{"id":"2b07bdd4-0aa6-43ac-9fa5-22b0ef375e9c","arxiv_id":"2411.17420","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A GAN combining pyramid convolution, channel attention, and self-attention generates PET images from sMRI on ADNI with slightly better similarity and classification metrics than compared baselines.","lead":"The authors train a GAN that converts structural MRI brain scans into synthetic PET images, adding pyramid convolution, channel attention, and self-attention to improve image quality. On the ADNI dataset they report small gains in similarity metrics and in downstream Alzheimer's classification accuracy, but the evaluation protocol has serious gaps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No proper held-out evaluation is described; the headline metrics appear to be reported on the training set, so the 'outperforms previous methods' claim is unsupported.","rationale":"The reader's weakest assumption points to the undisclosed train/test split and unreproduced baselines, which directly matches the core concern I identify. The manuscript text goes further: Section 4.3 explicitly says the metrics were achieved 'during training,' which is a stronger indication that no held-out evaluation was performed. The failure to describe a data split is not a minor omission; it is load-bearing because every quantitative claim, including the AD classification result, depends on the model's generalization to unseen subjects. The baseline comparison is also not independently checkable, as the authors state they only 'attempted to reproduce' prior methods and that full reimplementation was not completed. The contradictory loss definitions compound the issue by making even the training objective unclear. My proposed check—subject-level cross-validation with fixed folds and re-run baselines—would settle whether the reported metrics reflect genuine performance. If the numbers hold under this test, the central claim gains support; if not, the paper's main contribution is unverified. The reader's REJECT verdict is therefore appropriate, and no adjustment is needed.","tokens_in":11097,"tokens_out":2305,"duration_ms":22150,"concrete_test":"Obtain the code and the exact subject-level data partition; re-train PCSA-GAN with 5-fold subject-level cross-validation on the 356 paired subjects and report mean ± std MAE/PSNR/SSIM on held-out folds only. Re-run HGAN and GLA-GAN under the same training protocol and fixed test folds; if the held-out metrics differ substantially from 0.0194/29.65/0.9486 or the baselines cannot be reproduced, the outperformance claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PCSA-GAN generates high-quality PET images rests entirely on reported metrics: MAE 0.0194, PSNR 29.65, SSIM 0.9486. Section 4.3 states these were achieved 'during training on the publicly available ADNI dataset.' The paper never specifies a train/test split for the 356 paired subjects, nor any validation procedure, so the reported numbers may reflect in-sample performance rather than generalization. Tables 1 and 2 provide standard deviations for ablation variants, but they vary the architecture or loss, not the test set; no error bar is given for the final model on a fixed held-out partition. The comparison with prior methods is equally fragile: Section 4.7 says HGAN and GLA-GAN were 'attempted to reproduce' from their papers, and Section 4.8 admits full reimplementation was not completed. Without a common, independently specified protocol, the claimed superiority is not verifiable. Additionally, the adversarial loss in Section 3.4.1 is written identically to the voxel-wise L1 reconstruction loss in Section 3.4.2, making even the training objective ambiguous. These combined gaps mean the central empirical claim is not supported by the manuscript as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PCSA-GAN, a conditional GAN with a U-Net style generator that augments pyramid convolution with channel attention (PCCA blocks) and a self-attention layer for generating PET images from sMRI in Alzheimer's disease research. The method is evaluated on ADNI, reporting MAE 0.0194, PSNR 29.65, and SSIM 0.9486 for image generation, and 94.21% accuracy for AD classification when the generated PET is combined with sMRI. The authors claim these results outperform prior sMRI-to-PET synthesis methods.","tokens_in":11323,"tokens_out":5648,"duration_ms":49117,"significance":"If properly validated, the proposed architecture would be a modest but useful contribution to cross-modal medical image synthesis, combining multi-scale local feature extraction with global self-attention. The paper includes ablation studies over the PCA and SA modules and over loss functions, which is a strength, and it reports standard deviations for these ablations. However, the empirical claims are currently not supported because the evaluation protocol lacks a held-out test split, the baseline comparisons are incomplete, and the training objective is ambiguously specified. The paper also does not provide code or data to enable independent verification, limiting its reproducibility.","major_comments":[{"comment":"The paper