{"id":"6f85ec62-73b4-406e-883c-df3d781e496c","arxiv_id":"2505.17614","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A few-shot unsupervised method that combines contrastive losses on a normal-anatomy anchor bank with gradient-ascent synthetic anomalies to detect pathologies in brain MRI and chest X-rays.","lead":"PathoSCOPE is a method for finding diseases in medical scans using only two to eight healthy example images for training. It works by teaching the model what normal anatomy looks like and flagging anything that deviates from it as a possible pathology.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported few-shot SOTA may rest on an unverified evaluation protocol: BraTS2020 healthy-cohort construction, patient-level splits, and test-set η selection are unspecified, so the margin over GLASS-h may be leakage or tuning artifact.","rationale":"The abstract's claim is empirical: with 2–8 healthy samples, PathoSCOPE beats prior unsupervised methods. The most direct way for that claim to fail is not a conceptual gap in the loss, but an evaluation that does not actually measure few-shot generalization. Section 3.1 is the only description of the protocol, and it omits the two facts that determine validity: how 'healthy' was defined in BraTS2020, and whether splits are patient-disjoint. BraTS2020 is a glioma dataset without healthy controls, so '130 healthy samples' requires a nontrivial construction (e.g., tumor-free slices from tumor patients) that the paper does not disclose. If this construction is used, the healthy cohort is not a cohort of non-pathological subjects, and the data-efficiency claim is not tested on true healthy individuals. The same section does not state patient-level disjointness for ChestXray8. Additionally, Section 3.4 chooses η from test curves in Fig. 5, which is selective evaluation and can inflate the reported numbers; combined with no variance estimates, the margins over GLASS-h are not yet evidenced. If the requested metadata and validation protocol show clean separation and η selection, the empirical claim would be much stronger; hence the appropriate verdict remains conditional rather than accept or reject.","tokens_in":9160,"tokens_out":12216,"duration_ms":118345,"concrete_test":"Obtain from the authors the exact patient IDs, slice indices, and random seeds for all 10 repeats of Tables 1–2, and verify: (i) BraTS2020 'healthy' slices have zero tumor voxels by the official segmentation masks and come from distinct patients; (ii) no patient ID appears in both train and test in either dataset; (iii) the ChestXray8 training shots are unique healthy patients. Then re-run K=2–8 under a protocol in which all hyperparameters, especially η, are chosen on a held-out validation fold rather than on the test set, and report paired 95% CIs or Wilcoxon tests for Ours vs GLASS-h. If the AUROC advantage disappears or falls within noise, the central SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PathoSCOPE's central claim is that it reaches state-of-the-art few-shot UPD with 2–8 non-pathological training samples. For that claim to hold, the reported comparisons must be a clean measure of generalization to unseen healthy anatomy and real pathologies. Section 3.1 states only that training and test sets have 'no overlap'; it does not specify patient-level splits, subject IDs, or how BraTS2020—a glioma dataset with no healthy controls—was converted into 130 'healthy' test samples. If those healthy slices are tumor-free slices drawn from tumor patients, or if ChestXray8 repeats share patients across train/test, the few-shot setting is not a true normality-modeling test. In addition, Section 3.4 selects the key PiEG perturbation strength η by sweeping 'performance' (Figure 5), apparently on the test set; since η controls how far synthetic anomalies deviate from normal, test-set selection can inflate the reported AUROC. Without error bars or significance tests, the margins over GLASS-h (e.g., 89.19 vs 87.39 image AUROC at K=2; 72.23 vs 70.03 on ChestXray8) are not yet shown to be real. The reader's synthetic-realism concern compounds this: PiEG's gradient-ascent GPEs are validated only by a qualitative t-SNE (Fig. 2), so if the evaluation protocol is also loose, neither the data-efficiency nor the SOTA claim is secured.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PathoSCOPE, a few-shot unsupervised pathology detection framework that trains on only 2–8 non-pathological samples. The method combines a pre-trained feature extractor with a trainable feature adapter, a prototypical anchor bank, a Global-Local Contrastive Loss (GLCL), and a Pathology-informed Embedding Generation (PiEG) module that synthesizes pathological embeddings by gradient ascent on the discriminator loss. The authors report state-of-the-art results among unsupervised methods on BraTS2020 and ChestXray8, with image and pixel AUROC, PRO, and DICE metrics, as well as computational efficiency (2.48 GFLOPs, 166 FPS).","tokens_in":9476,"tokens_out":3561,"duration_ms":28911,"significance":"If the empirical claims hold, PathoSCOPE would be a meaningful step toward practical few-shot pathology detection, addressing the scarcity of healthy training