Pith. sign in

REVIEW 5 major objections 4 minor 17 references

Cepstrum-Based Texture Features for Melanoma Detection

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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

desk verdict Novel GLCM-on-cepstrum feature recipe with a useful ablation, but the headline AUC is selected on validation and no test AUC is reported. read the letter →

arxiv 2509.00669 v1 pith:SM6MIXJ2 submitted 2025-08-31 eess.IV

Keith Miller , Tristan Crawford , Jason Hagerty , William Stoecker , Ronald J. Stanley This is my paper · ORCID
classification eess.IV
keywords cepstrummelanomadetectionGLCMHaralicktexturefeaturesdermoscopyISIC2019XGBoostfrequency-domainanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

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.

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 (5)
  1. [IV, Fig. 7] 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.
  2. [Table 2, Section IV] 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.
  3. [IV, Table 2] 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.
  4. [II.A, Fig. 3] 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.
  5. [II.A-C] 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.
minor comments (4)
  1. [II.B] 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.
  2. [IV, Fig. 8] 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.
  3. [References] 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.
  4. [II.A, Fig. 3] 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.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline AUC 0.9697 is the validation-set maximum of the greedy feature selector, not an independent test measurement; the central performance claim is a selection objective reported as a result.

  1. fitted input called prediction [Section IV (Experimental Setup and Results), greedy feature selection algorithm and 'highest-performing model' sentence; Fig. 7]
    "b. Train an XGBoost classifier on the selected features and measure ROC AUC on a validation set. c. Select the feature that yields the greatest AUC improvement and add it to the selected set. ... Using the previously outlined greedy feature selection process significantly improved the XGBoost model's ROC AUC, with the highest-performing model achieving an AUC of 0.9697, utilizing fewer than 100 total features selected from all feature sets."

    The 0.9697 value is the maximum of the ROC-AUC curve computed on the validation set during greedy forward selection (Fig. 7: 'ROC AUC per iteration of the greedy feature selection algorithm'). The feature subset itself was chosen by maximizing this same validation ROC AUC at each step. Reporting the resulting maximum as the 'highest-performing model' is therefore reporting the selection objective, not an independent measure of generalization. No test-set AUC for the selected model is given anywhere in the paper, even though an 80/20 unique-lesion test split is described. Thus the headline improvement attributed to cepstral features is, by construction, the optimized criterion value; the claim that cepstral features 'offer complementary discriminatory information' rests on a number that was

full rationale

The cepstral feature definitions are fixed, label-independent functions of image pixels, so the feature derivation itself is not circular. The paper's use of the authors' own prior handcrafted feature families as baselines (refs [5], [12]–[17]) is not load-bearing circularity because these are external published comparators, not constraints that force the cepstral result. The substantive circularity is in the performance reporting: the headline AUC 0.9697 is the validation-set maximum of a greedy feature selector that explicitly optimizes ROC AUC on that validation set. The paper describes a unique-lesion 80/20 test split but never reports a held-out test AUC for the selected model, so the 'highest-performing model' number is the selection objective rather than a prediction. This matches the fitted-input-called-prediction pattern: the selector is fitted to the validation AUC and the same validation AUC is then presented as the result. The paper also concedes that when all feature sets were combined, adding cepstral features did not improve performance, which further tempers the complementary-information claim but is a limitation rather than a circular step. Overall, one central performance claim reduces by construction, yielding a score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard cepstral mathematics, on the quality of the segmentation and labels from ISIC 2019, and on several unspecified implementation choices (GLCM bins, cepstrum window, XGBoost settings). No new physical entities are introduced. The main fitting-to-data occurs in the greedy feature selection, which is tuned to validation AUC.

