acceptodds
Under review as a conference paper at ICLR 2027

BEAR-Med: From a Clinically Grounded Benchmark to Expert-Aligned Reasoning in Medical Image Forensics

Abstract

Clinical image forensics extends beyond binary detection: systems must distinguish local edits from global synthesis, localize affected regions, and justify verdicts with observable evidence. Existing benchmarks remain limited in scale, imaging diversity, forgery coverage, clinical plausibility, or spatial fidelity, while current methods often separate detection from localization, append post-hoc explanations, or derive chain-of-thought supervision from general-purpose multimodal models without explicit forensic evidence. We introduce BEAR-Med-1M, a million-image benchmark spanning four clinical imaging categories and five local-to-global mechanisms: copy-move, splicing, lesion insertion, lesion removal, and full-image synthesis. It targets medically meaningful, anatomically plausible regions and uses bidirectional restoration for lesion insertion and removal, producing spatially faithful local forgeries with low out-of-box spillover. We also introduce BEAR-Med, an expert-aligned model whose four frozen experts capture local traces, image-wide consistency, generation texture, and medical normality. Their evidence is distilled offline into expert-guided CoT targets, while inference uses only the original image. BEAR-Med generates a pre-hoc analyze-localize-decide sequence, placing evidence-based reasoning before localization and verdict rather than explaining predictions post hoc. A segmented objective separately balances reasoning, box, and verdict tokens, followed by GRPO with verifiable task and spatial rewards. BEAR-Med leads dense-localization and multimodal reasoning baselines in accuracy, F1, all reported localization metrics, and authentic-image error control, while retaining strong cross-dataset detection and localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
clinical image forensics, medical image forgery benchmark, forgery localization, multimodal large language models, expert-aligned reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_n7foOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents BEAR-Med, a benchmark for medical image forgery detection, localization, and evidence-based reasoning. The authors construct a dataset of one million images covering four medical imaging categories and five forgery types. They train a Qwen3-VL model using four frozen evidence experts, staged supervised fine-tuning, and GRPO-based optimization. While the model achieves strong in-distribution results, its localization performance drops substantially on the external MedForge-90K benchmark.

Critical strengths

1.The paper addresses a useful problem by combining forgery detection, localization, and evidence-based explanations in one framework. 2.BEAR-Med-1M is a substantial dataset covering multiple imaging modalities and forgery mechanisms. The authors also investigate whether editing affects regions outside the annotated boxes. 3.The experiments include comparisons with existing methods, ablations, robustness tests, and external evaluation. The failure analysis is also helpful. 4.The paper clearly states its limitations and does not claim that the results establish reliable deployment in real clinical settings.

Critical weaknesses

I have the following concerns: 1.The near-perfect in-distribution results may partly reflect artifacts specific to the forgery generation pipeline. The paper does not sufficiently establish generalization to unseen generators or editing methods.

2.External localization performance is weak. On MedForge-90K, Box [email protected] is only 0.083, and it remains low on the subset with contained edits. Annotation issues explain part of the gap but do not fully establish the model's generalization ability.

3.The reasoning evaluation relies mainly on LLM judges, without sufficient validation by medical experts. Fluent explanations do not necessarily reflect correct visual evidence, as some failure cases in the paper also demonstrate.

4.The contribution of expert evidence is not fully separated from privileged information used to generate training targets, including ground-truth labels and bounding boxes. More controlled ablations are needed.

5.The GRPO reward mainly encourages appropriate output length, format, correct verdicts, and box overlap. It does not directly measure the medical correctness or visual grounding of the reasoning, making the claims about reasoning improvement less convincing.

Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer n7fo (1/2)

Thank you for recognizing the value of joint detection, localization, and reasoning, the benchmark's coverage, and the spatial audits. We address your five concerns with unseen-generator expert evaluations, physician assessment of rationale quality, and reasoning-prefix interventions, alongside clarifications of the external localization results and the respective roles of expert-guided supervision and G-GRPO.

W1. "The near-perfect in-distribution results may partly reflect artifacts specific to the forgery generation pipeline. The paper does not sufficiently establish generalization to unseen generators or editing methods."

We additionally report leave-one-generator-out evaluation of the three trainable experts. Each fold excludes FLUX, SD3.5, or SDXL from both training and validation, retrains the experts, and selects checkpoints using the remaining generators. Testing covers the held-out generator's INS/REM/SYN images across all four imaging categories, against the same 100,008 authentic test images.

Table R1. Expert detection AUC on unseen generators. Mean is the equal-weight average across the three folds.

Held-out generatorLocal traceImage consistencyGeneration texture
FLUX.8404.9536.9975
SD3.5.9193.9589.9987
SDXL.8856.9536.9976
Mean.8817.9553.9979

Table R1 supports transferable expert detection cues, particularly from image consistency and generation texture, beyond editors encountered during training.

W2. "External localization performance is weak. On MedForge-90K, Box [email protected] is only 0.083, and it remains low on the subset with contained edits. Annotation issues explain part of the gap but do not fully establish the model's generalization ability."

Strict localization remains difficult even with target-domain supervision. In paper Table 3, the official MedForge model's [email protected] increases from .116 on ALL BOX to .177 on CONTAINED. This improvement is consistent with our independent spatial-fidelity audit identifying meaningful edits outside the released boxes. Nevertheless, .177 remains low despite MedForge training supervision. The remaining difficulty can reflect both precise-localization challenges and residual uncertainty in the reference annotations.

