acceptodds
Under review as a conference paper at ICLR 2027

Learning How Much and Where for 3D Vascular Reconstruction from Two Binary Projections

Abstract

Reconstructing sparse three-dimensional (3D) structures from two calibrated binary projections is a severely ill-posed inverse problem. In biplanar angiography, nonuniform contrast makes intensity unreliable for measuring vessel amount, motivating binary masks. These indicate ray–structure intersections but not occupied amounts or their depth distributions. We introduce Biplanar Amount–Allocation Reconstruction (BiAAR) around an explicit ray-wise decomposition: amount measures how much structure a pixel's ray contains; normalized allocation describes where it lies along that ray. Exact factors recover occupancy by multiplication, giving an information-preserving representation under complete calibrated coverage. From paired masks, signed distance fields, and cameras, BiAAR predicts view-normalized amounts and a coarse 3D occupancy prior supervised by the full target volume. This prior supplies allocation guidance; learned fusion and refinement combine both predictions with projection evidence rather than explicitly multiplying normalized factors. BiAAR achieves the highest mean Dice and clDice among evaluated methods on simulated ImageCAS and AVDNet-derived ASOCA projections. Each intermediate prediction improves reconstruction in ablations, with further gains when combined; fixed-capacity controls support their explicit supervision. These results support amount estimation and occupancy-based allocation guidance for two-view binary reconstruction. An anonymized repository containing the code is available at our project page.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, inverse problems, binary projections, sparse structures, geometry-aware learning, vascular reconstruction

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_anyZOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BiAAR reconstructs 3D vessels from two calibrated binary projections. It uses an amount–allocation formulation, supervises view-normalized amount maps and a 3D occupancy prior, and combines them with projection evidence through learned fusion and refinement. Experiments use simulated ImageCAS and AVDNet-derived ASOCA projections. The paper reports the highest mean Dice and soft clDice among the compared methods.

Critical strengths
  1. The ImageCAS gain over AutoCAR is substantial. Dice improves from 0.712 to 0.809 and cIDice from 0.783 to 0.879. HD95 and ASSD also improve.
  2. The fixed-capacity ablation is useful. Adding both auxiliary losses improves Dice from 0.774 to 0.809 without changing the architecture. The oracle amount result of 0.926 Dice also shows the amount branch matters.
Critical weaknesses
  1. I am not fully convinced of where the gain comes from. Shen et al. (CVPR 2019) already explored thickness/prior cues. Here, the allocation prior is a full 3D occupancy prediction. The ablation shows intermediate supervision helps. It does not separate the proposed fusion from a standard coarse-to-fine model. A capacity-matched baseline is needed.
  2. The ASOCA gains are small. Dice improves by 0.023 and cIDice by 0.017 over Two Projections Suffice. HD95 and ASSD are worse than all three non-Gaussian baselines. There are only four test patients. The reported std reflects test-pair variability, not repeated training runs. It is also unclear whether infinite distances from empty predictions were excluded.
Questions for authors
  1. Could you provide a capacity-matched coarse-to-fine baseline with the same amount and occupancy supervision, but with standard feature fusion and residual refinement? This would help separate the benefit of the proposed fusion from that of intermediate supervision.
  2. For ASOCA, please report per-patient Dice, soft clDice, HD95, and ASSD. Are the gains consistent across patients, and which cases account for the worse boundary distances? How many samples, if any, were omitted from the distance averages because of empty predictions?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer anyZ (1/3)

Dear Reviewer anyZ,

Thank you for recognizing the ImageCAS improvements and the value of the fixed-capacity and GT-amount ablations. Your comments helped us strengthen the fusion comparison and examine ASOCA at the patient and boundary levels. We address each weakness with its corresponding question below.

Reply to W1 and Q1: Contribution and the standard-fusion control

Concern: Thickness prediction has precedents, and the occupancy prior predicts the full target; does the proposed fusion improve upon a comparably supervised coarse-to-fine model?

The requested nearly parameter-matched CONCAT control shows a 1.15-point Dice advantage for our fusion with the same intermediate supervision and residual refinement. FULL retains the paper checkpoint, and CONCAT uses the same split, seed 42, and training settings. The evaluation covers 100 ImageCAS cases and 500 paired projections.

MetricBiAAR (FULL; paper checkpoint)CONCAT
Dice ↑0.8086070.797072
soft-clDice ↑0.8790140.870968
Amount FG-MSE ↓0.0110.011
Prior Dice ↑0.7893850.790294
Final Dice − Prior Dice0.0192230.006778

FULL/CONCAT have 13,883,723/13,876,379 parameters and 651.839/621.036 GFLOPs per paired sample, under the same counting convention. The parameter difference is 0.053%. The FLOPs reflect 34 versus 17 channels in the first refiner convolution. FULL uses [VR,VR⊙G,P,U][V^R,V^R\odot G,P,U], whereas CONCAT uses [VR,P][V^R,P]; both retain the refiner backbone, amount and prior losses, and prior-logit residual. Here G=0.1+0.9PG=0.1+0.9P and U=P(1−P)U=P(1-P).

FULL's Dice gain is 1.15 points, with case-paired 95% CI [0.89, 1.41] and improvement in 86/100 cases. With comparable reported amount FG-MSE and slightly lower prior Dice, its larger 1.92-versus-0.68-point refinement gain supports more effective use of the branch information. FULL recovers 21.01% versus 7.88% of prior false negatives and removes 30.45% versus 13.04% of prior false positives. The complete reconstruction and correction results, including harmful edits, are in Tables 13–14. (Changes and additions are marked in blue in the revised manuscript.)

A matched-capacity representation control also favors amount. We replace amount with ray-depth centroid while retaining prior supervision, fusion, residual refinement, and loss settings. Both variants have 13,883,723 parameters and 651.839 GFLOPs per pair.

MetricBiAAR (FULL; paper checkpoint)Centroid
Dice ↑0.808607 ± 0.0413950.797105 ± 0.041216
soft-clDice ↑0.879014 ± 0.0399350.869639 ± 0.042101
Prior Dice ↑0.789385 ± 0.0425410.776332 ± 0.041118
Final Dice − Prior Dice0.0192230.020772

These are case-level mean ± SD after averaging five pairs per case. FULL improves Dice by 1.150 points (95% CI [0.873, 1.419]; 82/100 cases). The advantage is already present in prior Dice, while Centroid has a slightly larger refinement gain. Together with the original fixed-capacity loss ablation, these controls test supervision, representation choice, and fusion. Complete Centroid results are in Table 15; Section 4.3 summarizes the findings.

