{"id":"6cf77ce0-5938-4c01-ab51-a5e4b7fa9301","arxiv_id":"2507.16429","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion-based medical image segmentation model with prototype contrastive consistency improves mIoU on Endoscapes2023 and on the new MOSXAV angiography benchmark.","lead":"This paper combines a diffusion-based segmentation decoder with prototype contrastive losses to train semi-supervised medical image segmentation models on limited labeled data plus pseudo-labeled images. It also introduces MOSXAV, a new manually annotated benchmark for multi-object segmentation in X-ray angiography videos.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SoTA comparison is uncontrolled: Ours uses ViT-Large while baseline backbones are unspecified, so the +5.15% margin over MedSegDiff may reflect model capacity rather than the proposed diffusion decoder and prototype contrastive losses.","rationale":"The reader's weakest_assumption is precisely the load-bearing concern: the comparison is uncontrolled because baseline backbones and training details are not specified, so the claimed margin may stem from ViT-Large capacity rather than the proposed contributions. This concern directly undermines the strongest claim about outperforming state-of-the-art methods. The paper provides no same-backbone baseline, no error bars, and no experiment that varies pseudo-label noise to test robustness, which the title and abstract emphasize. These are fixable experimental omissions, not fundamental contradictions, so a conditional verdict remains appropriate. No additional concern outweighs this one: the MOSXAV description inconsistency is real but secondary, and the method's internal logic is coherent. Therefore the reader's conditional verdict should stand unchanged.","tokens_in":8622,"tokens_out":3215,"duration_ms":38183,"concrete_test":"Run a controlled comparison on Endoscapes2023 under the identical GT+Pseudo protocol: (a) train the proposed model with ViT-Large; (b) train the same ViT-Large encoder and diffusion decoder but remove L_inter and L_intra; (c) train MedSegDiff and at least U-Net with ViT-Large encoders, the same MedSAM+SAMRefiner pseudo-labels, input resolution, batch size, and 40k iterations; report mean and std over three seeds. If (b) already reaches ~56.7 mIoU or (c) closes the gap to below 5.15, the headline margin is attributable to backbone capacity or training protocol, not the proposed method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 3.2 is that the method 'significantly outperforms' MedSegDiff by +5.15 mIoU under the semi-supervised setting. Section 3.1 states that the proposed main experiments use ViT-Large as the visual backbone and that 'all models are trained for 40,000 iterations' with batch size 8, but Table 1 does not report the backbone, input resolution, training iterations, or pseudo-label protocol used for any baseline. Standard baselines such as U-Net, Attention U-Net, TransU-Net, CMU-Net, and SwinU-Net have their own encoder architectures of varying capacity; MedSegDiff also does not necessarily use ViT-Large. Simply training all methods for the same number of iterations does not equalize model capacity or the quality of supervision. The paper's ablations use ViT-Base and the full objective reaches 54.47 mIoU, while the main result with ViT-Large reaches 56.68; no experiment isolates how much of the +5.15 margin over MedSegDiff comes from the larger backbone versus the proposed prototype contrastive consistency. Therefore the state-of-the-art claim rests on an uncontrolled comparison and is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffusion-based framework for semi-supervised medical image segmentation, combining a conditional diffusion decoder with prototype-anchored contrastive losses on label embeddings. It also introduces MOSXAV, a new multi-object X-ray angiography segmentation benchmark. The method is evaluated on Endoscapes2023 and MOSXAV, and the authors report state-of-the-art results, including a +5.15 mIoU gain over MedSegDiff under the GT+Pseudo setting on Endoscapes2023.","tokens_in":8864,"tokens_out":11968,"duration_ms":110660,"significance":"If the results hold, the contribution is potentially valuable: the paper releases a new public benchmark, applies dense label denoising with structured latent anchors, and provides algorithm pseudocode and a repository link. The dataset release and the external-benchmark evaluation are strengths. However, the current evidence for the state-of-the-art claim is not conclusive. The baseline comparison is uncontrolled with respect to backbone capacity, and the prototype-contrastive losses are applied to the clean ground-truth label embedding rather than to the denoised prediction, which contradicts the claimed mechanism. These issues are fixable but require additional experiments or a substantially revised description.","major_comments":[{"comment":"The prototype-contrastive losses are not applied to the denoising process. In Algorithm 1, z_latent = latent_encoder(label_encoder(gts)) is computed from the ground-truth label map before noise is added, and the inter/intra losses in Eq. (2)-(3) are evaluated on this clean embedding; z_crpt, the timestep t, and the predicted preds do not enter these losses. The abstract and Section 2.3 claim a constraint 'during the denoising diffusion process', which does not describe