Pith. sign in

REVIEW 4 major objections 4 minor 24 references

Asymmetric Lesion Detection with Geometric Patterns and CNN-SVM Classification

T0 review · 4 major / 4 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 This paper claims that a four-quadrant pixel-count rule detects dermatologically asymmetric lesions with 99% agreement on two dermoscopic datasets, and that a ResNet-101 plus SVM classifier reproduces the three shape labels with up to 97%…

desk verdict The ISIC2016 accuracy is circular because the algorithm helped build the ground truth; the PH2 result and the dataset idea are the real contributions. read the letter →

arxiv 2507.17185 v1 pith:DIFYZ363 submitted 2025-07-23 cs.CV cs.AI

M. A. Rasel , Sameem Abdul Kareem , Zhenli Kwan , Nik Aimee Azizah Faheem , Winn Hui Han , Rebecca Kai Jan Choong , Shin Shen Yong , Unaizah Obaidellah This is my paper
classification cs.CVcs.AI
keywords dermoscopicimageanalysismelanomaasymmetrylesionshapeclassificationgeometry-basedCNNfeatureextractionmulticlassSVMPH2datasetISIC2016
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 is trying to establish two things: a simple geometry-based rule can label a skin lesion's shape with nearly the same judgment as dermatologists, and a deep-learning classifier can reproduce those shape labels automatically. The geometry rule, called Geometry Shape-Based Asymmetry Analysis (GSAA), splits the binary lesion mask into four quadrants at its centroid and compares white-pixel counts between opposite quadrants, encoding each ratio as matching (0.9-1.1) or not. The authors report 99.00% agreement on PH2 (198 of 200) and 99.06% on ISIC2016 (1267 of 1279) against clinical labels, and they use the newly labeled ISIC2016 data plus the mirror-augmented PH2 set to train a ResNet feature extractor with a multiclass SVM; the CNN itself is trained on clinical ground truth, not on GSAA output. The best CNN-based result reaches a 94% kappa, 95% macro F1, and 97% weighted F1 on the APH2 test set. If correct, this gives non-experts a teachable, checkable reading of the asymmetry criterion and supplies labeled shape data for future melanoma-screening models.

What carries the argument

The load-bearing object is the Geometry Shape-Based Asymmetry Analysis (GSAA) rule: a binary lesion mask is split into four quadrants by perpendicular axes through the lesion centroid, and the white-pixel counts in the four parts are compared as ratios. The 0.9-to-1.1 acceptance window turns continuous area ratios into binary match or no-match decisions, and the pattern of matches across the four pairwise ratios yields the three shape labels of asymmetric, half-symmetric, and symmetric. The second mechanism is transfer learning: features are taken from the 'fc1000' layer of pretrained ResNet18, ResNet50, or ResNet101 and passed to a multiclass linear SVM trained with a one-versus-one coding design. The GSAA rule does the labeling work, while the ResNet-plus-SVM stack does the automatic classification work.

What would settle it

Take the 1279 ISIC2016 images, have a panel of dermatologists label lesion shape without access to the GSAA output or the pixel-ratio code, then recompute GSAA's agreement and the ResNet101-plus-SVM evaluation against those blind labels; if the kappa drops materially below the reported 98.5% and 94%, the claimed accuracy was partly an artifact of label-generation feedback.

Watch

Extended reading notes

Core claim

The paper's central claim is that lesion-shape asymmetry, the first criterion of the ABCDE melanoma rule, can be measured by a pixel-count ratio rather than by holistic visual judgment. GSAA divides a binary mask into four quadrants A, B, C, and D at the lesion centroid, counts white pixels in each, and forms four ratios ($A_p/B_p$, $A_p/D_p$, $B_p/C_p$, $C_p/D_p$); each ratio in the window 0.9 to 1.1 counts as a match. If all four ratios fail to match, the lesion is asymmetric; one or two matches means half-symmetric; three or more means symmetric. On PH2 this rule reproduces expert labels with 99.00% accuracy (kappa 98.2%), and on the newly expert-labeled ISIC2016 set with 99.06% accuracy (kappa 98.5%). The paper then argues that features extracted from the layer before ResNet101's classification layer, fed to a one-versus-one linear SVM, classify the three shape classes with 94% kappa, 95% macro F1, and 97% weighted F1 on APH2, and that this combination outperforms the decision-tree, histogram, neural-network, and prior CNN baselines included in the comparison.