Besides, CONTAINED screens for outside-box changes and it does not verify the accuracy of the box itself. Because the audit hides the GT rectangle, it cannot establish whether that rectangle tightly and correctly delineates the realized edit. Moreover, in our existing human assessment, the two reviewers still identified meaningful outside-box changes in an average of 49.0% of the 102 automatically CONTAINED pairs within the 240-pair MedForge sample (Supplementary Table S7). Although containment screening removes cases with detected outside-box changes, these human--automatic disagreements show that it reduces spatial-annotation inconsistencies rather than providing a strict annotation-quality filter.

We therefore place greater emphasis on CONTAINED as a diagnostic under reduced outside-box inconsistency, while interpreting absolute [email protected] as agreement with the released annotations. Per-sample verification of the full edit extent, followed by evaluation on reliably annotated samples, could further improve measured localization where annotation mismatch penalizes otherwise correct predictions. Such verification would help distinguish model errors from reference-box discrepancies. However, expert verification across MedForge-90K would require substantial annotation effort and is not feasible within our available time and resources.

Under paper Table 3's conditional localization protocol, BEAR-Med exceeds all six adapted baselines in both views. In particular, it achieves .150 on CONTAINED without MedForge training supervision, versus .080 for the strongest adapted baseline. This supports a comparative transfer advantage, while the audit explains why the absolute score should not be interpreted independently of annotation reliability.

AuthorsOpenReview ↗

Response to Reviewer n7fo (2/2)

W3. "The reasoning evaluation relies mainly on LLM judges, without sufficient validation by medical experts. Fluent explanations do not necessarily reflect correct visual evidence, as some failure cases in the paper also demonstrate."

To address the lack of medical-expert validation, we added physician assessment for this rebuttal. Three physicians with medical-image interpretation experience each rated 120 BEAR-Med rationales on selected images, with no overlapping images across physicians and one reader per image. They saw the image, output, and reference labels/boxes. Visual Evidence Grounding (VEG) assesses whether cues and locations match the image, and Forensic Correctness (FC) assesses whether they justify the conclusion. Table R2 summarizes this independent human assessment.

Table R2. Physician assessment of BEAR-Med rationales. Each dimension has 120 ratings on a 1-5 scale.

DimensionMeanRatings at least 4
VEG3.6867.5%
FC3.8369.2%

Because the reported failures show that correct predictions do not guarantee grounded explanations, we additionally tested whether reasoning context influences localization performance. We varied the rationale prefix while holding the checkpoint, images, instruction, and decoding fixed. Generic text approximately matched the original rationale's length. Table R3 reports F1 on 10,000 test images, [email protected] on all 4,000 local forgeries, and FBR on 5,000 authentic images.

Table R3. Effect of reasoning-prefix interventions.

ConditionF1[email protected]FBR
Natural generation.9929.84700.0016
Original-rationale replay.9930.84600.0016
Empty rationale.9342.82275.0630
Generic rationale.9615.81550.0368

Generic substitution lowers [email protected] by 3.15 percentage points (paired-bootstrap 95% CI: 2.50-3.80) and increases false-box rates. By contrast, reusing the model's original rationale as a prefix yields performance close to natural generation, suggesting that the decline with generic text reflects changed rationale content rather than the act of supplying a prefix itself. Together, these evaluations support physician-assessed visual grounding and a functional role for reasoning context in localization.

W4. "The contribution of expert evidence is not fully separated from privileged information used to generate training targets, including ground-truth labels and bounding boxes. More controlled ablations are needed."

Both the full and no-expert CoT conditions retain the same private GT labels and local boxes. The comparison therefore does not give privileged annotations exclusively to the full model. Instead, the ablation removes expert maps and spatial statistics together with expert-dependent guidance, rationale routing, and weighting, as detailed in Table R4.

Table R4. Controls in the no-expert CoT ablation.

FactorNo-expert condition
GT controlsSame labels and local boxes retained.
Clinical guidelinesCategory-specific guidance retained; expert-dependent rules removed or rewritten.
Rationale lengthStandard 50-120-word targets requested for all samples, replacing expert-dependent routing; lengths not individually matched.
Reasoning weightUniform 1.0 after verification, replacing expert-derived weighting.

Because both variants retain the same GT labels and boxes, the .795-to-.839 [email protected] improvement in Table 6 supports the benefit of expert-guided supervision beyond shared access to privileged annotations.

W5. "The GRPO reward mainly encourages appropriate output length, format, correct verdicts, and box overlap. It does not directly measure the medical correctness or visual grounding of the reasoning, making the claims about reasoning improvement less convincing."

The two training stages have different roles: expert-guided SFT supplies reasoning-content supervision, while G-GRPO optimizes verifiable task and spatial decisions. As specified in Section 4.4, the rationale reward checks length and nonrepetition; it is not a medical-correctness score. Likewise, GT-box IoU rewards spatial alignment, not the truth of every textual claim.

The corresponding ablation in paper Table 6 shows that adding G-GRPO raises [email protected] from .830 to .839 and reduces FPR and FBR from .002 to .001, with F1 unchanged at .995. These results support its role in localization and authentic-image error control.

Reasoning quality is evaluated separately through W3's physician assessment. Accordingly, we attribute the measured localization and false-positive improvements to G-GRPO, while the physician assessment supports the final model's rationale quality rather than an improvement attributable specifically to G-GRPO.

We believe our response has addressed most of your concerns and we respectfully ask if you'recomfortable to re-consider the rating in light of our efforts and dedication

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer n7fo,

Thank you for your time and your constructive feedback on our paper. As the rebuttal period will end soon, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns or questions remain. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Thank you again for your time.