never specifies a train/test split for the 356 paired subjects. Section 4.3 reports the headline metrics 'during training on the publicly available ADNI dataset,' which suggests the MAE, PSNR, and SSIM values may be computed on the training set. Without a held-out test partition, these numbers do not demonstrate generalization, and the claimed superiority over prior methods cannot be assessed. This is a load-bearing issue for the central empirical claim.","section":"Section 4.2, 4.3"},{"comment":"Equation (7) is labeled 'adversarial loss' but is identical to the voxel-wise L1 reconstruction loss in Eq. (8). The discriminator loss L_D that appears in Eq. (12) is never defined. The training objective is therefore ambiguous, and it is unclear how adversarial training contributes to the reported results. The authors should state the actual adversarial loss and the full training objective.","section":"Section 3.4.1, 3.4.2, Eq. (7)-(12)"},{"comment":"The comparison with HGAN and GLA-GAN is not verifiable. The authors state that they 'attempted to reproduce' these models and admit in Section 4.8 that it is 'not easy to implement these methods completely.' Without a complete, common implementation and a shared evaluation protocol, the reported superiority is unsupported. Additionally, the 328 sMRI-only subjects used for the classification experiment are not shown to be disjoint from the 356 paired subjects used to train the generator; if there is overlap, the classification results are optimistically biased.","section":"Section 4.7, 4.8"},{"comment":"Table 3 shows that the proposed method has a lower AUC (94.91%) than Pan et al. (96.95%) and Gao et al. (95.60%), yet the text in Section 4.8 and the abstract claim the method outperforms previous studies in AD diagnosis. The claim should be restricted to the metrics where the method is actually best (ACC, SPE, SEN, F1), or the AUC discrepancy must be explained. As written, the claim is internally inconsistent with the reported numbers.","section":"Table 3, Section 4.8"}],"minor_comments":[{"comment":"The caption of Table 2 repeats the caption of Table 1 and refers to 'the impact of the self-attention mechanism,' but the table rows list loss functions. The caption should be corrected.","section":"Table 2 caption"},{"comment":"The text contains 'as shown in Table ??' instead of a proper reference to Table 3.","section":"Section 4.8"},{"comment":"There are typesetting errors in equations: Eq. (3) is missing the pooling and activation symbols, and Eqs. (9) and (10) have misplaced or mismatched brackets. These should be cleaned up.","section":"Section 3.4.2, Eq. (3), (9), (10)"},{"comment":"Reference [24] is cited for the N3 bias-correction algorithm but points to a paper about the anterior commissure; reference [27] is cited for the HAMMER registration algorithm but the cited work is on histopathological image registration. The citations should be corrected to the original methodological sources.","section":"References"},{"comment":"The paper uses both 'MS-SSIM' and 'MM-SSIM' (the latter in Table 2). The abbreviation should be unified to 'MS-SSIM'.","section":"Throughout"},{"comment":"The rows of Table 1 are labeled 'Patch Size,' but the text describes applying self-attention at different scales of features. Please clarify what 'Patch Size' means and how it relates to the feature scale.","section":"Table 1, Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript lacks a proper evaluation protocol: no train/test split is stated, and the headline numbers are reported as 'during training.' In addition, the loss definition is ambiguous and the baseline comparisons are acknowledged to be incomplete. These are fixable with additional experiments, but they are substantial gaps. The paper might be suitable for a future version if the authors provide a rigorous held-out evaluation, complete baselines, and clarify the training objective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the PCSA-GAN manuscript. The architecture is a reasonable, incremental combination: pyramid convolution with channel attention for multi-scale local features, self-attention for global context, and a joint loss with MS-SSIM. As a package, I don't think this exact configuration has been published before, and the ablations suggest each block contributes something. The preprocessing pipeline is described in enough detail that a determined reader could roughly replicate it.\n\nBut the evaluation does not support the headline claim. Section 4.3 reports an MAE of 0.0194, PSNR 29.65, and SSIM 0.9486 \"during training\" on ADNI. There is no train/test split for the 356 paired subjects, no separate validation procedure, and no error bar for the final model on a fixed held-out set. The std devs in Tables 1 and 2 come from varying architecture or loss, not from repeated trials on a fixed test set. So the \"outperforms previous methods\" claim is not verifiable. The baseline comparison is also fragile: Section 4.7 says HGAN and GLA-GAN were \"attempted to reproduce\" from their papers, and Section 4.8 admits full reimplementation was not completed.