acceptodds
Under review as a conference paper at ICLR 2027

Progressive Task-Conditioned Correction for Semantic Segmentation

Abstract

Hierarchical segmentation decoders repeatedly combine encoder and decoder representations, but their discrepancy provides candidate refinement content rather than a task update by itself. We formulate decoding as progressive task-conditioned correction and introduce DiCoSeg. In the encoder, Consistency-Guided Directional Aggregation (CGDA) adapts multi-directional state aggregation while retaining the pretrained equal-path computation as its initialization. In the decoder, each Corrective Fusion Block (CFB) estimates where a correction is supported and its signed task-space direction, transports supported corrections while preserving their support, and uses the resulting Corrective State to modulate discrepancy-derived features. A Persistent Task State carries this corrective context across decoder stages, turning hierarchical decoding into an evolving correction process rather than a sequence of isolated fusions. Across medical-image segmentation and natural-image scene parsing, DiCoSeg consistently improves segmentation quality, while controlled analyses support the intended separation between observable discrepancy and corrective action. DiCoSeg therefore separates two operations often coupled implicitly in hierarchical fusion: identifying representational discrepancy and deciding how that discrepancy should update the task representation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image segmentation, representation learning, feature refinement, feature fusion, dense prediction

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

This paper views the layer-wise encoder–decoder fusion in segmentation networks as a progressive correction process. The main idea is that the difference between encoder and decoder features should not be used directly to update the features; instead, the current prediction should decide where and how to correct. Based on this, the authors propose DiCoSeg on top of VMamba. It adds CGDA to the encoder, which adaptively aggregates the different scan directions, and CFB to the decoder, which keeps a task state across stages and uses it to modulate the discrepancy features. Extra supervision on the correction process is also used during training. Experiments on four medical segmentation datasets and ADE20K show improvements over the compared methods, and the paper also includes fairly complete ablations and analysis of intermediate results.

Critical strengths
  1. The performance gains are fairly consistent. The method outperforms existing approaches on several medical segmentation datasets and also shows a solid improvement on ADE20K, which suggests it is not tailored to a single dataset. The improvement in boundary-related metrics is particularly noticeable, which matters in practice for medical segmentation. The extra parameters and computation are also small.

  2. The paper is generally well written. The motivation and design of each component are explained clearly, and the appendix gives enough implementation detail that reproducing the method should not be difficult. Beyond the main results, the experiments include a fairly comprehensive set of ablations, along with visualizations and statistics of intermediate quantities that help show what each module is doing. I also appreciate that the authors openly report unfavorable results, such as the slower inference and the cases where performance drops.

Critical weaknesses
  1. The ablation study is not fully convincing. The ablation results appear to come from a single run: they do not match the main table, and no standard deviations are reported, while many of the differences between variants are quite small. In addition, the reported intermediate statistics show that the learned CGDA directional weights stay very close to equal weighting, and the task state changes only slightly across decoder stages. It is therefore not entirely clear to me that the gains actually come from the correction mechanism described in the paper. I would suggest reporting multi-seed results for the key variants. I would also suggest inference-time interventions on the trained model (e.g., setting α to 0 or fixing the task state) to verify that the model actually relies on these components.

  2. The related work section misses several classic methods with closely related ideas, such as LRR, PraNet, Attention U-Net, PointRend, and Iterative Error Feedback, as well as SPN and Normalized Convolution, which use similar propagation and normalization schemes. Without this context, it is hard to judge what is genuinely new in this work. I would suggest adding a discussion of how the proposed method differs from these approaches.

Questions for authors
  1. (Clarification) The DiCoSeg result in Table 4 differs from the one in Table 1. Are the ablations based on a single run? If so, which run does this row correspond to, and how was it chosen?

  2. (Suggestion) The multi-seed experiments do not need to cover every variant. Running five key variants three times each and reporting mean and standard deviation would be sufficient: VSS Baseline, CGDA Only, CFB Only, Discrepancy-Only, and the full DiCoSeg.

    The inference-time interventions require no retraining. Starting from the trained DiCoSeg, one could apply each of the following at inference time and compare DSC/HD95 against the unmodified model:

    • set α to 0 in all CGDA blocks, recovering equal-weight aggregation;
    • use the initial task state T_0 at every stage, without stage-wise updates;
    • replace the modulation M with its spatial mean, removing the spatial selectivity of "where to correct".

