{"id":"d7043830-d510-419e-8a98-c00cbb95f992","arxiv_id":"2507.19076","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"SP-Mamba pairs a Mamba autoencoder with sliding-window anatomical prototypes and concentration/contrast anomaly scoring, reporting state-of-the-art unsupervised medical anomaly detection on three benchmarks.","lead":"This paper introduces SP-Mamba, a Mamba-based neural network for spotting diseases in medical images like X-rays and MRIs without needing labeled examples of diseases. It reports top results on three medical benchmarks while using less compute than rival transformer-based methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA rests on hyperparameters selected on the ZhangLab test set; the concentration/contrast scoring weights (beta, gamma) are tuned on that test set and no validation split confirms they transfer to CheXpert and Uni-Medical.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the concentration/contrast statistics are assumed to separate normal from abnormal across all modalities, and the weights beta/gamma are tuned on the ZhangLab test set (Section 4.4). I agree with that assessment. This is the primary correctness risk because the paper's largest reported gains in the ablation (Table 4) come from the S_concen and S_contra terms, and those terms are the ones whose weights are most directly tuned on test data. Without a validation split, the ZhangLab results cannot support a claim of state-of-the-art performance, and the transfer of the tuned weights to CheXpert and Uni-Medical is plausible but unverified. I do not see an internal inconsistency in the method itself; the issue is empirical validation. The appropriate verdict remains CONDITIONAL, pending a proper train/validation/test evaluation and cross-dataset validation of the scoring weights. I considered whether the 'first Mamba in medical AD' claim (contradicted by MambaAD) or the lack of error bars on Uni-Medical should be the primary objection, but both are secondary to the test-set tuning issue, which directly affects the reliability of every reported performance number. The concrete test proposed above would settle whether the reported SOTA persists when hyperparameters are selected without looking at the test labels.","tokens_in":14836,"tokens_out":3458,"duration_ms":36708,"concrete_test":"Re-run the ZhangLab experiment with a proper three-way split: train on the 1249 healthy images, tune K, alpha, p, beta, gamma, sigma, and k_sigma on a validation split (e.g., a random 20% subset of the normal and abnormal test images), and evaluate the final model once on the held-out test split. Report AUC/Acc/F1 for the full model and for the model without S_concen/S_contra. Additionally, re-tune beta and gamma on a CheXpert validation split and report whether the optimal values differ by more than a factor of 2 from the ZhangLab-tuned values and whether the CheXpert AUC of 80.4 is stable under that re-tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that SP-Mamba reaches a new state of the art on ZhangLab, CheXpert, and Uni-Medical. Section 4.4 explicitly tunes all hyperparameters, including beta=-0.025 and gamma=400 for Eqs. (4), (5), and (7), on the ZhangLab test set. The ZhangLab numbers in Table 1 therefore are not an out-of-sample estimate; they are the result of test-set selection and can be inflated. Table 4 shows that adding S_concen and S_contra is responsible for the largest single jump on CheXpert (AUC 73.1 -> 80.4, compared to 86.2 -> 92.0 on ZhangLab), so the transferability of these two statistics and their weights to brain MRI, liver CT, and retinal OCT is load-bearing for the Uni-Medical results. The paper offers only two example anomaly maps (Figure 4) as evidence that 'concentrated vs dispersed' holds across modalities; no quantitative distributional analysis, no per-modality validation, and no error bars on Uni-Medical are provided. If the concentration/contrast premise fails on any modality, or if the ZhangLab-tuned weights are not near-optimal for the other datasets, the reported improvements over SimSID and MambaAD could shrink or reverse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SP-Mamba, an unsupervised medical anomaly detection model based on a Mamba state-space decoder. It contributes (i) a Circular-Hilbert scanning scheme that prioritizes central image regions for the Mamba blocks, (ii) a window-sliding medical prototype module that matches normal features against learned prototypes within a local window, and (iii) an anomaly scoring function that combines reconstruction error, prototype distance, and two post-hoc statistics