acceptodds
Under review as a conference paper at ICLR 2027

MedVision-X: Unifying Medical Image Analysis and Synthesis through Generative Modeling

Abstract

Medical image analysis and synthesis are served by separate specialist models: segmentation networks predict label maps, restoration networks recover intensities, and translation or editing networks synthesize new images, each with its own architecture, objective, and output head. We ask how far a single generative model with one image output pathway can go across these heterogeneous tasks. We present , which casts segmentation, super-resolution, cross-modal translation, virtual immunohistochemical staining, CT denoising, and lesion editing as instruction-conditioned image generation. Masks are rendered as color-coded images with an optional text legend and continuous targets are generated directly, so every task is trained with the same flow-matching objective and decoded by the same image decoder, without task-specific heads or losses. Starting from the SenseNova-Vision unified multimodal model, is adapted in two stages: broad adaptation on 1.04M medical records, followed by balanced joint training on the training folds of 16 benchmarks. Across 14 test benchmarks and 62 scored baseline entries, obtains the best primary endpoint on 8 and ranks second on 2 more. The gains are largest for physically meaningful targets: one checkpoint reduces body HU MAE from 99.0 to 39.1 on low-dose CT and from 83.1 to 55.2 on MRI-to-CT synthesis, and improves PSNR by 4.0 dB on LoDoPaB-CT and 3.2 dB on IXI. Dataset-specific segmentation checkpoints lead on BUSI and Kvasir-SEG, while ISIC, DRIVE, and pixel-fidelity super-resolution remain behind the strongest specialists. A mask round-trip diagnostic shows that the image representation preserves thin structures, locating the remaining gap in conditional spatial precision rather than in the unified output format.

Primary area:
foundation or frontier models, including LLMs
Keywords:
medical image analysis, generative models, unified multimodal models, flow matching, medical image segmentation, image synthesis, modality translation, virtual staining, CT denoising, super-resolution

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_665pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MedVision-X adapts SenseNova-Vision to six medical imaging task families using a shared image-generation pathway. Segmentation masks are encoded as RGB images, while restoration, translation, staining, and editing produce image targets directly. Training consists of broad medical adaptation followed by balanced joint training on benchmark training folds. The paper reports results on 14 benchmarks, with particularly strong reported performance in CT restoration and cross-modal synthesis. Ten non-segmentation benchmarks use one checkpoint, while segmentation uses dataset-specific checkpoints or ensembles.

Critical strengths

Table 2 lists body HU MAE of 39.11 on LDCT projection and 55.2 on SynthRAD2023, both substantially below the strongest listed baselines at 99.0 and 83.11.

Critical weaknesses
  1. Section 6 states that the Stage-A data were not checked against the benchmark test splits. Earlier exposure to test data therefore remains possible. Table 2 also gives no evaluated sample counts. Appendix H does not show whether the updated scores were computed on the same cases across methods.
  2. The training recipe is not tested with controlled ablations. There is no direct comparison of broad medical adaptation, balanced joint training, and single-task adaptation from the same initialization. It is therefore unclear which part of the training strategy brings the gain. Restormer is also tested on LoDoPaB-CT with an out-of-domain checkpoint rather than a target-adapted model.
  3. MVG and UniMedVL are discussed as related medical generalists, but neither is evaluated. The advantage over existing medical generalist models is therefore unclear.
  4. The segmentation results use dataset-specific checkpoints or checkpoint ensembles. ISIC also uses test-time augmentation. The paper does not show that one checkpoint can reach the reported performance across all six task families.
Questions for authors
  1. Did any benchmark test images or patients appear in the Stage-A data? If no overlap audit was done, what evidence shows that Stage A did not expose the model to the test sets?
  2. How many test cases were evaluated for MedVision-X and each comparator in Table 2? Were all rankings computed on exactly the same cases?
  3. What results does the single Stage-B checkpoint achieve on the four segmentation datasets without checkpoint selection, ensembling, or test-time augmentation? Which claims in the paper are based on one checkpoint?
3 comments

Thank you for the review. Section numbers (§) refer to our General Response (Parts 1–3).

Correction of the headline numbers. Your strength cites 39.11 HU on LDCT projection against a 99.0 HU baseline. In re-auditing against the full result matrix we found complete baselines that the submitted tables omitted. On LDCT projection the strongest complete baseline is a target-trained text-guided unified model at 33.46 HU MAE (ours 39.11, bootstrap CI [24.2, 54.2] over 10 test patients), so that comparison is withdrawn. Against every complete baseline, MedVision-X leads on 5 of 14 benchmarks (SynthRAD2023 55.2 vs. 83.1 HU, IXI 25.58 vs. 22.43 dB, LoDoPaB-CT, MIST, BCI), and the revision limits claims of leading to those (§1).

Q1: did any test image appear in Stage A? The Stage-B training folds were audited image by image with exact pixel hashing and a calibrated near-duplicate detector: three ISIC test images are also listed in the training folder of the adopted split (0.79% of the split, effect ≤ 0.008 Dice), and four DRIVE validation eyes are in training while the test set is unaffected. Stage A draws on a large aggregate medical corpus that is not audited image by image. AAPM, whose source includes test patient L333, is marked as exposed and is not among the 14 main benchmarks, and the other benchmarks are described as not audited against Stage A. The stage-wise results bear on exposure: Stage A alone scores 0.358 / 0.658 / 0.088 / 0.011 Dice on ISIC / Kvasir / BUSI / DRIVE, 0.000 Edit Dice on ATLAS, 113.5 FID on BCI and 276.5 HU on SynthRAD, i.e. it does not reproduce test targets, whereas the final scores (0.881 / 0.775 / 0.774 / 0.772 Dice, 56.2 FID, 66.2 HU) arise after Stage B on the audited benchmark training folds (§4.1, §4.4).

Q2: test cases per method and matched cases. Every ranking uses the complete frozen test split, identical for every method and verified against the manifest digest. MedVision-X counts: ISIC 379, Kvasir 100, BUSI 78, DRIVE 20, BraTS-GLI 27, SynthRAD 36 subjects, IXI 57 subjects, AAPM 610, LoDoPaB 3,553, LDCT projection 2,461, fastMRI 8,852, SurgiSR4K 125, BraTS editing 363, ATLAS 21, MIST 4,000, BCI 977 (§2). The partial counts in Appendix H belonged to an earlier snapshot. 95% bootstrap CIs over patients/subjects: SynthRAD 55.2 [50.4, 60.6] HU, IXI 25.58 [25.15, 26.03] dB, Kvasir Dice 0.905 [0.869, 0.934], LDCT 39.1 [24.2, 54.2] HU.

