{"id":"6cbe7150-8d54-4a17-8679-7968f287bfd2","arxiv_id":"2411.18975","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A U-Net variant with Fourier-analysis and self-attention blocks reports tiny gains on skin-lesion segmentation, but the abstract's key metrics belong to a different model.","lead":"FAN-UNet adds Fourier analysis blocks and self-attention to the standard U-Net for biomedical image segmentation. The paper claims small improvements on two skin-lesion datasets, but its headline numbers actually belong to another model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on Table 1, but the abstract's headline metrics (96.07/78.83/88.16) exactly match Med-TTT [23] rather than FAN-UNet's own row (96.30/78.99/88.25), and no code or hyperparameters are given to verify the numbers.","rationale":"The reader correctly identifies the untested periodicity assumption as a weakness, but the more load-bearing concern is the integrity of the experimental evidence itself. The abstract of the paper reports numbers that are identical to Med-TTT's ISIC17 row, not FAN-UNet's own Table 1 row. This is a concrete, checkable inconsistency: if the headline results are actually a baseline's numbers, then the paper does not even state its own results consistently, and the reader cannot tell which numbers correspond to FAN-UNet. Combined with the total absence of training details and code, the empirical basis for 'state-of-the-art performance' collapses. The periodicity motivation is a secondary weakness: even if ISIC images lack periodic structure, the architecture could still perform well via its attention and convolution components, so the absence of a periodicity test does not by itself invalidate the empirical claim. But the unreproducible and misattributed metrics go directly to the central claim. My concern therefore agrees partially with the reader's weakest assumption: both point to inadequate support, but I locate the primary failure in the experimental reporting rather than in the domain motivation. The proposed test—retraining from the paper's own specification—would settle whether the Table 1 numbers are genuine; the paper currently provides no way to run it, which is itself the problem. The verdict of REJECT stands, as the central claim is not supported by verifiable evidence.","tokens_in":7830,"tokens_out":2957,"duration_ms":27525,"concrete_test":"Request the authors' training code and configuration, then re-train FAN-UNet on the ISIC17 training split using the architecture described in Sections 3.1–3.2 and the loss in Eq. (10) with α=0.5, evaluating on the standard test split. If the resulting mIoU and Dice do not match Table 1 (78.99% and 88.25%) within a plausible tolerance, or if the authors cannot supply code, logs, or a complete training specification, the SOTA claim should be treated as unverified. As a secondary check, cross-reference the abstract's metrics (96.07/78.83/88.16) against Med-TTT's published row to confirm the misattribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—FAN-UNet achieves state-of-the-art segmentation on ISIC17/ISIC18—depends entirely on the experimental numbers in Table 1. That evidence is unreliable on two counts. First, the abstract reports accuracy 96.07%, mIoU 78.83%, and DSC 88.16% as FAN-UNet's results; these are exactly the ISIC17 row for Med-TTT (reference [23], an earlier paper by the same author) in Table 1. The actual FAN-UNet row is 96.30/78.99/88.25. This internal inconsistency means the headline numbers in the paper are not the proposed model's numbers, and it raises the possibility that Table 1's FAN-UNet entries were carried over from a baseline rather than obtained by training FAN-UNet. Second, Section 4 provides no training configuration: no optimizer, learning rate, number of epochs, image size, augmentation, normalization, hardware, or random seed, and no code is released. Without these, no independent check of Table 1 is possible. The periodicity motivation (Sections 1 and 3.1) is also untested—no spectral analysis of ISIC images is offered—but even if that motivation were sound, the missing experimental verification would still leave the central empirical claim unsupported. The load-bearing premise is therefore not merely the periodicity assumption; it is that the reported performance numbers are genuine. That premise is currently unsupported and directly contradicted by the abstract/Table 1 mismatch.