acceptodds
Under review as a conference paper at ICLR 2027

AID-GenReg: Anatomical Identity-Preserving Depth Generation for Enhancing Preoperative-to-Intraoperative Liver Point Cloud Registration

Abstract

Accurate rigid registration between complete preoperative and partial intraoperative liver point clouds is essential for reliable augmented reality visualization of internal hepatic anatomy during laparoscopic liver resection. However, the scarcity of paired clinical data, coupled with the unavailability of ground-truth transformations, hinders the performance of registration models trained via self-supervision under surgical conditions. Diffusion-based generative data construction presents a promising solution to this data scarcity by synthesizing geometrically diverse observations using limited paired data, producing abundant point cloud pairs to advance downstream registration. Nevertheless, adapting this scheme to clinical registration poses a distinct challenge, as existing methods do not explicitly enforce patient-specific anatomical identity between synthetic intraoperative observations and their corresponding preoperative models, yielding unreliable pairs that deteriorate subsequent self-supervised learning. In this work, we present AID-GenReg, a novel framework that constructs rich anatomy-consistent synthetic pairs to enhance preoperative-to-intraoperative liver point cloud registration. Specifically, AID-GenReg reprojects intraoperative observations and employs a diffusion-based anatomical identity-preserving depth generator to inpaint the resulting missing regions. To preserve anatomical identity represented in the preoperative model, we devise a region-decoupled generation scheme comprising two synergistic mechanisms: Geometry Preservation in observable regions with valid depth and preoperative Anatomy-Guided Completion in complementary missing regions. Such mechanisms ensure that the synthetic pairs offer reliable supervision for registration. Extensive experiments on a public in vivo human liver dataset under three downstream training paradigms demonstrate that AID-GenReg outperforms competitive generative alternatives and achieves consistent gains across multiple registration models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Preoperative-to-intraoperative registration, point cloud registration, generative data construction, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oc7wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies data augmentation for preoperative-to-intraoperative liver point cloud registration, where paired in vivo data are scarce and ground-truth rigid transformations are generally unavailable. The proposed AID-GenReg framework generates additional intraoperative depth observations from existing clinical pairs. It first reprojects a real intraoperative depth map under a sampled rigid transformation and then completes the missing regions with a diffusion model. The main design is a region-decoupled generation strategy: Geometry Preservation keeps valid reprojected depth unchanged in observable regions, while Anatomy-Guided Completion uses the corresponding preoperative liver model to guide generation in the missing regions through a differentiable geometric energy. The resulting real and synthetic pairs are used for self-supervised registration training. The method is evaluated on the public P2I-LReg in vivo liver dataset using four registration backbones and three training paradigms. AID-GenReg consistently improves downstream registration compared with real-only fine-tuning and with the SIRA-PCR and PointRegGPT data-construction baselines. The paper also includes ablations on reprojection, Geometry Preservation, Anatomy-Guided Completion, synthetic-data ratio, and guidance strength.

Critical strengths

The main strength is that the paper addresses a real limitation of generative data construction for clinical registration rather than treating synthetic target generation as a generic augmentation problem. In this setting, a generated intraoperative surface is only useful if it remains geometrically compatible with the paired patient-specific preoperative model. The separation between preserving measured geometry in observable regions and guiding completion only in missing regions is well motivated for this problem. I also found the downstream evaluation reasonably convincing. The authors do not rely on a single registration architecture: the generated pairs are evaluated with CoFiNet, GeoTransformer, P2P, and Self-P2IR, and the gains are generally consistent across models. The comparisons with SIRA-PCR and PointRegGPT are particularly useful because they test whether the proposed anatomical guidance provides additional value beyond existing simulation- or diffusion-based pair construction. The experiments are further conducted on public in-vivo human liver data rather than only simulation or phantom data.

Critical weaknesses

My main concern is that the central claim of patient-specific anatomical identity preservation is only indirectly validated. The ablation shows that Anatomy-Guided Completion improves downstream registration, but this does not by itself establish that the improvement comes from preserving the identity of the correct patient anatomy. A particularly informative control would be to replace the paired preoperative model with a mismatched model from another patient during guidance. If the proposed mechanism truly captures patient-specific identity rather than providing a generic liver-shape prior, the performance and geometric consistency should deteriorate under mismatched guidance. At present, this distinction is not directly demonstrated. A second concern is that Anatomy-Guided Completion is not fully independent of the downstream registration model. According to the appendix, the anatomical energy uses the Sim-Pretrain version of the downstream registration backbone under evaluation to coarsely align the preoperative model before computing the guidance signal. This means that the synthetic data used to evaluate different registration models may themselves depend on those models. As a result, the cross-backbone experiments do not cleanly establish that a single model-agnostic synthetic dataset improves different registration architectures. Using one fixed coarse aligner to generate the same synthetic pairs for all downstream models would provide a cleaner test.

Questions for authors

Q1: The paper's central claim is patient-specific anatomical identity preservation, but it is supported mainly indirectly through downstream registration performance. Could the authors provide a direct identity-preservation experiment, for example, by comparing guidance from the correctly paired preoperative liver with guidance from a mismatched liver model from another patient? A clear degradation under mismatched guidance would provide much stronger evidence that AGC preserves patient-specific anatomy rather than acting as a generic liver-shape prior.

Q2: Could the authors directly evaluate the geometric quality of the generated synthetic depth maps, rather than only their downstream utility? In particular, it would be helpful to report separate consistency measures for the observable region and the completed missing region, and to compare these quantities with PointRegGPT and SIRA-PCR. This would help verify that the proposed Geometry Preservation and Anatomy-Guided Completion actually improve the intended geometric properties of the generated samples.