The conceptual contribution is the decomposition-guided choice and use of supervised intermediate representations. The ray-wise identity yv=A~vrayαvrayy_v=\widetilde A_v^{\mathrm{ray}}\alpha_v^{\mathrm{ray}} specifies amount and normalized allocation. BiAAR learns amount maps and a shared occupancy prior, then uses amount-conditioned features and prior-guided spatial weighting for refinement. The prior is supervised by the full occupancy target and retains allocation and amplitude information; the learned decoder consumes these cues directly.

As you noted, Shen et al. (CVPR 2019) provide an important thickness-prediction precedent. They predict thickness and sine/cosine-weighted maps, construct Fourier slices, and reconstruct through an inverse 3D Fourier transform. BiAAR combines calibrated lifting with occupancy-based allocation guidance and learned spatial fusion. We clarify the formulation-to-implementation correspondence in Section 3.1, with the prior-weighting derivation in Appendix C.3.

AuthorsOpenReview ↗

Reply to Reviewer anyZ (2/3)

Reply to W2 and Q2: ASOCA patient-level results, boundary errors, and exclusions

Concern: Are ASOCA gains consistent, which patients account for worse boundaries, and were empty predictions omitted?

ASOCA Dice improves in all four patients against every baseline; its overlap gains coexist with a boundary-accuracy disadvantage, and no predictions were excluded. The original benchmark evaluation means are:

MethodDice ↑soft-clDice ↑HD95 ↓ASSD ↓
BiAAR0.63610.58658.37941.7028
Two Projections Suffice0.61300.56877.29231.5092
AutoCAR0.57010.48575.92401.3898
DeepCA0.56730.43975.75661.3434
3DGR-CAR0.23030.207110.55273.2407

Patient-level results. Each patient contributes separate artery and vein volumes, with 20 pairs per volume. We average within each volume, then equally across the two volumes. Each row summarizes 40 pairs, while the independent sample size is four patients. Distances are voxels on the 1283128^3 grid. Below is the comparison with the strongest overlap baseline; bold marks the better value within each patient:

PatientMethodDice ↑soft-clDice ↑HD95 ↓ASSD ↓
3BiAAR0.63330.56716.82381.5208
3Two Projections Suffice0.61820.56645.89321.2867
9BiAAR0.64460.60119.85931.6481
9Two Projections Suffice0.62820.57738.97031.6060
10BiAAR0.68050.63505.75191.2297
10Two Projections Suffice0.63140.57957.99141.6371
36BiAAR0.58600.542711.08282.4128
36Two Projections Suffice0.57410.55146.31431.5070

BiAAR improves Dice in four patients and soft-clDice in three relative to Two Projections Suffice; both boundary metrics improve for patient 10. Patient 36 contributes most to the boundary disadvantage, with HD95/ASSD differences of +4.7685/+0.9058 voxels. Patient 9 has high absolute HD95 and a smaller gap to this baseline; patient 10 offsets part of the aggregate difference. The complete five-method patient table is in Appendix D.3, Table 6.

Directional distances identify incomplete branch coverage. For patient 36, the mean reference-to-prediction distance is 3.4024 voxels, compared with 0.7366 in the reverse direction; 21.88% of reference surface lies more than five voxels from the prediction. The complete directional diagnostics are in Table 7.

All 160 ASOCA predictions per method are included. There are zero missing records, empty predictions, empty references, or distance exclusions across all five methods. HD95 and ASSD denominators are 160 per method and 40 per patient; Table 8 provides the audit.

The original standard deviations describe test-pair variability for the selected checkpoint. Patient-level aggregation accounts for repeated projections, and the four Dice comparisons attain the minimum possible two-sided exact p-value, 2/16 = 0.125. Appendix D.2 reports the statistical protocol and finite-sample resolution.

ASOCA tests reconstruction with 32 training patients / 64 reference volumes, compared with 800 ImageCAS training cases, and detailed arterial and venous structures. Its AVDNet-derived masks and reconstruction targets come from the same volumes, maintaining input–target consistency for this synthetic task. The revised Section 4.6 identifies the reference provenance, patient count, and overlap–boundary tradeoff.

We hope the controlled fusion results and patient-level analysis address your questions about attribution and reliability. Thank you again for helping us strengthen the evidence and its presentation.

AuthorsOpenReview ↗

Reply to Reviewer anyZ (3/3)

References

  1. Shen W, Jia Y, Wu Y. 3D shape reconstruction from images in the frequency domain[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019: 4466-4474.
  2. Wang W, Xia Q, Yan Z, et al. AVDNet: Joint coronary artery and vein segmentation with topological consistency[J]. Medical Image Analysis, 2024, 91: 102999.
  3. Cafaro A, Dorent R, Haouchine N, et al. Two projections suffice for cerebral vascular reconstruction[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 722-731.
  4. Zhu Y, Wang Y, Di C, et al. Sparse and transferable three-dimensional dynamic vascular reconstruction for instantaneous diagnosis[J]. Nature Machine Intelligence, 2025, 7(5): 730-742.
  5. Wang Y, Banerjee A, Choudhury R P, et al. DeepCA: Deep learning-based 3D coronary artery tree reconstruction from two 2D non-simultaneous X-ray angiography projections[C]//2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2025: 337-346.
  6. Fu X, Li Y, Tang F, et al. 3dgr-car: Coronary artery reconstruction from ultra-sparse 2d x-ray views with a 3d gaussians representation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 14-24.
Reviewer_ZjriOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies supervised three-dimensional vascular occupancy reconstruction from two calibrated binary projection masks. It motivates an auxiliary representation that separates total occupied amount along a ray from its normalized depth allocation. The implementation predicts view-normalized amount maps and a full-volume occupancy prior, then combines lifted mask/SDF/amount features with prior-guided gating and residual refinement. It does not explicitly reconstruct occupancy by multiplying normalized amount and allocation factors. Experiments on simulated ImageCAS and AVDNet-derived ASOCA projections report improvements in Dice and soft clDice over adapted baselines, with component, fixed-capacity supervision, calibration, and transfer analyses. The method is best viewed as auxiliary-task-supervised coarse-to-fine volumetric reconstruction.

