acceptodds
Under review as a conference paper at ICLR 2027

Learning Structured Deviation from Normal Anatomy for Pancreatic Tumor Segmentation

Abstract

Pancreatic tumor segmentation on CT remains challenging because CT-only predictions may fail to reliably distinguish the pancreatic tumor from normal tissue. We investigate whether complementary evidence can be obtained from how a patient's anatomy deviates from plausible normal states. To this end, we introduce a Deviation-Aware Residual Adaptation (DARA) model, which consists of Star-chain Anatomy Graph (SAG) and Bounded Residual Reconciliation (BRR). SAG learns a multi-prototype normal anatomy universe from normal reference data and matches each patient's anatomy to class-compatible normal prototypes. SAG measures structural anatomical deviation between the patient and the normal reference and projects structural anatomical deviation into image space as a deviation field that complements the CT-only prediction. BRR trains a refiner using CT-only and SAG-aware out-of-fold (OOF) predictions and applies its bounded residual correction to the CT-only tumor logit. Effectiveness of the proposed DARA model has been validated by using three widely-used datasets (PanTSMini, MSD07, and PANORAMA) with five metrics (DSC, IoU, NSD, HD95, Recall). Experimental results show that the proposed model outperforms all competing methods in DSC, IoU, and NSD. In particular, 3.96%, 3.86%, 0.66% DSC gains can be achieved compared with the second-best method in each dataset respectively. These results suggest that learning what normal anatomy should look like provides a complementary cue to recognizing what is abnormal.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Pancreatic Tumor Segmentation, Normal Anatomy, Structural Anatomical Deviation, Normal Prototypes, Deviation Field

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YHMtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose DARA, a pancreatic tumor segmentation model that uses deviation from normal anatomy as extra evidence, with two components: SAG, which learns a normal anatomy model from tumor-free CTs and turns a patient's deviation from it into a voxel-level map, and BRR, which fuses the CT-only and SAG-aware predictions through a bounded logit correction. Experiments on PanTSMini, MSD07, and PANORAMA demonstrate DSC gains of 3.96%, 3.86%, and 0.66% over the second-best of ten segmentation baselines.

Critical strengths
  1. The motivation is clear and makes clinical sense. Comparing a patient's anatomy with a learned normal reference is a natural way to add evidence when the tumor itself is hard to see.
  2. The evaluation is extensive. It covers 2 datasets, including a large PanTS subset, and 10 recent segmentation baselines.
  3. The bounded residual design helps keep the final output anchored to the CT-only prediction and limits how much the deviation branch can change it.
Critical weaknesses
  1. Need more ablation. Most of the gain comes from fusing two models. A simple max of the two branches (36.54) already beats every baseline. Please run BRR on two CT-only nnU-Nets.
  2. The comparison is unfair. DARA uses extra reference CTs, organ labels, and 13 trained nnU-Nets, while each baseline is a single network trained on tumor labels only.
  3. Anatomy-aware baselines are missing, such as "Position-Aware Representation Learning with Anatomical Priors for Enhanced Pancreas Tumor Segmentation" (Dong et al., Neurocomputing 2025), which the authors cite but do not compare against.
Questions for authors
  1. BRR only corrects the tumor logit, so DARA's pancreas output should come from one of the two branches. Yet its pancreas results in Table 10 differ from nnU-Net's (e.g., 84.58 vs. 84.32 DSC on PanTSMini, 93.97 vs. 93.22 on PANORAMA). Which model produces DARA's pancreas mask? Is the CT-only branch the same model as the nnU-Net baseline in Table 1?
  2. How were the deviation weights (w_n, w_r, w_s, w_a) = (0.25, 0.75, 0.25, 1.0) chosen?
2 comments

We sincerely thank the reviewer for the careful and constructive comments. The suggestions helped us better separate the contribution of structural deviation from generic model fusion and clarify the resource, anatomy-aware comparison, and implementation issues. All newly added or revised content in the manuscript is highlighted in blue for ease of review.

