{"id":"5d6353fa-497b-43aa-a775-8acc86e3348c","arxiv_id":"2504.17628","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ADZUS uses self-attention maps from pretrained Stable Diffusion to segment diabetic foot ulcers with no labeled training data, achieving IoU 86.68% on the chronic wound dataset.","lead":"This paper tests an unsupervised segmentation pipeline that extracts wound boundaries from Stable Diffusion's internal attention maps, without training on labeled wound images. It reports results comparable to or better than supervised models on two diabetic foot ulcer datasets, with a text prompt as the only user input.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADZUS's reported superiority is unverifiable: the paper never specifies how the binary wound mask is extracted from its multi-proposal output, and the published ADZUS IoU/DSC values violate the exact identity DSC=2IoU/(1+IoU) that all other baselines satisfy.","rationale":"The reader's weakest assumption concerned the diffusion timestep t=300 and hand-chosen attention-merge hyperparameters. That is a real sensitivity question, but it is downstream of a more basic evaluation issue: before asking whether t=300 produces valid attention maps, one must know how the multi-proposal ADZUS output is turned into the binary wound mask that is scored. If that step uses the ground truth or is otherwise unspecified, no choice of t can rescue the zero-shot claim. I therefore focus the stress-test on the missing proposal-selection rule and on the concrete numerical inconsistencies in the reported metrics, which are checkable from the manuscript alone. The metric inconsistency is especially telling because all six supervised baselines in Table II satisfy DSC = 2IoU/(1+IoU) exactly, while the ADZUS row does not, and the custom-dataset medians also violate the same monotone identity. These are not accusations of misconduct; they are unresolved technical gaps that a code and mask release would settle. I agree with the reader's CONDITIONAL verdict: the idea is interesting and the authors did attempt a contamination check, but the paper is not accept-ready without a fully specified, label-free evaluation pipeline and consistent metrics. If the checks fail, the verdict should move to REJECT; if they pass, the conditions are met and the claim can be reconsidered. I mark agreement as partial because my load-bearing concern is the evaluation protocol rather than the specific t=300 hyperparameter, although both point to the need for a reproducible, sensitivity-tested implementation.","tokens_in":15007,"tokens_out":13852,"duration_ms":134634,"concrete_test":"Request the inference code and the exact per-image binary masks for the 200-image chronic wound test set and the 40-image custom set. Then (a) confirm that the binary mask is generated by a fixed, ground-truth-independent rule (for example, the proposal with the highest prompt-conditioned attention score), and (b) recompute IoU, DSC, precision, and recall both globally and per image, checking the identities DSC = 2IoU/(1+IoU), IoU = pr/(p+r-pr), and median(DSC) = 2*median(IoU)/(1+median(IoU)). If the published numbers do not reproduce, or if the ground truth is needed to select the wound proposal, the zero-shot superiority claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a zero-shot, label-free diffusion model beats supervised SOTA on DFU segmentation. That claim stands or falls with the binary masks used in Tables II and Figure 6. But Section II.B.3 states that ADZUS outputs multiple unlabeled object proposals and that users may 'interactively select the relevant region of interest'; no automatic rule is given for converting the proposal stack into the single wound mask that is scored against clinical ground truth. If the wound proposal was selected using the ground truth, or via an unspecified oracle, the reported IoU/DSC are not zero-shot results. This is not a pedantic gap: the reported numbers are internally inconsistent in a way that all other baselines are not. With precision 94.69% and recall 92.46%, the aggregate IoU must be pr/(p+r-pr) = 87.90% and DSC = 2pr/(p+r) = 93.56%, yet Table II lists IoU 86.68% and DSC 91.98%. Every other row in Table II satisfies these identities to rounding, and Section III.A's text even cites a DSC of 93.56 for ADZUS. On the custom dataset, the stated median DSC of ~75% and median IoU of ~68% cannot both be medians of the same per-image segmentations, because DSC = 2IoU/(1+IoU) is monotone and would force median DSC to be ~81% at median IoU 68%. These inconsistencies indicate the evaluation protocol is unspecified or