Critical strengths
  1. The paper isolates an important ambiguity in sparse-view reconstruction: binary masks indicate ray support but not depth occupancy. It tests explicit intermediate supervision rather than relying only on the final volume loss.

  2. The implementation and reporting expose important practical tradeoffs. The model is feed-forward, includes calibration perturbation tests, and reports all six reconstruction metrics. The paper distinguishes binary-input adaptations from the original baseline tasks, identifies the ASOCA reference-label source and statistical units, and acknowledges the gap between exact factorization and learned refinement.

  3. The empirical section includes component and fixed-capacity loss ablations, calibration perturbations, transfer to a second dataset, and an inference-time comparison. These checks make the benchmark more informative than a single aggregate score.

Critical weaknesses
  1. The distinctive algorithmic contribution of amount–allocation decomposition remains insufficiently established. The identity y=Aαy=A\alpha is valid, but does not make the factors identifiable from binary masks or establish a learning advantage. The occupancy prior is directly supervised by the full target volume and already solves the original prediction task in principle; the refiner consumes that full prior rather than enforcing a normalized factorization. The fixed-capacity controls establish the usefulness of auxiliary supervision, but do not isolate this particular representation from an otherwise comparable deeply supervised cascade. The 0.009 Dice increment over the prior-only configuration is encouraging, yet needs stronger attribution to support the conceptual contribution beyond established learned lifting and coarse-to-fine reconstruction.

  2. The predicted amount remains a major bottleneck: ground-truth amount maps raise ImageCAS Dice from 0.809 to 0.926. The main comparison is also limited to simulated/derived volumes and clean masks. Some baselines originally use grayscale or accumulated-intensity evidence, whereas this evaluation changes them to binary masks; that supports a narrow benchmark claim, not broad superiority over their published systems or clinical angiography.

Questions for authors
  1. The exact factorization is not identifiable from two binary views and is not enforced by the decoder. What controlled evidence shows that the amount/allocation formulation itself helps beyond a matched-capacity network with generic auxiliary occupancy or depth-bin supervision?

  2. Can the authors provide a matched-capacity comparison that removes each auxiliary target while keeping the same number of branches and refiner inputs, and quantify how much of the gain comes from the amount loss versus the occupancy prior?

  3. Please clarify the scope of the benchmark for the intensity-based baselines and the ASOCA targets: which published baseline capabilities are removed by the binary-mask adaptation, and how should readers interpret the AVDNet-derived reference labels relative to manual vessel annotations?

6 comments
AuthorsOpenReview ↗

Reply to Reviewer Zjri (1/4)

Dear Reviewer Zjri,

Thank you for recognizing the importance of binary-projection ambiguity, explicit intermediate supervision, and experimental reporting. Your comments helped us strengthen the attribution of the representation, supervision, and fusion choices. We address overlapping weaknesses and questions together below.

Reply to W1 and Q1: Contribution beyond generic auxiliary supervision and coarse-to-fine reconstruction

Concern: Does this particular representation help beyond a comparably supervised cascade, and what supports the approximately 0.009 Dice improvement?

Matched controls support the choice of amount representation and its spatially guided fusion, complementing the fixed-architecture supervision ablation. FULL improves Dice by approximately 1.15 points over both Centroid and CONCAT on the same 100 ImageCAS cases and 500 paired projections.

Representation control. We replace amount with ray-depth centroid, retaining the occupancy-prior branch, fusion, residual refinement, training settings, and loss form and weights. Both models have 13,883,723 parameters and 651.839 GFLOPs per paired sample.

MetricBiAAR (FULL; paper checkpoint)Centroid
Dice ↑0.808607 ± 0.0413950.797105 ± 0.041216
soft-clDice ↑0.879014 ± 0.0399350.869639 ± 0.042101
Prior Dice ↑0.789385 ± 0.0425410.776332 ± 0.041118
Final Dice − Prior Dice0.0192230.020772

After averaging five pairs within each case, the Dice improvement is 1.150 points, with 10,000 case-paired bootstrap 95% CI [0.873, 1.419] and improvement in 82/100 cases. Centroid learns meaningful depth information (foreground MSE 0.008, MAE 0.068642, Pearson 0.735744). FULL's advantage is already present in prior Dice, supporting the amount representation within the jointly trained system. The complete metrics are in Table 15. (Changes and additions are marked in blue in the revised manuscript.)

Fusion control. CONCAT retains both supervised targets, the refiner backbone, and prior-logit residual. It replaces FULL's [VR,VR⊙G,P,U][V^R,V^R\odot G,P,U] with direct concatenation [VR,P][V^R,P]. FULL uses the paper checkpoint; CONCAT uses the same split, seed 42, and training protocol.

MetricBiAAR (FULL; paper checkpoint)CONCAT
Dice ↑0.8086070.797072
soft-clDice ↑0.8790140.870968
Amount FG-MSE ↓0.0110.011
Prior Dice ↑0.7893850.790294
Final Dice − Prior Dice0.0192230.006778

FULL/CONCAT have 13,883,723/13,876,379 parameters and 651.839/621.036 GFLOPs per pair. This requested standard-fusion control closely matches capacity: the parameter difference is 0.053%. FLOPs follow the same convention and reflect 34 versus 17 input channels at the refiner's first convolution.

FULL gains 1.15 Dice points (case-paired 95% CI [0.89, 1.41]; 86/100 cases). Amount FG-MSE is equal at the reported precision and FULL's prior Dice is slightly lower, while its refinement gain reaches 1.92 versus 0.68 points. FULL recovers 21.01% versus 7.88% of prior false negatives and removes 30.45% versus 13.04% of prior false positives. This supports the proposed fusion's use of the intermediate evidence. Complete metrics and the correction audit are in Tables 13–14.

For the approximately 0.009 Dice increment noted in the review, original B6–B1 records give +0.008629, 95% CI [0.006255, 0.010979], improving 76/100 cases; Holm-adjusted p is 0.000900 across 90 ablation comparisons.

