acceptodds
Under review as a conference paper at ICLR 2027

Causally Guided Chromosomal Abnormality Detection for Unseen Configurations

Abstract

Chromosomal structural abnormality detection informs genetic diagnosis and counseling. However, scarce abnormal samples cannot fully cover naturally occurring configurations across rearrangement locations and intervals, while patient backgrounds and preparation and imaging conditions alter their appearance. Detection thus requires generalization to unseen configurations and robustness to environmental shifts. We observe progressive degradation under these shifts and near-chance performance on unseen configurations. Our causally guided two-stage framework combines homologous comparison with normal-statistics calibration, using abundant normal pairs as a baseline to recast normal-versus-abnormal classification as distinguishing normal from abnormal variation. This reduces reliance on chromosome-specific appearance and simplifies learning from scarce abnormal samples. Stage I combines skeleton-guided counterfactual generation, configuration generalization learning, and synthetic bias suppression to expand configuration coverage and learn transferable structural evidence. Stage II freezes the encoder and calibration statistics, refining the classification head through quality and reference constraints for environmental robustness. We introduce ChromoShift, a full-pipeline dataset comprising skeleton-centromere annotations, normal pretraining data, and real abnormality data under distribution shifts. Experiments show gains in both test domains, achieving 90.83% AUROC on unseen configurations. The framework supports clinical assessment of known abnormalities and screening for potentially novel configurations to guide further genetic research. Code: https://anonymous.4open.science/r/CausalChrom-3D66/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Causal Learning, Out-of-Distribution Generalization, Medical Image Analysis, Chromosomal Structural Abnormality

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_fUQ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper identifies the sharp performance degradation of deep learning models in detecting chromosomal structural abnormalities under environmental shifts and unseen rearrangement configurations. To address this challenge, the authors introduce a two-stage framework based on homologous pair comparison and group-wise normal-statistics calibration, recasting abnormality detection as distinguishing abnormal structural variation from normal baseline variation. Evaluated on their introduced dataset ChromoShift, the method achieves 90.83% AUROC on unseen configurations, outperforming existing baselines.

Critical strengths

1.This paper introduces the ChromoShift dataset and provides comprehensive ablation studies covering backbones, pair formulations, training dynamics, and generator types.

  1. The visibility-based sample filtering, real–artificial alignment LRAL_{RA} and sham controls LSL_{S} help prevent shortcut learning caused by synthetic artifacts.
Critical weaknesses
  1. The framework includes an upstream ResUNet, counterfactual generation, two distinct training stages, and numerous regularization hyperparameters (λB,λRA,λS,λQ,λP1,λP2\lambda_B, \lambda_{RA}, \lambda_S, \lambda_Q, \lambda_{P1}, \lambda_{P2}), making implementation and hyperparameter tuning complex.

  2. Lack of formal causal theoretical grounding: Although the authors show a causal diagram with a domain indicator E (in Figure 2), the framework lacks comprehensive formal causal proofs. There are no proofs or analyses regarding causal effect identifiability or do-calculus operations. In practice, the method relies on standard consistency regularization and stratified feature standardization. Labeling this framework as "Causally Guided" may be superficial conceptual inflation. From the overview of the proposed framework, I cannot find how the causal modeling works. The author should explain it.

  3. Limited Robustness Analysis: The paper lacks a systematic analysis of robustness to individual factors in isolation.

minors:

Using the possessive suffix 's after a parenthetical year is non-standard in academic writing (Lines 114, 116, 118, 140), Examples include "Li et al. (2024)'s HomNet" (Line 114), "Chen et al. (2024)'s BR-ChromNet" (Line 116). Suggested Revision like: "HomNet (Li et al., 2024)", "BR-ChromNet (Chen et al., 2024)".

Questions for authors
  1. The method uses the absolute difference ∣q−r∣|q-r| to represent feature differences between the target and reference chromosomes, then concatenates it with the target features to form [q,∣q−r∣][q, |q-r|] and resolve the symmetry ambiguity. Why not use the signed difference q−rq-r directly to avoid this ambiguity? The paper also evaluates alternatives such as direct concatenation [q,r][q,r], but the performance gains are limited. Could the authors provide a theoretical explanation for these findings?

