acceptodds
Withdrawn from ICLR 2027

Beyond Domain Alignment: Anatomically Constrained Unsupervised Adaptation for Optic Nerve Head Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Glaucoma causes irreversible vision loss, and accurate optic disc (OD) and optic cup (OC) segmentation from fundus images is essential for computing the vertical cup-to-disc ratio (vCDR) used in primary screening. Deep learning has automated OD/OC segmentation, but supervised models degrade under domain shift caused by heterogeneous devices, protocols, and populations. Existing unsupervised domain adaptation (UDA) methods align feature or output distributions but neglect OC-OD anatomical relationship, leading to clinically implausible predictions. We propose an anatomically constrained UDA framework for optic nerve head (ONH) segmentation that jointly addresses feature-level domain discrepancy and anatomical-level population variability. The framework combines automated ROI extraction, a segmentation backbone, and decoder adversarial alignment. To preserve valid ONH geometry without target annotations, a differentiable cup-to-disc area ratio is regularized by a two-sided extended log-barrier, with per-sample bounds adapted via class-conditional mechanism driven by a lightweight classifier. Ablation shows that anatomical regularization benefits only when the source-target cup-to-disc distributions differ significantly. We therefore enable it selectively using a two-sample hypothesis test on per-image ratios, evaluated without target labels. Trained on REFUGE as source domain, the gated framework achieves competitive cross-domain performance, with OC of 0.895 and 0.823 and OD of 0.982 and 0.974 on Drishti-GS and on RIM-ONE-r3, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Unsupervised domain adaptation, Medical image segmentation, Optic cup segmentation, Domain alignment, Fundus imaging

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xQeSOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an unsupervised domain adaptation framework for optic disc and cup segmentation, combining a SegFormer backbone, domain-specific batch normalization, and decoder-level adversarial alignment. Its main contributions are a differentiable cup-to-disc area-ratio regularizer with classifier-conditioned bounds, and a statistical gate that selectively activates this regularizer using a KS test between source ground-truth ratios and unlabeled target predictions.

Critical strengths
  1. The paper addresses a clinically relevant segmentation problem by incorporating the optic cup–disc relationship into UDA. Its differentiable cup-to-disc area ratio constraint combines a two-sided extended log-barrier with sample-specific bounds from an auxiliary classifier. This provides a meaningful complement to distribution alignment by incorporating anatomical knowledge while accommodating variation across patients.

  2. The evaluation covers all six directed transfers among REFUGE, RIM-ONE-r3, and Drishti-GS, with competitive results in the REFUGE-source comparisons. Beyond performance, the ablations examine when anatomical regularization helps or hurts, motivating its selective application. This strengthens the empirical support for the approach and clarifies its conditions of usefulness.

Critical weaknesses
  1. Table 2 states that the results with and without LbarL_{\mathrm{bar}} use “the same trained model,” although this term is a training-time loss. It is therefore unclear how these results were obtained and whether they isolate the effect of anatomical regularization. This ambiguity weakens the experimental evidence for a central component and its selective application through the gating mechanism.

  2. The classifier ablation is difficult to reconcile with the stated gradient flow. Section 3.5 blocks classification gradients from entering the shared encoder, so without anatomical constraints or other parameter coupling, the classifier should not directly affect segmentation learning. Yet adding it changes OC Dice from 0.835 to 0.819 on Drishti-GS and from 0.803 to 0.823 on RIM-ONE-r3. Without clarification or variability estimates, these results weaken confidence in the attribution of gains to individual components.

Questions for authors
  1. Table 2 presents results obtained with and without using LbarL_{bar} based on the same trained model; however, this loss is employed only during the training phase. Could the authors explain why different segmentation results are produced?

  2. Section 3.5 states that the classification loss is not backpropagated to the shared encoder, yet segmentation performance changes when the auxiliary classifier is introduced (as shown in Table 1). Could the authors explain the reason for this?

  3. The paper adopts the Cup-to-disc Area Ratio as an anatomical constraint but uses vCDR as one of the evaluation metrics. Could the authors further validate the effectiveness of this area ratio constraint regarding vertical anatomical structures?