The organizing representation is yv=A~vrayαvrayy_v=\widetilde A_v^{\mathrm{ray}}\alpha_v^{\mathrm{ray}}: amount specifies how much occupied structure lies along a ray, and allocation specifies where it lies. BiAAR supervises view-normalized amount and a shared occupancy prior whose ray profiles induce allocation. The prior also retains amplitude; it weights lifted amount–mask–SDF features before learned residual refinement. We clarify this correspondence in Section 3.1, and summarize the additional controls in Section 4.3 and Appendix D.6. Together with Q2, these experiments test representation choice, fusion, and direct supervision separately.

AuthorsOpenReview ↗

Reply to Reviewer Zjri (2/4)

Reply to Q2: Contributions of the two losses at fixed architecture

Question: Can each auxiliary target be removed while retaining all branches and refiner inputs, and how much does each loss contribute?

At fixed architecture, amount supervision adds 1.5 Dice points when prior supervision is present, and prior supervision adds 3.1 points when amount supervision is present. Table 3 changes only λA\lambda_A and λP\lambda_P, retaining every B6 branch, refiner input, parameter, and training/test setting. Both branches continue receiving final reconstruction gradients.

λA\lambda_AλP\lambda_PDice ↑soft-clDice ↑Amount FG-MSE ↓Prior Dice ↑
000.7740.8510.0430.350
010.7940.8690.0520.773
100.7780.8550.0120.000
110.8090.8790.0110.789

Joint supervision gives the best Dice and soft-clDice and adds 3.5 Dice points over neither loss. The table uses the original 500-pair means; amount FG-MSE measures normalized foreground amount, and prior Dice thresholds PP at 0.5. At (0,0)(0,0), final-loss training already yields amount MSE 0.043 and prior Dice 0.350, showing partial intermediate learning. Direct supervision improves target alignment and final overlap. Full reconstruction metrics appear in Table 3, with analysis in Section 4.3.

AuthorsOpenReview ↗

Reply to Reviewer Zjri (3/4)

Reply to W2: Benchmark scope, amount-estimation headroom, and robustness

Concern: Predicted amount remains a bottleneck, and simulated binary-input benchmarks support a task-specific comparison.

Our comparative claim concerns reconstruction under the shared binary-input protocol. Replacing grayscale or accumulated-intensity observations changes the available evidence; the results compare the adapted methods on this task. Q3 details the original capabilities and the projection/loss adaptations. Within this setting, the controls under W1/Q1 and Q2 support the representation, fusion, and supervision choices.

Amount-estimation headroom. Original B11 raises Dice/soft-clDice from 0.809/0.879 to 0.926/0.950 with accurate amount maps. This demonstrates the value of recovering the ray-total information discarded by binary projections and motivates better amount estimation, potentially using temporal sequences. The case-level relationship between amount error and reconstruction is reported in Appendix D.7, Table 20; the amount FG-MSE versus final Dice Spearman correlation is 0.003024 for the current checkpoint.

Non-ideal input simulations. Among reviewed public resources, we have not identified a dataset jointly providing real paired angiograms, calibrated cameras, and independent 3D references matched to acquisition. AutoCAR releases real videos and poses, without identifying such 3D reference sequences. We therefore test mask defects, unseen view angles, and 3D inter-view motion with known targets, alongside the original calibration tests. We would welcome pointers to a public dataset meeting these requirements.

All new input tests use frozen BiAAR and AutoCAR dense reconstruction checkpoints trained on clean masks and 90° pairs only, with the same 100 cases and five pairs per case. They compare reconstruction under the shared simulation protocol, with common historical BP support.

Mask perturbationBiAAR Dice / soft-clDice ↑AutoCAR Dice / soft-clDice ↑
Clean0.8086 / 0.87900.7128 / 0.7837
Boundary-adjacent 1% foreground addition0.8048 / 0.87560.7066 / 0.7784
Localized 5% foreground removal0.7436 / 0.80710.6586 / 0.7182
One-pixel erosion0.5535 / 0.65270.3066 / 0.3431
One-pixel dilation0.5580 / 0.69990.4679 / 0.6557

BiAAR achieves Dice 0.8048 with 1% foreground addition and 0.7436 with 5% localized removal. Erosion/dilation results are 0.5535/0.5580, compared with 0.3066/0.4679 for AutoCAR. At unseen 75°/60°/45° separations with correct cameras, BiAAR reaches Dice 0.8064/0.8005/0.7863, versus 0.7103/0.6997/0.6837 for AutoCAR. Complete severities and metrics are in Tables 9–10.

We also simulate motion by transforming the 3D anatomy before generating only the second projection, following DeepCA's precedent (Wang et al., 2025). The first-time-point reference remains fixed, cameras are correct, and second-view BP is regenerated. Local branch rotations of 1°/2°/3° give BiAAR Dice 0.78856/0.76818/0.75373, compared with AutoCAR's 0.70001/0.68405/0.67188. Whole-structure rotations give 0.75101/0.67496/0.60192, compared with 0.67999/0.62983/0.57738.

At local 3°, the Dice advantage is 8.18 points (case-paired 95% CI [7.75, 8.62]; 100/100 cases). Moving-region reference-time recall is 47.61%, with approximately 82.4% elsewhere; whole-structure 3° yields HD95/ASSD 4.657/1.025 for BiAAR and 3.696/0.925 for AutoCAR. These measurements identify the remaining local alignment challenge alongside higher overall overlap. The 7,000 predictions include no empty outputs or distance exclusions. Section 4.5 summarizes the added tests, and Appendix D.5 gives the complete motion results and diagnostics.

Reviewer_ZydYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BiAAR, a feed-forward method for reconstructing 3D vascular structures from two binary projections with known camera parameters. It predicts vessel amount maps and an initial 3D occupancy volume, then combines them through a refinement network. Experiments show good performance on simulated vascular reconstruction benchmarks.

Critical strengths
  1. Good Performance. The method achieves good reconstruction results, particularly on ImageCAS, and obtains the highest mean Dice and clDice among the evaluated methods on both datasets.

  2. Clear Motivation. Binary projections do not directly reveal vessel amount or its distribution along each ray, motivating the use of supervised intermediate predictions.

Critical weaknesses
  1. Limited Novelty. The model mainly combines established components. The benefit of the amount–allocation decomposition beyond auxiliary supervision and coarse-to-fine reconstruction remains insufficiently established.

  2. Decomposition-Implementation Gap. The allocation branch predicts a full occupancy volume, and refinement does not explicitly combine the proposed factors. It remains unclear whether the gains come from the decomposition or generic occupancy supervision.