  2. Are all training samples used in Stage 2 artificially perturbed versions of real samples? Does the test set include any real samples with perturbations?

  3. Table 2 shows substantial performance variability across folds. What accounts for this variability? Also, the sensitivity and specificity results in Table 2 do not differ significantly from those of the second-best method. Could the authors explain this observation?

  4. Why are the loss terms optimized in two stages rather than combined into a single objective for joint training?

5 comments
AuthorsOpenReview ↗

Rebuttal for Weaknesses (1/2)

Thank you for your constructive feedback. The revised PDF is uploaded, with additions and revisions marked in blue.


W1: Implementation Complexity and Hyperparameter Sensitivity

A: Our pipeline covers skeleton extraction, inversion generation, and abnormality detection; the first two are independently reusable or replaceable. The skeleton model is frozen downstream and can be replaced by conventional methods. Default geometric inversions are generated offline without training another generator; the appendix compares GAN and flow-matching alternatives. Stage I learns inversion structure; Stage II adjusts only classification-head weights for environmental robustness. These roles make the workflow clear.

To assess tuning sensitivity, we vary one loss weight at a time over 0.25–4 times its reference value, keeping other settings unchanged. Stage I examines B, RA, and S: across both test domains, the maximum absolute AUROC/AUPRC changes from the reference setting are 0.42/1.38 percentage points (pp). Stage II examines P1, P2, and Q; the corresponding changes are 0.18/0.35 pp. Performance is insensitive to fine weight adjustment within the tested ranges.


W2: How Causal Modeling Guides Method Design

A: “Causally Guided” means using causal analysis to understand how inversion configuration B, patient background P, and preparation/imaging conditions Q affect observations and predictions, thereby guiding sample construction and constraints: what to change, preserve, and constrain. Section 4.1 and Figure 2 model structure formation as M = f_M(A,B,P,U_M) and image acquisition as X = f_X(M,Q,U_X), where A denotes inversion status. This follows medical-imaging analyses of distribution shifts through the data-generating process [1].

Artificial inversions change local band order to supervise structural recognition. Sham controls preserve band order while introducing artifacts from the same processing pipeline, discouraging artifact dependence. For label-preserving comparisons, we perturb quality or replace the reference while fixing the target, constraining prediction variation. This parallels using semantic-preserving counterfactual image pairs for imaging robustness [2]. NSC uses normal homologous differences as a baseline for assessing departures from normal variation.

Regarding the identifiability and do-calculus issues you raised, our theoretical objective is prediction stability under these controlled comparisons. Appendix A.2.1 defines ideal structural interventions and relates target risk to synthetic-sample risk, configuration coverage error, and generation error. Equations (8)–(9) and Appendix A.2.2 show that, for label-preserving comparisons with mean squared logit change J, mean absolute probability change and the change in mean cross-entropy risk are bounded by √J/4 and √J, respectively. The appendix gives coverage conditions for unseen changes; A.3 analyzes NSC under location–scale assumptions. These analyses provide the rationale and conditions under which the designs apply; the main-text ablations demonstrate their detection gains.

[1] Castro, D. C., Walker, I., and Glocker, B. Causality matters in medical imaging. Nature Communications, 11:3673, 2020.

[2] Roschewitz, M., et al. Robust image representations with counterfactual contrastive learning. Medical Image Analysis, 105:103668, 2025.


W3: Single-Factor Robustness Analysis

A: Cross-environment testing evaluates new patients with seen configurations and natural acquisition differences: preparation and G-banding changes can affect band appearance even within one center [3]. Unseen-configuration testing additionally excludes test configurations from that fold's real training and validation data. Appendix E.7 adds controlled perturbations to isolate imaging factors.

We fix patients, configurations, and sample composition, separately applying brightness changes, contrast changes, blur, and Gaussian noise. Absolute AUROC/AUPRC changes from original images are averaged within each perturbation type, then equally across the four types. The table reports deep learning baselines, Stage I-only VA6, and Full. Values are estimates [patient-cluster 95% CI], in pp. Table E.11 gives all results.

