{"id":"a9b44a18-0bc0-4cb7-979d-a77a7d40d408","arxiv_id":"2509.00669","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cepstrum-based GLCM texture features give small, inconsistent improvements to melanoma versus nevus classification on ISIC 2019 in a feature-ablation study with no error bars.","lead":"This paper tests a new way to measure texture in skin lesion images: applying an established texture-analysis tool (GLCM) to a mathematical transform called the cepstrum, which highlights repeating patterns. On the ISIC 2019 dataset, adding these features to other handcrafted features gave modest but sometimes inconsistent improvements in melanoma versus nevus classification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline AUC is selected on validation rather than measured on a held-out test set; no test-set AUC is reported, so the central performance claim is unverified.","rationale":"I read the paper in good faith as proposing a genuinely novel feature construction—GLCM statistics applied to 2D cepstra—and testing it on a clinically meaningful binary melanoma-versus-nevus task. The feature extraction is plausible and parameter-light, and the use of unique-lesion splits to avoid duplication is a positive step. However, the empirical claim that these features 'offer complementary discriminatory information' depends entirely on performance numbers that are trustworthy and generalizable. The weakest link is the evaluation protocol: greedy feature selection is driven by validation AUC, and the headline 0.9697 is reported from that same selection run, with no held-out test AUC for the final model. This is not a mathematical inconsistency, but it is exactly the kind of selection-bias problem that makes reported performance optimistic. The reader's weakest_assumption identified the same issue, and I agree. I also note internal tensions in the paper: Table 2 contradicts the 'consistently improved' wording, and the authors acknowledge that adding cepstral features to the full feature set did not improve performance. These do not disprove the feature's value, but they show the reported improvement is not as clean as the abstract implies. A single concrete check—reporting test-set AUC for the frozen selected model—would settle whether the central claim survives. Because the paper's methodology is reasonable and the flaw is a missing evaluation step rather than an irreparable design error, the appropriate verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":7220,"tokens_out":2844,"duration_ms":37252,"concrete_test":"Run the existing 80/20 unique-lesion split with a strict two-stage protocol: (1) perform greedy forward selection on the training/validation portion exactly as described, freeze the feature subset and XGBoost model at the iteration that produced AUC 0.9697; (2) evaluate that frozen model once on the untouched test subset. Report ROC AUC, accuracy, and F1 with 95% bootstrap confidence intervals, for the model with and without cepstral features. If the test AUC is materially below 0.9697, or if the with-cepstral minus without-cepstral test AUC is not positive, the central contribution claim fails. If the test AUC is close to 0.9697 and the cepstral increment persists, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that cepstral features offer complementary discriminatory information, with the headline 'highest-performing model achieving an AUC of 0.9697' after greedy feature selection. Section IV describes greedy forward selection as evaluating features by ROC AUC on a validation set, and Fig. 7 plots 'ROC AUC per iteration of the greedy feature selection algorithm'—by construction, this is validation AUC. The 0.9697 value appears to be the maximum of this selection run, not an independent test-set measurement. The paper states an 80/20 split reserved unique lesions for testing and that the test set was unseen during training, but it never reports test-set performance for the final selected feature set. Because the validation set was used both to choose features and to report the headline number, the result is a selection maximum and is likely optimistic. This is load-bearing: if the validation-to-test generalization fails, the 'complementary discriminatory information' claim is unsupported. Internal evidence reinforces the concern: Table 2 shows Median Split F1 and AUC decrease with cepstral features, and the text admits that when all feature sets are combined, adding cepstral features did not improve performance. Thus the claim of consistent improvement is not robust as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 35-element cepstrum-derived feature set for melanoma versus nevus classification on ISIC 2019 dermoscopic images. The central novelty is applying GLCM Haralick statistics to the 2D real cepstrum of image channels in multiple color spaces, along with non-GLCM cepstral statistics such as entropy and radial profile metrics. These features are combined with several established handcrafted feature