  3. Limited Experimental Validation. Evaluation is restricted to synthetic projections. Real-data validation, a matched coarse-to-fine baseline, and clearer verification of intensity-based baselines under binary inputs are needed.

  4. Unclear Presentation. The connection between the mathematical decomposition and the actual network is difficult to follow. The method description and figures should clarify the intermediate predictions and fusion process.

Questions for authors
  1. What advantages does the amount–allocation decomposition offer over conventional coarse-to-fine reconstruction with auxiliary supervision?

  2. Why is allocation implemented as a fully supervised occupancy volume? Can the authors clarify how the factorization guides the actual refinement and whether predicted amounts remain consistent with the final reconstruction?

  3. How does the method perform on real angiographic inputs or under segmentation errors and non-orthogonal views? How were the projection models and losses of intensity-based baselines adapted to binary inputs?

  4. How much do cross-view attention and the epipolar bias contribute to amount prediction and reconstruction quality?

Overall, the paper presents good reconstruction results with a reasonably motivated design. However, my main concern is whether the proposed decomposition offers meaningful advantages beyond conventional coarse-to-fine reconstruction and auxiliary supervision. The gap between the formulation and implementation, together with evaluation limited to synthetic projections, also leaves its practical effectiveness unclear. Clarifications and additional evidence addressing these concerns could positively influence my assessment and lead me to increase my rating.

8 comments
AuthorsOpenReview ↗

Reply to Reviewer ZydY (1/6)

Dear Reviewer ZydY,

Thank you for recognizing the reconstruction performance and motivation for explicit intermediate predictions. Your comments helped us clarify the design and add controlled evidence for the representation, supervision, and fusion choices. We address overlapping weaknesses and questions together below.

Reply to W1 and Q1: Contribution beyond auxiliary supervision and coarse-to-fine reconstruction

Concern: Does amount–allocation decomposition provide a meaningful benefit beyond established components, auxiliary supervision, and a conventional coarse-to-fine cascade?

Our contribution is a decomposition-guided choice of supervised intermediate representations and their spatially guided fusion. New controls show approximately 1.15 percentage points higher Dice than both the requested nearly parameter-matched CONCAT baseline and a matched-capacity Centroid alternative.

Fusion control with the same supervision and residual refinement. CONCAT retains both intermediate branches, their losses, and the prior-logit residual, replacing our fusion with direct concatenation. FULL uses the paper checkpoint; CONCAT uses the same split, seed 42, and training settings. Evaluation covers 100 ImageCAS cases and 500 paired projections.

MetricBiAAR (FULL; paper checkpoint)CONCAT
Dice ↑0.8086070.797072
soft-clDice ↑0.8790140.870968
Amount FG-MSE ↓0.0110.011
Prior Dice ↑0.7893850.790294
Final Dice − Prior Dice0.0192230.006778

FULL and CONCAT have 13,883,723/13,876,379 parameters and 651.839/621.036 GFLOPs per paired sample, respectively. The requested standard-fusion baseline therefore closely matches model capacity. The parameter difference is 0.053%; the FLOPs reflect 34 versus 17 channels at the refiner's first convolution, under the same counting convention.

FULL improves Dice by 1.15 points (case-paired bootstrap 95% CI [0.89, 1.41]), with higher Dice in 86/100 cases. Its advantage emerges after refinement despite comparable amount FG-MSE at the reported precision and slightly lower prior Dice. The prior-to-final gain is 1.92 versus 0.68 points. FULL recovers 21.01% versus 7.88% of prior false negatives and removes 30.45% versus 13.04% of prior false positives. This supports more effective use of the two branches through the proposed fusion. The complete reconstruction metrics and correction audit, including harmful edits, are in Tables 13–14. (Changes and additions are marked in blue in the revised manuscript.)

Representation control with matched capacity. Replacing amount with ray-depth centroid retains the prior branch, fusion, residual refinement, loss form and weights, and training settings. Both models have 13,883,723 parameters and 651.839 GFLOPs per pair.

MetricBiAAR (FULL; paper checkpoint)Centroid
Dice ↑0.808607 ± 0.0413950.797105 ± 0.041216
soft-clDice ↑0.879014 ± 0.0399350.869639 ± 0.042101
Prior Dice ↑0.789385 ± 0.0425410.776332 ± 0.041118
Final Dice − Prior Dice0.0192230.020772

Values are case-level mean ± SD after averaging each case's five pairs. FULL improves Dice by 1.150 points (10,000 case-paired bootstrap resamples; 95% CI [0.873, 1.419]), with higher Dice in 82/100 cases. The advantage is already present in the prior; Centroid has a slightly larger refinement gain. This supports amount as a supervised representation within the jointly trained system. Complete metrics are in Table 15.

The geometric motivation is the ray-wise factorization

yv(p,t)=A~vray(p),αvray(t∣p).y_v(\mathbf{p},t)=\widetilde A_v^{\mathrm{ray}}(\mathbf{p})\\,\alpha_v^{\mathrm{ray}}(t\mid\mathbf{p}).

Amount specifies how much occupied structure lies along a ray; allocation specifies where that amount lies. BiAAR supervises view-normalized amounts and a shared occupancy prior that induces allocation profiles, then uses them to condition and weight lifted features. The amount branch operates on foreground 2D tokens with camera-aware attention; 3D backprojection is used by the allocation and refinement branches.

Thickness supervision has precedents. Shen et al. (CVPR 2019) predict thickness and weighted-thickness maps, construct Fourier slices, and apply an inverse 3D Fourier transform. BiAAR couples ray-wise amount with occupancy-based allocation guidance and learned spatial fusion from calibrated binary masks. The new controls separately test representation choice and fusion, complementing the supervision control below. We added these comparisons to Section 4.3 and Appendix D.6.

AuthorsOpenReview ↗

Reply to Reviewer ZydY (2/6)

Reply to W2 and Q2: How the factorization guides occupancy supervision and refinement

Concern: The allocation branch already predicts full occupancy; how does this implement the factorization, and are predicted amounts consistent with the final reconstruction?