Q3: The in vivo test set contains 92 pairs from only five patients. Could the authors report per-patient registration results, together with confidence intervals or a patient-level statistical analysis? This would help determine whether the improvements in Table 1 are consistent across patients or dominated by a small subset of cases.

Q4: Since Anatomy-Guided Completion depends on coarse preoperative-to-intraoperative alignment before computing the anatomical energy, how sensitive is the method to errors in this initial alignment? A controlled experiment with different coarse aligners or deliberately perturbed initial poses would help establish whether AGC remains beneficial when the coarse registration is inaccurate, which seems important for the intended clinical setting.

6 comments
AuthorsOpenReview ↗

Response to Reviewer oc7w (Part 1)

We are grateful for your thorough review and encouraging feedback, and for recognizing our clear presentation, well-motivated region-decoupled design, and convincing downstream evaluation. Below, we address your concerns by directly evaluating patient-specific anatomical identity preservation, demonstrating the generality of our augmentation across registration models, reporting patient-level results, and presenting additional experiments and analyses.


W1 & Q1: About Patient-Specific Anatomical Identity Preservation

  • W1: My main concern is that the central claim of patient-specific anatomical identity preservation is only indirectly validated. The ablation shows that Anatomy-Guided Completion improves downstream registration, but this does not by itself establish that the improvement comes from preserving the identity of the correct patient anatomy. A particularly informative control would be to replace the paired preoperative model with a mismatched model from another patient during guidance. If the proposed mechanism truly captures patient-specific identity rather than providing a generic liver-shape prior, the performance and geometric consistency should deteriorate under mismatched guidance. At present, this distinction is not directly demonstrated.
  • Q1: The paper's central claim is patient-specific anatomical identity preservation, but it is supported mainly indirectly through downstream registration performance. Could the authors provide a direct identity-preservation experiment, for example, by comparing guidance from the correctly paired preoperative liver with guidance from a mismatched liver model from another patient? A clear degradation under mismatched guidance would provide much stronger evidence that AGC preserves patient-specific anatomy rather than acting as a generic liver-shape prior.

A: To directly validate our central claim of patient-specific anatomical identity preservation, we compare AGC guided by correctly paired and mismatched CT models. In the mismatched condition, we align another training patient’s CT to the original intraoperative observation using the same frozen coarse aligner. All other generation and downstream training protocols remain unchanged, and the original patient’s correctly paired CT is retained for both registration pair construction and geometric evaluation.

Table A1: Mismatched versus patient-specific AGC guidance, evaluated by anatomical consistency and CoFiNet registration. Anatomical consistency is measured by unidirectional Chamfer (UCD) and 95th-percentile unidirectional Hausdorff (UHD95) distances from completed (synthesized) regions (C) to the correct CT under the same reference alignments. Accuracy is in % and distances in mm.

AGC guidanceUCD-C ↓UHD95-C ↓Acc@8 ↑Acc@15 ↑Mean Err. ↓
Mismatched20.7939.734.3564.1315.00
Patient-specific (Ours)19.5836.8511.9671.7414.07

As shown in Table A1, mismatched guidance increases UCD-C/UHD95-C by 1.21/2.88 mm compared with patient-specific guidance. It also lowers both Acc@8 and Acc@15 by 7.61 percentage points and raises mean registration error by 0.93 mm. This deterioration in geometric consistency and downstream performance supports the value of correctly paired patient anatomy beyond a generic liver-shape prior.

AuthorsOpenReview ↗

Response to Reviewer oc7w (Part 2)

W2: Decoupling AGC from the Evaluated Registration Backbone

A second concern is that Anatomy-Guided Completion is not fully independent of the downstream registration model. According to the appendix, the anatomical energy uses the Sim-Pretrain version of the downstream registration backbone under evaluation to coarsely align the preoperative model before computing the guidance signal. This means that the synthetic data used to evaluate different registration models may themselves depend on those models. As a result, the cross-backbone experiments do not cleanly establish that a single model-agnostic synthetic dataset improves different registration architectures. Using one fixed coarse aligner to generate the same synthetic pairs for all downstream models would provide a cleaner test.

A: In our default setting, AGC uses the Sim-Pretrain version of the evaluated registration backbone as its coarse aligner. To verify that a single synthetic dataset benefits different registration architectures, we decouple AGC from the evaluated registration backbone by fixing Self-P2IR (Sim-Pretrain) as the coarse aligner and using the same synthetic pairs for all four downstream models. Table A2 reports each registration model’s performance using this same synthetic dataset.

Table A2: Registration performance under Real-FT and RealSyn-FT. Chamfer accuracy is in % and mean Chamfer error in mm.

Registration ModelTraining ParadigmAcc@8 ↑Acc@15 ↑Mean Err. ↓
CoFiNetReal-FT8.7053.2616.07
RealSyn-FT8.7064.1314.03
GeoTransformerReal-FT11.9640.2217.96
RealSyn-FT21.7460.8714.83
P2PReal-FT43.4896.748.76
RealSyn-FT48.9196.748.55
Self-P2IRReal-FT45.6596.748.70
RealSyn-FT51.0997.838.48

Compared with Real-FT, all four registration models achieve lower mean error. In particular, CoFiNet and GeoTransformer improve Acc@15 by 10.87 and 20.65 percentage points, respectively. These results demonstrate that a single synthetic dataset generated by AID-GenReg benefits different registration architectures, underscoring its generality as a generative augmentation tool for clinical registration.


Q2: Direct Geometric Evaluation

Could the authors directly evaluate the geometric quality of the generated synthetic depth maps, rather than only their downstream utility? In particular, it would be helpful to report separate consistency measures for the observable region and the completed missing region, and to compare these quantities with PointRegGPT and SIRA-PCR. This would help verify that the proposed Geometry Preservation and Anatomy-Guided Completion actually improve the intended geometric properties of the generated samples.

