List of Papers Browse by Subject Areas Author List
Abstract
Multi-region severity scoring on chest X-rays (CXR) faces three structural challenges: extreme class imbalance across regions of interest (ROIs), ordinal label discontinuity, and inter-rater subjectivity. We argue that performance degradation in sparse high-severity grades stems from ordinal geometry collapse—the compounding interaction of Neural Collapse and Minority Collapse in the representation space—rather than from data scarcity alone.
To address this, we propose the Decoupled Ordinal Refinement with Geometric Alignment (DORGA), which operates within an explicitly designed Ordinal-Aligned Latent Space.
Our approach is threefold: (i) learnable anchors constrained by polar fixing and ordinal separation pin the severity coordinate system on the hypersphere a priori; (ii) ROI embeddings are orthogonally decoupled into severity and structure components, preventing severity similarity from distorting inter-ROI attention; and (iii) pattern-conditional Bayesian propagation refines ambiguous severity embeddings via confidence-gated spherical interpolation, guided by oracle-supervised structure-aware attention during training. DORGA achieves an average MAE of 0.301 on the Brixia consensus test set, compared to 0.441 (BS-Net) and 0.350 (PAFE). Without fine-tuning, it attains QWK 0.797 with the senior radiologist on the external Cohen dataset, matching inter-rater agreement (0.790).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4660_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/mai-inha/DORGA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JeoGeo_Decoupled_MICCAI2026,
author = { Jeong, Geon AND Kim, Jung Soo AND Lee, Hyun Gyu},
title = { { Decoupled Ordinal Refinement with Geometric Alignment for Chest X-ray Severity Scoring } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes DORGA, a novel framework for multi-region ordinal severity scoring on chest X-rays. The key contribution is the identification of ordinal geometry collapse as the main failure mode under class imbalance, and the introduction of a geometrically structured solution.
Specifically, the method:
Defines a severity axis using learnable anchors constrained on a hypersphere, Decouples ROI representations into severity and structure components, Introduces a pattern-conditional Bayesian message passing mechanism with oracle-guided attention, Performs confidence-gated spherical refinement (SLERP) for prediction correction.
This combination forms a coherent geometric framework addressing both representation collapse and inter-ROI ambiguity.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
[Well-motivated problem formulation] The paper identifies an important and often overlooked issue in ordinal prediction under imbalance. The concept of ordinal geometry collapse provides a unifying perspective that connects known phenomena (e. g. , Neural Collapse, Minority Collapse) to the specific challenges of multi-region ordinal prediction. [Shift from loss design to representation design] A key strength of this work is the decision to enforce ordinal structure directly in the latent space via anchor-based geometric constraints, rather than relying solely on loss functions. This perspective is both conceptually interesting and practically meaningful. [Decoupling of severity and structure representations] The separation between severity embedding and structure embedding is a particularly strong design choice. It directly addresses a known limitation of prior contrastive or ordinal methods, where severity similarity can distort relational modeling. The analysis in Fig. 3 provides useful evidence that this decoupling leads to more balanced attention behavior.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
[Limited empirical validation of the “ordinal geometry collapse” hypothesis] While the paper provides a compelling conceptual argument for ordinal geometry collapse, the empirical validation of this phenomenon within the learned representations is limited. Fig. 1 serves as an illustration rather than a measurement. Additional analyses such as angular distributions, cluster separability, or evolution of embeddings during training would strengthen the claim. [Lack of detailed dataset distribution analysis] The paper emphasizes severe class imbalance across ROIs as a core motivation, but does not provide detailed statistics (e. g. , per-ROI grade distributions, frequency of high-severity cases). Including such information would make the problem setting more concrete and better justify the design choices. [Interpretability of individual components could be improved] Although the ablation study demonstrates the importance of each module, the paper provides limited qualitative or case-based analysis explaining how specific components correct particular types of errors. More interpretive analysis would improve understanding of the method. [Figure 2 is difficult to interpret] Figure 2 attempts to summarize multiple aspects of the method in a single visualization, but the result is visually dense and difficult to follow. The text labels are small, and several concepts are presented simultaneously, making it challenging for readers to grasp the workflow at a glance. Splitting this figure into simpler, step-by-step diagrams or improving readability would significantly enhance clarity. [Limited representation-level analysis] Despite the central role of geometry in the proposed method, the paper provides limited direct visualization or analysis of the learned embedding space (e. g. , before/after alignment, anchor relationships, or disentanglement quality). Such analysis would better support the geometric interpretation. [Pattern-based prior design is underexplained] The use of K-means clustering to derive pattern-specific transition matrices is interesting, but the rationale for choosing K=7 and the interpretability of these patterns are not fully discussed. Additional analysis or sensitivity studies (ablation?) would strengthen this component.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Providing visualization of the embedding space (e. g. , t-SNE or angular plots) would greatly strengthen the geometric claims.
