PaperlayerView exact revision on arXiv

Source-linked AI summary

When Misalignment Becomes Supervision: Structured Label Noise in Supervised Synthetic CT Generation

Valentin Boussot, Cedric Hemon, Caroline Lafond, Jean-Claude Nunes, Jean-Louis Dillenseger

arXiv:2609.29387v1cs.CV

TL;DR

Voxel-wise supervised sCT relies on registered references that can contain structured geometric errors, complicating what its metrics measure. The study analyzes registration effects and evaluates SAM-based perceptual supervision and evaluation alongside voxel-wise approaches. Scores depend on matching registration conventions, while SAM-based supervision improves downstream anatomical metrics and image structure in the evaluated settings.

  • Problem

    Residual registration errors in paired MRI/CBCT–CT data make voxel-wise reference targets imperfect, raising the question of whether scores reflect synthesis fidelity or agreement with the registration pipeline.

    Source: §I · pp. 1–2, Contributions · p. 2

  • Method

    The study analyzes registration-induced supervision bias and evaluates SAM-based perceptual supervision and a separately calibrated SAM-based evaluation metric as anatomy-oriented complements to voxel-wise approaches.

    Source: Contributions · p. 2, D. SAM-based perceptual evaluation… · p. 6

  • Results

    Voxel-wise performance improves when training and evaluation share a registration convention, while higher-anatomical-consistency supervision is associated with lower prediction variability and better out-of-distribution performance; SAM-based supervision improves downstream anatomical metrics and image structure.

    Source: §VI · p. 15

  • Takeaways & Limitations

    With imperfectly aligned references, voxel-wise metrics can reward smoothing or adaptation to evaluation geometry, supporting anatomy-oriented evaluation alongside intensity agreement.

    Source: §VI · p. 15

    IMPACT is not a ground-truth alignment, and both registration conventions may retain residual errors.

    Source: §VI · p. 15

Abstract

from arXiv · show

Supervised synthetic CT (sCT) generation is commonly trained and evaluated as voxel-wise regression against registered reference CT images. In practice, MRI-CT and CBCT-CT pairs are aligned through registration procedures that leave residual misalignments. These residuals are not independent intensity noise but spatially coherent geometric discrepancies that act as structured label noise. We investigate how this registration-induced bias affects supervised MRI-to-CT and CBCT-to-CT synthesis on 1,784 paired patients covering five anatomical regions. Voxel-wise scores strongly depend on the consistency between the registration used to build the training targets and the one used for evaluation: models score best when both conventions match, showing that networks partly learn the geometric convention of the registration pipeline and that standard metrics reward it. Training on more anatomically consistent registrations reduces prediction variability and improves out-of-distribution robustness, and CT-only controls show that registration alone produces metric errors in the range of top challenge submissions. To mitigate the limits of voxel-wise supervision, we introduce a perceptual loss computed in the feature space of a pretrained Segment Anything encoder. Compared with MAE-only and VGG-based objectives, it improves downstream segmentation and yields sharper, more structurally coherent sCT. Perceptual and voxel-wise metrics disagree under imperfect alignment and agree when the evaluation geometry is reliable. These results identify registration-induced bias as a central confounder in supervised sCT generation and argue for complementing voxel-wise agreement with anatomy-oriented evaluation criteria.

I. INTRODUCTION

Voxel-wise supervised sCT assumes accurate voxel alignment, but residual registration errors create structured geometric label noise that can shape both learned predictions and their evaluation. The study examines this bias and introduces SAM-based perceptual supervision as a complementary anatomy-oriented approach.