    If these changes lead to a clear drop in performance, that would show more directly that the model relies on these components, and I would be more confident in maintaining or raising my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Wdbb — Weaknesses W1–W2

We thank the reviewer for the encouraging assessment and specific suggestions, which helped us strengthen the ablations, related work discussion, and mechanism analyses.

To W1:

Thank you for the specific suggestions. We repeated the five requested key variants over three independent runs on both Synapse and ISIC2018. Full DiCoSeg achieves the best mean on all four metrics. We also performed all three proposed inference interventions on the same trained checkpoint. These interventions reduce DSC by 0.115–0.151 percentage points, while tracking errors at each stage shows that reusing the initial task state increases the error after the final correction from 0.997% to 1.966%. The ablation table from three runs is provided in Q2 below. Q3 reports the inference interventions and task state measurements at each stage, which offer complementary evidence about the learned mechanism.

To W2:

Thank you for identifying these important references. We have added all seven to Related Work and compared their specific mechanisms directly with DiCoSeg.

Prior workPrincipal designTechnical distinction in DiCoSeg
LRRGated Laplacian-pyramid multiscale reconstructionDiscrepancy-derived decoder updates conditioned on corrective state
Attention U-NetDecoder-guided skip-feature attentionExplicit correction support/direction preceding selective feature correction
PraNetReverse-attention-guided progressive refinementSupported signed correction transport with persistent task-state updates
PointRendUncertainty-guided point-wise refinementDense corrective fusion inside hierarchical decoder stages
Iterative Error FeedbackIterative prediction-error updates in pose estimationPersistent task-state updating tied to encoder–decoder feature correction
SPNLearned affinity-based spatial propagationTransport of the joint field [Sl⊙Ql,Sl][S_l\odot Q_l,S_l]
Normalized ConvolutionCertainty-weighted normalized aggregationJoint transport and support-normalized recovery of corrective direction

Together, these comparisons clarify the role of the complete corrective formulation. During training, task residuals supervise where and how correction is needed. At inference, the predicted corrective signals form a shared state that guides current feature refinement and informs correction at the next decoder stage. This connects corrective estimation with task information accumulated throughout decoding. Matched Fusion and Matched Deep Supervision evaluate the value of the design, while stage-wise analyses and the T0T_0 reuse intervention examine its corrective behavior. The corresponding results are reported in §4.4 and Appendix A.12–A.15.

AuthorsOpenReview ↗

Response to Questions Q1–Q2

To Q1:

Thank you for catching the statistical distinction. Table 4 uses the preset default seed 42 for all detailed variants, providing a controlled comparison with the same seed. This seed was not chosen by ranking test performance. Its DiCoSeg result is 83.96 DSC / 15.45 HD95. Table 1 reports mean ± SD across three runs, 83.77 ± 0.18 DSC / 14.73 ± 1.54 HD95. Both select checkpoints by minimum training loss. We have clarified Table 4's caption and added the multi-run table below.

To Q2:

We completed the suggested ablations for all five variants across three independent runs on both Synapse and ISIC2018.

VariantSynapse DSC ↑Synapse HD95 ↓ISIC2018 mIoU ↑ISIC2018 DSC ↑
VSS Baseline82.05 ± 0.1822.63 ± 1.8581.16 ± 0.1689.60 ± 0.10
CGDA Only83.16 ± 0.1519.72 ± 1.2281.84 ± 0.0890.01 ± 0.05
CFB Only83.07 ± 0.1120.17 ± 1.1581.91 ± 0.0990.06 ± 0.06
Discrepancy-Only83.62 ± 0.0819.51 ± 0.5682.04 ± 0.0990.13 ± 0.05
Full DiCoSeg83.77 ± 0.1814.73 ± 1.5482.35 ± 0.1990.32 ± 0.11

Full DiCoSeg has the best mean on all four metrics. Relative to Discrepancy-Only, Synapse HD95 improves by 4.78, with a 0.15 percentage point DSC gain. Results are added in Appendix A.10, Table 10.

We also performed all three requested inference interventions on the same Synapse checkpoint selected by minimum training loss, without retraining.

