{"id":"0a807b39-1adc-43a4-81e5-b63995303358","arxiv_id":"2505.24687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TumorGen uses rectified flow matching to jointly synthesize 3D tumor images and masks from a rough bounding box, achieving lower FID and faster sampling than prior tumor synthesis models on two PET datasets.","lead":"A new AI method, TumorGen, generates realistic 3D tumor images and their segmentation masks at the same time, using only a rough bounding box to show where the tumor should be. It is faster than previous diffusion-based tumor synthesis methods because it uses rectified flow matching, and tests on two PET datasets show better image quality scores.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference-time bounding box acquisition is unspecified for healthy subjects; BA-PMG derives the box from the GT tumor mask, so Figure 4 and the 'automatic bbox' claim are unsupported.","rationale":"The central claim is that TumorGen outperforms existing tumor synthesis methods in efficiency and realism and can generalize to healthy subjects. The most load-bearing assumption is not the flow-matching mathematics, which is internally coherent, but the conditioning signal: the only spatial prior is a bounding box, and the box's source is ambiguous. BA-PMG explicitly requires the resized tumor mask m1 to locate the tumor voxels, so for tumor-bearing scans the box is ground-truth-derived rather than truly user-specified. For healthy subjects, no m1 exists, yet Figure 4 presents generated tumors without any described mechanism for obtaining the box. The reader's weakest_assumption identified exactly this gap, and I agree with that assessment. The conditional verdict already reflects this concern; no adjustment is needed. Secondary issues such as the absence of error bars and the binary nature of the final refined mask are real but less central than the unspecified inference-time conditioning.","tokens_in":8025,"tokens_out":4311,"duration_ms":55728,"concrete_test":"Require the authors to define and implement the healthy-subject bounding-box protocol: e.g., sample boxes in lung regions with a fixed deterministic rule or train a tumor-proposal network, then generate tumors in the healthy scans and report FID, DSC, and NSD for that protocol. Also compare the main-table FID when the bounding box is taken from the GT mask versus from the protocol on the tumor-bearing test set. If the protocol-dependent numbers are not reported or differ substantially from the GT-mask numbers, the Section 3.2 generalizability claim and the Section 2.1 'automatic' bounding-box claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1's BA-PMG computes the bounding box from a ground-truth tumor mask: it finds the tight box B' from 'minimum and maximum coordinates of tumor voxels' in m1 and then expands it. During synthesis for tumor-bearing scans, the claimed 'only a rough bounding box' input is therefore derived from the GT mask, and the healthy-subject experiment in Figure 4 has no described way to obtain B because no tumor mask exists. The Figure 2 caption calls this 'automatic tumor bounding box generation', but no mechanism for inference on healthy images is given. This is load-bearing because the practical contribution and the generalizability claim depend on being able to place a box without annotations. Without a specified protocol—user-drawn, random, or detector-based—the experiment is not reproducible and the method cannot be used as claimed. This concern does not invalidate the comparative FID results on tumor-bearing scans, but it undercuts the strongest downstream claim of synthesizing tumors in healthy subjects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TumorGen, a 3D PET tumor synthesis framework with three components: a Boundary-Aware Pseudo Mask Generation module (BA-PMG) that expands ground-truth tumor masks into bounding boxes, a Spatial-Constraint Vector Field Estimator (SC-VFE) that jointly synthesizes tumor latents and masks by rectified flow matching under coarse- and fine-grained spatial constraints, and a VAE-Guided Mask Refiner (VMR) that refines predicted masks with decoder features. Experiments on Hecktor and AutoPET report improved FID, DSC/NSD, and faster sampling relative to existing tumor synthesis baselines, including a 10-step rectified flow run at 0.218 s per sample. The paper also claims generalization to healthy subjects by synthesizing tumors in tumor-free images.","tokens_in":8239,"tokens_out":4361,"duration_ms":55701,"significance":"If the method works as described, it offers a practical advance in medical tumor synthesis: simultaneous image-mask generation with few-step rectified flow could substantially reduce synthesis cost while maintaining image-mask alignment. The use of two public datasets and standard metrics (FID, DSC, NSD) and the explicit inference-time comparison are strengths, as is the direct measurement of sampling speed. However, two central claims are currently undersupported: the inference-time bounding-box mechanism for healthy-subject synthesis is not defined, and the stated soft-boundary motivation is contradicted by the binary mask reconstruction loss. These issues