1. Contribution beyond generic two-model fusion

The strong maximum-fusion result (36.54 DSC) makes generic model diversity an important alternative explanation. Following the reviewer’s suggestion, we therefore evaluate two independently trained CT-only nnU-Nets with the same architecture, training data, supervision, and schedule, differing only in random seed. Their OOF predictions are used to train the same BRR refiner.

SettingDSCNSDRecall
CT-only30.6027.0832.30
CT$_1$ + CT$_2$ + BRR31.5428.2439.11
CT$_1$ + CT$_2$, maximum fusion30.6427.2432.41
CT + SAG, maximum fusion36.5432.1143.44
AG+BRR33.6030.1639.75
EmpiricalRef+BRR34.6230.8440.02
DARA38.2134.0248.01

The CT+CT control improves recall but remains substantially below DARA. Maximum fusion shows the same pattern: combining two CT-only models produces almost no DSC gain, whereas CT+SAG reaches 36.54 DSC. This indicates that generic model diversity alone does not reproduce the benefit obtained when one branch receives patient-to-normal structural-deviation information.

The matched AG+BRR control provides a complementary comparison. It retains the anatomy predictor, anatomical supervision, second segmentation branch, OOF procedure, and BRR, while removing only the structural-deviation input. DARA improves over AG+BRR by +4.61/+3.86/+8.26 pp in DSC/NSD/Recall. For each subject, we first average the metric across the three random seeds for each method and then perform a 10,000-sample paired bootstrap over subjects. The resulting 95% CIs are [+3.110,+6.179], [+2.259,+5.562], and [+6.290,+10.340], respectively.

We further include EmpiricalRef+BRR to separate access to the reference cohort from optimized structured-reference modeling. It uses the same reference subjects, frozen node encoder, K=4K=4 states, and the same SAG-aware/OOF/BRR pipeline, but replaces the optimized reference with K-means prototypes and empirical shape/relation averages. EmpiricalRef+BRR obtains 34.62 DSC, compared with 38.21 DSC for DARA.

Together, these controls distinguish generic two-model diversity, access to anatomical/reference resources, and the proposed optimized patient-to-normal structural-deviation modeling.

2. Resource comparability and anatomy-aware comparison

DARA uses additional anatomical information and additional training-time computation relative to a single segmentation network. Accordingly, Table 1 is described as an end-to-end system-level comparison, while the matched controls above are used for methodological attribution.

The 5-fold OOF models are used only to generate held-out predictions for BRR training and are discarded afterwards; they are not ensembled at inference. At test time, DARA uses one frozen anatomy predictor, one CT-only branch, one SAG-aware branch, and the lightweight BRR refiner.

The reference cohort is used only to construct the NAU, which is subsequently frozen, while organ annotations are used only to train the frozen anatomy predictor. These resources therefore provide additional anatomical knowledge rather than additional tumor-supervised segmentation cases.

We also add the anatomy-aware comparison suggested by the reviewer. We implement the position-aware anatomical-prior baseline under the same predefined split, nnU-Net backbone, available anatomical supervision, validation-based model selection, and three-seed evaluation protocol as DARA.

Anatomy-aware settingDSCNSDRecall
Position-aware anatomical-prior baseline33.8132.1642.11
AG+BRR33.6030.1639.75
EmpiricalRef+BRR34.6230.8440.02
DARA38.2134.0248.01

Under this matched evaluation, DARA achieves the strongest overall performance.

3. Pancreas-output construction

Thank you for pointing out this ambiguity. BRR applies the bounded residual correction to the CT-only tumor logit. The CT-only branch itself is the same trained nnU-Net used as the Table 1 baseline. The final multiclass output is subsequently constructed using the refined tumor prediction together with the pancreas prediction, so the reported final pancreas mask can differ slightly from the original CT-only pancreas output.

Accordingly, the pancreas DSC changes from 84.32 to 84.58 on PanTSMini, 85.78 to 85.79 on MSD07, and 93.22 to 93.97 on PANORAMA. These differences reflect the final multiclass output construction rather than a separately trained CT-only model or a direct residual correction to the pancreas logit.