Inference settingDSC (%) ↑ΔDSC (pp)HD95 ↓ΔHD95
Unchanged DiCoSeg83.9640.00015.4460.000
Set CGDA α=0\alpha=083.849−0.11515.451+0.005
Reuse initial T0T_0 at each stage83.820−0.14415.778+0.332
Replace MlM_l with its spatial mean83.813−0.15115.749+0.303

Single fixed checkpoint. Spatial averaging is applied to the post-tanh modulation.

Setting α=0\alpha=0 in all CGDA blocks produces a modest immediate change of −0.115 DSC points and +0.005 HD95. This measures the sensitivity of the converged model to restoring equal-weight directional aggregation while leaving the other learned parameters fixed. It differs from training a model with or without CGDA, where all parameters are learned under the corresponding aggregation rule.

The training ablations show CGDA's contribution from a complementary perspective. For seed 42, adding CGDA improves DSC from 82.03 to 83.13 without CFB, and from 82.93 to 83.96 with CFB. The comparisons across three independent runs also favor CGDA: VSS Baseline 82.05 ± 0.18 versus CGDA Only 83.16 ± 0.15, and CFB Only 83.07 ± 0.11 versus Full DiCoSeg 83.77 ± 0.18. CGDA starts from a pretrained solution with equal directional weighting and learns a small signed reweighting, so the modest effect of zeroing α\alpha at inference is consistent with its design. Together, the inference perturbation and retrained ablations offer complementary perspectives on CGDA’s role.

The other two interventions also lower DSC. Reusing T0T_0 and replacing spatial modulation with its mean increase HD95 by 0.332 and 0.303, respectively. To examine why retaining the task state matters, we compared intermediate predictions of the unchanged model and the T0T_0-reuse intervention at each decoder stage, using the same trained checkpoint.

StageSettingError before (%) ↓Error after (%) ↓Relative error reduction (%) ↑
1Full DiCoSeg2.2451.57529.84
1Reuse T0T_02.2451.57529.84
2Full DiCoSeg1.5291.00834.08
2Reuse T0T_02.4031.74727.29
3Full DiCoSeg1.5550.99735.86
3Reuse T0T_02.5691.96623.47

Relative error reduction is calculated within each stage as (before − after) / before. These are internal task-state predictions, not the final segmentation output.

Stage 1 is unchanged, as both settings begin with the same state. At later stages, reusing T0T_0 discards accumulated context and increases the incoming error. Each stage still performs corrective updates, but it cannot recover the accuracy of the persistent-state model. By Stage 3, the error after correction is 1.966% versus 0.997%, while the relative error reduction is 23.47% versus 35.86%. The reset variant can have larger absolute reductions because it starts with more errors. Persistence allows subsequent correction to begin from a more accurate task state and end at a substantially lower error rate.

The task state measurements from this checkpoint explain the role of persistence, while the inference interventions and retrained ablations test complementary aspects of the learned mechanism. Together, these results directly support the progressive task-conditioned correction described in the paper. The task state reuse comparison is included in Appendix A.14.

Thank you for the detailed response and the additional experiments. The multi-seed ablations and the inference-time interventions address my main concerns, and the expanded related work discussion makes the contribution clearer. I will keep my score.

Thank you for your thoughtful feedback and for considering our responses. We are glad that the additional experiments and clarifications have addressed your main concerns. We sincerely appreciate your time and constructive suggestions.

Reviewer_V5hgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DiCoSeg, a semantic segmentation framework that treats hierarchical decoding as a progressive task-conditioned correction process. The central idea is to separate encoder–decoder feature discrepancy from the decision of how this discrepancy should modify the current segmentation representation. The proposed framework consists of two main components: Consistency-Guided Directional Aggregation (CGDA), which adjusts the aggregation of directional state-space representations, and a Corrective Fusion Block (CFB), which estimates correction support and direction, propagates supported corrections, and maintains a persistent task state across decoder stages. The reported results are competitive, particularly in terms of HD95 on Synapse and mIoU on ADE20K.

Critical strengths
  • The formulation is intuitive and makes the role of the task state explicit.
  • The ablation study is relatively comprehensive.
Critical weaknesses
  • The paper's main contribution is the separation of representation discrepancy from task-conditioned corrective action. However, several components of CFB resemble existing approaches based on prediction-guided feature modulation, gated fusion, and iterative refinement.

  • Since the method explicitly constructs correction targets from prediction errors, it would be natural to evaluate how accurately the learned support predicts erroneous regions. Similarly, measuring the changes in prediction errors before and after each corrective stage would provide more direct evidence for the claimed progressive correction behavior.