are fixable in revision and do not by themselves invalidate the reported comparative FID scores on tumor-bearing scans, but they must be addressed before the paper can be accepted.","major_comments":[{"comment":"The inference-time bounding box is not specified. BA-PMG computes the tight box B' from the coordinates of tumor voxels in the ground-truth mask m1 and then expands it (Section 2.1). During inference on tumor-bearing scans it is unclear whether the ground-truth mask is still used to obtain B; for healthy subjects in Figure 4 no tumor mask exists, so there is no described mechanism to place the box. This makes the generalizability experiment in Figure 4 non-reproducible and the claim of synthesizing tumors in healthy subjects unsupported. Please specify the protocol (user-drawn box, random placement, detector, or ground-truth-derived box), report sensitivity to box placement, and temper the claims of automatic generalization accordingly.","section":"Section 2.1, Figure 4"},{"comment":"The boundary-aware/soft-boundary claim is contradicted by the mask objective. The paper's motivation (Section 1) emphasizes soft, gradual tumor boundaries, but the refined mask M' is trained with Lrec = ||f_theta(hat m1) - M||^2 against the binary ground-truth mask M (defined in the Overview), and the final output is a binary mask. The bounding box in BA-PMG provides surrounding context but does not model gradual transitions in the mask. This is load-bearing because the boundary-aware framing is central to the paper's contribution. Please either replace the binary reconstruction loss with a soft-boundary formulation (e.g., distance-transform encoded labels) and evaluate with boundary-aware metrics, or revise the claims to describe bounding-box-conditioned binary mask synthesis.","section":"Section 2.3, Eq. (6)"},{"comment":"Several implementation details needed for reproducibility are missing. The values of alpha (BA-PMG offset scaling) and lambda_1, lambda_2, lambda_3 (overall objective in Eq. (7)) are not reported, and the number of synthetic samples used to compute FID is not stated. These omissions do not invalidate the reported comparisons, but they prevent replication and make the sensitivity of the method to its free parameters impossible to assess. Please provide the hyperparameter values and the FID evaluation protocol in the revised manuscript.","section":"Section 3.1, Section 2"}],"minor_comments":[{"comment":"There is a citation inconsistency for MedFusion: Table 1 labels it [23], while Section 3.2 cites MedFusion as [18]; reference [23] in the bibliography is Singan-Seg, not MedFusion.","section":"Table 1, Section 3.2"},{"comment":"The caption calls BA-PMG 'automatic tumor bounding box generation', but the method computes the bounding box from a ground-truth tumor mask, so it is not automatic without annotation. Please rephrase the caption to match the actual procedure.","section":"Figure 2 caption"},{"comment":"Equation (2) uses the same symbol epsilon for z_t and m_t, which could imply that the same noise sample is used for the image and mask latents; please clarify whether the noises are shared or independently sampled.","section":"Equation (2)"},{"comment":"The fine-grained constraint uses SSIM on latent-space tensors, but the paper does not specify how SSIM is computed across the latent channels; please state the SSIM variant and parameter settings.","section":"Equation (5)"}],"recommendation":"major_revision","confidential_remarks":"The two load-bearing issues — the missing inference-time bounding-box protocol and the mismatch between the soft-boundary framing and the binary mask loss — are fixable in revision. If the authors provide a concrete inference protocol and reconcile the objective with the boundary-aware claim, the core contributions could be acceptable. I see no evidence of numerical overfitting or circularity in the main comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it combines rectified flow matching with a bounding-box condition to synthesize tumor image latents and masks jointly, and the reported FID and speed gains over DiffTumor and other baselines are consistent across Hecktor and AutoPET. Second, the paper's headline claim about generalizing to healthy subjects has a load-bearing hole: the BA-PMG module computes the bounding box from the ground-truth tumor mask, and there is no described mechanism for placing a box on a tumor-free scan. That gap undercuts the most interesting downstream use case, not the comparative FID numbers on tumor-bearing scans.\n\nWhat is genuinely new: the joint generation of image and mask in one rectified-flow pass with coarse and fine spatial constraints, plus the VAE-guided mask refiner. The efficiency number (0.218s per sample at 10 steps with FID 55.862) is credible given how rectified flow works, and the nnU-Net alignment check (DSC 0.694, NSD 0.741) is a sensible sanity check that most synthesis papers skip. The ablation study is complete and each component earns its keep. I also credit the authors for not overselling the method as a clinical tool; the cancer-diagnosis framing is just context.