I. INTRODUCTION

  • Residual registration errors are spatially coherent anatomical displacements, making them structured geometric label noise rather than independent intensity noise.The study distinguishes variable and systematic effects on the learned target distribution.Source: §I · p. 1, Contributions · p. 2
  • The study tests whether training and evaluation registration choices affect quantitative scores, out-of-distribution behavior, and predictive uncertainty.Its stated objective is to determine whether synthesis networks learn a geometric convention that reference-based metrics subsequently reward.Source: §I · p. 2
  • A SAM-based perceptual loss compares synthesized and reference CT in a pretrained encoder’s feature space, targeting spatially coherent anatomy beyond voxel-wise intensity matching.It is intended to reduce regression-induced blurring but does not eliminate geometric bias in imperfectly registered targets.Source: §I · p. 2, A. Image-to-Image Translation: Problem… · p. 3
  • Voxel-wise scores improve when training and evaluation use the same registration convention, so metrics can reward pipeline agreement rather than source-anatomy fidelity.The study evaluates MRI→CT and CBCT→CT synthesis across five anatomical regions in 1,784 paired patients from six centers.Source: Contributions · p. 2

A. Image-to-Image Translation: Problem Formulation

  • Variable residuals broaden the observed target distribution and can blur regression predictions, while systematic residuals can shift predictions toward the registration pipeline’s geometric convention.This target uncertainty originates in data construction and is distinct from intrinsic modality ambiguity and limited model knowledge.Source: A. Image-to-Image Translation: Problem… · p. 3
  • The study uses SAM representations for their multi-scale anatomical structure and boundary features, with the perceptual distance intended to be less sensitive to small residual displacements.The feature representation matters; the section notes that natural-image classification features may not optimally encode medically relevant structures.Source: A. Image-to-Image Translation: Problem… · p. 3
  • Perceptual and generative objectives may improve sharpness, but training on imperfectly aligned pairs does not remove systematic registration bias.Such objectives still learn from the registered target distribution and may reproduce biases embedded in the registration pipeline.Source: A. Image-to-Image Translation: Problem… · p. 3
  • Under structured or systematic misalignment, regression can smooth uncertain boundaries or shift anatomy, so intensity-based benchmarks may favor registration distortions over anatomical faithfulness.Joint synthesis-registration and weakly paired approaches address correspondence in different ways, but residual misalignment remains difficult to eliminate completely.Source: A. Image-to-Image Translation: Problem… · p. 3

III. SUPERVISED CROSS-MODALITY IMAGE SYNTHESIS: EXPERIMENTAL STUDY ON SYNTHRAD

The section describes supervised cross-modality synthesis as prediction of CT from CBCT or MRI using paired voxel-level targets, with evaluation based primarily on intensity similarity to reference CT.

III. SUPERVISED CROSS-MODALITY IMAGE SYNTHESIS: EXPERIMENTAL STUDY ON SYNTHRAD

  • Supervised cross-modality synthesis trains models to predict CT from CBCT or MRI using voxel-wise reconstruction losses on spatially aligned image pairs.Source: §III · p. 3
  • The section studies the standard supervised formulation adopted in recent benchmarks, pairing spatially aligned inputs with voxel-level CT targets.Source: §III · pp. 3–4
  • Current evaluation frameworks primarily assess performance through intensity-based similarity metrics computed against a reference CT.Source: §III · p. 4

A. SynthRAD challenge and datasets

The study uses paired SynthRAD data and alternative registration conventions to examine supervised synthetic CT generation. It compares a voxel-wise baseline with added SAM feature-space supervision.

A. SynthRAD challenge and datasets

  • SynthRAD2023 and SynthRAD2025 provide paired MR-to-CT and CBCT-to-CT data spanning five anatomical regions, enabling comparison of registration conventions.The anatomy coverage varies by task and challenge edition.Source: A. SynthRAD challenge and datasets · p. 4

B. Experimental setup

  • ELX and IMPACT are treated as alternative registration conventions, not ground-truth alignments; IMPACT showed higher anatomical consistency in segmentation-based analyses.ELX uses Elastix mutual-information deformable registration, whereas IMPACT-Reg uses deep semantic feature correspondences.Source: B. Experimental setup · p. 4

1) Training and inference:

