
Axial affinityτo=0.7\tau_o=0.7, so=0.1s_o=0.1Adjust the threshold and softness together if directional selectivity needs to change.
Magnitude affinityτv=0.05\tau_v=0.05, sv=0.02s_v=0.02Consider jointly with the directional-state scale and magnitude floor.
Coverage controlsϵc=0.1\epsilon_c=0.1, ϵw=10−4\epsilon_w=10^{-4}Use the reference coverage floor initially; keep the numerical safeguard fixed.
Context and graphαg=0.25\alpha_g=0.25, Dh=32D_h=32, Kc=4K_c=4, Ks=12K_s=12, strong-edge threshold 0.50.5Start from the reference architecture and keep its structural constants fixed when adjusting affinity parameters.
Prior weightsλpde=0.01\lambda_{pde}=0.01, λgv=0.05\lambda_{gv}=0.05, λseg=0.005\lambda_{seg}=0.005, λtube=0.005\lambda_{tube}=0.005Start from the reference weights; assess any adjustment using validation projection fidelity and available structural criteria.
State floorvf=0.08v_f=0.08Adjust jointly with magnitude affinity.
Reviewer_e48rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

RIVER reconstructs a dynamic vascular attenuation field from a small number of DSA projections with known camera geometry and acquisition times. The method uses Gaussians with fixed geometry and time-varying amplitudes to represent the vessels. It incorporates neighborhood information through a coverage graph for message aggregation and a confidence graph for structural regularization. Additional constraints, including directional state stabilization, transport, segment, and tubular priors, are introduced to improve reconstruction. Each acquisition is optimized independently.

Overall, the method is well motivated, and the experiments demonstrate the benefits of incorporating neighborhood information. While the evaluation is more convincing for projection quality than for the accuracy of the reconstructed 3D structure and temporal dynamics, the results provide reasonable support for the proposed approach. Additional quantitative evaluations and video demonstrations would further strengthen the paper. Considering the overall technical contribution and experimental results, my recommendation is Weak Accept.

Critical strengths
  1. Reasonable method and experimental design. Using neighboring primitives to coordinate attenuation predictions is well motivated under severe projection sparsity, and separating communication from structural regularization is a coherent design choice.

  2. Extensive and informative controls. The experiments distinguish neighborhood information from increased model capacity, additional processing of each primitive, and auxiliary losses. These comparisons provide meaningful support for the central contribution.

  3. Strong results under severe sparsity. At ten input views, private cohort PSNR improves from 29.76 to 31.40 dB over the strongest compared baseline, with gains also observed on both public acquisitions. The qualitative examples show improved projection detail and fewer reconstruction artifacts.

Critical weaknesses
  1. The improvements in 3D structure are less consistent. Projection quality and Hausdorff distance improve, but Chamfer distance and clDice show mixed advantages over baselines, leaving the benefit to overall shape accuracy and connectivity less clear.

  2. Limited direct evaluation of temporal reconstruction. The held-out projections vary in both angle and time, making it difficult to distinguish improvements in spatial reconstruction from improvements in temporal modeling. The predicted trajectories and consistency analyses do not resolve this distinction.

  3. Missing video demonstrations. The paper addresses 4D reconstruction but provides no supplementary videos, making it difficult to assess temporal consistency and the visual quality of the reconstructed dynamics.

Questions for authors
  1. Could the authors report quantitative metrics that assess the temporal quality of the reconstructed sequences and compare the results with baselines?

  2. How does the confidence graph avoid linking spatially close but disconnected structures, particularly around parallel segments and branch points?

  3. The current evaluation starts at 10 input views. Could the authors examine more extreme settings, such as 3 or 6 views, to show whether the gains persist and where reconstruction begins to fail?

6 comments
AuthorsOpenReview ↗

Rebuttals (1/4)

Q1: Could the authors report quantitative metrics that assess the temporal quality of the reconstructed sequences and compare the results with baselines?

Response.

Thank you for the suggestion. To better evaluate the temporal reconstruction quality from changes in viewing angle, we construct a synthetic dynamic vascular benchmark with known reference dynamics. This complements the clinical evaluation, where acquisition angle and time vary together and their effects cannot be directly separated.