families (median-split color, atypical pigment network, pink region, saliency, telangiectasia, semitranslucent area) and evaluated with XGBoost. The authors use greedy forward selection with validation ROC AUC, report improvements in accuracy, F1, and AUC when cepstral features are added to individual feature sets, and state a highest-performing model with AUC 0.9697. They conclude that cepstral features offer complementary discriminatory information. The paper does not, however, report a test-set evaluation of the selected model, and Table 2 shows inconsistent gains.","tokens_in":7610,"tokens_out":5042,"duration_ms":61147,"significance":"If the cepstral GLCM descriptor generalizes, it would be a reusable and interpretable texture feature for dermoscopic image analysis and potentially other medical imaging tasks. The proposed features are fixed and label-independent, so there is no circular derivation, and the use of unique lesion IDs to avoid training/test leakage is a good methodological choice. The ablation across multiple handcrafted feature families is also a strength. The main weakness is that the reported evidence does not currently establish generalizability: the headline AUC is a validation-selection maximum, no test-set AUC is reported, and the improvements are not consistent across feature sets. With a proper held-out evaluation and uncertainty quantification, the contribution could be valuable, but the paper in its current form does not support its central claim.","major_comments":[{"comment":"The headline result \"highest-performing model achieving an AUC of 0.9697\" is not a test-set measurement. Section IV describes greedy forward selection using ROC AUC on a validation set, and Fig. 7 plots \"ROC AUC per iteration of the greedy feature selection algorithm,\" so 0.9697 is the maximum of the validation selection trajectory. The paper reserves a test set of unique lesions and states it was unseen, but no test-set AUC, accuracy, or F1 is reported for the final selected feature set. Because the validation set is used both to select features and to report the headline number, the result is optimistically biased. The central claim that cepstral features offer complementary discriminatory information is load-bearing and remains unverified. Please report the final selected model's performance on the reserved test set, or use nested cross-validation.","section":"IV, Fig. 7"},{"comment":"The text states that \"cepstral features consistently improve diagnostic performance when combined with existing handcrafted descriptors,\" but Table 2 does not support consistency. For the Median Split feature set, ROC AUC decreases from 0.9488 to 0.9435 and F1 score decreases slightly from 0.5899 to 0.5886 when cepstral features are added. The text also admits that \"when all feature sets were combined, model performance did not improve with the addition of cepstral features.\" Thus the improvement is limited to certain feature sets (notably the weaker baselines such as Telangiectasia and Salient) and is not robust as reported. The claim should be tempered or supported by a statistical analysis that accounts for variability.","section":"Table 2, Section IV"},{"comment":"No repeated runs, cross-validation folds, bootstrap confidence intervals, or error bars are reported. All comparisons in Table 2 and Figs. 7-8 are point estimates from a single split, and the XGBoost hyperparameters are not specified. Many of the reported differences (e.g., F1 changes of 0.005-0.01, AUC changes of 0.001-0.01) are small and could easily be within noise. Please provide repeated stratified splits or bootstrap resampling with confidence intervals, and specify the classifier hyperparameters, so that the reader can judge whether the observed differences are meaningful.","section":"IV, Table 2"},{"comment":"The preprocessing relies on the Lama segmentation algorithm to isolate the lesion, but no validation of mask accuracy on ISIC 2019 is provided. The cepstrum is computed after masking, and if the masks are inaccurate, the features will capture perilesional skin texture rather than intra-lesion texture. This directly affects the interpretation of the cepstral features as lesion descriptors. Please report quantitative mask quality (e.g., Dice coefficient on a sample) or a sensitivity analysis comparing results with different segmentation settings.","section":"II.A, Fig. 3"},{"comment":"The exact cepstral feature extraction is not fully specified, which prevents reproducibility. The paper does not state the cepstrum size or window used, the padding before the Fourier transform, the number of gray-level bins used for GLCM computation, or how the \"normalized unsigned integer representation\" of the cepstrum is constructed. These are free parameters that can materially change the resulting features. Please provide the full parameter settings or release code, so that the proposed 35-element descriptor can be reproduced by others.","section":"II.A-C"}],"minor_comments":[{"comment":"The feature list includes \"Radial Area-Under-the-Curve,\" but the text describes it as \"area under the curve (AUC) for the receiver operating characteristic (ROC).