acceptodds
Under review as a conference paper at ICLR 2027

Falsifiable Claim–Region Grounding in 3D CT

Abstract

3D CT grounding usually evaluates where a supported finding is located, while verification evaluates whether a claim matches a scan. These separate tests miss a failure in which a real finding is displayed as support for a false claim. If a CT shows a finding in the left lower lobe but the claim places it in the left upper lobe, the finding is real yet does not support that claim. Positive-only grounding may still return its region, whereas a global verifier may reject the sentence without identifying the conflicting region. This paper introduces ReX-Verify, which keeps the CT and annotated source region fixed as location and finding edits change that region from support to annotation-relative counter evidence. The accompanying Claim-conditioned Evidence Grounding for Verification (CEG-Verify) model predicts separate support and counter fields, suppresses support for non-supported claims, and uses the predicted relation to select the returned field. Matched comparisons show that false-support supervision reduces exposed support after relation errors, mainly for cross-case unsupported claims, while also reducing returned maps for supported claims. Outcome decomposition attributes the reduction to map blocking after false admission. Exact-text and region analyses show weak image-conditioned discrimination and coarse counter localization.

Primary area:
interpretability and explainable AI
Keywords:
3D CT, claim-region grounding, claim verification, counter evidence, false-support exposure

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FKckOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates falsifiable claim–region grounding in 3D CT and introduces the ReX-Verify benchmark and CEG-Verify framework. The benchmark constructs supported and unsupported medical claims through location modifications, finding-type substitutions, and cross-case pairing. The proposed framework combines image–text feature encoding, separate support and counter-evidence prediction heads, relation classification, and false-support suppression to reduce misleading spatial evidence.

Experimental results show that false-support supervision reduces the false-support rate (FSR) from 0.130 to 0.032. However, the model achieves very low support and counter-evidence localization IoUs (0.010 and 0.018, respectively), and most improvements arise from cross-case unsupported claims rather than the core fixed-region claim edits.

Critical strengths

1. Interesting Problem Formulation.
The paper identifies a meaningful limitation of existing medical visual grounding methods: a correctly localized finding may still be incorrectly presented as evidence supporting a false medical claim. By distinguishing supporting and counter-evidence regions, the work introduces an interesting perspective that extends beyond conventional positive-only grounding and global claim verification.

2. Systematic Evaluation and Error Analysis.
The paper introduces the ReX-Verify benchmark and evaluates false-support exposure through controlled claim modifications. The ablation studies and error decomposition provide useful insights into the effect of false-support supervision, particularly the reduction in FSR from 0.130 to 0.032. Although the overall effectiveness remains limited, these analyses help clarify the proposed mechanism and provide a useful foundation for future research.

Critical weaknesses

1. Questionable Dataset Validity and Potential Textual Shortcuts. The automatically constructed negative claims may not necessarily be medically false, as modified findings or locations could still correspond to genuine abnormalities elsewhere in the CT. Moreover, reusing the original lesion masks as counter-evidence does not guarantee that these regions actually falsify the modified claims. The strong text-only performance (AUC = 0.939 for finding edits) further suggests potential linguistic shortcuts. These issues undermine the reliability of the benchmark and raise concerns about whether it genuinely evaluates image-grounded medical claim verification.

2. Limited Technical Contribution and Insufficient Experimental Validation. The proposed framework mainly combines existing image–text encoders, dual prediction heads, and suppression losses, without demonstrating clear advantages over simpler alternatives. Although false-support supervision reduces FSR, it also substantially decreases the exposure of valid support maps (from 0.340 to 0.124). Furthermore, the extremely low support and counter-evidence IoUs (0.010 and 0.018) and weak image-conditioned discrimination (AUC = 0.561) indicate that the model has not convincingly achieved reliable evidence grounding. These limitations significantly weaken the claimed methodological contribution and constitute a major reason for my rejection recommendation.

Questions for authors

How do the authors ensure that automatically generated negative claims are genuinely unsupported by the corresponding CT scans? For example, a modified location or finding type may still correspond to another existing abnormality. Were these samples independently validated by radiologists, and how was potential label noise addressed?

What is the justification for directly reusing the original lesion mask as counter-evidence for modified claims? Since the presence of the original finding does not necessarily falsify the modified claim, how do the authors ensure that these masks represent genuine counter-evidence rather than merely regions inconsistent with the edited description?

What specific advantages does the proposed dual-head architecture provide over simpler single-head alternatives? Given that CEG-Verify does not consistently outperform Single+gate or independent dual-head models, can the authors provide stronger evidence that the proposed design offers meaningful benefits beyond output suppression?

How can the authors demonstrate that the reduction in FSR reflects improved image-grounded verification rather than simply suppressing mask outputs? In particular, can the authors explain the substantial decrease in valid support-map exposure (from 0.340 to 0.124), the extremely low localization IoUs, and the weak image-conditioned discrimination (AUC = 0.561)? Would the improvement remain under matched support-map coverage?

Reviewer_gH6SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReX-Verify tests whether a fixed CT region changes from support to counter evidence after a claim edit. CEG-Verify learns separate support and counter fields with relation-controlled output and support suppression. Matched experiments show reduced false-support exposure, largely through blocking maps rather than improving relation decisions or precise localization.

Critical strengths
  • The distinction between relation rejection, blocked false admission, and exposed spatial support identifies a meaningful failure missed by positive-only localization.
  • Matched three-seed controls, exact-text tests, and explicit reporting of coverage loss and weak localization provide a candid, useful diagnosis of the method's behavior.
Critical weaknesses
  • The target relation is annotation-relative, yet inference receives only the CT and edited claim, with the source region hidden (Section 3.2). If another region satisfies the edited claim, the supplied inputs do not identify which source annotation determines its label. No new radiologist adjudication was collected. An ambiguity audit or an explicit region-conditioned formulation is needed to establish target validity.
  • The practical grounding advance is weak: support suppression lowers valid-map exposure from 0.340 to 0.124 (Table 3), exact-text discrimination is near chance (AUROC 0.561), and counter IoU is only 0.018 overall. The strongest architectural comparison is inconclusive. Lower false-support exposure alone therefore does not establish more useful evidence grounding.
Questions for authors
  1. How are edited or cross-case claims that remain true elsewhere in the CT identified and handled?
  2. Do existing threshold sweeps show an advantage at matched supported-map exposure, rather than matched relation recall alone?
Reviewer_yv8wOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a method for detecting the failures between 3D CT grounding and claim verification.

Critical strengths
  1. The problem formulation is clear, and relevant to the real clinical settings.
  2. The text-only control, the identical-string scan-swap control, and the phrase-matched structured-edit control are good diagnostics of shortcut learning.
Critical weaknesses
  1. If I understand correctly, the spatial ground barely works. The support and counter IoU are close to zero according to the experiment results. Given this result, can the proposed method still called a grounding method?

  2. The verifications mostly rely on text, and the labels did not go through clinical validation.

Questions for authors
  1. Can you provide curves of FSR against supported-map exposure and supported-map IoU over the map threshold, for the full model, the model without the suppression loss, and the single-head gated baseline?
  2. How were location and finding edits generated, and what fraction of edited claims are actually true elsewhere in the same CT?
  3. What is localization performance at the lobe or organ level, whether the peak voxel falls inside the source mask?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.