{"id":"94a69d98-9d0d-468c-87c6-ca2c07dc8913","arxiv_id":"2505.12203","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CTLformer, a hybrid CNN-transformer denoiser, reports marginally higher SSIM and lower RMSE than the CTformer baseline on a single Mayo Clinic LDCT test patient, with no code or statistical validation.","lead":"The authors propose CTLformer, a hybrid neural network that mixes convolutional layers with transformer attention to remove noise from low-dose CT scans. The paper reports slightly better scores than an existing transformer baseline, but only on one test patient and without code or error bars.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-patient metrics and absence of significance tests cannot support 'significantly outperforms'; the reported 1.85M vs 1.45M parameter count also undercuts the efficiency claim.","rationale":"The reader's verdict is REJECT with moderate confidence, and my stress-test supports keeping that verdict unchanged. The reader's weakest assumption was that a single test patient is sufficient to demonstrate general superiority; this is indeed the central evidential gap. I agree with that diagnosis. I would add a second, independent weakness that makes the central claim even more fragile: the paper's own Table 1 shows CTLformer at 1.85M parameters versus CTformer at 1.45M, while the abstract and conclusion claim superiority in model efficiency. Since the paper uses parameter count as its efficiency metric, this is an internal inconsistency, not merely a missing measurement. The attack is therefore not about disagreeing with an outside consensus; it is about the evidence presented being insufficient and partly self-contradictory. A concrete fix is straightforward: execute the described cross-validation, provide per-patient results with uncertainty and significance tests, and supply a consistent efficiency measure. If those results show consistent, statistically meaningful improvements and a genuine efficiency advantage, the central claim could be rehabilitated. Until then, the rejection stands. The paper also lacks architectural equations, code, and ablations, but those are secondary to the load-bearing evidential problem.","tokens_in":183,"tokens_out":3835,"duration_ms":44262,"concrete_test":"Run the full leave-one-patient-out cross-validation described in Section 3.1: for each of the 10 patients, train on the other 9 and evaluate on the held-out patient. Report PSNR, SSIM, and RMSE per fold, with mean and standard deviation across patients, and conduct a paired significance test (e.g., Wilcoxon signed-rank or bootstrap 95% CI) comparing CTLformer against CTformer. Also report parameter count, FLOPs, and wall-clock inference time per image. If the CI for the CTLformer-minus-CTformer difference includes zero on either SSIM or RMSE, or if CTLformer is not faster or lighter on a primary efficiency axis, the claims of significant superiority and improved efficiency are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in the abstract and Section 4, is that CTLformer 'significantly outperforms existing methods in both denoising performance and model efficiency.' The empirical support for this claim is a single test patient (L506, slice 057) reported in Section 3.4 and Table 1. Section 3.1 describes a 9-patient train / 1-patient test cross-validation, but no per-fold, mean, variance, or significance results are given. The reported advantage over the strongest baseline, CTformer, is SSIM +0.0021 and RMSE -0.0090; without any measure of variability or a paired test, this difference is indistinguishable from training noise and could reflect favorable patient selection. Moreover, the efficiency claim is directly at odds with Table 1: CTLformer uses 1.85M parameters while CTformer uses 1.45M, so on the paper's own parameter-based efficiency measure, CTLformer is not more efficient than the baseline it claims to surpass. No runtime, FLOPs, or throughput measurements are provided. Thus, the central claim rests on an unvalidated single-case comparison and an internally inconsistent efficiency metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CTLformer, a hybrid convolutional-transformer model for low-dose CT denoising. Two novel components are claimed: a multi-scale attention mechanism based on Token2Token and self-attention interaction, and a dynamic attention control mechanism that adapts attention to noise characteristics. The model is evaluated on the 2016 AAPM Mayo Clinic LDCT Challenge dataset against RED-CNN, WGAN-VGG, and CTformer, reporting SSIM, RMSE, and parameter counts. The abstract and conclusion claim that CTLformer significantly outperforms existing methods in both denoising performance and model efficiency.","tokens_in":5429,"tokens_out":4397,"duration_ms":43577,"significance":"If the reported architecture were fully specified and the performance claims were supported by rigorous evaluation, the proposed hybrid design could be a useful contribution to LDCT denoising, particularly the idea of noise-adaptive attention. However, the manuscript as written does not provide the necessary evidence: the evaluation