  • Appendix A.1 states that volume evaluation is performed every 100 epochs on the project-provided test_vol split and that the checkpoint with the highest volume DSC is retained for reporting. This suggests that the test set is used for model selection. If so, the reported Synapse performance is not based on a strictly held-out test evaluation.

  • I am not fully convinced that the experiments establish the necessity of the specific correction formulation over simpler task-guided fusion mechanisms. More importantly, the reported Synapse checkpoint-selection protocol raises a concern about the validity of a substantial portion of the experimental evidence. The considerable inference overhead also limits the practical appeal of the method.

  • According to Table 8, inference latency increases from 8.77 ms to 37.78 ms per image, corresponding to approximately a 4.3-fold slowdown, whereas FLOPs increase by only 2.67%. This is a significant practical cost for a segmentation model.

Questions for authors
  • Can the authors clarify whether the Synapse test volumes are used for checkpoint selection? If so, can the main results and ablations be evaluated using independently selected checkpoints?

  • How does DiCoSeg compare with a simpler decoder that uses the current coarse segmentation probabilities to gate encoder–decoder discrepancy features, without explicit correction transport or a persistent task state?

3 comments
AuthorsOpenReview ↗

Response to Reviewer V5hg — Weaknesses W1–W3

We thank the reviewer for the careful reading and insightful concerns about the corrective mechanism, its relationship to prior work, and the evaluation protocol.

To W1:

Thank you for highlighting the connections to prediction-guided modulation and iterative refinement. Our starting point is to formulate encoder–decoder interaction as progressive task-conditioned correction. Feature discrepancy provides candidate information for refinement, but does not itself specify the correction required by the current segmentation. Rather than using the current prediction primarily to guide feature fusion, DiCoSeg explicitly estimates what should be corrected and in which direction.

CFB predicts correction support SlS_l and signed direction QlQ_l from the incoming task state and encoder–decoder features. During training, task residuals supervise these predictions. At inference, SPCT forms a corrective state ClC_l from the predicted signals. The same ClC_l governs current feature refinement and updates the task state used to condition the next stage's correction. The distinction is thus the explicitly supervised corrective state and its coupled role across stages, rather than iterative refinement or gating alone. The matched fusion control in W4 and stage-wise measurements in W2 examine the value of this formulation.

To W2:

Thank you for requesting a direct measurement of the proposed correction process. Using one fixed, training-loss-selected Synapse checkpoint, we ranked Support scores against pre-correction task-state errors and compared predictions before and after each corrective update at the same stage resolution.

StageSupport AP ↑Precision@1% ↑Recall@5% ↑Error before → after ↓
10.778286.24%96.10%2.245% → 1.575%
20.664369.46%97.07%1.529% → 1.008%
30.738777.81%97.96%1.555% → 0.997%

These measurements use a single fixed checkpoint and describe intermediate task states, not the final segmentation output.

The top 5% of positions ranked by Support capture 96–98% of the errors before correction. Across the three stages, the percentages of corrected pixels (0.712%, 0.591%, 0.614%) exceed those of newly harmed pixels (0.042%, 0.070%, 0.057%). Each corrective update therefore reduces the error rate at that stage. Cross entropy also decreases and pooled foreground Dice increases. Detailed measurements appear in Appendix A.13, Tables 12–13.

To W3:

We thank the reviewer for identifying the incorrect description in Appendix A.1 and sincerely apologize for the confusion it caused. Contrary to what the original wording suggested, the reported Synapse checkpoints were selected by minimum epoch-averaged training loss, following the checkpoint-saving rule of the public VM-UNet Synapse training implementation, not by test DSC. Periodic volume-level evaluations every 100 epochs were for monitoring and were not used for checkpoint selection or hyperparameter tuning. We corrected Appendix A.1 to reflect the actual procedure. The reported results are unchanged.

We additionally conducted an independent validation experiment with 15 train, 3 validation, and 12 test cases, repeated over three runs. During training, only the validation cases were evaluated, and checkpoints were selected by validation DSC. The test cases were evaluated after that selection.

