{"id":"b9d07566-29d6-402d-8231-173dfc6641fd","arxiv_id":"2412.04106","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A text- and mask-conditioned diffusion model can synthesize training images for MRI modalities without masks, boosting downstream segmentation accuracy compared with augmentation and translation baselines.","lead":"This paper introduces MRGen, a diffusion-based engine that generates synthetic MRI images from text prompts and organ masks, and shows that segmentation models trained on these images improve accuracy on MRI modalities that lack manual annotations. The work also releases MRGen-DB, a large radiology image-text dataset for medical image generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paired-dataset experiments leak test-subject anatomy through source-domain test masks; the headline DSC gains are inflated, though cross-dataset evidence still shows positive but smaller gains.","rationale":"The reader's stated weakest assumption (target modalities must appear in the unlabeled Radiopaedia pretraining corpus) is a real limitation, but it is not the most load-bearing concern for the paper's central quantitative claim. The target modalities evaluated (T1, T2, T2-SPIR, ADC) are standard abdominal MRI protocols that very likely occur in Radiopaedia, and the paper's manual verification and FID results suggest the generator can produce them; so this assumption, while unverified, is not contradicted by the evidence. The more damaging issue is the training/test leakage in paired same-subject dataset pairs. Appendix B.3 explicitly states that the entire source-domain dataset, including test splits, is used to derive masks for synthetic generation, except for MSD-Prostate. For CHAOS and PanSeg, source and target modalities are from the same subjects, so source test masks coincide with target test anatomy. Training the downstream segmentation model on images generated from those masks yields artificially high DSC on the target test set. The clean cross-dataset pairs still show MRGen improving over source-only training and translation baselines, so the paper's main idea is plausible; however, the strongest headline numbers and the '8/10 best' statement are inflated by the leaked settings. The proposed concrete test—recomputing paired results with only source training masks—would settle whether the central claim holds without leakage. Given that the core idea has independent support but the reported magnitude is suspect, the reader's conditional verdict remains appropriate; the paper should also report variance and release code/data as the reader requested. We note Table 5 additionally shows that the largest gains require including real unlabeled target-domain images, so the contribution of the synthetic data engine alone is much smaller than the headline suggests.","tokens_in":23832,"tokens_out":9634,"duration_ms":87798,"concrete_test":"Rerun the four paired settings (CHAOS T1→T2-SPIR, CHAOS T2-SPIR→T1, PanSeg T1→T2, PanSeg T2→T1) using only source-domain training-split masks (e.g., CHAOS T1 train: 1,018 slices; PanSeg T1 train: 14,656 slices) to generate synthetic target data, with all else unchanged, and evaluate DSC on the same target test split. If the gains drop to the level seen in cross-dataset pairs (roughly +10–20 DSC over source-only), the reported +50–60 DSC improvements are largely leakage artifacts. Additionally, for one paired setting, run a control that keeps test-subject masks but removes organ-shape information (e.g., apply a random non-linear warp to the mask) to quantify how much of the gain comes from knowing test anatomy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix B.3 states that for synthetic-data construction, 'we use mask annotations from the entire source-domain dataset (including both training and test sets)' with the sole exception of MSD-Prostate. For the four paired settings (CHAOS T1↔T2-SPIR and PanSeg T1↔T2), the source and target test sets are the same subjects imaged in different protocols, so the source-domain test masks are the target test subjects' ground-truth anatomies. These masks are used to generate synthetic target-domain training images, and the downstream nnUNet/UMamba is trained on image-mask pairs whose masks come from the target test subjects. At evaluation, the model has effectively seen each test subject's organ geometry, inflating DSC. The paper's headline example (CHAOS T1→T2-SPIR nnUNet: 6.90→66.18) is exactly this leaked setting. The clean cross-dataset pairs (LiQA→CHAOS, CHAOS→LiQA, MSD-Prostate→PROMISE12, PROMISE12→MSD-Prostate) show smaller but consistent gains (roughly +10–40 DSC), so the central idea is not falsified, but the 'best in 8/10' claim and the magnitude of improvement over baselines rest substantially on the leaked paired settings.