Q3: single Stage-B checkpoint on the four segmentation datasets. The final Stage-B checkpoint, one generation, no checkpoint selection, ensembling or TTA, gives Dice 0.881 / 0.775 / 0.774 / 0.772 on ISIC / Kvasir / BUSI / DRIVE (§2). The claims resting on one checkpoint are the 16-benchmark table in §2, including the single-checkpoint leads on SynthRAD2023 (66.2 vs. 83.1 HU) and IXI (22.99 vs. 22.43 dB). The ensembled and TTA numbers are reported separately as an analysis of inference strategies. The dense segmentation adapter (§3) reaches 0.9396 / 0.9485 / 0.8616 / 0.82595 Dice, above the strongest complete baseline on all four datasets; it is a separate task-specific checkpoint and is not used to claim a single checkpoint across task families.

Recipe ablations. Stage-wise results under one protocol (§4.1): base model Dice 0.617 / 0.776 / 0.470 / 0.097, SynthRAD 2161 HU, IXI 6.26 dB; Stage A 0.358 / 0.658 / 0.088 / 0.011, 276.5 HU, 15.74 dB; Stage B 0.881 / 0.775 / 0.774 / 0.772, 66.2 HU, 22.99 dB. For joint vs. single-task, the segmentation adapter supervised on the four segmentation folds only reaches specialist-level masks on the same frozen backbone, while on continuous restoration and synthesis the jointly trained model leads without task-specific components (§4.2).

Restormer and generalists. The LoDoPaB-CT lead (37.99 vs. 35.53 dB) is against the text-guided unified baseline, not the out-of-domain Restormer checkpoint, and Table 2 includes target-trained CT and SR baselines (§5). MVG and UniMedVL differ from MedVision-X in task coverage, training data and output convention, and a numerical comparison requires identical frozen splits and benchmark-aligned prompts; outside that convention a generalist degenerates the way our own base-model rows do (DRIVE AUC 0.462, BUSI normal-FPR 0.999, SynthRAD 2161 HU).

Thank you for the detailed response. The new results help, but possible test-set overlap in Stage A and the lack of comparisons with other medical generalists still concern me. I am keeping my original score.

AuthorsOpenReview ↗

Response and Followup : the score in light of the full-cohort and single-checkpoint results

Dear Reviewer 665p,

Thank you for clarifying your remaining concerns. We agree that a complete audit of Stage A exposure and direct comparisons with MVG and UniMedVL would strengthen the evaluation. We respectfully ask you to reconsider the evaluaiton in light of the evidence addressing two other questions in your original review: whether the gains hold on the complete test cohorts and whether they require checkpoint selection or ensembling.

1. Stage A exposure and test-cohort coverage

We identified AAPM as exposed because its source includes test patient L333, and we excluded it from the 14 primary benchmarks. Stage A exposure remains unverified for those 14 benchmarks. For Stage B, we audited the training folds using exact pixel hashes and a calibrated near-duplicate detector. The audit identified three overlapping ISIC images among 379 test images, with a maximum effect of 0.008 on mean Dice (§4.4).

The stage-wise evaluation answers a separate empirical question. On all 36 SynthRAD subjects, body HU MAE fell from 276.5 after Stage A to 66.2 after Stage B, compared with 83.1 for the strongest evaluated baseline. On all 57 IXI subjects, PSNR rose from 15.74 to 22.99 dB, compared with 22.43 dB for the strongest evaluated baseline. Both Stage B results use the same fixed checkpoint and one generation per case, without ensembling or test-time augmentation (§§2, 4.1). These results establish that the gains hold on the complete frozen cohorts under a common inference protocol. Whether individual test patients appeared in Stage A remains an open provenance question.

2. What the baseline comparisons establish

The relative ranking of MedVision-X, MVG, and UniMedVL remains open. Our revised claim is narrower: one checkpoint with a shared decoder across six task families achieved better primary-endpoint values than the strongest evaluated baselines on SynthRAD and IXI under single-generation inference. We report results with benchmark-specific inference strategies separately. For those configurations, we corrected the overall lead count to 5 of 14 benchmarks and withdrew the LDCT superiority claim after identifying a stronger target-trained comparator (§1).

Given that the rebuttal directly answers the questions about cohort coverage and checkpoint selection, we ask you to reassess the evaluation. If either remaining limitation calls a specific revised conclusion into question, could you identify the benchmark and claim you consider unsupported? This would allow us to address the precise disagreement before the discussion closes.

Kind regards,
The Authors

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

This paper presents MedVision-X, a unified generative framework for six medical imaging task families: segmentation, super-resolution, CT denoising, cross-modal synthesis, virtual staining, and lesion editing. Built upon SenseNova-Vision, the framework formulates these tasks as instruction-conditioned image generation using a shared flow-matching objective and image decoder, without task-specific prediction heads. The model is trained through two stages: broad medical adaptation on approximately 1.04 million records, followed by balanced multi-task training across 16 benchmarks. Experiments on 14 benchmarks report the best primary metric on 8 tasks, with particularly promising improvements in CT restoration and cross-modal synthesis. However, the model remains less competitive on tasks requiring precise spatial predictions, such as fine-grained segmentation. Additionally, the segmentation results rely on dataset-specific checkpoints or ensembles rather than a single unified checkpoint.

