acceptodds
Under review as a conference paper at ICLR 2027

MieDB-100k: A Comprehensive Dataset for Medical Image Editing

Abstract

The scarcity of high-quality data remains a primary bottleneck in adapting multi-modal generative models for medical image editing. Existing medical image edit dataset often suffer from limited diversity, neglect of medical image understanding and inability to balance quality with scalability. To address these gaps, we propose MieDB-100k, a large-scale, high-quality and diverse dataset for text-guide medical image editing. It categorizes editing tasks into perspectives of Perception, Modification and Transformation, considering both understanding and generation abilities. We construct MieDB-100k via a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods, followed by rigorous manual inspection to ensure clinical fidelity. Extensive experiments demonstrate that model trained with MieDB-100k consistently outperform both open-source and proprietary multi-modal models while exhibiting strong generalization ability. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing.

Primary area:
datasets and benchmarks
Keywords:
Medical, Image Editing, Dataset

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

The paper presents MieDB-100k, a dataset of 104,267 text-guided editing triplets built from 48 public datasets across 10 imaging modalities. Perception tasks paint segmentation labels onto the image as color overlays. Modification tasks use modality-specific FLUX inpainting experts to turn lesion-bearing images into healthy counterparts, followed by automatic filtering and expert review. Transformation tasks use paired images from public repositories. OmniGen2 finetuned on the training split achieves the best results in Table 2 across all three perspectives.

Critical strengths

The Modification pipeline is a reasonable way to obtain counterfactual pairs with consistent backgrounds. Inpainting is confined to the lesion mask, and expert review after automatic filtering discarded 9,294 of 25,677 candidate pairs. The Turing test in Table 16 gives direct evidence on the realism of the inpainted tissue.

Critical weaknesses
  1. The benchmark is built from the test splits of the same source datasets and follows the same task construction used for training. OmniGen2-MIE is therefore evaluated on a distribution closely matched to its training data, while the other models are used zero-shot. For lesion removal, the Rubric-S reference is also generated by the FLUX-based inpainting pipeline. In Figure 5, Nano Banana Pro scores 58.1 on the unseen bone metastasis task, compared with 52.2 for OmniGen2-MIE. Table 15 only compares OmniGen2-MIE with OmniGen2, so the evidence for outperforming proprietary models mainly comes from the MieDB-100k benchmark.
  2. The paper presents MieDB-100k as a higher-quality alternative to existing medical editing datasets, but no model is trained on an existing dataset. The effect of the proposed curation is therefore difficult to separate from general in-domain finetuning.
  3. Section 4.2 links low Perception scores to a lack of medical knowledge. However, Table 3 lists a Perception ACC of 0.000 for the M-only model while its Modification RubricScore is 57.5. Low overlay performance may also reflect difficulty with the required output format.
Questions for authors
  1. Section 4.2 treats low Perception scores as evidence of missing medical knowledge. How do the authors reconcile this with the M-only model in Table 3, which has a Perception ACC of 0.000 but a Modification RubricScore of 57.5?
  2. The reported correlation of −0.905 appears to match the correlation between the ten model-level averages in Table 2. How many cases were ranked by the three evaluators, and how well do Rubric-S and human rankings agree at the case level?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer CKiS (1/2)

We sincerely thank the reviewer for the thorough review and constructive feedback, and for recognizing the soundness of our Modification pipeline, rigorous expert filtering, and clinical Turing test validation. Below, we address your questions and concerns point by point.


Benchmark Distribution, Ground-Truth Authenticity, and External Comparisons (W1)

We will refine our wording in the Abstract and Introduction to state more precisely that OmniGen2-MIE "surpasses open-source and proprietary models on the MieDB-100k benchmark while achieving highly competitive zero-shot generalization approaching SOTA proprietary models on unseen editing targets". We address the rest concern regarding in-distribution evaluation and FLUX-generated references from three perspectives:

  • Performance on 100% Real Clinical Ground-Truth Targets: In Modification, FLUX inpainting is only used to synthesize the healthy tissue (H). For Lesion Addition and Localized Tumor Simulation tasks (~50% of Modification), H is the input and the original, 100% real clinical lesion image is the reference target. Similarly, 100% of Perception and Transformation targets are real ground-truth annotations or physical reconstructions. Evaluating strictly on the Lesion Addition subset (100% real targets), OmniGen2-MIE still outperforms Nano Banana Pro and all other baselines.
  • External Validation (Tab. 15): We have evaluated on MedE-Bench (which consists of real longitudinal clinical image pairs completely disjoint from MieDB-100k). The performance increases after training, which proves the generalization ability beyond in-distribution assessment.
  • Fig. 5 Clarification: For our 7B OmniGen2-MIE, "bone metastasis" is a strictly zero-shot Modification target (never seen in Modification training, learned solely via cross-task transfer from Perception masks). Training on MieDB-100k boosts the 7B OmniGen2 backbone from 33.7 to 52.2 on this unseen task, bringing an open-source 7B model remarkably close to Nano Banana Pro (58.1)—a massive commercial system trained on orders of magnitude more proprietary data.

Isolating the Effect of Our Curation Pipeline from General In-Domain Finetuning (W2)

To directly separate the effect of our proposed curation pipeline from generic in-domain finetuning, we finetuned the exact same OmniGen2 backbone on Med-banana-50K using the same training recipe, and compared it against our models:

Training Dataset (Backbone: OmniGen2 7B)#Mod. TripletsCuration PipelineModification (Rubric-S ↑)
Baseline (Zero-shot)0—29.1
Med-banana-50K (In-domain finetuning)~50kGemini distillation41.5
MieDB-100k (M-only full, Tab. 3)41.6kOurs (Expert Inpaint + Filter + Audit)57.5
MieDB-100k (Full P + M + T)41.6k (+P,T)Ours (Unified P + M + T Curation)65.9

This controlled comparison attribute the performance gains into three distinct factors:

  1. Generic In-Domain Finetuning (29.1 → 41.5): Finetuning on an existing other medical editing dataset provides a moderate baseline boost from domain exposure.
  2. Superiority of Our Curation Pipeline (41.5 → 57.5): Under a single-task Modification-only setup with comparable data scale, our M-only model outperforms OmniGen2-MedBanana by a massive +16.0 points on Modification Rubric-S. Since both models share the same architecture and training objective, this +16.0 gap is directly attributable to our modality-specific healthy inpainting, nnUNet/VLM rejection sampling, and clinician quality audit as well as broader modality coverage (10 vs. 3).
  3. Synergy from Unified Task Formulation (57.5 → 65.9): Incorporating our curated Perception and Transformation tasks yields an additional +8.4 gain on Modification while unlocking strong Perception and Transformation capabilities in a single unified model.

We will include this OmniGen2-MedBanana comparison table and analysis in the revised manuscript.

AuthorsOpenReview ↗

Reply to Reviewer CKiS (2/2)

Reconciling Low Perception Scores of Zero-Shot Baselines vs. the M-only Model (Q1 & W3)

Thank you for raising this insightful point. The low Perception scores of zero-shot general-purpose baselines and our single-task M-only model arise from two fundamentally different mechanisms:

  • Zero-Shot Baselines Fail Due to Missing Radiological Knowledge, Not Format Unfamiliarity: Untouched baselines already possess the instruction-following ability to paint semi-transparent colored overlays. As shown in our modality-wise breakdown (Appendix E.1, Fig. 10), under the exact same prompt format, baselines achieve strong DICE scores (~0.60–0.80) on optical modalities resembling natural images, where OmniGen2 scores 0.248 overall. Furthermore, our least-squares alpha de-blending extraction (Appendix D.1, Eq. 2) is invariant to opacity variations. Their failure specifically on non-optical scans (CT, MRI, OCT) occurs because they paint the requested color onto the wrong anatomical structures (Fig. 4), exposing a lack of radiological domain knowledge.
  • M-only Scores 0.000 Due to Single-Task Mode Collapse (Format Forgetting): Unlike zero-shot baselines, M-only is full-parameter finetuned for 20,000 steps exclusively on 41,644 Modification triplets, where every target output is a natural-looking clinical scan without any artificial color overlays. Consequently, M-only experiences catastrophic forgetting of the color-painting output mode.