  • A 2.5D U-Net++ predicts central sCT slices from adjacent input slices, with retained model predictions averaged at inference.The architecture is a supervised baseline for studying registration bias, rather than a proposed new architecture.Source: 1) Training and inference · p. 4

C. SAM-based perceptual supervision

  • The SAM-based perceptual loss compares synthesized and reference CT in a pretrained encoder’s feature space, complementing voxel-level supervision.The frozen SAM 2 encoder supplies multi-scale features; the selected intermediate levels target anatomical structure while avoiding low-level noise sensitivity and overly coarse semantics.Source: C. SAM-based perceptual supervision · pp. 4–5
  • The final training objective combines the voxel-wise MAE and SAM-based perceptual losses with equal weighting.The formulation is intentionally simple to assess perceptual supervision without extensive hyperparameter tuning.Source: C. SAM-based perceptual supervision · p. 5

D. SAM-based perceptual evaluation metric

The SAM-based evaluation distance is calibrated separately from the training loss. Its learned channel weights are intended to emphasize synthesis-related structural defects while downweighting responses dominated by residual misalignment.

D. SAM-based perceptual evaluation metric

  • The calibrated SAM distance is an anatomy-oriented evaluation metric computed after training, not a synthesis objective or checkpoint-selection criterion.Separately learned channel-wise weights calibrate feature contributions, following the general spirit of LPIPS.Source: D. SAM-based perceptual evaluation… · p. 5
  • Channel weights are learned to distinguish sharp but misaligned deformed CT controls from MAE-trained predictions with blurring or structural inaccuracies.Calibration uses precomputed feature differences and a hinge-ranking objective; the weights are frozen for evaluation.Source: D. SAM-based perceptual evaluation… · pp. 5–6
  • The calibrated metric downweights responses dominated by residual misalignment and emphasizes channels sensitive to blurring, lost fine boundaries, and structural inconsistency.Intermediate SAM layers, particularly layers 2 and 3, were identified as the most discriminative representations.Source: D. SAM-based perceptual evaluation… · p. 6

E. Experimental configurations

The experiments vary training objectives and registration conventions, then test the models across in-distribution, out-of-distribution, expert-aligned, and simulated registration-free settings.

E. Experimental configurations

  • The study compares MAE-only and SAM-augmented objectives with ELX- and IMPACT-constructed training pairs, using held-out testing and five-fold cross-validation for model development.Fifteen percent of cases are reserved for final testing, and five models are ensembled at inference.Source: E. Experimental configurations · p. 6
  • MRI-to-CT evaluation includes 66 held-out SynthRAD2025 cases with ELX and IMPACT references, plus 50 SynthRAD2023 OOD cases evaluated under rigid ELX and non-rigid IMPACT alignment.The OOD set covers brain and pelvis cases.Source: E. Experimental configurations · p. 6
  • A separate external OOD test uses 24 T2-weighted MRI cases with expert-refined voxel-wise alignment and no T2-weighted MRI training data.Its single expert-refined reference reduces evaluation variability from different registration methods.Source: E. Experimental configurations · p. 6
  • CBCT-to-CT evaluation uses 103 real held-out cases with ELX and IMPACT references, alongside simulated CBCT generated directly from the same CT volumes to remove registration uncertainty.The simulation does not fully reproduce real acquisitions: Hounsfield units are not fully calibrated and some CT volumes are truncated, affecting field of view.Source: E. Experimental configurations · p. 6

F. Statistical analysis

Registration conventions affect measured sCT performance: matching training and evaluation yields the best voxel-wise scores, while IMPACT-based registrations show greater average anatomical consistency than ELX. Statistical comparisons use matched patients, with patients—not slices or folds—as the unit.

