Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Bluish Veil Detection and Lesion Classification using Custom Deep Learnable Layers with Explainable Artificial Intelligence (XAI)

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A deep network with learnable PReLU activation detects the melanoma blue-white veil at 85–95% accuracy on unfamiliar dermoscopic datasets, beating three earlier methods.

desk verdict A plausible cross-dataset BWV detector with a good evaluation design, but the abstract swaps datasets, the custom layer is just PReLU, and the ISIC comparisons have a circularity problem that needs fixing. read the letter →

arxiv 2507.07453 v1 pith:RMOKIBMW submitted 2025-07-10 cs.CV cs.AI

M. A. Rasel , Sameem Abdul Kareem , Zhenli Kwan , Shin Shen Yong , Unaizah Obaidellah This is my paper
classification cs.CVcs.AI
keywords melanomablue-whiteveildermoscopicimagedeepconvolutionalneuralnetworkPReLULIMEskinlesionclassificationcolorthresholdannotation
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

The paper sets out to show that a deep convolutional neural network can classify dermoscopic images by the presence or absence of the blue-white veil (BWV), the bluish, ground-glass haze that is a recognized warning sign for melanoma. Because one large public skin-lesion archive lacked BWV annotations, the authors first build an automatic labeling rule: cut each image into 16x16 patches, flag any patch whose RGB values fall in a BWV color range derived from 80 published color palettes, and label the whole image positive if it contains at least one such patch. Using those labels, they train a 31-layer network with a custom learnable activation (PReLU) on four dataset splits and report held-out test accuracies of 85.71% to 95.05%, consistently above three earlier BWV detection methods. They then apply LIME to show that the network focuses on the bluish and whitish parts of lesions when it predicts BWV. If the results hold, the method offers a way to label unannotated skin-lesion archives and build an automated, interpretable screening aid for a melanoma feature that is easy to miss.

What carries the argument

The load-bearing mechanism is a custom PReLU layer inserted after each convolution-and-normalization block of a 31-layer network: instead of a fixed activation, each channel learns a coefficient $\alpha_i$ that scales its negative inputs during training, so the network can adapt the nonlinearity to BWV-specific color and texture statistics. The second essential piece is the annotation pipeline: 80 color palettes, taken from earlier BWV work, are reduced to an RGB window (red 45–166, green and blue 73–98), and any $16\times16$ patch whose pixels fall inside that window marks the image as BWV-positive. LIME then segments each test image into superpixels, perturbs them, and fits a simple surrogate model to map which regions drive the network's prediction.

What would settle it

Have independent dermatologists provide patch-level BWV labels for the same 204 ISIC images used in the paper, then retrain and retest the model on those labels; if accuracy against clinician labels falls materially below the 99.51% agreement the paper reports against its own threshold, the reported accuracies partly measure agreement with a color rule rather than clinical BWV.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a purpose-built DCNN with Parametric Rectified Linear Unit (PReLU) activation layers detects the blue-white veil in dermoscopic images more accurately than existing threshold-based and machine-learning detectors, including on datasets the model has never seen during training. The reported evidence is four cross-dataset experiments: training on augmented PH2 and testing on ISIC gives 85.71% accuracy; training on augmented ISIC and testing on PH2 gives 95.00%; training on PH2 plus ISIC and testing on Derm7pt gives 95.05%; and training on Derm7pt and testing on PH2 plus ISIC gives 90.00%. Across those four test sets the PReLU model averages 91.44% accuracy and 86.49% AUC, and its LIME explanations point at the bluish, whitish regions of lesions as the features driving BWV decisions. The same experiments run with Leaky ReLU are lower but still above the three comparison methods, which the paper attributes to PReLU's per-channel learnable negative slope.

Load-bearing premise

The load-bearing premise is that the RGB threshold rule (a $16\times16$ patch with red 45–166 and green/blue 73–98 labeled as BWV, with one such patch labeling the whole image) is accurate enough to stand in for a clinician's judgment when generating the ISIC training and test labels.