is limited to a single patient slice, no statistical significance is established, and the efficiency claim is contradicted by the paper's own parameter counts. The strengths of the work—the qualitative architectural ideas—are not yet backed by reproducible or convincing experimental support.","major_comments":[{"comment":"The central claim of 'significantly outperforms existing methods' rests on metrics from a single slice of a single patient (L506, slice 057). Section 3.1 describes a 9-patient train / 1-patient test cross-validation, but no per-fold results, means, standard deviations, or paired significance tests are reported. The differences against the strongest baseline, CTformer, are SSIM +0.0021 and RMSE −0.0090, which are within the range expected from training noise and could reflect favorable patient selection. The paper must report results across all test patients (and ideally all slices), with variability estimates and a paired statistical test (e.g., Wilcoxon signed-rank), before any claim of significant superiority is made.","section":"§3.4, Table 1"},{"comment":"The abstract and conclusion claim that CTLformer outperforms existing methods in 'model efficiency,' but Table 1 reports CTLformer with 1.85M parameters versus CTformer with 1.45M parameters. On the paper's own parameter-based efficiency measure, CTLformer is less efficient than the baseline it claims to surpass. No runtime, FLOPs, throughput, or memory measurements are provided. The efficiency claim is therefore internally inconsistent and unsupported; either the claim must be retracted or substantiated with appropriate efficiency metrics.","section":"§3.4, Table 1, Abstract"},{"comment":"The two key contributions—the multi-scale attention mechanism and the dynamic attention control mechanism—are described only in qualitative prose. No equations, tensor shapes, layer configurations, or algorithmic details are given, making the architecture impossible to reproduce from the text. For example, it is unclear how the 'dynamic adjustment factors' are computed from noise patterns and how they reallocate attention weights. The paper should provide a precise mathematical specification of both modules, including the exact computation of the attention maps and the integration with the Token2Token representation.","section":"§2.1, §2.2"},{"comment":"Section 3.2 states that RMSE, PSNR, and SSIM are adopted as evaluation metrics, but Table 1 reports only SSIM and RMSE. PSNR is missing from the results. Furthermore, no ablation study is provided to isolate the contributions of the multi-scale attention and dynamic attention control components; without such an ablation, the observed (marginal) improvements cannot be attributed to the proposed mechanisms. The authors should report all three metrics and include ablation experiments that remove each of the two proposed components.","section":"§3.2, §3.4, Table 1"}],"minor_comments":[{"comment":"The keyword 'denoisng' is misspelled; it should be 'denoising.'","section":"Keywords"},{"comment":"The first row of Table 1 is labeled 'LCDT,' which appears to be a typo for 'LDCT' (low-dose CT). This should be corrected.","section":"§3.4, Table 1"},{"comment":"In §3.4 the text states 'CTLformer[12]' but reference [12] is the CTformer paper (Wang et al., PMB 2023), not the proposed CTLformer. This is a mis-citation and needs to be fixed.","section":"§3.4, Reference [12]"},{"comment":"The implementation details mention only hardware and the initial learning rate; the number of training epochs, loss function, optimizer, patch size, and denoising-specific preprocessing are not stated. These are necessary for reproducibility.","section":"§3.3, Implementation Details"},{"comment":"The dynamic attention control mechanism is described as using 'a lightweight network, a Fully connected layer' but in the introduction it is called 'MLP or convolution.' This inconsistency should be resolved, and the exact architecture of the module should be specified.","section":"§2.2"},{"comment":"The figures are referenced as 'Architecture of our method' and 'Architecture of Token2Token Block,' but in the manuscript text the figures are not described in sufficient detail to align with the prose. The figures should be self-contained with clear labels for each module and tensor resolution.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is currently far from publishable due to the extreme thinness of the experimental evaluation and the internal inconsistency in the efficiency claim. I would be willing to reconsider after the authors provide full cross-validation results with statistics, a complete architectural specification, and corrected claims. If the authors are unable or unwilling to add multi-patient evaluation and significance testing, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nQuick take: this is a thin empirical report that does not support its central claim. The reported gain over CTformer is SSIM +0.0021 and RMSE -0.009, on one test patient, with no error bars. The efficiency claim is contradicted by the paper's own Table 1 (1.85M vs 1.45M parameters). I would let this one go unless the authors return with a proper cross-validation.