F. Statistical analysis

  • Comparisons use patient-level means and standard deviations, matched-patient two-sided Wilcoxon signed-rank tests, and do not treat folds as independent observations.For the CV ensemble, predictions are averaged voxel-wise before metrics are computed; Mean CV instead averages fold-specific metrics per patient.Source: F. Statistical analysis · pp. 6–7

G. Qualitative comparison of IMPACT and ELX registrations

  • IMPACT registration raises mean Dice from 0.724 with ELX to 0.758, although the direction of the difference varies by region.Both methods provide visually plausible, globally coherent deformable alignments; the differences are refinements over an already credible ELX baseline.Source: G. Qualitative comparison of IMPACT… · p. 7, Table I · p. 7

A. Effect of Registration Consistency Between Training and Evaluation

  • Task 1 MAE rises from 63.55 to 72.47 HU and Task 2 MAE from 58.76 to 69.17 HU when IMPACT-trained models are evaluated against ELX rather than IMPACT references.Matched training and evaluation conventions produce the best voxel-wise performance; evaluation mismatch harms IMPACT-trained models more than ELX-trained models.Source: Table II · p. 7, A. Effect of Registration Consistency… · p. 7
  • IMPACT-trained models outperform ELX-trained models across OOD settings, with MAE improvements of 9.0%–15.0% across Task 2 simulated CBCT regions.The comparisons against IMPACT reach p < 0.001, and the pattern is not limited to evaluation against IMPACT references.Source: Table III · p. 8, A. Effect of Registration Consistency… · p. 8
  • ELX-trained models have higher prediction uncertainty than IMPACT-trained models, with aggregate increases of 29.7% in Task 1 and 12.5% in Task 2 in-distribution.Uncertainty is variance across 15 predictions per patient; the increase reaches 116.1% in the Task 1 OOD pelvis, while the brain increase is 2.7%.Source: Table IV · p. 8, A. Effect of Registration Consistency… · p. 8

B. SAM-based perceptual supervision

SAM-based perceptual supervision improves anatomical preservation in several evaluations, while reference-based challenge scores favor the ELX convention used to construct the evaluation references. The results distinguish source-anatomy preservation from agreement with a particular registered reference.

B. SAM-based perceptual supervision

  • SAM-based supervision improves aggregate Task 1 Dice to 0.738 from 0.711 with MAE and 0.712 with VGG; SAM features also better preserve sharp boundaries and fine anatomical structures than voxel-wise objectives.SAM obtains the best Dice in all three simulated CBCT regions, while SSIM remains similar to or occasionally lower than with MAE supervision.Source: B. SAM-based perceptual supervision · pp. 8–9
  • SAM improves dSAM and LPIPS but increases MAE in-distribution, whereas it improves MAE, dSAM, and LPIPS in Ext-T2 and simulated CBCT settings.The in-distribution trade-off appears for both individual models and ensembles; ensemble averaging primarily reduces MAE, with smaller and inconsistent effects on perceptual metrics.Source: B. SAM-based perceptual supervision · p. 9, Table VI · p. 11

C. Anatomical consistency

  • IMPACT-trained models require smaller deformations than ELX-trained models across evaluated regions, indicating greater anatomical consistency with the source CBCT.The ELX–IMPACT difference exceeds 120% in abdominal and head-and-neck regions; it is smaller but statistically significant in thoracic cases.Source: C. Anatomical consistency · p. 9, Table VII · p. 11

D. Official SynthRAD results as a benchmark-bias case study

  • The submitted method ranked third in both tasks, while the official protocol favored ELX training, with MAE decreasing from 75.82 to 68.20 HU in Task 1 and 56.05 to 52.87 HU in Task 2.The ELX advantage reflects compatibility with the challenge evaluation geometry; the submission used an ELX ensemble with SAM supervision.Source: D. Official SynthRAD results as a… · pp. 10–11, Table VIII · p. 11

E. Estimation of registration-induced sensitivity in sCT evaluation metrics

CT-only controls estimate how much registration differences can change evaluation metrics without introducing synthesis error. Their metric errors are comparable in magnitude to those of high-performing supervised methods, and the estimates are sensitive to the anatomy and registration setup.

