{"id":"4e1526f4-06fe-4953-b311-eebf3404b1ae","arxiv_id":"2603.01878","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A wavelet-frequency-spatial CNN reports 96.01% mAcc on a new ten-generator CT deepfake benchmark, but the training generator remains in the test set.","lead":"This paper presents a new benchmark of 75,990 chest CT images — half generated by ten AI models — and a detector that distinguishes real from synthetic scans with 96.01% mean accuracy and only a 0.99% accuracy drop under blur, cropping, JPEG, and noise. It matters because hospitals and data curators need automated ways to keep AI-generated medical scans out of clinical and training pipelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Domain confound (all real images from CT-RATE) and HA-GAN overlap in train/test undermine the claimed generalization; the 96.01% mAcc may reflect CT-RATE-specific cues rather than generic generative artifacts.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the real-versus-fake contrast is confounded with source domain, and HA-GAN appears in both training and testing. My stress-test pass confirms this from the manuscript text. This is not a disagreement with the paper's internal consistency; the ablations (Tables 3–4) are coherent and the robustness numbers are plausible under the stated protocol. But the abstract and conclusion generalize beyond the protocol: they claim superiority for detecting AI-generated CT images in real-world clinical settings, while the evaluation never presents a single real image from outside CT-RATE or a fake generator that is truly unseen given the training set. The most load-bearing weakness is therefore not the architecture but the evaluation design. A concrete test—external real data plus patient-level splits and removal of the training generator from test—would settle whether the 96.01% figure transfers. Since this concern is already reflected in the reader's CONDITIONAL verdict, no verdict change is needed; the recommendation remains conditional on the authors providing such a restructured evaluation or an explicit narrowing of the claim.","tokens_in":8660,"tokens_out":2755,"duration_ms":29670,"concrete_test":"Run the evaluation with (1) a patient-level split of CT-RATE, (2) an external real-CT set such as LIDC-IDRI, and (3) a test set that excludes HA-GAN (or reports per-generator results with HA-GAN omitted). If the external-real accuracy or the non-HA-GAN average drops more than 10 percentage points, the domain confound is material.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ESF-CTFD generalizes across ten unseen CT generative models. But the evaluation protocol does not separate the real-image domain from the fake-image domain. Every real test image is sampled from CT-RATE (Sec. 4), while the fake test images come from ten generators trained on heterogeneous datasets (LIDC, COVID-CT, etc.). A detector can therefore achieve high accuracy by recognizing CT-RATE-specific intensity statistics, windowing, or anatomical layout rather than by detecting generic generative artifacts. This confound is compounded by two concrete protocol choices: (1) HA-GAN, the generator used to create the training fakes, also appears in the test set, so the 'unseen generator' evaluation is partially self-referential (Table 1, HA-GAN = 100.00); (2) no patient-level split is reported, so the 23,000 real training slices and the test real slices may share patients, inflating separability. The robustness results in Table 5 are also measured on CT-RATE real images only, so they demonstrate robustness to perturbations of one real domain, not cross-dataset robustness. Consequently, the headline numbers (96.01% mAcc, 0.99% average drop) are internally consistent for the CTForensics benchmark as constructed, but they do not establish the real-world generalization claim made in the abstract and Sec. 7.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CTForensics, a benchmark of 75,990 2D CT images with a balanced 29,990-image test set spanning ten generative models (four GAN-based, six diffusion-based). It also proposes ESF-CTFD, a CNN combining a wavelet-enhanced stem, multi-scale spatial aggregation, and a frequency-process block. On the authors' benchmark, ESF-CTFD reports 96.01% mean accuracy and 99.96% mean average precision, outperforming existing detectors by a large margin, and shows only a 0.99% average mAcc drop under blur, cropping, JPEG compression, and noise. The paper's central claim is that ESF-CTFD generalizes across unseen CT generative models and is robust to realistic perturbations.","tokens_in":8958,"tokens_out":5835,"duration_ms":54396,"significance":"A multi-generator