Editorial extensions

If this is right

  • Unannotated dermoscopic archives can be converted into BWV training sets without manual pixel-level labeling, so the approach scales to larger collections.
  • Using a learnable activation in place of a fixed ReLU gives a consistent accuracy gain across all four cross-dataset experiments, suggesting the channel-adaptive nonlinearity is doing real work for this task.
  • Because the model is tested on data from different sources than its training data, the reported cross-dataset numbers are evidence the learned BWV representation transfers rather than memorizes one dataset.
  • LIME maps on held-out images show the network's BWV decisions concentrate on bluish and whitish lesion regions, which is the clinically expected behavior and supports use as an explainable screening tool.

Reading between the lines

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

  • Whether the accuracy numbers reflect clinically meaningful BWV depends on the threshold labels being correct; a natural extension is to rerun the identical pipeline with dermatologist-provided patch labels on ISIC, which the paper has not done.
  • The same patch-threshold recipe could be extended to other Seven-Point Checklist structures, such as regression structures or atypical vessels; nothing in the method ties it exclusively to BWV colors.
  • Because only 66 of 204 ISIC images carry BWV labels, the reported percentages are estimates with wide confidence intervals; a larger clinician-annotated test set would tighten them.
  • Since the model was trained on images that already contain near-BWV color distributions, its performance on very dark or heavily pigmented lesions, where the veil is hardest to see, remains an open question.
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

4 major / 5 minor

Summary. The paper proposes a pipeline for detecting the blue-white veil (BWV) in dermoscopic images. It first converts previously unlabeled ISIC archive images into labeled data by testing each image's 16x16 patches for pixel RGB values lying in a rectangular box derived from 80 Madooei-Drew palette colors; any qualifying patch makes the image BWV-positive. A 31-layer DCNN with a PReLU activation in custom layers is then trained on augmented PH2, augmented ISIC, and Derm7pt data, separately and in combination. Cross-dataset testing (e.g., trained on augmented PH2, tested on 10% of ISIC) yields reported accuracies of 85.71% to 95.05%, which the authors compare with three prior BWV-detection methods (Celebi et al. 2008; Madooei and Drew 2013; Madooei et al. 2019). LIME is used to visualize which image regions drive the model's decisions.

Significance. The cross-dataset evaluation design is a genuine strength: training and test sets come from different sources, which reduces the risk of simply memorizing a single dataset's artifacts. The comparison against three published baselines on the same test sets is also useful, and the authors make a reasonable effort to involve clinical input when constructing the ISIC labels. If the reported results were statistically robust and the labels were clinically meaningful, the paper would be a practical contribution to automated BWV detection. However, the significance is curtailed by two problems: the ISIC ground-truth rule shares its origin with one of the baselines, creating a circular comparison; and the test sets are too small, with no confidence intervals or repeated runs, to support the claimed superiority.