Reviewer_HB8YOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces BEAR-Med-1M, a million-image clinical forgery benchmark covering four imaging categories and five manipulation mechanisms. BEAR-Med distills four frozen experts into image-only reasoning, bounding-box localization, and authenticity prediction through segmented supervised fine-tuning and GRPO. Experiments evaluate detection, localization, rationale quality, robustness, and external transfer.

Critical strengths
  • The benchmark offers substantial scale and manipulation coverage, with explicit distinctions between authentic images, local edits, and global synthesis. Canonical preprocessing and source-grouped splits address several common evaluation shortcuts.
  • The offline expert-distillation design is practically appealing because deployment requires only the original image. Ablations support improvements in strict localization from expert-guided targets and segmented supervision.
  • Evaluation is extensive and includes authentic-image false-positive and false-box rates, degradation tests, and external transfer. The paired-image spatial audit and independent human assessment are particularly useful, and their disagreements are transparently reported.
Critical weaknesses
  • Benchmark provenance remains partly unresolved. Table S2 explicitly flags the authentic-label assignment of 4,586 CT2USforKidneySeg images as requiring verification. This ambiguity should be resolved for a benchmark centered on authenticity. Pooling overlapping medical datasets also requires clearer evidence of cross-source deduplication beyond grouping available patient identifiers.
  • External localization excludes important failures. Appendix G defines “ALL BOX” using only images where each method predicts forged and emits a valid box. Consequently, methods are evaluated on different subsets, and missed detections or missing boxes do not receive zero localization scores. Table 3 needs per-method coverage and unconditional localization results to substantiate the external comparison.
  • The reasoning claims exceed the validation provided. Generating text before a verdict establishes an output order, but does not establish faithful use of the stated evidence. Teachers receive labels and boxes, and rationale quality is assessed by MLLM judges supplied with reference metadata. The human study evaluates outside-box changes, leaving clinical correctness and rationale faithfulness without independent human validation.
Questions for authors
  • Please broaden the comparison with biomedical-forgery literature by discussing BioTamperNet (ICLR 2026), ReSCIND/Integscan (AAAI 2026), and MONet (ICIP 2022), alongside the already cited BioFors (ICCV 2021), M3Dsynth (ICASSP 2024), and MedForensics/DSKI (MICCAI 2025). Clarify differences in benchmark construction, clinical versus scientific-publication imagery, manipulation coverage, annotation granularity, and localization methodology. In particular, explain how BEAR-Med’s expert-guided framework advances beyond prior affinity-guided duplication localization and language-guided biomedical forgery analysis, while acknowledging differences in task scope.
  • Can you resolve the CT2USforKidneySeg provenance issue and clarify whether Med-Banana records use only original clinical images?
  • How were duplicate patients, images, and splicing donors reconciled across source datasets? Was source-image overlap with MedForge independently checked?
  • Please provide Table 3’s per-method localization denominators and results counting missed detections and missing boxes as zero.
  • In the ablation without expert-guided CoT, were guidelines, private label/box controls, rationale lengths, and supervision weights held fixed?
  • What evidence supports rationale faithfulness beyond output ordering and MLLM scoring? Are any existing clinician assessments available?
6 comments
AuthorsOpenReview ↗

Response to Reviewer HB8Y (1/5)

Thank you for recognizing the benchmark's coverage, the practical value of offline expert distillation, and the spatial audits. We address your concerns point by point.

W1. "Benchmark provenance remains partly unresolved" / "clearer evidence of cross-source deduplication beyond grouping available patient identifiers"

CT2US provenance. The CT2US release contains synthetic ultrasound images translated from CT using CycleGAN, rather than acquired ultrasound scans. We have already replaced its 4,586 entries one-for-one with authentic images from BUS-BRA (breast ultrasound) and FETAL_PLANES_DB (routine maternal-fetal ultrasound), leaving the C3 authentic count and overall dataset size unchanged. We confirm that no CM, SP, INS, or REM samples were derived from CT2US. Therefore, this replacement is confined to the authentic-image pool. We would like to clarify that Table S2 has not yet been updated to match this adjustment and this is purely a documentation oversight and will be fixed in the final version, with no impact on any experimental results.

Duplicate checks beyond patient identifiers. The frozen split records contain 129,637 local-edit source groups, with zero recorded groups crossing partitions. This keeps the linked source images and their local derivatives together, controlling leakage through those recorded relationships. Separately, an independent image-content audit compared all 500,000 authentic-labeled records against MedForge test sources using file-level and canonical-pixel SHA-256 hashes, followed by perceptual-hash candidate retrieval and SSIM, normalized-correlation, and difference-hash verification. Q3 reports the additional external-transfer evaluation after excluding all test records derived from identified overlapping sources, where BEAR-Med retains the highest BAcc and macro-F1 among the adapted models.

W2. "External localization excludes important failures" / "Table 3 needs per-method coverage and unconditional localization results"

Thank you for this request. Table R1 adds coverage and unconditional localization from the existing predictions, counting missed detections and missing or invalid boxes as zero. Every method uses the same 20,000 local forgeries in Panel A and the fixed 8,142-image CONTAINED subset in Panel B. Checkpoints, decision thresholds, and box-parsing rules are unchanged. BEAR-Med exceeds all adapted baselines in unconditional [email protected] on both populations.

Table R1. End-to-end external localization with failures scored as zero. Valid counts images predicted forged with a valid box; Coverage = Valid / N. FBR uses all 10,000 authentic test images. The six baselines are BEAR-Med-1M-adapted; the official MedForge model is a target-supervised reference. Higher localization scores and lower FBR are better.

