- 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
-
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.
-
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
-
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.
-
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
-
(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?
-
(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
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.
| Variant | Synapse DSC ↑ | Synapse HD95 ↓ | ISIC2018 mIoU ↑ | ISIC2018 DSC ↑ |
|---|---|---|---|---|
| VSS Baseline | 82.05 ± 0.18 | 22.63 ± 1.85 | 81.16 ± 0.16 | 89.60 ± 0.10 |
| CGDA Only | 83.16 ± 0.15 | 19.72 ± 1.22 | 81.84 ± 0.08 | 90.01 ± 0.05 |
| CFB Only | 83.07 ± 0.11 | 20.17 ± 1.15 | 81.91 ± 0.09 | 90.06 ± 0.06 |
| Discrepancy-Only | 83.62 ± 0.08 | 19.51 ± 0.56 | 82.04 ± 0.09 | 90.13 ± 0.05 |
| Full DiCoSeg | 83.77 ± 0.18 | 14.73 ± 1.54 | 82.35 ± 0.19 | 90.32 ± 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 setting | DSC (%) ↑ | ΔDSC (pp) | HD95 ↓ | ΔHD95 |
|---|---|---|---|---|
| Unchanged DiCoSeg | 83.964 | 0.000 | 15.446 | 0.000 |
| Set CGDA | 83.849 | −0.115 | 15.451 | +0.005 |
| Reuse initial at each stage | 83.820 | −0.144 | 15.778 | +0.332 |
| Replace with its spatial mean | 83.813 | −0.151 | 15.749 | +0.303 |
Single fixed checkpoint. Spatial averaging is applied to the post-tanh modulation.
Setting 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 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 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 -reuse intervention at each decoder stage, using the same trained checkpoint.
| Stage | Setting | Error before (%) ↓ | Error after (%) ↓ | Relative error reduction (%) ↑ |
|---|---|---|---|---|
| 1 | Full DiCoSeg | 2.245 | 1.575 | 29.84 |
| 1 | Reuse | 2.245 | 1.575 | 29.84 |
| 2 | Full DiCoSeg | 1.529 | 1.008 | 34.08 |
| 2 | Reuse | 2.403 | 1.747 | 27.29 |
| 3 | Full DiCoSeg | 1.555 | 0.997 | 35.86 |
| 3 | Reuse | 2.569 | 1.966 | 23.47 |
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 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.
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.
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 T0 reuse intervention examine its corrective behavior. The corresponding results are reported in §4.4 and Appendix A.12–A.15.