this implementation. Please either move the prototype losses to the predicted/denoised embedding and re-run the experiments, or clearly state that the regularizer is applied to the target label embedding and justify why it should improve denoised predictions.","section":"Section 2.3, Algorithm 1"},{"comment":"The baseline comparison in Table 1 is not controlled for backbone capacity or training protocol. Implementation Details report ViT-Large for the proposed method and state that all models are trained for 40,000 iterations with batch size 8, but no backbone, input resolution, or pseudo-label protocol is reported for U-Net, Attention U-Net, TransU-Net, CMU-Net, CMU-NeXt, SwinU-Net, or MedSegDiff. The +5.15 mIoU margin over MedSegDiff on Endoscapes2023 and the +3.11 mIoU on the MOSXAV test set may therefore stem from the ViT-Large backbone rather than from the proposed diffusion decoder and prototype losses. The ViT-Base ablation (54.47 mIoU) versus the ViT-Large main result (56.68 mIoU) confirms a non-negligible backbone effect, so the state-of-the-art claim requires same-backbone baselines or an explicit backbone-isolation ablation.","section":"Section 3.1, Table 1"},{"comment":"The MOSXAV dataset description contains an internal inconsistency. The text states that each video contains 33-70 frames and that dense annotations are provided every 5 frames across 50 sequences, totaling 2,335 frames; this implies about 46.7 annotated frames per sequence, which cannot be reconciled with 'every 5 frames' for the stated sequence lengths. Please specify the exact annotation density, the selection of the 50 sequences, and how the 488 test frames are drawn from the remaining sequences, and state the number of classes and the annotation/QA protocol for the benchmark.","section":"Section 3.1"},{"comment":"The MOSXAV evaluation does not actually demonstrate the semi-supervised setting. Table 1 reports only a GT column for MOSXAV val and test, with no Pseudo or GT+Pseudo columns, and Section 3.2 gives no pseudo-label protocol for MOSXAV. The paper's headline contribution is robustness to noisy pseudo-labels in semi-supervised learning, but on the new benchmark the model is only compared under full supervision. Please add the MOSXAV semi-supervised experiments, or explicitly restrict the semi-supervised claim to Endoscapes2023.","section":"Section 3.1, Table 1"},{"comment":"The prototype update mechanism is not specified. Algorithm 1 calls online_cluster(z_latent, gts, p), but the paper never states how p is initialized, updated, or normalized, nor how empty prototypes are handled. Because the method's novelty is the use of non-learnable prototype anchors, omitting this update rule prevents reproduction and leaves the stability of the online clustering unsupported.","section":"Section 2.3"}],"minor_comments":[{"comment":"The subsection headings appear swapped: Eq. (2) is a contrastive loss that pushes negative prototypes apart (inter-class separation), while Eq. (3) pulls pixels toward their assigned prototype (intra-class compactness). Please rename the headings accordingly.","section":"Section 2.3"},{"comment":"The reverse-process description says the denoising network is applied to z0 during inference; it should be applied to the current noisy state z_t, with z0 being the predicted target.","section":"Section 2.1"},{"comment":"In Algorithm 2, the time update t_next = max(1 - (1 + step + t_diff)/steps, 0) appears to skip two diffusion intervals when t_diff=1; please clarify the units of t_diff and check the schedule against DDP's asymmetric sampling.","section":"Algorithm 2"},{"comment":"No error bars or multiple-seed results are reported; given the limited labeled data and the noisy-pseudo-label setting, at least three runs with mean ± standard deviation should be provided for the headline numbers.","section":"Tables 1 and 2"},{"comment":"Figure 1 is hard to read at the current resolution; please enlarge it and label the prototype update and contrastive-loss blocks more clearly.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical imaging venue and the MOSXAV benchmark could be a useful community resource. The main concerns are that the headline SoTA claim rests on an uncontrolled baseline comparison and that the described loss placement does not match the implementation. These require new experiments or a substantially revised method description, not just textual edits. The annotation-count inconsistency in Section 3.1 should also be checked against the released repository before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a solid but uneven paper. The genuinely new piece is MOSXAV, a public benchmark for multi-object X-ray angiography segmentation, and the method reasonably combines diffusion label denoising with prototype contrastive losses. The main empirical claim, however, is not yet controlled enough to trust.\n\nWhat's actually new: the MOSXAV benchmark is a real contribution—62 videos with manual annotations of vascular structures, publicly linked. That alone merits attention. The technical combination is a sensible extension of DDP/MedSegDiff: anchor the latent space of label embeddings with non-learnable prototypes while training the diffusion decoder. The ablation shows each loss component adds a small, consistent gain, which is honest.