Critical strengths
  1. Broad coverage with promising results on continuous medical targets. The paper demonstrates a shared generative pathway across six medical task families. One checkpoint supports ten non-segmentation benchmarks, with substantial reported improvements in CT accuracy: body HU MAE decreases from 83.11 to 55.2 on SynthRAD2023 and from 99.0 to 39.11 on LDCT projection compared with the listed strongest baselines. If confirmed under fully comparable evaluation, these results would provide useful evidence for adapting general image generators to medical prediction tasks (Table 2).
  2. Thoughtful interpretation of task-specific metrics. The paper distinguishes visual realism from biomarker agreement and lesion realization from preservation of surrounding tissue. It also explicitly acknowledges that perfect outside-mask preservation on ATLAS follows from compositing. These discussions help readers understand the practical limits of the results (Sec. 5.3).
Critical weaknesses
  1. The validity and comparability of the main results remain insufficiently established. Appendix H states that the latest results lack evaluated-sample counts and do not establish full or matched test-cohort coverage; earlier evaluations covered only 17/36 SynthRAD subjects and 20/57 IXI subjects. In addition, Stage-A training data were not audited for overlap with benchmark test sets (Sec. 6; Appendix C). These disclosures are appreciated, but they leave uncertainty about whether the reported improvements hold on the same unseen cases. Some comparisons also use out-of-domain checkpoints—for example, the Restormer baseline supporting the LoDoPaB improvement uses a Gaussian-denoising checkpoint. Missing uncertainty estimates further limit interpretation of small gains, such as the 0.0012 Dice improvement on Kvasir-SEG.
  2. The advance over existing generalists and the benefit of joint training are not sufficiently demonstrated. SenseNova-Vision and Vision Banana already formulate dense prediction through generated image representations, including color-coded segmentation [1,2]. UniMedVL and MVG already unify several medical analysis and generation tasks [3,4]. The contribution therefore rests mainly on medical adaptation and convincing empirical progress. However, the paper provides neither direct comparisons with these medical generalists nor ablations separating the effects of the pretrained backbone, adaptation stages, task balancing, and joint versus single-task training. Moreover, segmentation uses dataset-specific checkpoints and ensembles, leaving the performance of one fixed checkpoint across all six task families unresolved. The small mask round-trip diagnostic supports representational feasibility, but the stronger conclusion that it identifies the cause of segmentation errors is not fully established.
Questions for authors
  1. Can you provide the evaluated case IDs and counts for every method, confirm matched test cohorts, and audit Stage-A/test overlap? How do the main results change when restricted to confirmed unseen patients or slides?
  2. How does one fixed checkpoint perform across all 16 registered benchmarks? Please distinguish its performance from checkpoint ensembling, generation averaging, and postprocessing, and provide uncertainty estimates for the main comparisons.
  3. What evidence isolates the contribution of the medical adaptation recipe? Comparisons with UniMedVL/MVG and controlled ablations of the adaptation stages and joint versus single-task training would materially strengthen the assessment. Reporting training and inference costs would also help interpret the gains.

[1] Han, X., et al. Vision as Unified Multimodal Generation. arXiv:2607.06560, 2026.

[2] Gabeur, V., et al. Image Generators are Generalist Vision Learners. arXiv:2604.20329, 2026.

[3] Ning, J., et al. UniMedVL: Unifying Medical Multimodal Understanding and Generation through Observation-Knowledge-Analysis. arXiv:2510.15710, 2025.

[4] Ren, S., et al. Medical Vision Generalist: Unifying Medical Imaging Tasks in Context. arXiv:2406.05565, 2024.

3 comments

Thank you for the careful review. Section numbers (§) refer to our General Response (Parts 1–3).

Correction of the headline numbers. Your strengths cite 8 benchmarks with the best primary metric and a 99.0 → 39.11 HU improvement on LDCT projection. In re-auditing against the full result matrix we found complete baselines that the submitted tables omitted. Against every complete baseline, the submitted configurations lead on 5 of 14 benchmarks (SynthRAD2023, IXI, LoDoPaB-CT, MIST, BCI). On LDCT projection the strongest complete baseline is a target-trained text-guided unified model at 33.46 HU MAE (ours 39.11, bootstrap CI [24.2, 54.2] over 10 test patients), so the 99.0 → 39.1 HU claim is withdrawn. The SynthRAD2023 lead (55.2 vs. 83.1 HU) and IXI lead (25.58 vs. 22.43 dB) hold (§1).

Q1: coverage, matched cohorts, overlap. Every row uses the complete frozen test split, verified against the manifest digest; the earlier partial counts in Appendix H belonged to an earlier snapshot, and counts are now in every table (SynthRAD 36 subjects, IXI 57 subjects, full test sets elsewhere, §2). Stage-B training folds were audited image by image (exact pixel hashing plus a calibrated near-duplicate detector). The only test overlap found is three ISIC images, 0.79% of the split, which shift the mean Dice by at most 0.008. Four DRIVE validation eyes are in training; the DRIVE test set is unaffected. Stage A draws on a large aggregate corpus; AAPM, whose source includes test patient L333, is marked as exposed and is not among the 14 main benchmarks, and the other benchmarks are described as not audited against Stage A. Stage A alone scores 0.358 / 0.658 / 0.088 / 0.011 Dice on the four segmentation sets, 0.000 Edit Dice on ATLAS, 113.5 FID on BCI and 276.5 HU on SynthRAD, so it does not reproduce test targets, and the benchmark-level scores arise after Stage B on the audited training folds (§4.4).

Q2: one fixed checkpoint and uncertainty. The final Stage-B checkpoint, one generation, no TTA, checkpoint ensembling, seed averaging, residual correction or compositing, on all 16 test sets (§2): Dice 0.881 / 0.775 / 0.774 / 0.772 on ISIC / Kvasir / BUSI / DRIVE; SynthRAD 66.2 HU (best baseline 83.1); IXI 22.99 dB (best baseline 22.43). Ensembling and averaging add gains on most benchmarks (e.g. Kvasir 0.775 → 0.905), and the revision reports them in a separate table. 95% bootstrap CIs over patients/subjects: SynthRAD 55.2 [50.4, 60.6] HU, IXI 25.58 [25.15, 26.03] dB, Kvasir Dice 0.905 [0.869, 0.934], LDCT 39.1 [24.2, 54.2] HU. The 0.0012 Kvasir margin is within these intervals and is withdrawn together with the Kvasir ranking (4 of 7 after correction).

Q3: contribution of the recipe, generalists, costs. The architecture and objective are inherited from SenseNova-Vision, and the revision positions MedVision-X as a study of the generated-image interface for medical tasks, building on SenseNova-Vision, Vision Banana, MVG and UniMedVL. Stage-wise results (§4.1): base model Dice 0.617 / 0.776 / 0.470 / 0.097, SynthRAD 2161 HU, IXI 6.26 dB; Stage A 0.358 / 0.658 / 0.088 / 0.011, 276.5 HU, 15.74 dB; Stage B 0.881 / 0.775 / 0.774 / 0.772, 66.2 HU, 22.99 dB. For joint vs. single-task, the segmentation adapter supervised on the four segmentation folds only reaches 0.9396 / 0.9485 / 0.8616 / 0.82595 Dice on the same frozen backbone, above the strongest complete baseline on all four, while on continuous tasks the jointly trained model leads without task-specific components (§3, §4.2). MVG and UniMedVL differ in task coverage, training data and output convention, and a numerical comparison requires identical frozen splits and benchmark-aligned prompts (§5). Cost: 14.61B stored parameters (about half active per token), about 2.2 s for one 20-step 512×512 CT slice on a GH200, about 330 A100 GPU-hours for Stage B (2000 updates).

Out-of-domain Restormer. The LoDoPaB-CT lead (37.99 vs. 35.53 dB) is against the text-guided unified baseline, not the out-of-domain Restormer checkpoint.