\n\nWhere it is soft: The \"boundary-aware\" claim is overstated. The final mask is binary; the soft-boundary language refers to the pseudo-mask and loss constraints, not the output. Also, the bounding-box issue is exactly as the stress-test note says. In training, B is derived from the mask; at inference on tumor-bearing images you might argue a user could draw a box, but the paper never says that, and for healthy subjects no mask exists. The Figure 2 caption calls it 'automatic tumor bounding box generation', which is misleading. This is fixable with a short protocol description, but it has to be fixed. Minor concerns: no error bars, no code, and only a single run of the self-citation Polyp-Gen in related work (not load-bearing).\n\nThe math and experiments are internally consistent; the FID comparisons are on standard datasets and the architecture choices are grounded. The citation pattern is fine. I do not see a fatal flaw, but I do see a reproducibility gap in the main generalization experiment.\n\nWho is this for? Researchers working on tumor synthesis for training segmentation models, especially in PET. It deserves a serious referee. I would send it out, with a request for the bounding-box protocol and a softened boundary claim.\n\nRecommendation: accept for peer review, conditional on the authors specifying how a bounding box is obtained on healthy subjects and reporting variance across seeds.","headline":"A plausible and efficient joint tumor-image-and-mask synthesis method that deserves peer review, but the healthy-subject generalization claim rests on an unspecified inference-time bounding-box procedure.","tokens_in":8762,"tokens_out":1438,"would_cite":true,"duration_ms":19082,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TumorGen claims that 3D tumor images and their soft boundary masks can be synthesized jointly from a rough bounding box using rectified flow matching, matching or beating two-stage diffusion baselines in realism at roughly one-sixth the…","keywords":["tumor synthesis","rectified flow matching","3D medical imaging","PET imaging","generative models","boundary-aware masks","data augmentation","image-mask alignment"],"falsifier":"Run a held-out experiment in which tumors are synthesized into healthy PET scans using bounding boxes from three independent radiologists and from a purely automatic detector, then measure FID and the DSC/NSD of the generated image-mask pairs against real tumors; if realism and alignment collapse when the box is not derived from ground-truth mask expansion, the 'only a rough bounding box' claim fails.","tokens_in":7839,"feed_emoji":"🩻","tokens_out":4914,"duration_ms":54507,"temperature":0.7,"pith_summary":"The paper aims to make synthetic 3D tumor data practical for training diagnostic models: instead of first generating a pixel-perfect tumor mask and then an image in a slow two-stage diffusion pipeline, TumorGen generates image and mask at the same time from a single rough bounding box. It replaces strict binary masks with a boundary-aware pseudo-mask that allows soft, gradual tumor edges, and uses rectified flow matching to cut sampling from hundreds of steps to ten. On the PET datasets Hecktor and AutoPET, the authors report lower Fréchet Inception Distance than prior tumor-synthesis methods, and generated image-mask pairs remain segmentable by a pre-trained nnU-Net. The significance, if the results hold, is a far cheaper pathway to synthetic tumors for data-hungry cancer imaging models.","feed_headline":"Tumor synthesis drops to 10 steps and 0.218 seconds per sample","feed_subtitle":"A boundary-aware flow model beats two-stage diffusion baselines on PET tumor realism while pairing each image with its mask.","key_machinery":"Three components carry the argument. BA-PMG (Boundary-Aware Pseudo Mask Generation) derives a training-time bounding box by expanding the tight box around the ground-truth tumor with random offsets proportional to image dimensions; this is the object that relaxes strict binary supervision. SC-VFE (Spatial-Constraint Vector Field Estimator) jointly predicts tumor-image latents and mask latents with rectified flow matching, which learns a straight interpolation path between noise and data so only few sampling steps are needed, and applies two spatial losses: an $\\ell^2$ constraint inside the bounding box and an SSIM constraint inside the generated tumor mask. VMR (VAE-Guided Mask Refiner) upsamples the coarse latent mask by adding hierarchical features from the 3D-VAE decoder, recovering boundary detail lost in the latent space.","core_discovery":"The central discovery is that tumor synthesis does not need a predefined binary segmentation; a flexible bounding box plus joint flow-matching generation of latents and masks is enough to produce realistic 3D tumors with soft boundaries, and this recipe is substantially faster than the previous diffusion-based two-stage approaches. More specifically, TumorGen reports average FID 52.035 on Hecktor versus 66.821 for DiffTumor and 72.433 versus 84.946 on AutoPET, with 0.218 seconds per 10-step sample, and the generated masks align with images closely enough that a segmentation network achieves DSC 0.694 and NSD 0.741 on Hecktor.","pith_inferences":["The authors do not compare