{"id":"945f0f2d-595a-4d88-ae35-54b4a590ab78","arxiv_id":"2501.12488","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A comparison of 18 generic pre-trained CycleGAN models for bidirectional brain MR-CT translation finds that the iphone2dslr flower checkpoint yields the best metrics and near-real perceptual ratings, with performance attributed to visual similarity between flowers and brain tissue.","lead":"Researchers tested 18 pre-trained image-translation models, originally trained on non-medical images, for converting brain MR scans to CT scans and back, and found that the model trained on flower photos gave the best quality scores. A perceptual study with two radiologists rated the synthetic images almost as realistic as real scans, suggesting that generic models may transfer to medical imaging.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fine-tuning protocol in §III.B is architecturally incoherent: FC+Softmax head on a fully convolutional CycleGAN cannot output 256×256×3 images, and the claimed 'top scores in all four evaluation scales' contradicts Table III.","rationale":"I read the paper in good faith as a small empirical comparison of 18 generic CycleGAN checkpoints for bidirectional MR-CT synthesis. For the central claim to hold, the fine-tuning procedure must actually transfer features, the model must be able to generate images, and the reported metrics must support the ranking. The weakest link is §III.B: the described insertion of two fully connected layers with Softmax into a fully convolutional generator, with all original weights frozen, is architecturally not a CycleGAN image-to-image translation setup. The adversarial and cycle-consistency losses in Eqs. (1)-(3) require image-sized outputs, so the stated objective and the stated architecture are incompatible. This goes beyond the reader's 'no architectural justification' point: the method as written has no obvious way to produce the reported images at all. The absence of a from-scratch baseline and the absence of an ablation of the added layers mean the paper never establishes that pretrained initialization, rather than random initialization or selective hyperparameter tuning, drives the result. I also verified that the paper's own tables undercut the 'top scores in all four evaluation scales' statement: in Table III, iphone2dslr flower has UQI=0.00 and VIF=0.01 for MR-CT, so the quantitative support is internally contradictory. These issues do not change the reader's REJECT verdict; they strengthen it. The concrete forward-pass test would settle whether my architectural reading is correct: if the reconstructed model does output 256×256×3 images, then the concern about incoherence would not land, though the missing baseline and metric contradictions would remain.","tokens_in":19841,"tokens_out":5414,"duration_ms":57247,"concrete_test":"Reconstruct the exact model graph described in §III.B for the iphone2dslr flower checkpoint and run a forward pass on a single 256×256 MR image. If the output tensor is not 256×256×3 with plausible image-range values, the described FC-Softmax adaptation cannot implement the reported image translation, and the central transfer-learning mechanism fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that fine-tuning a generic CycleGAN checkpoint transfers useful features for bidirectional MR-CT synthesis. The procedure in §III.B is the load-bearing step, and as written it is internally inconsistent. CycleGAN's generator is fully convolutional (c7s1-64, d128, d256, residual blocks, u128, u64, c7s1-3) and outputs a 256×256×3 image. The authors say they insert two randomly initialized fully connected layers with Softmax and freeze all pre-existing layers. Flattening the generator output gives 196,608 values, and an FC(256)-FC(256)-Softmax head cannot produce a 256×256×3 image. If the FC layers are instead placed elsewhere, the stated use of the pretrained generator and the CycleGAN losses in Eqs. (1)-(3) is not specified. The adversarial and cycle-consistency losses are defined over images, so this modified generator cannot be trained by the objective as described. No tensor shapes, model summary, or code are provided, and no from-scratch baseline is included, so the reported ranking has no verified mechanism. Moreover, the paper's own Table III contradicts the 'top scores in all four evaluation scales' claim: iphone2dslr flower has UQI=0.00 and VIF=0.01 in the MR-CT direction, with other rows showing higher UQI/VIF values. Without a coherent architecture or a baseline that isolates initialization, the headline result is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that transfer learning from 18 generic non-medical CycleGAN checkpoints can perform bidirectional MR-CT brain image translation, and reports that the iphone2dslr flower checkpoint is the best, with radiologists rating its outputs nearly as realistic