acceptodds
Withdrawn from ICLR 2027

RIR: Relative Intensity Relations for Single-Source Generalization under Unseen Pathology Shifts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Medical domain generalization typically addresses shifts in acquisition conditions, institutions, or patient cohorts, yet costly annotation and fragmented clinical datasets may leave deployment conditions beyond the pathology coverage available during source training. We study brain tumor segmentation under such external shifts and ask which source-derived evidence remains useful when absolute intensity, morphology, and ordinal intensity position may change across tumor categories. We propose Relative Intensity Relation (RIR), a source-only framework that distinguishes absolute intensity from image-internal contrast organization. RIR derives within-volume intensity rank and local intensity ordering from each input, reducing direct dependence on absolute intensity scale while encoding image-internal relations. Because these relations are recomputed from each input, RIR does not transfer a fixed source intensity or morphology template across cohorts. Source segmentation supervision grounds these rank-derived cues into tumor consistency, tumor-background separation, and boundary evidence, which is integrated with the original image representation through a bounded residual correction. The framework further selects its deployment checkpoint and decision threshold using held-out source validation only, without target-domain feedback. Across five source-supervision sizes, RIR achieves the highest mean two-cohort external macro Dice among the compared methods and the highest mean sensitivity on both BraTS-Africa cohorts at all five sizes. The broader comparison also shows substantial degradation beyond the source glioma category across multiple DG strategies. A representation-level diagnostic further shows smaller source-to-target displacement of source-defined tumor scores in the pre-grounding RIR space than in absolute FLAIR. Together, these results support the utility of the complete RIR framework and provide evidence that rank-derived relational structure remains useful under pathology-associated external shifts.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
domain generalization, single-source domain generalization, medical image segmentation, relative intensity relations, relational representation learning, brain tumor segmentation, pathology shift, MRI

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5K9HOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Relative Intensity Relations (RIR), a framework that leverages within-volume intensity ranks and local intensity relationships to improve single-source generalization for brain tumor segmentation under unseen pathology-associated shifts. The method integrates relational features with conventional image representations and incorporates source-only model selection. While the problem is relevant, the current evaluation is limited in scope and does not sufficiently substantiate the broader generalization claims.

Critical strengths

The study addresses single-source domain generalization under incomplete pathology coverage, a meaningful challenge in medical image segmentation. Modeling relative intensity relationships provides a reasonable complement to absolute intensity features and may improve robustness to distributional variation.

Critical weaknesses
  1. The title, “RIR: Relative Intensity Relations for Single-Source Generalization under Unseen Pathology Shifts,” suggests a general methodology for single-source generalization under pathology shifts. However, the experiments are limited to brain tumor segmentation on FLAIR MRI, with a single glioma source cohort and two external cohorts from BraTS-Africa. In particular, the difference between the source glioma cohort and the Other Neoplasms cohort does not isolate pathology shift from other sources of distributional variation, such as differences in tumor burden, morphology, and spatial distribution. Moreover, both external cohorts remain within the relatively narrow setting of brain tumor segmentation. The current evidence therefore does not establish generalization across pathology shifts in a broader sense, nor does it demonstrate applicability to substantially different medical imaging tasks or clinical settings. The title and claims should be narrowed to reflect the actual experimental scope.

  2. The evaluation relies too heavily on Dice and lacks a sufficiently comprehensive assessment of segmentation quality. The experimental analysis primarily emphasizes Dice similarity coefficient, with sensitivity reported as an additional metric. This is insufficient to fully characterize segmentation performance, particularly when the method is intended to improve robustness under substantial distribution shifts. Boundary-based metrics, such as HD95 and average surface distance, are important for assessing contour accuracy and the spatial extent of segmentation errors. Furthermore, the manuscript does not establish whether the observed improvements translate into benefits for clinically relevant downstream tasks, such as tumor burden quantification, lesion characterization, or treatment planning. Without more comprehensive geometric evaluation and downstream validation, the practical significance of the reported improvements remains unclear.

  3. The comparison with established segmentation methods and domain generalization approaches is insufficient. The main quantitative comparison includes BASE, RandConv, and DSU, which provides a limited view of the proposed method's relative performance. The study lacks comparisons with strong segmentation baselines, such as nnU-Net, and a broader range of competitive domain generalization methods under matched experimental conditions. These comparisons are necessary to determine whether RIR offers a meaningful advantage over well-established segmentation pipelines and existing generalization strategies, rather than simply outperforming a small subset of baselines. The broader comparison presented in the manuscript does not fully resolve this concern unless the methods are evaluated under consistent data splits, supervision budgets, and training protocols.