erroneous, so the superiority claim cannot be verified from the paper alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADZUS, a zero-shot unsupervised segmentation method for diabetic foot ulcers built on self-attention maps from a pretrained Stable Diffusion model. The method aggregates attention tensors from multiple resolutions, iteratively merges attention maps using KL divergence, applies non-maximum suppression to obtain object proposals, and optionally incorporates text prompts to guide the segmentation. On the chronic wound dataset (200 test images) the authors report IoU 86.68%, precision 94.69%, recall 92.46%, and DSC 91.98%, claiming to outperform supervised models including FUSegNet. On a custom-curated dataset of 40 white-light images, they report a median DSC of about 75% and IoU of about 68% for ADZUS, versus about 45% and 50% for FUSegNet. The paper also demonstrates text-guided segmentation qualitatively.","tokens_in":15400,"tokens_out":7777,"duration_ms":72301,"significance":"The underlying idea of exploiting pretrained diffusion self-attention for label-free medical image segmentation is timely and potentially valuable, and the authors include a data-contamination check using a public search interface, which is a useful precaution. If the quantitative claims could be verified, the result that a zero-shot, label-free method is competitive with supervised state-of-the-art on diabetic foot ulcer segmentation would be of clear interest to the biomedical imaging community. However, the central quantitative claims are not currently verifiable: the binary mask extraction protocol is unspecified, the reported metrics are internally inconsistent, and the key hyperparameters are not documented. The paper also provides no code, no sensitivity analysis, and no uncertainty quantification, so the headline claim of outperforming supervised models is not supported at the current level of reporting.","major_comments":[{"comment":"The conversion from the multi-proposal ADZUS output to the single binary wound mask scored in Tables II and Figure 6 is never specified. Section II.B.3 states that ADZUS 'generates a comprehensive segmentation mask that delineates multiple regions' and that 'clinicians or users ... interactively select the relevant region of interest.' If the proposal used for evaluation was selected with knowledge of the ground truth, or by an unspecified oracle, then the reported IoU/DSC values are not zero-shot results. The authors must provide the exact automatic selection rule, or a fully specified human protocol that is blind to the ground truth, and ideally release code; without this, the central superiority claim cannot be verified from the paper alone.","section":"Section II.B.3, Section III.A, Table II"},{"comment":"The metrics reported for ADZUS are internally inconsistent. Given precision 94.69% and recall 92.46%, the pooled IoU must be 87.90% and DSC 93.56%, yet Table II lists IoU 86.68% and DSC 91.98%, while the text in Section III.A cites a DSC of 93.56. Every other row in Table II satisfies the identities IoU = PR/(P+R-PR) and DSC = 2IoU/(1+IoU) to rounding. If the metrics are per-image averages rather than pooled values, the aggregation protocol must be stated explicitly and applied uniformly; as written, the table is internally contradictory and the comparison to the baselines cannot be interpreted.","section":"Table II; Section III.A"},{"comment":"The custom-curated dataset results are also numerically impossible as stated. Because DSC = 2IoU/(1+IoU) is monotone, the median DSC must equal 2 times the median IoU divided by (1 plus the median IoU); a median IoU of about 68% would imply a median DSC of about 81%, not 75%, and FUSegNet's median IoU of about 50% would imply a median DSC of about 67%, not 45%. The authors need to report the paired per-image statistics or clarify which quantity is the median; the current numbers cannot both be medians of the same segmentations.","section":"Section III.B, Figure 6"},{"comment":"The method relies on several hand-set hyperparameters, including the diffusion timestep t about 300, the resolution weights R_k, the KL threshold tau, the anchor grid size M, and the iteration count N, but the paper does not report their values or provide a sensitivity analysis. The assertion that t=300 places clean images in the reverse-diffusion regime where attention maps reflect semantic grouping is central to the method and is stated without evidence. Without this information, the results are not reproducible and there is no check that the authors did not tune these parameters on the test set.","section":"Section