These support the core of our work: "Perception Matters for Modification". While M-only learns implicit localization for trained lesion types, explicitly retaining and injecting anatomical understanding via joint training can further boost Modification Rubric-S from 57.5 to 65.9, and enables zero-shot concept transfer to unseen lesions (Fig. 5).


Case-Level Agreement and Sample Size for Human Preference Ranking (Q2)

In Sec. 3.3.2, the reported ρ=−0.905 indeed measures the model-level Spearman rank correlation across the 10 benchmarked models. Here we clarify the evaluation scale and case-level agreement:

  • Number of Ranked Cases: All three clinical evaluators independently ranked all 618 test cases in the Modification benchmark (comparing all 10 model outputs per case, totaling 6180 generated images evaluated per clinician).
  • Case-Level Correlation: When computing the Spearman rank correlation between VLM Rubric-S and each of the three clinical evaluators' preference rankings within each individual test case and averaging across all cases, we obtain a case-level correlation ρ of -0.687, -0.700, -0.673 ( with all p<1×10−3). This confirms high case-level agreement between our rubric-guided VLM evaluator and individual clinical experts.

We hope these analyses and additional experiments fully address your concerns.

Thanks for the response and the extra experiments. This addresses several of my initial questions, though I still have two quick follow-ups before finalizing my score. You noted that OmniGen2-MIE still beats Nano Banana Pro on the Lesion Addition subset with real targets. Could you post the exact Rubric-S and Pref-Rank numbers broken down into Lesion Addition vs. Lesion Removal? Also, the Med-banana-50K comparison is still tested on MieDB-100k, and Table 15 only compares against untuned OmniGen2. Could you run OmniGen2-MedBanana and at least one strong baseline such as Nano Banana Pro or Qwen-Image-Edit on MedE-Bench?

AuthorsOpenReview ↗

Reply to Reviewer CKiS (Round II)

Thank you for your prompt reply and constructive follow-up questions. Here we supplement the detailed breakdown:


1. Lesion Addition vs. Lesion Removal

ModelLesion Addition (100% Real Target LL) Rubric-S (↑\uparrow) / Pref-Rank (↓\downarrow)Lesion Removal (Inpainted Target HH) Rubric-S (↑\uparrow) / Pref-Rank (↓\downarrow)
Nano Banana Pro67.78 / 2.0358.94 / 2.05
OmniGen2-MIE (Ours)69.09 / 1.3862.53 / 1.40

OmniGen2-MIE consistently outperforms Nano Banana Pro across both directions. Specifically, on the Lesion Addition subset (the ground-truth targets (LL) are 100% real clinical scans), OmniGen2-MIE still achieves a clear lead, confirming that its superiority holds on authentic clinical pathology targets rather than merely matching FLUX-inpainted references.


2. Expanded External Evaluation on MedE-Bench

ModelSizeCLIP (↑\uparrow)PSNR (↑\uparrow)SSIM (↑\uparrow)EA (↑\uparrow)CP (↑\uparrow)VQ (↑\uparrow)
OmniGen27B0.23414.60.8980.350.640.51
Qwen-Image-Edit27B0.26315.60.9220.470.660.52
OmniGen2-MedBanana7B0.26115.00.9350.420.680.54
OmniGen2-MIE (Ours)7B0.26715.40.9510.460.770.57

We note that because the data distribution and longitudinal collection setup of MedE-Bench differ from our dataset, the performance margins are naturally more moderate than those measured on the in-distribution benchmark. Nevertheless, two clear conclusions emerge:

  1. Direct Dataset Comparison: Evaluated on an independent, real-pair external benchmark using the exact same 7B backbone, OmniGen2-MIE outperforms OmniGen2-MedBanana across all six metrics, directly validating the superior quality and generalizability of our curation pipeline over Med-banana-50K.
  2. Comparison with Strong Baselines: Despite having nearly 4×4\times fewer parameters than Qwen-Image-Edit (7B vs. 27B), OmniGen2-MIE achieves the best overall performance.

We will incorporate both tables into the revised manuscript and thank you again for your constructive engagement!