  • For the dataset, we analytically generate two vascular trees containing ten segments and four junctions on a 1283128^3 grid with 0.50.5 mm isotropic spacing. Contrast attenuation follows prescribed curves that rise and subsequently decay, with arrival times increasing with distance from each root and different peak attenuations between the trees. An independent LEAP projector generates 96 rotating-view DSA projections, of which ten are used for reconstruction, together with a separate fixed-camera sequence at the same 96 timestamps. The known attenuation fields, topology, and fixed-camera sequence are reserved for evaluation.

  • For the quantitative metrics, we use NRMSE to quantify reconstruction and temporal-curve errors, Pearson correlation to assess agreement in temporal variation, and peak-time error to assess timing accuracy. PSNR and SSIM provide complementary measures of projection fidelity. For direct 3D evaluation, we additionally report branch NRMSE, the predicted-to-reference sampled peak ratio, and pairwise NRMSE and correlation for inter-branch attenuation differences. These metrics distinguish projection agreement, temporal timing, attenuation amplitude, and inter-branch heterogeneity.

  • For the temporal quality, Table 1 assesses reconstructed dynamics directly at the 3D branch level, complementing the projection-based evaluation. RIVER shows a modest advantage over the internal controls in the reported branch-level error measures. These results provide complementary evidence that its reconstruction benefits extend to individual vascular branches in this controlled setting.

Table 1. Branch-resolved 3D attenuation evaluation on the controlled synthetic dynamic benchmark.

MethodBranch NRMSE↓\downarrowSampled peak ratioPair NRMSE↓\downarrowPair correlation r↑r\uparrow
RIVER0.49960.12810.93640.5065
Frozen states0.49980.12720.94230.5157
Image-derived graph0.50160.12450.94560.5119

Table 2 compares actual acquisition timestamps with a frozen temporal query across ten clinical cases. Both conditions evaluate the same model, differing only in whether the temporal query follows the acquisition timestamp or remains fixed. Using actual time reduces vessel NRMSE from 0.1872±0.02010.1872 \pm 0.0201 to 0.0800±0.01460.0800 \pm 0.0146 and curve NRMSE from 0.0714±0.02010.0714 \pm 0.0201 to 0.0193±0.00610.0193 \pm 0.0061. This supports the contribution of temporal conditioning to agreement with measured DSA sequences.

Table 2. Clinical comparison between actual and frozen temporal queries on the private cohort.

ConditionFull NRMSE↓\downarrowVessel NRMSE↓\downarrowCurve NRMSE↓\downarrowCorrelation r↑r\uparrow
Actual time0.0356 ±\pm 0.00300.0800 ±\pm 0.01460.0193 ±\pm 0.00610.9776 ±\pm 0.0176
Frozen time0.0648 ±\pm 0.00410.1872 ±\pm 0.02010.0714 ±\pm 0.02010.7024 ±\pm 0.2759
  • For the baseline comparisons, Table 3 shows that RIVER achieves better temporal reconstruction quality than the reported 4DRGS output on the controlled fixed-camera benchmark. Its reconstructed sequences more closely follow the reference temporal curves and successive-frame changes, alongside improved projection fidelity. Comparisons with frozen relational states and an image-derived graph also favor RIVER across the reported metrics, with modest gains over these internal controls, providing complementary support for its effectiveness in reconstructing time-varying vascular projections.

Table 3. Fixed-camera projection and temporal evaluation on the controlled synthetic dynamic vascular benchmark. All metrics use the same reference vascular ROI. Temporal-change NRMSE (EΔE_{\Delta}) compares successive-frame changes and remains defined for constant predictions.

MethodFG NRMSE$\downarrow$Curve NRMSE$\downarrow$EΔ↓E_{\Delta} \downarrowPSNR$\uparrow$SSIM$\uparrow$
RIVER0.261060.161370.02289511.66530.4181
Frozen states0.263190.162180.02296911.59450.4178
Image-derived graph0.264490.163260.02319711.55180.4028
4DRGS0.365990.273990.0261768.73060.3754
AuthorsOpenReview ↗

Rebuttals (2/4)

Q2: The clinical evaluation couples camera angle and acquisition time, so improved held-out projections do not by themselves establish temporal reconstruction accuracy. How is this limitation addressed?

Response.

