{"id":"d7ea9134-1f3a-44e8-aedd-ed5a412f6ca1","arxiv_id":"2505.16659","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SD-MAD adapts CLIP with shift adapters, an anchor-based loss, and automatic sign-prompt selection to detect and label multiple anomaly types from one or a few medical images.","lead":"This paper presents SD-MAD, a CLIP-based system that detects and classifies several different medical anomalies from just a few labeled images by aligning images with text descriptions of radiological signs. A general reader might care because few-shot multi-anomaly detection could help clinical screening tools work in hospitals with scarce labeled data and for rarer conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The anchor-loss and sign-selection mechanism assumes the unverified separation ordering in Remark 3.1 and the well-defined prompt labeling in Remark 3.3; the paper's own Limitation concedes prompts can be multi-category, so the central theoretical motivation is not established.","rationale":"The reader's weakest assumption is the right one, and the manuscript's own Limitation makes it more acute than the reader stated. The anchor loss and sign-selection inference are the two components that distinguish SD-MAD from a generic adapter on CLIP; both are justified by the separation/uniqueness postulates in Remarks 3.1 and 3.3. Those postulates are strong: for each category c, every image must be closer to every one of its own prompts than to the normal anchor, and the anchor must be closer than every prompt of every other category. In a high-dimensional CLIP embedding with diverse LLM-generated prompts, there is no a priori reason this holds, and the paper gives no empirical check. The finite-prompt version of Remark 3.3 is also computation-dependent: D_inf is defined via the unknown labeling function h, and no algorithm is given for estimating it, so sign selection is not fully specified. The paper's own limitation that prompts may correspond to more than one anomaly type directly violates the uniqueness required for h_text. I do not treat the missing error bars as the decisive issue: point estimates can be checked, and the reader already flagged them, but the theoretical mechanism is more load-bearing because it underwrites the design choices and the interpretation of Table 2. A direct empirical test of Remark 3.1's ordering, before and after training, would settle whether the claimed mechanism operates. If it does not, the method may still work, but the paper's central explanation would need to be revised, hence the verdict remains conditional rather than accept. The reader's conditional verdict is therefore unchanged.","tokens_in":11648,"tokens_out":8090,"duration_ms":68611,"concrete_test":"Rerun the pipeline on one fastMRI+ slice (e.g., slice 0) with the exact generated prompts and CLIP ViT-L/14 features. For every held-out test image, compute the three quantities in Remark 3.1 and report the fraction of images satisfying sup_i d(f_img, f_text,i) ≤ d(f_img, f_anchor) ≤ inf_{k≠c,j} d(f_img, f_text,j), both before and after SD-MAD training. Also compute subset accuracy separately on the satisfying versus violating subsets. If the satisfied fraction is low, or if subset accuracy is not higher on the satisfying subset, the anchor-loss mechanism and Eq. 8's decision rule are not validated, and the central claim should be conditioned on either relaxing Remark 3.1 or showing the loss still enforces the intended separation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's anchor loss (Eq. 4) and Section 3.3's binary rule (Eq. 8) are justified by Remark 3.1, which postulates that for every image feature of category c, sup_i d(f_img, f_text,i) ≤ d(f_img, f_anchor) ≤ inf_{k≠c,j} d(f_img, f_text,j). The paper provides no evidence that CLIP ViT-L/14 features satisfy this ordering, before or after training. This is not a harmless technicality: if the ordering fails, the hinge terms in Eq. 4 penalize geometrically meaningless directions, and Eq. 8's comparison against f_anchor is not a valid decision boundary. The same concern applies to Remark 3.3 and Eq. 6, which require a labeling function h_text with h_text(f_text)=c. The Conclusion's Limitation explicitly states that 'some prompts may correspond to more than one anomaly type,' so h_text is not well defined for the very prompts sign selection is meant to filter. Thus the method's central 'sign-driven' mechanism rests on an unverified and partially self-contradicted geometric postulate. The reported gains in Table 2 could in principle come from the adapters alone, with the anchor loss and sign selection contributing little of the claimed alignment; the ablations do not isolate the anchor loss from the shift adapter.