free parameters (4)
  • GLCM gray-level quantization (bins)
    The number of bins used to quantize the normalized cepstrum before computing GLCM is not specified, and it directly changes all Haralick statistics.
  • Greedy feature-selected subset = fewer than 100 features
    The subset of cepstral and handcrafted features used in the final model is selected on validation ROC AUC, i.e., fitted to the evaluation metric.
  • XGBoost hyperparameters
    Tree count, depth, learning rate, subsampling and seed are not reported, and they materially affect AUC and F1.
  • Cepstrum size/window and padding
    Image resizing and DFT windowing before the 2D cepstrum are not specified; these determine which spatial frequencies are represented.
assumptions (5)
  • standard math Fourier/log/inverse-Fourier cepstrum properties (Eq. 1): time delays become impulses.
    Standard DSP result used to justify that cepstra capture periodic patterns; not in dispute.
  • domain assumption The Norsang Lama segmentation algorithm yields correct lesion masks on ISIC 2019 images.
    Section II.A: masks are used to isolate the lesion before cepstral extraction; incorrect masks would mix background texture into the cepstral features.
  • domain assumption The ISIC 2019 melanoma/nevus labels and the lesion-ID deduplication correctly define the binary task.
    Section IV: class selection and duplicate handling assume the metadata is accurate and that matching lesion IDs fully prevents leakage.
  • ad hoc to paper GLCM statistics computed on a normalized unsigned cepstrum are meaningful texture descriptors (distance 1, four directions).
    Section II.C: the choice of distance 1 and the directionality max/mean quotient are introduced without validation against other distances or prior benchmarks.
  • domain assumption Greedy forward selection on validation ROC AUC produces a feature set that generalizes to an unseen test set.
    Section IV: the procedure can overfit the validation set; the paper does not report a separate test AUC for the final selected model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cepstrum-Based Texture Features for Melanoma Detection." pith.science (2026). https://pith.science/paper/SM6MIXJ2

@misc{pith2026250900669,
  author       = {Pith},
  title        = {Pith review of: Cepstrum-Based Texture Features for Melanoma Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SM6MIXJ2}},
  note         = {Machine review of arXiv:2509.00669}
}
read the original abstract

This paper introduces a set of cepstrum-based texture features for melanoma classification and validates their performance on dermoscopic images from the ISIC 2019 dataset. We propose applying gray-level co-occurrence matrix (GLCM) statistics to 2D cepstral representations, a novel approach in image analysis. Combined with established handcrafted lesion descriptors, these features were evaluated using XGBoost models. Incorporating select cepstral features improved the area under the receiver operating characteristic curve, accuracy, and F1 score for binary melanoma vs. nevus classification. Results suggest that cepstral GLCM features offer complementary discriminatory information for melanoma detection.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

  1. [1]

    The Quefrency Alanysis [sic] of Time Series for Echoes: Cepstrum, Pseudo Autocovariance, Cross-Cepstrum and Saphe Cracking,

    B. P. Bogert, M. J. Healy, J. W. Tukey, “The Quefrency Alanysis [sic] of Time Series for Echoes: Cepstrum, Pseudo Autocovariance, Cross-Cepstrum and Saphe Cracking,” Proceedings of the Symposium on Time Series Analysis, pp. 209-243, 1963

    work page 1963

  2. [2]

    Nonlinear filtering of multiplied and convolved signals,

    A. V. Oppenheim, R. W. Schafer, T. G. Stockham, “Nonlinear filtering of multiplied and convolved signals,” Proceedings of the IEEE, vol. 56, no. 8, pp. 1264-1291, Aug. 1968

    work page 1968

  3. [3]

    Novel Mixed Domain Hand-Crafted Features for Skin Disease Recognition Using Multiheaded CNN,

    A. Kumar, A. Vishwakarma, V. Bajaj, S. Mishra, “Novel Mixed Domain Hand-Crafted Features for Skin Disease Recognition Using Multiheaded CNN,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1-13, 2024

    work page 2024

  4. [4]

    Textural Features for Image Classification,

    R. M. Haralick, K. Shanmugam, I. Dinstein, “Textural Features for Image Classification,” IEEE Transactions on Systems, Man, and Cybernetics, vol. SMC-3, no. 6, pp. 610-621, Nov. 1973

    work page 1973

  5. [5]

    Deep learning and handcrafted method fusion: Higher diagnostic accuracy for melanoma dermoscopy images,

    J. R. Hagerty, R. J. Stanley, H. Almubarak, N. Lama, R. Kasmi, P. Guo, R. J. Drugge, H. S. Rabinovitz, M. Oliviero, W. V. Stoecker, “Deep learning and handcrafted method fusion: Higher diagnostic accuracy for melanoma dermoscopy images,” IEEE Journal of Biomedical and Health Informatics, vol. 23, no. 4, pp. 1385-1391, 2019

    work page 2019

  6. [6]