II.A, Section II.B"},{"comment":"The claimed superiority over FUSegNet rests on a 0.28 percentage-point IoU difference (86.68% versus 86.40%) with no error bars, confidence intervals, or significance test. Given the metric inconsistencies described above, the statement that ADZUS 'surpasses' supervised state-of-the-art models is not supported at the current level of reporting; the authors should provide uncertainty quantification or qualify the claim as competitive rather than superior.","section":"Section III.A, Table II"}],"minor_comments":[{"comment":"The definitions should state whether TP/FP/FN are pooled over the dataset or computed per image, and how per-image metrics are aggregated.","section":"Appendix A, Eqs. (A1)-(A4)"},{"comment":"The rendered equation is incomplete; the aggregation weights R_k and the upsampling notation should be defined explicitly so that the method is reproducible.","section":"Section II.B.1, Eq. (2)"},{"comment":"The figure should report the underlying distribution, for example as box plots with n=40, and the text should state whether the reported IoU and DSC values are medians of the same per-image scores.","section":"Section III.B, Figure 6"},{"comment":"Reference [37] is described as the 'LAION-5B search tool,' but 'Have I Been Trained?' is a retrieval interface; the authors should describe its search modality and the limitations of the contamination check.","section":"Section I.B, Reference [37]"},{"comment":"The abstract contains the typo 'text-gui ded' and should read 'text-guided.'","section":"Abstract"},{"comment":"The text-guided workflow in Section II.B.4 repeats the three-step description already given in Sections II.B.1-II.B.3; consider consolidating the two descriptions to avoid redundancy.","section":"Section II.B.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's quantitative claims are not verifiable in its current form. I recommend major revision rather than rejection because the proposal-selection protocol and the metric definitions are in principle fixable; however, if a revision does not reconcile the reported numbers, supply the selection rule and hyperparameters, and provide uncertainty quantification, I would not recommend acceptance. The novelty relative to DiffuMask and DAS should also be clarified in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper applies the DAAS/DiffuMask trick—aggregating self-attention maps from a pretrained Stable Diffusion model and merging them into proposals—to diabetic foot ulcer segmentation. The application is new and clinically motivated, and the authors checked LAION-5B for training contamination, which is responsible. If the method works, it would ease the annotation bottleneck in wound care.\n\nBut the central claim, that zero-shot ADZUS beats supervised FUSegNet, is not verifiable from the reported numbers. Table II lists ADZUS IoU 86.68, precision 94.69, recall 92.46, DSC 91.98. Those precision/recall values imply DSC ≈ 93.56 and IoU ≈ 87.90. The text even cites a DSC of 93.56, so the table is internally inconsistent. On the custom dataset, median DSC ~75% and IoU ~68% cannot both be medians if DSC=2IoU/(1+IoU) is computed per image. That points to an erroneous or unreported evaluation protocol.\n\nThe bigger gap is that ADZUS outputs multiple unlabeled proposals, and the paper says users can \"interactively select the relevant region of interest.\" There is no automatic rule for choosing the wound proposal that gets scored. If the selection used ground truth or an oracle, the zero-shot claim collapses. This needs to be fixed before the results are trustworthy.\n\nOther soft spots: the method is a minor variant of DAAS; hyperparameters (t=300, merge threshold, anchor grid) are hand-picked with no sensitivity analysis; the custom dataset is small (40 images) and the FUSegNet comparison is confounded by domain shift. To be fair, the stress-test's claim that every other row in Table II satisfies the identity is not quite right—LinkNet's row doesn't—but the ADZUS inconsistency is real.\n\nDespite these issues, the idea is worth pursuing and the problems are addressable. I'd send it to a serious referee, but with a clear request to specify the mask-selection procedure and report consistent metrics. Needs major revision before reliable.