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MRGen, a two-stage diffusion-based data engine that generates MRI images conditioned on text prompts and segmentation masks, together with MRGen-DB, a curated dataset of roughly 250,000 2D slices with modality, region, and organ text annotations and about 18,000 mask-annotated slices. The method is evaluated by using MRGen to synthesize target-modality training images from source-domain masks and then training nnUNet and UMamba on the resulting synthetic pairs across five cross-modality dataset pairs. The authors report that MRGen achieves the best DSC in 8 of 10 settings, with the headline example CHAOS T1 to T2-SPIR nnUNet rising from 6.90 to 66.18.","tokens_in":24066,"tokens_out":5405,"duration_ms":49913,"significance":"If the empirical claims survive a corrected evaluation, the paper makes a useful contribution: it introduces a large-scale radiology image-text dataset, a controllable latent diffusion engine for mask-conditioned MRI synthesis, and an automatic filtering pipeline based on SAM2. The release of code, models, and data would enable direct reproducibility and follow-up work. However, the current evaluation contains a train/test leakage in exactly the settings that drive the headline numbers, so the magnitude and ranking of the reported improvements are not yet established.","major_comments":[{"comment":"The synthetic-data statistics state that masks from the entire source-domain dataset (including both training and test sets) are used as conditions to generate target-domain images, with the only exception being MSD-Prostate. For the paired settings CHAOS T1↔T2-SPIR and PanSeg T1↔T2, the source and target test sets are the same subjects imaged under different protocols. Using source test masks to generate synthetic training pairs therefore trains the downstream nnUNet/UMamba on the target test subjects' ground-truth anatomies, inflating the DSC scores in Table 3. In particular, the headline CHAOS T1→T2-SPIR nnUNet result (6.90 to 66.18) and the 'best in 8 of 10' claim rest on these leaked settings. The clean cross-dataset pairs (LiQA→CHAOS, CHAOS→LiQA, MSD-Prostate→PROMISE12, PROMISE12→MSD-Prostate) show smaller but still positive gains, so the central idea may survive, but the paper must re-run the experiments with source test masks excluded from synthetic-data construction for all settings and recompute the summary claims.","section":"Appendix B.3, Table 7"},{"comment":"Section 4.1 states that 'the test data for segmentation is strictly not used for training our generative data engines to prevent information leakage.' This is contradicted by Appendix B.3, which uses source-domain test masks as conditioning inputs for the generated training pairs. The contradiction is not merely a wording issue: in the paired CHAOS and PanSeg settings, the source test masks coincide with the target test subjects' anatomy, so the segmentation model effectively sees the test geometry during training. The protocol must be clarified and fixed so that no test-subject information enters the synthetic training data.","section":"Sec. 4.1 vs Appendix B.3"},{"comment":"All FID and DSC values are reported as point estimates without error bars, multiple seeds, or significance tests. Some settings use very small test sets (e.g., CHAOS has only 4 test volumes for T2-SPIR), so the reported differences may not be statistically reliable. The paper should report mean and standard deviation over at least three training runs of the downstream segmentation models, or provide an equivalent statistical analysis.","section":"Tables 2 and 3"},{"comment":"The paper uses the term 'zero-shot' and claims that MRGen generalizes to underrepresented modalities lacking mask annotations. However, the text-guided pretraining stage requires the target modality's appearance to be present in the unlabeled Radiopaedia corpus. The paper does not verify whether T2-SPIR, ADC, T1, and T2 actually occur in the pretraining data, so the method's behavior for truly unseen protocols is untested. Please either provide modality-coverage statistics for MRGen-DB or evaluate on a modality that is absent from the pretraining corpus, and adjust the 'zero-shot' terminology accordingly.","section":"Sec. 3.1, Sec. 4.1, Appendix E"}],"minor_comments":[{"comment":"Figure 7 contains the typo 'Groud Truth' and Figure 8 contains 'stuggles'; these should be