\n\nWhere it's soft: first, the key comparison is uncontrolled. The proposed method uses ViT-Large; Table 1 lists no backbones for any baseline, so the +5.15 mIoU margin over MedSegDiff could be a capacity effect. There are no error bars, so we can't judge stability. Second, the paper never compares against the semi-supervised methods it cites, such as FixMatch or Cross Pseudo Supervision, even though the setting is explicitly semi-supervised. Third, the MOSXAV description is internally inconsistent: 'dense annotations every 5 frames across 50 sequences' should give roughly 500-700 frames, not 2,335. Fourth, the prototype contrastive loss is applied to the clean label embedding before noise is added, not during the denoising; the abstract's phrasing overstates the connection. Finally, the title promises robustness to noisy pseudo-labels, but there is no experiment that actually perturbs pseudo-label noise.\n\nNone of these are fatal. Circularity isn't a concern—the evaluation uses external benchmarks, and the self-citations aren't load-bearing. The method is plausible and the benchmark is useful. But the state-of-the-art claim needs same-backbone baselines, error bars, a direct pseudo-noise robustness test, and a corrected dataset description before it's convincing.\n\nWho this is for: researchers working on semi-supervised medical segmentation, particularly with diffusion models, and anyone needing a new angiography benchmark. It deserves a serious referee, because the benchmark and the idea are worth engaging with, even if the current evidence is conditional.\n\nRecommendation: send it to peer review, with a request for the missing controlled experiments and dataset documentation.","headline":"Useful new benchmark and a plausible diffusion-segmentation combination, but the SoTA claim needs same-backbone baselines and error bars; worth a referee.","tokens_in":9407,"tokens_out":3870,"would_cite":true,"duration_ms":39394,"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":"A prototype-contrastive diffusion decoder beats prior semi-supervised segmentation methods by +5.15 mIoU on Endoscapes2023.","keywords":["semi-supervised learning","medical image segmentation","diffusion models","pseudo-label noise","prototype contrastive learning","X-ray angiography","Endoscapes2023","MOSXAV"],"falsifier":"Retrain all Table 1 baselines with the same ViT-Large backbone, 512x512 input resolution, 40,000 iterations, batch size 8, and the same MedSAM plus SAMRefiner pseudo-label protocol. If the mIoU gap over MedSegDiff shrinks to near zero while the prototype losses are removed from the proposed model, the core claim fails.","tokens_in":8406,"feed_emoji":"🩻","tokens_out":4965,"duration_ms":50007,"temperature":0.7,"pith_summary":"This paper tries to make semi-supervised medical image segmentation robust to noisy pseudo-labels by restructuring the latent space of a diffusion-based segmenter. The core idea is to treat segmentation as a label-denoising problem, then anchor learned pixel embeddings to fixed class prototypes with two contrastive losses, one pulling same-class pixels together and one pushing different classes apart. The authors report that this beats state-of-the-art baselines, including MedSegDiff by +5.15 mIoU on Endoscapes2023, and introduce MOSXAV, a new multi-object X-ray angiography segmentation benchmark. If the result holds, noisy pseudo-labels from automatic tools become less damaging, and data-efficient clinical segmentation improves.","feed_headline":"Diffusion model beats MedSegDiff by 5.15 mIoU","feed_subtitle":"Prototype anchors tame noisy pseudo-labels, and a new X-ray benchmark ships with the work.","key_machinery":"The central object is the diffusion decoder that predicts dense labels by denoising corrupted label embeddings: a label encoder maps a one-hot mask to a normalized embedding z, Gaussian noise with a cosine schedule corrupts it to z_t, and the decoder, conditioned on features from a ViT backbone and visual-specific branches, reconstructs the clean embedding. The second load-bearing piece is the set of fixed, non-learnable class prototypes {p_{c,k}}, updated by online clustering, which serve as anchors in the projected feature space Z = φ(z). The contrastive losses L_inter and L_intra pull each pixel toward its assigned prototype and away from other classes' prototypes, which is what protects dense predictions from the misleading gradients of noisy pseudo-labels. Inference uses DDIM to iterate from noise to a predicted mask.","core_discovery":"The paper claims that a conditional diffusion decoder, guided by multi-level visual features, can learn semantic label distributions even when supervision mixes ground truth and noisy pseudo-labels, provided the latent embedding space is explicitly structured. The structuring device is a set of non-learnable class prototypes that act as fixed anchors: pixel embeddings are assigned to prototypes by online clustering, and a prototype-anchored contrastive loss enforces intra-class compactness and inter-class separability. With this mechanism, the model reaches 56.68 mIoU on Endoscapes2023 with GT plus pseudo labels, 55.33 with GT only, and 43.04 on the MOSXAV test set, surpassing MedSegDiff on every setting. The same paper releases MOSXAV, a public benchmark with 62 X-ray angiography video sequences and manually annotated ground truth.","pith_inferences":["An implication the paper leaves implicit is that