major comments (4)
  1. [Section 4.2 and Table 8] The ISIC ground-truth labels are generated by a color-threshold rule built from the palettes of Madooei and Drew (2013), and that same publication is one of the baselines in Table 8. Consequently, the comparison on the ISIC test portions (Expt. 1 and the ISIC portion of Expt. 4) is partially circular: the DCNN is trained to approximate a rule derived from the Madooei-Drew palette, and then the two are compared on outputs of that same rule. The reported superiority over Madooei and Drew (2013) on those ISIC tests is therefore expected and does not, by itself, demonstrate clinical superiority. The PH2 and Derm7pt test sets are less affected because those labels come from expert annotation. To support the claim, the authors should re-evaluate the ISIC test sets using the independent dermatologist labels described in Table 4, or explicitly reframe the ISIC results as measuring agreement with the automated labeling rule rather than with clinical BWV.
  2. [Section 3.7 and Table 7/Table 8] The AUC values in Table 8 are not computable from the information provided. A receiver operating characteristic (ROC) curve requires the model's continuous scores or a range of classification thresholds; a single confusion matrix gives only one (false-positive rate, true-positive rate) operating point. The text states that 'those confusion matrices are used' to draw the ROC curve, which is mathematically impossible. The AUC columns therefore appear to be arbitrary conversions of accuracy or another scalar metric and cannot be used to rank methods. The authors should remove the AUC columns or recompute them from per-image predicted probabilities using a proper ROC analysis, and if test sets remain small, report bootstrapped confidence intervals for the AUC.
  3. [Section 3.3] The reported test sets are very small: Expt. 1 uses 10% of 204 ISIC images (about 21 test images), Expt. 2 uses 10% of 200 PH2 images (about 20), Expt. 4 uses 10% of 404 PH2+ISIC images (about 40), and Expt. 3 uses 5% of 2013 Derm7pt images (about 101). With n≈20, an accuracy difference of 10–20 percentage points is easily within sampling error. The paper reports a single split with no confidence intervals, no repeated runs, and no significance tests. The 'Number of folds = 5' column in Table 7 does not appear to produce independent test estimates for the reported numbers. To support the conclusions, the authors should repeat the cross-dataset experiments over multiple random splits (or use leave-one-out where feasible) and report means with standard deviations or exact binomial confidence intervals, and ideally a paired test when comparing models on the same test sets.
  4. [Section 3.3] The automated labeling rule is quite permissive: a single 16x16 patch containing any pixel whose RGB channels all fall in the ranges R∈[45,166], G∈[73,98], B∈[73,98] is enough to label the entire image as BWV-positive. This could be triggered by shadows, ink marks, compression artifacts, or other incidental blue-grey regions that are not clinically BWV. The 99.51% agreement with clinician opinion in Table 4 is helpful, but the protocol says discrepancies were 'discussed and resolved', so the clinician labels may not be fully independent of the algorithm. Moreover, Table 4 reports only image-level agreement, not patch-level agreement, which is the granularity at which the rule operates. The authors should report patch-level agreement, describe the resolution process more transparently, and perform a sensitivity analysis (e.g., varying the RGB ranges or requiring at least two patches) to show that the reported accuracies are not an artifact of one particular threshold.
minor comments (5)
  1. [Table 8] The abstract misreports which test set produces each accuracy: 85.71% is on the ISIC test set after training on augmented PH2, 95.00% is on the PH2 test set after training on augmented ISIC, 95.05% is on Derm7pt after training on PH2+ISIC, and 90.00% is on PH2+ISIC after training on Derm7pt. The current wording attributes these numbers to the training datasets instead of the test sets.
  2. [Section 3.6] The 'Dataset' column in Table 8 lists training datasets (AuP, AuIa, etc.), while the remainder of the table describes performance on the testing datasets (ISIC 10%, PH2 10%, etc.). This is confusing and should be relabeled, for example 'Training set (test set)'.
  3. [Section 4.3] The paper repeatedly calls PReLU a 'custom deep learning layer', but PReLU is a standard activation function with learnable parameters introduced by He et al. The novelty lies in its integration and experimental comparison, not in the layer itself; the wording should be adjusted to avoid overstating the contribution.
  4. [Table 6] The LIME visualizations in Figures 10–14 are qualitative. There is no quantitative assessment of explanation quality, such as whether the highlighted regions correspond to dermatologist-annotated BWV areas. Adding a localization metric or at least a few more examples with expert labeling would strengthen the interpretability claim.
  5. [Section 3.3/Table 4] In Table 6, layer 28 (Custom) has missing entries for dilation factor, padding, and stride. The table should be completed or the missing values explained as 'not applicable'.

Circularity Check

1 steps flagged · score 4.0 of 10

ISIC ground truth is built from Madooei and Drew's color palettes, and that same work is one of the baselines compared on ISIC-containing test sets, so part of the reported advantage measures agreement with the palette rule rather than independent clinical BWV.