Mask round-trip diagnostic. The DRIVE round trip (0.976 → 0.963 Dice after the VAE round trip) shows that the representation carries thin structures. The revision states it as a feasibility result; the adapter in §3 adds that the frozen backbone supports specialist-level masks.

AuthorsOpenReview ↗

Followup on rebuttal

Dear Reviewer; 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.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer;

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.

Authors

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

The paper presents a unified generative framework for six medical imaging tasks, including segmentation, super-resolution, CT denoising, cross-modal synthesis, virtual staining, and lesion editing. It builds on SenseNova-Vision and represents different tasks as instruction-conditioned image generation with a shared flow-matching objective.

The idea is interesting, and the results are promising. However, several aspects of the methodological contribution and experimental evaluation require further clarification.

Critical strengths

x. The paper potentially reduces the dependence on separate task-specific models through a shared generative framework.

x. The scope is relatively broad, covering several imaging modalities and six task families. The results on CT restoration and MRI-to-CT synthesis are particularly encouraging.

x. The authors include additional measures for segmentation, virtual staining, and lesion editing.

Critical weaknesses

x. The model tries to do everything and the methodological novelty is somewhat limited

The proposed framework largely inherits its architecture and generative formulation from SenseNova-Vision. The main contributions appear to be the medical task representation and the two-stage adaptation strategy. The paper does not include sufficient ablation studies to assess the contribution of balanced multi-task training or the shared output representation.

From my view, I would suggest comparing Stage A and Stage B separately, as well as multi-task training against task-specific fine-tuning under comparable settings. This would help demonstrate whether the shared training strategy provides benefits beyond adapting an existing large generative model to medical datasets.

x. The central claim of a single unified model is not fully supported by its experimental protocol

A major motivation is to replace multiple specialist models with one shared model. However, the four segmentation benchmarks use dataset-specific checkpoints, including ensembles of three or four checkpoints for BUSI and Kvasir-SEG. Only the remaining ten benchmarks are evaluated using the same checkpoint.

So, why not one fixed checkpoint on all 14 benchmarks, without dataset-specific checkpoint selection or ensembling? The current results could then be retained as an additional evaluation of task-specific inference strategies.

x. The comparison with existing methods is not sufficiently controlled

The paper compares MedVision-X with a mixture of target-trained models, publicly released checkpoints, and reimplementations. Some baseline checkpoints are applied outside their original training domains, while MedVision-X benefits from extensive medical pretraining and a large model capacity. The inference settings also differ substantially. The computational cost and inference budgets of the competing methods are not systematically compared. Direct comparisons with closely related medical generalist frameworks, such as MVG and UniMedVL, are missing.

x. Potential training-data overlap and incomplete evaluation coverage

The authors state that the Stage-B training data exclude the benchmark test splits. However, the much larger Stage-A dataset has not been fully audited for overlap with the evaluation data. In addition, the updated evaluation results do not clearly establish the number of test samples evaluated for each benchmark. Earlier results were based on partial test-set coverage for several datasets.

x. Limitation in 2D method

2D approach is clearly not sufficient to be SOTA. Most latest works work on 2.5D or 3D for generation quality in 3D.

Questions for authors

Some minor comments

x. The preprocessing and intensity normalization procedures deserve further explanation, particularly the use of 8-bit image representations for tasks evaluated in physical HU units.

x. The computational requirements of the 14.6B-parameter model should be reported more clearly. A unified framework is attractive, but its practical benefit also depends on inference efficiency and deployment cost.

3 comments

Thank you for the constructive review. Section numbers (§) refer to our General Response (Parts 1–3).

Novelty and ablations. The architecture and generative formulation are inherited from SenseNova-Vision. The contribution of this work is the medical task representation (RGB-coded masks, HU encoding, instruction format, balanced multi-task sampling) and its evaluation across six task families. Following your suggestion, we compare the stages separately under one protocol (§4.1):

ISICKvasirBUSIDRIVESynthRAD HU↓IXI dB
SenseNova-Vision base0.6170.7760.4700.09721616.26
Stage A0.3580.6580.0880.011276.515.74
Stage B0.8810.7750.7740.77266.222.99

Stage B produces benchmark-level accuracy, and Stage A supplies broad medical initialization on continuous tasks. For multi-task vs. task-specific training: the segmentation adapter, supervised on the four segmentation folds only, reaches 0.9396 / 0.9485 / 0.8616 / 0.82595 Dice (ISIC / Kvasir / BUSI / DRIVE) on the same frozen backbone, whereas the jointly trained generative path gives 0.881 / 0.775 / 0.774 / 0.772. On continuous restoration and synthesis, where the generated image is the native output, the jointly trained model leads on SynthRAD2023 and IXI with no task-specific component. We read this as transfer through the shared generation pathway on continuous targets, with the segmentation gap sitting in the generative mask path rather than in the backbone.

One fixed checkpoint on all benchmarks. We adopt it as the main protocol (§2). The final Stage-B checkpoint, one generation and no post-processing, is evaluated on all 16 test sets and leads on SynthRAD2023 (66.2 vs. 83.1 HU) and IXI (22.99 vs. 22.43 dB). The dataset-specific, ensembled and TTA results are retained as a separate analysis of inference strategies, as you suggest. The segmentation adapter (§3) is reported as a separate task-specific extension.

Controlled comparisons and cost. Baselines are marked [T] target-trained, [C] released checkpoint or [R] re-implementation, and claims of leading are limited to the 5 of 14 benchmarks where MedVision-X leads the strongest complete baseline (§1). MedVision-X stores 14.61B parameters (7.62B understanding/shared, 6.53B generation experts, 0.40B ViT), about half active per token. One 20-step generation of a 512×512 CT slice takes about 2.2 s on a GH200, and Stage B (2000 updates) costs about 330 A100 GPU-hours. The main table uses a single generation per sample. MVG and UniMedVL differ from MedVision-X in task coverage, training data and output convention, and a numerical comparison requires identical frozen splits and benchmark-aligned prompts (§5).

Training-data overlap and coverage. Every row covers the complete frozen test split, verified against the manifest digest, with counts in the tables (§2). Stage-B training folds were audited image by image with exact pixel hashing and a calibrated near-duplicate detector; the only test overlap found is three ISIC images (0.79% of the split, effect ≤ 0.008 Dice), and four DRIVE validation eyes are in training while the test set is unaffected. Stage A is described as not audited. Stage A alone scores 0.358 / 0.658 / 0.088 / 0.011 Dice on the four segmentation sets, 0.000 Edit Dice on ATLAS and 276.5 HU on SynthRAD, so it does not reproduce test targets, and the benchmark-level scores arise after Stage B on the audited training folds (§4.4).

