{"id":"2f8fa0f0-1fa1-4400-99cb-b59072656fc4","arxiv_id":"2607.29541","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frequency-domain residual called the K-Space Signature, obtained by subtracting a global mean Log-PSD of real scans, detects synthetic 3D MRIs with high accuracy and zero-shot scanner generalization.","lead":"The authors detect fake medical MRI scans by converting volumes into frequency spectra, subtracting an average \"real brain\" spectrum, and training a neural network on the leftover pattern. The method reports above 0.99 accuracy on synthetic scans and up to 0.93 accuracy on scanners never seen during training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global Log-PSD mean is a hardware-weighted mixture, not an anatomical prior; KSS residual may retain anatomy and subtract part of the synthetic signal.","rationale":"The reader's weakest assumption is precisely the load-bearing point: the global mean subtracted in Eq. (4) is treated as a pure anatomical prior, but it is computed from a finite set of authentic volumes from six scanner models. In Log-PSD space, an individual volume approximately decomposes into anatomy plus scanner hardware plus noise; averaging over the training set yields the mean anatomy plus a hardware mixture, not anatomy alone. Consequently, the KSS residual contains both individual anatomical deviations and a scanner-dependent offset. The paper's empirical results may still hold if the downstream classifier learns to separate real from synthetic despite this contamination, but the claimed 'isolation' mechanism is not validated. The supplementary ablation (Table 2) is framed around scanner classification, so it does not refute this concern for the binary real-vs-fake task. A concrete computational experiment—recomputing the baseline under leave-one-scanner-out and comparing to no subtraction—would directly quantify how much of the detection signal relies on the baseline's specific hardware composition. This does not force a change of the overall verdict: the paper remains conditionally acceptable, but the condition should include demonstrating that the global prior's scanner composition is not load-bearing for detection performance.","tokens_in":32397,"tokens_out":12288,"duration_ms":191837,"concrete_test":"Retrain the KSS MLP-Mixer ArcFace detector under three conditions: (a) proposed \\tilde P_global computed from all six training scanners; (b) leave-one-scanner-out baselines (six variants, each computed without one scanner); (c) no subtraction, i.e., use raw Log-PSD \\tilde P_i directly. Compare open-set ROC-AUC and the L2 norm of the synthetic residual vectors across conditions. If condition (c) performs comparably to (a), the subtraction is not doing the claimed disentanglement; if condition (b) shifts open-set ROC-AUC by more than 0.05, the 'anatomical prior' is scanner-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (4) defines \\tilde P_global as the empirical mean Log-PSD over authentic training volumes from six scanner models. This is not an anatomical prior: it is a hardware-weighted average. For a volume from scanner s, the residual is \\tilde P_i - \\tilde P_global ≈ (A_i - A_mean) + (H_s - H_mean), so individual anatomical deviations and scanner-specific hardware traces both remain. More importantly, Fig. 3 shows synthetic volumes are strongly Siemens-correlated; if the training cohort's Siemens component is substantial, subtracting \\tilde P_global partially cancels the very generative signature the method aims to detect. The paper's ablation in Supplementary Table 2 evaluates scanner classification, not deepfake detection, so it does not directly show that the baseline's scanner composition is immaterial for the detection claim. The central assertion that KSS 'isolates hardware and generative traces while suppressing anatomy' is therefore unproven; at present KSS is a centered Log-PSD, not a disentangled signature. This matters for both the method's interpretability and its zero-shot generalization, since an unseen scanner's hardware trace will be superimposed on the residual rather than cleanly removed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the K-Space Signature (KSS), a frequency-domain representation for detecting fully synthetic 3D MRI volumes. The method computes the Log-PSD of each volume, subtracts the empirical mean Log-PSD over authentic training volumes (Eq. 4), and classifies the residual KSS tensor with a 3D MLP-Mixer equipped with an ArcFace metric-learning head. Experiments on ADNI/PPMI real scans and MAISI/Med-DDPM/SuperSynth synthetic volumes report near-perfect closed-set detection and moderate-to-high open-set accuracy on unseen scanners, alongside scanner-attribution and visual-explainability analyses. The paper claims that KSS isolates hardware and generative traces while suppressing anatomy, and that the MLP-Mixer-ArcFace combination yields exceptional detection and zero-shot generalization.","tokens_in":32669,"tokens_out":9857,"duration_ms":144069,"significance":"If the empirical claims hold, KSS is a simple, cheap spectral residual with potentially useful generalization benefits across scanners. The manuscript has several strengths: strict patient-level splitting, extensive supplementary ablations (23 baseline