DiffTumor at 10 diffusion steps, so part of the speed gap is sampling-schedule choice; a head-to-head at equal step count would isolate the rectified-flow contribution.","A testable extension is to couple TumorGen with an automatic lesion-proposal network, which would test whether the bounding box can be fully automatic rather than user-supplied on healthy subjects.","The same three-component design could be adapted to CT or MRI, since the VAE latent space and flow-matching objective are modality-agnostic; the BA-PMG box offsets would need re-tuning to organ-specific tumor sizes.","Because VMR conditions mask refinement on decoder features, the refined mask inherits image details, suggesting the module could double as an anomaly-localization head for generated images."],"forward_implications":["A single trained model outputs both the tumor image and a high-resolution mask in the same forward pass, so downstream segmentation or detection training no longer needs a separate mask-proposal stage.","Tumor synthesis for rare lesion types becomes practical on clinical budgets: 10-step sampling at 0.218 seconds per volume allows dataset expansion on a single GPU in minutes.","The healthy-subject generalization shown in Figure 4 implies the model can populate tumor-free PET scans with plausible lesions, a route to balancing class-imbalanced cancer datasets.","The boundary-aware pseudo-mask replaces binary supervision, so synthesized tumors exhibit graded transitions between diseased and healthy tissue rather than hard edges."],"supporting_citations":[{"why":"Defines the DiffTumor two-stage synthesis baseline that TumorGen compares against for FID and inference-time gains.","marker":"[4]"},{"why":"Supplies the rectified flow matching objective and straight-path latent-space sampling that SC-VFE adapts for joint image-mask synthesis.","marker":"[6]"},{"why":"Provides the pre-trained 3D KL-VAE used to encode and decode tumor volumes and supplies decoder features for the mask refiner.","marker":"[8]"},{"why":"SynTumor is a representative two-stage tumor synthesis baseline; the comparison establishes the realism improvement.","marker":"[12]"},{"why":"nnU-Net is the pre-trained segmenter used to predict masks from synthesized images, grounding the reported DSC/NSD alignment values.","marker":"[14]"},{"why":"TotalSegmentator is used to extract and crop lung regions on the AutoPET dataset, defining the preprocessing for that benchmark.","marker":"[27]"},{"why":"The Hecktor challenge dataset supplies the primary training and testing PET volumes for the main experiments.","marker":"[1]"}],"fun_headline_variants":["Boundary-aware tumor masks in one flow step","Rectified flow cuts tumor synthesis to 10 steps","TumorGen: joint latent-mask flow for PET synthesis","Faster tumor synthesis with soft-boundary masks","No binary masks needed for realistic tumor synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method presupposes that a rough 3D bounding box around the tumor is available or provided at generation time; the paper only explains how the box is derived when the ground-truth mask is known, leaving the healthy-subject case without a specified box-proposal mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Boundary-aware tumor masks in one flow step","Rectified flow cuts tumor synthesis to 10 steps","TumorGen: joint latent-mask flow for PET synthesis","Faster tumor synthesis with soft-boundary masks","No binary masks needed for realistic tumor synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":1926,"prompt_tokens":868,"completion_tokens":1058,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":984}},"tokens_in":484,"tokens_out":1058,"duration_ms":10741,"temperature":1.0,"reasoning_tokens":984,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:14:40.973443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a held-out experiment in which tumors are synthesized into healthy PET scans using bounding boxes from three independent radiologists and from a purely automatic detector, then measure FID and the DSC/NSD of the generated image-mask pairs against real tumors; if realism and alignment collapse when the box is not derived from ground-truth mask expansion, the 'only a rough bounding box' claim fails.","supporting_citations":[{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Defines the DiffTumor two-stage synthesis baseline that TumorGen compares against for FID and inference-time gains."},{"cited_title":"CoRR (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained 3D KL-VAE used to encode and decode tumor volumes and supplies decoder features for the mask refiner."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"SynTumor is a representative two-stage tumor synthesis baseline; the comparison establishes the realism improvement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"nnU-Net is the pre-trained segmenter used to predict masks from synthesized images, grounding the reported DSC/NSD alignment values."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TotalSegmentator is used to extract and crop lung regions on the AutoPET dataset, defining the preprocessing for that benchmark."}],"review_version":1}