Including detailed dataset statistics (class distribution per ROI) would improve the clarity of the problem formulation.
Simplifying or restructuring Figure 2 would significantly enhance accessibility for readers.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a well-motivated and thoughtfully designed framework for ordinal severity scoring in chest X-rays. The key strength lies in its geometric perspective on ordinal learning and the integration of multiple components into a coherent system.
The experimental results, particularly on consensus annotations and external validation, demonstrate the practical effectiveness of the approach. The decoupling strategy and anchor-based alignment are especially compelling.
However, the paper would benefit from stronger empirical validation of its central hypothesis (ordinal geometry collapse), more detailed dataset analysis, and improved interpretability of its components. Additionally, the presentation of some figures (notably Figure 2) could be improved.
Overall, the work is above the acceptance threshold due to its conceptual clarity, solid experimental results, and potential impact, but there is room for improvement in analysis and presentation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes DORGA, a framework for automated multi-region ordinal severity scoring on chest X-rays using the Brixia score. The core argument is that performance failures on rare high-severity grades are not simply a data scarcity problem. The authors argue it is a geometry problem: under class imbalance, neural collapse and minority collapse interact in the hyperspherical representation space and destroy the ordinal coordinate system that severity grading needs. They call this ordinal geometry collapse. The solution is built around constructing an explicit ordinal geometry on the unit hypersphere before training can distort it, using learnable grade anchors with polar and ordinal separation constraints. ROI embeddings are then decoupled into a severity component and a structure component that is orthogonal to the severity axis, so that inter-ROI graph attention is based on anatomical relationships rather than grade similarity. Stage 3 adds pattern-conditional belief propagation with oracle-supervised attention that teaches the model which neighboring ROIs are corrective during training, and a confidence-gated SLERP step that refines severity embeddings toward the inter-ROI consensus. What makes this different from prior work like SCOL and HCOR is that those methods enforce ordinal margins but keep severity and structure entangled in a shared embedding, which the paper shows distorts GNN attention toward grade similarity. DORGA is the first framework that explicitly separates these two concerns while also constructing the ordinal geometry from scratch rather than relying on the loss function to recover it implicitly.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The strongest aspect of the paper is the tight correspondence between the theoretical diagnosis and each architectural component. Every design choice directly addresses a specific failure mode, and the ablation in Table 1 confirms that the components are mutually dependent rather than independently additive. The MAE progression from S2-only (0.347) to removing oracle (0.337) to removing decoupling (0.329) to full DORGA (0.301) tells a coherent story, and the non-trivial finding that removing oracle supervision hurts more than removing decoupling reveals a genuine interaction between the two stages.
2.The Part A versus Part B asymmetry is the most convincing empirical result. The improvement over BS-Net is modest on the single-annotator blind test set (MAE 0.426 vs 0.441) but becomes large on the 5-radiologist consensus test set (MAE 0.301 vs 0.424). This is exactly what the ordinal geometry hypothesis predicts, and it is not something that can be explained by overfitting because the consensus reference was not available during training.
3.The attention distribution analysis in Figure 3 is also a genuine strength. The same/other attention ratio of 1.057 with decoupling versus 0.559 without it is a direct mechanistic validation that the structure embedding successfully excludes severity information from neighbor selection. This kind of intermediate diagnostic evidence is more convincing than a downstream metric improvement alone.
