acceptodds
Withdrawn from ICLR 2027

FLARE: Selectively Routing Lesion Evidence for Comprehensive Volumetric Radiology Reporting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent multimodal large language models have enabled radiology report generation from full 3D CT volumes, but clinically important focal lesions can remain underrepresented within global volumetric context. A natural remedy is to augment the global input with lesion-focused visual tokens. However, our experiments show that this naive global-plus-local approach does not yield consistent improvements, suggesting that simply providing local evidence does not ensure its effective use. We introduce FLARE, a post-training framework that explicitly controls when local lesion evidence influences report generation. FLARE pairs a frozen global report generator with a crop expert that extracts lesion-specific evidence from local crops. A router decides whether this evidence should influence the global prediction at each decoding step of full-report generation. On the ReXGroundingCT held-out test set, FLARE shows overall improvements across clinically oriented metrics over both full-report supervised fine-tuning and naive local augmentation, while maintaining comparable textual similarity to the global-only baseline. These results support selective lesion-evidence routing as an effective approach for strengthening focal finding descriptions within comprehensive volumetric reports.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
radiology report generation, multimodal large language models, volumetric CT, lesion-aware reasoning, evidence routing, medical image analysis

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 4 reviews, above 27% of the others.
Reviewer_a98hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the underrepresentation of focal lesions in full-report generation from 3D CT volumes. The authors propose FLARE, a post-training framework that separates lesion-evidence extraction from selective evidence routing. A frozen MedGemma 1.5 model, fine-tuned for full-report generation, provides global predictions. A crop expert extracts local evidence through the next-token logit difference between a lesion ROI and a null visual input, while a learned router determines whether this evidence should modify the global prediction at each decoding step. On the 100-volume ReXGroundingCT held-out test set, FLARE with ground-truth ROIs improves Micro-F1 from 37.94% to 46.64% and Macro-F1 from 20.20% to 31.04% over the global SFT baseline, while maintaining broadly comparable textual similarity. Experiments using automatically localized ROIs on an additional cohort of 3,324 CT scans show smaller improvements in several clinically oriented metrics. These results support the potential of the framework for incorporating focal findings into comprehensive reports, although the independent contribution of selective routing and its advantage over alternative local-augmentation methods in the automatic setting remain insufficiently established.

Critical strengths
  1. Consistent improvements in clinically oriented metrics under controlled backbone comparisons. The paper evaluates FLARE against not only off-the-shelf models but also a global SFT baseline, zero-shot crop augmentation, and a trained global-plus-crop baseline using the same model backbone. With ground-truth ROIs, FLARE achieves the best point estimates in Micro-F1, Macro-F1, RadGraph F1, RaTEScore, and RadCliQ, whereas directly adding crops produces mixed results. These comparisons provide meaningful evidence that the complete framework can make better use of local lesion information during full-report generation, which is a central consideration supporting acceptance.

  2. Additional analyses test whether the gains depend on lesion-relevant visual content. In Appendix Table 7, replacing the ground-truth ROI with a region from the same CT that does not contain the target lesion reduces FLARE's Micro-F1 from 46.64% to 37.56% and Macro-F1 from 31.04% to 21.66%. This supports a dependence on relevant visual evidence rather than merely the presence of an additional crop. The automatic localization experiments also demonstrate operation without manual ROIs at inference, while patient-level bootstrap analysis quantifies uncertainty on the small primary test set. Together, these analyses strengthen confidence in the empirical findings.