CT forgery benchmark and a CT-specific detector would be a useful contribution if the evaluation protocol substantiated the generalization claim. The paper offers per-generator results, ablations of the multi-scale and wavelet/central components, and a robustness comparison. However, the current protocol does not yet support the headline claim: the training fake-generator HA-GAN appears among the test generators, all real images are drawn from a single source (CT-RATE) while fakes come from heterogeneous datasets, and no patient-level split is reported. The dataset has potential, and the method's internal design is coherent, but the evidence for unseen-generator and cross-domain generalization needs to be established by a revised evaluation.","major_comments":[{"comment":"The task definition in Sec. 3 states that the model is trained on fake images from one generative method and evaluated on multiple unseen forgery types. Sec. 4 says the training fakes are generated using HA-GAN, yet HA-GAN is also one of the ten test generators in Table 1, where it achieves 100.00% accuracy. This makes the 'unseen generator' evaluation circular for one test column. Remove HA-GAN from the test set, retrain on a different generator, or report mAcc excluding the training generator; otherwise the 96.01% figure is not an unbiased generalization estimate.","section":"Sec. 3 vs. Sec. 4 and Table 1"},{"comment":"All real images in both training and testing come from CT-RATE, while the fake images from the ten generators were produced by models trained on other CT datasets (e.g., LIDC, COVID-CT). The real-vs-fake decision is therefore confounded with source domain: a detector can succeed by recognizing CT-RATE-specific intensity statistics, windowing, or anatomy rather than generic generative artifacts. To support the claimed generalization, evaluate the same detector on real images from a second dataset (and, ideally, pair each generator with the held-out real domain), or explicitly control for source domain.","section":"Sec. 4, Table 1, and Fig. 1"},{"comment":"No patient-level split is reported. The 23,000 real training slices and the real test slices are both from CT-RATE, and if slices from the same patients appear in both sets, the reported accuracy is inflated by patient-level leakage. Please specify the patient/volume-level split and verify that no patient overlaps between training and test; if such a split already exists, state it explicitly.","section":"Sec. 4 (train/test split)"},{"comment":"The frequency-process block is motivated by average frequency spectra computed over the test images shown in Fig. 3. Using test-set statistics to select the architecture or feature transformations means the reported generalization is not fully blind. This should be acknowledged as a limitation, and the method should be validated on a separate, untouched test set or an external dataset.","section":"Sec. 5.4 and Fig. 3"}],"minor_comments":[{"comment":"'It comprises of there components' should be 'three components'; the Sec. 6.2 heading 'Performamce' should be 'Performance'.","section":"Sec. 1; Sec. 6.2"},{"comment":"The text says 'hh, hl, and lh sub-bands ... while lh is handled by WTConv'; lh is listed twice and the low-frequency sub-band should be ll. Correct the notation.","section":"Sec. 5.2"},{"comment":"Define mAcc and mAP precisely. Clarify whether mAcc is the average of per-generator accuracies and whether the real test images are shared across generators.","section":"Sec. 6.1"},{"comment":"State whether perturbations are applied to both real and fake images, and give the exact ranges in the caption. Also 'A verage Drop' contains a typo.","section":"Table 5"},{"comment":"Report the CT preprocessing pipeline (windowing, intensity normalization, PNG conversion) and any filtering/quality checks used in dataset construction; this is needed for reproducibility.","section":"Sec. 4 and Sec. 6.1"},{"comment":"All tables report point estimates without confidence intervals. Consider repeated runs with mean and standard deviation, especially for the per-generator results.","section":"Tables 1, 2, 5"}],"recommendation":"major_revision","confidential_remarks":"The proposed dataset and detector are potentially valuable, but the review should insist on a revised evaluation protocol. The HA-GAN test overlap is a straightforward fix, but the CT-RATE-only real-image domain is a more fundamental confound that requires either a larger multi-domain real set or a restrained claim. The per-generator table is useful; I would be willing to look at a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2603.01878. It's a useful contribution in a niche area — CT forgery detection — and the dataset is genuinely new: 75,990 2D CT slices spanning ten generative models, broader than the existing one- or two-generator benchmarks. The proposed ESF-CTFD detector, while assembled from known components (wavelet enhancement, central difference convolution, fast Fourier convolution, multi-scale fusion), is reasonably engineered and the internal ablations are coherent. If the benchmark were clean, the claimed 96% mAcc and near-perfect mAP would be impressive.