  4. The experimental setting does not sufficiently support the emphasis on pathology generalization. Although the manuscript frames unseen pathology as a central challenge, the evaluation is confined to a specific MRI segmentation task and a limited set of related cohorts. The external Other Neoplasms cohort introduces variation beyond tumor category, making it difficult to attribute the observed performance differences specifically to pathology shifts. In addition, the manuscript does not adequately investigate whether the proposed relative-intensity representation remains useful across substantially different acquisition conditions, imaging modalities, anatomical regions, or disease categories. Consequently, the current experiments support, at most, the utility of RIR in a particular external brain tumor segmentation setting, rather than the broader pathology-generalization capability suggested by the title.

  5. The qualitative visualizations require substantial improvement in both presentation and technical clarity. The figures appear insufficiently polished for a conference submission of this level. More importantly, Figure 4 appears to present the task as a simplified binary tumor-versus-background segmentation problem, without adequately clarifying how the evaluation relates to the different tumor subregions or annotation categories in the BraTS setting. As currently presented, the figure makes it difficult to assess the method's behavior on different lesion components and to verify that the visual comparisons are fully representative of the underlying segmentation task.

Questions for authors

Refer to weakness

Reviewer_64HROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies single-source domain generalization (SSDG) for brain tumor segmentation when deployment data extend beyond the pathology category observed during source training. Instead of relying exclusively on absolute MRI intensities or conventional appearance perturbations, the authors propose Relative Intensity Relations (RIR), which represent each voxel through its within-volume intensity rank and eight local directional rank comparisons. Source-domain annotations then supervise local and region-level affinities, tumor-background separation, and boundary evidence. These relational features are integrated with U-Net image features through two stages of local cross-attention and bounded residual corrections. The framework additionally includes a source-only checkpoint and threshold selection procedure that prioritizes sensitivity among checkpoints satisfying a source-validation Dice constraint. Experiments use FLAIR-only, 2D whole-tumor segmentation, with 10–102 labeled source training cases, and evaluation on BraTS-Africa Glioma and Other Neoplasms cohorts. The full framework achieves the highest mean two-cohort external macro Dice among BASE, RandConv, DSU, and RIR at all five source-supervision sizes.