\n\nRecommendation: engage, but require fixes on the evaluation protocol.","headline":"A clinically motivated zero-shot segmentation pipeline built on existing diffusion-attention methods, but internally inconsistent metrics and an unspecified mask-selection step make the headline claim unverifiable.","tokens_in":15925,"tokens_out":5461,"would_cite":false,"duration_ms":46838,"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":"The paper claims that zero-shot, label-free segmentation of diabetic foot ulcers can beat supervised state-of-the-art models by reading self-attention maps from a pretrained diffusion model.","keywords":["diabetic foot ulcer segmentation","zero-shot learning","self-attention","diffusion models","text-guided segmentation","medical image segmentation","unsupervised segmentation","wound assessment"],"falsifier":"Sweep the diffusion time-step from $t=100$ to $t=500$ and vary the resolution weights and merge threshold on the same 200-image chronic-wound set, then compare each resulting mask to the ground-truth tracings. If IoU and precision drop steeply or the masks shift off the wound under plausible settings, the zero-shot attention maps are not robustly encoding wound anatomy and the reported leaderboard numbers would not generalize.","tokens_in":14823,"feed_emoji":"🩹","tokens_out":6490,"duration_ms":59215,"temperature":0.7,"pith_summary":"This paper proposes that the internal self-attention maps of a pretrained image-generation model, Stable Diffusion, already contain enough object-grouping information to outline diabetic foot ulcers in real photographs without any labeled training data. The authors build ADZUS, a pipeline that extracts those maps, combines them across resolutions, merges similar regions, and turns the result into a segmentation mask guided by a text prompt. On the public chronic-wound test set, ADZUS reports the highest intersection-over-union (86.68%) and precision (94.69%) among compared models, narrowly ahead of the supervised FUSegNet; on a small clinician-curated white-light DFU set, it reports a median dice score of 75% versus about 45% for FUSegNet. If these numbers hold, zero-shot segmentation could remove the annotation bottleneck for wound assessment and let clinicians steer segmentation by describing what they are looking for.","feed_headline":"Label-free diffusion segmentation beats supervised wound models on DFUs","feed_subtitle":"A text-guided Stable Diffusion attention pipeline hits 86.68% IoU on chronic wounds and 75% median DSC on curated DFUs.","key_machinery":"The load-bearing mechanism is the self-attention tensor of the diffusion U-Net: for each spatial location, a 2D attention map records how strongly every other location responds to it, and the paper assumes locations belonging to the same object have similar maps. ADZUS aggregates the 16 tensors from four resolutions ($8\\times8$, $16\\times16$, $32\\times32$, $64\\times64$) by bilinear upsampling and weighted summation, then merges nearby maps whose KL divergence is below a threshold, and applies non-maximum suppression to produce a binary mask. The aggregation weights proportional to resolution, the chosen threshold, and the time-step $t=300$ are hand-set hyperparameters; text guidance enters through cross-attention to condition which groups are emphasized.","core_discovery":"The paper's central claim is that the self-attention layers inside a pretrained Stable Diffusion U-Net encode spatial object groupings that are good enough to segment medical wounds without any training. ADZUS treats a real wound photograph as if it were a near-final denoised generated image, runs one unconditioned diffusion pass at a large time-step ($t=300$), and collects the 16 self-attention tensors produced at resolutions of 8, 16, 32, and 64. These tensors are upsampled, aggregated with resolution-proportional weights, merged iteratively using KL divergence between attention maps, and collapsed by non-maximum suppression into a final mask. A text prompt supplied through the model's cross-attention path shapes which regions the merged attention emphasizes. The paper reports that this label-free procedure reaches an IoU of 86.68% and precision of 94.69% on the chronic-wound benchmark, surpassing FUSegNet, and a median DSC of about 75% on its own 40-image DFU dataset, far above FUSegNet's 45%.","pith_inferences":["A sensible stress test would sweep $t$ and the merge threshold; if performance is stable across a wide range, the attention maps are genuinely carrying wound structure, and if it is not, the reported numbers may be tuned to the two datasets.","The text-guided examples show different prompts producing different masks, which suggests an untapped evaluation axis: measuring whether prompt semantics align with the selected tissue zones, rather than only measuring boundary overlap with a single ground-truth tracing.","Because ADZUS uses a frozen pretrained model, it could be