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FAN-UNet, a U-Net variant in which the standard encoder blocks are replaced by Vision-FAN blocks combining window-based self-attention with a 2D Fourier Analysis layer (FANLayer2D). The method is evaluated on the ISIC 2017 and ISIC 2018 skin lesion segmentation benchmarks and compared with CNN-, Transformer-, and Mamba-based baselines. The paper claims state-of-the-art results on ISIC17 with 78.99% mIoU, 88.25% Dice, and 96.30% accuracy, and reports an ablation study on ISIC18. The central claim is that explicitly modeling periodicity via Fourier features improves segmentation performance.","tokens_in":8112,"tokens_out":6469,"duration_ms":53098,"significance":"If the reported results were reproducible and the Fourier component were shown to be responsible for the improvement, the contribution would be a modest but useful addition to medical image segmentation: integrating FAN into U-Net is a simple and potentially principled way to capture periodic structures. However, the experimental evidence as presented is internally inconsistent and lacks reproducibility information. The abstract reports metrics identical to the Med-TTT baseline rather than the FAN-UNet row in Table 1, one ablation row is numerically identical to a Table 1 baseline, and no training configuration or code is provided. Until these issues are resolved, the paper cannot support its central claim of state-of-the-art performance.","major_comments":[{"comment":"The abstract and the contribution bullet in §1 report FAN-UNet's ISIC17 results as accuracy 96.07%, mIoU 78.83%, and DSC 88.16%. These are exactly the Med-TTT row in Table 1 (reference [23]), whereas FAN-UNet's own row in Table 1 is 96.30%, 78.99%, and 88.25%. The paper's headline numbers are therefore not the proposed model's numbers, contradicting Table 1 and undermining the credibility of the reported experimental results.","section":"Abstract and §1"},{"comment":"The claimed state-of-the-art advantage over Med-TTT and HC-Mamba is tiny (e.g., +0.16 mIoU over Med-TTT on ISIC17), and the paper provides no error bars, confidence intervals, or significance tests. With only single-point comparisons and no information about the number of runs or variance, the claimed superiority is not statistically supported.","section":"§4.2, Table 1"},{"comment":"Section 4 gives no training details: no optimizer, learning rate, batch size, number of epochs, input size, data augmentation, normalization, hardware, random seed, or train/validation protocol for ISIC17/ISIC18. No code is released. This makes every number in Tables 1 and 2 impossible to reproduce or verify, leaving the central empirical claim unsupported.","section":"§4.1–§4.2"},{"comment":"The row 'Proposed w/o Vision-FAN Block' in Table 2 is numerically identical to the UNet row for ISIC18 in Table 1 (77.86/87.55/94.05/96.69/85.86). If the ablation was conducted by replacing the block with a convolution, exact equality with the pre-existing UNet baseline is implausible without additional explanation; the result suggests the ablation numbers may have been copied from the baseline table rather than produced by an actual experiment.","section":"§4.3, Table 2"},{"comment":"The periodicity premise—that skin lesion images exhibit consistent periodic patterns—is asserted without evidence. The only citation is [15], which concerns quasi-periodic patterns in resting-state fMRI, a different imaging modality. No spectral analysis of ISIC images, frequency-domain diagnostics, or controlled experiment isolating the contribution of the Fourier terms is provided, so the architectural motivation is not validated empirically.","section":"§3.1 and §1"}],"minor_comments":[{"comment":"The phrase 'we presents' is ungrammatical and should be corrected to 'we present'.","section":"Abstract"},{"comment":"The title uses inconsistent spacing ('FAN-U NET' in the running head versus 'FAN-UNet' in the text); please unify.","section":"Title and running head"},{"comment":"Mathematical notation is inconsistent (e.g., Wp in Eq. (7) vs. W'_p in Eq. (6), and the capitalized 'Compute' mid-sentence). Please standardize notation and proofread the equations.","section":"§3.2, Eqs. (7)–(9)"},{"comment":"Equation (10) is not clearly typeset and does not define Input_b, Target_b, epsilon, or B; these should be specified for the formula to be meaningful.","section":"§3.3, Eq. (10)"},{"comment":"The paper states 'Extensive experiments on various medical imaging datasets' but evaluates on only two ISIC datasets, both dermoscopic skin lesion datasets; this overstates the scope.","section":"§4.1"},{"comment":"The abstract and conclusion mention a 'favorable balance between model complexity and performance,' but no model complexity metrics (parameters, FLOPs, inference time) are reported.","section":"§4.2"}],"recommendation":"reject","confidential_remarks":"The exact numerical coincidences—abstract metrics identical to Med-TTT and the 'Proposed w/o Vision-FAN Block' ablation row identical to the UNet baseline—raise serious concerns about the integrity of the experimental record. I recommend that the editor request raw logs, trained models, or code before considering any revision. Additionally, the two strongest baselines (HC-Mamba and Med-TTT) are the author's own prior papers, and the claimed improvements over them are within 0.2 mIoU with no variance estimates, which further limits independent verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a U-Net with a Fourier-analysis block grafted in. That combination is new as far as the citations go, and the method section is clear enough that someone could reimplement it. The ablation on ISIC18 gives some support for the Vision-FAN block. Credit where it's due: the architecture is simple, and the writing is direct.