4.The inter-rater agreement results on the Cohen dataset are solid. DORGA reaches QWK 0.797 with the senior radiologist without any fine-tuning, which is essentially at human inter-rater level (0.790). More importantly, when the two radiologists disagree, DORGA consistently sides with the senior radiologist, and the gap widens as disagreement grows larger, reaching 44.2% versus 9.3% at the largest disagreement level.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper explicitly argues in the related work that SCOL and HCOR fail because they entangle severity and anatomical identity, distorting GNN attention. HCOR is from MICCAI 2025 and targets precisely the same problem. Yet neither method appears in the experimental comparison. This is the method used to evaluate the paper’s central theoretical claim. Without this comparison, the claim that geometric anchor alignment outperforms contrastive ordinal approaches has zero empirical support. The fact that S2-only already matches PAFE makes this omission more serious, because PAFE is a weak baseline and the apparent advantage in Table 1 may not hold against a properly implemented HCOR.
2.The paper motivates DORGA by arguing that each ROI has a different dominant class and therefore a different collapse direction. Yet the severity projector phi and the four anchors are shared across all six ROIs, while region-specific projectors are used only for structure embeddings v_r. The component that varies per ROI (severity distribution) gets shared parameters, while the component that is already ROI-specific by definition (anatomical structure) gets dedicated projectors. The Part A degradation on ROI 3 and ROI 6, the lower-lobe regions with the most atypical grade distributions, is the most likely symptom of this inconsistency, and the paper does not discuss it.
3.DORGA uses ViT-B/16 initialized with MRM pre-trained weights, a domain-specific self-supervised pre-training method for chest X-rays that jointly reconstructs image patches and clinical text. BS-Net, the primary baseline, uses a CNN backbone with no comparable medical domain pre-training. The paper never ablates the contribution of MRM initialization by comparing it against a version of the same ViT-B/16 backbone trained with standard cross-entropy and no geometric components. Without this baseline, it is impossible to know how much of the improvement over BS-Net comes from the stronger backbone initialization versus the DORGA framework itself. This is critical because in recent medical imaging literature, domain-specific pre-training alone routinely produces improvements of comparable magnitude to the gains reported in Table 1.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The two most decisive factors for this score are: First, the theoretical framing combined with the Part A /Part B asymmetry result is genuinely interesting and the paper makes a real scientific point about how ordinal geometry collapse manifests under single-annotator vs. consensus evaluation. This is not a marginal improvement paper and the gap on the consensus test set is large and the framing is coherent. Second, the absence of comparisons against directly relevant ordinal contrastive baselines (SCOL, HCOR) is a real weakness. HCOR is described in the related work as addressing the same problem. Without this comparison, the claim of superiority over the state of the art is not fully established. If the authors can add this comparison in the rebuttal or camera-ready, the score would move to 5.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes DORGA, a geometrically principled framework for multi-region ordinal severity scoring on chest X-rays that explicitly constructs an ordinal-aligned latent space on the hypersphere. The method uses learnable class anchors with polar/ordinal constraints to establish a severity axis, orthogonally decouples severity and structural embeddings to avoid attention bias, and refines predictions via pattern-conditional probabilistic message passing with structure-aware attention and confidence-gated SLERP updates. On the Brixia consensus test set, DORGA achieves an average MAE of 0.301 (improving over BS-Net and PAFE), and without fine-tuning matches senior radiologist agreement (QWK 0.797) on the external Cohen dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Introduces an ordinal-aligned latent geometry with learnable anchors constrained by polar fixing and equal angular spacing. Decouples severity and structure via orthogonal projections, a well-motivated design to avoid severity-driven attention shortcuts in multi-ROI settings. 2.Develops a pattern-conditional prior over inter-ROI grade transitions and uses belief propagation to compute consensus targets, combined with confidence-gated SLERP refinement that preserves hyperspherical geometry. Uses a training-time oracle attention target that supervises structure-driven neighbor selection without leaking severity at inference, aligning with the overall geometric rationale. 3.Evaluates on the Brixia blind test and consensus gold standard, and additionally on the external Cohen dataset without fine-tuning, demonstrating strong generalization and alignment with expert graders. 4.The geometric motivation and the failure mode (ordinal geometry collapse) are clearly articulated with intuitive figures and modular stages. Loss definitions, training pipeline, and inference behavior are mostly well-specified, and the role of stop-gradients and temperatures is explained.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The central “ordinal geometry collapse” hypothesis is compelling but not deeply substantiated beyond conceptual figures and aggregate performance; more direct evidence would strengthen the claim. 