\n\nBut the benchmark is not clean, and the problems are load-bearing. First, HA-GAN is used to generate the training negatives and also appears in the test set. The paper explicitly frames the task as generalization to unseen generators; that claim is partially self-referential for HA-GAN. Second, all real images come from CT-RATE, while the fakes come from a variety of sources trained on different datasets. A detector can achieve high accuracy simply by recognizing CT-RATE-specific intensity statistics, windowing, or anatomy, without learning anything generic about generative artifacts. That confound alone undermines the cross-generator generalization claim as stated.\n\nTwo more issues: no patient-level split is reported, so train/test real slices may share patients; and the frequency process block is motivated by Fig. 3, which averages spectra over the test set — test-set leakage in the design process, not necessarily fatal, but it weakens the claim that frequency cues are intrinsic.\n\nTo be fair: the dataset construction is real work, the writing is clear, and the method section is reproducible in principle. These are fixable flaws, but they need to be fixed before the numbers mean what the abstract says. I'd send this to peer review — it's worth referee time — but I would not accept it in its current form. A revision that excludes HA-GAN from the test set, adds a held-out real dataset, and reports patient-level splits would make the contribution solid.\n\nWho is this for? People working on medical image forensics and generalization evaluation. It's also a good reading-group case study in how easily source domain can be confounded with class label.\n\nRecommendation: major revision.","headline":"Useful new CT forgery benchmark, but the headline generalization claim is undercut by a domain confound and a train/test generator overlap.","tokens_in":9496,"tokens_out":2687,"would_cite":false,"duration_ms":24818,"reading_group":"yes","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 claims that a detector built around CT-specific wavelet, spatial, and frequency cues can identify AI-generated CT images with 96% accuracy across ten generative models, outperforming prior methods by more than 10 percentage point","keywords":["AI-generated image detection","CT forensics","medical deepfakes","computed tomography","generative adversarial networks","diffusion models","wavelet transform","frequency analysis"],"falsifier":"Run the same training/evaluation protocol but replace the real test images with CT slices from a different institution or scanner (keeping the same ten generators); if mean accuracy drops sharply while the detector still tags the training generator perfectly, the model was keyed to dataset-specific cues, not universal forgery artifacts. A patient-level split of CTForensics would also show whether overlapping patients inflate the scores.","tokens_in":8507,"feed_emoji":"🩻","tokens_out":8672,"duration_ms":73832,"temperature":0.7,"pith_summary":"The paper sets out to show that AI-generated CT images can be reliably detected when the detector is built around CT-specific imaging physics rather than adapted from natural-photo forensics. To test this, it introduces CTForensics, a benchmark of 75,990 2D CT slices with a balanced test set of 29,990 images from ten generative models spanning GANs and diffusion models. It then presents ESF-CTFD, a convolutional neural network that combines wavelet, spatial, and frequency analysis across three image scales. The paper reports 96.01% mean accuracy and 99.96% mean average precision on this benchmark, with only a 0.99% average drop under blur, cropping, JPEG compression, and noise. If correct, this gives clinical settings a practical way to flag synthetic scans and a shared benchmark to compare future detectors.","feed_headline":"CT detector identifies AI-generated scans with 96% accuracy","feed_subtitle":"Trained on one GAN, it flags nine unseen generators and drops less than 1% under blur, cropping, JPEG, and noise.","key_machinery":"The key mechanism is a three-branch architecture tuned to CT-specific traces: a Wavelet-Enhanced Central Stem decomposes the image into sub-bands and amplifies