Reviewer_U7G1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MieDB-100k, a large-scale medical image editing dataset covering 10 modalities and 63 editing targets. The dataset unifies Perception, Modification, and Transformation tasks through rule-based construction and modality-specific expert models. Experiments demonstrate substantial improvements in medical image editing performance, supported by ablation studies and external evaluations.

Critical strengths
  1. Comprehensive and practically useful dataset. MieDB-100k covers diverse medical modalities and editing tasks, extending beyond conventional image modification. Its scalable construction pipeline and expert quality inspection provide a valuable resource for medical image editing research.

  2. Extensive experimental validation. The paper evaluates multiple open-source and proprietary models and provides task-specific ablations, cross-architecture experiments, and external validation. The results demonstrate clear benefits of domain-specific editing data.

Critical weaknesses
  1. Potential evaluation bias in perception tasks. Perception tasks require models to generate segmentation overlays with specific colors and transparency. General-purpose editing models may perform poorly because they are unfamiliar with this output format, rather than lacking medical understanding.

  2. Uncertainty regarding the clinical validity of synthetic editing targets. Modification data are constructed using synthetic healthy images obtained through lesion inpainting, with lesion addition and removal pairs generated bidirectionally. Although expert inspection supports visual quality, it remains unclear whether these counterfactual targets consistently preserve clinically meaningful anatomy and pathology, particularly for lesion addition.

Questions for authors
  1. How can the authors distinguish failures in medical image understanding from failures to follow the specific mask-overlay format in Perception tasks?

  2. How do the authors validate the clinical correctness of synthetic lesion addition and removal targets beyond visual plausibility?

  3. Can the authors provide further evidence that the gains from joint Perception and Modification training arise from improved medical understanding rather than general multi-task regularization?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer U7G1 (1/2)

We sincerely thank the reviewer for the encouraging feedback and recognition of the comprehensiveness, scalability, and extensive experimental validation of MieDB-100k. Below, we provide point-by-point responses to address your insightful questions and concerns.

Q1: Distinguishing failures in medical image understanding from failures to follow the specific mask-overlay format in Perception tasks.

We agree that disentangling medical understanding from format-following is critical. Both our evaluation metric design and experimental observations directly rule out format unfamiliarity as the primary cause of baseline failures:

  1. Visual Evidence: Qualitative examples (Fig. 4 and Appendix F) confirm that baseline models successfully produce colored overlays of the requested hue, but paint them onto incorrect anatomical structures.
  2. Modality-Wise Control Experiment: As shown in Appendix E.1 (Fig. 10), under the exact same mask-overlay prompt format, baseline models can achieve strong DICE scores (~0.60–0.80) on optical modalities resembling natural images. However, their performance collapses on non-optical clinical modalities (e.g., CT, MRI, Ultrasound, OCT). This serves as a natural control experiment: general-purpose models fully understand and can execute the color-overlay instruction, and their failure on radiological scans stems genuinely from a lack of specialized medical domain knowledge.
  3. Tolerance to Transparency/Color Drift via Least-Squares Alpha De-blending: Baselines are not penalized for failing to match the exact ground-truth transparency. As detailed in Appendix D.1 (Eq. 2), we recover the binary mask by projecting the observed pixel color shift onto the target color vector via a least-squares solution, binarized at a generous threshold. Thus, as long as a model paints the requested primary color (Red/Green/Blue), our extraction algorithm accurately recovers the spatial mask.

Q2: Validating the clinical correctness of synthetic lesion addition and removal targets beyond visual plausibility.

We ensure clinical and pathological fidelity—especially for lesion addition—through our asymmetric bidirectional construction design and multi-stage clinical verification:

  • Target Pathology in "Lesion Addition" is 100% Real Clinical Data: we never synthesize lesions during data construction. As described in Sec. 3.2.2, we start with a real clinical lesion-bearing image (L) and inpaint only the masked lesion region using a modality-specific expert model trained exclusively on healthy anatomy to obtain the healthy counterfactual (H). For Lesion Addition tasks, H serves as the input and the real clinical image L serves as the ground-truth target O. Consequently, the editing targets for lesion addition preserve 100% authentic, clinically verified pathology and anatomical relationships by construction.
  • Objective Algorithmic Verification: For the synthesized healthy tissue (H, the target for lesion removal), before human review, we apply strict rejection sampling (Sec. 3.2.2 Stage III & Appendix B.3). For radiological modalities, we use 5-fold cross-validated nnUNet models to objectively verify that zero residual lesion features remain detectable in the inpainted parenchyma.
  • Rigorous Clinical Audit & Turing Test: All auto-filtered pairs underwent full manual inspection by three board-certified clinicians (Appendix B.4), who discarded 36.2% (9,294 pairs) showing any anatomical discontinuity or unnatural texture. Furthermore, in a blind Clinical Turing Test (Appendix E.10.1, Tab. 16), clinicians rated 85%–92% of our inpainted healthy images (H) as indistinguishable from or superior to real healthy clinical scans.
AuthorsOpenReview ↗

Reply to Reviewer U7G1 (2/2)

To provide fine-grained clinical verification beyond Good/Fair/Poor ratings, we conducted an additional structured audit by 3 clinical experts across the Modification benchmark reference pairs on a 5-point Likert scale (1 = Clinically Unacceptable, 3 = Acceptable with Minor Flaws, 5 = Clinically Indistinguishable from Native Anatomy):

ModalityLesion EradicationAnatomical ContinuityHealthy Tissue FidelityArtifact FreeOverall Mean
CT4.824.774.784.754.78
MRI4.854.764.834.854.82
X-ray4.924.884.944.924.92
Ultrasound4.904.904.884.884.89
Dermoscopy4.884.794.834.764.82
Endoscopy4.764.784.804.764.78
All Modalities (Mean)4.864.814.844.824.83 / 5.00

Our benchmark counterfactuals achieve a high overall clinical fidelity of 4.83 / 5.00, with every dimension scoring ≥4.75\ge 4.75 across all six modalities. Notably, even on structurally complex cross-sectional scans (CT: 4.78, MRI: 4.82) and specular-heavy Endoscopy (4.78), clinicians confirm that the inpainted references faithfully preserve native anatomical continuity and physical signal characteristics.


Q3: Evidence that gains from joint Perception and Modification training arise from improved medical understanding rather than general multi-task regularization.

We provide three lines of evidence demonstrating that the performance gains stem specifically from semantic medical grounding rather than generic multi-task regularization:

  1. Additional Experiment on M+T: To test this directly, we trained an M+T model and compared it against P+M model (both with controlled data size of 40K and the same Modification samples). Pairing M with T provides multi-task training diversity yet yields negligible improvement on Modification (55.9 →\rightarrow 56.3), while incorporating Perception significantly boost the Modification performance (55.9 →\rightarrow 61.2). This isolates semantic medical understanding (P) from generic multi-task regularization.
  2. Zero-Shot Cross-Task Concept Transfer (Sec. 4.5, Fig. 5): We evaluated OmniGen2-MIE on Bone Metastasis Addition and Removal in CT. Crucially, "bone metastasis" was completely excluded from Modification training data and was only seen by the model during Perception training. Despite never seeing how to add or remove bone metastasis, OmniGen2-MIE successfully localizes the bone and synthesizes/removes the lesion with high clinical fidelity, vastly outperforming the baseline. Generic multi-task regularization cannot explain zero-shot transfer of a specific unseen clinical concept from localization to generation.
  3. Empirical Explanation: Unlike masked inpainting, text-guided Lesion Addition/Removal requires the model to implicitly localize the target organ or pathology before modifying pixels. Training on Perception explicitly injects pixel-level anatomical grounding into the DiT backbone, directly preventing the "right-edit-in-the-wrong-location" failures common in M-only and baseline models.

We hope these clarifications thoroughly resolve your concerns.

Thank the authors for the detailed rebuttal and additional experiments. It has addressed most of my concerns. I will keep my score.

Reviewer_AtYZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MieDB-100K, a large-scale dataset for text-guided medical image editing, comprising three task categories: Perception, Modification, and Transformation. The dataset is constructed using modality-specific expert models and rule-based synthesis, followed by manual quality inspection to ensure clinical fidelity. Experimental results demonstrate that models trained on MieDB-100K outperform existing open-source and proprietary models and exhibit strong generalization capabilities.