as ground truth. The authors describe fine-tuning by freezing pre-existing layers and adding two fully connected layers with Softmax, report PSNR, SSIM, UQI, and VIF results for all 18 checkpoints, and include a radiologist perceptual study and latent-space visualization. The central claim is that a generic flower-photo model outperforms all other non-medical checkpoints and approaches clinical realism.","tokens_in":20047,"tokens_out":1622,"duration_ms":17790,"significance":"If the central claim held, the paper would provide a striking data point about which properties of pre-trained image-translation models transfer to medical modalities and would support a practical recipe for low-data MR-CT synthesis. The study is also useful as a broad benchmark of 18 public CycleGAN checkpoints on a fixed paired brain dataset, and the radiologist study is a commendable attempt to go beyond pixel metrics. However, the significance is undermined by the paper's own tables, which contradict the headline claim, and by an insufficiently specified fine-tuning procedure that prevents the reader from verifying that any transfer actually occurred.","major_comments":[{"comment":"The fine-tuning protocol is not architecturally coherent. CycleGAN's generator is fully convolutional (c7s1-64, d128, d256, residual blocks, u128, u64, c7s1-3) and outputs a 256×256×3 image. Placing two randomly initialized fully connected layers with Softmax after this generator, while freezing all original layers, cannot produce a 256×256×3 output from the image-sized generator output unless the FC layers are preceded by a flattening and followed by a reshaping that is not specified. As written, Eqs. (1)-(3) are defined over images and cannot train such a modified generator. The paper gives no tensor shapes, no model summary, and no code, so the described mechanism cannot be reproduced or verified.","section":"§III.B, Algorithm 1"},{"comment":"The stated conclusion that iphone2dslr flower achieves 'top scores in all four evaluation scales' is contradicted by Table III. In the MR-CT direction, iphone2dslr flower has UQI = 0.00 and VIF = 0.01, while winter2summer yosemite has UQI = 0.04 and zebra2horse has VIF = 0.05; summer2winter yosemite also has higher SSIM. In the CT-MR direction, several models have higher VIF (e.g., winter2summer yosemite at 0.14) and higher UQI (e.g., winter2summer yosemite at 0.04). The paper itself states in the text that summer2winter yosemite has the highest SSIM and UQI and zebra2horse has the highest VIF in the CT-MR table. The headline claim is therefore not supported by the reported data.","section":"Table III"},{"comment":"No from-scratch CycleGAN baseline is included. The paper's central claim is that pre-trained generic checkpoints transfer useful features for MR-CT synthesis, but without training the same architecture from random initialization on the same data, the reported ranking of 18 checkpoints could reflect initialization artifacts, optimization path differences, or even the added FC layers rather than transfer of learned representations. This missing baseline is load-bearing for the conclusion that transfer learning, rather than the fine-tuning procedure itself, is responsible for the results.","section":"§IV.A, §IV.C"},{"comment":"The explanation that iphone2dslr flower succeeds because flower images resemble brain structures is post hoc. The paper does not report any pre-registered hypothesis or any quantitative measure of visual similarity between the flower dataset and brain images; it presents the resemblance as an observation made after identifying the winner. As stated, this is a rationalization rather than a tested explanation, and it cannot be distinguished from overfitting to the specific test split or from random checkpoint variation.","section":"§V, Discussion"},{"comment":"The perceptual study is reported only for the best model, selected using the same test data and metrics reported in Tables III-IV. Since the selection was made on the evaluation set, the radiologist realism scores are not an independent validation of the model's quality; they confirm only that the already-selected checkpoint produces images that radiologists find realistic. The paper does not report whether the same radiologists rated outputs of other checkpoints or a from-scratch baseline, so the perceptual evidence cannot support the claim that transfer learning from this checkpoint is uniquely beneficial.","section":"§IV.D, Table V"}],"minor_comments":[{"comment":"The abstract and discussion state that results provide 'compelling evidence' of exceptional performance, but the internal contradictions in Tables III-IV should be acknowledged and reconciled before making such strong claims.","section":"Abstract and §V"},{"comment":"The table headers are misaligned: 'A VG STD' appears to be a corrupted rendering of 'AVG', and the columns under