local intensity differences; Multi-Scale Spatial Aggregation fuses features at 112, 224, and 448 pixel resolutions; and a Frequency Process Block applies Fast Fourier Convolution to expose spectral artifacts. The CTForensics benchmark itself is the enabling asset, providing ten generative models (four GAN-based, six diffusion-based) with balanced real-fake test sets, so generalisation can actually be measured.","core_discovery":"The central claim is that a detector explicitly modeling CT artifacts in three complementary domains—high-frequency wavelet residuals, local pixel correlations, and global spectral statistics—generalizes across unseen generative models. Trained on fake images from a single GAN, ESF-CTFD achieves 96.01% mean accuracy and 99.96% mean average precision on the CTForensics test set, which includes nine additional generators, and outperforms existing natural-image forensics methods by more than 10 percentage points. The paper further claims the design makes the detector resilient to realistic distortions, with only a 0.99% average accuracy drop under Gaussian blur, random cropping, JPEG compressio","pith_inferences":["Because all real images come from a single public CT dataset, the 96% figure may partly encode a dataset fingerprint; testing on real scans from other scanners or hospitals would show whether the detector learns a general 'real' distribution.","The training generator also appears in the test set, so the perfect 100% score on it may lift the mean; a leave-one-generator-out protocol would isolate true cross-generator generalization.","The frequency-artifact analysis is tied to the ten generators studied; new generators with different spectral signatures, or post-processing that alters the spectrum, could require the benchmark to be refreshed.","The same wavelet–spatial–frequency architecture could transfer to other modalities where synthetic images are a growing concern, such as MRI or X-ray, if real data are collected from multiple sources."],"forward_implications":["A model trained on fake images from a single GAN can identify images from nine unseen generators, covering both GAN and diffusion architectures.","Combining wavelet, spatial, and frequency cues yields large gains over any single branch: removing both wavelet and central-correlation components drops accuracy from 96.01% to 88.37%.","Multi-scale processing (112, 224, and 448 pixels) is necessary; dropping any scale reduces both mean accuracy and mean average precision.","The detector loses only 0.99% accuracy on average under Gaussian blur, random cropping, JPEG compression, and injected noise, even when all are applied together.","CTForensics provides a fixed benchmark with ten generative models, enabling direct comparison of future CT forgery detectors."],"fun_headline_variants":["CT forensics dataset and detector catch AI fakes at 96%","AI-generated CT scans exposed by new forensic detector","96% accuracy spotting synthetic CT images across 10 models","Detector trained on one GAN flags nine unseen CT generators","New CT detector generalizes to unseen AI generators with 96% mAcc"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that 'real' CT images are well represented by a single public dataset, so the detector can learn a general real-image signature; if that dataset's quirks—noise, windowing, anatomy—are what the model exploits, the reported accuracy will not carry over to other hospitals or scanners.","fun_headline_variants_meta":{"raw":{"variants":["CT forensics dataset and detector catch AI fakes at 96%","AI-generated CT scans exposed by new forensic detector","96% accuracy spotting synthetic CT images across 10 models","Detector trained on one GAN flags nine unseen CT generators","New CT detector generalizes to unseen AI generators with 96% mAcc"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000461,"raw_usage":{"total_tokens":2174,"prompt_tokens":804,"completion_tokens":1370,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1283}},"tokens_in":548,"tokens_out":1370,"duration_ms":9418,"temperature":1.0,"reasoning_tokens":1283,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:29:40.139863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training/evaluation protocol but replace the real test images with CT slices from a different institution or scanner (keeping the same ten generators); if mean accuracy drops sharply while the detector still tags the training generator perfectly, the model was keyed to dataset-specific cues, not universal forgery artifacts. A patient-level split of CTForensics would also show whether overlapping patients inflate the scores.","supporting_citations":[],"review_version":1}