A. All local forgeries (N = 20,000).

MethodValidCoveragemIoU[email protected][email protected]FBR
PSCC-Net11,61058.05%.02632.01046.00397.2643
SparseViT18,46092.30%.16976.11674.03012.0782
IML-ViT18,01190.06%.11337.08070.02585.2160
TruFor19,67998.40%.19248.15308.04208.5840
Mesorch13,95069.75%.10821.07699.02454.0158
SIDA13,07970.40%.09530.05957.00981.0016
BEAR-Med15,63978.20%.16314.14909.06509.0043
MedForge (reference)19,95299.76%.31282.27015.11575.0064

B. Fixed CONTAINED subset (N = 8,142).

MethodValidCoveragemIoU[email protected][email protected]
PSCC-Net472658.04%.02486.00957.00397
SparseViT6,95685.43%.20496.18985.06040
IML-ViT7,21688.63%.15269.13586.05071
TruFor7,93097.40%.25688.24336.07807
Mesorch5,50367.59%.13433.12591.04990
SIDA5,70770.09%.11382.08658.01699
BEAR-Med6,30977.49%.21891.24601.11659
MedForge (reference)8,11099.61%.36974.34795.17588

The original conditional scores remain informative about box quality given a forged verdict and a valid box, while coverage shows how often that condition is met. We will therefore use unconditional localization as the main comparison in Table 3, retain the conditional scores as a supplementary diagnostic, and clarify both definitions in Appendix G.

AuthorsOpenReview ↗

Response to Reviewer HB8Y (2/5)

W3. "Generating text before a verdict ... does not establish faithful use of the stated evidence" / "without independent human validation"

We agree that output order alone does not establish faithful use of evidence. We therefore added two complementary evaluations for this rebuttal: physician assessment of whether the stated evidence is visually supported and the argument is valid, and a controlled intervention testing whether the reasoning context affects subsequent predictions.

Physician assessment. Three physicians each assessed 120 BEAR-Med rationales on selected images, with no overlapping images across physicians and one reader per image. Model identity was hidden; readers saw the image, model output, and reference labels/boxes. Visual Evidence Grounding (VEG) assessed whether stated cues and locations were supported by the image; Forensic Correctness (FC) assessed whether those cues justified the conclusion. Mean scores were 3.68 VEG and 3.83 FC on a 1-5 scale, with 67.5% and 69.2%, respectively, scoring at least 4. These ratings assess visual support and argument validity, rather than verdict agreement alone, extending the original outside-box image audit to model rationales.

Functional use of reasoning. To examine whether reasoning context influences predictions, we followed the CoT-intervention approach of Lanham et al. (2023). On 10,000 BEAR-Med-1M test images, we kept the checkpoint, images, instruction, and decoding settings fixed and compared natural generation, original-rationale replay, an empty prefix, and an approximately length-matched generic prefix. Table R2 reports detection F1 on all 10,000 images, [email protected] on all 4,000 local forgeries, and FBR on all 5,000 authentic images.

Table R2. Reasoning-prefix interventions on the same fixed test subset. Higher F1 and [email protected] and lower FBR are better.

Rationale conditionF1[email protected]Authentic FBR
Natural generation.9929.84700.0016
Original-rationale replay.9930.84600.0016
Empty.9342.82275.0630
Generic.9615.81550.0368

Generic substitution lowers F1 from .9929 to .9615, reduces [email protected] by 3.15 percentage points (paired-bootstrap 95% CI: 2.50-3.80), and increases FBR from .0016 to .0368. In contrast, replaying the original rationale preserves similar aggregate performance, helping distinguish the effect of replacing rationale content from that of prefix-based continuation itself. These results support a functional role for reasoning context in detection, localization, and false-box control, beyond its position before the verdict.

Together, these evaluations support physician-assessed visual grounding and a measurable functional role for reasoning context. Establishing a faithful correspondence between every verbalized cue and the model's internal computations is a distinct, technically challenging research problem in CoT interpretability, requiring evidence beyond output-quality assessment and prefix interventions (Tutek et al., 2025).

AuthorsOpenReview ↗

Response to Reviewer HB8Y (3/5)

Q1. "Please broaden the comparison with biomedical-forgery literature" / "explain how BEAR-Med's expert-guided framework advances beyond prior affinity-guided duplication localization and language-guided biomedical forgery analysis"

Thank you for these relevant references. We will expand Related Work to include BioTamperNet, ReSCIND/Integscan, and MONet. Table R3(a) distinguishes contribution types and manipulation coverage; Table R3(b) compares specific image types, construction, and localization approaches.

Table R3(a). Contributions, manipulation coverage, and annotation granularity.

WorkPrimary contributionIntra-dup.Inter-dup.InsertionRemovalGlobal synthesisAnnotations
BioForsBenchmark●●○○○Pixel masks
MONetMethod○●○○○Pixel masks
BioTamperNetMethod●●○○○Pixel masks
ReSCIND / IntegscanBenchmark + method●●○●○Pixel masks
M3DsynthBenchmark○○●●○Voxel masks
MedForensics / DSKIBenchmark + method○○○○●Image labels
BEAR-Med-1M / BEAR-MedBenchmark + method●●●●●Image labels + local boxes

●: explicitly reported coverage; ○: not a separately reported task, rather than a capability limitation. Intra-/inter-dup. denote within-/across-image duplication, including splicing. Method rows describe their published evaluation scope; benchmark contributions include baseline evaluation.