formulations, interpolation robustness, RAW-vs-REG spectral integrity, zero-shot gender classification), and a commitment to release code. The frequency-domain residual idea is plausible and worth pursuing. However, the central interpretability claim—that the global mean is an 'anatomical prior'—is not established, and the experimental comparison is incomplete: there are no comparisons to existing medical deepfake detectors, no error bars or repeated-seed statistics, and several accuracy/AUC inconsistencies. The contribution is potentially valuable but not yet rigorously supported.","major_comments":[{"comment":"The central claim that KSS 'isolates hardware and generative traces while suppressing anatomy' is not established. P~global is the empirical mean Log-PSD over authentic training volumes from six scanner models, so it is a hardware-weighted average, not purely anatomical. For a volume from scanner s, the residual is KSS_i ≈ (A_i − A_mean) + (H_s − H_mean); both anatomical deviations and scanner hardware remain. Moreover, Fig. 3 shows synthetic volumes are strongly correlated with Siemens spectra; if Siemens is a substantial component of Rtrain, subtracting P~global partially cancels the generative signature the method aims to detect. The ablation in Supplementary Table 2 is a 6-class scanner-classification task, not a deepfake-detection task, so it does not show that the global prior's scanner composition is immaterial for detection. Please either demonstrate (e.g., by recomputing P~globa","section":"§2, Eq. (4)–(5)"},{"comment":"The headline 'exceptional detection performance' is not benchmarked against any existing medical deepfake detector. The manuscript cites Li et al. (2025) and Grabovski et al. (2025) as related methods but never compares to them; the baselines are generic 3D CNNs and MLP-Mixer variants only. Also, no ablation isolates the effect of the log transform from the subtraction: 'K-Space' magnitude is compared to 'KSS', but the latter differs by both log-scaling and mean subtraction. Add comparisons to at least one medical-deepfake-specific method and a Log-PSD-without-subtraction baseline to support the claim that the KSS representation itself is responsible for the reported gains.","section":"§3, 'Medical Deepfake Detection'; Table 3"},{"comment":"No error bars, repeated seeds, or significance tests are reported. Given the small open-set (120 volumes) and near-saturated closed-set numbers, single-run results are not convincing. Report mean ± standard deviation over at least 5 random seeds (ideally with different patient splits) and, for key zero-shot comparisons, a paired significance test.","section":"§3, 'Experimental Setup'; Tables 1–3"},{"comment":"The reported Accuracy and AUC values are not coherent as presented. In Table 3, the KSS MLP-Mixer-ArcFace model trained on MAISI has Open Dataset Accuracy 0.973 but ROC-AUC 0.769. In Table 1, the Med/PHILIPS row reports Target Fakes Accuracy 55% despite PR-AUC and ROC-AUC 100%, and Open Set Accuracy 49% with PR-AUC 81% and ROC-AUC 68%. Such patterns imply the Open Set test sets are imbalanced or the decision threshold is chosen without calibration, neither of which is documented. The abstract's 'up to 0.93 Accuracy' does not match Table 1, where the maximum Open Set Accuracy is 99% and the maximum PR-AUC is 93%; the 0.93 value appears to be PR-AUC. Please clarify class balance, threshold selection, and correct the abstract.","section":"Abstract; Table 1; Table 3"},{"comment":"The protocol does not state whether P~global is recomputed from the single training manufacturer or from the full multi-scanner authentic training set. If the global prior of Eq. (4) is used, then the KSS residuals seen during training contain contributions from the held-out scanner manufacturers, undermining the zero-shot generalization claim. The supplementary Zero-Shot Domain Adaptation section explicitly recomputes P~global on source domains; the main leave-one-manufacturer-out experiment must do the same (or justify why it need not).","section":"§3, 'Robustness and Scanner-Invariant Generalization'; Table 1"}],"minor_comments":[{"comment":"The phrase 'mathematically orthogonal inter-vendor domains' is inaccurate for negative correlations (e.g., -0.75 to -0.77); orthogonality corresponds to zero correlation. Suggest 'oppositely oriented' or 'near-orthogonal' with an explicit reference to the zero baseline.","section":"Fig. 3 and accompanying text"},{"comment":"The ArcFace loss is written for a K-class problem, but the cross-dataset experiments use BCEWithLogitsLoss for binary classification. Clarify how the ArcFace margin is implemented for K=2 and why BCE is used for the main results.","section":"§2, Eq. (8); Supplementary B"},{"comment":"The cosine similarity is computed on averaged P~ tensors. Specify the number of volumes averaged per scanner and whether the same computation on the KSS tensors (after mean subtraction) would change the qualitative conclusions.","section":"§3, 'Pairwise Similarity Evaluation', Eq. (9)"},{"comment":"The x-axis is categorical with arbitrary ordering; the overlaid linear trend lines are misleading. Consider per-generator box plots or a clearer visualization that does not imply a monotone relationship.","section":"Fig. 4"},{"comment":"The 'Open