the same prototype-anchored contrastive constraint could be grafted onto non-diffusion segmentation heads, offering a cheaper way to test whether the mechanism, rather than the diffusion decoder, is responsible for the gains. A natural next step is to measure inter-observer agreement on MOSXAV's manual annotations, since the paper does not report it"],"forward_implications":["Semi-supervised medical segmentation can be improved by anchoring latent semantics to fixed prototypes instead of relying only on mask-level consistency. If the reported gains are reproducible, noisy pseudo-labels become far less damaging during training. MOSXAV gives the community a new public benchmark for multi-object X-ray angiography segmentation, with 2,335 densely annotated training and val"],"supporting_citations":[{"why":"MedSegDiff is the main diffusion-based baseline the proposed method must beat.","marker":"[32]"},{"why":"DDP supplies the label-denoising formulation and asymmetric time intervals used in the diffusion decoder.","marker":"[9]"},{"why":"ViT is the visual backbone that produces the multi-level features conditioning the decoder.","marker":"[6]"},{"why":"DDIM is the inference update rule that iteratively converts noise into a predicted mask.","marker":"[28]"},{"why":"MedSAM generates the pseudo masks from object detection boxes for Endoscapes2023.","marker":"[16]"},{"why":"SAMRefiner refines the MedSAM masks into higher-quality pseudo labels.","marker":"[12]"},{"why":"SimCLR motivates the nonlinear projection head used for contrastive learning over feature maps.","marker":"[3]"},{"why":"The cosine schedule controls the noise level in the forward diffusion process.","marker":"[23]"},{"why":"Endoscapes2023 is the public dataset used for the main comparison.","marker":"[22]"},{"why":"CADICA is the source of 40 of the 62 MOSXAV video sequences.","marker":"[10]"}],"fun_headline_variants":["New X-ray benchmark and diffusion model for semi-supervised medical segmentation","Prototype anchors improve diffusion robustness to noisy pseudo-labels","Diffusion model beats SOTA on semi-supervised medical segmentation","Prototype anchors beat noisy pseudo-labels in diffusion model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stated gains assume that the baseline methods were trained under settings as favorable as the proposed method's; if the baselines used smaller backbones, lower input resolution, or fewer iterations, the +5.15 mIoU gap could reflect training budget rather than the prototype-contrastive mechanism.","fun_headline_variants_meta":{"raw":{"variants":["New X-ray benchmark and diffusion model for semi-supervised medical segmentation","Prototype anchors improve diffusion robustness to noisy pseudo-labels","Diffusion model beats SOTA on semi-supervised medical segmentation","Prototype anchors beat noisy pseudo-labels in diffusion model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3381,"prompt_tokens":955,"completion_tokens":2426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2356}},"tokens_in":571,"tokens_out":2426,"duration_ms":18455,"temperature":1.0,"reasoning_tokens":2356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:09:44.052336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain all Table 1 baselines with the same ViT-Large backbone, 512x512 input resolution, 40,000 iterations, batch size 8, and the same MedSAM plus SAMRefiner pseudo-label protocol. If the mIoU gap over MedSegDiff shrinks to near zero while the prototype losses are removed from the proposed model, the core claim fails.","supporting_citations":[{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"MedSegDiff is the main diffusion-based baseline the proposed method must beat."},{"cited_title":"In: Proceedings of the IEEE International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"DDP supplies the label-denoising formulation and asymmetric time intervals used in the diffusion decoder."},{"cited_title":"In: International Conference on Learning Representations (2021) 4, 8","cited_arxiv_id":null,"evidence_quote":"ViT is the visual backbone that produces the multi-level features conditioning the decoder."},{"cited_title":"In: International Conference on Learning Representations (2021) 3, 6","cited_arxiv_id":null,"evidence_quote":"DDIM is the inference update rule that iteratively converts noise into a predicted mask."},{"cited_title":"Nature Communications15(1), 654 (2024) 2, 8 Robust Noisy Pseudo-label Learning Using Diffusion Model 11","cited_arxiv_id":null,"evidence_quote":"MedSAM generates the pseudo masks from object detection boxes for Endoscapes2023."},{"cited_title":"In: International Conference on Learning Representations (2025) 8","cited_arxiv_id":null,"evidence_quote":"SAMRefiner refines the MedSAM masks into higher-quality pseudo labels."},{"cited_title":"In: Proceedings of the International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"SimCLR motivates the nonlinear projection head used for contrastive learning over feature maps."},{"cited_title":"In: Proceed- ings of the International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"The cosine schedule controls the noise level in the forward diffusion process."},{"cited_title":"Expert Systems41(12), e13708 (2024) 7","cited_arxiv_id":null,"evidence_quote":"CADICA is the source of 40 of the 62 MOSXAV video sequences."}],"review_version":1}