2D formulation. The 2D formulation is the scope of this work, and the paper makes no 3D-consistency claim.

8-bit representation and HU. CT is encoded as (HU + 1024)/4095 in 8 bits (16 HU per level). A perfect 8-bit predictor incurs 4.0 HU body MAE, an order of magnitude below our 39–66 HU error, and evaluation outputs are kept in floating point.

thank you for the rebuttal. If there is not much technical contribution for ICLR, I am not sure including more dense prediction tasks into the generative modeling framework would be relevant for the medical community. I am not seeing a clear benefit of unifying all tasks without discussing the continual learning setting (as when having more tasks beyond the initial ones or new datasets, one would need to work on adaptation of the unified model)

I am concerning its practical impact (that's why I mentioned about 3D setting as a lot of dense prediction tasks are 3D).

AuthorsOpenReview ↗

What the fixed-checkpoint results establish about unification*

Dear Reviewer aBPi,

Thank you for clarifying your concern. We agree that adding more tasks, by itself, would say little about the value of unification. Your suggestion to evaluate one fixed checkpoint provides a direct test of our central question: can the same weights and image decoder produce both discrete anatomical masks and quantitatively accurate continuous images across medical imaging tasks?

1. What one checkpoint can achieve. MedVision-X inherits the SenseNova-Vision architecture. Our contribution is the way we represent medical targets and train them together under one flow-matching objective, including color-coded masks and CT intensities encoded for evaluation in Hounsfield units. Following your suggestion, we evaluated one fixed checkpoint on all 16 test sets, using one generation per input, without checkpoint selection, ensembling, or postprocessing. On SynthRAD2023, that checkpoint achieved 66.2 HU body MAE, compared with 83.1 HU for the strongest evaluated baseline across all 36 test subjects. The same weights achieved 22.99 dB PSNR on IXI, compared with 22.43 dB across all 57 subjects (§2). The practical benefit demonstrated here is specific: these two results can be obtained while retaining a single set of weights for all six task families. Whether joint training improves on compute-matched single-task training remains open.

2. What happens when a new task or dataset arrives. We agree that this is an important question for a model intended to evolve after training. Our adapter experiment provides one relevant observation. With the medical backbone frozen, training a dense segmentation adapter on the four segmentation training folds raised Kvasir-SEG Dice from 0.775 to 0.949, while preserving the original generative pathway (§3). This shows that the learned backbone can support a specialized output for an existing task. A continual-learning evaluation would require introducing a previously unseen task or dataset and measuring both performance on the addition and retention on earlier tasks after each update. We agree that this distinction belongs in the discussion.

3. Where 2D evaluation is informative. Your concern about 3D is particularly important for MRI-to-CT synthesis. Body HU MAE measures intensity accuracy, but even a low error can coexist with anatomical discontinuities between adjacent slices. Our SynthRAD result therefore supports a claim about intensity accuracy on the evaluated subjects. Volumetric consistency requires a separate evaluation. The 2D formulation also has direct relevance to tasks whose benchmark targets are themselves images, including endoscopic polyp masks, retinal vessel maps, and histopathology staining.

We would appreciate it if you could reconsider your assessment in light of the fixed-checkpoint experiment you proposed. It establishes where a shared decoder is already competitive and, through the segmentation results, where a task-specific output still adds substantial value.

Kind regards,
The Authors

Reviewer_UPb1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper explores a unified flow-matching generative model for multiple medical image analysis and synthesis tasks. Built on SenseNova-Vision, the method formulates segmentation, super-resolution, CT denoising, cross-modal synthesis, virtual staining, and lesion editing as instruction-conditioned image generation, with a shared generation pathway and image decoder. The model is first adapted on approximately 1.04 million medical records and then jointly trained on the training splits of 16 benchmarks using task-balanced sampling.

Critical strengths

Medical image analysis and synthesis typically rely on different models and training objectives. This paper explores a shared flow-matching generation pathway for six task families, spanning discrete label prediction, continuous intensity restoration, and local image editing. The research question is meaningful, and the unified task interface has practical value.

Critical weaknesses
  1. The model’s competitiveness is uneven across tasks, with clear weaknesses on several benchmarks. According to the main table and Appendix H, none of the reported metrics on ISIC 2018 or DRIVE is the best: Dice and IoU on ISIC both rank sixth among seven evaluated methods, while Dice, IoU, and AUC on DRIVE all rank third. For super-resolution, although fastMRI SR achieves the best LPIPS, NMSE, PSNR, and SSIM all rank only fourth among five methods; none of the metrics on SurgiSR4K is the best. Edit Dice ranks second on both lesion-editing datasets. On ATLAS v2, FID ranks only third, and the best outside-region preservation scores result from mask compositing rather than directly demonstrating the model’s own preservation ability. For BCI virtual staining, although FID and KID are the best, HER2 agreement ranks only second and SSIM ranks last among four methods, indicating that the advantage in distributional realism does not fully translate into consistency with paired images and marker expression. Thus, the advantages on restoration and synthesis do not consistently extend to fine-grained segmentation, super-resolution, and lesion editing. Meanwhile, segmentation results rely on dataset-specific checkpoints, checkpoint ensembles, or test-time augmentation, and some non-segmentation tasks use averaging over eight generations, residual correction, or mask compositing. Consequently, the final scores reflect both the model and the inference strategies. Without complete results using a fixed checkpoint and a uniform, simplified inference setup, the evidence is insufficient to establish consistent competitiveness across tasks with a single set of weights. Generalization to unseen data distributions also lacks direct validation.

  2. The paper does not provide systematic comparisons involving the original SenseNova-Vision base model, single-task versus joint training, or the two medical adaptation stages. It therefore remains unclear whether joint training yields positive transfer across tasks, whether weaker segmentation, super-resolution, and lesion-editing performance is related to interference between tasks, and to what extent the stronger results arise from the proposed training design, the base model’s existing capabilities, model scale, or additional medical data. Since the architecture and primary optimization objective are inherited from the base model, these controls are particularly necessary to establish the independent contribution of the proposed method.

  3. Although the paper compares against multiple specialist methods, it does not directly compare with relevant medical generalist models such as MVG and UniMedVL, making it difficult to assess its actual progress over existing unified medical multitask approaches. In addition, some specialist baselines use released out-of-domain checkpoints, whereas MedVision-X is adapted on the training splits of the corresponding benchmarks. For example, Restormer on LoDoPaB-CT uses an out-of-domain checkpoint, whose performance may not represent what the architecture could achieve after target-domain training. The current comparisons therefore do not sufficiently establish performance advantages over related generalist models or adequately trained specialists.