\" This is confusing: the radial AUC is the area under the radial profile, not the ROC AUC. Please clarify the terminology.","section":"II.B"},{"comment":"The text says the highest-performing model uses \"fewer than 100 total features\" while Fig. 8 states \"100 features were selected for each model.\" Please reconcile whether the final model uses exactly 100 or fewer features.","section":"IV, Fig. 8"},{"comment":"Reference [1] has the typo \"Alanysis\" in the title (should be \"Analysis\"); reference [10] is cited for the GLCM trace, but the cited paper appears to be a general GLCM paper and may not be the source of the trace statistic. Please verify the citations.","section":"References"},{"comment":"Please specify how the lesion mask is applied before the cepstral computation: zeroing the background, cropping to the bounding box, or something else. This affects the frequency content and the resulting cepstral features.","section":"II.A, Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline manuscript. The core idea is plausible, and the label-independent feature construction plus unique-lesion splitting are good practices. However, the evaluation as reported does not support the headline claim, and the paper heavily reuses the authors' own feature families. The main fix is a proper held-out test-set evaluation of the selected model with confidence intervals; if that is provided and the gains persist, the paper could become acceptable for a methods-oriented venue. If the authors cannot demonstrate test-set improvement, the central contribution would be significantly weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is applying GLCM/Haralick statistics to the 2D cepstrum of dermoscopic lesions. That combination is not in the cited literature, and the directionality max/mean quotient is a small, credible idea. The feature definitions are fixed and label-independent, so there is no circularity problem. The authors test their features against six established handcrafted feature families, which is useful ablation work. Most of Table 2 shows real gains, particularly for weaker baseline features like telangiectasia and saliency.\n\nThe soft spots are real and mostly fixable. The headline 0.9697 AUC comes from a greedy forward-selection run that evaluates on a validation set, and the paper never reports a held-out test AUC for the final selected model. So the central \"complementary\" claim is not yet verified. Table 2 also contradicts the word \"consistently\": median-split AUC drops from 0.9488 to 0.9435, and F1 drops a hair. The text itself admits that combining all feature sets did not improve performance. No error bars or repeated runs appear anywhere, so we have no sense of variance. And there is no deep-learning baseline, which matters because the ISIC task is currently dominated by deep models.\n\nNone of these flaws destroys the core idea; they prevent the paper from being accepted as-is. The segmentation masks from Lama's method are an unverified assumption, but that is a standard preprocessing choice, not a fatal issue.\n\nThe paper is not a breakthrough. It is an honest, modest addition to handcrafted feature engineering, with an interpretable recipe that could plausibly fuse into deep pipelines. The evaluation needs to be tightened: report test AUC for the final selected model, give error bars or repeated runs, clarify which numbers come from validation versus test, and soften the \"consistent improvement\" wording. If the authors do that, the result is a legitimate contribution to the dermoscopy CAD literature.\n\nMy recommendation: send it to peer review. It deserves a serious referee, and the fixable evaluation issues are exactly what review is for.","headline":"Novel GLCM-on-cepstrum feature recipe with a useful ablation, but the headline AUC is selected on validation and no test AUC is reported.","tokens_in":8003,"tokens_out":2546,"would_cite":true,"duration_ms":29995,"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":"A 35-element cepstrum-derived feature set — GLCM statistics computed on the 2D cepstrum of dermoscopic lesions — offers complementary discriminatory information for melanoma detection, reaching ROC AUC 0.9697 when fused with handcrafted fea","keywords":["cepstrum","melanoma detection","GLCM","Haralick texture features","dermoscopy","ISIC 2019","XGBoost","frequency-domain analysis"],"falsifier":"Run the same ISIC 2019 melanoma/nevus images through the same greedy selection but score the final 50- and 100-feature models on the held-out unique-lesion test set, comparing test AUC against the handcrafted-only baseline; if the cepstral-augmented model does not beat the baseline on that untouched partition, the complementary-information claim fails. A second check: recompute cepstral features on a sample of lesions using