of the anomaly map: a concentration score and a DoG-based contrast score. The method is evaluated on ZhangLab Chest X-ray, CheXpert, and Uni-Medical (brain/liver/retinal), reporting state-of-the-art AUC/Acc/F1 or AUC/AP/F1 with about 25.8M parameters and 8.3 GFLOPs. The central empirical claim is that SP-Mamba outperforms all compared methods, including SQUID, SimSID, and MambaAD, on all three benchmarks.","tokens_in":15216,"tokens_out":6066,"duration_ms":55790,"significance":"If the reported results are reproducible, SP-Mamba is a valuable contribution: it achieves competitive or better accuracy than transformer-based methods at roughly half the parameters and FLOPs, and it offers a set of design ideas (scanning order, prototypes, scoring statistics) that are clearly described and ablated. The paper ships code and the ablations in Table 4 isolate the contribution of each component. However, the main empirical claim is currently weakened by the fact that key hyperparameters, in particular the weights beta and gamma of the concentration and contrast scores, are selected on the ZhangLab test set with no validation split; this makes the reported SOTA on ZhangLab an optimistic, selection-biased estimate, and the transfer of those weights to CheXpert and Uni-Medical is not demonstrated. The significance of the work therefore depends on fixing this evaluation protocol.","major_comments":[{"comment":"Hyperparameters K, alpha, p, sigma, k_sigma, beta, and gamma are selected using AUC on the ZhangLab test set, and no validation split or cross-validation is described. Consequently, the ZhangLab numbers in Table 1 are not an independent estimate of generalization; the reported 92.0 AUC is the result of test-set selection. Please add a validation protocol (e.g., a held-out split of the training set, or repeated cross-validation on the training set) and report test-set performance only after fixing all hyperparameters. Also state explicitly whether the same hyperparameters are used for CheXpert and Uni-Medical or whether they were re-tuned per dataset.","section":"Section 4.4, Figure 6"},{"comment":"The ablation shows that adding S_concen and S_contra is the largest single improvement on CheXpert (AUC 73.1 to 80.4), yet the premise that anomalous maps are concentrated and high-contrast while normal maps are dispersed and low-contrast is supported only by two visual examples in Figure 4. Because beta and gamma are tuned on the ZhangLab test set, the CheXpert and Uni-Medical gains could reflect overfitting to the tuning statistic rather than a general property of medical anomaly maps. Please provide quantitative evidence: per-dataset distributions of S_concen and S_contra for normal versus abnormal images (e.g., box plots or histograms), and a sensitivity analysis of the Table 2 and Table 3 results to beta and gamma over a grid, to show the reported ranking is not an artifact of the chosen weights.","section":"Table 4, Eqs. (4), (5), (7)"},{"comment":"The statement that SP-Mamba \"outperforms all the comparative methods\" on CheXpert is not supported by the F1 column: SP-Mamba reports F1 = 75.8±1.5 while SQUID reports 75.9±5.7, so SP-Mamba's F1 is not higher. Please restrict the SOTA claim to AUC and Acc, or provide a statistical comparison (e.g., paired tests across runs) that justifies a claim of overall superiority.","section":"Table 2, Section 4.2"}],"minor_comments":[{"comment":"The recurrence for the Hilbert matrix H_{n+1} is typeset in a garbled, hard-to-read way; please rewrite it cleanly and verify the formula against a standard Hilbert curve construction.","section":"Eq. (3)"},{"comment":"The text says G is \"a two-dimensional Gaussian function with mean sigma,\" but the formula has zero mean and standard deviation sigma; please correct the wording to \"standard deviation sigma.\"","section":"Section 3.4, Eq. (5)"},{"comment":"The conclusion claims SP-Mamba \"attempts to firstly apply Mamba in medical AD,\" but MambaAD [11] already applies Mamba to medical anomaly detection on the Uni-Medical benchmark; please reword this claim to avoid inaccuracy.","section":"Section 5"},{"comment":"Uni-Medical results are reported as means without standard deviations or error bars; please add standard deviations across runs or specify that results come from a single run.","section":"Table 3"},{"comment":"The choice of the loss weight epsilon (set to 25) is stated but not analyzed; please include epsilon in the hyperparameter study or provide a brief justification for the selected value.","section":"Section 4.4 / Eq. (2)"},{"comment":"The captions for Figure 6 should state the dataset and metric (AUC) and should specify which hyperparameters are held fixed while the plotted one is varied.","section":"Figure 6"},{"comment":"The heading \"Affect of Medical-Prototype Module\" should read \"Effect of Medical-Prototype Module.