corrected.","section":"Figure captions"},{"comment":"The header 'Source Datset' contains a typo, and the abbreviations CM., MP., PS., LQ., and PR. are only defined in the table caption; please ensure they are introduced in the text as well.","section":"Table 2"},{"comment":"The column layout of Table 3 is difficult to parse, with the DualNorm column separated from the UMamba and nnUNet groups; consider reorganizing the table so that each segmentation framework is clearly nested under its own header.","section":"Table 3"},{"comment":"DualNorm is evaluated with a customized UNet while all other baselines use nnUNet or UMamba; this makes the DualNorm comparison non-controlled, and the main text should explicitly acknowledge this limitation.","section":"Sec. 4.1"},{"comment":"The comparison between SDM-ft and MRGen is not controlled: MRGen uses a higher latent dimension and a different text encoder, so the individual contributions of these components are not isolated.","section":"Table 4"},{"comment":"The ablation on segmentation performance would be easier to interpret if the source-only nnUNet baseline were included in the table for each setting, as it is currently only reported in Table 3.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The leakage in Appendix B.3 is the main obstacle to publication. If the corrected cross-dataset experiments still show consistent gains, the paper could be a solid contribution to medical image synthesis and segmentation. The 'zero-shot' claim should be substantially tempered unless the authors verify modality coverage in the pretraining corpus and evaluate on a truly unseen protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: MRGen is a real contribution — a large MRI image-text dataset (245K slices) and a two-stage diffusion pipeline that can synthesize target-modality images from masks and text. The clean cross-dataset experiments show that this synthetic data consistently improves segmentation over source-only baselines, so the core idea holds. But the headline numbers, especially the CHAOS T1->T2-SPIR jump from 6.90 to 66.18 DSC, are inflated by a leakage in the paired settings, and that should be fixed before anyone relies on the quantitative claims.\n\nWhat's genuinely new: MRGen-DB is one of the larger public MRI datasets with rich text prompts and a masked subset. The two-stage training — text pretraining on unlabelled Radiopaedia data, then mask-conditioned finetuning — is a sensible way to extend controllable generation to modalities that lack masks. The ablation showing both stages matter is useful. I also appreciate that the appendix is explicit about using 'the entire source-domain dataset (including both training and test sets)' when generating synthetic pairs; the problem is not hidden.\n\nWhere it's soft: for the paired CHAOS and PanSeg settings, source and target test sets are the same subjects imaged in different protocols. Using source test masks to generate synthetic target training images means the model has seen each test subject's organ geometry during training. That is exactly the setup producing the largest gains. The clean cross-dataset pairs — LiQA to CHAOS, MSD-Prostate to PROMISE12 — show smaller but positive gains, which is the real evidence. On top of that, FID and DSC are reported without error bars or multiple seeds, so we don't know the variance. And the code/data are only 'will be released,' so the results aren't independently checkable right now.\n\nA lesser concern: the method's ability to generate a target modality depends on that modality appearing in the unlabelled Radiopaedia corpus; the paper doesn't verify that T2-SPIR, ADC, T1, T2 are actually represented there. For a 'zero-shot' claim that matters.\n\nWho should read this: anyone working on medical image synthesis and cross-modality segmentation. It's a serious piece of work with a useful dataset and a plausible method, but the evaluation needs cleaning up. I'd send it to peer review, with a strong request to rerun the paired experiments using only source training masks, report variance, and release artifacts. The core idea survives; the headline numbers won't.","headline":"Good dataset and plausible method, but the paired-experiment leakage inflates the headline gains; clean cross-dataset results still support the core idea.","tokens_in":24632,"tokens_out":2504,"would_cite":true,"duration_ms":24377,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MRGen claims that a diffusion generator conditioned on text and masks can synthesize MRI for unannotated modalities, lifting segmentation DSC from 6.90 to 66.18 and