    The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions,

    P. Tschandl, C. Rosendahl, H. Kittler, “The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions,” Scientific Data, vol. 5, p. 180161, 2018

    work page 2018

  7. [7]

    N. C. Codella, D. Gutman, M. E. Celebi, B. Helba, M. A. Marchetti, S. W. Dusza, A. Kalloo, K. Liopyris, N. Mishra, H. Kittler, A. Halpern, “Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin 8 Imaging Collaboration (ISIC),” arXiv preprint arXiv:1710...

    arXiv 2017

  8. [8]

    BCN20000: Dermoscopic Lesions in the Wild,

    M. Combalia, N. C. Codella, V. Rotemberg, B. Helba, V. Vilaplana, O. Reiter, A. C. Halpern, S. Puig, J. Malvehy, “BCN20000: Dermoscopic Lesions in the Wild,” arXiv preprint arXiv:1908.02288, 2019

    arXiv 1908

Show all 17 references
  1. [9]

    Deep Learning Techniques for Image Segmentation in Dermoscopic Skin Cancer Images,

    N. Lama, “Deep Learning Techniques for Image Segmentation in Dermoscopic Skin Cancer Images,” Doctoral Dissertations, Missouri University of Science and Technology, 2024

    2024

  2. [10]

    Grey Level Co-Occurrence Matrices: Generalisation and Some New Features,

    B. Vadakkenveettil, “Grey Level Co-Occurrence Matrices: Generalisation and Some New Features,” International Journal of Computer Science, Engineering and Information Technology, vol. 2, pp. 151-157, 2012

    2012

  3. [11]

    Mahotas: Open source software for scriptable computer vision,

    L. P. Coelho, “Mahotas: Open source software for scriptable computer vision,” Journal of Open Research Software, vol. 1, no. 1, p. e16, 2013

    2013

  4. [12]

    Implication and Applications of Machine Learning on Biomedical Images,

    J. Hagerty, “Implication and Applications of Machine Learning on Biomedical Images,” Doctoral Dissertations, Missouri University of Science and Technology, 2025

    2025

  5. [13]

    The Median Split Algorithm for Detection of Critical Melanoma Color Features,

    K. V. Ghantasala, R. J. Stanley, W. V. Stoecker, “The Median Split Algorithm for Detection of Critical Melanoma Color Features,” VISAPP 2013 - Proceedings of the International Conference on Computer Vision Theory and Applications, vol. 1, pp. 492-495, 2013

    2013

  6. [14]

    Automated classification of malignant melanoma based on detection of atypical pigment network in dermoscopy images of skin lesions,

    N. K. Mishra, “Automated classification of malignant melanoma based on detection of atypical pigment network in dermoscopy images of skin lesions,” Doctoral Dissertations, Missouri University of Science and Technology, 2014

    2014

  7. [15]

    Real-time supervised detection of pink areas in dermoscopic images of melanoma: importance of color shades, texture and location,

    R. Kaur, P. P. Albano, J. G. Cole, J. Hagerty, R. W. LeAnder, R. H. Moss, W. V. Stoecker, “Real-time supervised detection of pink areas in dermoscopic images of melanoma: importance of color shades, texture and location,” Skin Research and Technology, vol. 21, no. 4, pp. 466-473, 2015

    2015

  8. [16]

    A deep learning approach to detect blood vessels in basal cell carcinoma,

    A. Maurya, R. J. Stanley, N. Lama, S. Jagannathan, D. Saeed, S. Swinfard, J. R. Hagerty, W. V. Stoecker, “A deep learning approach to detect blood vessels in basal cell carcinoma,” Skin Research and Technology, vol. 28, no. 4, pp. 571-576, Jul. 2022

    2022

  9. [17]

    Detection of basal cell carcinoma using color and histogram measures of semitranslucent areas,

    W. V. Stoecker, K. Gupta, B. Shrestha, M. Wronkiewiecz, R. Chowdhury, R. J. Stanley, J. Xu, R. H. Moss, M. E. Celebi, H. S. Rabinovitz, M. Oliviero, J. M. Malters, I. Kolm., “Detection of basal cell carcinoma using color and histogram measures of semitranslucent areas,” Skin R...

    2009

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.