each metric are not clearly labeled in the manuscript text. This makes the tables harder to read than necessary.","section":"Table III and IV"},{"comment":"The PSNR equation is written as PSNR = 20 log10(R²/MSE), which is dimensionally incorrect; the standard definition is 10 log10(R²/MSE). This appears to be a typographical error, but it should be corrected.","section":"§IV.B, Eq. (4)"},{"comment":"There are numerous typos and incomplete sentences, including 'Translation from MR–CT brain tumors presented' at the start of Section IV and 'pre-trained model' inconsistently capitalized in Section III.C. The paper would benefit from careful proofreading.","section":"Throughout"},{"comment":"Several references appear to be cited inconsistently: [15] is invoked for multiple different works, and reference [43] (VIF) is not discussed in the text with the same depth as the other metrics. The reference list should be checked for accuracy and completeness.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper has a useful empirical scope—18 public checkpoints on a fixed paired brain dataset—but the central claim is not supported by the authors' own tables, and the fine-tuning description in §III.B is architecturally incomplete to the point of being non-reproducible. The missing from-scratch baseline and the post hoc flower-brain similarity explanation further weaken the transfer-learning conclusion. These are load-bearing issues that cannot be fixed by local edits; a revision would need to re-specify the architecture, add baselines, and re-derive the ranking claims. I recommend rejection, though a substantially revised version with a corrected architecture description, a from-scratch baseline, and an honest reporting of per-metric rankings could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague:\n\nThe useful core of this paper is the systematic scan: 18 public CycleGAN checkpoints, pre-trained on flowers, horses, maps, cityscapes, and so on, fine-tuned on the same paired MR-CT brain dataset and evaluated with four FR-IQA metrics plus two radiologists. That is a legitimate, limited-scope question—whether the source domain of a generic checkpoint matters for a small medical target set. The paper does run the experiment, reports per-model tables, and names the checkpoints and dataset specifically enough that the scan could be reproduced. The perceptual study is a nice addition.\n\nThe problem is that the central claim is contradicted by the paper's own numbers. The iphone2dslr flower model has PSNR 30.98 and SSIM 0.65 in the MR-CT direction, and 34.36 and 0.83 in CT-MR, but its UQI is 0.00 in both directions and VIF is 0.01/0.10. Several other rows have higher UQI and VIF. So the statement that it achieves \"top scores in all four evaluation scales\" is false. The abstract's \"exceptional performance\" and \"premier pre-trained model\" language are overclaims.\n\nThe more serious issue is the fine-tuning protocol in Section III.B. The generator is fully convolutional and outputs 256×256×3; adding two randomly initialized fully connected layers with Softmax and freezing all existing layers cannot produce that output, and the CycleGAN losses are defined on images. As written, the architecture is incoherent. There is no from-scratch CycleGAN baseline and no comparison to the cited uagGAN medical baseline, so the ranking of checkpoints has no demonstrated mechanism—it could be initialization artifacts, or something about the added layers, or just noise. The flower-brain resemblance explanation in the discussion is post hoc: the hypothesis appears only after observing the winner. And the conclusion states that medical-specific models performed better even though all 18 are non-medical, which is an internal contradiction the authors missed.\n\nWho is this for? Medical imaging practitioners curious about checkpoint initialization for small synthesis datasets might get value from the tables, but only after the claims are corrected and the architecture is properly specified. In current form it is not citable as evidence for transfer learning. If this crossed my desk as an editor, I would send it to peer review with a strong request for revision: clarify the fine-tuning, add a from-scratch baseline and a medical baseline, fix the metric claims, and report the perceptual study per model or justify selection. There is a real empirical core here, but it needs substantial work before the conclusions can be trusted.","headline":"A useful checkpoint-scan experiment buried under an unsupported headline claim and an incoherent fine-tuning description.","tokens_in":20659,"tokens_out":2579,"would_cite":false,"duration_ms":28983,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generic CycleGAN pretrained on flower photos, after light fine-tuning, outperforms 17 other non-medical checkpoints for bidirectional MR-CT brain synthesis and yields