{"id":"3efbda84-c2ae-4022-a6a0-8f6cdc0472ce","arxiv_id":"2505.07364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Synthetic FDG PET images generated from T1 MRI by a 3D patch-based Cycle-GAN with MSE loss can train an unsupervised anomaly detector that localizes epileptogenic lesions in MRI-negative epilepsy patients with 74% sensitivity, outperforming a detector trained on real PET (42%).","lead":"The paper used a type of generative AI to create fake brain PET scans from MRI images of healthy volunteers, then trained an unsupervised anomaly-detection model on those fake scans. When tested on real epilepsy patients' scans, the model found subtle lesions more often than a model trained on real PET scans, suggesting synthetic imaging data can substitute for scarce and costly nuclear medicine exams.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central comparison is confounded: UAD1 trains on true PET from DBC1 while UAD2/3 train on synthetic PET generated from a different T1-only dataset, so the 74% vs 42% gain cannot be attributed to synthetic PET.","rationale":"The reader identified the risk that the synthetic-trained improvement is an artifact (denser latent distribution, lenient top-10 evaluation) but did not identify the more fundamental experimental-design confound: the 'real' and 'synthetic' UAD models are trained on different normative subjects (DBC1 vs DBC2), so the central comparative claim cannot be evaluated from the current data. This is a structural issue, not just a statistical-power problem. The paper's feasibility result, that a UAD trained on synthetic PET can reach 74% sensitivity on a small homogeneous cohort, remains a useful proof-of-concept. However, the title-level claim that synthetic PET 'can serve to improve performance' relative to true PET, and the conclusion that it can 'efficiently replace' true PET in normative databases, rest entirely on this confounded comparison. The proposed within-DBC1 ablation would isolate the effect of the PET modality and settle whether the claimed improvement is real. Since the required evidence is absent, the central claim is unverdictable in its current form, rather than conditionally acceptable pending external validation alone.","tokens_in":23375,"tokens_out":7234,"duration_ms":68781,"concrete_test":"Controlled within-DBC1 ablation: generate synthetic PET for the 35 DBC1 T1s using the best 3D-patch GAN (retrained in the same 4-fold scheme so the target subjects are unseen during GAN training), then train UAD-fake on real T1 + synthetic PET and compare it with UAD-real trained on real T1 + true PET, using identical UAD hyperparameters, cluster threshold (at most 10 clusters), and ranking criterion; evaluate on DBe_p with bootstrap confidence intervals for sensitivity and mean rank. If UAD-fake does not match or exceed UAD-real, or if the 74%-vs-42% margin disappears, the headline improvement is explained by the DBC1-vs-DBC2 training-set difference rather than by synthetic PET.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The paper's main evidence for the replacement claim is the §5.3 / Table 3 comparison between UAD1 (42%, 8/19) and UAD3 (74%, 14/19). But Table 1 and §4.2 show these models are not trained on the same normative subjects: UAD1 is trained on 35 true T1+PET exams from DBC1, while UAD2/UAD3 are trained on 35 true T1 exams from DBC2 paired with synthetic PET generated from those T1s. DBC2 is a different set of healthy controls and has no true PET, so 'synthetic vs real' is entangled with 'DBC2 vs DBC1'. The observed improvement could therefore arise from the different T1 training population or the slightly different training-set size, not from properties of synthetic PET. The paper calls this a fair comparison in §4.2, but fairness would require true and synthetic PET for the same normative subjects, or a within-DBC1 control. In addition, no confidence interval or significance test is reported for the 14/19 vs 8/19 difference; a McNemar-style exact test on the discordant pairs in Table 3 gives p≈0.07, so even under the current design the superiority claim is not statistically established. The §6 Discussion appropriately cautions that the denser-latent explanation is speculative and that the cohort is small, but it does not acknowledge this training-set confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes to synthesize FDG PET images from T1-weighted MRI using GAN-based architectures, with the stated goal of using such synthetic PET data as normative training data for an unsupervised anomaly detection (UAD) model for epileptogenic zone screening. The authors compare several GAN configurations (2.5D vs 3D-patch, simple GAN vs Cycle-GAN, with and without an additional MSE loss), evaluate visual quality using PSNR/SSIM/LPIPS and task-oriented out-of-distribution