2.The oracle attention uses ground-truth labels to build targets during training; while not used at inference, there remains a risk of inadvertently encouraging structure embeddings to correlate with severity unless carefully controlled and measured. 3.Multiple bespoke loss terms and components may make the method sensitive to hyperparameters, robustness analyses are limited. Ablations do not isolate key design choices such as anchor initialization vs. standard angular classification, SLERP vs. simpler logit or vector averaging, or dynamic pattern-conditional prior vs. a global prior. 4.Some notation is dense, and a few equations show minor rendering artifacts, details on the construction, parameterization, and regularization of the transition matrices could be expanded.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses an important medical image analysis problem and proposes a technically interesting framework grounded in a coherent geometric view of ordinal representation learning. The combination of ordinal anchor construction, severity/structure decoupling, and graph-based refinement is well motivated, and the empirical evaluation includes both ablations and external validation. The results on the Brixia consensus test set and the Cohen dataset are promising and suggest that the method may be particularly useful in the presence of noisy or subjective labels. My reservations are mainly about the complexity of the full pipeline, the limited direct evidence for the claimed “ordinal geometry collapse” phenomenon, and the need for stronger comparison to broader baselines. In addition, the paper is somewhat dense and could communicate its key ideas more clearly. Overall, however, I find the paper marginally above the acceptance threshold because the problem is important, the approach is nontrivial, and the reported empirical gains are meaningful.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
1: HCOR / SCOL HCOR and DORGA address ordinal learning from different perspectives. HCOR focuses on ordinal margin learning over a shared embedding, whereas DORGA emphasizes explicit geometry construction and severity/structure disentanglement for multi-ROI reasoning. In DORGA, the severity coordinate system is explicitly constrained on the hypersphere, while inter-ROI attention operates only on structure embeddings. The two approaches are therefore complementary in principle, and HCOR’s ordinal margin could act as an auxiliary loss on top of DORGA. We agree that a direct empirical comparison with HCOR would further strengthen the framework’s positioning. Nevertheless, our central claim concerns explicit geometry construction and decoupled multi-ROI attention, which are orthogonal to HCOR’s shared-embedding ordinal margin formulation.
2: Geometry hypothesis empirical support We acknowledge that the representation-level evidence could have been communicated more directly. Figure 1 was intended as a conceptual illustration rather than a direct measurement. The primary mechanistic evidence is Figure 3 (Same/Other attention ratio 0.559 vs 1.057), showing that severity-structure decoupling substantially alters attention behavior. At the outcome level, DORGA’s Blind→Consensus improvement (Δ=0.125) is substantially larger than BS-Net’s (Δ=0.017), consistent with the geometry-denoising hypothesis. We agree that additional embedding visualizations and angular analyses would further clarify the geometric interpretation, and we will improve this presentation in the camera-ready version.
3: Shared severity projector Brixia grades (0–3) have the same semantic meaning across all ROIs, motivating a shared severity projector, whereas anatomical structure is ROI-dependent and therefore modeled with ROI-specific structure projectors.
4: MRM pretraining: We agree that fully separating backbone and framework contributions would require additional controlled experiments. However, all ablations are performed under the same MRM-pretrained ViT-B/16 backbone, where performance improves from S2-only (0.347) to full DORGA (0.301), indicating an incremental framework contribution under fixed initialization.
5: Oracle attention and severity leakage Oracle attention is used only during training; inference does not use ground-truth labels. To reduce severity leakage, structure embeddings (v_r) are learned under an orthogonality constraint (L_{orth}) with respect to the severity axis (w), and oracle attention is computed only on (v_r). The oracle signal is therefore intended to supervise structural neighbor selection while minimizing the encoding of severity information within the structure embeddings.
6: Dataset statistics Per-ROI grade distributions will be added in the camera-ready to better illustrate severe imbalance and ROI heterogeneity.
7: Loss terms / pattern prior: Each loss targets a distinct failure mode rather than acting as an arbitrary stacked heuristic. Table 1 further suggests non-additive interaction between components, as removing oracle supervision (0.337) degrades performance more than removing decoupling (0.329). For the pattern prior, the dynamic prior produced lower validation MAE than a global prior, and K=7 was selected empirically as the smallest value preserving distinct transition patterns.
8: Readability / Notation In the camera-ready version, Figure 2 will be split into separate panels to improve readability, notational density will be reduced for consistency, and equation rendering/transition-matrix details will be clarified more explicitly.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
All three reviewers have appreciated the methodological innovation and the experimental results presented in the submission and have recommended accepting the work.