\"","section":"Section 4.4 heading"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the test-set hyperparameter selection: beta, gamma, and several other hyperparameters are tuned on the ZhangLab test set, so the reported SOTA on ZhangLab is not a clean out-of-sample estimate, and the transfer to CheXpert and Uni-Medical is not demonstrated. If the authors add a proper validation split, show that the concentration/contrast weights transfer across modalities, and fix the overstated F1 claim on CheXpert, the paper would be suitable for publication. The novelty relative to MambaAD is incremental, but the efficiency and the empirical gains are potentially interesting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: SP-Mamba is a competent, incremental architecture paper whose headline SOTA numbers are partly an artifact of tuning hyperparameters on the ZhangLab test set. The core method is not circular, but the reported gains from the concentration and contrast scores are.\n\nWhat's actually new: the Circular-Hilbert scanning, the window-sliding prototype matching, and the concentration/contrast scoring. Each piece is a variation of known ideas (Hilbert curves, PatchCore-style memory, hand-crafted scoring), but the package is new and the ablations show each component contributes. The efficiency story is real: 25.8M params and 8.3 GFLOPs, about half of SQUID/SimSID, with better AUC on ZhangLab and CheXpert. Code is available.\n\nThe main soft spot is exactly what the stress-test note flags. Section 4.4 tunes K, p, alpha, beta, gamma, sigma, k_sigma on the ZhangLab test set. So the ZhangLab numbers in Table 1 are not an out-of-sample estimate; they are post-selection. Table 4 shows that on CheXpert the largest jump comes from adding S_concen and S_contra (73.1 to 80.4 AUC), so the transferability of those two scores and their ZhangLab-tuned weights is load-bearing. There are no error bars on Uni-Medical, no per-modality hyperparameter checks, and only two example anomaly maps as evidence that 'concentrated vs dispersed' holds across brain MRI, liver CT, and retinal OCT. That premise might hold, but it is unvalidated.\n\nThe 'first Mamba in medical AD' claim in the conclusion is also hard to defend, since MambaAD's Uni-Medical benchmark is medical. That claim can be dropped without hurting the method.\n\nThe base Mamba/prototype components are not circular, and the method itself is sound enough to deserve referee time. What I'd demand in review: a proper validation split or nested CV for hyperparameters, error bars on all three datasets, and a sensitivity analysis for beta/gamma on CheXpert and Uni-Medical. If those hold, this is a useful contribution for efficient medical AD. As it stands, I'd treat the ZhangLab result as promising but not proven.\n\nVerdict: send to peer review, but make the validation protocol a condition of acceptance.","headline":"Solid incremental method with a real test-set-selection problem: the ZhangLab SOTA is partly an artifact of tuning scoring weights on the test set, but the core architecture is worth reviewing.","tokens_in":15663,"tokens_out":2651,"would_cite":false,"duration_ms":23527,"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":"SP-Mamba claims state-of-the-art unsupervised medical anomaly detection on three benchmarks, using roughly half the compute of transformer-based rivals by making anatomical spatial structure explicit.","keywords":["unsupervised medical anomaly detection","state space models","Mamba","Circular-Hilbert scanning","window-sliding prototype learning","chest X-ray analysis","anomaly scoring","reconstruction-based detection"],"falsifier":"On a held-out dataset of diffuse chest disease (for example, widespread ground-glass opacities rather than a focal pneumonia), compute the total score with and without the $S_{\\mathrm{concen}}$ and $S_{\\mathrm{contra}}$ terms while keeping the ZhangLab-tuned weights fixed; if the two terms do not raise image-level AUC, or if their class separation collapses when the lesion is spread out, the scoring premise fails. A second check: re-select $\\beta$ and $\\gamma$ on the CheXpert training set and compare with the fixed ZhangLab values; if fixed values are clearly worse, the reported state of the art rests on test-set tuning.","tokens_in":14648,"feed_emoji":"🩻","tokens_out":10294,"duration_ms":99581,"temperature":0.7,"pith_summary":"The