images radiologists rate nearly as real as ground truth.","keywords":["image translation","transfer learning","pre-trained models","CycleGAN","brain tumor","magnetic resonance imaging","computed tomography","medical image synthesis"],"falsifier":"Train the same CycleGAN architecture with the same paired MR-CT data, splits, and hyperparameters from randomly initialized weights, and compare PSNR, SSIM, UQI, and VIF on the same test slices; if the from-scratch model matches or exceeds the flower checkpoint's PSNR of about 30.98 for MR-to-CT and 34.36 for CT-to-MR, the claimed transfer advantage is not real. A complementary check is to rerun the fine-tuning without the two added fully connected layers and with the pretrained layers unfrozen, to see whether the reported ranking survives changes to the adaptation layer.","tokens_in":19528,"feed_emoji":"🧠","tokens_out":7900,"duration_ms":76131,"temperature":0.7,"pith_summary":"The paper tries to show that generic, non-medical pretrained CycleGAN checkpoints can be transferred to bidirectional brain MR-CT image synthesis, and that the choice of source images matters. Using 18 pretrained models covering artistic styles, animals, landscapes, photography, satellite maps, and urban scenes, the authors fine-tune each on 367 paired MR-CT brain slices and evaluate PSNR, SSIM, UQI, and VIF. The iphone2dslr flower checkpoint, trained on iPhone and DSLR flower photos, ranks first on all four metrics in both translation directions, with PSNR 30.98 for MR-to-CT and 34.36 for CT-to-MR. In a blinded perceptual study, two radiologists classified its synthetic images as real about 98% of the time, close to the ground-truth rate, and the model showed clean separation of MR and CT in latent space. If right, the result suggests that a cheap, publicly available non-medical checkpoint whose textures resemble brain tissue can reduce the data and compute needed for clinically plausible synthesis.","feed_headline":"Flower-photo AI model tops brain MR-CT translation","feed_subtitle":"Radiologists rated its synthetic scans as real in about 98% of cases, nearly matching ground truth.","key_machinery":"The machinery is a CycleGAN backbone, an unsupervised generative network with two generators and two discriminators trained with adversarial loss plus cycle-consistency loss so a translated image can be mapped back to its source, combined with a transfer-learning protocol. Each pretrained checkpoint is loaded with its original weights; two fully connected layers, FC1 and FC2 with 256 neurons each and softmax activation, are added and randomly initialized; all original layers are frozen; and the model is trained for 200 epochs on the paired MR-CT dataset with lambda values of 9 to 11 depending on the source category and a learning rate that linearly decays after epoch 100. The frozen pretrained layers are meant to retain generic visual features while the added layers adapt the model to the MR and CT domains, and the latent-space analysis is offered as evidence that the best model separates the two modalities cleanly.","core_discovery":"The central claim is that transfer learning from generic pretrained models is a viable and effective route for bidirectional MR-CT brain image translation, and that the iphone2dslr flower checkpoint is the best of the 18 generic models for the task. After fine-tuning all checkpoints under the same protocol, adding two 256-neuron fully connected layers with softmax, freezing the pretrained weights, and training for 200 epochs with category-dependent lambda, the flower model achieves the top score in all four evaluation scales in both directions, including PSNR 30.98 plus or minus 0.119 and SSIM 0.65 plus or minus 0.031 for MR-to-CT, and PSNR 34.36 plus or minus 0.072 and SSIM 0.83 plus or minus 0.022 for CT-to-MR. Radiologists rated its outputs 3.79 out of 4 for MR-to-CT and 3.69 out of 4 for CT-to-MR, with 97.91% and 97.7% of the images judged real, against 98.58% and 98.37% for ground truth. The paper attributes the advantage to the high quality and structural similarity of flower images to brain tissue, including petal patterns and convoluted surfaces that resemble cerebral structures, and argues that careful selection of representative training images is decisive for medical synthesis.","pith_inferences":["A decisive control the paper leaves implicit is training the same CycleGAN from scratch on the same data; without it, the ranking is compatible with the added fully connected layers and training schedule doing most of the work.","The flower model's advantage may come from low-level texture statistics shared between petals and brain tissue; repeating the 18-model ranking on lung or cardiac CT-MR data would test whether the advantage transfers to other anatomies.","The clean MR-CT separation in the latent space suggests a cheap screening procedure: measure latent separability