Load-bearing premise

The reported accuracies assume the ISIC2016 ground truth was set by independent expert judgment, but the paper says the pixel-ratio imaging algorithm was used to augment those manual annotations, so the 99% agreement may partly compare the algorithm with labels it helped produce.

Editorial extensions

If this is right

  • A non-expert can now point to a concrete, repeatable numerical rule for the asymmetry criterion instead of relying only on subjective visual inspection.
  • The labeled ISIC2016 shape classes (667 asymmetric, 344 half-symmetric, 268 symmetric) become a training resource for future dermoscopic shape models.
  • A pretrained ResNet with a linear SVM can classify three shape classes with high agreement using a small 5% test split, suggesting a data-efficient path for shape classification.
  • GSAA can generate ground-truth shape labels for unannotated datasets, reducing the bottleneck of expert annotation when training data are scarce.

Reading between the lines

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

  • If the ISIC2016 labels were co-produced by the same pixel-ratio logic, the 99%-class agreement figures reflect consistency between GSAA and its own label-augmentation step as much as agreement with independent experts; a blind re-annotation study would separate the two.
  • The 0.9-to-1.1 threshold is a free parameter, so sweeping it against blind expert labels could produce a different optimal operating point and the reported rates are not invariant to that choice.
  • The paper's own suggested eight-sector split is a natural stress test: if 45-degree sectors reproduce the three classes with similar agreement, the method generalizes, and if not, the four-quadrant rule may be tuned to the datasets used.
  • Because GSAA operates on binary masks, its ceiling is set by segmentation quality; on imperfect segmentations the pixel-ratio method will inherit the segmenter's errors even if the ratio logic itself is sound.
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 / 4 minor

Summary. The manuscript proposes a two-stage method for classifying dermoscopic lesion shape into three classes (asymmetric, half-symmetric, symmetric). First, a pixel-count-based geometric algorithm (GSAA) divides binary lesion masks into four quadrants at the centroid and thresholds pairwise white-pixel ratios to assign a shape label; it is applied to PH2 and to a newly annotated ISIC2016 symmetry set. Second, features from three pretrained ResNets are fed to a multiclass linear SVM trained on augmented PH2 (APH2) and ISIC2016, with experiments in four train/test configurations. The headline results are 99.00% GSAA accuracy on PH2, 99.06% on ISIC2016, and 94% Kappa / 95% Macro F1 / 97% Weighted F1 for R101+SVM on the APH2 same-dataset test.

Significance. If the reported numbers are valid, the paper would provide a useful explainable support tool for non-experts and the first three-class symmetry annotation for ISIC2016, with a straightforward CNN-SVM pipeline. The PH2 evaluation is a genuine external check for GSAA, and the confusion matrices allow verification of the aggregate metrics. However, the ISIC2016 ground truth may have been influenced by GSAA itself, and the APH2 evaluation suffers from image-level leakage; until these are resolved, the empirical support for the central claims is limited. The paper's strength is its transparency: the GSAA rules are fully specified, the datasets are public, and the CNN-SVM protocol is reproducible.