Critical strengths
  1. The paper addresses an important problem in medical image editing, where high-quality training data remain limited and general-purpose generative models may struggle with domain-specific anatomical and clinical requirements.
  2. The dataset is large-scale and diverse.
  3. The authors evaluate multiple open-source and proprietary generative models.
Critical weaknesses
  1. The proposed dataset construction pipeline mainly combines existing techniques, including pretrained expert models, rule-based synthesis, and manual inspection. The methodological novelty beyond dataset scale and diversity is limited.
  2. The dataset relies heavily on synthetic medical images generated through lesion removal and inpainting. It remains unclear whether these images preserve anatomical consistency and clinical plausibility. More rigorous expert validation is needed.
  3. The evaluation mainly relies on PSNR, SSIM, DICE, and VLM-based rubric scores, which may not adequately capture clinically significant errors. The reliability of automated evaluation against expert clinical judgments requires further investigation.
  4. Despite covering multiple medical modalities, the paper lacks comprehensive analysis of dataset imbalance and performance across individual modalities.
Questions for authors
  1. Compared with Med-banana-50K and MedGEN-Bench, what are the key methodological contributions of MieDB-100K beyond its larger scale, broader modality coverage, and additional task categories? Could the authors provide a more systematic comparison using matched training settings?
  2. What are the individual contributions of modality-specific expert models, rejection sampling, and manual inspection? Could the authors provide ablation experiments investigating how each component affects data quality and downstream performance?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer AtYZ (1/3)

We sincerely thank reviewer for the constructive evaluation and for recognizing the importance of the problem we address, the scale and diversity of MieDB-100k, and our comprehensive benchmarking across open- and closed-source models. Below, we address each weakness and question in detail.


Methodological Contributions Beyond Scale and Matched-Setting Comparison (W1 & Q1)

While scale and diversity (10 modalities, 63 targets) are foundational for dataset works, MieDB-100k introduces three core methodological contributions over Med-banana-50K and MedGEN-Bench:

  1. Unified Task Formulation Bridging Medical Understanding and Generation: Prior datasets are restricted to Modification or low-level Transformation. We are the first to formulate medical image understanding (Perception) within a unified text-guided editing paradigm alongside Modification and Transformation.
  2. Controlled Empirical Discovery of Understanding–Generation Synergy (Appendix E.4, Tab. 11): Holding the total training budget strictly constant at 40,000 triplets, joint training on Perception and Modification (20k P + 20k M) outperforms single-paradigm training (40k M-only) while retaining strong anatomical localization. This indicates a structural cross-task grounding effect rather than a data-scale effect.
  3. Asymmetric Counterfactual Synthesis Paradigm (Preserving 100% Real Pathology): Unlike Med-banana-50K—which directly distills a general-purpose model (Gemini) to synthesize lesions from scratch—we formulate an asymmetric bidirectional paradigm (Sec. 3.2.2). Because healthy parenchyma follows much more tractable anatomical priors than pathology, we train modality-specific expert models exclusively to inpaint healthy tissue. By reversing pairs for Lesion Addition, the target images remain 100% authentic clinical lesion scans, structurally eliminating synthetic pathology hallucinations.

Matched-Setting Comparison with Med-banana-50K: To provide a systematic, matched-setting comparison, we finetuned the exact same OmniGen2 (7B) backbone on Med-banana-50K (~50k Modification triplets) using the identical training recipe and compared it against our models:

Training Dataset (Backbone: OmniGen2 7B)#Mod. TripletsCuration PipelineModification (Rubric-S ↑)
Untouched Baseline (Zero-shot)0—29.1
Med-banana-50K (In-domain finetuning)~50kGemini distillation41.5
MieDB-100k (M-only full, Tab. 3)41.6kOurs (Expert Inpaint + Filter + Audit)57.5
MieDB-100k (Full P + M + T)41.6k (+P,T)Ours (Unified P + M + T Curation)65.9

Under matched Modification-only settings with comparable scale (~41.6k vs. ~50k), our M-only model surpasses OmniGen2-MedBanana by +16.0 Rubric-S (57.5 vs. 41.5), and our full unified dataset widens the lead to +24.4 (65.9 vs. 41.5).