\n\nThe loss function section compounds the problem. Eq. 7 is labeled \"adversarial loss\" but is numerically identical to the voxel-wise L1 in Eq. 8. And Eq. 12 includes an LD term that is never defined. That's not a typo-level issue; it means the training objective is ambiguous. There are also smaller signs of carelessness: Table 2 is titled \"The impact of the self-attention mechanism\" but reports loss-function ablations, and several references don't match the method they are meant to support ([24] is about anterior commissure anatomy, not the N3 algorithm; [27] and [28] are similarly off-target).\n\nI want to give credit where it's due: the idea is plausible, the architecture description is mostly clear, and the authors identify a real clinical problem. But for an empirical paper, the evidence needs to be reproducible. A serious referee would hit the same wall within an hour. I would not send this to review in its current form. The right next step is a revision with a proper held-out protocol, code or data release, and a corrected loss description. If that lands, it could be a solid incremental contribution, but right now the claims outrun the evidence.","headline":"A plausible architecture package, but the headline numbers are effectively unreviewable: no held-out split, an adversarial loss written as L1, and baselines only partially reproduced.","tokens_in":11891,"tokens_out":3308,"would_cite":false,"duration_ms":29785,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A GAN that generates PET images from structural MRI beats previous methods on ADNI quality and AD diagnosis.","keywords":["Alzheimer's disease","cross-modal medical image generation","PET synthesis from MRI","pyramid convolution","channel attention","self-attention","generative adversarial network","ADNI dataset"],"falsifier":"Rerun the three main models (PCSA-GAN, HGAN, and GLA-GAN) on a fixed public train/test split of the 356 paired ADNI subjects, using the authors' preprocessed 64x64x64 gray-matter masks, and check whether PCSA-GAN still gets the best MAE, PSNR, and SSIM on the same held-out test set; if it does not, the outperformance claim collapses.","tokens_in":10882,"feed_emoji":"🧠","tokens_out":7955,"duration_ms":70633,"temperature":0.7,"pith_summary":"The paper aims to show that missing PET images can be synthesized from structural MRI scans well enough for clinical and diagnostic use. The proposed generative model, PCSA-GAN, combines pyramid convolution with channel attention to capture multi-scale local features from sMRI, and injects global relationships with self-attention before reconstructing the PET image. On the ADNI dataset the authors report a mean absolute error of 0.0194, a peak signal-to-noise ratio of 29.65, and structural similarity of 0.9486, all better than the previous methods they compared against. When the generated PET images are combined with the original sMRI in an Alzheimer's disease classifier, they report 94.21% accuracy, above sMRI alone and above prior sMRI-to-PET pipelines. If these results hold, the method would make multimodal diagnosis feasible in settings where PET is unavailable.","feed_headline":"Synthetic PET from MRI hits 94.21% AD classification accuracy","feed_subtitle":"Synthetic PET from structural MRI, good enough to aid Alzheimer's diagnosis.","key_machinery":"The load-bearing mechanism is the PCCA block (pyramid convolution plus channel attention) in the generator's contraction path, followed by a self-attention layer before the deconvolution stages. Pyramid convolution runs several 3D convolutions with different kernel sizes in parallel and concatenates their output, so the network sees both fine details and broad context; channel attention reweights those multi-scale features to suppress redundancy. Self-attention then computes global correlations across spatial positions of the deepest feature map, which the authors argue preserves the overall structure of the brain. The generator is trained with a joint loss combining adversarial loss, voxel-wise L1 loss, and multi-scale SSIM loss, each responsible for a different aspect of fidelity.","core_discovery":"The central claim is that a U-Net-style conditional GAN whose encoder uses pyramid convolution (parallel 3D convolutions with kernels of size 7, 5, and 3) plus channel attention, and whose decoder begins with a self-attention layer, produces PET images that quantitatively and visually match real FDG-PET scans better than earlier CNN-, U-Net-, and CycleGAN-based sMRI-to-PET methods. The authors report MAE 0.0194, PSNR 29.65, and SSIM 0.9486 on the ADNI dataset, and they report that these synthesized PET images, paired with the source sMRI, give 94.21% accuracy in an AD-versus-normal classification task. They also demonstrate via ablation that removing either the pyramid-channel-attention module or the self-attention module degrades the metrics, and that adding L1 and multi-scale SSIM losses on top of adversarial loss improves output quality.","pith_inferences":["Editorial inference: if the reported results replicate on an independent patient-disjoint split, the method could reduce the need for real PET scans in AD research cohorts, since the