MethodMean absolute AUROC change ↓Mean absolute AUPRC change ↓
HomNet9.768 [6.696,12.730]2.105 [1.081,13.927]
DCF3.652 [1.946,8.215]0.342 [0.106,1.010]
CHROMA1.918 [0.098,6.795]1.109 [0.249,2.067]
VA6 (Stage I)1.332 [0.414,4.989]0.822 [0.123,2.911]
Full0.544 [0.032,2.220]0.537 [0.367,0.855]

Full reduces mean absolute AUROC/AUPRC changes by 0.788/0.285 pp relative to VA6, supporting Stage II's role in reducing imaging sensitivity with patients and configurations fixed.

[3] Henegariu, O., et al. Improvements in cytogenetic slide preparation: Controlled chromosome spreading, chemical aging and gradual denaturing. Cytometry, 43(2):101–109, 2001.

AuthorsOpenReview ↗

Rebuttal for Questions (2/2)

Thank you for your questions.

Q1: Homologous Pair Representation

A: 1. Retaining q. q addresses absolute-difference symmetry and preserves the assessed member's state. Signed differences distinguish input order but lose shared features. For a feature component a added to both,

(q + a) − (r + a) = q − r.

Different pairs can yield identical differences: q − r and |q − r| alone cannot preserve the full target state, nor can swapping inputs restore it. Bidirectional [q, |q − r|] and [r, |q − r|] let us assess each member using its own features and homologous differences.

2. Absolute differences and NSC. With q retained, we choose absolute differences to work with NSC. For N equally weighted normal homologous pairs in one group, dᵢ is pair i's feature difference. Counting both directions equally gives

μ_signed = Σᵢ[dᵢ + (−dᵢ)]/(2N) = 0; μ_abs = Σᵢ|dᵢ|/N.

Signed means cancel; μ_abs captures typical normal difference magnitudes. NSC calibrates dimensionwise: z = (|q − r| − μ_abs)/σ_abs, with σ_abs the normal scale in Eq. (5). Signed differences allow scale calibration, but mean centering cannot provide the same magnitude baseline.

[q, q − r] and [q, r] are linearly interconvertible, hence our original omission of the former. Before calibration:

  • [q, r]: η = uᵀq + vᵀr + b.
  • [q, q − r]: η = uᵀq + vᵀ(q − r) + b = (u + v)ᵀq − vᵀr + b.
  • [q, |q − r|]: η = uᵀq + Σⱼvⱼ|qⱼ − rⱼ| + b.

The first two have equal linear expressivity. Absolute differences add piecewise-linear magnitude features per dimension. With NSC, the head directly uses departures from normal differences, easing this comparison with limited abnormal samples.

3. Ablations (Table E.3). Adding q to q − r improves all four metrics. Versus [q, q − r], our representation improves cross-environment/unseen-configuration AUPRC by 13.42/0.89 pp. Versus [q, r], unseen-configuration AUROC is similar (90.83% vs 90.94%); cross-environment AUPRC differs most (56.10% vs 39.24%).


Q2: Stage II Samples and Testing

A: Stage II classification uses original real training samples. Q compares predictions before/after quality perturbations; P compares predictions before/after reference replacement. It uses real samples and derived pairs, without artificial inversions.

Main tests use uniformly preprocessed real clinical images, retaining natural patient/preparation/imaging differences without added perturbations. Appendix E.7 separately reports controlled test-image perturbations to isolate imaging effects.


Q3: Fold Variation, Sensitivity and Specificity

A: 1. Fold variation. Pericentric inversions may change relative centromere position and arm ratios; paracentric inversions mainly change local band order. Short intervals or similar bands offer fewer recognition cues. Tables B.2–B.3 list 2–3 patients per cross-environment fold and 3–4 per unseen-configuration fold. Each fold has 2.17% positive images, but patient/configuration composition differs, so a few difficult cases can affect fold metrics. Figure 5 reports between-method differences and paired patient-cluster bootstrap 95% CIs using pooled five-fold OOF predictions.