ModelDSC (%) ↑HD95 ↓
VSS Baseline80.19 ± 0.2023.85 ± 1.45
DiCoSeg81.63 ± 0.1917.59 ± 1.28

This protocol trains on 15 rather than 18 cases and selects checkpoints by validation DSC, so its absolute scores should be interpreted within that setting rather than directly compared with the standard 18/12 results. Under this shared validation-based protocol, DiCoSeg achieves higher mean DSC and lower mean HD95 than VSS Baseline. The experiment is now documented in Appendix A.1.

AuthorsOpenReview ↗

Response to Weaknesses W4–W5

To W4:

We implemented the suggested gate based on current coarse probabilities over encoder–decoder discrepancy as a direct alternative. The gate is predicted from the pre-fusion decoder feature and modulates discrepancy-derived features before residual fusion. The Light and Matched variants retain the CGDA encoder and VSS decoder while replacing the entire CFB, including corrective transport and persistent task state. Their shared fusion rule is:

G_l=\sigma(\operatorname{Conv}_{1\times1}(P_l)),$$ $$X_l=\operatorname{LN}_E(E_l)-\operatorname{LN}_D(D_l),\quad F_l=D_l+E_l+\gamma_l[G_l\odot\Phi_l(X_l)].$$ | Model | Params (M) | FLOPs (G) | Synapse DSC (%) ↑ | HD95 ↓ | |---|---:|---:|---:|---:| | Simple Fusion Light | 18.590 | 4.535 | 83.09 ± 0.22 | 21.22 ± 1.70 | | Simple Fusion Matched | 18.798 | 4.621 | 83.10 ± 0.18 | 19.18 ± 1.10 | | DiCoSeg | 18.782 | 4.611 | 83.77 ± 0.18 | 14.73 ± 1.54 | At closely matched parameters and FLOPs, DiCoSeg improves DSC by 0.67 points and reduces HD95 by 4.45. We also isolated the effect of supervision within the unchanged DiCoSeg architecture: ordinary deep supervision obtains 82.68 DSC / 19.63 HD95, compared with 83.96 / 15.45 using corrective supervision (fixed seed). The alternative fusion design and the supervision control with an unchanged architecture address complementary questions and support the value of the full corrective formulation. Further details are in §4.4 and Tables 4–5. ### **To W5:** We acknowledge the initial inference overhead. We **optimized DiCoSeg's complete execution path by combining equivalent batching, vectorization, and parallelization of corrective computations with CUDA Graph replay to reduce execution overhead**. This preserves the mathematical operations, trained weights, parameters, and FLOPs. CUDA Graph also benefits VSS Baseline, and the optimized implementation is included in the supplementary materials. | Model | Initially reported latency (ms) | Optimized latency (ms) ↓ | Optimized FPS ↑ | |---|---:|---:|---:| | VSS Baseline | 8.77 | 4.354 ± 0.002 | 229.66 ± 0.09 | | DiCoSeg | 37.78 | 8.349 ± 0.002 | 119.78 ± 0.03 | On an NVIDIA A800, optimized DiCoSeg runs at 8.349 ms/image (119.78 FPS), below the 8.77 ms/image originally reported for VSS Baseline, demonstrating a manageable inference cost after optimization. Under the optimized execution setup, VSS Baseline achieves 4.354 ms/image (§4.5 and Appendix B).
AuthorsOpenReview ↗

Response to Questions Q1–Q2

To Q1:

The main results and detailed Synapse ablations use checkpoints saved by minimum training loss, not by test_vol DSC. In the additional 15/3/12 protocol, only validation volumes are evaluated during training, and checkpoint selection uses validation DSC. Test metrics are computed after selection. DiCoSeg achieves 81.63 ± 0.19 DSC / 17.59 ± 1.28 HD95, compared with 80.19 ± 0.20 / 23.85 ± 1.45 for VSS Baseline across three runs.

To Q2:

We implemented exactly this type of replacement for the entire CFB, including a version matched in parameters and FLOPs. It obtains 83.10 ± 0.18 DSC / 19.18 ± 1.10 HD95, versus 83.77 ± 0.18 / 14.73 ± 1.54 for DiCoSeg at closely matched parameters and FLOPs. The architecture and the three-run comparison are shown directly in W4 above.

Reviewer_5QWCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DiCoSeg, a semantic segmentation framework that formulates hierarchical decoding as progressive task-conditioned correction. It introduces Consistency-Guided Directional Aggregation (CGDA) for adaptive directional feature aggregation and a Corrective Fusion Block (CFB) that selectively refines encoder-decoder representations using a persistent task state. Experiments on medical segmentation benchmarks and ADE20K demonstrate improved segmentation performance over existing methods.