A: We directly quantify anatomical identity preservation in observable and completed (synthesized) regions using paired preoperative CT surfaces as references. Table A3 compares anatomical consistency in both regions for PointRegGPT, our method and its variant without AGC.

Table A3: Anatomical consistency with paired preoperative CT surfaces under the same coarse alignments, measured by unidirectional Chamfer (UCD) and 95th-percentile unidirectional Hausdorff (UHD95) distances (mm). O/C denote observable/completed (synthesized) regions.

MethodUCD-O ↓UHD95-O ↓UCD-C ↓UHD95-C ↓
PointRegGPT16.8032.7328.3754.46
Ours w/o AGC16.7132.7023.9545.69
Ours16.7132.6919.5836.85

Observable-region consistency remains comparable. In synthesized regions, our variant without AGC lowers UCD-C/UHD95-C by 4.42/8.77 mm compared with PointRegGPT, and adding AGC yields further reductions of 4.37/8.84 mm. These results demonstrate the effectiveness of GP and AGC in improving anatomical consistency in synthesized regions.

We do not include SIRA-PCR in Table A3 because it adapts CT-rendered point clouds using a GAN and thus has no corresponding division into observable and completed regions. Although its whole-cloud UCD/UHD95 values are low (9.47/18.14 mm), these are obtained using known rendering poses and are not directly comparable to the regional metrics computed under estimated coarse alignments.

AuthorsOpenReview ↗

Response to Reviewer oc7w (Part 3)

Q3: Patient-Level Registration Results

The in vivo test set contains 92 pairs from only five patients. Could the authors report per-patient registration results, together with confidence intervals or a patient-level statistical analysis? This would help determine whether the improvements in Table 1 are consistent across patients or dominated by a small subset of cases.

A: To examine whether the gains in Table 1 generalize across patients, we report per-patient results for CoFiNet and Self-P2IR, comparing Real-FT with AID-GenReg (Tables A4 and A5). Uncertainty is quantified using 95% confidence intervals (CIs), with two-sided Wilson intervals for Acc@8/Acc@15. For mean error, we use non-parametric percentile bootstrap resampling of each patient's test pairs (2,000 iterations), drawing with replacement and taking the 2.5th and 97.5th percentiles of the resampled means as CI bounds.

Table A4: Per-patient CoFiNet registration results, reported as value [95% CI]. The first row for each patient is the Real-FT baseline. Chamfer accuracy is in % and mean Chamfer error in mm.

Patient (#Pairs)MethodAcc@8 ↑Acc@15 ↑Mean Err. ↓
01 (22)-0.00 [0.00, 14.87]31.82 [16.36, 52.68]18.69 [16.10, 21.25]
Ours4.55 [0.81, 21.80]68.18 [47.32, 83.64]13.92 [12.18, 15.77]
02 (20)-0.00 [0.00, 16.11]35.00 [18.12, 56.71]19.57 [16.20, 23.68]
Ours0.00 [0.00, 16.11]45.00 [25.82, 65.79]20.32 [15.93, 25.28]
03 (21)-14.29 [4.98, 34.64]57.14 [36.55, 75.53]15.73 [12.61, 19.29]
Ours19.05 [7.67, 40.00]85.71 [65.36, 95.02]12.46 [10.08, 15.54]
04 (11)-0.00 [0.00, 25.88]63.64 [35.38, 84.83]13.45 [10.79, 16.53]
Ours9.09 [1.62, 37.74]54.55 [28.01, 78.73]13.04 [10.96, 14.94]
05 (18)-27.78 [12.50, 50.87]88.89 [67.20, 96.90]11.01 [9.50, 12.57]
Ours27.78 [12.50, 50.87]100.00 [82.41, 100.00]9.79 [8.73, 10.90]

Table A5: Per-patient Self-P2IR registration results, with notation and units following Table A4.

Patient (#Pairs)MethodAcc@8 ↑Acc@15 ↑Mean Err. ↓
01 (22)-18.18 [7.31, 38.52]86.36 [66.67, 95.25]10.85 [9.71, 12.07]
Ours27.27 [13.15, 48.15]90.91 [72.19, 97.47]10.45 [9.33, 11.60]
02 (20)-35.00 [18.12, 56.71]100.00 [83.89, 100.00]9.52 [8.64, 10.53]
Ours35.00 [18.12, 56.71]100.00 [83.89, 100.00]9.40 [8.61, 10.26]
03 (21)-61.90 [40.88, 79.25]100.00 [84.54, 100.00]7.15 [6.40, 7.96]
Ours61.90 [40.88, 79.25]100.00 [84.54, 100.00]7.06 [6.26, 7.83]
04 (11)-90.91 [62.26, 98.38]100.00 [74.12, 100.00]6.99 [6.37, 7.72]
Ours90.91 [62.26, 98.38]100.00 [74.12, 100.00]6.90 [6.40, 7.47]
05 (18)-44.44 [24.56, 66.28]100.00 [82.41, 100.00]7.99 [7.15, 8.84]
Ours61.11 [38.62, 79.69]100.00 [82.41, 100.00]7.68 [6.87, 8.58]

On CoFiNet, AID-GenReg achieves higher Acc@15 and lower mean error in 4/5 patients for each metric, while maintaining or improving Acc@8 in 5/5 patients. On Self-P2IR, it reduces mean error in 5/5 patients and maintains or improves both accuracy metrics for every patient. Together, these findings support generally consistent augmentation benefits across patients with both registration backbones.

Reviewer_K3YFOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AID-GenReg to alleviate the shortage of clinical training data for preoperative-to-intraoperative liver point cloud registration. It synthesizes novel-view depth maps through reprojection and diffusion-based completion, preserving observed geometry and using the preoperative liver model to guide missing regions. The resulting point cloud pairs improve self-supervised registration across four backbones.