Critical weaknesses
  1. The independent contribution of selective routing is not sufficiently established, and comparisons with related decoding-time control methods are missing. DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts, ACL 2021, already adjusts base-model predictions using the difference between expert and anti-expert logits. Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding, CVPR 2024, similarly uses differences between predictions under different visual inputs to modify generation. FLARE introduces a lesion-specific crop expert and a learned router, but the relationship to these established decoding strategies warrants a clearer account of its incremental contribution. In particular, the paper does not compare selective routing with adding the same ROI-null evidence at every decoding step using a fixed weight, while holding the crop expert unchanged. The positive-only variant in Appendix A.5.2 simultaneously removes the ROI-null difference and unrelated-ROI negative supervision, so it cannot isolate the router's contribution. The results therefore establish the effectiveness of the complete system more clearly than the necessity of learning when to incorporate evidence, despite this being the central claim of the paper.

  2. The incremental advantage and clinical relevance of the fully automatic setting remain insufficiently validated. The strongest improvements are obtained with ground-truth ROIs on the 100-volume test set. On the larger cohort of 3,324 scans using predicted ROIs, Micro-F1 increases from 35.83% to 38.34%, Macro-F1 from 27.16% to 28.50%, and the LLM evaluation score from 38.11 to 38.47, while RadCliQ slightly worsens. More importantly, this automatic experiment compares FLARE only with the global-only baseline, without evaluating competing local-augmentation methods using the same predicted ROIs. It therefore remains unclear whether the gains in this setting are specific to selective evidence routing. The reported patient-level bootstrap analysis concerns the 100-volume primary test set and does not directly quantify uncertainty in these smaller gains on the larger cohort. Because the paper emphasizes full-report generation without manual lesion annotations at inference, stronger evidence of the method's incremental advantage under this condition would materially affect the assessment of its practical contribution.

Questions for authors
  1. The independent contribution of selective routing. Have the authors compared router-based selective evidence incorporation with adding the same evidence at every decoding step using a fixed weight, while keeping the crop expert, ROI–null logit difference, and other settings unchanged? The positive-only comparison in Appendix A.5.2 removes both the ROI–null difference and unrelated-ROI negative supervision, so it does not isolate the contribution of selective routing. This comparison would help determine whether the gains specifically depend on learning when to use local evidence.

  2. The distinction from existing decoding-time control methods. FLARE’s logit adjustment has a formal connection to DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts and Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding. Could the authors clarify the key distinctions in lesion-evidence construction and routing, and identify which experimental results support an additional benefit from these distinctions? This would help establish the incremental contribution over existing decoding strategies.

  3. The reliability of the gains with automatic localization. In the automatic localization experiment on 3,324 CT scans, have the authors evaluated a Global + Crop baseline using the same predicted ROIs as FLARE? Are patient-level paired confidence intervals available for the differences between methods on this cohort, particularly for the relatively small improvements in Macro-F1 and the LLM evaluation score? These results would help determine whether the improvements are robust and whether they arise from introducing local information or from FLARE’s specific evidence-incorporation mechanism.

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

This paper proposes FLARE, a selective lesion evidence routing framework for 3D CT radiology report generation. The authors find that directly combining local lesion crops with global CT features does not consistently improve report quality. To address this, FLARE introduces a Crop Expert that extracts lesion-related evidence via the logit difference between the ROI and a Null Input, and a Token-level Router that decides when to inject this evidence into the global model’s predictions.

Experiments show that FLARE achieves favorable clinical metrics on ReXGroundingCT and maintains certain performance improvements when using automatically localized ROIs.

Critical strengths
  1. This paper proposes a reasonable solution to the problem that global CT representations tend to overlook local lesions, with clear clinical motivation.

  2. Separating local evidence extraction from Token-level Routing is a flexible and effective design that can leverage local information without modifying the global report generation model.

Critical weaknesses

1. Multiple ROI Handling is unclear

The problem definition allows a CT scan to contain multiple lesion ROIs, but Eq. (3)–(4) only describe evidence extraction and routing for a single ROI. When multiple ROIs are present, the paper does not clearly explain how their evidence is selected or fused, or how potential conflicts are handled when routing for multiple ROIs is activated simultaneously. For example, when a CT contains multiple lesion crops corresponding to different diseases, how are the crops selected? Will the ground-truth connections between the generating sentences and the crops be used?