data in medical imaging. The paper contributes a dual contrastive formulation that regularizes normal embeddings while synthesizing hard pathological embeddings, and it evaluates against a broad set of baselines on two medical datasets. The inclusion of efficiency metrics is a practical strength. However, the central claim of state-of-the-art few-shot performance rests on evaluation and hyperparameter-selection details that are currently underspecified, so the results are not yet fully supported.","major_comments":[{"comment":"The statement that training and test sets have 'no overlap' is insufficient for a few-shot normality-modeling claim. BraTS2020 is a glioma dataset without healthy control subjects; the manuscript does not state whether the 130 'healthy' test slices come from tumor patients (e.g., tumor-free slices from tumor patients) and whether patient identities are disjoint between training and test. If slices from the same patient appear in both training and test, the few-shot generalization results would be inflated. Please specify the exact construction of the healthy cohort and confirm patient-level disjointness.","section":"Section 3.1"},{"comment":"The perturbation strength η in Eq. (3) is selected per dataset by sweeping AUROC in Fig. 5, with no indication of a held-out validation split. Because η directly controls how far synthetic anomalies deviate from normal, selecting it on the test set can inflate the reported AUROC. Please report how η was chosen, provide validation curves with error bars, and clarify whether the reported test numbers correspond to a single selected η or to an oracle selection.","section":"Section 3.4 / Fig. 5"},{"comment":"All experiments are repeated 10 times but only averages are reported. Given the small margins over GLASS-h (e.g., ChestXray8 K=2: 72.23 vs 70.03; BraTS K=8: 87.84 vs 87.19), standard deviations and significance tests are needed to support the claim that PathoSCOPE is state of the art. Without variance information, the reported improvements may not be statistically meaningful.","section":"Tables 1–4"},{"comment":"As written, Eq. (1) divides by Σ Mp(h,w); for non-pathological embeddings ν_n the anomaly mask Mp is all zeros, so dlocal(ν_n) is undefined. The mask used for normal embeddings needs to be specified; otherwise the Local Contrastive Loss in Eq. (5) cannot be computed as described.","section":"Eq. (1)"},{"comment":"The authors acknowledge in the Conclusion that synthetic local pathologies 'lack explicit anatomical coherence,' which directly limits their resemblance to real localized pathologies. Since PiEG trains the discriminator against synthetic embeddings generated from the discriminator's own gradients, there is a risk that the model learns synthesis artifacts rather than real disease. The t-SNE evidence in Fig. 2 is qualitative only. Please add quantitative evidence that the synthetic-to-real transfer works, for example by ablating PiEG/GLCL and measuring performance on real pathologies, or by testing cross-synthesis robustness.","section":"Section 2.2 and Conclusion"}],"minor_comments":[{"comment":"The caption contains a typo: 'ChestXrat8' should be 'ChestXray8'.","section":"Fig. 5 caption"},{"comment":"In the sentence introducing Eq. (6), 'PGEs' should be 'GPEs' (Global Pathological Embeddings).","section":"Section 2.4"},{"comment":"The word 'auxillary' should be 'auxiliary'.","section":"Conclusion"},{"comment":"The coreset anchor budget is not specified; please state how many prototypes are retained in ΦA.","section":"Section 2.2"},{"comment":"The threshold τ used to classify an image as pathological is mentioned but never defined; please specify how it is set, especially for DICE computation.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed method is interesting, but the evaluation protocol has several load-bearing gaps: the BraTS healthy-cohort construction, patient-level split, test-set η selection, and lack of error bars. These are fixable with additional experiments and clarifications, so I recommend major revision rather than rejection. I would also encourage the editor to require the authors to release code and a precise evaluation pipeline, given that the reported margins are small."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: PathoSCOPE is a reasonable few-shot extension of the GLASS-style synthetic-anomaly recipe to medical imaging, with a new contrastive loss that seems to help. But the headline SOTA claim is not yet verifiable because the evaluation protocol has several unspecified choices that could inflate the numbers.\n\nWhat's actually new: the two-part Global-Local Contrastive Loss with a prototypical anchor bank (coreset-selected from ImageNet features) is a clean way to regularize both local and global embedding distances. The PiEG module is basically gradient-ascent synthesis from GLASS, but with the global loss as the guide, which is a sensible adaptation. The paper is also honest about the biggest weakness: the local synthetic patches drawn from DTD lack anatomical coherence, and they say so in the conclusion. Efficiency numbers (2.48 GFLOPs, 166 FPS) are nice and match SimpleNet/GLASS.