The revised manuscript now states the BRR output and final multiclass construction more explicitly.

4. Deviation-weight selection and sensitivity

The deviation weights were selected from a small set of candidate configurations on the PanTSMini validation set and then fixed across all three datasets without further tuning. The selected configuration is

(wn,wr,ws,wa)=(0.25,0.75,0.25,1.0).(w_n,w_r,w_s,w_a)=(0.25,0.75,0.25,1.0).

We additionally evaluate several alternative configurations while keeping all other components unchanged:

(wn,wr,ws,wa)(w_n,w_r,w_s,w_a)DSCNSDRecall
(1,1,1,1)(1,1,1,1)37.4733.1546.58
(0.25,0.75,0.25,1.0)(0.25,0.75,0.25,1.0)38.2134.0248.01
(0.5,0.5,0.5,1.0)(0.5,0.5,0.5,1.0)36.2333.4846.80
(0.25,1.0,0.25,0.75)(0.25,1.0,0.25,0.75)37.9333.7646.26

The selected configuration performs best among the tested settings, while the alternatives remain competitive. The same weights are used unchanged on MSD07 and PANORAMA.

We sincerely thank the reviewer again for the constructive suggestions. The additional CT+CT control, matched anatomy-aware comparison, and implementation clarifications have substantially strengthened the evaluation. We hope these additions have adequately addressed the reviewer’s main concerns, and we would be deeply grateful if the reviewer could kindly reconsider the current rating.

Reviewer_68qrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DARA segments pancreatic tumors by comparing a patient’s predicted anatomy with learned normal references. SAG turns these differences into a spatial map, and BRR uses predictions with and without this map to adjust the final segmentation. The paper reports improvements in DSC, IoU, and NSD on PanTSMini, MSD07, and PANORAMA. The idea is interesting, and the reported results are promising. However, the benefit of the learned reference, the computational cost, and performance beyond the current task remain unclear. After these issues are addressed, the work may be a better fit for a medical imaging conference such as MICCAI, given its focus on a specific medical task.

Critical strengths
  1. The method provides a clear way to add anatomical context to CT-based tumor segmentation. The bounded correction offers a simple way to use this information to adjust the original prediction.

  2. This paper reports results across three datasets and three random seeds. It also includes tumor-size analysis, missed tumors, and false positives, giving a fuller picture of performance than average Dice alone.

Critical weaknesses
  1. Table 4 compares different anatomical structures, but does not show whether the learned normal prototypes are needed. Keeping the same masks, segmentation branches, and BRR in this comparison would show what the learned reference adds.

  2. DARA requires an anatomy predictor, two segmentation branches, and five-fold training to prepare BRR inputs. The paper does not fully report the segmentation and baseline training settings or the total runtime and memory costs. These details matter when judging the practical benefit, especially given the 0.66-point DSC gain over MedNeXt on PANORAMA.

  3. Limited task scope and no OOD evaluation. All experiments focus on pancreatic tumor segmentation, so the benefit for other organs or diseases remains unclear. Each dataset uses separately trained models and normal references, leaving performance on unseen datasets untested. Another segmentation task and testing without retraining on an unseen dataset would better support the method’s wider use.

Questions for authors

See weaknesses.

2 comments

We sincerely thank the reviewer for the careful and constructive comments. The feedback helped us better isolate the contribution of the learned normal reference, clarify the computational trade-off, and narrow the scope of our claims. All newly added or revised content in the manuscript is highlighted in blue for ease of review.

1. Contribution of the learned normal reference

To directly separate learned structured-reference modeling from simply having access to reference CTs, we add a matched EmpiricalRef+BRR control. It uses the same reference subjects, frozen node encoder, number of normal states (K=4K=4), SAG-aware segmentation, OOF procedure, and BRR pipeline as DARA. The only difference is reference construction: node prototypes are obtained by K-means, while shape and relation references are computed by empirical averaging rather than optimized with the NAU objective.