AuthorsOpenReview ↗

Reply to Reviewer AtYZ (2/3)

Expert validation of the synthetic Modification data (W2)

We would like to clarify that comprehensive expert and distributional validations are already established in our study, and we further supplement them with a structured clinical audit.

Existing Multi-Faceted Validation in the Paper:

  1. 100% Full-Population Expert Review (Appendix B.4 & Fig. 6): Every single Modification pair in MieDB-100k (all 25,677 auto-filtered candidates, not just a small sample) was reviewed by three board-certified clinicians under Good/Fair/Poor criteria, achieving >76% unanimous (3-out-of-3) agreement and discarding all 9,294 Poor pairs.
  2. Blind Clinical Turing Test (Appendix E.10.1, Tab. 16): Across 100 randomized side-by-side pairs of real healthy scans vs. our inpainted healthy scans, three clinical experts rated 85%–92% of our generated images as Same (indistinguishable quality, ≥77%) or Good (more realistic than real scans, 4%–9%).
  3. Feature Distribution Consistency: The Maximum Mean Discrepancy (MMD2MMD^2, RBF kernel on Inception features) between our inpainted healthy images and real healthy images is 0.021 across all six Modification modalities, matching the real-vs-real baseline variability (0.017) computed between random splits of real healthy scans.

Additional Structured Clinical Audit on Benchmark References: To provide fine-grained clinical verification beyond Good/Fair/Poor ratings, we conducted an additional structured audit by 3 clinical experts across the Modification benchmark reference pairs on a 5-point Likert scale (1 = Clinically Unacceptable, 3 = Acceptable with Minor Flaws, 5 = Clinically Indistinguishable from Native Anatomy):

ModalityLesion EradicationAnatomical ContinuityHealthy Tissue FidelityArtifact FreeOverall Mean
CT4.824.774.784.754.78
MRI4.854.764.834.854.82
X-ray4.924.884.944.924.92
Ultrasound4.904.904.884.884.89
Dermoscopy4.884.794.834.764.82
Endoscopy4.764.784.804.764.78
All Modalities (Mean)4.864.814.844.824.83 / 5.00

Our benchmark counterfactuals achieve a high overall clinical fidelity of 4.83 / 5.00, with every dimension scoring ≥4.75\ge 4.75 across all six modalities. Notably, even on structurally complex cross-sectional scans (CT: 4.78, MRI: 4.82) and specular-heavy Endoscopy (4.78), clinicians confirm that the inpainted references faithfully preserve native anatomical continuity and physical signal characteristics.


Individual Contributions and Ablation of the Three Pipeline Components (Q2)

To strictly isolate the individual contribution of each pipeline component, we evaluate the intermediate Modification datasets produced at each cumulative stage of our pipeline. For each stage, we measure (i) Data Distributional Fidelity (MMD2↓\text{MMD}^2 \downarrow) (RBF kernel on Inception features between generated healthy images HH and real healthy images. The Real-vs-Real lower bound across random splits of real healthy images is about 0.017) and (ii) Downstream Editing Performance after finetuning OmniGen2 on the respective data:

Pipeline StageMean MMD2\text{MMD}^2 vs. Real (↓\downarrow)Rubric-S (↑\uparrow)
0. Zero-shot Gen (Vanilla Pretrained FLUX)0.16027.8
1. BeforeFilter (+ Modality-Specific Experts)0.06143.4
2. BeforeManual (+ Auto Rejection Sampling)0.03755.0
3. MieDB-100k (+ Manual Clinician Audit)0.02157.5

These quantitative results directly display the individual contributions of each stage and justify our data curating pipeline.