Table R3(b). Image types, benchmark/training-data construction, and localization methodology.

WorkImage typesConstructionLocalization methodology
BioForsMicroscopy, blot/gel, FACS, macroscopyReal-world scientific figures extracted from publicationsDuplication and anomalous-boundary localization baselines
MONetBioFors image typesSynthetic duplication pairsHierarchical multi-scale overlap matching
BioTamperNetBioFors image typesAugmented and GAN-generated patchesAffinity-guided self-/cross-attention with state-space modeling; duplication and boundary localization
ReSCIND / IntegscanMicroscopy, blot/gel, FACS, macroscopyImage-processing, generative, and VLM-guided editsPrompt-conditioned state-space modeling with dense manipulation and anomalous-boundary localization
M3DsynthLung CT volumesGAN/diffusion-based nodule insertion and removalDense 3D localization evaluated against volumetric targets
MedForensics / DSKIX-ray, CT, MRI, ultrasound, endoscopy, histopathologyTwelve medical image generatorsImage-level classification using spatial/noise adaptation and forensic retrieval
BEAR-Med-1M / BEAR-MedX-ray, CT, MRI, ultrasound, endoscopy, dermoscopy, fundusClinically constrained CM/SP, restoration-based INS/REM, and full-image synthesis; paired spatial auditsOffline expert distillation into image-only reasoning, box localization, and verdict generation

Beyond affinity-guided duplication localization. BEAR-Med integrates local traces, image consistency, generation texture, and authentic-reference medical-normality evidence, addressing duplicated content alongside lesion removal and synthesis, where a duplicated counterpart need not exist. Its contribution is therefore to transfer complementary forensic evidence into a unified clinical-forensics model, extending the task beyond correspondence-based localization.

Beyond language-guided analysis. Integscan uses language to condition dense localization, while DSKI adapts a vision-language representation and retrieves forensic knowledge for classification. BEAR-Med instead uses four frozen experts offline to construct evidence-guided rationale targets, followed by segmented SFT and spatially grounded policy optimization. This transfers expert guidance into the student's parameters, enabling reasoning, localization, and authenticity prediction from the image and a fixed instruction, without deployment-time experts or evidence retrieval.

Together, our benchmark and method combine diverse clinical image types, conventional and generative manipulations, and training-time evidence transfer, complementing scientific-publication integrity analysis and specialized volumetric localization.

Reviewer_4R2POpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on identifying and localizing forged clinical images and introduces two contributions. The first is BEAR-Med-1M, a benchmark of 500K authentic and 500K forged medical images across four imaging categories (radiography, CT/MRI, ultrasound, optical) and five mechanisms (copy-move, splicing, lesion insertion, lesion removal, full synthesis), with 25K images per category-mechanism cell. The second is BEAR-Med, a Qwen3-VL-8B policy that emits a reasoning span, a box span, and a verdict span from the raw image. Four frozen experts (local trace, image consistency, generation texture via NPR+DIRE, medical normality via DINOv2+PatchCore) are used offline to produce feature-specific forge detection evidence. A Qwen3.7-Plus teacher that sees the image, rendered expert maps, guideline rules, and the private GT label and box to write a rationale, and the student is then trained with per-span-normalized SFT (rationale, box, and verdict weighted w_r, 3, and 3) followed by GRPO with format, length, verdict, scope, and IoU rewards. On the BEAR-Med-1M test split, it reaches .995 F1 and .839 Box [email protected] versus .992 and .777 for the adapted MedForge. On MedForge-90K, it reaches .769 balanced accuracy and .083 [email protected] (.150 on audit-contained pairs). Ablations, a two-MLLM-judge rationale audit on 1,200 images, and a degradation sweep are also provided.

Critical strengths
  1. The paired outside-box audit is a useful contribution on its own. The finding that 59% of MedForge-90K lesion-edit pairs by the automatic audit, and 75% by two human reviewers with kappa 0.55, have meaningful changes outside the annotated box (Tables S4-S6) is well documented and generator-stratified (Table S5). This justifies that restricting inpainting to the mask and auditing the result is the right fix.

  2. The evaluation protocol is explicit, with predeclared thresholds, frozen prompts, full-split evaluation, per-cell and per-expert diagnostics, FPR/FBR, and scope accuracy. The paper reports the diagnostics a reader needs to judge saturation and shortcuts, as discussed below.

Critical weaknesses
  1. It is not clear what the benchmark contributes beyond scale and a particular set of generators, and the clinically grounded framing is limited in implementation in construction or evaluation.