Reference settingDSCNSDRecall
No structural reference (AG+BRR)33.6030.1639.75
Empirical K=4K=4 reference + BRR34.6230.8440.02
Learned NAU + BRR (DARA)38.2134.0248.01

DARA improves over EmpiricalRef+BRR by +3.59 DSC, +3.18 NSD, and +7.99 Recall points. Because the node encoder is shared, this comparison specifically supports the benefit of the optimized prototype/relation/shape reference downstream of a common representation, rather than attributing the gain merely to access to reference CTs.

We also retain the matched AG+BRR control, which keeps the anatomy predictor, anatomical supervision, second segmentation branch, OOF procedure, and BRR unchanged while removing the structural-deviation input. DARA improves over AG+BRR by +4.61/+3.86/+8.26 pp in DSC/NSD/Recall. For each subject, we first average the metric across the three random seeds for each method and then perform a 10,000-sample paired bootstrap over subjects. The resulting 95% CIs are [+3.110,+6.179], [+2.259,+5.562], and [+6.290,+10.340], respectively.

Together, these controls distinguish the effect of access to anatomical/reference resources from the contribution of optimized patient-to-normal structural deviation.

2. Computational cost and practical trade-off

We additionally benchmark end-to-end inference on the same hardware using 200 cases per method on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition. For DARA, the measurement includes the anatomy predictor, CT-only branch, SAG-aware branch, deviation computation, and BRR.

MethodSeconds/casePeak GPU memory (GB)
nnU-Net13.806.05
SwinUNETR6.124.00
MedNeXt75.062.83
SegMamba5.693.17
UNETR++19.901.45
DiffUNet6.652.85
MASS5.575.21
DINO-Med3D23.7339.88
SegMamba-V25.124.03
VoCo19.504.98
DARA22.077.61

DARA therefore incurs additional computation relative to a single nnU-Net, but the measured test-time cost remains within the range of several compared 3D segmentation systems. Importantly, the 5-fold OOF models are used only to generate held-out predictions for BRR training and are discarded afterwards; they are not ensembled at inference.

The revised manuscript makes this training-versus-inference distinction explicit and frames DARA as trading additional computation for complementary structural-deviation evidence rather than as a computationally efficient method.

On PANORAMA, the DSC gain over the second-best method is 0.66 pp; accordingly, the revised manuscript states that the practical trade-off can be dataset-dependent rather than equally favorable in every setting.

3. Scope and broader generality

The present work focuses on pancreatic tumor segmentation. Our central claim is correspondingly specific: within this task, patient-to-normal structural deviation provides complementary evidence beyond CT appearance alone.

The three datasets provide cross-dataset replication under different pancreatic cohort and reference conditions. PanTSMini and PANORAMA use separate tumor-negative reference cohorts, whereas MSD07 uses normal-appearing anatomy after tumor-margin exclusion; the available surrounding anatomical context also differs across datasets. Despite these differences, the same normal-reference/deviation/reconciliation principle consistently improves the CT-only formulation.

We therefore interpret the current evidence as cross-dataset replication within pancreatic tumor segmentation, rather than zero-shot domain transfer or demonstrated cross-organ generalization. Extending the same principle to fully frozen cross-dataset transfer and other organs or diseases is an important direction for future work.

We sincerely thank the reviewer again for the thoughtful feedback. The new matched reference control, end-to-end efficiency benchmark, and revised scope have substantially strengthened the attribution and presentation of our method. We hope these additions adequately address the reviewer’s main concerns, and we would be deeply grateful if the reviewer could kindly reconsider and raise the current rating.

Reviewer_eChzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Deviation-Aware Residual Adaptation (DARA) for pancreatic tumor segmentation in CT. Its central idea is to use deviations from plausible normal anatomy as complementary evidence to image appearance. The Star-chain Anatomy Graph (SAG) represents patient anatomy, compares it with a learned multi-prototype normal anatomy universe, and projects node, relation, shape, and appearance deviations into a voxel-level field. Bounded Residual Reconciliation (BRR) then combines CT-only and SAG-aware predictions through an out-of-fold-trained refiner that applies a bounded correction to the CT-only tumor logit. Experiments on PanTSMini, MSD07, and PANORAMA report the highest mean DSC, IoU, and NSD among the compared methods.