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SD-MAD, a few-shot multi-anomaly detection method for medical images built on CLIP. It generates multiple radiological-sign text prompts per anomaly category via an LLM, inserts lightweight shift adapters into selected CLIP encoder layers, and trains with an image-text loss plus an inter-anomaly anchor loss that uses the normal-category text feature as a separating hyperplane. At inference, it selects reliable prompts by re-labeling text features based on nearest-neighbor distances and produces both continuous scores and binary predictions. The authors introduce three evaluation protocols (general AD AUROC, multi-label Hamming/subset accuracy, and category-wise AUROC) and report experiments on seven datasets, with the main multi-anomaly results on brain MRI slices from fastMRI+.","tokens_in":11987,"tokens_out":6917,"duration_ms":54354,"significance":"If validated, the paper makes a useful contribution: it defines a clinically motivated few-shot multi-anomaly task, proposes a concrete VLM adaptation, and evaluates it with three protocols over seven datasets. The use of LLM-generated sign prompts is interesting, and the 1-shot subset-accuracy gains on slice 0 are striking. However, the current evidence is not conclusive because the geometric assumptions underlying the anchor loss and sign selection are unverified and partly contradicted by the paper's own limitation statement, the component contributions are not isolated by ablation, and all few-shot numbers are single runs.","major_comments":[{"comment":"The inter-anomaly loss relies on the stipulated ordering sup_i d(f_c_img, f_c_text,i) ≤ d(f_c_img, f_anchor) ≤ inf_{k≠c,j} d(f_c_img, f^k_text,j). This is asserted rather than derived or verified. If this ordering does not hold for CLIP ViT-L/14 features, the hinge terms in Eq. (4) can push embeddings in directions unrelated to the intended separation, and the binary rule in Eq. (8) has no valid decision boundary. Please provide empirical verification of the ordering on held-out images before and after training (e.g., quantiles of the relevant distance differences), or a proof under the adapter parameterization. This is load-bearing because both training and inference use the anchor.","section":"§3.2, Remark 3.1 and Eq. (4)"},{"comment":"The sign-selection step assumes a labeling function h_text with h_text(f^c_text)=c for every prompt of category c. Yet the Conclusion's Limitation states that 'some prompts may correspond to more than one anomaly type.' For such prompts h_text is not well defined, so Definition 3.2 and Remark 3.3 do not apply to exactly the prompts that sign selection is intended to filter. Please specify how multi-category prompts are assigned, or restrict the selection to prompts for which h_text is a function, and quantify the effect of that restriction on the reported numbers.","section":"§3.3, Remark 3.3 and Eq. (6)"},{"comment":"All few-shot results are reported without standard deviations or number of seeds. With K=1, the choice of the single training sample is likely to dominate the outcome. Moreover, the benefit of sign selection is inconsistent: on slice 10 the average category-wise AUROC drops from 67.1% without sign selection to 61.5% with it, and on slice 5 the subset accuracy drops from 29.0% to 27.3%. Please report multiple seeds/random splits with error bars and significance tests, and temper the claim that sign selection yields consistent improvements.","section":"§4.3, Tables 2 and 3"},{"comment":"The comparison between CLIP/MedCLIP and 'Ours (no SS)' conflates the shift adapter with the inter-anomaly anchor loss. There is no ablation that trains the adapter with only L_img-text, without Lanchor, so the specific contribution of the anchor loss to the observed gains is not established. Please add this ablation, and also report a variant with Lanchor but no adapter if feasible.","section":"§4.3, Table 2 and §3.2"},{"comment":"The abstract claims 'consistent improvements over state-of-the-art baselines' across datasets. Table 1 does not support this at the individual-dataset level: MVFA is higher on BrainMRI (92.4 vs 91.4), RESC (96.2 vs 95.2), and HIS (82.7 vs 81.6), and on all three pixel-level rows MVFA is higher or equal. Only the average is higher. Please either report pairwise significance across runs or revise the claim to 'average improvement.'","section":"Abstract and Table 1"}],"minor_comments":[{"comment":"The baseline naming is inconsistent: the text mentions BRA and BGAD, but the table shows DRA, BGAD, MVFA, and Ours; please clarify which methods are included and use a consistent notation.","section":"§4.2 and Table 1"},{"comment":"The final row has a formatting error: '92.292.5' should be '92.2 / 92.5'.","section":"Table 1"},{"comment":"Please explain why the category 'Small vessel chronic white matter ischemic