The prior provides a shared spatial field for allocation guidance, and joint supervision improves final reconstruction at fixed architecture. Its normalized ray profiles are αvP(x∣p)=wv(p,x)P(x)/∑x′wv(p,x′)P(x′)\alpha_v^P(\mathbf x\mid\mathbf p)=w_v(\mathbf p,\mathbf x)P(\mathbf x)/\sum_{\mathbf x'}w_v(\mathbf p,\mathbf x')P(\mathbf x'), with zero profiles for zero denominators. At P=YP=Y, these are the GT allocations, also computable directly from YY. Occupancy supervision supplies a common field across both views and retains occupancy amplitude.

In the implemented fusion, amount conditions the lifted features and PP weights their spatial assignment. Prior gating admits the allocation-weighted lifting identity derived in Appendix C.3. The full refiner performs learned fusion and residual correction rather than explicitly multiplying the two predicted factors. We clarify these distinct roles in Section 3.1 and the correspondence under W4 below.

Fixed-capacity supervision. Table 3 keeps every branch and refiner input, changing only λA\lambda_A and λP\lambda_P. All branches receive gradients through the final reconstruction loss.

λA\lambda_AλP\lambda_PDice ↑soft-clDice ↑Amount FG-MSE ↓Prior Dice ↑
000.7740.8510.0430.350
010.7940.8690.0520.773
100.7780.8550.0120.000
110.8090.8790.0110.789

With prior supervision present, amount supervision adds 1.5 Dice points; with amount supervision present, prior supervision adds 3.1 points. Joint supervision adds 3.5 points over neither loss. The table uses the original 500-pair means; prior Dice thresholds PP at 0.5, and amount FG-MSE uses foreground normalized amounts. The complete metrics remain in Table 3. The CONCAT and Centroid controls above separately examine how predictions are combined and what is supervised.

Amount–reconstruction consistency. Final reprojections have lower target error than the amount-head predictions while remaining different from them. Using the paper checkpoint, we compare the direct amount with the view-normalized projection of the thresholded final occupancy:

ComparisonForeground MSE
Predicted amount vs. GT amount0.011
Binary final reprojection vs. GT amount0.006
Predicted amount vs. binary final reprojection0.013

Soft final reprojection gives the same reported errors, 0.006 against GT and 0.013 against the direct amount. Binary reprojection support has mean IoU 0.927987 with the input masks, measured over the whole image without mask clipping. These results describe the consistency attained by learned reconstruction; Appendix D.7, Table 19, gives all comparisons and definitions.

Furthermore, the GT-amount ablation raises Dice/soft-clDice from 0.809/0.879 to 0.926/0.950. With accurate amount supplied, the resulting Dice exceeds 0.90 and all evaluated binary-input baselines. This identifies amount estimation as a promising direction: temporal sequences may provide additional evidence for more reliable amount maps and better biplanar reconstruction.

AuthorsOpenReview ↗

Reply to Reviewer ZydY (3/6)

Reply to W3 and Q3: Validation scope, input robustness, and baseline adaptations

Concern: Synthetic-only evaluation, the missing matched cascade control, and unclear binary-input baseline adaptations limit the evidence.

The requested standard-fusion control favors BiAAR, and the frozen model retains higher overlap under tested non-ideal inputs. W1/Q1 reports the nearly parameter-matched comparison with unchanged intermediate supervision and residual refinement. Here we summarize the additional input tests and the actual binary-loss adaptations.

Among the public resources reviewed, we have not identified a clinical dataset jointly providing real biplanar angiograms, calibrated cameras, and independent 3D references matched to the acquisition state. We therefore simulate segmentation errors, unseen viewing geometries, and inter-view motion with known 3D structures, complementing the paper's calibration tests. We would welcome pointers to a public dataset meeting these requirements. AutoCAR's documented release includes real videos and poses; independent time-matched 3D references are not identified in that release.

Segmentation and viewing geometry. Training used only clean masks and 90° pairs; all input tests use the original frozen checkpoints on the same 100 cases and 500 pairs per condition. SDFs, tokens, and BP support follow the perturbed masks.

Mask perturbationBiAAR Dice / soft-clDice ↑AutoCAR Dice / soft-clDice ↑
Clean0.8086 / 0.87900.7128 / 0.7837
Boundary-adjacent 1% foreground addition0.8048 / 0.87560.7066 / 0.7784
Localized 5% foreground removal0.7436 / 0.80710.6586 / 0.7182
One-pixel erosion0.5535 / 0.65270.3066 / 0.3431
One-pixel dilation0.5580 / 0.69990.4679 / 0.6557

With 1% added foreground or 5% localized removal, BiAAR reaches Dice 0.8048/0.7436. One-pixel erosion/dilation changes about 30% of foreground and yields Dice 0.5535/0.5580, above AutoCAR's 0.3066/0.4679. At unseen, correctly calibrated 75°/60°/45° separations, BiAAR reaches 0.8064/0.8005/0.7863, compared with AutoCAR's 0.7103/0.6997/0.6837. Section 4.5 summarizes these tests; Appendix D.4, Tables 9–10, reports all metrics and severities.

Inter-view motion. We deform the 3D vessel before generating the second view while retaining the first-time-point GT, correct cameras, and regenerated second-view BP. DeepCA provides a precedent for this simulation (Wang et al., 2025), and measured heart rotations motivate the 1°–3° range (Shechter et al., 2004).

Second-view 3D motionBiAAR Dice / soft-clDice ↑AutoCAR Dice / soft-clDice ↑
Local branch 1°0.78856 / 0.855220.70001 / 0.76758
Local branch 2°0.76818 / 0.829870.68405 / 0.74460
Local branch 3°0.75373 / 0.811600.67188 / 0.72693
Whole-structure 1°0.75101 / 0.829450.67999 / 0.75292
Whole-structure 2°0.67496 / 0.756020.62983 / 0.70073
Whole-structure 3°0.60192 / 0.670590.57738 / 0.63559

The experiment comprises 7,000 predictions, all included with zero empty outputs or distance exclusions. At local 3°, BiAAR has 8.18 points higher Dice (case-paired 95% CI [7.75, 8.62]; 100/100 cases). Its moving-region reference-time recall is 47.61%, compared with approximately 82.4% elsewhere, identifying local alignment as the principal remaining issue. Whole-structure 3° gives HD95/ASSD 4.657/1.025 for BiAAR and 3.696/0.925 for AutoCAR. Complete motion results and diagnostics are in Appendix D.5, Tables 11–12. Our intended synchronized setting is illustrated by the Siemens biplane system in Fig. 1; rotational coronary scanning instead requires motion compensation or gating (Blondel et al., 2004).