Critical strengths
  1. Relevant and clearly motivated evaluation setting. The paper explicitly addresses missing pathology coverage in SSDG rather than focusing only on scanner or acquisition shifts. Including a same-category external glioma cohort alongside a heterogeneous beyond-glioma cohort highlights an important practical challenge.
  2. Reasonable representation and integration design. Within-volume ranks and directional comparisons reduce direct dependence on absolute intensity magnitude under order-preserving transformations. The affinity and boundary objectives make these generic cues task-relevant, while the residual image anchor helps avoid fully replacing useful appearance information.
  3. Transparent source-only deployment and repeated evaluation. The authors describe a target-blind checkpoint/threshold selection rule, evaluate five labeled-source sizes with three seeds, and report a matched-trajectory Select ablation. The supplementary implementation details and explicit limitations improve interpretability and reproducibility.
Critical weaknesses
  1. Insufficient component-level evidence for the claimed mechanism. Represent, Ground, and Integrate are trained jointly, but the manuscript does not isolate the benefits of volume ranking, eight-direction ordinal cues, local/regional affinity objectives, boundary supervision, or cross-attention integration. It is therefore unclear whether the improvements stem specifically from relative-intensity relations, from standard auxiliary structural supervision, or simply from additional capacity and regularization. The statement that these components are coupled does not remove the need for controlled comparisons (Sections 3 and 4; Appendix H.1).
  2. Deployment selection is a substantial confounding factor in method comparisons. The full RIR method uses sensitivity-constrained checkpoint selection and source-optimized thresholds, while BASE/RandConv/DSU are not shown with the same selection rule. At N = 102, RIR without Select obtains 0.555 external macro Dice versus 0.570 for BASE, whereas full RIR reaches 0.589 (Table H.1). Thus, the observed full-framework advantage at this size cannot be attributed to relational learning alone. Applying identical selection rules to all baselines is essential for a clean comparison.
  3. The pathology-shift interpretation is not causally isolated. BraTS-Africa Other Neoplasms differs from the source in pathology mix as well as potentially acquisition, site, and cohort characteristics. The authors acknowledge this limitation, but the experiments cannot determine whether RIR is robust specifically to unseen pathology or mainly to correlated imaging/cohort shifts. A more controlled cross-pathology evaluation or pathology-stratified analysis would strengthen the central motivation (Section 4.1; Appendix A).
  4. Limited robustness evidence for ordinal assumptions and 3D context. RIR ranks are computed from the entire 3D volume, although segmentation uses a 2D backbone. This supplies global volume-derived statistics not obviously available in the same form to the slice-based baselines. Moreover, rank relations are invariant to strictly monotonic intensity transformations but not to arbitrary pathology-dependent contrast reordering, field-of-view changes, or intensity inversions. The paper does not directly test these failure modes or compare volume-wise and slice-wise ranking (Appendix B).
  5. Limited external breadth and uncertainty for modest improvements. The controlled study considers one FLAIR-only 2D task, one nested source-subset sequence, two external cohorts, and three seeds. At N = 51, full RIR improves over DSU by only 0.003 macro Dice, and at N = 102 by 0.016 (Table 1). These small gaps should be interpreted cautiously given run-to-run variation; the study provides descriptive means and standard deviations rather than statistical evidence of superiority. The one-dimensional representation diagnostic is suggestive but does not establish general pathology invariance (Appendix G).
Questions for authors
  1. Matched Select for all methods. Could you apply the identical source-only checkpoint and threshold selection procedure to BASE, RandConv, and DSU and report the resulting external Dice and sensitivity? This seems particularly important at N = 102, where RIR without Select underperforms BASE.
  2. Ablations that isolate the representation. Could you report controlled variants using (a) rank cues without the affinity/boundary objectives, (b) the same affinity/boundary supervision without rank-derived inputs, and (c) simple concatenation or residual fusion without the two-stage cross-attention? Where feasible, please match model capacity and deployment selection.
  3. Ordinal robustness tests. How sensitive is RIR to contrast-order reversals, nonlinear intensity transformations that do not preserve local ordering, different valid-field-of-view masks, and volume-wise versus slice-wise rank estimation? Such tests would more directly support the proposed invariance rationale.
  4. Disentangling pathology from cohort changes. Can the results be stratified by diagnostic category within Other Neoplasms, or supplemented with a setting in which imaging/acquisition characteristics are better matched across differing pathologies?
  5. Statistical reliability and source sampling. Are the improvements consistent across alternative source-subset draws, rather than one fixed nested subset sequence? Could you provide paired per-seed differences or confidence intervals for the small RIR–DSU gains?
  6. Practical cost and inputs. What are the additional parameter count, inference latency, and memory cost of volume-wise quantile preprocessing plus the two attention modules? Does inference require access to the full volume and an explicitly defined valid-field-of-view mask?
Reviewer_YVfQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles single-source domain generalization for brain tumour segmentation: a model is trained on one source dataset and evaluated on external cohorts without access to target images or labels. The task is 2D whole-tumour segmentation on FLAIR images. Models are trained on BraTS-GLI and tested on two BraTS-Africa cohorts (Glioma and Other Neoplasms). As expected, without any intervention, performance drops on the external cohorts (about 0.83 Dice in-domain versus roughly 0.4–0.6 externally). The proposed method, RIR, has four components: Represent (each voxel's intensity rank within the volume, plus comparisons with its 8 neighbours), Ground (affinity and boundary losses from the tumour mask teach which rank cues mean tumour), Integrate (a small gated attention correction added to U-Net features), and Select (checkpoint and threshold chosen using source validation only). RIR reports improved Dice score against baseline, RandConv and DSU.

Critical strengths

The problem is well motivated and the writing is comparatively readable.

Critical weaknesses

Use of only single-modality FLAIR MRI Images does not mimic clinical workflow and weakens the clinical framing although the datasets provide T1, T1ce and T2 (available in both BraTS-GLI and BraTS-Africa) all of which provide complementary information and essential in clinical workflow, the proposed method and compared baselines only use T2-FLAIR images to accomodate the proposed method. Hence, does not translate to real-world benefit specially for such clinically motivated problem.

Experiment rigor and baseline comparison most of the improvement in dice score seem to come from the Select step (adhoc checkpoint selection step) rather than the core intensity-based pixel ranking (rank-transform) (section H.1 appendix). Without the Select step, the proposed method seem to be within the CI of the baseline. Stronger baselines (DCON, TSIAA, AdverIN) appear only in Fig. 1b as problem motivation (reduction of performance in external cohort), and is surprisingly absent in Table 1 comparison of proposed RIR to these single-source DG methods. Atleast, DCON and AdverIN seem to be relevant methods. Also, no citation/reference to these methods, only names.

Questions for authors
  1. (C) Please report BASE, RandConv, DSU and the stronger methods in Fig. 1b (DCON, TSIAA, AdverIN) under the same Select rule.

  2. (C) Why FLAIR only, when T1, T1ce and T2 are available? What stops RIR from extending to multimodal and 3D input?

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

Related papers

Discussion (0)

Sign in to comment.