change' cannot be evaluated in slices 5 and 10, as this exclusion affects the comparability of the slice-wise averages.","section":"§4.3.2"},{"comment":"The figure lacks axis labels and a caption stating the exact dataset and slice setting; the text mentions the 5th slice but does not restate the protocol.","section":"Figure 3"},{"comment":"The paper repeatedly refers to the Appendix for dataset construction, prompt details, and metric definitions, but no appendix is included in the submitted manuscript; please provide it.","section":"Appendix"},{"comment":"References [10]/[11], [23]/[24], and [44]/[45] are duplicated; please consolidate the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript is a preprint without code or an appendix, and the central geometric postulates are plausibly checkable. I recommend major revision rather than rejection because the core idea is interesting and the requested additions—empirical verification of Remarks 3.1 and 3.3, component ablations, and multi-seed results—are feasible within the manuscript's scope. Please ask the authors to provide reproducibility materials with the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth reading for the task it defines, not for the mechanism it proposes. Few-shot multi-anomaly detection in medical images is a real gap, and the authors deserve credit for setting up three sensible evaluation protocols and for showing that a CLIP-based pipeline with adapters, LLM-generated sign prompts, and an inter-anomaly loss can move 1-shot subset accuracy on their Brain MRI benchmark from near zero to a usable level. The general AD numbers on BMAD are competitive with MVFA, even if not uniformly better.\n\nThe soft spots are real but fixable. The biggest one is theoretical: Remark 3.1 just postulates that CLIP features satisfy a separating ordering around the anchor, and Remark 3.3 assumes a clean labeling function for prompts that the authors themselves concede does not exist — the Limitation says some prompts map to more than one anomaly type. The paper frames these as Remarks rather than assumptions, and the binary prediction rule in Eq. 8 stands or falls on that ordering. The authors should either validate the ordering empirically on CLIP features or relax the claims and describe the loss as a heuristic.\n\nThe experimental evidence is also thinner than the prose. All 1-shot results are single runs with no variance, which matters a lot for results that are this stochastic. Sign selection — the headline inference contribution — reduces average category-wise AUROC on slices 5 and 10 relative to the no-sign-selection variant. The paper says this out loud, which is honest but also undermines the claim that sign selection is a core part of the method. The ablation on λ doesn't isolate the anchor loss from the shift adapter, so we don't know how much of the gain comes from the adapter alone. And the claim of 'consistent improvements' over baselines in the contributions is overstated: Table 1 has several rows where MVFA or others are better.\n\nWho should read it: people working on few-shot medical AD, especially anyone thinking about multi-label anomaly settings. The benchmark protocols themselves are a useful reference. It deserves a serious referee, but with the expectation of heavy revision — add error bars, isolate the components, and either prove or soften the geometric story.","headline":"A genuinely new few-shot multi-anomaly task with a reasonable CLIP-based method; evidence is thinner than the claims, and the sign-selection mechanism doesn't consistently help and rests on unproved geometric assumptions.","tokens_in":12505,"tokens_out":2786,"would_cite":false,"duration_ms":22819,"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":"SD-MAD aligns medical images with LLM-generated radiological sign prompts and an anchor-based inter-anomaly loss to detect and classify multiple anomalies from a few labeled examples, lifting 1-shot subset accuracy from 0.4% to 60.8%.","keywords":["few-shot anomaly detection","multi-anomaly detection","medical imaging","vision-language model","radiological signs","prompt selection","inter-anomaly loss"],"falsifier":"Take a held-out set of abnormal images with known categories, run the trained model, and compute how often the inequality in Remark 3.1 is violated—an image whose distance to its own category's signs exceeds its distance to the normal anchor, or whose anchor distance exceeds its distance to another category. If a substantial share of correctly classified images violates this ordering, the geometric assumption is not the actual source of the accuracy gain; a more direct test would ablate the anchor loss and compare subset accuracy on the same folds.","tokens_in":11442,"feed_emoji":"🩻","tokens_out":9568,"duration_ms":72380,"temperature":0.7,"pith_summary":"The