on a validation split to pick a pretrained checkpoint before committing to full fine-tuning.","The roughly 98% radiologist realism rate comes from two readers at one institution on one dataset, so it supports further study rather than immediate clinical deployment."],"forward_implications":["If the central claim holds, a checkpoint trained on non-medical images that resemble brain textures can replace a large medical training set as the starting point for MR-CT synthesis.","The iphone2dslr flower checkpoint becomes the default choice among the 18 tested generic models for this bidirectional brain-synthesis task.","Synthetic CT generated from MR could reduce unnecessary radiation exposure by supplying missing modalities without new scans.","The latent-space separation observed for the winning model offers a diagnostic signal for whether a pretrained model has actually adapted to the MR and CT domains."],"supporting_citations":[{"why":"Provides the CycleGAN architecture, cycle-consistency loss, and the pretrained checkpoint family that the transfer-learning study is built on.","marker":"[59]"},{"why":"Supplies the 367 paired MR-CT brain images from 18 patients used for training and testing, and the prior bidirectional synthesis approach this work extends.","marker":"[2]"},{"why":"Supplies the large image database used to train the animal and landscape checkpoints, anchoring what those generic models learned.","marker":"[12]"},{"why":"Supplies the street-scene training set for the cityscapes photo-to-label and label-to-photo checkpoints.","marker":"[10]"},{"why":"Cited as the source of the satellite and map image collection used to train the sat2map and map2sat checkpoints.","marker":"[19]"},{"why":"Supplies the facade image database used to train the facades photo-to-label and label-to-photo checkpoints.","marker":"[49]"}],"fun_headline_variants":["Flower-AI beats all in brain MR-CT translation","Flower pretraining wins bidirectional brain synthesis","Radiologists rate flower-AI scans real ~98%","Generic flower model best for brain scan translation","Flower-photo AI outdoes 18 models in brain MR-CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that adding two randomly initialized fully connected layers with softmax to a frozen generic CycleGAN checkpoint is genuine transfer learning; the paper gives no ablation of those added layers and no comparison to training the same CycleGAN from scratch, so the reported ranking could reflect initialization artifacts rather than transferred features.","fun_headline_variants_meta":{"raw":{"variants":["Flower-AI beats all in brain MR-CT translation","Flower pretraining wins bidirectional brain synthesis","Radiologists rate flower-AI scans real ~98%","Generic flower model best for brain scan translation","Flower-photo AI outdoes 18 models in brain MR-CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00133,"raw_usage":{"total_tokens":5494,"prompt_tokens":1114,"completion_tokens":4380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":730,"completion_tokens_details":{"reasoning_tokens":4301}},"tokens_in":730,"tokens_out":4380,"duration_ms":34061,"temperature":1.0,"reasoning_tokens":4301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:08:31.630872+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same CycleGAN architecture with the same paired MR-CT data, splits, and hyperparameters from randomly initialized weights, and compare PSNR, SSIM, UQI, and VIF on the same test slices; if the from-scratch model matches or exceeds the flower checkpoint's PSNR of about 30.98 for MR-to-CT and 34.36 for CT-to-MR, the claimed transfer advantage is not real. A complementary check is to rerun the fine-tuning without the two added fully connected layers and with the pretrained layers unfrozen, to see whether the reported ranking survives changes to the adaptation layer.","supporting_citations":[{"cited_title":"Paired-unpaired Unsupervised Attention Guided GAN with transfer learning for bidi- rectional brain MR-CT synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the 367 paired MR-CT brain images from 18 patients used for training and testing, and the prior bidirectional synthesis approach this work extends."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"Supplies the large image database used to train the animal and landscape checkpoints, anchoring what those generic models learned."},{"cited_title":"The cityscapes dataset for se- mantic urban scene understanding","cited_arxiv_id":null,"evidence_quote":"Supplies the street-scene training set for the cityscapes photo-to-label and label-to-photo checkpoints."},{"cited_title":"Spatial pattern tem- plates for recognition of objects with regular struc- ture","cited_arxiv_id":null,"evidence_quote":"Supplies the facade image database used to train the facades photo-to-label and label-to-photo checkpoints."}],"review_version":1}