Critical strengths
  • The paper addresses a clinically relevant data scarcity problem.
  • The evaluation includes real human data, multiple registration models, and comparisons with generative baselines.
Critical weaknesses

1. Limited methodological novelty (main concern). The overall pipeline closely resembles PointRegGPT [1], which also combines novel-view depth reprojection and diffusion-based completion to generate point cloud registration pairs. The broader idea of anatomy-preserving liver image synthesis has also been explored, for example in [2].

Moreover, the paper's two core components are closely related to established techniques, as the authors themselves acknowledge:

  • Geometry Preservation (Eq. 1) follows the known-region reinjection principle used in RePaint [3].
  • Anatomy-Guided Completion (Eqs. 2–5) uses a Chamfer-distance energy to guide the predicted clean depth. In particular, Eq. 3 follows the clean-estimate gradient guidance shortcut of MPGD [4].

While the application is meaningful, the method appears to be primarily an integration of existing techniques rather than a sufficiently novel algorithmic contribution.

2. Limited evidence of generalization. The additional gains from anatomical guidance are modest in some cases. For example, on Self-P2IR, adding AGC improves Acc@8 from 48.91% to 51.09%, while mean Chamfer error slightly increases from 8.47 to 8.48 mm (Table 2). More importantly, the clinical test set includes only five patients (92 pairs). Without patient-level results or uncertainty estimates, it is difficult to judge how well the gains generalize across patients and surgical conditions.

Questions for authors
  1. What is the key algorithmic contribution beyond PointRegGPT [1] and existing diffusion guidance techniques [3, 4]? A baseline combining PointRegGPT with straightforward anatomy-based gradient guidance would help clarify this.
  2. Can anatomical identity preservation be evaluated directly in synthesized regions, rather than primarily through downstream registration performance?
  3. Can the authors provide patient-level results or evaluate on more independent clinical cases to substantiate generalization?

References

[1] Chen et al. PointRegGPT: Boosting 3D Point Cloud Registration Using Generative Point-Cloud Pairs for Training. ECCV, 2024.

[2] Nguyen et al. Boosting 3D Liver Shape Datasets with Diffusion Models and Implicit Neural Representations. MICCAI, 2025.

[3] Lugmayr et al. RePaint: Inpainting Using Denoising Diffusion Probabilistic Models. CVPR, 2022.

[4] He et al. Manifold Preserving Guided Diffusion. ICLR, 2024.

6 comments
AuthorsOpenReview ↗

Response to Reviewer K3YF (Part 1)

We appreciate your thoughtful review and are glad that you found our paper clear and helpful in deepening your understanding of the clinical problem. Below, we address your concerns by clarifying our methodological contributions, providing patient-level analysis, directly quantifying anatomical identity preservation, and presenting additional experiments and analyses.


W1 & Q1: About the Methodological Novelty

1. Key Contributions Beyond Prior Approaches

  • W1: The overall pipeline closely resembles PointRegGPT [1], which also combines novel-view depth reprojection and diffusion-based completion to generate point cloud registration pairs. The broader idea of anatomy-preserving liver image synthesis has also been explored, for example in [2].
  • Q1: What is the key algorithmic contribution beyond PointRegGPT [1]? A baseline combining PointRegGPT with straightforward anatomy-based gradient guidance would help clarify this.

[1] Chen et al. PointRegGPT: Boosting 3D Point Cloud Registration Using Generative Point-Cloud Pairs for Training. ECCV, 2024.
[2] Nguyen et al. Boosting 3D Liver Shape Datasets with Diffusion Models and Implicit Neural Representations. MICCAI, 2025.

A: Compared with PointRegGPT, our key contributions lie in a novel methodological concept of anatomical identity-preserving synthesis tailored for reliable clinical registration and a heuristic region-decoupled inference scheme incorporating Anatomy-Guided Completion. We highlight key differences below:

  • Conceptual contribution. PointRegGPT constrains observed regions but leaves completion to a learned depth prior, which may produce geometry that deviates from the paired patient’s preoperative anatomy. To tackle this issue, we propose anatomical identity-preserving synthesis, which additionally requires completed geometry to remain compatible with the paired patient-specific preoperative anatomy.

  • Algorithmic contribution. We realize this concept through a region-decoupled inference scheme that coordinates geometric anchoring in observable regions with patient-specific anatomical guidance in missing regions in a cascaded manner.

    • Geometry Preservation (GP) uses noised reprojected observations to anchor observable regions before each denoiser prediction, whereas PointRegGPT constrains the predicted clean depth via DDNM. We argue that repeated input anchoring may limit error propagation into missing regions. Compared with PointRegGPT in Table A1, our variant without AGC reduces UCD-C and UHD95-C by 4.42 mm and 8.77 mm, respectively, and raises CoFiNet Acc@15 from 52.17% to 63.04%. These results demonstrate GP’s tangible algorithmic advantages for anatomical identity-preserving registration pair synthesis.

    • Anatomy-Guided Completion (AGC) formulates a patient-specific anatomical energy measuring one-way 3D discrepancy to the paired preoperative surface and minimizes it through gradient-descent updates restricted to missing regions. This guides newly synthesized geometry toward the paired patient’s anatomy while retaining measured geometry.

Table A1: Further comparisons on anatomical consistency of completed (i.e., synthesized) regions with paired preoperative CT surfaces under the same coarse alignments, and downstream registration using CoFiNet. Metrics are unidirectional Chamfer (UCD-C) and 95th-percentile unidirectional Hausdorff (UHD95-C) distances (mm), Chamfer accuracy (%), and mean Chamfer error (mm).