2. The definition of Clinically Informative Tokens and Negation Handling are unclear

The Binding Loss of the Crop Expert relies on the average log-probability of clinically informative finding tokens, but the paper does not clearly specify how these tokens are selected. In particular, it is unclear whether negation words, disease presence status, anatomical location, and severity information are included in this objective.

In addition, Negative ROI is mainly used to distinguish correct from irrelevant lesion–finding pairs, and is not directly equivalent to clinical Presence/Absence supervision. This is particularly important for report generation. For example, when the prefix already contains an erroneous negative expression (e.g., "There is no pleural ..."), increasing the probability of disease-entity-related tokens may not correct the erroneous clinical statement.

Questions for authors
  1. When a CT contains multiple ROIs, how does FLARE select or fuse the evidence from each ROI? Does inference require any GT finding–ROI correspondence information?

  2. How are clinically informative tokens defined and extracted? Do they include tokens related to Negation, Presence/Absence, Localization, and Severity? If the presence/absence in the prefix is incorrect, what impact does adding the crop have?

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

This paper proposes FLARE, a post-training framework for selectively incorporating lesion-specific evidence into 3D chest CT radiology report generation. FLARE combines a crop expert, which extracts evidence from localized lesion crops, with a routing module that determines at each decoding step whether the local evidence should modify the global report-generation prediction. The authors argue that simply augmenting a global CT representation with local lesion crops may not ensure that the local information is used appropriately. Experiments on ReXGroundingCT compare FLARE with global-only, crop-augmented, and several volumetric medical VLM baselines using clinical report-generation metrics, text-similarity metrics, and LLM-based evaluation. The paper also evaluates automatically localized ROIs and provides ablation studies on the crop expert, routing supervision, and lesion-evidence specificity. The results suggest that FLARE can improve lesion-related clinical metrics while maintaining broadly comparable textual similarity to the global-only baseline.

Critical strengths
  1. The paper addresses a meaningful problem in 3D CT radiology report generation: focal lesions may be underrepresented when the model processes the entire volume. Separating lesion-evidence extraction from evidence routing is a reasonable and practically motivated design, and the qualitative examples support the intuition that local evidence can be incorporated more accurately than with naive global-plus-crop augmentation.

  2. The experimental study includes several useful controlled comparisons and ablations, including global-only generation, crop augmentation, positive-only versus positive-negative ROI supervision, random-ROI replacement, and automatically localized ROIs. These analyses provide evidence that the proposed router is sensitive to lesion-relevant local content rather than merely to the presence of an additional crop.

Critical weaknesses
  1. The methodological novelty is limited. FLARE combines a crop expert with a token-level router that conditionally adds local evidence to the global logits. While this is a reasonable engineering design, the core mechanism appears relatively straightforward and is closely related to existing gated fusion, conditional routing, and logit-level guidance strategies. The paper does not provide sufficiently direct comparisons against these alternatives, making it difficult to establish that the proposed formulation represents a substantial methodological advance.

  2. The empirical validation is not sufficiently strong for the claims made. The main evaluation relies on only 100 held-out CT volumes from 98 patients and, more importantly, uses ground-truth lesion ROIs for the primary FLARE results. This creates a substantial gap between the reported main results and a realistic deployment setting. The automatically localized ROI experiment is useful, but its improvements are more modest and some metrics degrade. Consequently, the robustness and practical effectiveness of the method remain insufficiently demonstrated.

Questions for authors
  1. Clarification: Please clarify the exact role of ground-truth ROIs in the main FLARE results. Were the ROIs available at inference time for every test case, and how many ROIs were provided per volume? Please also clarify whether the automatically localized ROI experiment uses exactly the same test cohort and reference reports as the main evaluation.