Dataset' column should specify its composition—how many real scans from unseen scanners and how many synthetic samples from each generator—and the class ratio. This is needed to interpret Accuracy/AUC inconsistencies.","section":"Supplementary Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is interesting, but the current submission overstates the interpretability of the KSS prior and lacks baseline and statistical rigor. The issues raised are fixable with additional experiments and revised claims; I see no evidence of misconduct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper gives a workable method for detecting fully synthetic 3D MRI by subtracting a global mean Log-PSD from each volume and feeding the residual to a 3D MLP-Mixer with ArcFace. The idea is not deep but it is sensible, and the experiments are more thorough than most in this space. The headline numbers – over 0.99 in closed-set – are believable in that setup. The open-set numbers are more modest and vary a lot depending on training scanner, which the paper acknowledges.\n\nWhat's genuinely new: applying residual Log-PSD analysis to volumetric MRI deepfake detection, plus the specific architecture pairing. The cross-model similarity analysis showing that synthetic data clusters with Siemens hardware is a nice observation and suggests the method is picking up real generative artifacts.\n\nWhere it wobbles: the central interpretive claim – that subtracting an empirical mean over authentic training volumes yields a purely anatomical prior that 'isolates hardware and generative traces' – is not supported. That mean is a hardware-weighted average from six scanner models. The residual still contains anatomy and scanner-specific traces. The ablation in Table 2 is a scanner classification task, not a detection test, so it doesn't validate the prior's purity. The stress-test note is correct. This matters because the method's generalization story depends on the residual being hardware-invariant, and the open-set results (AUC around 0.77 in some rows) suggest the invariant story is only partially true.\n\nOther soft spots: no comparisons against the medical deepfake detectors they cite (e.g., Li et al. 2025, Grabovski et al. 2025) – only generic backbones. No error bars or repeated seeds. Some open-set accuracy/AUC inconsistencies are hand-waved with a threshold argument; that explanation could be tested with calibration plots. The paper also overstates 'zero-shot generalization' for a model that still needs a precomputed baseline from authentic data.\n\nNet: the core method is plausible and the experiments are reproducible in spirit, but the disentanglement narrative needs to be toned down, and the evaluation needs a baseline comparison and uncertainty quantification. I would send this to peer review with a request for major revision. It is a serious paper for the medical forensics community, not a solved problem.","headline":"Plausible frequency-domain detector for 3D MRI deepfakes, but the 'anatomical prior' framing is overreaching and the evaluation needs baseline comparisons and error bars.","tokens_in":33165,"tokens_out":2927,"would_cite":false,"duration_ms":41213,"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":"This paper argues that subtracting the average Log-PSD of authentic MRI volumes from a volume's spectrum isolates the scanner hardware trace and the generative fingerprint, and that a token-mixing 3D network reading this residual can detect","keywords":["K-Space Signature","Medical deepfake detection","Frequency-domain forensics","Log-PSD residual","MRI scanner attribution","MLP-Mixer","ArcFace","Generative-model artifacts"],"falsifier":"Train the KSS pipeline with the global baseline computed from only two of the three manufacturers, then test on the third. If detection accuracy on that held-out vendor's authentic volumes (or on synthetic volumes matched to that vendor's spectrum) drops far below the reported ~0.93, the 'anatomy-only' prior is falsified and the residual is carrying scanner content rather than purely generative traces.","tokens_in":32289,"feed_emoji":"🧠","tokens_out":5785,"duration_ms":77089,"temperature":0.7,"pith_summary":"This paper tries to show that a simple spectral subtraction can separate the machine that took an MRI scan from the brain being scanned, and that this separation makes it possible to spot fully synthetic medical images. The proposed method, the K-Space Signature (KSS), replaces each volume by its logarithmic power spectrum and subtracts the average logarithmic spectrum of authentic training volumes. What remains is claimed to be a residual containing hardware noise and generative artifacts but almost no anatomy. A token-mixing 3D network with an angular-margin classification head then reads these residuals and, according to the authors, exceeds 0.99 accuracy and ROC-AUC on multi-generator synthetic data, and keeps up to 0.93 accuracy on scanners never seen in training. If true, the approach would give hospitals a way to verify that an MRI came from a real machine rather than from a generative model.","feed_headline":"Frequency-domain fingerprint spots synthetic MRI at 99% accuracy","feed_subtitle":"The residual spectrum isolates machine and generator fingerprints, catching fakes from unseen scanners