Questions for authors

Q1: Can the authors report results from the same final checkpoint on all 14 benchmarks and compare performance using a single generation, without checkpoint ensembling or additional postprocessing? These results would help distinguish the unified model’s own capabilities from gains due to inference strategies. Q2: Can the authors provide results for the original base model, after Stage A, and after Stage B, and compare single-task and joint training under comparable data and training budgets? Can these controls explain whether weaker tasks suffer from interference between tasks and whether stronger tasks benefit from positive transfer? Q3: Can the authors add comparisons with MVG and UniMedVL on benchmarks where tasks and data splits can be aligned, and provide target-domain-trained results for key specialist baselines? These comparisons would directly affect the assessment of the paper’s relative progress and performance advantages. Q4: Can the authors add side-by-side visualizations of the final checkpoint and representative baselines, including magnified segmentation boundaries and thin structures, virtual staining consistency, and successful and failed lesion-editing cases? This would help determine whether the quantitative metrics correspond to actual task quality.

3 comments

Thank you for the detailed review. Section numbers (§) refer to our General Response (Parts 1–3).

Uneven competitiveness across tasks (Weakness 1). We re-audited all comparisons against the full result matrix. Against every complete baseline, the submitted configurations lead the primary endpoint on 5 of 14 benchmarks (SynthRAD2023, IXI, LoDoPaB-CT, MIST, BCI), and the revision restricts all claims of leading to those (§1). On ISIC, Kvasir, BUSI, DRIVE, the SR datasets and lesion editing, MedVision-X is competitive with, but not above, the strongest specialist in the single-checkpoint setting.

Q1: one fixed checkpoint, one generation (§2). The final Stage-B checkpoint with one generation, no TTA, checkpoint ensembling, seed averaging, residual correction, compositing or tiling, evaluated on all 16 test sets with full coverage:

  • Segmentation Dice (ISIC / Kvasir / BUSI / DRIVE): 0.881 / 0.775 / 0.774 / 0.772.
  • SynthRAD2023 HU MAE 66.2 (best baseline 83.1) and IXI PSNR 22.99 dB (best baseline 22.43): the single checkpoint leads on both.
  • LoDoPaB 34.61 dB, LDCT projection 51.7 HU, fastMRI SR 33.78 dB, SurgiSR4K 29.33 dB, MIST FID 64.3, BCI FID 56.2.

Ensembling, seed averaging and residual correction add gains on most benchmarks (e.g. Kvasir 0.775 → 0.905, LoDoPaB 34.61 → 37.99 dB). The revision uses the single-checkpoint protocol as the main table and moves the inference strategies to a separate table. For segmentation, the dense adapter in §3 reaches 0.9396 / 0.9485 / 0.8616 / 0.82595 Dice, above the strongest complete baseline on all four datasets; it is a separate task-specific checkpoint and is presented as a separate extension.

ATLAS and BCI. Without compositing the same checkpoint reaches Edit Dice 0.717 against MedEdit's 0.743, with outside-mask MAE of 0.151; the composited and non-composited rows are both reported, and perfect outside-region preservation is attributed to compositing. On BCI, FID is 39.1 vs. 47.3 for the best baseline; HER2 agreement (second) and SSIM (last among four) are reported alongside.

Q2: base, Stage A, Stage B, joint vs. single-task (§4.1–4.2). Under one protocol: zero-shot base model Dice 0.617 / 0.776 / 0.470 / 0.097 on ISIC / Kvasir / BUSI / DRIVE, SynthRAD 2161 HU, IXI 6.26 dB; Stage A 0.358 / 0.658 / 0.088 / 0.011, 276.5 HU, 15.74 dB; Stage B 0.881 / 0.775 / 0.774 / 0.772, 66.2 HU, 22.99 dB. Stage B produces benchmark-level accuracy and Stage A supplies broad medical initialization on continuous tasks. For joint vs. single-task, the segmentation adapter is supervised on the four segmentation folds only and reaches specialist-level masks on the same frozen backbone, while on continuous restoration and synthesis the jointly trained model leads without task-specific components. The shared generation pathway therefore transfers well to continuous targets, and the segmentation gap sits in the generative mask path rather than in the backbone. The 8-bit representation is not the limiting factor for restoration: a perfect 8-bit predictor of the CT target incurs 4.0 HU MAE, an order of magnitude below our 39–66 HU error.

Q3: generalists and specialists (§5). MVG and UniMedVL differ from MedVision-X in task coverage, training data and output convention. A numerical comparison requires the same frozen splits and benchmark-aligned prompts; outside that convention a generalist degenerates the way our own base-model rows do (DRIVE AUC 0.462, BUSI normal-FPR 0.999, SynthRAD 2161 HU). Table 2 includes target-trained CT and SR baselines (text-guided unified, WIA-LD2ND, VQ-SCD, DiffMSR, GaussM2ASR, HDW-SR). The LoDoPaB lead (+2.46 dB) is against the text-guided unified baseline, not the out-of-domain Restormer checkpoint, and one target-trained baseline beats us on LDCT projection (33.5 vs. 39.1 HU).

Q4: qualitative comparisons. The reported measures target the qualities you list: HD95 and DRIVE sensitivity/AUC for boundaries and thin structures, HER2 agreement alongside FID/KID for staining consistency, and outside-mask MAE alongside Edit Dice for lesion-edit preservation.

Generalization to unseen distributions. Stage-B training folds were audited image by image: the only test overlap is three ISIC images (effect ≤ 0.008 Dice); Stage A is described as not audited rather than unseen (§4.4).

Reviewer_UPb1OpenReview ↗

Assessment of the Additional Experiments

Thank you for the detailed response and the additional experiments. The evaluation with one fixed checkpoint and one generation per input directly addresses my main question. MedVision-X performs well on SynthRAD2023 and IXI, but it still trails the strongest baselines on the primary metrics for segmentation, super-resolution, and lesion editing. The dense adapter substantially improves segmentation using a separate task-specific checkpoint. The comparisons across the base model, Stage A, and Stage B clarify the effects of adaptation, and the target-trained baselines make the restoration comparisons more informative. The response also explains how the reported metrics capture output quality, though the qualitative evidence remains limited. Overall, the fixed-checkpoint results still show uneven performance across tasks, so I will keep my rating at 2.