  2. Clarification: How were the router threshold τ\tau, evidence scale α\alpha, and router checkpoint selected? Were these values determined exclusively on a validation set before accessing the held-out test set? Please also report the corresponding confidence intervals or paired statistical tests for the improvements over MedGemma1.5-SFT.

  3. Suggestion for strengthening the paper: The novelty claim would be more convincing if the authors compared FLARE against direct alternatives such as continuous evidence gating, hidden-state or attention-level gating, and ungated logit steering using the same crop expert and training data. Please clarify whether such comparisons were considered and how they affect the conclusions.

Reviewer_uV6EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FLARE targets full-report generation from 3D chest CT, where small focal lesions are easily lost in global context. A frozen MedGemma-1.5 report generator is paired with a “crop expert” (3.3M-param adaptor trained on ReXGroundingCT lesion crops with a local LM loss and a margin-based binding loss). At each decoding step, lesion evidence Δ is the logit difference between the crop expert conditioned on the ROI crop and on a null input; a 0.75M-param router, trained with target vs. unrelated ROIs, decides whether to add αΔ to the global logits. A SAM3-based coarse localizer supplies ROIs when masks are unavailable. With GT ROIs, FLARE improves clinical metrics over SFT and naive global+crop baselines on a 100-volume held-out set; with predicted ROIs it improves RadBERT F1 on that set and on 3,324 CT-RATE scans.

Critical strengths
  1. Clinically meaningful problem with a clean decomposition: a crop expert supplies what the lesion contributes, a router decides when it is used, and the global generator stays frozen. The pipeline also runs with automatically predicted ROIs (Table 3), so it is not purely oracle-dependent.

  2. Evaluation is broad: clinical and LLM-based metrics, patient-level bootstrap, ROI-sensitivity controls (Tables 6–7), and token-level analysis (Fig. 10).

Critical weaknesses
  1. Novelty is overstated and the relevant literature is absent. g_t + α(z_ROI − z_null) with a learned token-level gate is classifier-free guidance / proxy-tuning; its visual-region forms already exist (VCD; CRG’s ROI-vs-blacked-out contrast; HALC’s crop/FOV contrast triggered on object tokens; AGLA; MARINE), as do medical versions (ARCD, CAST, CCD, and LEAD’s learned gating of expert signals for RRG). Learned token-level deferral is Co-LLM / FusionRoute. None are cited or compared, so the stated contribution—controlling local evidence at the next-token level—is not positioned against work that does exactly this.

  2. The central claim (“when” matters, not just “what”) is untested, and realistic-setting gains are small and mixed. No always-on baseline (fixed α, no router), no τ/α sweep, no routing-rate statistics (τ = 0.1 is permissive); the positive-only ablation removes two components at once and is itself far below global-only (Micro F1 0.27 vs 0.38). With predicted ROIs, gains concentrate in RadBERT F1 while RadGraph/RaTE/RadCliQ/LLM are flat on the 100-volume set; on 3,324 scans RadCliQ and all text metrics worsen, and LLM-judged faithfulness falls below SFT (9.06 vs 9.65). Paired-bootstrap difference CIs are computed but not reported.

Questions for authors
  1. Add an always-on variant (fixed α, no router), a τ/α sweep, and the fraction of routed tokens on the test set.
  2. Report 95% paired-bootstrap CIs for FLARE − {SFT, GT-crop zero-shot, GT-crop trained} on each clinical metric.
  3. How exactly is “GT crop (trained)” trained—same adaptor, same localized-finding supervision? If not, add a supervision-matched baseline.
  4. Add a two-stage baseline: global report + crop-expert local finding, merged by rule or LLM.
  5. How are multiple ROIs handled at inference (Eq. 3 is single-ROI), and how many crops come from the 85×16×16 localizer map?
  6. Why does FLARE + Localizer lower faithfulness below SFT? Break down Table 3’s LLM score by dimension and report false-positive findings on normal scans.

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

Related papers

Discussion (0)

Sign in to comment.