metrics, and then train a siamese-autoencoder/OC-SVM UAD model on either 35 real paired T1/PET exams (UAD1) or on 35 T1 exams from a different control dataset paired with synthetic PET generated from those T1 scans (UAD2/UAD3). On 17 epilepsy patients with 19 subtle lesions, the model trained on 3D-patch synthetic PET achieves 74% sensitivity (14/19) versus 42% (8/19) for the model trained on real PET, with mean ranks of 2.1 and 3.9 respectively. The authors conclude that synthetic PET could efficiently replace true PET in multi-modality normative databases for training UAD models.","tokens_in":23631,"tokens_out":5147,"duration_ms":45500,"significance":"If the central claim were established, this would be a valuable contribution: it would relax a significant practical bottleneck for UAD training, namely the difficulty of acquiring normative PET databases, and it would extend prior work by evaluating synthetic data on a concrete downstream detection task rather than only with visual metrics. The paper has clear strengths: the code and normative image dataset are released, the visual-quality comparison includes formal statistical tests (Wilcoxon signed-rank), and the clinical evaluation targets a well-defined, difficult problem (MRI-negative epilepsy with subtle hypometabolism). The task-oriented OOD metrics are a reasonable attempt to bridge visual quality and utility, although their self-referential nature limits their probative value. However, the central claim—that synthetic PET can replace true PET—is supported by a comparison that is confounded by the training population, and the effect size is not statistically established on the small lesion cohort. Because these issues bear directly on the main conclusion, the current evidence supports a more cautious claim than the one stated in the Conclusion.","major_comments":[{"comment":"The comparison between UAD1 (trained on real PET) and UAD2/UAD3 (trained on synthetic PET) is confounded: UAD1 is trained on the 35 true T1/PET exams of DBC1, while UAD2 and UAD3 are trained on 35 T1 exams of DBC2 paired with synthetic PET generated from those T1 scans. The observed 74% vs 42% difference in Table 3 therefore conflates the effect of synthetic PET with differences in the normative T1 training population (DBC2 vs DBC1) and any differences in training-set composition. The paper calls this a 'fair comparison' in Section 4.2, but fairness would require true and synthetic PET for the same normative subjects, e.g., by generating synthetic PET from DBC1 T1 scans (the GANs are trained on DBC1) or by obtaining true PET for DBC2. This confound is load-bearing for the replacement claim and needs to be resolved with a within-dataset control.","section":"Section 4.2 and Table 1"},{"comment":"The key result, 14/19 vs 8/19 detections, is reported without confidence intervals or a significance test. A McNemar exact test on the discordant pairs from Table 3 yields p≈0.07, so even under the current (confounded) design the superiority of the synthetic-trained model is not statistically established. The authors should report an appropriate paired test or a bootstrap interval, and should frame the 74% vs 42% difference as an exploratory finding given the small lesion count.","section":"Table 3 and Section 5.3"},{"comment":"The explanation for the improvement—that synthetic PET induces a 'denser latent distribution'—is acknowledged by the authors as speculative, but the evaluation protocol itself may contribute to the gap: clusters are thresholded per score map to yield at most 10 clusters (Appendix C), and a lesion counts as detected if it lies in any of the top-10 clusters. If the synthetic-trained models produce more or smaller clusters on this cohort, the sensitivity difference could be inflated. The paper should report the number of clusters per patient per model and evaluate sensitivity to the cluster threshold; without this, the 74% vs 42% difference cannot be confidently attributed to a genuinely more useful normative representation.","section":"Section 6 and Appendix C"}],"minor_comments":[{"comment":"The Gaussian smoothing FWHM (1.5 mm) and the histogram-matching reference image are selected on the validation set; a brief sensitivity analysis or justification would clarify how dependent the reported visual metrics are on these post-processing choices.","section":"Section 3.1.3"},{"comment":"In the comparison sentence, 'the real paired T1 MRI and FDG PET of D_B3Dpatch_C1' should read 'of DBC1'; the subscript appears to be a typo.","section":"Section 5.3"},{"comment":"The scatter plots are dense and the three groups overlap; consider plotting marginal distributions or increasing transparency, and state the number of points per group in the caption.","section":"Figure 5"},{"comment":"The cluster-ranking criterion