Critical strengths
  1. Connecting normal-reference modeling with structured anatomical representations is an appealing approach to pancreatic tumor segmentation. Patient-to-normal anatomical deviation provides a plausible complementary signal when tumor appearance is heterogeneous, low contrast, or poorly delineated.
  2. SAG constructs anatomical deviation evidence, while BRR explicitly accounts for the fact that this evidence need not be tumor-specific. Restricting its influence through a bounded residual correction is a reasonable design choice.
  3. Evaluation across multiple datasets with relevant ablations. The paper evaluates three datasets and includes component, anatomical-structure, and reconciliation ablations.
Critical weaknesses
  1. The specific contribution of learned normal-reference modeling remains insufficiently isolated. Tables 2 and 6 provide evidence that structural deviation benefits the proposed pipeline. In particular, removing structural deviation reduces DSC from 38.21 to 33.60 on PanTSMini. Table 4 also includes a pancreas anatomical prior baseline, which achieves 36.04 DSC. These are useful controls, but their implementation is not described sufficiently to establish exactly which information and computational components are held fixed. More importantly, the experiments do not directly establish whether learning the proposed normal anatomy universe is necessary, rather than using a simpler anatomical reference.

  2. Training-resource comparability and differences between dataset-specific configurations require clearer analysis. DARA uses an anatomy predictor trained with organ annotations, a separate reference cohort, two segmentation branches, and five-fold OOF prediction generation for refiner training. The main comparison does not sufficiently establish whether competing methods receive comparable auxiliary supervision or training resources. Furthermore, the three datasets instantiate materially different settings. PanTSMini and PANORAMA use tumor-negative reference cohorts, whereas MSD07 uses normal-appearing regions obtained by excluding annotated tumors and a 5-mm margin. Table 7 also shows that MSD07 uses only pancreatic chain nodes, without global organ nodes.

  3. Robustness to errors in anatomical grounding is not sufficiently evaluated. SAG depends on predicted organ masks, and the pancreatic chain is constructed by ordering predicted pancreas voxels along the longest physical axis and partitioning them into 12 regions. Although missing structures are excluded and confidence weighting is used, the manuscript does not quantify how upstream anatomical errors propagate to tumor segmentation.

  4. The main evaluation should make failure cases and negative-case behavior easier to interpret. Appendix A.1 explicitly states that HD95 is averaged only over cases with finite values and appropriately cautions against interpreting it independently of the contributing case count. However, these counts appear in Table 11 rather than alongside HD95 in Table 1. The contributing subsets differ substantially across methods. For example, on PanTSMini, Table 11 reports 175 finite cases for DINO-Med3D, 223 for nnU-Net, and 242 for DARA, out of 250 tumor-positive subjects. The reported HD95 averages do not summarize the same case population.

  5. The representation analysis is underspecified, and the regional analysis supports a narrower conclusion than a mechanistic explanation. Table 5 reports layer-wise (R2R^2) values for Real-D and Zero-D, but the manuscript does not clearly specify the prediction target, fitted model or probe, fitting and evaluation splits, or uncertainty across runs. Figure 3 provides evidence that the deviation field is associated with tumor location, with higher mean deviation in tumor regions than in non-tumor pancreatic regions in 85% of the analyzed cases. However, this does not establish how the field responds to benign anatomical abnormalities, normal variation, or errors in predicted anatomy.

Questions for authors
  • How exactly are “pancreas anatomical prior” in Table 4 and “BRR without structural deviation” in Tables 2 and 6 implemented? Which branches, anatomy inputs, OOF procedures, and training budgets are retained?

  • Can a simpler reference model built from the same reference cohort achieve comparable gains? How sensitive is performance to reference-cohort composition and the MSD07 tumor-exclusion margin?

  • How does DARA perform when anatomical masks are inaccurate? Can the authors provide an oracle-mask comparison or a controlled perturbation analysis?