To address this coupling, we construct a synthetic dynamic vascular benchmark with known reference dynamics and compare the reference and reconstructed projections from the same fixed camera at 96 timestamps (Table 3). This removes viewpoint changes from the temporal projection comparison. We additionally evaluate attenuation directly at the 3D branch level (Table 1) and compare actual versus frozen temporal queries on the clinical sequences (Table 2), with the latter assessing the contribution of temporal conditioning to acquisition consistency.

The benchmark construction, evaluation metrics, quantitative results, and their interpretation are provided in our response to Question 1.

Q3: How does the confidence graph avoid linking spatially close but disconnected structures, particularly around parallel segments and branch points?

Response.

Thanks for the suggestions. We first explain how confidence is assigned to candidate connections, and then evaluate erroneous connections and junction retention in the two situations highlighted by the reviewer.

  • In avoiding linking spatially close but disconnected structures of confidence graph, RIVER does not treat every nearby pair of Gaussians as a vessel connection. It first constructs a local candidate set using a radius and a kk-nearest-neighbour search. For each candidate directed edge (i,j)(i,j), the confidence weight combines three quantities:

w_{ij}=s_{ij}^{\mathrm{spatial}}s_{ij}^{\mathrm{orient}}\left[\rho+(1-\rho)s_{ij}^{\mathrm{speed}}\right],\tag{1}

where sijspatials_{ij}^{\mathrm{spatial}} decreases with Euclidean distance, sijorients_{ij}^{\mathrm{orient}} measures sign-invariant agreement of the learned directional states, and sijspeeds_{ij}^{\mathrm{speed}} measures whether both states have sufficient magnitude. The sign-invariant term uses ∣vi⊤vj∣|\mathbf{v}_i^{\top}\mathbf{v}_j| because the direction vectors have an axis ambiguity.

Spatial proximity therefore determines the candidate set, while the confidence weight additionally accounts for directional agreement and state magnitude. Soft weights are used for message passing, whereas the reported strong graph retains only edges with wij>0.5w_{ij}>0.5. This separates information exchange over uncertain connections from the selection of strong graph edges. These confidence factors reduce the reliance on spatial proximity, but do not provide a hard guarantee of anatomical connectivity.

  • For nearby parallel segments, RIVER applies the confidence mechanism described above to combine spatial proximity, directional agreement, and state magnitude. Candidates with low directional or magnitude-based confidence are downweighted, and only edges with wij>0.5w_{ij}>0.5 enter the reported strong graph. Thus, spatial proximity alone does not determine whether two segments are strongly connected. We evaluate erroneous connections on the known-topology synthetic benchmark containing nearby but disconnected vascular structures and four labelled junctions. We compare RIVER with frozen-state and image-derived-graph controls. All variants use the same input, initialization, optimization settings, and evaluation protocol. We measure the proportion of mapped edges that incorrectly connect different vascular trees, together with the number of retained junctions, at shared confidence thresholds. Table 4 reports the wrong-edge rates and junction retention of the three variants. At q=0.50q=0.50, RIVER has 70 wrong edges among 25,970 mapped edges (0.2695%), compared with 146/26,156 (0.5582%) for frozen states, while both retain all four junctions. This supports reduced erroneous connections without sacrificing measured junction retention in this setting. The image-derived graph retains no mapped strong edge or junction at q=0.50q=0.50; at q=0.25q=0.25, its wrong-edge rate is zero, but only two junctions remain. These findings highlight the importance of evaluating false connections and true-junction preservation together.

Table 4. Post-hoc confidence-graph evaluation on the controlled synthetic vascular benchmark. Each row reports one graph variant across three moderate strong-edge thresholds. The measurements assess internal graph topology, rather than PSNR, SSIM, or final 3D reconstruction quality.

Graph variantWrong-edge rate (%, lower), q=0.10q=0.10Wrong-edge rate (%, lower), q=0.25q=0.25Wrong-edge rate (%, lower), q=0.50q=0.50Junctions kept (of 4, higher), q=0.10q=0.10Junctions kept (of 4, higher), q=0.25q=0.25Junctions kept (of 4, higher), q=0.50q=0.50
RIVER0.26950.26950.26954/44/44/4
Frozen states0.58670.58500.55824/44/44/4
Image-derived graph0.71370.0000–4/42/40/4
AuthorsOpenReview ↗

Rebuttals (3/4)

  • At branch points, RIVER retains a coverage floor for message passing, allowing information exchange between candidate primitives even when their directional agreement is low. Confidence weighting then selects the strong connections. This combines soft neighborhood communication with confidence-based structural filtering around bifurcations.