2. Sensitivity and specificity. With low positive prevalence, interpret both fixed-threshold metrics together. In Table 2's unseen-configuration results, our sensitivity is similar to DCF's (45.44% vs 45.08%), with higher specificity (98.77% vs 55.44%). Versus CHROMA, specificity is slightly lower (98.77% vs 99.58%) and sensitivity higher (45.44% vs 21.77%). Means/SDs alone cannot establish paired significance. Figure 5 compares Sens@95%Spec under the same false-positive constraint, reporting paired differences and patient-cluster 95% CIs. Ours achieves 63.22% versus CHROMA's 31.80%: +31.42 pp [19.43, 43.75].


Q4: Why Two-stage Optimization?

A: In §4.5, configuration learning must preserve subtle abnormal structures, while P/Q seeks to reduce sensitivity to patient/imaging variation. These goals may interfere in joint encoder updates. Stage I learns structural representations and updates normal statistics; Stage II freezes the encoder, statistics and head bias, adjusting only classification weights so environmental constraints act on how learned structural evidence is used.

Table E.4 reports five-fold mean ± sample SD (%); CE/UC denote cross-environment/unseen-configuration.

StrategyCE AUROCCE AUPRCUC AUROCUC AUPRC
Single-stage89.65±3.8337.24±16.7288.65±5.2736.72±16.84
Two-stage93.41±5.3656.10±21.3390.83±5.7446.77±25.32

Higher means for all four metrics support structure learning before environmental adjustment. Table E.6 shows that adding P/Q improves cross-environment AUROC/AUPRC by 3.36/2.46 pp over classification-only Stage II, beyond additional classification training.


Minor Comments

A: We corrected the citation formats and checked similar usage throughout.

Thank you again for your constructive feedback.

AuthorsOpenReview ↗

Sincere Thanks and a Brief Follow-up on Our Response

Dear Reviewer fUQ1,

We hope your research is going well. Thank you for the careful thought you gave to the reasoning behind our method. Your questions about how causal analysis guides our design choices and why we train in two stages helped us explain more clearly how the theory informs those choices and how the experiments test them.

As the discussion draws to a close, we wanted to follow up on our earlier response. If you have a moment, we would welcome your thoughts on whether the explanations and experiments clarify the reasoning behind these choices.

Thank you again for your careful review and constructive suggestions.

With best wishes,
The Authors

Reviewer_cTT5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a causally guided two-stage framework for chromosomal structural abnormality detection, focusing on inversions. By combining homologous comparison, normal-statistics calibration, skeleton-guided counterfactual generation, and environment-invariant constraints, the method targets generalization to unseen inversion configurations and robustness under acquisition shifts. The authors also release ChromoShift, a dataset with skeleton/centromere annotations, normal pretraining data, and 25 real inversion configurations from 84 patients. The reported 90.83% AUROC on unseen configurations is a clear improvement over strong baselines.

Critical strengths
  1. Clear diagnosis of the real failure mode: near-chance on unseen configurations, not just class imbalance.
  2. Normal-statistics calibration is conceptually simple and empirically the biggest lever (82→90 AUROC).
  3. Homologous pairing + bidirectional labels exploit cytogenetic structure instead of treating images as i.i.d.
  4. Dataset design (patient-disjoint + config-unseen) should become a benchmark for distribution-shift cytogenetics.
Critical weaknesses
  1. Ablations show NSC + CCRL + constraints all help, but there is no run that removes the “causal” inductive bias while keeping the same data. Is the gain from causality, or from (a) homologous pairs, (b) massive normal pretraining, (c) counterfactual augmentation?
  2. AUROC/AUPRC are reported, but clinical deployment cares about sensitivity at fixed specificity (e.g., 95% spec). F1 at 0.5 threshold is unstable with 25 configs. Per-config confidence intervals are missing.
  3. Male X/Y use standard ideograms; ideogram is drawn from hg38. Need to clarify ideogram statistics are excluded from normal-baseline leakage and that Stage II doesn’t overfit to ideogram style.
Questions for authors
  1. If you replace the causal calibration with a non-causal pairwise contrastive head (same encoder, same pairs), how much does unseen-config AUROC drop?
  2. Can you report sensitivity/specificity at 95% specificity and per-config 95% CI, not just macro AUROC?
  3. How do you validate that synthetic inversions are cytogenetically plausible? Any expert rating or band-correlation metric vs real ones?
  4. In Stage II, freezing stats from Stage I—how sensitive is the result to encoder drift if Stage I trains 1 more epoch?
  5. For clinical use, what is the false-referral rate when used as a pre-screen before manual karyotyping?
1 comment
AuthorsOpenReview ↗