major comments (4)
  1. [Section 3.1, Table 5] The ISIC2016 ground truth is not demonstrably independent of the GSAA algorithm. Section 3.1 states that "to augment our manual annotations, we developed an imaging algorithm designed to quantitatively assess lesion symmetry based on pixel counts within the segmented masks," and Section 3.3 later evaluates GSAA against this same ground truth, reporting 99.06% accuracy in Table 5. If the final labels were influenced by GSAA during labeling, the agreement is circular and the ISIC2016 rows of Tables 5, 6, 7, 10, and 11 are correspondingly inflated. The sentence in Section 3.3 that "the output of GSAA is not used" for CNN training does not address the possibility that the labels themselves incorporate GSAA. The authors must either disclose a labeling protocol in which GSAA played no role in the final expert decisions, or re-annotate a held-out subset for independent evaluation.
  2. [Section 3.1, Table 8] The APH2 evaluation uses a random split after mirror augmentation, which leaks identical lesions across the train/test boundary. Each of the 200 PH2 images is mirrored vertically and horizontally to produce 600 images, so each original lesion appears in three near-duplicate copies. If the 95/5 split is performed on the 600 images, the same original lesion can appear in both training and test sets; with 600 images and a 5% test fraction, most of the 200 originals will be split across the sets. This makes the same-dataset results in Fig. 8 and Table 8 (97% weighted F1, 94% Kappa) unreliable as estimates of generalization. The split should be stratified at the original-lesion level, e.g., by assigning all mirrored copies of each PH2 image to the same fold.
  3. [Section 4.1, Tables 6-9] The test sets are small and the results are given for a single random split. APH2 contributes 31 test images and ISIC2016 63 test images; with three classes, the per-class counts are in single digits for some classes (e.g., the half-symmetric class in the APH2 test set appears to have only 4-5 instances). The headline Kappa and F1 values therefore have very wide confidence intervals, and the repeated conclusion that "R101+SVM consistently emerges as the top-performing model" is not supported by any significance test. Please report results over multiple random splits (e.g., 5-fold or repeated 95/5 splits) with means and standard deviations, or at least provide the exact test-set CMs and class counts so readers can assess the uncertainty.
  4. [Section 4.2, Table 10] The comparative study does not support the claim that the proposed method "outperforms state-of-the-art methods." The proposed method's numbers are taken from different experimental configurations (the table itself cites "From tables 7&9"), while the comparison methods use their own datasets, class numbers (2 vs 3), and evaluation protocols. Accuracy values for different class counts are not directly comparable, and the APH2 accuracy is inflated by the leakage described above. A fair comparison requires a common benchmark and protocol, or at least matching class definitions and held-out splits.
minor comments (4)
  1. [Section 3.2, Eq. (7)] Equation (7) defines accuracy in binary terms (TP, TN, P, N) but is applied to three-class confusion matrices; please state that accuracy is computed as the fraction of correctly classified samples on the diagonal.
  2. [Section 3.4] The test set size for APH2 is given as 31 images, but 5% of 600 is 30; please verify the rounding or the augmentation count.
  3. [Table 10] Table 10 does not state which of Tables 6-9 each proposed-method accuracy value comes from; please add explicit table references so the reader can trace the provenance of each number.
  4. [Section 3.3] The GSAA ratio threshold of 0.90-1.10 is introduced without justification; please explain how this tolerance was chosen and whether the reported PH2 accuracy is sensitive to it.

Circularity Check

2 steps flagged · score 6.0 of 10

ISIC2016 ground-truth labels were augmented by the same pixel-count algorithm later scored at 99.06% on them, so the ISIC2016 GSAA accuracy and the CNN results measured against those labels are not fully external.

  1. self definitional [Section 3.1 (Data Acquisition); GSAA evaluation in Section 3.3, Table 5]
    "Furthermore, to augment our manual annotations, we developed an imaging algorithm designed to quantitatively assess lesion symmetry based on pixel counts within the segmented masks. This algorithm splits each lesion into four distinct sections, calculating the proportion of pixels within each segment relative to the total lesion area."

    This imaging algorithm is the same four-quadrant pixel-count method later presented as GSAA in Section 3.3. The ISIC2016 ground-truth labels were therefore not independent expert labels: they were augmented by GSAA during construction. When GSAA is then scored against that ground truth, as reported in Section 3.3 ('These outputs match the ground truth with an accuracy of 99.06%, as presented in the CM in Table 5'), the comparison is between the algorithm and a benchmark it helped create. The 99.06% ISIC2016 accuracy is therefore partly determined by the labeling procedure itself rather than by fully external clinical consensus.

  2. other [Section 3.1 vs. Section 3.3 and Section 4.1 (Tables 6 and 7)]
    "By combining manual annotations from clinical dermatologists with computational techniques, we provide ground truths about lesion shape and introduce a methodology for objectively quantifying symmetry characteristics. ... Since PH2 and ISIC2016 have clinical-based ground truth, which are more reliable than GSAA generated ground truth. For this reason, the clinical-based ground truth of PH2 and ISIC2016 are used to train and evaluate the proposed CNN."