E. Estimation of registration-induced sensitivity in sCT evaluation metrics

  • The CT-only controls compare registration-modified CT geometries under each benchmark’s available evaluation setup, isolating metric sensitivity to registration without a synthesis model.The study treats disagreement between ELX and IMPACT as a proxy for geometric uncertainty, not as evidence that either method is ground truth.Source: E. Estimation of registration-induced… · p. 11
  • Registration-only controls produce aggregated MAEs of 60.02 HU for Task 1 and 53.42 HU for Task 2 in SynthRAD2025 AB/HN/TH, comparable in magnitude to high-performing supervised results.The controls contain no synthesis error; SynthRAD2023 brain/pelvis control MAEs are 59.18 HU and 39.50 HU for Tasks 1 and 2, respectively.Source: E. Estimation of registration-induced… · p. 12
  • For small residual displacement, metric effects depend on local image gradients; air, bone, teeth, and thin soft-tissue boundaries are especially sensitive.This helps explain larger metric variations in regions such as head-and-neck and pelvis.Source: E. Estimation of registration-induced… · p. 12
  • Some supervised models obtain better MAE, PSNR, or SSIM than CT-derived controls, yet the controls generally preserve substantially higher Dice scores.Local smoothing or attenuation of uncertain boundaries can reduce voxel-wise penalties without establishing better anatomical fidelity.Source: E. Estimation of registration-induced… · p. 12
  • The controls should be interpreted conservatively because they capture disagreement between two plausible regularized registrations and likely underestimate correspondence uncertainty.The SynthRAD2023 deformations were constrained, particularly for brain cases, and small deformation-field changes rapidly increased MAE.Source: E. Estimation of registration-induced… · p. 12

V. DISCUSSION

Registration conventions shape both the targets learned by supervised sCT models and the references used to score them. The discussion connects this bias to source-anatomy preservation and argues that benchmark scores need anatomy-oriented context.

A. Registration bias as structured label noise

  • Changing only the evaluation reference raises IMPACT-trained models’ aggregate MAE from 63.55 to 72.47 HU in Task 1 and from 58.76 to 69.17 HU in Task 2.Thus voxel-wise scores reflect both synthesis fidelity and compatibility with the selected registration geometry.Source: A. Registration bias as structured… · p. 12
  • Under inconsistent voxel-wise supervision, uncertain target locations can lead ℓ1 optimization toward smoother intermediate boundaries, attenuated high-frequency detail, or locally altered anatomy.The discussion attributes this to spatially variable targets broadening the conditional distribution and the objective minimizing expected reconstruction error.Source: A. Registration bias as structured… · p. 12
  • Higher intensity agreement with a registered reference does not necessarily mean better structural preservation: CT-derived controls generally retain higher Dice despite some supervised models’ better MAE, PSNR, or SSIM.The discussion links this distinction to possible smoothing, boundary attenuation, or adaptation to evaluation geometry.Source: A. Registration bias as structured… · pp. 12–13, Table IX · p. 13
  • Benchmark rankings are protocol-dependent, so the discussion recommends documenting registration conventions and complementing voxel-wise scores with anatomy-oriented and task-specific criteria.The local source-preservation analyses and the ELX-consistent SynthRAD server favor different training conventions.Source: A. Registration bias as structured… · p. 13, B. Anatomical consistency and source… · p. 13

B. Anatomical consistency and source preservation

  • ELX-trained models require larger deformations than IMPACT-trained models to align synthesized images with source CBCT anatomy, with differences exceeding 120% in abdominal and head-and-neck regions.The thoracic difference is smaller but statistically significant, and the magnitude varies by region.Source: B. Anatomical consistency and source… · p. 13

C. Consequences on uncertainty and OOD generalization

Registration inconsistency is associated with greater predictive variability and weaker OOD performance, while voxel-wise ERM can blur anatomy under imperfect alignment. SAM-based supervision improves structural preservation, with perceptual and voxel-wise metrics aligning more closely when correspondence is reliable.