AuthorsOpenReview ↗

Reconsideration in light of the fixed-checkpoint results*

Dear Reviewer UPb1,

Thank you for confirming that the evaluation with one fixed checkpoint and one generation per input directly addresses your main question. We agree with your assessment that performance varies across tasks. We ask you to consider what that variation tells us about the paper’s central question: how far can one shared image decoder go across six medical imaging task families?

Under the protocol you requested, the same checkpoint achieved 66.2 HU body MAE on SynthRAD2023 across all 36 test subjects, compared with 83.1 HU for the strongest evaluated baseline. On IXI, it achieved 22.99 dB PSNR across all 57 test subjects, compared with 22.43 dB. Both results use one generation per input, without checkpoint selection, ensembling, or postprocessing (§2). They establish that a model trained to perform all six task families can exceed the evaluated baselines on these two synthesis benchmarks using a single set of weights.

We also agree that the gaps deserve equal visibility. On fastMRI super-resolution, the fixed checkpoint achieved 33.78 vs. 36.76 dB PSNR. On ATLAS lesion editing, it achieved 0.717 vs. 0.743 Edit Dice. The segmentation gap was larger on Kvasir-SEG, at 0.7749 vs. 0.9333 Dice. Here the adapter experiment adds a specific finding: a separately trained dense adapter on the frozen medical backbone achieved 0.9485 Dice on the same benchmark (§3). The backbone therefore supports accurate polyp segmentation, even though the shared generative pathway falls well short of that performance. We report the adapter separately throughout.

We acknowledge that the qualitative evidence remains limited, particularly for understanding individual failure cases. The quantitative evidence nevertheless supports a concrete conclusion: one shared decoder achieved the strongest evaluated result on two synthesis benchmarks, and the same controlled evaluation identifies where specialization still offers substantial gains. Since the fixed-checkpoint experiment resolves the main uncertainty in your original review, we respectfully ask whether you would reconsider the evaluation; empirically supported contribution.

Kind regards,
The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response (1/3): complete comparison and single-checkpoint results

We thank all four reviewers. Their reviews converge on four questions: (1) what one fixed checkpoint achieves without inference-time aggregation; (2) which part of the recipe produces the gains (base model, Stage A, Stage B, joint vs. single-task); (3) how the model compares with medical generalists and target-trained specialists; (4) whether the evaluation is sound (sample coverage, Stage-A/test overlap, uncertainty).

We re-audited every comparison against the full result matrix and updated the comparative claims accordingly (§1). We also report single-checkpoint results on all 16 test sets (§2), a segmentation adapter that reaches specialist-level masks (§3), and an analysis of what each training stage and the joint formulation contribute (§4). Parts 2 and 3 of this response contain §3–§5; each reviewer's individual reply follows.

1. Complete comparison against all baselines

With all complete target-trained and released baselines included, the submitted configurations lead the primary endpoint on 5 of 14 benchmarks: SynthRAD2023, IXI, LoDoPaB-CT, MIST and BCI. The leads are largest on continuous physical targets:

  • SynthRAD2023: 55.2 vs. 83.1 body HU MAE.
  • IXI: 25.58 vs. 22.43 dB PSNR.
  • LoDoPaB-CT: 37.99 vs. 35.53 dB PSNR (+2.46 dB, against the text-guided unified baseline).

On the remaining benchmarks MedVision-X is competitive with the strongest specialist:

Dataset (primary)OursStrongest complete baselineRank
Kvasir-SEG (Dice)0.9046EMCAD [T] 0.93334 / 7
BUSI (Dice)0.8271MedCLIPSeg [C] 0.85912 / 7
ISIC 2018 (Dice)0.8918EMCAD [T] 0.92696 / 7
LDCT projection (HU MAE)39.11Text-guided unified [R] 33.462 / 5

On LDCT projection (10 test patients) the 95% bootstrap CI for our result is [24.2, 54.2] HU, which covers the 33.5 HU baseline. The revision updates Table 2, the abstract, the introduction and Appendix H to this complete comparison ([T] target-trained, [C] released checkpoint, [R] re-implementation). Claims of leading are scoped to the benchmarks above, against the strongest complete baseline irrespective of provenance.

2. One fixed checkpoint, one generation

The final Stage-B checkpoint, one generation (seed 42), no TTA, checkpoint ensembling, seed averaging, residual correction, compositing or tiling, on all 16 test sets. Every test sample is covered.

Benchmarkn (test)One ckpt, one generationWith inference strategyBest complete baseline
ISIC 2018 Dice3790.88060.89180.9269
Kvasir-SEG Dice1000.77490.90460.9333
BUSI Dice780.77380.82710.8591
DRIVE Dice200.77170.79110.8252
BraTS-GLI WT Dice270.592——
SynthRAD2023 HU MAE36 subj.66.255.283.1
IXI PSNR57 subj.22.9925.5822.43
AAPM HU MAE61031.024.820.2
LoDoPaB PSNR3,55334.6137.9935.53
LDCT projection HU MAE2,46151.739.133.5
fastMRI SR PSNR8,85233.7835.8636.76
SurgiSR4K PSNR12529.3330.9231.83
BraTS editing Edit Dice3630.44450.44660.4849
ATLAS v2 Edit Dice210.71680.59360.7428
MIST FID4,00064.331.744.9
BCI FID97756.239.147.3

One checkpoint with one generation leads on SynthRAD2023 and IXI, and its ATLAS Edit Dice (0.717) is close to MedEdit's 0.743. Ensembling, seed averaging and residual correction add further gains on most benchmarks. The revision uses this single-checkpoint protocol as the main table and reports the inference strategies in a separate table.

AuthorsOpenReview ↗

General Response (2/3): segmentation adapter and recipe analysis

3. Dense segmentation adapter

To address the segmentation gap raised by all reviewers, we attach a lightweight dense segmentation adapter (task-routed LoRA, spatial-prior CNN and dense mask decoder) to the frozen medical backbone. It is trained on the four segmentation training folds only and leaves the generative path unchanged.

DatasetAdapterStrongest complete baseline
ISIC 2018Dice 0.9396 / IoU 0.8907 / HD95 39.27EMCAD 0.9269 / 0.8728 / 48.46
Kvasir-SEGDice 0.9485 / IoU 0.9099 / HD95 21.42EMCAD 0.9333 / 0.8854 / 27.47
BUSIDice 0.8616 / IoU 0.8030 / HD95 29.93MedCLIPSeg 0.8591 / 0.7917 / 31.86
DRIVEDice 0.82595 / IoU 0.70380 / AUC 0.98095nnWNet 0.82521 / 0.70270 / 0.98085

The adapter exceeds the strongest complete baseline on all four segmentation primary endpoints. It is a separate task-specific checkpoint, so the revision presents it as its own section and keeps the unified-model count in §1 unchanged.