MethodUCD-C ↓UHD95-C ↓Acc@8 ↑Acc@15 ↑Mean Err. ↓
PointRegGPT28.3754.467.6152.1715.91
PointRegGPT + AGC26.8752.056.5260.8714.12
Ours w/o AGC23.9545.6910.8763.0414.45
Ours19.5836.8511.9671.7414.07

Comparison with anatomy-guided PointRegGPT. We construct PointRegGPT + AGC by retaining DDNM's observed-region correction and applying AGC to missing regions of the predicted clean depth before reverse sampling. Table A1 shows that adding AGC improves PointRegGPT's anatomical consistency, Acc@15 and mean error, while AID-GenReg achieves further gains on all five metrics. These results substantiate the algorithmic contribution of our synergistic design of GP and AGC.

Clarification regarding [2]. We respectfully disagree that [2] addresses the same patient-specific preservation objective. This approach combines unconditional diffusion with implicit neural representations to model anatomical variation and expand liver-shape datasets. Instead, our formulation requires patient-specific anatomical identity preservation between synthesized intraoperative geometry and its paired preoperative model for reliable registration pair construction.

AuthorsOpenReview ↗

Response to Reviewer K3YF (Part 2)

W1 & Q1: About the Methodological Novelty

2. Algorithmic Contribution Beyond Existing Techniques

  • W1: Moreover, the paper's two core components are closely related to established techniques, as the authors themselves acknowledge:

    • Geometry Preservation (Eq. 1) follows the known-region reinjection principle used in RePaint [3].
    • Anatomy-Guided Completion (Eqs. 2–5) uses a Chamfer-distance energy to guide the predicted clean depth. In particular, Eq. 3 follows the clean-estimate gradient guidance shortcut of MPGD [4].

    While the application is meaningful, the method appears to be primarily an integration of existing techniques rather than a sufficiently novel algorithmic contribution.

  • Q1: What is the key algorithmic contribution beyond existing diffusion guidance techniques [3, 4]?

[3] Lugmayr et al. RePaint: Inpainting Using Denoising Diffusion Probabilistic Models. CVPR, 2022.
[4] He et al. Manifold Preserving Guided Diffusion. ICLR, 2024.

A: We respectfully disagree with the claim that AID-GenReg is primarily an integration of existing techniques. Our main algorithmic contributions are a heuristic region-decoupled inference strategy and the AGC mechanism for anatomical identity-preserving clinical registration pair synthesis. The cascaded coordination of GP and AGC preserves measured geometry while guiding newly synthesized regions toward the paired patient’s preoperative anatomy.

  • GP and RePaint. GP adapts RePaint’s known-region replacement for image inpainting to geometric anchoring across viewpoints, using reprojected intraoperative observations to constrain novel-view depth synthesis. A final clean-state replacement ensures exact consistency with observed depths.

  • AGC and MPGD. We clarify the differences in objectives and algorithmic design between the two techniques. MPGD focuses on data-manifold preservation by constraining guidance directions to manifold tangent spaces during conditional generation, while AGC introduces a patient-specific anatomical energy and region-restricted guidance for anatomical identity-preserving synthesis.

    • Anatomical energy. We formulate a patient-specific anatomical energy (Eq. 2) measuring one-way 3D discrepancy from synthesized intraoperative points to the paired preoperative surface. This formulation reflects the clinical prior that the intraoperative surface is a partial observation of the same patient’s preoperative anatomy.
    • Anatomical guidance from the preoperative surface. At each guided step, we minimize the anatomical energy on the predicted clean depth using gradient descent (Eq. 3), which is a natural and widely used method for minimizing differentiable objectives in machine learning. Our guidance design (Eqs. 4-5) masks the energy gradient to missing regions, normalizes it, and scales it by the missing-depth norm, thereby controlling the relative update magnitude while leaving observable regions unchanged.

In summary, these contributions establish the methodological novelty of AID-GenReg beyond an integration of existing techniques.

AuthorsOpenReview ↗

Response to Reviewer K3YF (Part 3)

W2 & Q3: Anatomical Guidance Gains and Patient-Level Generalization

  • W2: The additional gains from anatomical guidance are modest in some cases. For example, on Self-P2IR, adding AGC improves Acc@8 from 48.91% to 51.09%, while mean Chamfer error slightly increases from 8.47 to 8.48 mm (Table 2). More importantly, the clinical test set includes only five patients (92 pairs). Without patient-level results or uncertainty estimates, it is difficult to judge how well the gains generalize across patients and surgical conditions.
  • Q3: Can the authors provide patient-level results or evaluate on more independent clinical cases to substantiate generalization?

A: Anatomy-Guided Completion (AGC) plays a key role in preserving anatomical identity, working in synergy with Geometry Preservation within our region-decoupled inference strategy. We present additional empirical evidence below to demonstrate the gains of our anatomical guidance.

  • Direct anatomical consistency. We compare the anatomical consistency of synthesized regions produced by our method and its variant without AGC (Table A1). Adding AGC lowers UCD-C by 4.37 mm and UHD95-C by 8.84 mm, bringing synthesized geometry closer to the paired patient’s preoperative anatomy.

  • Ablation in downstream registration. We supplement the original CoFiNet and Self-P2IR ablations (Table 2) with GeoTransformer. With AGC, GeoTransformer’s Acc@8/Acc@15 rise from 14.13%/59.78% to 17.39%/61.96%, while mean Chamfer error falls by 0.93 mm. CoFiNet also gains 8.70 percentage points in Acc@15 (Table 2). Although the extra gain on Self-P2IR is modest, the geometric consistency improvements and multi-backbone ablations support AGC’s effectiveness.