Critical strengths
  1. The paper presents an intuitive motivation for separating feature discrepancies from task-conditioned corrective updates.

  2. The proposed method achieves competitive performance across multiple medical segmentation datasets and ADE20K.

  3. The ablation studies investigate several key components, including corrective supervision, correction transport, and the persistent task state.

Critical weaknesses
  1. Limited novelty. The proposed framework mainly combines existing concepts, such as adaptive feature fusion, spatial propagation, and auxiliary supervision. The distinction from prior task-adaptive segmentation methods is not sufficiently convincing.

  2. Potential evaluation bias. Appendix A.1 indicates that checkpoint selection on Synapse is performed using the test_vol split. Selecting the best checkpoint based on test performance raises concerns about evaluation fairness and potential overfitting to the test set.

  3. Insufficient hyperparameter analysis. The method introduces multiple loss weights, thresholds, and correction-related parameters. However, sensitivity analyses for these hyperparameters are largely missing, making it difficult to assess the robustness of the proposed framework.

  4. Limited scope of ablation studies. Most ablations are conducted only on Synapse. Furthermore, comparisons with simpler task-conditioned fusion mechanisms under matched computational budgets would better justify the proposed architectural complexity.

  5. Significant inference overhead. Despite only a 2.67% increase in FLOPs, inference latency increases by approximately 4.3× compared to the baseline (Table 8). This substantial computational overhead raises concerns about practical efficiency.

Questions for authors
  1. Could the authors clarify the checkpoint selection protocol for Synapse and provide results using an independent validation set?

  2. How sensitive is the performance to the corrective supervision weights and support-related thresholds?

  3. Can the authors provide comparisons with simpler task-conditioned fusion alternatives under matched computational budgets?

  4. Can the key ablations be extended to additional datasets to demonstrate robustness and generalizability?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 5QWC — W1–W3

We thank the reviewer for the thoughtful assessment and constructive questions, which helped us clarify the technical contribution and strengthen its experimental validation.

To W1:

We appreciate the reviewer's concern about novelty relative to existing task-adaptive methods. We formulate hierarchical segmentation decoding as progressive task-conditioned correction. Encoder–decoder discrepancy supplies potential refinement information, but does not by itself determine where or how the current segmentation should be corrected. This distinction motivates treating the corrective action as an explicit computational object rather than simply using discrepancy for feature aggregation.

During training, task residuals supervise correction support SlS_l and signed direction QlQ_l. At inference, the predicted signals form a transported corrective state ClC_l. The same corrective state governs both the discrepancy-derived feature update and the persistent task-state update, with the accumulated task state conditioning correction at the next decoder stage. This supervised and coupled computation is the key distinction from a coarse-prediction gate that modulates current feature fusion. CGDA complements the decoder formulation through consistency-guided directional aggregation.

Matched Fusion, Matched Deep Supervision, and the T0T_0 reuse intervention provide complementary tests of the complete corrective design, correction-specific supervision, and cross-stage state accumulation. We have clarified the technical relationships with prior refinement and propagation methods in §2 and Appendix A.15.

To W2:

We thank the reviewer for carefully examining the Synapse protocol. We sincerely apologize for the incorrect statement in Appendix A.1, which suggested selecting checkpoints by the highest test-set DSC. That statement describes neither our implementation nor the model-selection rule used to produce the reported results.

On the established 18 training / 12 testing Synapse split, we follow the checkpoint-saving rule of the public VM-UNet Synapse training implementation: the checkpoint with the lowest epoch-averaged training loss is used for the reported evaluation. Periodic volume evaluation every 100 epochs is used only for monitoring. Its DSC and HD95 measurements do not inform checkpoint selection or hyperparameter tuning. We have corrected Appendix A.1 accordingly. The error was in the manuscript description, not in the actual checkpoint-selection procedure or the reported results.

For additional independent evidence, we conducted three runs using a separate split of 15 training, 3 validation, and 12 test cases. During training, only the three validation volumes were evaluated, and checkpoints were selected using validation DSC. The 12 test cases were evaluated after checkpoint selection.