\n\nThe soft spots are mostly in the experimental validation. Three things bother me.\n\nFirst, the construction of the BraTS2020 \"healthy\" test set is unexplained. BraTS is a glioma dataset; the paper says they sampled slices 60-90 and got 130 healthy samples, but it doesn't say whether these are tumor-free slices from tumor patients, or how they were labeled. If the \"normal\" test slices come from the same patients' tumor-free regions, the test is not a clean normality-modeling evaluation.\n\nSecond, the perturbation strength η is selected by sweeping performance (Fig. 5) with no mention of a held-out validation set. If that sweep is on the test set, the reported AUROC is optimistically biased, and the margin over GLASS-h (about 2 points) could be tuning artifact.\n\nThird, no error bars or significance tests despite \"all experiments repeated 10 times.\" With standard deviations, the 1-2 point gaps might vanish. The paper also never specifies the image-level threshold τ, which is needed for classification, and no code is released. Those are not fatal to the core idea, but they keep the central claim from being fully verified.\n\nMy take: the method is worth a serious referee, but only after the authors provide patient-level split details, a clear validation protocol for η, and error bars. As it stands, I'd treat the SOTA claim cautiously. I'd bring it to a reading group as an example of a few-shot UPD approach, but I wouldn't cite the numbers as established.","headline":"A plausible few-shot UPD method with a clean contrastive loss, but the evaluation protocol has enough holes that I'd want a revision before trusting the SOTA claim.","tokens_in":10032,"tokens_out":2595,"would_cite":false,"duration_ms":18904,"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's central claim is that an unsupervised pathology detector can be trained from as few as two healthy scans and still beat prior unsupervised methods at localizing brain tumors and chest diseases on BraTS2020 and ChestXray8.","keywords":["unsupervised pathology detection","few-shot learning","contrastive learning","synthetic embeddings","anomaly detection","medical imaging","BraTS2020","ChestXray8"],"falsifier":"Run the same 2-shot protocol on a third public benchmark with real localized lesions (for example, lung nodules on CT) and compare against GLASS-m with the identical backbone and compute budget; if PathoSCOPE does not match or beat GLASS-m there, the claim that its contrastive anchoring and pathology-informed synthesis enable few-shot detection fails to generalize.","tokens_in":8936,"feed_emoji":"🩻","tokens_out":11410,"duration_ms":118151,"temperature":0.7,"pith_summary":"The paper tries to establish that pathology detection in medical imaging does not need large healthy training sets or labeled disease examples. It introduces PathoSCOPE, which builds a normality reference from a few healthy scans and flags deviations as disease, claiming strong results with as few as two training scans on brain MRI and chest X-ray data. The method's two pillars are a contrastive loss that both compresses healthy-feature variation and enlarges deviations, and a synthesis module that invents pathological embeddings by pushing healthy features away along the discriminator's own gradient. If the claim holds, the practical bottleneck that hospitals lack representative healthy cohorts would be largely removed, and novel or rare diseases could be screened with minimal data.","feed_headline":"Two healthy scans may be enough to flag disease","feed_subtitle":"A contrastive anchor-bank method beats unsupervised baselines on brain MRI and chest X-ray with 2–8 normal samples.","key_machinery":"The carrying mechanism is a fixed prototypical anchor bank: a coreset-sampled set of healthy feature embeddings extracted by a pretrained ResNet18, which gives the model stable anatomical references. Two contrastive losses act on distances to the nearest anchor. The Local Contrastive Loss uses the Tritanh loss, $\\mathcal{L}_C(d_\\text{pull}, d_\\text{push}) = \\frac{e^{\\lambda_0 d_\\text{pull}} - e^{\\lambda_1 d_\\text{push}} + \\epsilon}{e^{\\lambda_0 d_\\text{pull}} + e^{\\lambda_1 d_\\text{push}} + \\epsilon}$, to pull non-pathological local embeddings toward the anchors and push local pathological embeddings (texture-overlaid patches) away. The Global Contrastive Loss does the same for whole-embedding distances, with global pathological embeddings produced by the PiEG module: twenty steps of normalized gradient ascent on the global loss, starting from Gaussian-perturbed healthy embeddings. The discriminator is then trained with binary cross-entropy on these synthetic positives and healthy negatives.","core_discovery":"On its own terms, the paper's claim is that few-shot unsupervised pathology detection can be made to work with as few as two non-pathological training scans. The method anchors healthy features to a prototypical bank, uses a local contrastive loss to suppress normal anatomical variation, and uses a global contrastive loss on gradient-ascent-synthesized pathological embeddings to sharpen the boundary between normal and diseased. Tested on BraTS2020 and ChestXray8 at 2, 4, 6, and 8 shots, PathoSCOPE reports the best image- and