    The second passage claims that the ISIC2016 CNN labels are clinical-based ground truth, but the first passage states that the ISIC2016 ground truths were produced by combining expert annotations with a computational imaging algorithm, i.e., the same pixel-count approach later called GSAA. Consequently, the CNN in-dataset ISIC2016 results in Table 6 and the ISIC2016-testing leg in Table 7 are evaluated against labels that GSAA helped define. This does not make the CNN scores tautological, but it removes their status as independent external validation; the ISIC2016 CNN numbers inherit the circularity of the ISIC2016 ground-truth construction.

full rationale

The circularity is real but partial. The PH2 path is independent: PH2 has pre-existing clinical lesion-shape ground truth, so the 99.00% GSAA agreement on PH2 (Table 3) and the best CNN headline figures from the APH2 in-dataset experiment (Table 8: 94% Kappa, 95% Macro F1, 97% Weighted F1) do not reduce to the algorithm's own construction. The problem is confined to ISIC2016: Section 3.1 explicitly says the authors augmented their manual annotations with a pixel-count imaging algorithm, and Section 3.3 then reports 99.06% GSAA agreement against those augmented labels and uses them as 'ground truth' for CNN training and evaluation. The later statement that 'the output of GSAA is not used' for CNN training does not resolve the ambiguity, because the labels themselves may have been shaped by GSAA. Cross-dataset experiments (APH2-to-ISIC2016 and ISIC2016-to-APH2) provide some independent evidence, since one leg uses independent PH2-derived labels, but the ISIC2016 leg still depends on possibly algorithm-augmented labels. There are no load-bearing self-citations or imported uniqueness theorems; the cited prior work is external. Overall, the ISIC2016 GSAA success and the ISIC2016 CNN scores are partially circular, giving a score of 6 rather than a higher score because the PH2-based claims remain externally grounded.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a hand-set tolerance, a fixed quadrant comparison scheme, and several domain assumptions about masks, centroid splitting, and feature transferability. No new physical entities are introduced.

free parameters (3)
  • GSAA quadrant ratio tolerance = 0.90 to 1.10
    Hand-set threshold classifying two quadrants as equal in pixel count; not derived from data or external theory.
  • GSAA quadrant pairing scheme = Ap/Bp, Ap/Dp, Bp/Cp, Cp/Dp
    The choice of which quadrant pairs are compared is fixed by design and affects the classification rule outcome.
  • Train/test split ratio = 95/5
    Random split chosen by the authors; the small test set (31 APH2 images, 63 ISIC2016 images) affects the stability of reported scores.
assumptions (4)
  • domain assumption Binary lesion masks accurately represent the lesion shape relevant to clinical asymmetry.
    GSAA operates only on masks and ignores color and structure, even though the ABCDE criteria include color and structure asymmetry.
  • domain assumption The centroid-based split into four quadrants is a valid operationalization of dermatological symmetry.
    The paper assumes that comparing quadrant pixel counts captures the clinical notion of shape asymmetry.
  • domain assumption ImageNet-pretrained features transfer to dermoscopic lesion shape classification.
    The pipeline uses ResNet activations trained on ImageNet without domain-specific fine-tuning.
  • domain assumption Expert labels for PH2 and ISIC2016 are correct and independent of the proposed algorithm.
    The paper's evaluation treats clinical ground truth as the reference, but Section 3.1 states the GSAA was used to augment ISIC2016 labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Asymmetric Lesion Detection with Geometric Patterns and CNN-SVM Classification." pith.science (2026). https://pith.science/paper/DIFYZ363

@misc{pith2026250717185,
  author       = {Pith},
  title        = {Pith review of: Asymmetric Lesion Detection with Geometric Patterns and CNN-SVM Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIFYZ363}},
  note         = {Machine review of arXiv:2507.17185}
}
read the original abstract

In dermoscopic images, which allow visualization of surface skin structures not visible to the naked eye, lesion shape offers vital insights into skin diseases. In clinically practiced methods, asymmetric lesion shape is one of the criteria for diagnosing melanoma. Initially, we labeled data for a non-annotated dataset with symmetrical information based on clinical assessments. Subsequently, we propose a supporting technique, a supervised learning image processing algorithm, to analyze the geometrical pattern of lesion shape, aiding non-experts in understanding the criteria of an asymmetric lesion. We then utilize a pre-trained convolutional neural network (CNN) to extract shape, color, and texture features from dermoscopic images for training a multiclass support vector machine (SVM) classifier, outperforming state-of-the-art methods from the literature. In the geometry-based experiment, we achieved a 99.00% detection rate for dermatological asymmetric lesions. In the CNN-based experiment, the best performance is found with 94% Kappa Score, 95% Macro F1-score, and 97% Weighted F1-score for classifying lesion shapes (Asymmetric, Half-Symmetric, and Symmetric).