manually corrected masks and compare predictions, isolating how much segmentation error inflates the reported scores.","tokens_in":7190,"feed_emoji":"🔬","tokens_out":9162,"duration_ms":96108,"temperature":0.7,"pith_summary":"This paper sets out to establish that texture statistics computed on the 2D cepstrum of a lesion, rather than on the image itself, carry signals useful for telling melanoma from nevus in dermoscopic images. The proposed 35-element feature set applies Haralick's gray-level co-occurrence matrix statistics plus simple summary metrics to the cepstral representation across several color spaces, capturing repeating and anisotropic structures such as pigment networks and streaks. Fused with established handcrafted descriptors and an XGBoost classifier, the cepstral features raise ROC AUC, accuracy, and F1 on the ISIC 2019 binary task, with the best greedy-selected model reaching AUC 0.9697. The claim matters because it offers an interpretable, low-cost texture modality that can be added to existing pipelines without tuning a pixel-distance parameter. The paper also reports that when all feature groups were combined unconditionally, cepstral features did not help; selection was required.","feed_headline":"Cepstrum texture features lift melanoma detection to 0.9697 AUC","feed_subtitle":"Fourier-domain texture adds discriminating power to standard dermoscopic features.","key_machinery":"The 2D real cepstrum of a masked lesion channel — the inverse Fourier transform of the log-magnitude Fourier transform — converts repeating spatial structures (pigment networks, globules, streak patterns) into impulses and periodic peaks, isolating texture that is hard to describe in the spatial domain. On top of this representation, gray-level co-occurrence matrices are computed for pixel pairs at fixed distance one in four directions, yielding Haralick's 13 statistics plus the matrix trace; per-feature means over directions and max/mean quotients encode anisotropy ('directionality'). The pipeline adds seven non-GLCM metrics per channel (mean, standard deviation, skew, kurtosis, entropy, ra","core_discovery":"On the paper's own terms, the discovery is that fusing cepstral and GLCM analysis produces a viable texture descriptor for skin lesions: the real 2D cepstrum of each color channel turns periodic spatial patterns into peaks, and Haralick statistics computed on that cepstrum quantify their strength, directionality, and homogeneity. Correlation and mutual-information analysis singles out novel directionality features — sum-entropy differences across GLCM angles in Lab and YCrCb channels — as among the most melanoma-associated descriptors. In ablation, adding cepstral features improved ROC AUC for four of the six handcrafted feature families and raised F1 for nearly all, with the largest jumps o","pith_inferences":["If the cepstrum genuinely encodes periodic dermal structures, the same 35-feature recipe should transfer to other pigmented-lesion tasks (e.g., multiclass ISIC 2019, basal cell carcinoma vs. nevus) and to non-dermoscopic modalities such as histopathology or OCT, where periodicity is also diagnostic — a direct, cheap test.","The headline 0.9697 AUC is measured during greedy selection on a validation split; evaluating the final selected feature set on a fully held-out, never-seen test partition would settle how much of the gain is real generalization versus selection optimism. The paper describes an 80:20 unique-lesion test split but does not report its AUC for the selected model.","A deeper question the paper leaves open is whether cepstral features capture the same periodic cues that convolutional filters learn implicitly; if so, they could serve as a compact, interpretable substitute for deep features in low-data settings."],"forward_implications":["Cepstral GLCM features become a plug-in descriptor: any existing handcrafted or learned dermoscopic feature set can be augmented with the 35-element recipe and re-run through a classifier.","Feature families that are weak on their own — telangiectasia and saliency cues — gain the most from cepstral fusion, suggesting the transform supplies exactly the structural information those cues miss.","Because the method adds more to F1 than to AUC, its main practical payoff would be fewer missed melanomas in screening triage.","The fixed-distance, no-tuning GLCM computation makes the descriptor reproducible across datasets without parameter search."],"supporting_citations":[{"why":"Introduces the cepstrum and the property that time delays become impulses, the mathematical basis for the whole feature family.","marker":"[1]"},{"why":"Extends cepstral analysis to images, supplying the 2D cepstrum definition used for texture extraction.","marker":"[2]"},{"why":"Recent work feeding cepstra to CNNs; establishes the gap that interpretable cepstrum-derived features are unexplored.","marker":"[3]"},{"why":"Defines the 13 GLCM texture statistics that the paper computes on the cepstrum.","marker":"[4]"},{"why":"Source