Sincere Thanks and a Brief Follow-up on Our Response

Dear Reviewer cTT5,

We hope your research is going well. Thank you for bringing a clinical perspective to our evaluation. Your questions about differences across inversion configurations and your suggestion to report sensitivity at 95% specificity helped us focus on what matters to clinical readers. Your attention to uncertainty also helped us make the analysis more rigorous.

With the discussion period drawing to a close, we would be grateful if you could take a look at our point-by-point response when you have a moment. We would especially value your perspective on whether the additional results address the evaluation questions you raised.

Thank you again for your thoughtful review and helpful suggestions.

With best wishes,
The Authors

Reviewer_Xff1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies chromosome inversion detection across unseen patients and configurations absent from real training data. Stage I learns structural representations through homologous comparison, normal-statistics calibration, synthetic inversions, and synthetic bias suppression. Stage II freezes the encoder and statistics and refines the classifier using imaging perturbations and reference-substitution constraints.

The authors introduce ChromoShift, whose real-abnormality component includes 84 patients and 25 inversion configurations. Under patient-disjoint five-fold evaluation, the method achieves an AUROC/AUPRC of 90.83±5.74%/46.77±25.32% on unseen configurations, compared with 75.04±10.44%/25.72±21.36% for CHROMA, the strongest comparator in this domain in the main table. Its cross-environment AUROC is 93.41±5.36%.

Critical strengths

This paper studies chromosome inversion detection across unseen patients and configurations absent from real training data. Stage I learns structural representations through homologous comparison, normal-statistics calibration, synthetic inversions, and synthetic bias suppression. Stage II freezes the encoder and statistics and refines the classifier using imaging perturbations and reference-substitution constraints.

The authors introduce ChromoShift, whose real-abnormality component includes 84 patients and 25 inversion configurations. Under patient-disjoint five-fold evaluation, the method achieves an AUROC/AUPRC of 90.83±5.74%/46.77±25.32% on unseen configurations, compared with 75.04±10.44%/25.72±21.36% for CHROMA, the strongest comparator in this domain in the main table. Its cross-environment AUROC is 93.41±5.36%.

Critical weaknesses
  1. The stability of the gains and the incremental contribution of the full framework remain unclear. Unseen-configuration AUPRC is 46.77±25.32% and sensitivity is 45.44±22.36%, with considerable variation across folds. Assessing stability requires per-fold patient and configuration composition, positive prevalence, and paired differences between methods.

Removing CCRL while retaining NSC and Stage II yields an AUROC/AUPRC of 87.69%/42.30%, corresponding to mean gains of 3.14/4.47 percentage points for the full method. Both the groupwise-centering experiment and the complete Stage I retain other training components, leaving a simpler model using only homologous comparison, NSC, and conventional supervision untested. The gains from difficult-group supervision, real–artificial alignment, and sham controls are also not isolated.

  1. Validation of within-chromosome-type configuration transfer and imaging robustness remains limited. In Table E.6, adding another chr9 inversion configuration does not improve ResNet50 recognition, but only one configuration pair and three random seeds are tested. In Table E.7, the full method’s result with same-type abnormal training support comes from one configuration in Fold 1; the subset without such support accounts for 96.74% of abnormal test chromosome instances.

Synthetic coverage of test configurations is not quantified, making it difficult to distinguish coverage from transfer to uncovered structures. Cross-environment evaluation combines patient and natural acquisition variation, without independently validating Stage II imaging robustness.