  1. fitted input called prediction [Section 3.3 (Establishing Ground Truth of BWV on ISIC Archive), Table 2; Table 7 Expt. 1; Table 8 Expt. 1]
    "To establish ground truth, 80 previously published color palettes are utilized (Madooei and Drew, 2013) ... These palettes informed the determination of pixel values for each color ... the minimum and maximum values for red, green, and blue channels are identified, which are respectively 45, 73, 73 (minimum) and 166, 98, 98 (maximum), constituting the considered color range for BWV. ... Employing this technique, ground truth is established for 204 dermoscopic images within the ISIC archive dataset, comprising 66 BWV and 138 non-BWV lesions."

    The ISIC archive labels are generated by an RGB-threshold rule whose bounds are taken from Madooei and Drew's 80 BWV color palettes. These labels are then used as the test reference for the ISIC archive portion of the experiments, including Expt 1's test set of 10% of the ISIC archive. In Table 8, 'Madooei and Drew (2013)' is one of the three baseline approaches 'applied to the same datasets in four experiments.' Evaluating Madooei-Drew against labels constructed from Madooei-Drew's own published palette is not an independent comparison: the reference standard and the baseline share the same color-rule source. The DCNN's predictions on ISIC are therefore partly scored for agreement with a rule derived from the competing method rather than with an independent clinical standard.

full rationale

The main derivation chain is not circular in its architecture: the proposed DCNN with custom PReLU layers is trained and tested on held-out images, and the PH2 and Derm7pt evaluations use expert-derived labels, giving independent content. The significant circular component is confined to the ISIC archive labeling pipeline. Section 3.3 defines BWV-positive as containing a 16x16 patch whose RGB values fall in a box derived from the 80 Madooei and Drew palette colors, and then uses that rule to establish ground truth for the 204 ISIC archive images. Table 8 then compares the proposed DCNN against Madooei and Drew (2013) on test sets that include these same ISIC archive labels, so the ISIC-based accuracy and the reported margin over that baseline partly measure agreement with a rule whose source is the baseline method itself. This is not a fully forced circularity because the final labels also underwent clinician adjudication (Table 4 reports 99.51% agreement with clinician opinion), and because independent expert-labeled test sets in PH2 and Derm7pt support the general claim. No load-bearing self-citation chain was found: the Rasel et al. (2022) citation motivates architecture choices without being used to forbid alternatives. Overall score 4 reflects partial, not total, circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on algorithmically generated ISIC labels, the transferability of RGB thresholds across three datasets, and a large unreleased set of trained network parameters. The clinical validation in Table 4 provides independent support for the ISIC labels, which lowers the circularity burden, but the thresholds themselves are inherited from the same published palette family used by one baseline.

free parameters (3)
  • RGB threshold range for BWV labeling = R=45-166, G=73-98, B=73-98
    Derived from 80 published color palettes in Section 3.3 and used to create ISIC ground truth; if this range is wrong or dataset-specific, all ISIC-based metrics shift.
  • DCNN weights and PReLU coefficients = not released
    Millions of parameters estimated by SGDM on augmented datasets in Section 3.7; the classifier behavior and reported accuracies depend on them and they are not provided.
  • Training hyperparameters and architecture dimensions = learning rate 0.01, 250 epochs, SGDM, 31 layers
    Selected by ablation and fine-tuning in Sections 3.5 and 3.7; no ablation results or sensitivity analysis are reported, so performance is conditional on these choices.
assumptions (5)
  • domain assumption The 80 color palettes from Madooei and Drew (2013) capture the clinically relevant range of BWV colors across ISIC images.
    Section 3.3 uses these palettes to set RGB thresholds; the ISIC labels and test labels inherit this assumption.
  • domain assumption A single 16x16 patch within the RGB range is sufficient to label an entire image as BWV-positive.
    Section 3.3 and Table 3 define the patch rule; the clinical validation in Table 4 is the only check on this threshold-based labeling decision.
  • domain assumption Dermoscopic images across PH2, ISIC, and Derm7pt are captured under constant or normalized illumination, so raw RGB thresholds transfer across datasets without color calibration.
    Section 3.3 states ISIC images were captured under constant conditions, but no explicit color normalization or cross-dataset calibration is described for PH2 and Derm7pt.
  • domain assumption Clinicians' BWV opinions on the 204 ISIC images are a valid external ground truth.
    Table 4 relies on clinical experts, but no inter-rater agreement statistic is provided, and the test splits used in Table 8 are not independently re-annotated.
  • standard math PReLU as defined in Eq. (7) with channel-wise learnable slopes is a valid differentiable activation optimized by standard CNN backpropagation.
    Sections 3.6 and 3.7 assume the usual training dynamics; no formal proof or formal verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bluish Veil Detection and Lesion Classification using Custom Deep Learnable Layers with Explainable Artificial Intelligence (XAI)." pith.science (2026). https://pith.science/paper/RMOKIBMW