Reviewer_PTQ5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a new methodology to create 3D blood vessel reconstructions from two 2D projections. It is based on the idea of decomposing the ill-posed reconstruction problem into amount (how much vessel each ray touches) and allocation (where along the ray the vessel lies) quantities, and leveraging learnable priors.

Critical strengths
  1. The writing is clear.
  2. The proposed decomposition into amount and allocation is sensible.
  3. Standard deviations are reported with results, which is especially important for small test sets.
  4. Most reported performance shows improvements over baselines.
  5. Six different evaluation metrics are used, making the evaluation more comprehensive.
Critical weaknesses
  1. Evaluation on simulated projections, AI-made annotations (ASOCA), a small ASOCA test set (4 patients), and internal validation are significant limitations. However, these are mainly limitations of the field and available public datasets. In any case, a clear limitations section explaining these problems is important, because these limitations strongly limit claims on clinical usability.
  2. Conclusions mention that “Clinical validation should cover real angiograms, mask errors, diverse biplane geometries, and motion compensation for rotational coronary scans.” At least the resistance to mask error could be evaluated with the current datasets. Motion could also be simulated.
  3. Standard deviations in Table 1 overlap. Statistical tests should be used to show whether results significantly surpass baselines.
Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer PTQ5 (1/3)

Dear Reviewer PTQ5,

Thank you for recognizing the clear presentation and sensible amount–allocation formulation. Your valuable comments help us better report validation limits, robustness, and statistical evidence.

Reply to W1: Dataset limitations and clinical scope

Concern: Simulated projections, derived ASOCA labels, few independent test patients, and internal validation limit clinical claims.

Our results support reconstruction under the tested binary-projection protocols, and the revised main text explicitly defines the validation scope. (Changes and additions are marked in blue in the revised manuscript.) Among the public resources reviewed, we have not identified a dataset jointly providing real biplanar angiograms, calibrated camera parameters, and independent 3D vessel annotations matched to the acquisition state. We would welcome pointers to a public dataset meeting these requirements.

Test setIndependent cases/patientsReference volumesPaired projectionsReference source
ImageCAS100100500ImageCAS vessel annotations
ASOCA-derived48160AVDNet artery/vein predictions

Both benchmarks synthesize projections from reference volumes and use internally held-out patients. ASOCA's 160 projection pairs represent four independent patients. Its masks and reconstruction targets come from the same volumes, providing a consistent synthetic inverse problem with detailed arterial and venous structures.

To simulate non-ideal acquisition and segmentation conditions with known 3D references, we evaluate mask defects, unseen view angles, inter-view motion, and camera-calibration errors. For example, BiAAR reaches Dice 0.7436 with localized 5% foreground removal and 0.7863 at an unseen 45° separation; the relevant tests are detailed under W2. AutoCAR's public release provides real videos and poses, while its documentation does not identify independent time-matched 3D references for those videos.

We added Section 4.6, Limitations and Validation Scope, and report the dataset audit and patient-level evidence in Appendices D.1–D.3. The revised text identifies the derived labels, internal splits, independent sample counts, and clinical validation needs.

AuthorsOpenReview ↗

Reply to Reviewer PTQ5 (2/3)

Reply to W2: Mask errors, unseen view angles, and motion

Concern: Evaluate segmentation errors and diverse biplane geometries, and clarify the missing motion evaluation.

BiAAR retains higher reconstruction overlap than AutoCAR under the tested mask and inter-view motion conditions and transfers to unseen viewing angles. All tests use the original frozen BiAAR and AutoCAR dense reconstruction checkpoints, trained only on clean masks and 90° pairs. Each condition uses the same 100 ImageCAS cases and five pairs per case, with identical observations for both methods.

Mask errors. We regenerate SDFs, tokens, and BP support from perturbed inputs. The key results are:

Mask perturbationBiAAR Dice / soft-clDice ↑AutoCAR Dice / soft-clDice ↑
Clean0.8086 / 0.87900.7128 / 0.7837
Boundary-adjacent 1% foreground addition0.8048 / 0.87560.7066 / 0.7784
Localized 5% foreground removal0.7436 / 0.80710.6586 / 0.7182
One-pixel erosion0.5535 / 0.65270.3066 / 0.3431
One-pixel dilation0.5580 / 0.69990.4679 / 0.6557

Adding 1% foreground yields BiAAR Dice 0.8048, and removing 5% locally yields 0.7436. One-pixel erosion and dilation change approximately 30% of foreground in these thin masks; their final Dice scores are 0.5535/0.5580, compared with AutoCAR's 0.3066/0.4679. Complete severities, six metrics, and evaluation counts are in Appendix D.4, Table 9.

Unseen viewing angles. After training only on 90° pairs, BiAAR achieves Dice 0.8064/0.8005/0.7863 at correctly calibrated 75°/60°/45° separations. AutoCAR achieves 0.7103/0.6997/0.6837. We regenerate the projections and supply the new camera parameters. The full results are in Table 10.

Simulated inter-view motion. We completed 7,000 predictions: 100 cases × five pairs × two models × seven conditions. The first view and its reference remain fixed; 3D whole-structure or smoothly blended local branch rotation precedes second-view projection, and its BP is regenerated. Cameras remain correct, and evaluation uses the first-time-point reference without registration. This follows DeepCA's 3D motion simulation precedent (Wang et al., 2025); measured respiratory heart rotations motivate the 1°–3° range (Shechter et al., 2004).

Second-view 3D motionBiAAR Dice / soft-clDice ↑AutoCAR Dice / soft-clDice ↑
Local branch 1°0.78856 / 0.855220.70001 / 0.76758
Local branch 2°0.76818 / 0.829870.68405 / 0.74460
Local branch 3°0.75373 / 0.811600.67188 / 0.72693
Whole-structure 1°0.75101 / 0.829450.67999 / 0.75292
Whole-structure 2°0.67496 / 0.756020.62983 / 0.70073
Whole-structure 3°0.60192 / 0.670590.57738 / 0.63559

At local 3° motion, BiAAR's Dice advantage is 8.18 percentage points (case-paired 95% CI [7.75, 8.62]; 100/100 cases). Its moving-region reference-time recall is 47.61%, and recall elsewhere remains approximately 82.4%. This identifies local branch alignment as a remaining challenge despite the higher whole-volume overlap. Whole-structure 3° motion produces HD95/ASSD of 4.657/1.025 voxels for BiAAR and 3.696/0.925 for AutoCAR. Every motion prediction is included, with zero empty outputs or distance exclusions.