    • The paper never states what challenge BEAR-Med-1M poses that M3Dsynth, MedForensics, or MedForge-90K do not, and its own numbers suggest the task is not hard. Two fine-tuned baselines reach at least 0.9 F1 and 0.8 mIoU (Table 1), and four of the five rows in Table 6 for ablation share an F1 of .995. Forged image identification is saturated, and the reported differences reduce to Box [email protected] in a .77-.84 band.
    • The clinically specific parts of construction are mostly a lesion-to-organ placement map, an affinity grouping for SP donors, and a VLM pathology check inside the box (A.3). The first two are necessary engineering for any usable forgery pipeline rather than a contribution, since a patch placed outside its organ would be implausible by construction, and the third is a yes-or-no abnormality check that does not assess whether the inserted or removed finding is clinically plausible. Beyond these, the inpainting models are standard FLUX, SDXL, and SD3 fine-tunes. As a result, there's no solid clinical plausibility check. The only human involvement is the 240-pair spillover check. The paper therefore has no evidence that its forgeries are clinically plausible, which is the property the title and introduction rest on.
    • The diagnostics point to a low-level shortcut rather than clinical implausibility. The generation-texture expert, whose CNN branches over NPR difference channels and DIRE reconstruction error are fused by a calibration MLP (B.4), reaches .999 AUC on INS and REM while its [email protected] is at most .007 (Table 4) and its mean box IoU is .06-.11 (Table S25), against .54-.67 for the trace and consistency experts. In the C2/REM example of Figure S5, its response is diffuse and largely misses the GT box, which the caption itself calls a "broader generation-texture response". Meanwhile, copy-move in C1/CM hardly triggers it. My reading is that this could be a whole-image generator fingerprint, which would arise if the inpainting output is kept rather than composited back onto the source outside the mask.
    • Two further confounds go unexamined. Authentic images are 54% CT/MRI (Tables S2, S3), while forged images are 25% per category, and INS/REM sources must carry lesion annotations, so category and source identity correlate with label. Every image is also bicubically resized to 512x512, so a radiograph or dermoscopy image might be downscaled 4-10x while a CT slice is barely resampled, and the paper does not say whether forgeries are applied before or after this resize. If before, artifact visibility differs by category, and the C1-C4 comparisons partly measure resampling. If after, authentic and forged images do not share one resize path, which is the potential format shortcut Section 3.3 says canonicalization removes. Either way, the experts' hand-designed inputs are tied to this.
  2. The paper does not justify why an MLLM is needed, and the MLLM-specific contributions are neither isolated nor shown to be load-bearing.

    • The task as evaluated is a verdict plus one box and does not necessarily depend on text. A 5x5 convolution over the four expert maps already gives .985 AUC (Table 5), and Mesorch reaches .945 F1 and .790 [email protected] (Table 1). The 8B policy adds .05 on each at far higher cost, and the rationale it uniquely adds is never shown to improve the box or to be clinically useful. The expert numbers also create a dilemma. Either the experts are genuinely this strong and generalizable, with .977-.999 AUC for hand-designed features (Table 4), so the policy is unnecessary, or, as I argue in point 3, they are overfit to this benchmark, so the CoT targets distilled from them encode in-distribution shortcuts and lead to the reduced .769 BAcc on MedForge-90K. The experiment that would distinguish the two, expert performance on external data, is not reported.
    • The gains attributed to the rationale are small, and the ablations cannot attribute them. Over the adapted MedForge, the margin is .003 F1 and .062 [email protected], and each one-at-a-time ablation in Table 6 moves [email protected] by only .009-.044 from a single run with no seeds. Within that margin, nothing isolates the rationale itself. No variant is trained to emit a box and verdict without a rationale, and the variant without expert-guided CoT still uses a teacher that sees the GT box and label (C.3), so it changes the rationale's style rather than removing guidance. Two additional observations that do bear on the question point the other way. Feeding the rendered expert panel directly to the policy lowers [email protected] by .07, and Figure S14 shows a box with IoU .803 next to a rationale both judges score 2 on grounding.
    • The rationale evaluation uses two MLLM judges and no clinician. Gains over the strongest baseline are .065-.098 on a 1-5 scale (Table S32), and the printed rationales are near-identical across methods and images. SIDA and BEAR-Med share whole clauses verbatim in Figures S9 and S11, and the same clause recurs for an ultrasound in S9 and a radiograph in S14. My reading, which the paper gives no data to confirm or refute, is that this reflects the task itself. A rationale has to express one of two verdicts, one region, and a handful of artifact types, so the training targets are likely heavily repetitive, and if so, the text channel may add little beyond the box and verdict. The abstract's claim of strong cross-dataset performance is also not fully supported by Table 3 (.769 BAcc, .083 [email protected]), and unreliable MedForge-90K boxes could explain the localization numbers but not the identification drop.
  3. The medical-normality expert is weakly justified, and all four experts are feature-engineered detectors never tested outside their training distribution.

    • The expert scores DINOv2 patches against a PatchCore memory of authentic training patches (B.5), but the authentic pool includes the lesion-bearing INS/REM sources and is 54% CT/MRI drawn heavily from BraTS editions, so normality means resemblance to the admitted sources rather than clinical normality. The paper concedes that authentic pathology can trigger it (4.2, B.5), and the numbers agree. It reaches .70-.74 AUC on C1/C2 and CM/SP, at most .028 [email protected] (Table 4), and a mean top-10% response of .69 on authentic test images versus .84-.95 on forged ones (Tables S23-S24), while the other experts respond at .01-.02 on authentic images. Its Figure S5 maps are scattered blobs with hardly any anatomical correspondence.
    • All four experts take certain hand-designed inputs (median residual, Sobel, NPR shifts, DIRE error, k-center DINOv2 memories) tuned to this benchmark's generators and canonicalization, and none is evaluated outside BEAR-Med-1M. A CNN on residuals and edges reaching .977-.985 AUC across all five mechanisms could be the kind of result one sees when a detector has learned a fixed set of editors rather than manipulation in general. I would hypothesize that the experts transfer limitedly and that the student's MedForge-90K result partly reflects this, since its CoT targets were organized around in-distribution expert evidence.
    • There is no leave-one-expert-out ablation of CoT construction, so the paper cannot say whether the medical-normality or generation-texture evidence changes any training target. The teacher is told to describe visible cues without naming experts, and the guideline bank (Tables S10-S14) is a generic checklist, so the interpretability claim rests on fluent prose rather than anything traceable to the expert designs.
Questions for authors
  1. For INS/REM, is the inpainting model's full decoded output kept, or is it composited back onto the authentic source outside the mask, and if so, with what blending? Relatedly, how does the generation-texture expert's response inside the GT box compare with its response outside it on INS/REM, and how well does NPR alone, DIRE alone, or a plain ResNet-50 on raw pixels separate forged from authentic per mechanism? These would tell me whether INS/REM detection rests on a whole-image generator trace.
  2. Is each forgery applied at native resolution before canonicalization or on the 512x512 canvas after it? What are the native resolutions and downscale factors per source dataset and per class, and have the authors observed any effect of the canonicalization path on final performance?
  3. Have the four experts been evaluated on any data outside their training distribution, for example on a split of BEAR-Med-1M that holds out one of the three inpainting editors, or on images produced by an editor or acquisition pipeline not used in construction? I could not find any expert evaluation beyond the in-distribution test split, and the .98-.999 in-distribution AUCs make this the main open question about whether the distilled CoT targets encode anything transferable.
  4. Was a variant with the same backbone, LoRA, data, and segmented loss but no rationale span ever trained, and if so, how did its F1 and [email protected] compare? Likewise, was the Table 5 fusion head ever trained on box targets at full scale? What happens to [email protected] if the box span is forced after an empty or generic rationale at inference?
  5. Were the full model and the Table 6 variants trained with more than one seed? Given the variability of RL and the size of the gains, I would like to know how the [email protected] differences compare with run-to-run variation.
  6. Has any human attempted to tell forged from authentic BEAR-Med-1M images, and if so, what was their accuracy by mechanism and category? What fraction of INS/REM masks are axis-aligned rectangles?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 4R2P (1/6)