  • What is the regression target and evaluation protocol for the (R2R^2) analysis in Table 5? Are its conclusions stable across seeds and held-out subjects?

  • Can the authors report false-positive behavior separately for tumor-negative subjects and provide uncertainty estimates for the principal ablation differences?

2 comments

We sincerely thank the reviewer for the careful and constructive comments. These suggestions helped us clarify the contribution of learned normal-reference modeling, resource usage, and anatomical grounding. We have added matched controls and additional analyses accordingly. All newly added or revised content in the manuscript is highlighted in blue for ease of review. 1. Learned normal-reference modeling

For the original BRR w/o structural deviation setting, now renamed AG+BRR, the anatomy predictor, anatomical supervision, second segmentation branch, branch architecture/training budget, 5-fold OOF prediction generation, and BRR refiner are all retained; only the structural-deviation input is removed.

AG+BRR achieves 33.60/30.16/39.75 DSC/NSD/Recall, compared with 38.21/34.02/48.01 for DARA, corresponding to gains of +4.61/+3.86/+8.26 pp. For each subject, we first average the metric across the three random seeds for each method, and then perform a 10,000-sample paired bootstrap over subjects. The resulting 95% CIs are [+3.110,+6.179], [+2.259,+5.562], and [+6.290,+10.340], respectively. This matched control isolates the contribution of structural deviation within the same anatomical and reconciliation pipeline.

The revised manuscript also clarifies the pancreas anatomical prior in Table 4. This control uses the predicted pancreas as a single anatomical prior, without surrounding global-organ nodes or the 12-node pancreatic chain; the downstream segmentation and reconciliation pipeline is otherwise unchanged.

To directly compare the learned structured reference with a simpler alternative, we additionally evaluate EmpiricalRef+BRR. It uses the same reference subjects, frozen node encoder, number of normal states (K=4K=4), SAG-aware segmentation, OOF procedure, and BRR pipeline as DARA. The only change is reference construction: node prototypes are obtained by K-means over the shared reference embeddings, while shape and relation references are formed by empirical averaging instead of optimization with the NAU objective.

Because the node encoder is shared, this control specifically evaluates the benefit of the optimized prototype/relation/shape reference downstream of a common representation, rather than isolating every learned component of NAU.

EmpiricalRef+BRR achieves 34.62/30.84/40.02 DSC/NSD/Recall, while DARA achieves 38.21/34.02/48.01, giving +3.59/+3.18/+7.99 pp.

ComparisonΔ\DeltaDSC (pp)Δ\DeltaNSD (pp)Δ\DeltaRecall (pp)
DARA −- EmpiricalRef+BRR+3.59+3.18+7.99

These results support the benefit of optimizing the structured normal reference relative to a matched K-means/empirical construction.

2. Training-resource comparability and dataset-specific settings

DARA uses additional anatomical information and additional training-time computation relative to a single tumor-segmentation network. Accordingly, Table 1 is described as an end-to-end system-level comparison, while the matched controls above are used for methodological attribution.

The reference cohort is not used as additional tumor-segmentation training data. It is used only to construct the NAU, which is frozen before downstream segmentation training and evaluation. Similarly, organ annotations are used only to train the anatomy predictor, which is also frozen before downstream tumor-segmentation training. Thus, the additional resources provide anatomical knowledge rather than additional tumor-supervised segmentation cases.

The dataset-specific configurations follow the anatomical information available in each dataset. PanTSMini and PANORAMA provide tumor-negative reference cohorts and surrounding-organ annotations, allowing both global organ nodes and pancreatic chain nodes. MSD07 lacks an independent healthy cohort and comparable surrounding-organ annotations; we therefore use normal-appearing regions outside the annotated tumor and construct only the 12 pancreatic chain nodes.

We further evaluate sensitivity to the MSD07 exclusion margin using the same downstream pipeline:

MSD07 exclusion marginDSCNSDRecall
0 mm58.6648.5460.27
5 mm (main setting)59.6551.8065.44
10 mm59.0249.7964.52

The 5-mm setting performs best, while the overall conclusion remains consistent across the three tested margins.