combined with active learning: a handful of labeled wound images could select the best prompt and hyperparameters per clinic, keeping the zero-shot core while adapting to local imaging conditions."],"forward_implications":["If the reported results are correct, automated wound segmentation no longer needs a task-specific annotated training set; a general-purpose diffusion model can be pointed at a wound photograph and a descriptive prompt.","Text-guided customization becomes practical: changing the prompt to emphasize infection, inflammation, or granulation changes the output mask, so a clinician could request the region of interest without retraining.","The same zero-shot attention pipeline should transfer to other anatomical targets such as retinal vessels or tumors with prompt engineering alone, a claim the paper explicitly lists as future work.","On the benchmark, a label-free method matching a supervised specialist model implies the main cost of applying AI to new wound-imaging setups shifts from annotation to choosing prompts and hyperparameters."],"supporting_citations":[{"why":"Provides the pretrained latent diffusion model whose 16 self-attention tensors ADZUS extracts and aggregates.","marker":"[38]"},{"why":"Shows cross-attention layers in diffusion models group objects into segments, the observation ADZUS extends to unconditional self-attention.","marker":"[41]"},{"why":"Defines FUSegNet, the supervised state-of-the-art baseline that ADZUS reports surpassing in IoU and precision.","marker":"[43]"},{"why":"Supplies the public chronic-wound dataset of 810 training and 200 test images used for the main benchmark.","marker":"[26]"},{"why":"Used to verify that the test images do not appear in Stable Diffusion's LAION-5B training corpus, supporting the zero-shot claim.","marker":"[37]"},{"why":"Provides the related unsupervised zero-shot segmentation approach with Stable Diffusion attention that ADZUS builds on.","marker":"[42]"}],"fun_headline_variants":["Zero-shot diffusion beats supervised wound segmentation","Text-guided attention maps segment wounds without labels","Diffusion self-attention reaches 86.7% IoU on diabetic ulcers","ADZUS: zero-shot wound masks from a simple text prompt"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the premise that the hand-picked diffusion settings, especially the time-step $t=300$ and the resolution weights, place real wound photographs in the attention regime where the self-attention maps mark actual wound boundaries, and the paper does not report sensitivity to these settings.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot diffusion beats supervised wound segmentation","Text-guided attention maps segment wounds without labels","Diffusion self-attention reaches 86.7% IoU on diabetic ulcers","ADZUS: zero-shot wound masks from a simple text prompt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000356,"raw_usage":{"total_tokens":1972,"prompt_tokens":1026,"completion_tokens":946,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":879}},"tokens_in":642,"tokens_out":946,"duration_ms":10560,"temperature":1.0,"reasoning_tokens":879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:34:33.059127+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep the diffusion time-step from $t=100$ to $t=500$ and vary the resolution weights and merge threshold on the same 200-image chronic-wound set, then compare each resulting mask to the ground-truth tracings. If IoU and precision drop steeply or the masks shift off the wound under plausible settings, the zero-shot attention maps are not robustly encoding wound anatomy and the reported leaderboard numbers would not generalize.","supporting_citations":[{"cited_title":"High-Resolution Image Synthesis With Latent Diffusion Models,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained latent diffusion model whose 16 self-attention tensors ADZUS extracts and aggregates."},{"cited_title":"Fully automatic wound segmentation with deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the public chronic-wound dataset of 810 training and 200 test images used for the main benchmark."},{"cited_title":"Have I been Trained ?","cited_arxiv_id":null,"evidence_quote":"Used to verify that the test images do not appear in Stable Diffusion's LAION-5B training corpus, supporting the zero-shot claim."},{"cited_title":"Diffuse Attend and Segment: Unsupervised Zero-Shot Segmentation using Stable Diffusion,","cited_arxiv_id":null,"evidence_quote":"Provides the related unsupervised zero-shot segmentation approach with Stable Diffusion attention that ADZUS builds on."}],"review_version":1}