C. Consequences on uncertainty and OOD generalization

  • ELX-trained models show 41.2% higher variability in Task 1 abdomen, 29.7% across in-distribution regions, 116.1% in OOD pelvis, and 61.8% on Ext-T2; Task 2 increases are 27.6% and 12.5%.The authors associate higher variability with less anatomically consistent supervision and describe the more consistent convention as associated with more stable predictions.Source: C. Consequences on uncertainty and OOD… · p. 14
  • On Ext-T2, MAE falls from 97.43 to 88.54 HU (p < 0.001) for IMPACT-trained versus ELX-trained models; IMPACT training also improves MAE across Sim-CBCT regions.Sim-CBCT abdomen, head-and-neck, and thorax MAE changes from 71.09 to 60.41 HU, 149.66 to 136.18 HU, and 109.15 to 95.99 HU, respectively.Source: C. Consequences on uncertainty and OOD… · p. 14

D. Origin of blurring under voxel-wise ERM

  • Residual geometric uncertainty shifts registered structures, so voxel-wise ERM favors intermediate configurations that reduce reconstruction error without necessarily preserving source anatomy.Perceptual objectives can improve sharpness and coherence but do not remove bias in imperfectly aligned targets; MAE may still favor smoothing or registration-convention matching.Source: D. Origin of blurring under voxel-wise… · p. 14

E. Why does SAM improve anatomical preservation?

  • SAM compares learned features that encode local structures in anatomical context; its compressed, boundary-aware representation may discourage locally averaged intensity patterns more effectively than VGG features.The proposed explanation is that the feature-space conditional median is taken over spatially coherent configurations rather than individual voxels.Source: E. Why does SAM improve anatomical… · p. 14
  • SAM-based perceptual supervision produces sharper, more structurally coherent sCT and improves downstream segmentation, while voxel-wise losses can lose thin structures and sharp interfaces.SAM-trained predictions are also closer to CT-derived controls in perceptual feature spaces, while MAE optimization can remain competitive in voxel-wise error.Source: E. Why does SAM improve anatomical… · pp. 14–15
  • In registration-free Sim-CBCT, SAM supervision improves ensemble MAE from 111.88 to 106.33 HU, dSAM from 20.67 to 17.99, and LPIPS from 6.15 to 5.15.With reliable correspondence, sharp coherent structures can also match the reference spatially, reducing the conflict between perceptual and voxel-wise criteria.Source: E. Why does SAM improve anatomical… · p. 15

F. Clinical relevance under task-specific constraints

The paper distinguishes dose calculation, where attenuation accuracy may make imperfect paired references acceptable, from anatomy-driven tasks that need stronger structural fidelity. Its conclusion links registration quality and perceptual supervision to anatomy preservation, and calls for evaluation beyond voxel-wise metrics alone.

F. Clinical relevance under task-specific constraints

  • Dose-based performance may remain acceptable with imperfect paired references if the generated attenuation map is accurate, while segmentation and deformable registration require stronger preservation of patient-specific anatomy.The paper reports strong dose-based performance for BreizhCT in the official SynthRAD results, but cautions against extending that conclusion to all downstream tasks.Source: F. Clinical relevance under… · p. 15

VI. CONCLUSION

  • Models score better when training and evaluation use the same registration strategy, while less consistent training targets are associated with lower OOD robustness and greater prediction variability.IMPACT-aligned supervision is associated with better source-anatomy preservation, lower variability, and improved OOD performance in the evaluated settings, but is not ground-truth alignment.Source: §VI · p. 15
  • Compared with MAE-only and VGG-based objectives, SAM-based supervision improves downstream anatomical metrics and produces sharper, more structurally coherent sCT.The authors suggest segmentation-learned feature representations may provide a more appropriate supervisory signal for anatomy-preserving synthesis than voxel-wise intensity losses.Source: §VI · p. 15
  • With imperfectly aligned references, voxel-wise metrics may reward smoothing or adaptation to evaluation geometry rather than patient-specific anatomical fidelity.The conclusion recommends anatomy-oriented criteria alongside reference-based intensity metrics.Source: §VI · p. 15