\n\nBut the paper fails on its own evidence. The abstract reports 96.07% accuracy, 78.83% mIoU, and 88.16% DSC as FAN-UNet's results. Those are exactly the Med-TTT row in Table 1. FAN-UNet's own row is 96.30/78.99/88.25. This is not a rounding difference; it's the wrong model's numbers in the summary. That alone makes it hard to trust any of the reported comparisons.\n\nThe rest is not much better. No code, no hyperparameters, no data preprocessing, no error bars. The gains over Med-TTT are a fraction of a percent—0.16 mIoU on ISIC17, 0.15 on ISIC18—and without variance estimates they're indistinguishable from noise. The motivation is asserted, not tested. The paper leans on a fMRI citation to claim skin lesions have periodic structure, but never shows any spectral analysis of the ISIC images. The author's own prior papers are two of the three strongest baselines, which is fine, but it does mean the numbers come from an unverifiable in-house pipeline.\n\nThe ablation table is a bright spot, though even there the 'w/o Vision-FAN Block' row exactly matches the UNet column in Table 1, which may be legitimate but deserves a footnote.\n\nMy take: this is a routine architecture combination that could be a workshop paper if the numbers were corrected and the code released. As a full paper, the abstract mismatch and the missing reproducibility make the central claim unsupported. I would not send it to a serious referee until those are fixed. If the authors clean it up, it might be a citable data point, but not in its current form.","headline":"A U-Net plus Fourier layer combo that is clearly described but undermined by a headline numbers mismatch and no reproducibility.","tokens_in":8701,"tokens_out":3000,"would_cite":false,"duration_ms":26856,"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":"FAN-UNet adds Fourier-analysis blocks to U-Net and reports top segmentation scores on two skin-lesion benchmarks.","keywords":["Fourier Analysis Network","U-Net","medical image segmentation","self-attention","periodicity modeling","skin lesion segmentation","ISIC2017","ISIC2018"],"falsifier":"Retrain FAN-UNet on ISIC17 with the FANLayer2D activation replaced by an equal-capacity nonlinearity; if mIoU and Dice do not drop, the Fourier component is not the source of the gain. Also, compute dominant-frequency or autocorrelation statistics on lesion crops: if periodic content is no stronger than natural images, the periodic-pattern premise fails.","tokens_in":7551,"feed_emoji":"🩺","tokens_out":6903,"duration_ms":54774,"temperature":0.7,"pith_summary":"The paper proposes FAN-UNet, a U-Net variant whose encoder blocks replace plain convolutions with Vision-FAN blocks: window-based self-attention for global context plus FANLayer2D, a layer that applies cosine and sine activations to learnable linear projections so the network can explicitly represent periodic structure. The stated goal is to fix two weaknesses at once—CNNs' poor long-range dependency modeling and standard networks' difficulty modeling periodic patterns—while keeping the model practical. On the ISIC17 and ISIC18 skin-lesion segmentation benchmarks, FAN-UNet reports the best mIoU and Dice among the compared models, including 78.99% mIoU, 88.25% Dice, and 96.30% accuracy on ISIC17. A sympathetic reader would care because the paper offers a concrete recipe for adding Fourier-style inductive biases to a standard segmentation backbone, and it suggests that explicit periodicity modeling can matter for biomedical images.","feed_headline":"Fourier analysis lifts U-Net segmentation on skin-lesion data","feed_subtitle":"Adding sine-cosine Fourier features and self-attention to U-Net edges out state-of-the-art models on ISIC17 and ISIC18.","key_machinery":"The Vision-FAN Block is the core mechanism. It applies positional encoding to the downsampled feature map, runs window-based self-attention to capture long-range dependencies, and passes the result to FANLayer2D. FANLayer2D computes $\\mathbf{P} = \\mathbf{W}_p \\ast \\mathbf{X} + \\mathbf{b}_p$ and $\\mathbf{G} = \\mathbf{W}_g \\ast \\mathbf{X} + \\mathbf{b}_g$, applies sine and cosine activations to $\\mathbf{P}$ and a nonlinear activation to $\\mathbf{G}$, concatenates the periodic and nonlinear features along the channel dimension, and returns them as the output. Residual connections and layer normalization preserve input information and stabilize training. The block is inserted into the U-Net downsampling path, and the training loss is a batch-level combination of cross-entropy and Dice loss with mixing weight $\\alpha = 0.5$.","core_discovery":"The paper's central claim is that integrating FANLayer2D—a layer built from the Fourier-series expansion $f_S(\\mathbf{x}) = B + \\mathbf{W}^c \\cos(\\mathbf{W}^{in}\\mathbf{x}) + \\mathbf{W}^s \\sin(\\mathbf{W}^{in}\\mathbf{x})$—into a U-Net, together with window-based self-attention, yields segmentation accuracy that surpasses recent CNN, Transformer, and Mamba-based models on ISIC17 and ISIC18. The paper attributes the gain to two complementary mechanisms: self-attention models long-range spatial dependencies, while the sine/cosine activations