paper tries to establish that few-shot medical anomaly detection can be expanded from a one-class normal-versus-abnormal task to a multi-class task that distinguishes several anomaly types using only a handful of labeled images. Its proposal, SD-MAD, anchors each anomaly category to a set of textual descriptions of radiological signs generated by a large language model, aligns image features to those sign descriptions, and then prunes unreliable signs at inference. The method's most striking evidence is a 1-shot brain MRI experiment where subset accuracy rises from 0.4% for the raw vision-language model to 60.8% for SD-MAD; across seven datasets and three evaluation protocols the paper reports improvements over baselines, with the largest gains in multi-label prediction. If the approach holds up, it would let clinicians separate conditions that need different treatments, such as tumor versus pneumonia, without large annotated datasets.","feed_headline":"Sign prompts lift 1-shot multi-anomaly detection from 0.4% to 60.8%","feed_subtitle":"LLM-written radiological signs let a vision-language model tell apart multiple anomalies with one labeled scan per type.","key_machinery":"The load-bearing object is the normal-prompt anchor $f_{\\text{anchor}}$, the text feature of the normal image prompt, which the paper treats as a separating hyperplane between each anomaly category and all others. Training uses a hinge-style loss (Eq. 4) with two terms: it pulls an image's features toward its own category's sign prompts only until they are closer than the anchor, and it pushes other categories' sign prompts away only until they are farther than the anchor. The second mechanism is sign selection at inference: a text feature is kept only if its distance to its own category's decision region is strictly smaller than its distance to every other category (Remark 3.3), and each category's score is the supremum cosine similarity over the surviving signs. The whole pipeline is wrapped in a shift adapter that interpolates learned features with the original transformer features in four image-encoder layers and one text-encoder layer, preserving the vision-language model's prior knowledge while adapting to the few-shot data.","core_discovery":"The central claim is that multi-anomaly detection in medical images can be driven by text: each anomaly category is represented by a set of radiological signs written by a large language model, and the model is trained so that an image embedding is closer to its own category's sign embeddings than to a normal-text anchor, which in turn is closer than to any other category's sign embeddings. This anchor-based inter-anomaly loss amplifies the separation between anomaly categories in the shared vision-language embedding space. At inference, a second geometric condition (Remark 3.3) is used to discard sign prompts that lie too far from their own category's decision region, reducing the noise from signs that match multiple categories. The paper argues that together these two mechanisms convert a frozen vision-language model into a few-shot multi-anomaly detector, and reports that on slice 0 of a brain MRI dataset the subset accuracy rises from 0.4% for vanilla CLIP to 60.8% with the full method, while the Hamming score rises from 80.2% to 87.2%.","pith_inferences":["A likely failure mode not tested in the paper: if the normal-prompt anchor does not consistently sit between an anomaly category and the others in real embedding spaces, the anchor loss can pull features in the wrong direction; measuring the rate of Remark 3.1 violations on held-out images would quantify this risk.","The sign-selection criterion (Remark 3.3) is defined entirely in text-feature space, so it could be applied to any multimodal classifier that generates multiple textual descriptions per class, such as few-shot object detection from captions, without retraining.","The paper's observation that sign selection sometimes hurts category-wise AUROC suggests that outlier prompts can still be informative for some test images; an alternative is to soften selection into a confidence-weighted average rather than a hard discard, which the authors do not explore.","The multi-anomaly experiments are built from one anatomical site (brain MRI); extending the same three protocols to chest X-rays or pathology would test whether the sign-alignment mechanism transfers across modalities and sign vocabularies."],"forward_implications":["A single vision-language model can handle multiple anomaly categories at once, so updating the system for a new anomaly type requires only new textual signs, not a new image dataset or retraining from scratch.","The sign-selection step makes the detector more robust to imperfect LLM-generated prompts: prompts that