of the mixed-domain handcrafted feature fusion approach that forms the comparison baseline.","marker":"[5]"},{"why":"Supplies the segmentation algorithm whose masks isolate the lesion before cepstral analysis; mask accuracy is load-bearing.","marker":"[9]"},{"why":"The mahotas library that computes the Haralick features; the feature pipeline depends on it.","marker":"[11]"},{"why":"Collection of the handcrafted dermoscopic feature groups (APN, pink, saliency, telangiectasia, semitranslucent) that cepstral features are fused with and compared against.","marker":"[12]"},{"why":"The ISIC 2019 dataset components (HAM10000, ISBI 2017, BCN20000) supply all dermoscopic images and labels.","marker":"[6][7][8]"}],"fun_headline_variants":["Cepstral GLCM features improve melanoma detection metrics across the board","Fusing cepstrum and GLCM delivers complementary melanoma discrimination","New cepstrum texture features sharpen melanoma vs nevus classification","Melanoma AUC jumps with Fourier-domain texture descriptors","Directionality in cepstral GLCM strongly associates with melanoma"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The headline 0.9697 AUC is the best value seen during greedy forward selection on a validation split, and the paper reports no test-set AUC for the final selected model, so if selection has overfit the validation split the true gain on unseen lesions would be smaller; the other load-bearing assumption is that the automated lesion masks are accurate, since wrong masks would let background skin texture leak into the cepstrum.","fun_headline_variants_meta":{"raw":{"variants":["Cepstral GLCM features improve melanoma detection metrics across the board","Fusing cepstrum and GLCM delivers complementary melanoma discrimination","New cepstrum texture features sharpen melanoma vs nevus classification","Melanoma AUC jumps with Fourier-domain texture descriptors","Directionality in cepstral GLCM strongly associates with melanoma"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1434,"prompt_tokens":628,"completion_tokens":806,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":372,"completion_tokens_details":{"reasoning_tokens":719}},"tokens_in":372,"tokens_out":806,"duration_ms":9875,"temperature":1.0,"reasoning_tokens":719,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:20:11.820083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ISIC 2019 melanoma/nevus images through the same greedy selection but score the final 50- and 100-feature models on the held-out unique-lesion test set, comparing test AUC against the handcrafted-only baseline; if the cepstral-augmented model does not beat the baseline on that untouched partition, the complementary-information claim fails. A second check: recompute cepstral features on a sample of lesions using manually corrected masks and compare predictions, isolating how much segmentation error inflates the reported scores.","supporting_citations":[{"cited_title":"The Quefrency Alanysis [sic] of Time Series for Echoes: Cepstrum, Pseudo Autocovariance, Cross-Cepstrum and Saphe Cracking,","cited_arxiv_id":null,"evidence_quote":"Introduces the cepstrum and the property that time delays become impulses, the mathematical basis for the whole feature family."},{"cited_title":"Nonlinear filtering of multiplied and convolved signals,","cited_arxiv_id":null,"evidence_quote":"Extends cepstral analysis to images, supplying the 2D cepstrum definition used for texture extraction."},{"cited_title":"Novel Mixed Domain Hand-Crafted Features for Skin Disease Recognition Using Multiheaded CNN,","cited_arxiv_id":null,"evidence_quote":"Recent work feeding cepstra to CNNs; establishes the gap that interpretable cepstrum-derived features are unexplored."},{"cited_title":"Textural Features for Image Classification,","cited_arxiv_id":null,"evidence_quote":"Defines the 13 GLCM texture statistics that the paper computes on the cepstrum."},{"cited_title":"Deep learning and handcrafted method fusion: Higher diagnostic accuracy for melanoma dermoscopy images,","cited_arxiv_id":null,"evidence_quote":"Source of the mixed-domain handcrafted feature fusion approach that forms the comparison baseline."},{"cited_title":"Deep Learning Techniques for Image Segmentation in Dermoscopic Skin Cancer Images,","cited_arxiv_id":null,"evidence_quote":"Supplies the segmentation algorithm whose masks isolate the lesion before cepstral analysis; mask accuracy is load-bearing."},{"cited_title":"Mahotas: Open source software for scriptable computer vision,","cited_arxiv_id":null,"evidence_quote":"The mahotas library that computes the Haralick features; the feature pipeline depends on it."},{"cited_title":"Implication and Applications of Machine Learning on Biomedical Images,","cited_arxiv_id":null,"evidence_quote":"Collection of the handcrafted dermoscopic feature groups (APN, pink, saliency, telangiectasia, semitranslucent) that cepstral features are fused with and compared against."}],"review_version":1}