@misc{pith2026250707453,
  author       = {Pith},
  title        = {Pith review of: Bluish Veil Detection and Lesion Classification using Custom Deep Learnable Layers with Explainable Artificial Intelligence (XAI)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMOKIBMW}},
  note         = {Machine review of arXiv:2507.07453}
}
read the original abstract

Melanoma, one of the deadliest types of skin cancer, accounts for thousands of fatalities globally. The bluish, blue-whitish, or blue-white veil (BWV) is a critical feature for diagnosing melanoma, yet research into detecting BWV in dermatological images is limited. This study utilizes a non-annotated skin lesion dataset, which is converted into an annotated dataset using a proposed imaging algorithm based on color threshold techniques on lesion patches and color palettes. A Deep Convolutional Neural Network (DCNN) is designed and trained separately on three individual and combined dermoscopic datasets, using custom layers instead of standard activation function layers. The model is developed to categorize skin lesions based on the presence of BWV. The proposed DCNN demonstrates superior performance compared to conventional BWV detection models across different datasets. The model achieves a testing accuracy of 85.71% on the augmented PH2 dataset, 95.00% on the augmented ISIC archive dataset, 95.05% on the combined augmented (PH2+ISIC archive) dataset, and 90.00% on the Derm7pt dataset. An explainable artificial intelligence (XAI) algorithm is subsequently applied to interpret the DCNN's decision-making process regarding BWV detection. The proposed approach, coupled with XAI, significantly improves the detection of BWV in skin lesions, outperforming existing models and providing a robust tool for early melanoma diagnosis.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 25 canonical work pages

  1. [4]

    https://doi.org/10.2174/1874372201004010110 Garrison, Z.R., Hall, C.M., Fey, R.M., Clister, T., Khan, N., Nichols, R., Kulkarni, R.P.,

    work page doi:10.2174/1874372201004010110

  2. [5]

    https://doi.org/10.1038/sdata.2018.161 Tsuneki, M.,

    doi:10.1038/sdata.2018.161 2018

  3. [10]

    V., Moss, R.H., Rabinovitz, H.S., Argenziano, G., Soyer, H.P.,

    https://doi.org/10.3390/app10124351 Celebi, M.E., Iyatomi, H., Stoecker, W. V., Moss, R.H., Rabinovitz, H.S., Argenziano, G., Soyer, H.P.,

    work page doi:10.3390/app10124351

  4. [12]

    https://doi.org/10.1109/SSD.2009.4956770 Fabbrocini, G., Betta, G., Di Leo, G., Liguori, C., Paolillo, A., Pietrosanto, A., Sommella, P., Rescigno, O., Cacciapuoti, S ., Pastore, F., De Vita, V., Mordente, I., Ayala, F.,

    work page arXiv 2009

  5. [15]

    https://doi.org/10.3390/technologies12020015 Maniatopoulos, A., Mitianoudis, N.,

    work page doi:10.3390/technologies12020015

  6. [17]

    https://doi.org/10.1007/978-3-319-46418-3_37 Landa, E., Fairchild, M.,

    Blue whitish veil, atypical vascular pattern and regress ion structures detection in skin lesions images, in: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). https://doi.org/10.1007/978-3-319-46418-3_37 Landa, E., Fairchild, M.,

    work page doi:10.1007/978-3-319-46418-3_37

  7. [21]

    https://doi.org/10.1109/JBHI.2018.2835405 Manakitsa, N., Maraslidis, G.S., Moysis, L., Fragulis, G.F.,

    arXiv 2018

  8. [23]

    https://doi.org/10.1109/JBHI.2018.2824327 Kropidlowski, K., Kociolek, M., Strzelecki, M., Czubinski, D.,

    arXiv 2018

Show all 36 references
  1. [25]

    https://doi.org/10.1109/EMBC.2013.6610779 Olayah, F., Senan, E.M., Ahmed, I.A., Awaji, B.,