uses equal weights for cluster size and average score, and the OC-SVM RBF kernel is used without a reported hyperparameter search; a brief description of how these choices were made, or a sensitivity check, would improve reproducibility.","section":"Appendix C"},{"comment":"The claim that GANs 'outperform transformer or diffusion models' rests on the cited review [6] and not on experiments in this paper; this limitation is acknowledged in the Discussion but could be stated more carefully in the Conclusion.","section":"Introduction and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The confound in Section 4.2 is the main obstacle to the central claim. It is fixable within the scope of the manuscript by adding a within-DBC1 comparison, for example by generating synthetic PET from DBC1 T1 scans and retraining UAD on those, or by otherwise matching the normative training population across conditions. I would also ask the authors to add a statistical test for the primary detection comparison and to report cluster counts per patient. The paper's strengths—released code and data, a clearly defined clinical task, and statistically tested visual metrics—make it worth revising rather than rejecting, but the current evidence does not support the strong replacement claim as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read. First, the paper's central comparison is not as clean as it looks: UAD1 trains on 35 true T1+PET exams from DBC1, while UAD2/UAD3 train on 35 T1 exams from DBC2 paired with synthetic PET generated from those T1s. DBC2 is a different set of healthy controls with no true PET. So 'synthetic vs real' is entangled with 'DBC2 vs DBC1', and the 74% vs 42% gap cannot be attributed to synthetic PET alone. Second, even under the given design, no confidence interval or significance test accompanies that gap; a McNemar-style exact test on the discordant pairs gives p≈0.07, so the superiority claim is weak statistically.\n\nWhat the paper does well: the GAN engineering is careful. Adding an MSE loss to Cycle-GAN with paired T1/PET clearly helps visual quality, with significant improvements on SSIM/PSNR/LPIPS; the 3D-patch variant is the best. The task-oriented OOD metrics (MSE in image space, Mahalanobis distance in the siamese-autoencoder latent space) are a sensible way to check distribution shift before committing to downstream training. Code and the normative dataset are released. The 74% lesion-level sensitivity on MRI-negative epilepsy patients is worth reporting as a proof-of-concept, even if the comparison with the real-PET model is what needs scrutiny.\n\nSoft spots in proportion: the confound is load-bearing for the replacement claim, though not for the weaker claim that synthetic PET can train a UAD at all. The evaluation is lenient — a lesion counts as detected if it appears anywhere among the top ten clusters, with the score-map threshold set to yield at most ten clusters. The OOD metrics are computed with the same UAD architecture that later performs detection, so as an independent quality check they are partly circular; as a sanity check they are fine. The abstract says GANs outperform transformer/diffusion models, but no such comparison is run; the Discussion admits this. The authors' own 'denser latent distribution' explanation is speculative and they flag it as such, but they do not flag the training-set confound, and that is the main omission.\n\nWho it is for: people working on cross-modal synthesis and unsupervised anomaly detection in neuroimaging, and readers who want a worked example of why dataset provenance matters in evaluation. It deserves a serious referee. My advice: send it to review with a request for a within-DBC1 or same-subjects control for the synthetic-vs-real comparison, plus a significance test or confidence interval on the sensitivity difference. If the authors cannot run that, the conclusion needs softening from 'can efficiently replace' to 'may be usable as a training surrogate, pending validation.'","headline":"Useful proof-of-concept, but the headline comparison (synthetic vs real PET for UAD training) is confounded by switching normative datasets, so the 74% vs 42% result does not support the replacement claim as stated.","tokens_in":24285,"tokens_out":3250,"would_cite":false,"duration_ms":29760,"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":"Synthetic FDG PET images generated from T1 MRI by GANs can replace real PET scans in the normative database used to train an unsupervised detector of epilepsy lesions.","keywords":["Medical image synthesis","Cycle-GAN","PET MRI","Unsupervised anomaly detection","out-of-distribution (OOD)","epileptogenic zone detection"],"falsifier":"Take the same UAD pipeline and the same 17-patient test set, but train the detector on a larger real-PET normative