Given the cumbersome clinical data acquisition process (as discussed in the Introduction), we instead assess generalization through patient-level analysis (Table A2). Specifically, we report per-patient results for CoFiNet, comparing Real-FT, our variant without AGC and AID-GenReg. Uncertainty is quantified using 95% confidence intervals (CIs), with two-sided Wilson intervals for Acc@8/Acc@15. Mean-error CIs are obtained using non-parametric percentile bootstrap resampling of each patient’s test pairs (2,000 iterations). We resample with replacement, recompute mean error, and take the 2.5th and 97.5th percentiles as the 95% CI bounds.

Table A2: Per-patient CoFiNet registration results, reported as value [95% CI]. The first row for each patient is the Real-FT baseline. Chamfer accuracy is in % and mean Chamfer error in mm.

Patient (#Pairs)MethodAcc@8 ↑Acc@15 ↑Mean Err. ↓
01 (22)-0.00 [0.00, 14.87]31.82 [16.36, 52.68]18.69 [16.10, 21.25]
Ours w/o AGC9.09 [2.53, 27.81]59.09 [38.73, 76.74]13.80 [11.95, 15.52]
Ours4.55 [0.81, 21.80]68.18 [47.32, 83.64]13.92 [12.18, 15.77]
02 (20)-0.00 [0.00, 16.11]35.00 [18.12, 56.71]19.57 [16.20, 23.68]
Ours w/o AGC0.00 [0.00, 16.11]45.00 [25.82, 65.79]18.56 [15.85, 21.30]
Ours0.00 [0.00, 16.11]45.00 [25.82, 65.79]20.32 [15.93, 25.28]
03 (21)-14.29 [4.98, 34.64]57.14 [36.55, 75.53]15.73 [12.61, 19.29]
Ours w/o AGC19.05 [7.67, 40.00]76.19 [54.91, 89.37]12.54 [10.24, 15.25]
Ours19.05 [7.67, 40.00]85.71 [65.36, 95.02]12.46 [10.08, 15.54]
04 (11)-0.00 [0.00, 25.88]63.64 [35.38, 84.83]13.45 [10.79, 16.53]
Ours w/o AGC9.09 [1.62, 37.74]27.27 [9.75, 56.56]18.87 [14.64, 24.02]
Ours9.09 [1.62, 37.74]54.55 [28.01, 78.73]13.04 [10.96, 14.94]
05 (18)-27.78 [12.50, 50.87]88.89 [67.20, 96.90]11.01 [9.50, 12.57]
Ours w/o AGC16.67 [5.84, 39.22]94.44 [74.24, 99.01]10.17 [9.04, 11.42]
Ours27.78 [12.50, 50.87]100.00 [82.41, 100.00]9.79 [8.73, 10.90]

Table A2 shows that adding AGC improves Acc@15 in 4/5 patients. Compared with Real-FT, AID-GenReg achieves higher Acc@15 and lower mean error in 4/5 patients for each metric, while matching or improving Acc@8 in 5/5 patients. In summary, these findings show that AGC’s accuracy gains and the benefits of generative augmentation are generally consistent across patients.

Reviewer_kwKKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The goal is to align a patient's liver model from pre-surgery CT with the partial liver surface seen through a laparoscope during surgery. Only 252 real training pairs exist, so the authors generate more. They move the camera virtually, re-render the depth the laparoscope would see, and use a diffusion model to fill the newly exposed regions. The fill is guided so that it agrees with that patient's CT liver. The extra pairs improve four registration networks.

Critical strengths

The paper uses the patient's paired preoperative (CT) model as a UCD target to steer the completion of synthetic intraoperative depth. This is a sensible, application-specific formulation. Prior pipelines such as PointRegGPT constrain only the observed region and complete the rest from the learned depth prior alone. AID-GenReg additionally ties the completed region to the anatomy it will be paired with.

Critical weaknesses
  1. The central claim that enforcing anatomical identity is why it beats existing methods isn't directly supported.
  • 1.A Consistency with the CT model is never measured directly. The paper asserts that existing methods "do not explicitly enforce patient-specific anatomical identity", yielding unreliable pairs. However, no experiment measures how consistent any method's (including this paper) synthetic depth is with the preoperative model. Possible measurements include the UCD between generated depth and the aligned preoperative surface, or the error in the completed regions on the simulation subset of P2I-LReg, where the true geometry is known.

  • 1.B The paper doesn't say how SIRA-PCR and PointRegGPT were adapted to liver data, and they're compared in only one setting (RealSyn-FT, 1:3 ratio; Table 1).

  • 1.C The Table 1 vs. Table 2 comparison: the variant without anatomy guidance already accounts for most of the gain over PointRegGPT. For Self-P2IR, anatomy guidance adds only 2 of 92 test pairs.

  1. Geometry Preservation has essentially no algorithmic novelty. Its mechanism is the known-region replacement of RePaint (Lugmayr et al., 2022) cited in the paper. Also, the main baseline PointRegGPT already enforces consistency in the observed region through DDNM's range–null-space decomposition.
Questions for authors
  1. How were SIRA-PCR and PointRegGPT adapted to this setting?

  2. Could the authors quantify anatomical consistency directly for each construction method? For example, report the UCD between the generated depth and the aligned preoperative surface.

6 comments
AuthorsOpenReview ↗

Response to Reviewer kwKK (Part 1)

We sincerely thank you for your careful assessment and constructive feedback, and for recognizing our patient-specific anatomical guidance as a “sensible, application-specific formulation”. Below, we address your concerns by directly quantifying anatomical consistency, clarifying our methodological contributions, and presenting additional experiments and analyses.


W1: About Anatomical Identity Preservation

The central claim that enforcing anatomical identity is why it beats existing methods isn't directly supported.

A: We address this concern through additional experiments and clarifications in W1.A-W1.C.