    PH2 - A dermoscopic image database for research and benchmarking, in: Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS. https://doi.org/10.1109/EMBC.2013.6610779 Olayah, F., Senan, E.M., Ahmed, I.A., Awaji, B.,

    2013

  2. [27]

    https://doi.org/10.1109/ACCESS.2022.3196911 Ribeiro, M.T., Singh, S., Guestrin, C.,

    2022

  3. [28]

    Why should I trust you?

    “Why should I trust you?” Explaining the predictions of any classifier, in: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. https://doi.org/10.1145/2939672.2939778 Seidenari, S., Longo, C., Giusti, F., Pellacani, G.,

  4. [29]

    https://doi.org/10.1046/j.1524-4725.2003.29260.x Di Leo, G., Fabbrocini, G., Paolillo, A., Rescigno, O., Sommella, P.,

    2003

  5. [32]

    https://doi.org/10.1016/j.compmedimag.2008.08.003 Celebi, M.E., Kingravi, H.A., Aslandogan, Y.A., Stoecker, W. V.,

    2008 doi

  6. [33]

    Journal of Oral Biosciences 64, 312 –320

    Deep learning models in medical image analysis. Journal of Oral Biosciences 64, 312 –320. https://doi.org/10.1016/j.job.2022.03.003 Wadhawan, T., Hu, R., Zouridakis, G.,

    2022 doi

  7. [46]

    https://doi.org/10.1364/ol.422930 De Giorgi, V., Massi, D., Trez, E., Salvini, C., Quercioli, E., Carli, P.,

    doi

  8. [63]

    https://doi.org/10.3399/bjgp13X667213 Wróblewska-Łuczka, P., Cabaj, J., Bargieł, J., Łuszczki, J.J.,

    doi

  9. [86]

    https://doi.org/10.1109/5.726791 Madooei, A., Drew, M.S.,

    doi

  10. [93]

    https://doi.org/10.1511/2005.55.975 LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.,

    2005 doi

  11. [105]

    https://doi.org/10.1016/j.adengl.2014.07.015 Codella, N. C. F., Gutman, D., Celebi, M. E., Helba, B., Marchetti, M. A., Dusza, S. W., Kalloo, A., Liopyris, K., Mishra, N. , Kittler, H., Halpern, A.,

    2014 doi

  12. [154]

    https://doi.org/10.1111/j.1365-2133.2006.07165.x Soyer, H.P., Argenziano, G., Zalaudek, I., Corona, R., Sera, F., Talamini, R., Barbato, F., Baroni, A., Cicale , L., Di Stefani, A., Farro, P., Rossiello, L., Ruocco, E., Chimenti, S.,

    2006

  13. [208]

    https://doi.org/10.1159/000075042 Switzer, B., Puzanov, I., Skitzki, J.J., Hamad, L., Ernstoff, M.S.,

    doi

  14. [513]

    Final published version: https://doi.org/10.1016/j.compbiomed.2024.108758 ©

    https://doi.org/10.3390/info12120513Henning, J.S., Stein, J.A., Yeung, J., Dusza, S.W., Marghoob, A.A., Accepted Manuscript This is the peer-reviewed version of the article accepted for publication in Computers in Biology and Medicine. Final published version: https://doi.org/...