  4. Could the authors provide a sensitivity analysis of key hyperparameters (such as tt, TT, and λbar\lambda_{bar}) to verify the stability of the proposed method?

  5. Compared to BEAL, this method employs more parameters, yet some performance gains are limited. Could the authors further discuss the trade-off between model complexity and performance?

Reviewer_kQ32OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies unsupervised domain adaptation for optic-disc and optic-cup segmentation. It combines a SegFormer backbone, domain-specific normalization and adversarial alignment with a differentiable cup-to-disc area-ratio barrier. A source-supervised glaucoma classifier determines sample-specific bounds, and a KS test comparing source ground-truth ratios with target predicted ratios selects whether to enable the constraint. Experiments cover six directed transfers among three datasets and report overlap and vCDR metrics.

Critical strengths
  1. Table 2 shows both improvements and negative transfer from anatomical regularization across six directions. This gives a useful reason to study when such constraints should be applied.
  2. Table 1 provides within-backbone evidence of methodological benefit. On REFUGE-to-Drishti-GS, cup Dice increases from 0.737 to 0.895; a fixed barrier alone reaches 0.874. These comparisons support a benefit beyond the backbone alone. Class-conditional bounds also recover OD performance relative to some fixed-barrier configurations.
  3. Reporting vCDR error alongside overlap makes the evaluation more informative for the intended task. The paper also explicitly acknowledges limitations of KS test power, its single-ratio prior and ROI extraction.
Critical weaknesses
  1. I could not reconstruct the training comparison in Table 2. It says its minus/plus-barrier columns use the same trained model (p.8, lines 378–383), while the barrier is a training objective. Section 4.1 also applies a static constraint during the first 30 epochs (p.6, lines 302–306). It is unclear whether the comparison branches from a shared warm-up, compares separately trained models, or uses different checkpoints. Table 1's no-barrier Base+Disc+Cls configuration obtains 0.819 cup Dice on Drishti-GS, whereas Table 2 reports 0.847 without the barrier. Different runs or training trajectories could explain the difference. Please specify the timelines, initialization and checkpoint selection so that the comparison can be interpreted.

  2. The evaluation sets are unclear. Appendix A, Table 4 lists 39 glaucoma and 85 healthy RIM-ONE-r3 images, totaling 124. However, pOSAL, Table I and CLR, Section 3 use a 99/60 RIM-ONE-r3 training/test split, 50/51 for Drishti-GS, and 400 REFUGE source-training images. Their published scores match the corresponding Table 3 entries. Please explain whether suspect cases are excluded, whether Table 4 is only descriptive, and which actual samples are used for adaptation and evaluation. The submission also needs baseline provenance, preprocessing and model-selection details. These omissions leave the comparability of the results uncertain. The 0.003 Drishti cup-Dice margin over CLR lacks uncertainty estimates, although the paper's description of OC performance as competitive is reasonable.

  3. The KS test compares source labels with target predictions after warm-up (p.5, lines 264–269), so its result can reflect prediction error as well as anatomical differences. The proxy may still be useful, but the claim that statistical distinguishability governs benefit (p.8, lines 393–397) is stronger than the observed association across three reused datasets. Please distinguish evidence used to formulate the rule from evidence testing a fixed rule, and report uncertainty for the decisive paired comparisons. Evaluation-only diagnostics could support the anatomical interpretation. Alternatively, a narrower claim about the predictive proxy, backed by a controlled test, would address this concern.

  4. The broad statement that prior UDA neglects the OC–OD anatomical relationship (p.1, lines 17–20) overlooks pOSAL, which combines morphology-aware segmentation with output-space adaptation. The same reference is also incorrectly grouped with image translation on p.2, lines 99–101. The disease-conditioned area bounds and KS gate remain distinct contributions. Similarly, CLR regularizes pixel-category representations, rather than providing the same glaucoma-conditioned bounds. Please explain and support this specific increment.