ModelDSC (%) ↑HD95 ↓
VSS Baseline80.19 ± 0.2023.85 ± 1.45
DiCoSeg81.63 ± 0.1917.59 ± 1.28

This protocol uses 15 rather than 18 training cases and selects checkpoints by validation DSC, so its absolute scores should be interpreted within that setting. DiCoSeg outperforms VSS Baseline under the same validation protocol. We now also report this experiment in Appendix A.1, while retaining the standard 18/12 split for comparisons with prior work.

To W3:

Thank you for suggesting a sensitivity analysis. We added three-run Synapse experiments covering both the overall strength of corrective supervision and the support-target threshold.

FactorSettingDSC (%) ↑HD95 ↓
Corrective loss scale0.5×83.10 ± 0.1715.53 ± 1.48
Corrective loss scale1×83.77 ± 0.1814.73 ± 1.54
Corrective loss scale2×83.47 ± 0.1317.41 ± 1.03
Support threshold0.1083.17 ± 0.1818.70 ± 1.20
Support threshold0.2083.77 ± 0.1814.73 ± 1.54
Support threshold0.3083.64 ± 0.1214.28 ± 1.01

Loss scaling applies jointly to the four corrective supervision terms. The support threshold controls the soft training target.

Across the tested supervision strengths, the default setting gives the highest mean DSC and the lowest mean HD95. Increasing the support threshold to 0.30 yields a modest HD95 improvement with a small DSC trade-off. Both observations are reported in Appendix A.11, Table 11.

AuthorsOpenReview ↗

Response to Weaknesses W4–W5

To W4:

We extended the five key variants to both Synapse and ISIC2018, with three independent runs per variant. Full DiCoSeg achieves the best mean on all four reported metrics:

VariantSynapse DSC ↑Synapse HD95 ↓ISIC2018 mIoU ↑ISIC2018 DSC ↑
VSS Baseline82.05 ± 0.1822.63 ± 1.8581.16 ± 0.1689.60 ± 0.10
CGDA Only83.16 ± 0.1519.72 ± 1.2281.84 ± 0.0890.01 ± 0.05
CFB Only83.07 ± 0.1120.17 ± 1.1581.91 ± 0.0990.06 ± 0.06
Discrepancy-Only83.62 ± 0.0819.51 ± 0.5682.04 ± 0.0990.13 ± 0.05
DiCoSeg83.77 ± 0.1814.73 ± 1.5482.35 ± 0.1990.32 ± 0.11

We also tested Light and Matched variants of fusion guided by coarse predictions as simpler alternatives. Both retain the CGDA encoder and VSS decoder while replacing the entire CFB with a gate that uses the current coarse prediction to modulate encoder–decoder discrepancy. Matched closely matches DiCoSeg's parameter and FLOPs budgets, without CFB's corrective state and task state evolution across decoder stages.

ModelParams (M)FLOPs (G)Synapse DSC (%) ↑HD95 ↓
Simple Fusion Light18.5904.53583.09 ± 0.2221.22 ± 1.70
Simple Fusion Matched18.7984.62183.10 ± 0.1819.18 ± 1.10
DiCoSeg18.7824.61183.77 ± 0.1814.73 ± 1.54

At matched budgets, DiCoSeg gains 0.67 DSC points and reduces HD95 by 4.45. Separately, for seed 42, replacing corrective supervision with ordinary deep supervision in the unchanged DiCoSeg architecture gives 82.68 DSC / 19.63 HD95, versus 83.96 / 15.45 with corrective supervision. The gated replacement and the supervision control with an unchanged architecture address complementary aspects of the design. Definitions and results are in §4.4, Appendix A.10–A.12, and Tables 4–5 and 9.

To W5:

We acknowledge the substantial inference overhead of our initial implementation. We optimized DiCoSeg's complete execution path by combining equivalent batching, vectorization, and parallelization of corrective computations with CUDA Graph replay to reduce execution overhead. These changes preserve the mathematical operations, trained weights, parameters, and FLOPs. We also apply CUDA Graph to VSS Baseline, allowing both models to benefit from the optimized execution setup. The optimized implementation is included in the supplementary materials.

ModelInitially reported latency (ms)Optimized latency (ms) ↓Optimized FPS ↑
VSS Baseline8.774.354 ± 0.002229.66 ± 0.09
DiCoSeg37.788.349 ± 0.002119.78 ± 0.03

On an NVIDIA A800, optimized DiCoSeg achieves 8.349 ms/image (119.78 FPS), below the 8.77 ms/image originally reported for VSS Baseline. This makes the remaining inference cost practical. With the same optimized execution setup, VSS Baseline achieves 4.354 ms/image. Full efficiency results are in §4.5 and Appendix B.