Table 4 also reports junction retention alongside the cross-component wrong-edge rate. At q=0.50q=0.50, RIVER retains all four labelled junctions with a wrong-edge rate of 0.2695%, compared with 0.5582% for frozen states, which also retains four junctions. The image-derived graph retains only two junctions at q=0.25q=0.25 and none at q=0.50q=0.50. These results show that RIVER reduces erroneous cross-component connections while preserving the measured true junctions.

Q4: The current evaluation starts at 10 input views. Could the authors examine more extreme settings, such as 3 or 6 views, to show whether the gains persist and where reconstruction begins to fail?

Response.

Thanks for the reviewer’s suggestion. We compare RIVER and 4DRGS on a single case with 3 and 6 input views, while retaining the same reconstruction budget and evaluation protocol as the main experiment. Because the clinical sequence has no 3D ground-truth mask, we evaluate held-out measured DSA projections using fixed-scale PSNR, SSIM, raw RMSE, and adjacent-frame change NRMSE.

Table 5. Extreme-sparsity projection and temporal-consistency stress test on single case with 3 and 6 input views. All metrics are computed against held-out measured DSA projections; no independent 3D ground truth is assumed. Boldface marks the better method.

ViewsMethodPSNR$\uparrow$SSIM$\uparrow$Raw RMSE$\downarrow$Change NRMSE$\downarrow$
34DRGS18.29190.64660.210940.06639
3RIVER19.29260.65880.193230.04987
64DRGS23.64230.77040.116740.04283
6RIVER25.69010.79850.094860.03237

At 6 views, RIVER improves all four reported projection and temporal metrics over 4DRGS, the same trend holds at 3 views. The 3-view setting remains a severe stress test, and the conclusion is limited to held-out projection fidelity and temporal consistency because no independent 3D mask is available. These results show that RIVER’s advantages over 4DRGS persist at both 3 and 6 input views.

Q5: The improvements in 3D structure are less consistent. Projection quality and Hausdorff distance improve, but Chamfer distance and clDice show mixed advantages over baselines, leaving the benefit to overall shape accuracy and connectivity less clear.

Response.

Thank you for highlighting the differences across structural metrics.

  • For the overall shape accuracy, table 6 compares CD, HD, and clDice across ten clinical cases under a common registration protocol. The measurements quantify agreement with the historical reference volume and retain the differences between methods across metrics. We compare RIVER with 4DRGS and R$^2$-Gaussian using CD, HD, and clDice under the same evaluation protocol. As shown in Table 6, RIVER achieves lower mean HD at both view counts, with a particularly pronounced reduction relative to R$^2$-Gaussian at 10 views. Its CD is close to that of 4DRGS at 10 views and improves at 20 views. Compared with R$^2$-Gaussian, RIVER obtains lower CD and HD in both settings, while its clDice remains close to that of R$^2$-Gaussian. These results indicate improved surface agreement alongside comparable centerline overlap with the registered reference.

Table 6. Registered-reference morphology agreement on the private cohort. Values are mean ±\pm standard deviation across ten cases.

ViewsMethodCD (mm)↓\downarrowHD (mm)↓\downarrowclDice↑\uparrow
10R$^2$-Gaussian5.96 ±\pm 2.4543.66 ±\pm 13.260.442 ±\pm 0.163
104DRGS4.03 ±\pm 0.6233.11 ±\pm 11.020.339 ±\pm 0.098
10RIVER4.07 ±\pm 0.3531.42 ±\pm 9.290.433 ±\pm 0.154
20R$^2$-Gaussian3.90 ±\pm 0.3034.96 ±\pm 13.060.461 ±\pm 0.139
204DRGS4.07 ±\pm 0.7431.63 ±\pm 5.600.407 ±\pm 0.124
20RIVER3.70 ±\pm 0.2628.49 ±\pm 5.620.445 ±\pm 0.127
  • As for the connectivity, our response to Question 3 provides complementary connectivity evidence using the synthetic benchmark with known component identities and labelled junctions. As reported in Table 4, at q=0.50q=0.50, RIVER reduces the cross-component wrong-edge rate from 0.5582% for frozen states to 0.2695%, while both retain all four reference junctions. This comparison shows that the learned confidence graph reduces erroneous connections without sacrificing measured junction retention at the reported operating point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.