Our intended setting uses synchronized biplanar acquisition, illustrated by the Siemens ARTIS icono system in Fig. 1. Asynchronous acquisition motivates additional motion handling. We added the motion protocol, all conditions and metrics, and regional diagnostics to Appendix D.5, Tables 11–12, and summarize these input tests in Section 4.5.

AuthorsOpenReview ↗

Reply to Reviewer PTQ5 (3/3)

Reply to W3: Statistical evidence at the anatomy level

Concern: Overlapping standard deviations require paired statistical tests of the improvements over baselines.

ImageCAS Dice improvements remain significant after multiplicity correction, and ASOCA Dice improves in all four patients against every baseline. The independent unit is the anatomy: five pairs are averaged within each ImageCAS case, and the two reference volumes and their projection pairs are averaged within each ASOCA patient.

ImageCAS baselineBiAAR − baseline Dice (pp)95% CIImproved casesHolm-adjusted p
Two Projections Suffice+13.67[13.19, 14.14]100/1000.000240
AutoCAR+9.64[9.21, 10.07]100/1000.000240
DeepCA+17.34[16.83, 17.86]100/1000.000240
3DGR-CAR+57.01[56.12, 57.89]100/1000.000240

These comparisons use the original benchmark records, with 10,000 case-paired bootstrap resamples and 100,000 Monte Carlo sign-flip draws. Holm correction covers 24 baseline–metric comparisons per dataset. The repeated ImageCAS p-values reflect the finite Monte Carlo resolution and common adjustment; the confidence intervals quantify the different effect sizes.

For ASOCA, the following shows patient-level Dice / soft-clDice; bold marks the best value for each metric within a patient:

ASOCA patientBiAARTwo Projections SufficeAutoCARDeepCA3DGR-CAR
30.6333 / 0.56710.6182 / 0.56640.5724 / 0.50400.5861 / 0.45540.2558 / 0.2267
90.6446 / 0.60110.6282 / 0.57730.5669 / 0.46250.5685 / 0.43830.1829 / 0.1711
100.6805 / 0.63500.6314 / 0.57950.5784 / 0.46520.5846 / 0.45290.2662 / 0.2317
360.5860 / 0.54270.5741 / 0.55140.5629 / 0.51110.5299 / 0.41220.2164 / 0.1988

BiAAR improves Dice in all four patients and soft-clDice in three relative to Two Projections Suffice. Four patients give 16 exact sign assignments; the Dice comparisons attain the minimum two-sided p-value, 2/16 = 0.125. We report the patient-level results and this finite-sample resolution in Appendices D.2–D.3, Tables 5–6, with the main findings in Section 4.2. The statistics summarize test-anatomy variation for the evaluated checkpoints.

Thank you again for helping us strengthen the validation and make the evidence easier to assess.

References

  1. Wang Y, Banerjee A, Choudhury R P, et al. DeepCA: Deep learning-based 3D coronary artery tree reconstruction from two 2D non-simultaneous X-ray angiography projections[C]//2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2025: 337-346.
  2. Shechter G, Ozturk C, Resar J R, et al. Respiratory motion of the heart from free breathing coronary angiograms[J]. IEEE transactions on medical imaging, 2004, 23(8): 1046-1056.
  3. Wang W, Xia Q, Yan Z, et al. AVDNet: Joint coronary artery and vein segmentation with topological consistency[J]. Medical Image Analysis, 2024, 91: 102999.
  4. Zhu Y, Wang Y, Di C, et al. Sparse and transferable three-dimensional dynamic vascular reconstruction for instantaneous diagnosis[J]. Nature Machine Intelligence, 2025, 7(5): 730-742.
  5. Cafaro A, Dorent R, Haouchine N, et al. Two projections suffice for cerebral vascular reconstruction[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 722-731.
  6. Fu X, Li Y, Tang F, et al. 3dgr-car: Coronary artery reconstruction from ultra-sparse 2d x-ray views with a 3d gaussians representation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 14-24.

Comments on the paper

1 comment

Our core contribution is to use the ray-wise factorization yv(p,t)=A~vray(p)αvray(t∣p)y_v(\mathbf p,t)=\widetilde A_v^{\mathrm{ray}}(\mathbf p)\alpha_v^{\mathrm{ray}}(t\mid\mathbf p) to design explicit intermediate supervision and fusion for two-view binary reconstruction. Amount describes how much occupied structure lies along a ray; normalized allocation describes where it lies. The exact factors jointly determine the covered 3D structure. We separately supervise view-normalized amount maps and a shared occupancy prior. For intuition, a foreground pixel's ray can traverse both vessel and background voxels. Amount conditions the pixel features before backprojection, while the prior assigns larger weights of 2D-derived features to likely occupied voxels along that ray, providing a more elegant way. The refiner fuses raw and weighted features to reconstruct the volume.

On 100 ImageCAS test cases, BiAAR gains 1.15 Dice percentage points over each of two controls: replacing amount with centroid depth, or replacing prior-guided fusion with direct concatenation (case-paired 95% CIs: 0.87–1.42 and 0.89–1.41 points, respectively). Fixed-architecture ablations also favor supervising both branches.

BiAAR reconstructs a 128³ volume in 0.12 s per forward pass (batch size one, RTX 4090D). Trained on clean 90° pairs (Dice 0.809), the same checkpoint achieves 0.786 at an unseen 45° separation, 0.799 with a 0.3° rotational calibration error bound, and 0.744 after localized removal of 5% of mask foreground. These tests quantify sensitivity under simulated acquisition conditions.

PTQ5 found the decomposition sensible and highlighted the evaluation across six metrics. ZydY recognized the practical value of the reconstruction accuracy and efficiency, and anyZ found that the matched CONCAT control supported the fusion design. Both ratings moved from 2 to 3, bringing the total to three weak acceptances. We thank all four reviewers and ACs for helping us test the design more directly and clarify its scope.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.