The ISIC test split contains three images (ISIC_0000279, ISIC_0001184, ISIC_0001385) that the adopted MedCLIPSeg split also lists in its training folder; they are 0.79% of the split and shift the mean by at most 0.008 Dice.

4. What the evidence says about the recipe

4.1 Base → Stage A → Stage B

Under the same single-generation protocol and the same test sets:

ISICKvasirBUSIDRIVESynthRAD HU↓IXI dBLoDoPaB dBfastMRI dBATLAS Edit DiceBCI FID↓
SenseNova-Vision base (zero-shot)0.6170.7760.4700.09721616.26————
Stage A (broad medical)0.3580.6580.0880.011276.515.7421.1917.050.000113.5
Stage B (final)0.8810.7750.7740.77266.222.9934.6133.780.71756.2

Stage B produces benchmark-level accuracy. Relative to the base model it adds +0.26 / +0.30 / +0.67 Dice on ISIC / BUSI / DRIVE, reduces SynthRAD HU MAE from 2161 to 66.2 and raises IXI PSNR from 6.26 to 22.99 dB. Stage A supplies the broad medical initialization (SynthRAD 2161 → 276.5 HU, IXI 6.26 → 15.74 dB), and Stage B aligns the model with each benchmark's instruction and output convention.

4.2 Joint training and the role of Stage A

Joint vs. single-task. The six task families share one decoder but differ in label space and output convention (RGB-coded masks, 8-bit continuous intensities, edited images), so cross-task effects have two channels: positive transfer through shared visual priors (edges, contrast, anatomical layout), and interference through competing prompt and output formats. Two observations bear on which dominates.

  • On continuous restoration and synthesis, where the generated image is the native output, the jointly trained model leads on SynthRAD2023 and IXI with no task-specific component (§2).
  • The segmentation adapter (§3) is supervised on the four segmentation folds only and reaches 0.9396 / 0.9485 / 0.8616 / 0.82595 Dice on ISIC / Kvasir / BUSI / DRIVE, against 0.881 / 0.775 / 0.774 / 0.772 for the joint generative path. Segmentation-only supervision on this backbone therefore yields specialist-level masks, so the weaker generative-path segmentation numbers are not a limit of the backbone's representation.

We read this as: the shared generation pathway transfers well to continuous targets, and the segmentation gap sits in the generative mask path rather than in the backbone. The revision states this as an analysis of the measured results.

Role of Stage A. Stage A alone does not follow benchmark instructions or output conventions (segmentation Dice 0.01–0.36, ATLAS Edit Dice 0.000, SynthRAD 276.5 HU), so benchmark-level accuracy is produced by Stage B on the benchmark training folds. Stage A's measurable contribution is the initialization advantage on continuous tasks (e.g. 2161 → 276.5 HU on SynthRAD), and the stage progression in §4.1 is the direct evidence for it.

Base model vs. proposed recipe. The base model's zero-shot failures are of a specific kind (degenerate masks, DRIVE AUC 0.462, BUSI normal-FPR 0.999; out-of-range continuous output, SynthRAD 2161 HU, IXI 6.26 dB), and they disappear after medical adaptation. The architecture and objective are inherited from SenseNova-Vision; the contribution of this work is the medical task representation (mask coding, HU encoding, instruction format, balanced sampling) and its evaluation across six task families.

AuthorsOpenReview ↗

General Response (3/3): coverage, overlap, generalists, cost

4.3 Coverage and uncertainty

Every row covers the complete frozen test split, verified against the manifest digest, and counts are in every table. 95% bootstrap CIs over the patient/subject/slide unit:

  • SynthRAD HU MAE 55.2 [50.4, 60.6]
  • IXI PSNR 25.58 [25.15, 26.03]
  • Kvasir Dice 0.905 [0.869, 0.934]
  • LDCT projection 39.1 [24.2, 54.2]

The primary protocol uses one generation and one seed. On these benchmarks, between-case variation is the dominant source of spread (per-case Dice from near-empty to saturated masks), which is why the intervals resample cases.

4.4 Training/test overlap

  • Stage B training folds were audited image by image with exact pixel hashing and a calibrated near-duplicate detector. The only test overlap found is the three ISIC images in §3 (effect ≤ 0.008 Dice). Four DRIVE validation eyes (25, 29, 31, 39) are in Stage-B training; the DRIVE test set is unaffected.
  • Stage A draws on a large aggregate medical corpus. AAPM, whose source includes test patient L333, is marked as exposed and is not among the 14 main benchmarks. The stage-wise scores bear on the other benchmarks: Stage A alone scores 0.358 / 0.658 / 0.088 / 0.011 Dice on the four segmentation sets, 0.000 Edit Dice on ATLAS, 113.5 FID on BCI and 276.5 HU on SynthRAD, i.e. it does not reproduce test targets, and the benchmark-level scores in §2 arise after Stage B on the audited training folds. The revision describes these benchmarks as not audited against Stage A and presents this stage-wise evidence alongside.

5. Generalists, specialists, cost

  • MVG / UniMedVL. These generalists differ from MedVision-X in task coverage, training data and output convention. A numerical comparison depends on running each under the same frozen splits and benchmark-aligned prompts, because a generalist evaluated outside the benchmark's prompt and output convention degenerates in the way our own base-model rows do (DRIVE AUC 0.462, BUSI normal-FPR 0.999, SynthRAD 2161 HU).
  • Specialists. Table 2 includes target-trained CT and SR baselines (text-guided unified, WIA-LD2ND, VQ-SCD for CT; DiffMSR, GaussM2ASR for fastMRI; HDW-SR for SurgiSR4K). The LoDoPaB lead in §1 is against the text-guided unified baseline, not the out-of-domain Restormer checkpoint. Target-domain training raises a specialist's in-domain score, so the revision's claims of leading are restricted to the 5 benchmarks in §1, where the comparison is against the strongest complete baseline.
  • Cost. MedVision-X stores 14.61B parameters (7.62B understanding/shared, 6.53B generation experts, 0.40B ViT), about half active per token. One 20-step generation of a 512×512 CT slice takes about 2.2 s on a GH200. Stage B (2000 updates) costs about 330 A100 GPU-hours. The main table uses a single generation per sample.
  • Task quality beyond headline metrics. The reported measures target the qualities the reviewers list: boundary error (HD95) and thin-vessel sensitivity/AUC for segmentation, HER2 agreement alongside FID/KID for staining, and outside-mask MAE alongside Edit Dice for lesion editing (0.151 for ATLAS without compositing).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.