  5. Two Dice/IoU pairs in Table 5 do not satisfy the relation expected for the same masks, images and averaging weights. Appendix B defines standard binary-mask Dice and IoU. For each image, IoU = Dice/(2−Dice); convexity implies mean IoU ≥ mean Dice/(2−mean Dice) when both metrics use the same images and weights. However, Table 5 reports OC Dice/IoU of 0.816/0.658 for Drishti-to-REFUGE and 0.769/0.587 for Drishti-to-RIM. The corresponding lower bounds are approximately 0.689 and 0.625, respectively, and three-decimal rounding cannot explain the discrepancies. Different aggregation schemes could account for this, but they are not specified. Please provide the per-image metrics and aggregation definitions, or correct the table and explain whether the substantive conclusions change.

Questions for authors
  1. (W1) What does "same trained model" mean in Table 2? Please give the shared or separate warm-up, gate checkpoint, subsequent training branches and checkpoint selection, and reconcile 0.819 versus 0.847 across Tables 1 and 2.

  2. (W2) What are the actual source training/validation and target adaptation/evaluation sets? How are RIM-ONE-r3 suspect cases handled, and are the Table 3 baselines evaluated on the same samples? Please distinguish quoted scores from reruns and specify whether evaluation is transductive.

  3. (W3) What evidence tests the fixed gate separately from the observations used to motivate it? Please report the validation design and uncertainty for paired comparisons; existing records that establish this separation would help.

  4. (W4) How were the normal/glaucoma bounds and temperature selected without target evaluation labels, and what evidence supports the classifier-conditioned bounds? Please distinguish this mechanism from pOSAL's morphology loss and CLR's pixel-category regularization.

  5. (W5) How were Dice and IoU aggregated in Table 5? Can you provide the per-image confusion counts or metric pairs for the two Drishti-source transfers, reconcile the stated bounds, and identify any corrections needed elsewhere?

  6. (W1) Please clarify two implementation details: Equation (2), as written, scales hard-label BCE terms by 0.9 rather than including both soft-label BCE terms; and the detached classifier should not directly change segmentation when the barrier is absent. Does the implementation match these descriptions, and are the no-barrier changes in Table 1 due to different runs or another pathway?

  7. In Equation (9), the shrinkage factor approaches one for highly confident predictions, including incorrect ones. How would these errors affect the anatomical bands? Does existing analysis show when interpolation helps or harms segmentation? A diagnostic or limitation discussion should distinguish reducing the influence of low-confidence predictions from correcting miscalibration.

  8. With OD defined as rim plus cup, cup inclusion holds by construction, but an area ratio does not ensure a plausible boundary shape, connectivity or vCDR. Do existing outputs show cases where the ratio is within the selected band but these properties remain poor? Do the constraints improve such cases? Qualitative examples or a discussion of this limitation would help.

Reviewer_eQ7AOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an unsupervised domain adaption method for optic disc/cup segmentation for retinal images. It trains SegFormer with labeled source images and unlabeled target images with adversarial decoder feature alignment (which is a discriminator to discriminate source and target domain decoder features). The main addition is an anatomical loss on the cup-disc area ratio, with bounds determined by an auxiliary glaucoma classifier. It performs experiments on three retinal image dataset and showed good results.

Critical strengths

The paper is very clear and method straightford. The results are good, and the use of that anatomical prior (differentiable soft cup-to-disc area ratio) seems interesting and new.

Critical weaknesses

My major concern is the limited novelty and broader impact of the proposed method. The underlying adversarial adaptation framework is established, and I do not see a substantial methodological contribution to UDA. The main potentially distinctive component is the soft cup-to-disc area constraint and its selective activation. However, this remains a task-specific heuristic, and the experiments do not establish a broader principle for domain adaptation or medical image segmentation. Also this loss seems to improve the segmentation network as well, not something specific for the UDA. The evaluation is restricted to three retinal datasets and one SegFormer backbone, providing limited evidence that the proposed approach generalizes beyond this particular setting. Although the anatomical bounds are clinically motivated, their construction and interpolation are largely hand-specified. So in overall this is a pratically good paper but the impact and generalizability are not convincing and personally I don't think it reaches the ICLR bar.

Questions for authors

None.

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

Related papers

Discussion (0)

Sign in to comment.