too.","key_machinery":"The load-bearing mechanism is the KSS residual: for a volume V, compute P~ = ln(1+|F(V)|^2) (the logarithmic power spectral density), then subtract the empirical mean P~_global computed over authentic training volumes. The subtraction is designed to cancel the shared low-frequency anatomical content, leaving a residual that concentrates hardware acquisition noise and generative artifacts. A second mechanism is the 3D MLP-Mixer, which mixes tokens across the whole volumetric spectrum (rather than local convolutional patches) and is paired with an ArcFace spherical metric head (an additive angular margin loss) that forces real and synthetic signatures into separated angular clusters. The paper","core_discovery":"The central discovery claim is that the residual KSS_i = P~_i − P~_global, where P~_global is the empirical mean Log-PSD over authentic training volumes, constitutes a representation that isolates non-anatomical traces: scanner hardware fingerprints and generative-model artifacts. The authors support this with pairwise cosine-similarity analysis showing strong intra-manufacturer clustering (GE models at 0.97, Siemens at 0.96) and marked inter-manufacturer orthogonality (negative correlations), and showing that synthetic volumes exhibit a strong positive correlation with Siemens hardware, interpreted as a learned bias from generative training distributions. Paired with a 3D MLP-Mixer plus Arc","pith_inferences":["A natural extension, not pursued in the paper, is to use the KSS residual to localize where in the spectrum the generative trace lives; the reverse-mapping visualizations hint at band-selective signatures that could support partial-forgery detection rather than whole-volume binary classification.","If the anatomy-only assumption on P~_global is relaxed, an even stronger test would be to compute the baseline from a different patient population or a non-brain modality; the framework's portability to CT or PET will hinge on whether an analogous 'global anatomical prior' exists for those modalities.","The threshold-collapse cases (e.g., Med vs Philips) suggest practical deployments should use calibrated thresholds or anomaly scores rather than a fixed 0.5 cutoff, since the ranking of real vs synthetic is preserved even when raw accuracy drops.","The reported correlation between synthetic data and Siemens hardware implies a potential privacy leak: a generative model trained on a specific hospital's Siemens-heavy data may embed that vendor bias, and KSS analysis could be used to infer the composition of training data of an unknown generative model."],"forward_implications":["If the KSS residual indeed suppresses anatomy, forensic models trained on it cannot cheat by memorizing ventricle shapes or cortical folding; any correct classification is attributable to acquisition or generation traces.","The reported performance (Accuracy and ROC-AUC above 0.99) indicates that state-of-the-art 3D generative models, including diffusion-based ones, leave detectable spectral traces when the global baseline is subtracted.","The strong positive correlation between synthetic volumes and one vendor's hardware implies that generative models inherit and replicate the spectral properties of their training corpora; this could be used to attribute a synthetic volume to the scanner family that dominated the generator's training data.","Leave-one-scanner-out experiments suggest that a detector trained on one manufacturer's scans transfers to other manufacturers, up to the reported ~0.93 accuracy, provided the decision threshold is chosen appropriately."],"fun_headline_variants":["K-Space signature exposes synthetic MRI at 99%+ accuracy","Spectral fingerprints unmask fake MRI even from new scanners","Frequency-domain residual spectrum catches medical deepfakes","Zero-shot fake MRI detection via K-Space analysis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's central assumption is that the average Log-PSD of authentic training volumes is a purely anatomical prior; if scanner hardware traces do not cancel out in that average, the subtracted baseline contaminates the residual and the claimed hardware-invariant generalization weakens.","fun_headline_variants_meta":{"raw":{"variants":["K-Space signature exposes synthetic MRI at 99%+ accuracy","Spectral fingerprints unmask fake MRI even from new scanners","Frequency-domain residual spectrum catches medical deepfakes","Zero-shot fake MRI detection via K-Space analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1411,"prompt_tokens":757,"completion_tokens":654,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":588}},"tokens_in":501,"tokens_out":654,"duration_ms":9242,"temperature":1.0,"reasoning_tokens":588,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:14:49.030451+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the KSS pipeline with the global baseline computed from only two of the three manufacturers, then test on the third. If detection accuracy on that held-out vendor's authentic volumes (or on synthetic volumes matched to that vendor's spectrum) drops far below the reported ~0.93, the 'anatomy-only' prior is falsified and the residual is carrying scanner content rather than purely generative traces.","supporting_citations":[],"review_version":2}