3. Robustness to anatomical-grounding errors

We additionally evaluate sensitivity to anatomical-grounding errors by perturbing each predicted organ mask with increasing geometric corruption while keeping the trained tumor-segmentation model fixed. Mild, moderate, and severe perturbations apply independent random translations of ±2\pm2, ±5\pm5, and ±10\pm10 mm per axis, respectively, together with random erosion or dilation of radius 1, 3, and 5 mm. We compare these settings with predicted and oracle anatomy:

Anatomy inputDSCNSDRecall
Oracle39.6035.4450.07
Predicted38.2134.0248.01
Mild perturbation37.5433.4046.85
Moderate perturbation36.1231.8144.43
Severe perturbation34.2729.6440.89

Performance decreases progressively with stronger anatomical corruption, while oracle anatomy improves DSC from 38.21 to 39.60. This indicates graceful degradation under the tested geometric perturbations, while the oracle result also suggests remaining headroom from improved anatomical grounding.

4. Failure cases, tumor-negative behavior, and uncertainty

To make HD95 directly comparable across methods, the main evaluation now uses HD95$_{200}$. For tumor-positive subjects with non-finite HD95 due to an empty prediction, we assign a fixed penalty of 200 mm and average over the full tumor-positive cohort. The corresponding mean and standard deviation are recomputed under this unified definition, so every method is evaluated over the same number of subjects.

We also report false-positive behavior separately on tumor-negative PanTSMini subjects:

MethodFP lesions / negative case ↓\downarrowNegative subjects with any predicted tumor ↓\downarrow
AG+BRR1.2846.84%
DARA1.2042.62%

Thus, relative to the matched AG+BRR control, DARA's higher tumor recall is not accompanied by an increase in either measured negative-case false-positive statistic.

For the principal matched ablation, we first average each metric across the three random seeds for every subject and then perform a 10,000-sample paired bootstrap over subjects. The resulting intervals are:

  • DSC: +4.611 pp, 95% CI [+3.110,+6.179]
  • NSD: +3.861 pp, 95% CI [+2.259,+5.562]
  • Recall: +8.263 pp, 95% CI [+6.290,+10.340]

These intervals provide subject-level statistical support for the improvement over the matched AG+BRR control.

5. Representation analysis and regional structural deviation

The revised manuscript makes the protocol for Table 5 explicit. The regression target is the original structural-deviation field DSAGD_{\mathrm{SAG}}. The segmentation network is frozen, and features are extracted from the early encoder, deep encoder, bottleneck, and final decoder. After spatial alignment, a ridge linear probe is fitted to predict DSAGD_{\mathrm{SAG}}, using disjoint subjects for probe fitting and evaluation.

LayerReal-D R2R^2Zero-D R2R^2ΔR2\Delta R^2
Early encoder0.8710.335+0.536
Deep encoder0.5650.528+0.037
Bottleneck0.1430.130+0.013
Final decoder0.7410.603+0.139

We use a more precise interpretation of this analysis: DSAGD_{\mathrm{SAG}} is more linearly recoverable from the corresponding representations when the deviation field is explicitly provided. In particular, the large early-encoder difference partly reflects the fact that DSAGD_{\mathrm{SAG}} is itself an input to the SAG-aware branch. Table 5 is therefore presented as a representation diagnostic rather than mechanistic evidence.

Similarly, Fig. 3 supports an association between structural deviation and tumor location: in 85% of analyzed cases, the tumor region shows larger deviation than the non-tumor pancreatic region within the same patient. Elevated deviation may also reflect benign anatomical variation, other abnormalities, or anatomical-grounding errors; this scope is stated explicitly in the revised manuscript.

We sincerely thank the reviewer for the detailed and constructive feedback. The added controls and analyses have substantially strengthened the experimental isolation and interpretation of the method. If the reviewer finds that these additions address the main concerns, we would be sincerely grateful if they could reconsider and raise the current rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.