match multiple categories or are underfit by few-shot training are discarded automatically at inference.","The anchor loss is additive, so it can be combined with pixel-level anomaly detection losses, as the paper does with a pixel-level baseline, to keep segmentation capability while adding multi-category separation.","The three proposed protocols—general anomaly detection, multi-label prediction, and category-wise AUROC—define a way to measure multi-anomaly detection performance that existing one-class scoring functions cannot support."],"supporting_citations":[{"why":"Supplies the pre-trained vision-language backbone whose image and text embeddings the method aligns and adapts; it is also the primary zero-shot baseline in Table 2.","marker":"[35]"},{"why":"MVFA is the few-shot medical anomaly detection baseline that SD-MAD is compared against in Table 1 and whose losses are combined with the inter-anomaly loss to obtain pixel-level scores.","marker":"[23]"},{"why":"BMAD defines the six benchmark datasets and evaluation conventions (image- and pixel-level AUROC) used for the general anomaly detection experiments.","marker":"[3]"},{"why":"fastMRI+ provides the brain MRI volumes with multiple anomaly categories and slice-level labels used to build the multi-anomaly detection experiments.","marker":"[48]"},{"why":"MedCLIP is the medical vision-language baseline compared in the multi-label and category-wise protocols, and it exemplifies the uncertain sign-alignment problem that motivates sign selection.","marker":"[40]"},{"why":"The CARES benchmark documents unreliable and noisy prompt-image alignment in medical vision-language models, which the paper cites as the reason not all radiological sign prompts are trustworthy.","marker":"[41]"},{"why":"Supplies the automatic prompt augmentation and selection approach that the inference-time sign selection strategy adapts.","marker":"[38]"}],"fun_headline_variants":["LLM signs power 1-shot multi-anomaly detection to 60.8% accuracy","Sign prompts lift multi-anomaly detection from 0.4% to 60.8%","Text-driven signs turn CLIP into few-shot multi-anomaly detector","One scan per type: sign anchors boost anomaly detection to 60.8%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a single 'normal' text prompt always sits between each anomaly category and every other category in the similarity space: every abnormal image is closer to its own category's signs than to normal, and closer to normal than to any other category's signs.","fun_headline_variants_meta":{"raw":{"variants":["LLM signs power 1-shot multi-anomaly detection to 60.8% accuracy","Sign prompts lift multi-anomaly detection from 0.4% to 60.8%","Text-driven signs turn CLIP into few-shot multi-anomaly detector","One scan per type: sign anchors boost anomaly detection to 60.8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000485,"raw_usage":{"total_tokens":2433,"prompt_tokens":1023,"completion_tokens":1410,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":1318}},"tokens_in":639,"tokens_out":1410,"duration_ms":10753,"temperature":1.0,"reasoning_tokens":1318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:56:39.924222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of abnormal images with known categories, run the trained model, and compute how often the inequality in Remark 3.1 is violated—an image whose distance to its own category's signs exceeds its distance to the normal anchor, or whose anchor distance exceeds its distance to another category. If a substantial share of correctly classified images violates this ordering, the geometric assumption is not the actual source of the accuracy gain; a more direct test would ablate the anchor loss and compare subset accuracy on the same folds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BMAD defines the six benchmark datasets and evaluation conventions (image- and pixel-level AUROC) used for the general anomaly detection experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"fastMRI+ provides the brain MRI volumes with multiple anomaly categories and slice-level labels used to build the multi-anomaly detection experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MedCLIP is the medical vision-language baseline compared in the multi-label and category-wise protocols, and it exemplifies the uncertain sign-alignment problem that motivates sign selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The CARES benchmark documents unreliable and noisy prompt-image alignment in medical vision-language models, which the paper cites as the reason not all radiological sign prompts are trustworthy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the automatic prompt augmentation and selection approach that the inference-time sign selection strategy adapts."}],"review_version":1}