W1.A & Q2: Direct Anatomical Consistency Measurement

  • Consistency with the CT model is never measured directly. The paper asserts that existing methods "do not explicitly enforce patient-specific anatomical identity", yielding unreliable pairs. However, no experiment measures how consistent any method's (including this paper) synthetic depth is with the preoperative model. ...
  • Could the authors quantify anatomical consistency directly for each construction method? For example, report the UCD between the generated depth and the aligned preoperative surface.

A: As shown in Table A1, we directly evaluate the anatomical consistency of synthesized regions produced by PointRegGPT and our method against paired preoperative CT surfaces. We do not include SIRA-PCR in this synthesized-region comparison because it uses a GAN to adapt CT-rendered point clouds to the real intraoperative domain rather than synthesize new regions from real intraoperative observations.

Table A1: Anatomical consistency of completed (i.e., synthesized) regions with paired preoperative CT surfaces under the same coarse alignments, measured by unidirectional Chamfer (UCD-C) and 95th-percentile unidirectional Hausdorff (UHD95-C) distances (mm).

MethodUCD-C ↓UHD95-C ↓
PointRegGPT28.3754.46
Ours19.5836.85

Compared with PointRegGPT, our method reduces UCD-C and UHD95-C by 8.79 mm and 17.61 mm, respectively, providing direct geometric evidence of improved anatomical identity preservation. Besides, the domain-adapted point clouds generated by SIRA-PCR exhibit high anatomical consistency (whole-cloud UCD/UHD95 of 9.47/18.14 mm) due to their CT-rendered source geometry. However, its reliance on rendered data still limits downstream registration performance compared with our method (as discussed in Related Work and shown in Table 1).

AuthorsOpenReview ↗

Response to Reviewer kwKK (Part 2)

W1.B & Q1: Generative Baseline Adaptation and Broader Evaluation

  • The paper doesn't say how SIRA-PCR and PointRegGPT were adapted to liver data, and they're compared in only one setting (RealSyn-FT, 1:3 ratio; Table 1).
  • How were SIRA-PCR and PointRegGPT adapted to this setting?

A: We adapted both generative baselines to the liver setting using data from the training patients. For SIRA-PCR, we trained a GAN for sim-to-real adaptation using CT-rendered liver point clouds as the source and real intraoperative point clouds as the target. During inference, we used the trained GAN to adapt CT-rendered point clouds from training patients to real intraoperative style, then paired the outputs with the corresponding preoperative CT models to form synthetic registration pairs. For PointRegGPT, we retrained the depth diffusion model on our real intraoperative liver depth maps. During generation, we first reprojected these observations to new viewpoints, then completed missing regions through DDNM-constrained diffusion, and finally paired the resulting point clouds with the corresponding preoperative CT models to form synthetic registration pairs.

We have also incorporated a broader comparison of generative data construction methods across different settings (i.e., synthetic data scales and downstream training paradigms), as shown in Tables A2 and A3.

Table A2: Comparison of generative data construction methods across real-to-synthetic ratios using CoFiNet under RealSyn-FT, measured by Chamfer accuracy (%) and mean Chamfer error (mm). The first row is the Real-FT baseline.

Real:SynMethodAcc@8 ↑Acc@15 ↑Mean Err. ↓
1:0-8.7053.2616.07
1:1SIRA-PCR1.0953.2617.99
1:1PointRegGPT6.5253.2615.87
1:1Ours8.7060.8715.17
1:2SIRA-PCR3.2646.7417.71
1:2PointRegGPT8.7057.6115.12
1:2Ours10.8765.2214.60
1:3SIRA-PCR5.4351.0917.64
1:3PointRegGPT7.6152.1715.91
1:3Ours11.9671.7414.07

Table A3: Comparison of generative data construction methods under different training paradigms using CoFiNet at a real:syn ratio of 1:3, measured by Chamfer accuracy (%) and mean Chamfer error (mm).

Training paradigmMethodAcc@8 ↑Acc@15 ↑Mean Err. ↓
RealSyn-FTSIRA-PCR5.4351.0917.64
RealSyn-FTPointRegGPT7.6152.1715.91
RealSyn-FTOurs11.9671.7414.07
Syn→Real-FTSIRA-PCR2.1746.7418.97
Syn→Real-FTPointRegGPT6.5256.5216.38
Syn→Real-FTOurs11.9658.7015.77
  • Synthetic data scale. Table A2 compares CoFiNet performance using different generative methods under varying real-to-synthetic point cloud pair ratios. AID-GenReg achieves the best results on all metrics at every ratio, with average Acc@8 gains of 7.25 and 2.90 percentage points (pp) over SIRA-PCR and PointRegGPT, respectively. These results underscore the effectiveness of AID-GenReg as a generative augmentation tool, which consistently improves registration performance with varying amounts of synthetic data.

  • Downstream training paradigm. Table A3 compares CoFiNet performance using different generative methods under RealSyn-FT and Syn→Real-FT paradigms. In Syn→Real-FT, CoFiNet is initialized from Sim-Pretrain, further pretrained on each method's synthetic pairs, and then fine-tuned on real pairs. AID-GenReg achieves the best results in both settings, demonstrating its effectiveness and generality under different downstream training paradigms.

AuthorsOpenReview ↗

Response to Reviewer kwKK (Part 3)

W1.C: Effectiveness of Our Anatomical Guidance

The Table 1 vs. Table 2 comparison: the variant without anatomy guidance already accounts for most of the gain over PointRegGPT. For Self-P2IR, anatomy guidance adds only 2 of 92 test pairs.

A: Anatomy-Guided Completion (AGC) plays a key role in preserving anatomical identity, working in synergy with Geometry Preservation within our proposed region-decoupled inference strategy. We present additional empirical evidence below to demonstrate the effectiveness of our anatomical guidance.

Table A4: Anatomical consistency of synthesized regions with and without AGC (mm).