synthesized PET carries most of the classification information.","Editorial inference: the same pyramid-plus-attention generator could be tried for other missing-modality pairs such as MRI-to-CT or across PET tracers, since the method is not tied to AD-specific image content.","Editorial inference: the strongest public test would be to release train/test subject identifiers and the reimplemented baseline code, allowing the outperformance claim to be checked without relying on the paper's in-house reproduction."],"forward_implications":["Generated PET could substitute for real PET in downstream multimodal classification, raising accuracy from 90.48% with sMRI alone to 94.21% in the paper's experiments.","Clinics without PET capacity could still obtain PET-equivalent metabolic information from the MRI scans they already collect.","The architecture suggests that multi-scale local features plus global self-attention is a workable recipe for cross-modal medical image synthesis.","The joint loss result implies that voxel-level and structural losses add complementary training signal to adversarial loss for medical image generation."],"supporting_citations":[{"why":"Supplies the first 3D-CNN baseline for sMRI-to-PET translation.","marker":"[8]"},{"why":"Introduces the 3D-UNet approach that the generator's U-Net structure extends.","marker":"[9]"},{"why":"Provides the conditional adversarial image-translation framework the generator builds on.","marker":"[12]"},{"why":"Defines conditional GAN, the base model for the proposed generator.","marker":"[17]"},{"why":"Supplies CycleGAN, one of the compared cross-modal translation baselines.","marker":"[18]"},{"why":"Gives the GLA-GAN baseline that combines global and local context for MRI-to-PET.","marker":"[21]"},{"why":"Gives the task-induced pyramid and attention GAN baseline for imputation and classification.","marker":"[22]"},{"why":"Gives the CycleGAN-based sMRI-to-PET baseline used in AD classification comparison.","marker":"[23]"},{"why":"Provides the HAMMER registration algorithm used to align gray-matter masks to the template.","marker":"[27]"},{"why":"Provides the MNI 152 template space used for spatial normalization.","marker":"[28]"}],"fun_headline_variants":["MRI-to-PET synthesis reaches 94.21% AD diagnostic accuracy","Pyramid attention network generates PET from MRI for AD diagnosis","Synthetic PET from MRI aids AD classification at 94.21%","Cross-modal PET generation from MRI with pyramid attention","High-fidelity PET from MRI achieves 94.21% AD classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers rest on the assumption that the test subjects were not used in training and that HGAN and GLA-GAN were run under the same conditions, but the paper admits in Section 4.8 that it did not fully reimplement those methods.","fun_headline_variants_meta":{"raw":{"variants":["MRI-to-PET synthesis reaches 94.21% AD diagnostic accuracy","Pyramid attention network generates PET from MRI for AD diagnosis","Synthetic PET from MRI aids AD classification at 94.21%","Cross-modal PET generation from MRI with pyramid attention","High-fidelity PET from MRI achieves 94.21% AD classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00124,"raw_usage":{"total_tokens":5105,"prompt_tokens":976,"completion_tokens":4129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":4041}},"tokens_in":592,"tokens_out":4129,"duration_ms":27779,"temperature":1.0,"reasoning_tokens":4041,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:07:11.443304+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the three main models (PCSA-GAN, HGAN, and GLA-GAN) on a fixed public train/test split of the 356 paired ADNI subjects, using the authors' preprocessed 64x64x64 gray-matter masks, and check whether PCSA-GAN still gets the best MAE, PSNR, and SSIM on the same held-out test set; if it does not, the outperformance claim collapses.","supporting_citations":[{"cited_title":"In: Medi- cal Image Computing and Computer-Assisted Intervention–MICCAI 2014: 17th International Conference, Boston, MA, USA, September 14-18, 2014, Proceedings, Part III 17, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the first 3D-CNN baseline for sMRI-to-PET translation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the 3D-UNet approach that the generator's U-Net structure extends."},{"cited_title":"IEEE journal of biomedical and health informatics 26(1), 36–43 (2021)","cited_arxiv_id":null,"evidence_quote":"Gives the task-induced pyramid and attention GAN baseline for imputation and classification."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part III 11, pp","cited_arxiv_id":null,"evidence_quote":"Gives the CycleGAN-based sMRI-to-PET baseline used in AD classification comparison."},{"cited_title":"Pattern Recognition Letters 152, 210–217 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides the HAMMER registration algorithm used to align gray-matter masks to the template."},{"cited_title":"Brain Structure and Function 228(5), 1045–1067 (2023) 18","cited_arxiv_id":null,"evidence_quote":"Provides the MNI 152 template space used for spatial normalization."}],"review_version":1}