AuthorsOpenReview ↗

Reply to Reviewer AtYZ (3/3)

Clinical Validity of Evaluation Metrics and Alignment with Expert Judgment (W3)

We address the clinical validity and expert alignment of our evaluation protocol across each of the three perspectives:

  • Perception (Deterministic & Verifiable): Because ground-truth organ/lesion segmentation masks are annotated by clinical experts in the source repositories, spatial overlap and Perception Accuracy directly and deterministically measure whether a model localizes the exact clinical target without altering background anatomy (B-PSNR/B-SSIM).
  • Modification (Strong Alignment with Clinical Experts & Cross-Evaluator Robustness): (1) Our VLM scoring rubric (Appendix D.2.1) explicitly evaluates 8 clinical dimensions, including Anatomical Plausibility & Clinical Coherence, Edit Area Morphology, and Intensity/Attenuation Consistency. (2) Among our three clinical evaluators, pairwise Spearman inter-rater correlations are 0.963, 0.989, and 0.957. The VLM Rubric-S exhibits a strong correlation with human clinical preference rankings (-0.905 at the model level, and -0.687 on average at the case level), as well as high cross-VLM stability between GPT-5.2 and Gemini-3-flash (ρ=0.951, Appendix D.2.3, Tab. 8).
  • Transformation (Downstream Clinical Task Validation Beyond PSNR/SSIM): We fully agree that PSNR/SSIM alone cannot capture all clinically significant structural errors, although they remain standard necessary metrics for low-level restoration. That is precisely why we also conducted a downstream clinical segmentation validation on CT Metal Artifact Removal (Appendix E.6, Tab. 12). By applying a pretrained all-organ segmentation model to the restored CT scans and measuring DICE against ground-truth anatomical masks, we directly evaluate whether the edited images preserve clinically diagnostic structures The DICE improvement confirms that training on MieDB-100k not only improves pixel-level PSNR/SSIM, but also faithfully restores anatomical structures required for downstream clinical diagnosis.

Dataset imbalance and per-modality performance (W4)

  1. Per-Modality Performance Breakdown Across All Three Perspectives: While Appendix E.1 (Fig. 10) reports the modality-wise breakdown for Perception, we supplement the per-modality evaluation for Modification (Rubric-S) and Transformation (PSNR / SSIM) below:
Perspective / ModalityBagelQwen-Image-EditGPT-Image-1Nano Banana ProOmniGen2 (Base)OmniGen2-MIE (Ours)
Modification (Rubric-S ↑\uparrow)
CT22.519.423.666.323.768.6
MRI32.520.548.458.06.867.6
Ultrasound40.743.058.962.329.465.5
X-ray41.343.161.462.237.264.4
Dermoscopy46.849.763.265.042.166.2
Endoscopy48.845.947.656.536.557.3
Transformation (PSNR / SSIM ↑\uparrow)
CT13.9 / 0.51521.1 / 0.71514.4 / 0.46822.1 / 0.7518.6 / 0.33428.1 / 0.842
MRI14.6 / 0.40217.6 / 0.51814.4 / 0.41018.0 / 0.52111.4 / 0.37218.3 / 0.561

While general-purpose baselines suffer severe performance drops on non-optical clinical modalities compared to optical ones, OmniGen2-MIE achieves consistently strong and balanced improvements across all modalities, including lower-resource ones.

  1. Dataset Imbalance Analysis & Our Two-Stage Mitigation Strategy:

To rigorously quantify the distribution across all 10 modalities and 63 targets, we compute the distribution statistics for both the Training Split and the Benchmark Split:

Split (#Triplets)Max / Min Modality CountMax-to-Min Ratio (↓)Normalized Shannon Entropy (↑)Gini Coefficient (↓)
Training Split (100,870)41,870 (CT) / 274 (Fundus)152.8 : 10.7570.552
Benchmark Split (3,397)1,066 (CT) / 59 (OCT)18.1 : 10.8830.390

The training distribution reflects the inherent availability of public clinical repositories, where 3D volumetric scans (CT, MRI) naturally yield far more 2D slices than 2D optical acquisitions (e.g., Fundus). During training split curation (Fig. 6), we applied quality-control and class-balancing downsampling to prevent extreme dominance by a single source while preserving data scale for pretraining.

Crucially, despite having fewer training samples in the long tail, lower-resource modalities achieve massive performance gains after unified training on MieDB-100k. This demonstrates that unified multi-modality, multi-task training enables strong representation sharing that effectively lifts tail modalities.

Nevertheless, we have listed data imbalance and rare conditions in Appendix H as Limitations and will incorporate the statistics and discussion above into the revised manuscript.


We hope these analyses and additional experiments fully address your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.