Questions for authors
  1. Analysis of existing results: Please provide the patient counts, configuration counts, positive and negative chromosome-instance counts, positive prevalence, and paired differences between the full method and major ablations for each fold. Could uncertainty be assessed at the patient or configuration level while accounting for within-patient dependence?

  2. Additional ablations: With the same encoder and pretraining, how does a simpler model retaining only homologous comparison, NSC, and conventional classification training perform? What are the individual gains from difficult-configuration-group supervision, real–artificial alignment, and sham controls?

  3. Analysis of existing results: Please characterize the proximity of synthetic configurations to real test breakpoints and provide patient counts, positive and negative instance counts, and uncertainty estimates for Table E.7 and Figure 9. Can existing data support further within-chromosome-type configuration-transfer evaluation of the full method?

  4. Mechanism validation: With cases held fixed, does Stage II improve robustness to controlled imaging perturbations?

1 comment
AuthorsOpenReview ↗

Sincere Thanks and a Brief Follow-up on Our Response

Dear Reviewer Xff1,

We hope your research is going well. Thank you for reviewing our experimental design so carefully. Your suggestions helped us design more rigorous experiments to assess the stability of our gains, examine generalization across inversion configurations on the same chromosome, and measure how well the synthetic data cover the test configurations.

As the author–reviewer discussion period draws to a close, we wanted to follow up on our point-by-point response. If you have a moment to read it, we would welcome your thoughts on whether the additional experiments and analyses address your concerns.

Thank you again for the time and care you have devoted to reviewing our work.

With best wishes,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Area Chair and Reviewers,

Thank you for the time and care you have devoted to reviewing and discussing our paper. Your feedback has helped us strengthen the work and clarify its presentation. We have uploaded the revised manuscript, with additions and revisions highlighted in blue. For ease of reference, we summarize the main changes below.

  • Fold-to-fold variation and performance differences

    • Tables B.2–B.3 report the inversion configurations, patient counts, numbers of normal and abnormal images, and positive prevalence in each fold. (Xff1 Q1, fUQ1 Q3)
    • Figure 5 reports paired differences between our method and each comparator in AUROC, AUPRC, and sensitivity at 95% specificity, with 95% confidence intervals from a patient-cluster bootstrap. (Xff1 W1, cTT5 Q2)
  • Component contributions and homologous pair representations

    • Table 3 adds the requested baseline using homologous pairs, normal-statistics calibration (NSC), and standard classification supervision. Table E.5 examines the individual contributions of difficult-configuration weighting, real–artificial alignment, and the sham constraint. (Xff1 Q2)
    • Table E.2 evaluates a conventional pairwise contrastive head as a replacement for NSC. (cTT5 Q1)
    • Table E.3 adds signed-difference controls to the comparison of pair representations. (fUQ1 Q1)
  • Cross-configuration generalization within the same chromosome type (Xff1 W2, Xff1 Q3)

    • Table E.8 describes the six additional test-only cases, each with a distinct new inversion configuration, and lists their fold assignments.
    • Table E.9 extends the controlled comparison to seven test configurations, with five random seeds per configuration. Table E.10 reports the full method’s performance grouped by whether training included a different real inversion on the same chromosome type.
  • Structural plausibility and number of synthetic inversions

    • Figures 11–12 compare normal homologs, synthetic inversions, and real inversions. The images and aligned band profiles show how local reversal changes the band pattern between the breakpoints. (cTT5 Q3)
    • Figure 4 shows how the size of the synthetic inversion pool affects performance in both test domains.
  • Two-stage training and robustness to imaging changes

    • Table E.4 compares joint and two-stage training. Table E.6 compares classification-only fine-tuning in Stage II with fine-tuning that also includes the environmental constraints. (fUQ1 Q4)
    • Table E.11 compares the Stage I model with the full method under four controlled imaging perturbations. We vary brightness, contrast, blur, and Gaussian noise separately while holding patients and inversion configurations fixed. (Xff1 Q4, fUQ1 W3)

Thank you again for your thoughtful review and constructive suggestions. We hope this guide makes the revisions easy to locate, and we would be happy to clarify any remaining questions.

Warm regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.