winning 8 of 10 cross-modality settings.","keywords":["MRGen","synthetic medical images","MRI segmentation","diffusion models","mask-conditioned generation","text-guided generation","underrepresented modalities","cross-modality segmentation"],"falsifier":"Take a target MRI protocol that a modality audit shows is absent from the unlabeled pretraining corpus, generate synthetic training data with MRGen, and measure downstream Dice score: if gains over source-only training stay near zero, the claim that MRGen generalizes to genuinely unseen modalities is falsified.","tokens_in":23590,"feed_emoji":"🩻","tokens_out":7145,"duration_ms":64088,"temperature":0.7,"pith_summary":"This paper claims that one diffusion-based generator can supply training data for segmentation models on MRI modalities that have no manual masks at all. MRGen is trained in two stages: first on roughly 250,000 unlabeled image-text slices to learn what each MRI modality looks like, then on a smaller set of mask-annotated slices so it can obey an organ mask while generating. Given a templated text prompt and a mask, it synthesizes a realistic image of the requested target modality, and training standard segmentation frameworks on those synthetic images lifts target-domain Dice scores sharply, for example CHAOS T1 to T2-SPIR climbs from 6.90 with source-only training to 66.18 with synthetic data. A sympathetic reader would care because rare but clinically important MRI protocols are exactly those where hand annotation is too costly and scarce to collect.","feed_headline":"Synthetic MRI lifts mask-free modality segmentation from 6.9 to 66.2","feed_subtitle":"A text-and-mask conditioned diffusion engine synthesizes training data for rare MRI protocols, winning 8 of 10 settings.","key_machinery":"The load-bearing object is the two-stage latent diffusion engine $I=\\Phi_{\\mathrm{MRGen}}(T,M;\\Theta,\\Theta_c)$, with templated text prompts $T$ and organ masks $M$ as conditions. A VAE maps slices into a latent space; a UNet denoiser is first trained with cross-attention to a biomedical text encoder so the model associates modalities, attributes, regions, and organs with images, and then a mask encoder (initialized from the diffusion UNet and downsampled) is added as a residual into the decoder blocks while the rest stays frozen. An off-the-shelf SAM2-based autofilter checks the mask-faithfulness of each candidate by comparing IoU and confidence scores before synthetic samples enter segmentation training. This arrangement carries the argument because modality knowledge comes from the unlabeled corpus while controllability comes from the frozen diffusion weights plus a small mask-annotated subset, so the two sources of information never compete for the same data.","core_discovery":"The paper's central claim is that mask-conditioned controllability can be transferred to MRI modalities that never had masks, as long as their appearance was seen during text-guided pretraining. Concretely, MRGen is a latent diffusion model that takes a text prompt (modality name, tissue signal attributes, body region, and organ list) plus a segmentation mask, and returns a 2D MR slice of the requested modality. The two-stage schedule, text pretraining on image-text pairs followed by mask-conditioned finetuning with the diffusion backbone frozen, is what lets the mask controller generalize beyond annotated modalities. In ten cross-modality settings, synthetic MRGen data improves both nnUNet and UMamba, giving the best Dice score in eight experiments; the strongest case is CHAOS T1 to T2-SPIR, where nnUNet's Dice rises from 6.90 to 66.18.","pith_inferences":["If a target protocol is absent from the unlabeled pretraining corpus, the method probably degrades to source-only performance, so auditing corpus modality coverage should come before deploying MRGen for a new protocol.","The proposal effectively converts annotation scarcity into a corpus-coverage problem: the engine's ceiling is set by the breadth of text-image pairs it can pretrain on, not by the number of masks.","One could test whether SAM2-based filtering is replaceable by any zero-shot segmenter, or whether the filter itself limits the method to organs SAM2 can recognize.","A direct extension is 3D volume synthesis or non-MRI protocols such as CT, where the same text-and-mask conditioning could be retrained."],"forward_implications":["Segmentation models can be trained for target MRI modalities without a single manually annotated target-domain mask, turning a previously zero-shot scenario into a usable one.","The benefit transfers across segmentation