\n\nWhat is actually new: not much. The hybrid CNN-transformer direction is established; the Token2Token mechanism is taken from CTformer (ref 12). The two 'innovations' are described in prose with no equations or precise definitions, so a reader cannot tell what was actually built. The paper does use a standard public dataset (Mayo L506) and standard metrics, which is the right starting point.\n\nSoft spots, in order of severity:\n1. Single-patient evaluation. Section 3.1 describes a 9-patient train / 1-patient test cross-validation, but results appear only for L506. No mean, standard deviation, or per-fold numbers are given. A 0.002 SSIM difference could easily be training noise.\n2. Efficiency claim is wrong as stated. Table 1 shows CTLformer uses 1.85M parameters versus CTformer's 1.45M, so the abstract's 'model efficiency' claim is undercut by the paper's own numbers. No runtime, FLOPs, or throughput are provided.\n3. The architecture is underspecified. No equations for the multi-scale attention or dynamic adjustment, no labeled architectural diagram, no code. This makes the work impossible to reproduce or verify.\n4. Sloppy citation and table errors: Table 1 labels CTLformer as '[12]' (ref 12 is CTformer), and the baseline 'LCDT' is undefined. These are minor but signal a lack of polish.\n\nOn the positive side, the motivation is clear, the dataset is appropriate, and the metrics (SSIM, RMSE, PSNR) are standard for LDCT denoising. The idea of adapting attention based on noise characteristics is worth exploring, but it is not developed enough here to count as a contribution.\n\nWho is this for? Maybe an early-stage researcher wanting a quick survey of LDCT denoising approaches. It is not ready for publication as is, and I would not cite it.\n\nMy recommendation: desk reject, but encourage resubmission if the authors provide full per-patient results, ablations, a precise architecture specification, and code. The central flaw is fixable, but the current version does not meet the threshold.\n\nCheers,\n[Your name]","headline":"Single-patient metrics and an internally inconsistent efficiency claim sink an otherwise routine hybrid CNN-transformer LDCT denoising paper.","tokens_in":6021,"tokens_out":2243,"would_cite":false,"duration_ms":21149,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid CNN-transformer model claims sharper low-dose CT images at 1.85M parameters","keywords":["low-dose CT denoising","transformer","convolutional neural network","multi-scale attention","dynamic attention control","medical image reconstruction","Token2Token","image quality assessment"],"falsifier":"Run the same 9-patient/1-patient cross-validation for all ten patients in the 2016 low-dose CT challenge dataset and report SSIM, RMSE, and PSNR per patient; if CTLformer does not beat CTformer and RED-CNN on a clear majority of patients, the paper's central claim is falsified.","tokens_in":4976,"feed_emoji":"🩻","tokens_out":7711,"duration_ms":71907,"temperature":0.7,"pith_summary":"This paper proposes CTLformer, a hybrid model that combines convolutional layers with transformer self-attention for low-dose CT image denoising. The authors' central claim is that two additions solve the core trade-off in LDCT reconstruction: a multi-scale attention mechanism built on Token2Token and self-attention interaction preserves both fine textures and large-scale anatomy, while a dynamic attention control module shifts focus toward noisy regions and away from clean ones. They report that on the benchmark low-dose CT challenge data, CTLformer achieves the best SSIM and RMSE among the compared models while using 1.85 million parameters, which they take as evidence that the hybrid design improves both denoising quality and efficiency. If the claim holds, it points to a practical recipe for building smaller, more clinically deployable denoising models.","feed_headline":"Hybrid model claims sharper low-dose CT images at 1.85M parameters","feed_subtitle":"It pairs convolutions with self-attention to beat CNN-only and transformer-only baselines.","key_machinery":"The argument is carried by two interlocking attention mechanisms inside a hybrid architecture. The multi-scale attention mechanism uses a Token2Token module to produce tokens at several resolutions and a self-attention interaction module that treats local and global attention ranges separately, alternately boosting one and then the other; this is what lets the network keep small textures and large structures in the same representation. The dynamic attention control mechanism is a lightweight fully connected layer that reads the input image's noise and texture pattern and produces per-region adjustment factors, which are used to reallocate attention toward high-noise regions while preserving detail in low-noise regions. Convolutional layers perform local feature extraction, and overlapping inference reduces boundary artifacts at inference time. Together these parts supply the mechanism by which CTLformer aims to beat both convolutional-only RED-CNN and transformer-only CTformer.","core_discovery":"The