AuthorsOpenReview ↗

Response to Questions Q1–Q4

To Q1:

The main Synapse experiments select checkpoints by training loss, not test DSC. In our additional 15 train / 3 validation / 12 test experiment, only validation volumes were evaluated during training, checkpoint selection used validation DSC, and the test split was evaluated after selection. Across three runs, DiCoSeg obtains 81.63 ± 0.19 DSC / 17.59 ± 1.28 HD95, versus 80.19 ± 0.20 / 23.85 ± 1.45 for VSS Baseline. This complements the established 18/12 benchmark used in the main paper.

To Q2:

Across three runs, the default loss scale gives 83.77 ± 0.18 DSC / 14.73 ± 1.54 HD95, versus 83.10/15.53 at 0.5× and 83.47/17.41 at 2×. Changing the support threshold from 0.20 to 0.30 gives 83.64 DSC / 14.28 HD95. The complete mean ± SD results are provided in W3 above.

To Q3:

Yes. We tested Light and computationally Matched alternatives that replace CFB with fusion guided by coarse probabilities. The Matched control uses 18.798M parameters / 4.621G FLOPs, compared with 18.782M / 4.611G for DiCoSeg, and reaches 83.10 DSC / 19.18 HD95 versus DiCoSeg's 83.77 / 14.73. The full controlled results are shown in W4 above.

To Q4:

Yes. We repeated all five key variants in three independent runs on Synapse and ISIC2018. Full DiCoSeg obtains the best mean in every reported metric, including 82.35 ± 0.19 mIoU and 90.32 ± 0.11 DSC on ISIC2018. The full results are displayed in W4 above and documented in Appendix A.10, Table 10.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We sincerely thank the reviewers for their time, careful evaluation, and constructive feedback. We also appreciate the Area Chair and Senior Area Chairs for their efforts in coordinating the review process. In response, we have clarified DiCoSeg's technical contribution, corrected the Synapse checkpoint description, and strengthened the experimental evidence and implementation.

1. Technical novelty. DiCoSeg formulates hierarchical segmentation decoding as progressive task-conditioned correction. The central question is not merely how encoder and decoder features should be fused, but which observed discrepancies warrant correction and in what direction, given the current segmentation state. We address this through correction estimates supervised by task residuals and a shared corrective state that couples current feature refinement with persistent task-state updates. This makes corrective action an explicit part of hierarchical decoding rather than relying on feature fusion alone.

2. Synapse evaluation. We apologize for the inaccurate checkpoint-selection description in the original manuscript. Under the standard 18/12 Synapse protocol, checkpoints were selected using the lowest epoch-averaged training loss, not test DSC. We corrected the manuscript and, following the reviewers' suggestion, conducted an additional three-run experiment with 15 training, 3 validation, and the original 12 test cases. The results further confirm DiCoSeg's advantage over the VSS Baseline.

3. Practical efficiency. We optimized DiCoSeg's complete execution path through equivalent computational restructuring and CUDA Graph replay. On an NVIDIA A800, DiCoSeg achieves 8.349 ms/image (119.78 FPS). The optimized code is included in the supplementary materials.

4. Controlled experiments. We added fusion controls matched in parameters and FLOPs, a deep supervision control using the same architecture, three-run ablations on Synapse and ISIC2018, and sensitivity analyses.

5. Mechanism verification. Support localization, stage-wise error measurements, and inference interventions, including task-state reuse, provide direct evidence for the proposed progressive correction mechanism.

6. Related work. We expanded comparisons with classical refinement and propagation methods to clarify the technical distinctions of our coupled corrective formulation.

Area_Chair_Kp3vOpenReview ↗

Reminder: Please Follow Up on the Author Rebuttal

Dear Reviewers,

As we are approaching the end of the discussion period, I would like to remind you to review the authors' rebuttal and assess whether your concerns have been adequately addressed.

If you have any remaining concerns, please raise them as soon as possible to allow the authors sufficient time to respond. If your concerns have been resolved, please leave a brief acknowledgment to let the authors know that you have read and considered their responses. Thank you for your time and efforts in the review process.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.