architectures: both nnUNet and UMamba improve, and MRGen ranks first in 8 of 10 cross-modality experiments.","Adding unlabeled target-domain images during training and using SAM2-based autofiltering both further raise Dice scores, so the pipeline degrades gracefully instead of depending on one trick.","Synthetic in-domain data can nearly match real annotated data for segmentation, suggesting the engine could also serve privacy-preserving data sharing.","Preliminary cross-protocol experiments from CT to MRI indicate the same engine may extend beyond MRI to other underrepresented imaging settings."],"supporting_citations":[{"why":"Supplies the latent diffusion architecture used as the backbone for text-conditioned generation.","marker":"[50]"},{"why":"Provides the biomedical text encoder used for templated prompts and for region classification during dataset curation.","marker":"[63]"},{"why":"Generates the free-text modality attribute descriptions that let the model distinguish fine-grained MRI protocols.","marker":"[1]"},{"why":"Runs the automatic filtering pipeline that keeps only synthetic images faithful to the conditioning masks.","marker":"[49]"},{"why":"Is the primary segmentation framework whose target-domain performance is improved by synthetic MRGen data.","marker":"[23]"},{"why":"Is the second segmentation architecture used to show that MRGen's synthetic data transfers across frameworks.","marker":"[41]"},{"why":"Provides the CHAOS-MRI T1 and T2-SPIR pairs used in the headline cross-modality segmentation result.","marker":"[27]"},{"why":"Provides the MSD-Prostate T2 and ADC pairs used for cross-modality generation and segmentation evaluation.","marker":"[3]"}],"fun_headline_variants":["Diffusion engine lifts MRI segmentation from 6.9 to 66.2 Dice","Text-mask diffusion turns 6.9 Dice into 66.2 on rare MRI","Synthetic MRI boosts mask-less modality Dice to 66.2","MRGen: Synthetic data lifts Dice 6.9 to 66.2 for rare MRIs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The generator can only synthesize a target modality whose appearance is present somewhere in the unlabeled Internet radiology corpus used for pretraining, and the paper does not verify that its test modalities (T2-SPIR, ADC, T1, T2) actually appear there.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion engine lifts MRI segmentation from 6.9 to 66.2 Dice","Text-mask diffusion turns 6.9 Dice into 66.2 on rare MRI","Synthetic MRI boosts mask-less modality Dice to 66.2","MRGen: Synthetic data lifts Dice 6.9 to 66.2 for rare MRIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00115,"raw_usage":{"total_tokens":4770,"prompt_tokens":951,"completion_tokens":3819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":3728}},"tokens_in":567,"tokens_out":3819,"duration_ms":22569,"temperature":1.0,"reasoning_tokens":3728,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:23:01.099765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a target MRI protocol that a modality audit shows is absent from the unlabeled pretraining corpus, generate synthetic training data with MRGen, and measure downstream Dice score: if gains over source-only training stay near zero, the claim that MRGen generalizes to genuinely unseen modalities is falsified.","supporting_citations":[{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion architecture used as the backbone for text-conditioned generation."},{"cited_title":"A multimodal biomedical foun- dation model trained from fifteen million image–text pairs","cited_arxiv_id":null,"evidence_quote":"Provides the biomedical text encoder used for templated prompts and for region classification during dataset curation."},{"cited_title":"Sam 2: Segment anything in images and videos","cited_arxiv_id":null,"evidence_quote":"Runs the automatic filtering pipeline that keeps only synthetic images faithful to the conditioning masks."},{"cited_title":"nnu-net: a self-configuring method for deep learning-based biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"Is the primary segmentation framework whose target-domain performance is improved by synthetic MRGen data."},{"cited_title":"Chaos challenge-combined (ct-mr) healthy abdominal organ seg- mentation","cited_arxiv_id":null,"evidence_quote":"Provides the CHAOS-MRI T1 and T2-SPIR pairs used in the headline cross-modality segmentation result."},{"cited_title":"The medical segmentation decathlon.Nature Communications, 13(1):4128, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the MSD-Prostate T2 and ADC pairs used for cross-modality generation and segmentation evaluation."}],"review_version":1}