  15. [974]

    https://doi.org/10.3390/life13040974 Gutman, D., Codella, N. C. F., Celebi, E., Helba, B., Marchetti, M., Mishra, N., Halpern, A.,

    doi

  16. [1314]

    https://doi.org/10.3390/diagnostics13071314 Rasel, M.A., Obaidellah, U.H., Kareem, S.A.,

    doi

  17. [2006]

    https://doi.org/10.1117/12.655779 Ciudad-Blanco, C., Avilés -Izquierdo, J.A., Lázaro -Ochaita, P., Suárez -Fernández, R.,

    Detection of blue -white veil areas in dermoscopy images using machine learning techniques, in: Medical Imaging 2006: Image Processing. https://doi.org/10.1117/12.655779 Ciudad-Blanco, C., Avilés -Izquierdo, J.A., Lázaro -Ochaita, P., Suárez -Fernández, R.,

    2006 doi

  18. [2008]

    Arch Dermatol

    CASH algorithm for dermoscopy revisited. Arch Dermatol. https://doi.org/10.1001/archderm.144.4.554 Mendonca, T., Ferreira, P.M., Marques, J.S., Marcal, A.R.S., Rozeira, J.,

    doi

  19. [2009]

    Towards an automatic diagnosis system for skin le sions: Estimation of blue-whitish veil and regression structures, in: 2009 6th International Multi -Conference on Systems, Signals and Devices, SSD

    2009

  20. [2011]

    https://doi.org/10.1109/ISSPIT.2011.6151559 Cacciapuoti, S., Di Leo, G., Ferro, M., Liguori, C., Masarà, A., Scalvenzi, M., Sommella, P., Fabbrocini, G.,

    2011

  21. [2012]

    https://doi.org/10.1109/BHI.2012.6211628 Walter, F.M., Prevost, A.T., Vasconcelos, J., Hall, P.N., Burrows, N.P., Morris, H.C., Kinmonth, A.L., Emery, J.D.,

    2012

  22. [2013]

    https://doi.org/10.2352/cic.2013.21.1.art00036 Madooei, A., Drew, M.S., Hajimirsadeghi, H.,

    A colour palette for automatic detection of blue -white veil, in: Final Program and Proceedings - IS and T/SID Color Imaging Conference. https://doi.org/10.2352/cic.2013.21.1.art00036 Madooei, A., Drew, M.S., Hajimirsadeghi, H.,

    2013 doi

  23. [2016]

    1605.01397

    Skin lesion analysis towa rd melanoma detection: A challenge at the international symposium on biomedical imaging ( ISBI) 2016, hosted by the international skin imaging collaboration (ISIC). 1605.01397. Kawahara, J., Daneshvar, S., Argenziano, G., Hamarneh, G.,

    2016 arXiv

  24. [2017]

    1710.05006

    Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedi cal imaging (ISBI), hosted by the international skin imaging collaboration (ISIC). 1710.05006. Codella, N., Rotemberg, V., Tschandl, P., Celebi, M. E., Dusza, S., Gutman,...

    2017 arXiv

  25. [2018]

    1902.03368

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (ISIC). 1902.03368. Combalia, M., Codella, N. C. F., Rotemberg, V., Helba, B., Vilaplana, V., Reiter, O., Carrera, C., Barreiro, A., Halpern, A. C., Puig, S....

    2018 arXiv

  26. [2019]

    arXiv:1908.02288

    BCN20000: Dermoscopic lesions in the wild. arXiv:1908.02288. Crnjanski, J., Krstić, M., Totović, A., Pleros, N., Gvozdić , D.,

    1908 arXiv

  27. [2022]

    JCO Oncology Practice 18, 335–351

    Managing Metastatic Melanoma in 2022: A Clinical Re view. JCO Oncology Practice 18, 335–351. https://doi.org/10.1200/op.21.00686 Tschandl, P., Rosendahl, C., Kittler, H.,

    2022 doi

  28. [2023]

    Pharmacological Reports 75, 1115–1125

    Anticancer effect of terpenes: focus on malignant melanoma. Pharmacological Reports 75, 1115–1125. https://doi.org/10.1007/s43440-023-00512-1

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.