MethodUCD-C ↓UHD95-C ↓
Ours w/o AGC23.9545.69
Ours19.5836.85
  • Direct anatomical consistency. We compare the anatomical consistency of synthesized regions produced by our method and its variant without AGC. As shown in Table A4, adding AGC reduces UCD-C and UHD95-C by 4.37 mm and 8.84 mm, respectively, showing that AGC makes synthesized geometry more consistent with the paired patient’s anatomy.

  • Ablation in downstream registration. We extend the CoFiNet and Self-P2IR ablations (Table 2) to the widely used GeoTransformer to further assess AGC’s downstream benefits. On GeoTransformer, adding AGC raises Acc@8/Acc@15 from 14.13%/59.78% to 17.39%/61.96% and reduces mean Chamfer error from 15.45 mm to 14.52 mm. On CoFiNet, AGC increases Acc@15 by 8.70 pp (Table 2). These ablations show that AGC improves Acc@8 on all three backbones at the stringent 8-mm threshold. Moreover, patient-level analysis using CoFiNet shows that AGC improves Acc@15 in four of five patients, indicating its benefits extend to multiple patients. Although the additional gain on Self-P2IR is modest, the direct geometric improvements, gains on multiple backbones and patient-level evidence collectively demonstrate AGC’s effectiveness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response Summary and Manuscript Updates

Dear Reviewers, ACs, and PCs,

We sincerely thank all reviewers for their valuable comments and insightful suggestions.

Strength summary

Generally, the reviewers recognize the clinical significance of our work and the value of patient-specific anatomical guidance, and appreciate our evaluation on real human data.

Specifically:

  • Reviewer kwKK describes our patient-specific anatomical guidance as a “sensible, application-specific formulation”.
  • Reviewer K3YF appreciates our clear, informative presentation and comprehensive evaluation on real human data with multiple registration models and generative baselines.
  • Reviewer oc7w recognizes our clear presentation, well-motivated region-decoupled design, and convincing downstream evaluation.

Concerns and discussion summary

We have responded to all concerns point by point. Reviewer oc7w reported increased confidence in our work and maintained the positive rating.

We welcome feedback from the other two reviewers on whether our responses adequately address their concerns.

Common concerns:

  • c1. Direct validation of our central claim of anatomical identity preservation (Reviewers kwKK, K3YF, oc7w).

    • We directly measure consistency in observable and synthesized regions against paired CT surfaces. AID-GenReg lowers synthesized-region UCD/UHD95 by 8.79/17.61 mm versus PointRegGPT, providing direct geometric evidence of improved anatomical identity preservation.
    • Adding Anatomy-Guided Completion (AGC) reduces synthesized-region UCD/UHD95 by 4.37/8.84 mm versus the variant without AGC, supporting its contribution to anatomical identity preservation.
  • c2. Clarification of methodological contributions (Reviewers kwKK, K3YF).

    • Our novel region-decoupled inference scheme uses Geometry Preservation (GP) to anchor denoising inputs to observed geometry and the proposed AGC mechanism to guide missing-region synthesis using patient-specific anatomical energy. We have clarified AID-GenReg’s algorithmic contributions and its distinctions from existing techniques (RePaint and MPGD).
    • Ablation studies on PointRegGPT and our method demonstrate AGC’s geometric and registration benefits, supporting its tangible algorithmic contribution.
    • AID-GenReg surpasses PointRegGPT + AGC on all five geometric and registration metrics, substantiating the synergistic design of GP and AGC.
  • c3. Patient-level generalization (Reviewers K3YF, oc7w).

    • We report per-patient CoFiNet and Self-P2IR results with 95% confidence intervals. Compared with the Real-FT baseline, CoFiNet improves Acc@15 and reduces mean error in 4/5 patients for each metric, while Self-P2IR lowers mean error in all five patients.
    • Adding AGC improves CoFiNet Acc@15 in 4/5 patients, supporting generally consistent AGC benefits across patients.

Specific concerns:

  • s1. Correct versus mismatched CT guidance (Reviewer oc7w).

    • Mismatching the guidance CT in AGC degrades geometric consistency and registration performance, providing empirical evidence that AGC preserves patient-specific anatomy rather than acting as a generic liver-shape prior.
  • s2. Decoupling AGC from the evaluated registration backbone (Reviewer oc7w).

    • With a fixed coarse aligner in AGC, the same synthetic dataset benefits all four registration models, underscoring AID-GenReg’s generality as a generative augmentation tool for clinical registration.
  • s3. AGC sensitivity to coarse aligner choice (Reviewer oc7w).

    • With GeoTransformer as the downstream registration model, adding AGC lowers mean error under all four coarse aligners, suggesting that AGC is generally beneficial under different coarse alignment conditions.
  • s4. Generative baseline adaptation and broader evaluation (Reviewer kwKK).

    • We have clarified both baselines’ liver adaptation. SIRA-PCR adapts CT-rendered clouds to intraoperative style using a GAN, while PointRegGPT uses a diffusion model trained on real intraoperative depth maps and retains its original reprojection and DDNM completion.
    • AID-GenReg outperforms SIRA-PCR and PointRegGPT at different synthetic data scales and under different downstream training paradigms (RealSyn-FT and Syn→Real-FT).

Manuscript updates

We have revised the manuscript and appendix, with changes highlighted in blue:

  1. Added direct geometric validation of anatomical identity preservation (Sec. 4.4, Table 4).
  2. Clarified generative baseline adaptation to our liver setting (Appendix A.3).
  3. Evaluated generative methods under different synthetic data scales and downstream training paradigms to assess generality (Appendix A.4, Tables 7-8).
  4. Included per-patient results with CIs to assess generalization (Appendix A.5, Tables 9-10).
  5. Assessed AGC’s robustness under different coarse alignment conditions (Appendix A.6, Table 11).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.