Thank you for recognizing the spatial audit and evaluation protocol. We respond point by point, quoting your comments.

W1.1. "The paper never states what challenge BEAR-Med-1M poses that M3Dsynth, MedForensics, or MedForge-90K do not" / "Forged image identification is saturated"

BEAR-Med-1M extends beyond scale through imaging diversity, local-to-global mechanisms and spatial auditing (Table R1).

Table R1. Benchmark imaging and manipulation coverage.

BenchmarkImaging typesManipulations
M3Dsynth3D lung CTNodule insertion/removal
MedForensicsX-ray, CT, MRI, ultrasound, endoscopy, histopathologyFull synthesis
MedForge-90KChest X-ray, brain MRI, fundusLesion insertion/removal
BEAR-Med-1MRadiography, CT/MRI, ultrasound, optical imaging including endoscopy, dermoscopy and fundusCM, SP, INS, REM and SYN in every category

Each category covers all five mechanisms, enabling local/global detection, local-edit localization and false-box control on authentic/full-synthesis images. The paired audit checks meaningful changes beyond the box.

Regarding saturation, near-ceiling detection is also reported by MedForge (99.23% accuracy, 98.98% F1), M3Dsynth (97.3-99.5% accuracy, matched generators), FaceForensics++ (99.26% accuracy, RAW) and FakeVLM (98.6% accuracy, 98.1% F1, FakeClue). However, detection does not imply localization: SIDA reaches .994 AUC but .196 [email protected]; adapted MedForge and BEAR-Med exceed .99 F1 yet reach .777/.839 [email protected] (paper Table 1).

In high-stakes clinical forensics, lesion edits can affect diagnosis and treatment. A binary verdict leaves the altered finding unidentified, making localization central, not auxiliary. We will clarify this remaining spatial challenge in the revision.

AuthorsOpenReview ↗

Response to Reviewer 4R2P (2/6)

W1.2. "The first two are necessary engineering for any usable forgery pipeline rather than a contribution" / "The paper therefore has no evidence that its forgeries are clinically plausible"

We agree that anatomical plausibility is a prerequisite, but implementing it across imaging categories and mechanisms is not automatic. M3Dsynth uses nodule-centered lung-CT placement, whereas MedForge relies on medical editing instructions and iterative VLM verification. BEAR-Med explicitly constrains construction: INS/REM learn lesion/healthy-tissue restoration, respectively; lesion-to-organ mappings and clinical donor groups constrain placement and source compatibility, with organ-mask checks where available. Thus, the contribution is a reproducible protocol across four categories and five mechanisms, complemented by spatial-fidelity auditing, not a claim that anatomical consistency itself is novel.

Automated pathology checks and spatial audits alone do not establish clinical plausibility. We therefore added two physician-evaluation tasks for this rebuttal. Three physicians experienced in image interpretation assessed non-overlapping subsets of a stratified, investigator-screened set:

  • Task A: authenticity discrimination. Distinguish authentic images from local manipulations or full synthesis, with dataset identities and authenticity labels hidden.
  • Task B: clinical plausibility. Score anatomy, morphology and tissue compatibility (1-5) before source images or labels are revealed; after locking ratings, inspect references to verify edit success.

Images were stratified by dataset, category and authenticity/mechanism. Table R2 gives the quotas; Task B uses a subset of Task A.

Table R2. Physician-study composition.

TaskBEAR-MedMedForge controls
AEach of four categories: 50 authentic + 10 per mechanism; 400 imagesEach of chest X-ray/brain MRI: 20 authentic + 10 INS + 10 REM; 80 images
BEach of four categories: 10 authentic + 10 per mechanism; 240 imagesEach of chest X-ray/brain MRI: 10 authentic + 10 INS + 10 REM; 60 images

Task A shows near-chance BEAR-Med discrimination (Table R3). Similar detection and false-positive rates indicate little separation of authentic from forged images; 74.5% of forgeries were judged authentic. MedForge has a higher detection rate.