pixel-level scores among the unsupervised baselines it compares with, including 89.19% image AUROC on BraTS2020 at two shots and 72.23% on ChestXray8, at a cost of 2.48 GFLOPs and 166 FPS.","pith_inferences":["Beyond the paper, the anchor-bank-plus-gradient-ascent recipe likely transfers to non-medical few-shot anomaly detection, such as inspecting manufactured parts with very few known-good samples, because it does not encode disease-specific priors.","The modality-dependent optimal perturbation strength (0.01 for brain MRI, 0.05 for chest X-ray) suggests a future adaptive scheduler could set synthesis difficulty per image or per region rather than as a global constant.","If the authors' stated limitation is taken seriously, replacing DTD texture overlays with a learned local generator trained from the discriminator's gradients could both remove reliance on an auxiliary texture dataset and improve anatomical realism.","The few-shot property may enable privacy-preserving multi-site collaboration, where sites share only compact healthy anchor banks instead of raw patient images."],"forward_implications":["Clinical sites with only a handful of normal scans per population could bootstrap a screening model without assembling large healthy cohorts.","Because outputs are pixel-level anomaly maps, the method can localize the suspicious region, not merely classify the image.","The training pipeline needs no pathology labels, so a new disease or imaging modality can be addressed as soon as two healthy examples are available.","Reported throughput of 166 FPS at 2.48 GFLOPs indicates the trained model could run in real time on clinical hardware, supporting use in resource-constrained settings."],"supporting_citations":[{"why":"Supplies the gradient-ascent synthetic-embedding strategy that PiEG extends, and provides the GLASS-h and GLASS-m baselines that PathoSCOPE claims to outperform.","marker":"[5]"},{"why":"Provides the DTD texture dataset used to generate local pathological images during training.","marker":"[6]"},{"why":"Supplies the coreset sampling algorithm that selects the prototypical anchor bank of healthy embeddings.","marker":"[23]"},{"why":"It supplies the Tritanh (mean-shifted) contrastive loss used in the global-local contrastive objective.","marker":"[21]"},{"why":"Defines the discriminative self-supervised synthesis-to-discriminator paradigm and serves as a comparison baseline.","marker":"[29]"},{"why":"Contributes the projection-plus-discriminator design PathoSCOPE builds on and is a direct baseline.","marker":"[17]"},{"why":"ChestXray8 is one of the two evaluation benchmarks used for the few-shot pathology detection claims.","marker":"[25]"},{"why":"These references define the BraTS2020 benchmark on which the image-, pixel-, and region-level metrics are measured.","marker":"[19,2,3]"}],"fun_headline_variants":["Two healthy scans may flag disease: PathoSCOPE","Two healthy scans enough to flag brain and chest disease","Contrastive learning spots disease with two normal scans","PathoSCOPE: pathology detection from just two normal scans","Two-shot unsupervised pathology detection via contrastive learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fake diseases the method creates—texture patches stuck onto healthy scans and healthy examples artificially altered to look abnormal—are close enough to real diseases that a model trained only on fakes will flag the real ones.","fun_headline_variants_meta":{"raw":{"variants":["Two healthy scans may flag disease: PathoSCOPE","Two healthy scans enough to flag brain and chest disease","Contrastive learning spots disease with two normal scans","PathoSCOPE: pathology detection from just two normal scans","Two-shot unsupervised pathology detection via contrastive learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001034,"raw_usage":{"total_tokens":4345,"prompt_tokens":927,"completion_tokens":3418,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":3340}},"tokens_in":543,"tokens_out":3418,"duration_ms":19783,"temperature":1.0,"reasoning_tokens":3340,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:43:25.288001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 2-shot protocol on a third public benchmark with real localized lesions (for example, lung nodules on CT) and compare against GLASS-m with the identical backbone and compute budget; if PathoSCOPE does not match or beat GLASS-m there, the claim that its contrastive anchoring and pathology-informed synthesis enable few-shot detection fails to generalize.","supporting_citations":[{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"It supplies the Tritanh (mean-shifted) contrastive loss used in the global-local contrastive objective."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Defines the discriminative self-supervised synthesis-to-discriminator paradigm and serves as a comparison baseline."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Contributes the projection-plus-discriminator design PathoSCOPE builds on and is a direct baseline."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"ChestXray8 is one of the two evaluation benchmarks used for the few-shot pathology detection claims."}],"review_version":1}