explicitly encode the regular, repetitive texture of tissue and lesion regions. In the paper's numbers, FAN-UNet reaches 78.99% mIoU and 88.25% Dice on ISIC17 and 78.74% mIoU and 88.11% Dice on ISIC18, and the ablation shows that removing the Vision-FAN Block lowers mIoU from 78.74% to 77.86%.","pith_inferences":["A direct test of the periodic-pattern premise would be to replace only the sine/cosine branch in FANLayer2D with an equivalently sized random nonlinearity; if mIoU and Dice stay flat, the gain comes from added capacity, not from Fourier structure.","The paper cites quasi-periodic brain fMRI patterns [15] as evidence that medical features are periodic, but never measures periodicity in dermatoscopic images; computing dominant frequencies or autocorrelation over lesion crops would settle whether the premise holds.","If the premise does hold, the same block may reduce the amount of training data needed for segmentation in modalities with repetitive texture, and may extend naturally to 3D volumes (CT/MRI) by replacing 2D convolutions with 3D Fourier layers.","A per-class or boundary-focused evaluation would show whether the Fourier block specifically improves edges and textured regions, which the aggregate metrics in the paper do not reveal."],"forward_implications":["If the claim holds, adding explicit frequency-domain activations to a U-Net is a low-cost way to improve segmentation on structured, repetitive textures.","The hybrid attention-plus-Fourier block is a recipe that could transfer to other biomedical segmentation tasks where anatomy or pathology appears with regular spatial repetition.","The reported margins over recent models—1.11% mIoU over HC-Mamba and 2.01% over U-Net on ISIC17—place Fourier feature modeling in the same range as recent Mamba-based global modeling.","The ablation numbers support the block's contribution: removing the Vision-FAN Block drops ISIC17 mIoU from 78.74% to 77.86%, and removing positional encoding drops it to 78.42%."],"supporting_citations":[{"why":"Provides the U-Net encoder-decoder architecture that FAN-UNet modifies.","marker":"[5]"},{"why":"Introduces Fourier Analysis Networks, the source of FANLayer2D's periodic-modeling layer.","marker":"[14]"},{"why":"Supplies the claim that medical imaging features exhibit consistent periodic patterns, used to justify the Fourier design.","marker":"[15]"},{"why":"Defines the ISIC2017 skin-lesion dataset, one of the two evaluation benchmarks.","marker":"[16]"},{"why":"Defines the ISIC2018 challenge dataset, the second evaluation benchmark.","marker":"[17]"},{"why":"Describes the HAM10000 dataset, a data source underlying ISIC2018.","marker":"[18]"},{"why":"HC-Mamba is the strongest recent baseline that FAN-UNet reports beating on ISIC17.","marker":"[22]"},{"why":"Med-TTT is the closest baseline on ISIC18 in the comparison table.","marker":"[23]"}],"fun_headline_variants":["Fourier features teach U-Net long-range context for skin lesions","Sine-cosine layer boosts U-Net on ISIC17 and ISIC18","Fourier analysis helps U-Net beat CNNs and Transformers in segmentation","FAN-UNet adds Fourier insight to U-Net for medical images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that skin-lesion images contain consistent periodic patterns, so that explicitly encoding sine and cosine features helps segmentation; if that premise is false, the reported improvements could come from added parameters rather than from Fourier analysis.","fun_headline_variants_meta":{"raw":{"variants":["Fourier features teach U-Net long-range context for skin lesions","Sine-cosine layer boosts U-Net on ISIC17 and ISIC18","Fourier analysis helps U-Net beat CNNs and Transformers in segmentation","FAN-UNet adds Fourier insight to U-Net for medical images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1334,"prompt_tokens":941,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":309}},"tokens_in":557,"tokens_out":393,"duration_ms":4485,"temperature":1.0,"reasoning_tokens":309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:40:15.818390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FAN-UNet on ISIC17 with the FANLayer2D activation replaced by an equal-capacity nonlinearity; if mIoU and Dice do not drop, the Fourier component is not the source of the gain. Also, compute dominant-frequency or autocorrelation statistics on lesion crops: if periodic content is no stronger than natural images, the periodic-pattern premise fails.","supporting_citations":[{"cited_title":"Quasi- periodic patterns (qpp): large-scale dynamics in resting state fmri that correlate with local infraslow electrical activity","cited_arxiv_id":null,"evidence_quote":"Supplies the claim that medical imaging features exhibit consistent periodic patterns, used to justify the Fourier design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ISIC2017 skin-lesion dataset, one of the two evaluation benchmarks."},{"cited_title":"Med-TTT: Vision Test-Time Training model for Medical Image Segmentation","cited_arxiv_id":"2410.02523","evidence_quote":"Med-TTT is the closest baseline on ISIC18 in the comparison table."}],"review_version":1}