Table R3. Blinded physician authenticity discrimination.

DatasetAccuracyForgery detectionAuthentic FPR
BEAR-Med48.75%25.5%28.0%
MedForge60.00%35.0%15.0%

Complementing discrimination, Task B evaluates clinical plausibility directly (Table R4). BEAR-Med's absolute differences from authentic-control means are below .08 points, even with slightly higher anatomy and morphology scores. Moreover, on shared chest-X-ray/brain-MRI INS/REM tasks, BEAR-Med receives higher means than MedForge on all three dimensions.

Table R4. Clinical-plausibility ratings, scored before reference reveal.

ImagesAnatomyMorphologyTissue compatibility
BEAR-Med: all forgeries4.0553.9603.805
Authentic controls: both datasets4.0173.8833.833
BEAR-Med: shared X-ray/MRI INS/REM3.8503.8503.725
MedForge: shared X-ray/MRI INS/REM3.5253.2752.975

After ratings were locked, source-forgery inspection assessed insertion/removal success. Edit confirmation reached 82.5% versus 55.0% for MedForge on shared tasks (difference: 27.5 percentage points; 95% bootstrap CI: 5.0-40.0), retaining uncertain responses in the denominator.

Together, near-chance discrimination, scores close to authentic controls and higher edit-confirmation rates provide direct physician evidence for clinical plausibility on the evaluated forgeries, beyond automated pathology checks and outside-box auditing.

AuthorsOpenReview ↗

Response to Reviewer 4R2P (3/6)

W1.3. "this could be a whole-image generator fingerprint"

Generation texture is established forensic evidence (NPR, DIRE). Clinical forensics combines acquisition-related texture, noise and image consistency with anatomy and clinical plausibility. Our generation-texture expert targets the former: a strong response indicates generative processing, not a particular lesion edit.

Thank you for motivating a shortcut check. Generative models also support reconstruction for legitimate restoration/super-resolution, as in DiffIR; their involvement alone should not constitute forgery. We therefore added a reconstruction diagnostic for this rebuttal.

The controls cover four categories and FLUX, SD3.5 and SDXL VAEs, 40 images per VAE, without intentional lesion edits. Fixed-set re-evaluation gives 117 authentic verdicts, three forged verdicts and zero local boxes. Some accepted controls have cached expert responses of 1.000: strong generation evidence need not dictate a verdict or box.

The three false positives cite global texture/noise inconsistent with natural acquisition, not implausible lesions; their no-box outputs reflect that global concern rather than a local clinical edit.

W1.4. "category and source identity correlate with label" / "the potential format shortcut Section 3.3 says canonicalization removes"

CT/MRI constitutes 53.52% of authentic images versus 25% of forgeries. However, paper Table 2 reports within-category F1 of 1.000/.980/1.000/.999. These within-category results show that category proportions alone cannot explain the strong detection performance.

To address annotation-bearing source priors, we added a fixed paired diagnostic for this rebuttal: 120 test-only source/edit pairs across four categories, INS/REM and three generators. Each source supplies both labels, balancing category and source-dataset identity. All 240 verdicts are correct, with .877 mIoU/.875 [email protected] on edits. Thus, source/category identity alone cannot explain this paired discrimination.

Our pipeline standardizes final inputs. Authentic images are square-padded/exported at 512 x 512; INS/REM edit prepared 512 x 512 inputs, while CM/SP edit source/recipient canvases before canonicalization. All share 512 x 512, 8-bit RGB, lossless PNG representation. Hence stored resolution, color format and bit depth have no class-specific specifications; we will clarify preparation, editing and standardization in the revision.

Beyond this format standard, the original resampling test retains .853 F1/.706 [email protected] after 512-to-256-to-512 resampling on the fixed 10K subset (Section 5.7/Figure 3; Appendix I/Table S38). Thus, detection and strict localization persist after changing the sampling representation, rather than requiring original canonical pixels. Together with within-category and source-matched results, this provides empirical evidence beyond normalization alone to address the shortcut concern.

W2.1. "The task as evaluated is a verdict plus one box and does not necessarily depend on text." / "already gives .985 AUC" / "never shown to improve the box or to be clinically useful"

Clinical image forensics is particularly high-stakes. Unlike object detection, the issue is why a finding should be considered manipulated, not simply what is present and where. An MLLM connects detection/localization to a rationale a clinical reviewer can inspect.

Beyond text, full-split, box-supervised 5 x 5 fusion reaches .985 AUC but .484 [email protected]/.067 FBR, versus .839/.001 for BEAR-Med (paper Table 5); Mesorch reaches .790 [email protected]. For the complete framework, .777-to-.839 [email protected] over adapted MedForge means about 28% fewer localization failures at IoU .75, despite near-ceiling F1. Hence strong detection does not make precise localization or the policy redundant. The 8B model costs more than compact localizers, but deployment needs no experts.

We added physician assessment for this rebuttal: three physicians rated 120 BEAR-Med rationales in non-overlapping subsets. Readers saw images, rationales, predicted verdicts/boxes and reference labels/boxes, with model identity hidden. Visual Evidence Grounding (VEG) assessed image support for observations/locations; Forensic Correctness (FC) assessed whether they justified the conclusion. Both used 1-5 ratings; a correct verdict alone was insufficient.

Scores were 3.68 VEG and 3.83 FC; 67.5%/69.2% scored at least 4, indicating supported observations or valid arguments with at most minor issues. Thus, physician assessment extends validation beyond MLLM judges and supports making the visual basis of an allegation available for scrutiny.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.