database that is scanner-matched to the patients; if its sensitivity reaches or exceeds 74%, the synthetic advantage is not caused by better normative fidelity. Independently, recompute detection counts using only the top-1 or top-3 clusters instead of the top-10 rule described in Appendix C; if the synthetic-trained sensitivity drops below the real-trained 42%, the reported gain is an artifact of the cluster-counting protocol.","tokens_in":23126,"feed_emoji":"🧠","tokens_out":6854,"duration_ms":62245,"temperature":0.7,"pith_summary":"The paper's claim is that synthetic FDG PET images, generated from T1 MRI by GAN-based architectures, can replace true FDG PET images in a multimodal normative database used to train an unsupervised anomaly detection model. It reports that the best detector trained on synthetic PET from a 3D-patch Cycle-GAN finds 14 of 19 subtle epilepsy lesions (74% sensitivity, mean rank 2.1) among 17 surgical patients, versus 8 of 19 (42%, mean rank 3.9) for the same detector trained on real PET. The practical stake is direct: real FDG PET control databases are costly and hard to collect, while T1 MRI is routine, so synthesizing the missing modality could make PET-based anomaly detection feasible where it currently is not. The paper also introduces task-oriented metrics, reconstruction error and Mahalanobis distance in latent space, to test whether synthetic data behave as in-distribution for the detector.","feed_headline":"Synthetic PET outperforms real PET for training epilepsy detector","feed_subtitle":"GAN-generated PET lifts lesion detection from 42% to 74% sensitivity in 17 epilepsy surgery patients.","key_machinery":"The machinery is a Cycle-GAN for T1-to-PET translation, augmented with a mean-squared-error loss between paired real and synthetic PET, and configured either as 2.5D slice triplets or 3D patches; the best version is the 3D-patch model, followed by Gaussian smoothing and histogram matching of the reconstructed volumes. The downstream UAD model is a siamese convolutional autoencoder whose latent vectors feed a per-voxel one-class SVM, producing ranked anomaly clusters. The paper's new quality metrics, global reconstruction MSE and Mahalanobis distance in the autoencoder latent space, check whether fake PET behaves as in-distribution for the detector before the detection experiment is run.","core_discovery":"The central claim, stated directly in the conclusion, is that synthetic FDG PET data generated from T1 MRI can efficiently replace true FDG PET images in normative databases for training unsupervised anomaly detection models. The paper supports this with three linked results: a Cycle-GAN with an added mean-squared-error loss, trained on paired T1 and PET data, produces synthetic PET with SSIM around 0.9 and PSNR around 23.8; the synthetic PET falls in-distribution relative to real PET under the proposed OOD metrics; and the UAD model trained on real T1 plus synthetic 3D-patch PET detects more subtle epileptogenic lesions than the same model trained on real PET. The authors caution that their explanation for the improvement, a denser synthetic latent distribution due to lower inter-individual variability, should be interpreted with caution.","pith_inferences":["Because the authors themselves flag the denser-latent-distribution explanation as tentative, a direct test would be to measure the covariance of latent vectors for synthetic versus real control PET: a substantially narrower synthetic distribution would predict that the detection gain shrinks on more heterogeneous patient data.","The top-ten cluster counting rule makes a hit easy, since a lesion counts as detected if it appears anywhere among the ten most suspicious clusters; recomputing the comparison with only the top-ranked cluster would show whether the synthetic-trained advantage reflects sharper localization or simply more candidate clusters.","If the effect generalizes, the same replacement strategy could apply to other radiotracers or modalities where normative databases are scarce, such as amyloid or dopamine-transporter PET, provided a paired MRI-based synthesis model can be trained.","A natural next experiment is mixing real and synthetic PET in the normative database; if detection improves monotonically with the amount of real data, the benefit is likely due to data scale rather than something unique to synthetic images."],"forward_implications":["A center with only T1 MRI of healthy subjects could build a PET-trained unsupervised anomaly detector without collecting a real FDG PET control database.","The 3D-patch Cycle-GAN with MSE loss is the recommended synthesis configuration, giving the best visual metrics and the best downstream detection performance.","The proposed OOD metrics offer a fast pre-screening check for whether a new patient's