paper sets out to show that a Mamba-based autoencoder can beat CNN- and transformer-based methods at unsupervised medical anomaly detection by making anatomical spatial structure explicit rather than treating images as flat sequences. Because radiographs follow fixed imaging protocols, normal anatomy is highly repeatable, yet individual patients shift and tilt organs; the paper argues that a good detector should exploit both the consistency and the tolerated variation. SP-Mamba does this with a window-sliding prototype memory that matches each location to nearby patches, and a Circular-Hilbert scanning scheme that serializes the image center with a locality-preserving curve while de-emphasizing the less informative borders. On top of reconstruction error and prototype distance, it adds anomaly-map statistics: a concentration score and a contrast score. The reported results claim new state-of-the-art numbers on two chest X-ray datasets and the best mean results on a three-modality medical benchmark, with model size and compute roughly half those of transformer-based rivals.","feed_headline":"Mamba-based detector sets new medical anomaly-detection record","feed_subtitle":"SP-Mamba hits 92% AUC on chest X-rays using half the compute of transformer rivals.","key_machinery":"The load-bearing mechanism is the Circular-Hilbert scanning Mamba block inside a pyramidal auto-encoder. For each feature map, the inner $\\frac{h'}{2}\\times\\frac{w'}{2}$ patches are serialized along an $n$-order Hilbert curve, which keeps spatially adjacent patches adjacent in the one-dimensional sequence, while the outer ring is scanned circularly; scanning clockwise or counterclockwise, from corners, and inward or outward yields eight directions in the experiments and up to sixteen in principle. This concentrates the state-space model's long-range modeling on the lesion-relevant image center and reduces interference from structurally random borders. The second mechanism is the Medical-Prototype module, which learns $K$ prototypes and matches each prototype patch through a sliding $p\\times p$ window, taking the minimum cosine distance, so that normal anatomical variation is absorbed rather than flagged. The third is the scoring rule $S_{\\mathrm{total}}=S_{\\mathrm{org}}+\\alpha S_{\\mathrm{p-dist}}+\\beta S_{\\mathrm{concen}}+\\gamma S_{\\mathrm{contra}}$, where $\\beta$ is negative and $\\gamma$ positive in the selected configuration, encoding the observed fact that abnormal anomaly maps are more concentrated and higher-contrast than normal ones. These mechanisms carry the argument because each one turns an anatomical or statistical regularity into an explicit inductive bias.","core_discovery":"The central claim is that the spatial regularity of medical images can be converted into a measurable anomaly-detection advantage. Trained only on normal images, the model is supposed to reconstruct normal features accurately and fail on anomalies; the paper's contribution is to steer that failure toward lesions. The window-sliding prototype learning computes, for each patch of a learned prototype, the minimum cosine distance against all patches in a $p\\times p$ window around the corresponding position, so shifted anatomy is matched instead of scored as anomalous. The Circular-Hilbert scanning method applies a Hilbert curve to the central half of each feature map and circular scans to the surrounding region, producing eight scan directions used by the Mamba decoder. At test time the anomaly score combines reconstruction distance, prototype distance, the concentration score $S_{\\mathrm{concen}}$, and the difference-of-Gaussians contrast score $S_{\\mathrm{contra}}$. On the paper's benchmarks, the full system is reported to reach 92.0% AUC on the ZhangLab chest X-ray dataset, 80.4% AUC on CheXpert, and the best mean image-level and pixel-level results on Uni-Medical.","pith_inferences":["A testable extension the paper leaves implicit: the concentration and contrast scores are post-hoc statistics computed from whatever anomaly map the model produces, so they could be attached to other reconstruction-based detectors and evaluated for the same gain.","The window-sliding prototype distance is close to a soft registration step; one could push this further by reading out the displacement that achieves the minimum distance as an explicit anomaly signal, which might catch lesions that manifest as gross anatomical displacement rather than intensity change.","The Circular-Hilbert scan's division of labor (locality-preserving center, circular periphery) suggests a natural 3D extension for volumetric CT or MRI: Hilbert-scan in-plane and sweep slices circularly, preserving linear complexity."],"forward_implications":["If the reported numbers hold, Mamba-based detectors become a practical choice for chest X-ray screening: higher AUC and accuracy than transformer-based baselines with roughly half the parameters and FLOPs.","The same three-component recipe (prototype matching, locality-preserving scan, and anomaly-map statistics) transfers across modalities: the paper reports gains on brain MRI, liver CT, and retinal OCT under the Uni-Medical protocol.","The anomaly scoring module is additive: the ablation shows that adding the contrast score and then the concentration score improves AUC and F1 on both chest X-ray datasets, implying the two statistics capture complementary signal.","Because the model runs at 25.8M parameters and 8.3 GFLOPs, deployment on modest hardware or higher-resolution inputs becomes plausible without the quadratic attention cost of transformers."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space backbone whose linear complexity the framework inherits.","marker":"[8]"},{"why":"Provides the 2D cross-scan idea that the Circular-Hilbert scan extends and contrasts with.","marker":"[18]"},{"why":"Supplies the Mamba-based baseline, the Uni-Medical benchmark construction, and the efficiency comparison protocol.","marker":"[11]"},{"why":"Defines the structural-consistency approach and the strongest chest X-ray baseline SP-Mamba must beat.","marker":"[31]"},{"why":"Introduces space-aware memory and inpainting for medical radiography and the evaluation protocol on the chest X-ray datasets.","marker":"[32]"},{"why":"Provides the ZhangLab Chest X-ray dataset used for training, testing, and hyper-parameter selection.","marker":"[15]"},{"why":"Provides the CheXpert dataset used as the second chest X-ray benchmark.","marker":"[14]"}],"fun_headline_variants":["SP-Mamba: Mamba model spots anomalies with 92% AUC on chest X-rays","Mamba framework detects medical anomalies at linear cost","Spatial-aware Mamba sets new benchmark in unsupervised anomaly detection","Circular-Hilbert Mamba learns anatomy for accurate anomaly detection","Unsupervised detector uses Mamba to hit 92% AUC on ZhangLab"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the concentration and contrast statistics of the anomaly map separate normal from abnormal images in every modality the model is applied to, and that the scoring weights $\\beta=-0.025$ and $\\gamma=400$, chosen by searching on the ZhangLab test set, keep working on CheXpert and Uni-Medical without retuning; the paper supports the statistical separation with two visual examples and reports no validation-set-based selection of the weights.","fun_headline_variants_meta":{"raw":{"variants":["SP-Mamba: Mamba model spots anomalies with 92% AUC on chest X-rays","Mamba framework detects medical anomalies at linear cost","Spatial-aware Mamba sets new benchmark in unsupervised anomaly detection","Circular-Hilbert Mamba learns anatomy for accurate anomaly detection","Unsupervised detector uses Mamba to hit 92% AUC on ZhangLab"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000585,"raw_usage":{"total_tokens":2762,"prompt_tokens":968,"completion_tokens":1794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1699}},"tokens_in":584,"tokens_out":1794,"duration_ms":11152,"temperature":1.0,"reasoning_tokens":1699,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:00:59.843245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out dataset of diffuse chest disease (for example, widespread ground-glass opacities rather than a focal pneumonia), compute the total score with and without the $S_{\\mathrm{concen}}$ and $S_{\\mathrm{contra}}$ terms while keeping the ZhangLab-tuned weights fixed; if the two terms do not raise image-level AUC, or if their class separation collapses when the lesion is spread out, the scoring premise fails. A second check: re-select $\\beta$ and $\\gamma$ on the CheXpert training set and compare with the fixed ZhangLab values; if fixed values are clearly worse, the reported state of the art rests on test-set tuning.","supporting_citations":[{"cited_title":"SQUID: Deep Feature In-Painting for Unsupervised Anomaly Detection","cited_arxiv_id":"2111.13495","evidence_quote":"Introduces space-aware memory and inpainting for medical radiography and the evaluation protocol on the chest X-ray datasets."},{"cited_title":"Kermany, Michael Goldbaum, Wenjia Cai, Carolina C","cited_arxiv_id":null,"evidence_quote":"Provides the ZhangLab Chest X-ray dataset used for training, testing, and hyper-parameter selection."}],"review_version":2}