central discovery claimed by the paper is that a convolution-free Transformer for low-dose CT denoising, CTformer, can be improved by adding convolutional feature extraction and two attention-control mechanisms without growing the model. In the proposed CTLformer, multi-scale tokens are generated from the input through a Token2Token module, and a self-attention interaction unit alternately strengthens local and global attention so that edges and textures are not lost while whole-organ structure is preserved. A second module, the dynamic attention control, uses a lightweight fully connected layer to estimate per-region noise characteristics from the input and reallocates attention weights accordingly, concentrating suppression on high-noise areas. On the held-out test patient used in the paper, CTLformer reports SSIM 0.9141 versus 0.9120 for CTformer and RMSE 9.0133 versus 9.0223, with the same parameter count as RED-CNN at 1.85 million, and the authors interpret these numbers as demonstrating superiority over pure CNN and pure transformer baselines.","pith_inferences":["Editorial extension: the reported advantage on one test patient is too small to establish general superiority by itself; averaging across the ten cross-validation folds would show whether the architecture consistently wins or only on that anatomy.","Extension: an ablation that switches off the dynamic attention control module would reveal how much of the gain comes from that mechanism rather than from the multi-scale tokens; the paper does not include one.","Extension: because the paper reports only parameter count as an efficiency measure, a natural follow-up is to measure per-image runtime and memory on the same GPU to see whether the claimed efficiency holds in practice.","Extension: the same dynamic attention idea could be tested on artificially noised full-dose CT images with known noise levels to see whether the model's attention maps actually shift toward high-noise regions as intended."],"forward_implications":["Convolution and self-attention can be combined at 1.85 million parameters without sacrificing either local detail or global context, suggesting that smaller hybrid models are viable for LDCT denoising.","A single model with dynamic attention control could in principle handle CT images with varying noise levels across patients, reducing the need for per-dose retraining.","The reported improvement over CTformer implies that transformer-only denoisers lose texture detail to over-smoothing and that adding convolutional layers is a direct remedy.","Overlapping inference as a post-processing strategy could be applied to other patch-based medical image restoration models to reduce stitching artifacts.","If the mechanism transfers, hybrid local-global attention could improve other image restoration tasks where noise is spatially non-uniform, such as low-dose MRI or ultrasound."],"supporting_citations":[{"why":"supplies the low-dose and full-dose CT image pairs used for training and testing, plus the RED-CNN baseline architecture that CTLformer is compared against.","marker":"[10]"},{"why":"defines the CTformer baseline, the main comparison point whose reported SSIM and RMSE the paper improves on.","marker":"[12]"},{"why":"provides the self-attention mechanism that forms the transformer branch of the hybrid architecture.","marker":"[7]"},{"why":"cited as the basis for the multi-scale attention with local-global interaction in the proposed model.","marker":"[9]"}],"fun_headline_variants":["Hybrid CT denoiser tops pure CNN and transformer in low-dose test","CTLformer: hybrid attention cuts noise, keeps details in low-dose CT","1.85M parameters: hybrid model beats CNN/transformer on LDCT","Self-attention plus convolutions: sharper low-dose CT at same cost","Multi-scale attention and dynamic control improve CT denoising"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results are presented for one held-out patient and largely one representative slice; if other test patients show smaller or reversed differences, the claimed superiority would not hold.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid CT denoiser tops pure CNN and transformer in low-dose test","CTLformer: hybrid attention cuts noise, keeps details in low-dose CT","1.85M parameters: hybrid model beats CNN/transformer on LDCT","Self-attention plus convolutions: sharper low-dose CT at same cost","Multi-scale attention and dynamic control improve CT denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1541,"prompt_tokens":1014,"completion_tokens":527,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":430}},"tokens_in":630,"tokens_out":527,"duration_ms":4729,"temperature":1.0,"reasoning_tokens":430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:38:37.938862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 9-patient/1-patient cross-validation for all ten patients in the 2016 low-dose CT challenge dataset and report SSIM, RMSE, and PSNR per patient; if CTLformer does not beat CTformer and RED-CNN on a clear majority of patients, the paper's central claim is falsified.","supporting_citations":[],"review_version":1}