multimodal images are in-distribution for a trained UAD model.","For localizing subtle epileptogenic zones, training on synthetic rather than real PET is reported to raise sensitivity from 42% to 74% in this 17-patient cohort."],"supporting_citations":[{"why":"Supplies the UAD architecture, a siamese autoencoder with per-voxel one-class SVM, that the paper adapts from T1/FLAIR MRI to T1 plus PET.","marker":"[1]"},{"why":"Provides the comparative review supporting the claim that GAN-based architectures outperform transformer and diffusion models for MRI-to-PET synthesis.","marker":"[6]"},{"why":"Provides the Mahalanobis-distance OOD method that the paper adapts into a task-oriented quality metric for synthetic PET.","marker":"[8]"},{"why":"Defines the least-squares GAN objective used to train the generators and discriminators.","marker":"[13]"},{"why":"Supplies the paired normal T1-MRI and FDG-PET control database used to train and cross-validate the synthesis models.","marker":"[15]"},{"why":"Provides the earlier GAN-based pseudo-normal PET synthesis approach for hypometabolism localization that this work extends to normative training data.","marker":"[33]"},{"why":"Provides the Cycle-GAN and PatchGAN baseline code and architecture on which the synthesis models are built.","marker":"[36]"},{"why":"Is the authors' preliminary study, extended here with task-oriented quality metrics, histogram matching, and the epilepsy patient evaluation.","marker":"[37]"}],"fun_headline_variants":["Fake PET tops real PET for epilepsy AI training","Synthetic PET generated from MRI improves epilepsy detection","GAN-synthesized PET images train better epilepsy detector than real","MRI-to-PET synthetic data surpass real data for epilepsy AI","Fake PET from MRI lifts epilepsy detection sensitivity to 74%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic-trained detector's higher sensitivity reflects a genuinely useful denser normative latent distribution, not an artifact of reduced inter-individual variability that happens to produce more or smaller candidate clusters on this particular 17-patient cohort.","fun_headline_variants_meta":{"raw":{"variants":["Fake PET tops real PET for epilepsy AI training","Synthetic PET generated from MRI improves epilepsy detection","GAN-synthesized PET images train better epilepsy detector than real","MRI-to-PET synthetic data surpass real data for epilepsy AI","Fake PET from MRI lifts epilepsy detection sensitivity to 74%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000467,"raw_usage":{"total_tokens":2419,"prompt_tokens":1127,"completion_tokens":1292,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":743,"completion_tokens_details":{"reasoning_tokens":1210}},"tokens_in":743,"tokens_out":1292,"duration_ms":9997,"temperature":1.0,"reasoning_tokens":1210,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:18:48.519770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same UAD pipeline and the same 17-patient test set, but train the detector on a larger real-PET normative database that is scanner-matched to the patients; if its sensitivity reaches or exceeds 74%, the synthetic advantage is not caused by better normative fidelity. Independently, recompute detection counts using only the top-1 or top-3 clusters instead of the top-10 rule described in Appendix C; if the synthetic-trained sensitivity drops below the real-trained 42%, the reported gain is an artifact of the cluster-counting protocol.","supporting_citations":[{"cited_title":"Regularized siamese neural network for unsupervised outlier detection on brain multiparametricmagneticresonanceimaging:applicationtoepilepsy lesion screening","cited_arxiv_id":null,"evidence_quote":"Supplies the UAD architecture, a siamese autoencoder with per-voxel one-class SVM, that the paper adapts from T1/FLAIR MRI to T1 plus PET."},{"cited_title":"Least squares generative adversarial networks, in: Proceedings of the IEEE conference on computer vision (ICCV), pp","cited_arxiv_id":null,"evidence_quote":"Defines the least-squares GAN objective used to train the generators and discriminators."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier GAN-based pseudo-normal PET synthesis approach for hypometabolism localization that this work extends to normative training data."},{"cited_title":"Unpaired image- to-image translation using cycle-consistent adversarial networks, in: Computer Vision (ICCV), 2017 IEEE International Conference on","cited_arxiv_id":null,"evidence_quote":"Provides the Cycle-GAN and PatchGAN baseline code and architecture on which the synthesis models are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the authors' preliminary study, extended here with task-oriented quality metrics, histogram matching, and the epilepsy patient evaluation."}],"review_version":1}