VII. DATA AND CODE AVAILABILITY

  • IMPACT registration files are publicly released, while original SynthRAD images are not redistributed and restricted challenge-center cases are excluded; reproduction code and configurations are available for both SynthRAD2025 tasks.The registration files are released under CC BY-NC 4.0 and can be applied to original SynthRAD images with Transformix.Source: §VII · pp. 15–16

APPENDIX A SYNTHRAD CHALLENGE AND BENCHMARK DETAILS

The appendix describes the SynthRAD challenge settings, datasets, evaluation protocol, official rankings, and implementation used for the study. It combines SynthRAD2025 primary analyses with complementary SynthRAD2023 evaluation and reports BreizhCT as third-ranked in both synthesis tasks.

A. Challenge design and objectives

  • SynthRAD uses containerized submissions, hidden validation and test references, and remote evaluation to standardize comparisons and separate training from testing.The evaluation pipeline is described transparently, which the paper notes can also facilitate metric-oriented optimization.Source: A. Challenge design and objectives · p. 17

B. Dataset characteristics

  • SynthRAD2025 contains 2,362 cases from five European centers across MRI-to-CT and CBCT-to-CT tasks; complementary SynthRAD2023 evaluation adds 1,080 paired acquisitions from three Dutch centers covering brain and pelvis.The primary analyses use SynthRAD2025, while SynthRAD2023 provides complementary OOD evaluation.Source: B. Dataset characteristics · p. 17

C. Evaluation protocol

  • Official evaluation combines MAE, PSNR, and MS-SSIM, segmentation Dice and Hausdorff distance, and dosimetric measures; SynthRAD2025 adds deformable registration at evaluation, unlike SynthRAD2023 rigid alignment.The appendix treats the two benchmark editions separately because their evaluation registration differs.Source: C. Evaluation protocol · pp. 17–18

D. Official challenge rankings

  • BreizhCT ranked third in both image-based and dosimetric MR-to-CT evaluations and third overall for CBCT-to-CT, with competitive image-similarity, anatomical, and dose measures.The reported rankings support that the registration analyses are not based on a deliberately weak synthesis baseline.Source: D. Official challenge rankings · p. 18

APPENDIX B IMPLEMENTATION DETAILS

  • The implementation uses a five-slice 2.5D U-Net++ with a ResNet-34 encoder; preprocessing, optimization, checkpoint selection, and inference procedures were kept consistent across configurations.Models were trained on fixed-size patches, selected using validation MAE, then ensembled across folds and test-time augmentations.Source: Appendix B · p. 18

APPENDIX C COMPLETE REGION-WISE RESULTS

The appendix reports region-wise support for registration-dependent performance and SAM-based supervision. It also compares perceptual and voxel-wise metrics across training objectives and registration settings.

A. Registration consistency

  • Across regions, performance is generally highest when the training and evaluation registration conventions match, though the effect varies with anatomy and local image gradients.Tables XIV and XV provide the detailed Task 1 and Task 2 combinations of training and evaluation registration.Source: A. Registration consistency · p. 18, Table XIV · p. 21, Table XV · p. 21

B. SAM-based supervision

  • SAM supervision generally improves downstream Dice, particularly in OOD regions, although SSIM does not always improve; displayed IMPACT registrations also show sharper, more spatially consistent boundary correspondences than ELX in several cases.The figure shows MR, CT, overlay, and checkerboard views across abdomen, head-and-neck, and thorax regions.Source: B. SAM-based supervision · pp. 18, 20, Figure 4 · p. 19

C. Perceptual and ensemble results

Loading 2609.29387v1…