Figures

Figures reproduced from arXiv: 2507.17185 by the authors.

Figure 1
Figure 1. (a) Asymmetric (do not match); and (b) Symmetric (almost match) lesions are from the ISIC2016 dataset [4]. The requirement for precise and prompt evaluation has driven the incorporation of artificial intelligence in medical settings, resulting in inventive approaches that enable the computerized examination of dermoscopic images. In recent work, researchers have used CNNs to create advanced algorithms capable of dis… view at source ↗
Figure 2
Figure 2. Dermoscopic images (a, c) and corresponding binary images of lesions (b, d), respectively from PH2 (a, b) and ISIC2016 (c, d) datasets. Each binary (mask) image is split into 4 portions using two perpendicular axes at the lesion centroid. As a result, 4 different split images are found in a single image [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. (a), (b), and (c) are briefly presented architecture of ResNet18, ResNet50, and ResNet101. The 'MiniBatchSize' is set 32 to ensure that the model and image data fit into the system’s GPU memory. Also, the activation output is arranged as columns. This helps speed up the multiclass linear SVM training that follows [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: shows the multiclass SVMs classification process for lesion shape detection. This choice helps speed up training, especially when working with high-dimensional ResNet feature vectors, as linear SVMs are computationally efficient and can handle large feature spaces effe…
Figure 6
Figure 6. Figure 6: (a), (b), and (c) are the CMs of ResNets+SVM, when training on ISIC2016 and testing on ISIC2016. (a) R18+SVM (b) R50+SVM (c) R101+SVM [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: (a), (b), and (c) are the CMs of ResNets+SVM, when Training on APH2 and testing on ISIC2016. (a) R18+SVM (b) R50+SVM (c) R101+SVM [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: (a), (b), and (c) are the CMs of ResNets+SVM, when training on APH2 and testing on APH2. (a) R18+SVM (b) R50+SVM (c) R101+SVM [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: (a), (b), and (c) are the CMs of ResNets+SVM, when training on ISIC2016 and testing on APH2 [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages

  1. [1]

    ABCDE —An Evolving Concept in the Early Detection of Melanoma

    Rigel, D.S., Friedman, R.J., Kopf, A.W., Polsky, D., 2005. ABCDE —An Evolving Concept in the Early Detection of Melanoma. Archives of Dermatology 141. https://doi.org/10.1001/archderm.141.8.1032

    work page doi:10.1001/archderm.141.8.1032 2005

  2. [2]

    Three -point checklist of dermoscopy: an open internet study

    Zalaudek, I., Argenziano, G., Soyer, H.P., Corona, R., Sera, F., Blum, A., Braun, R.P., Cabo, H., Ferrara, G., Kopf, A.W., La ngford, D., Menzies, S.W., Pellacani, G., Peris, K., Seidenari, S., 2005. Three -point checklist of dermoscopy: an open internet study. British Journal of Dermatology 154, 431–437. https://doi.org/10.1111/j.1365-2133.2005.06983.x

    work page arXiv 2005

  3. [3]

    CASH Algo rithm for Dermoscopy Revisited

    Henning, J.S., Stein, J.A., Yeung, J., Dusza, S.W., Marghoob, A.A., Rabinovitz, H.S., Polsky, D., Kopf, A.W., 2008. CASH Algo rithm for Dermoscopy Revisited. Archives of Dermatology 144. https://doi.org/10.1001/archderm.144.4.554

    work page doi:10.1001/archderm.144.4.554 2008

  4. [4]

    F.; Celebi, Emre; Helba, Brian; Marchetti, Michael; Mishra, Nabin; Halpern, Allan

    Gutman, David; Codella, Noel C. F.; Celebi, Emre; Helba, Brian; Marchetti, Michael; Mishra, Nabin; Halpern, Allan. "Skin Lesi on Analysis toward Melanoma Detection: A Challenge at the International Symposium on Biomedical Imaging (ISBI) 2016, hosted by the Intern ational Skin Imaging Collaboration (ISIC)". eprint arXiv:1605.01397. 2016. Accepted Manuscrip...

    arXiv 2016

  5. [5]

    Structural asymmetry as a dermatoscopic indicator of malignant melanoma – a latent class analysis of sensitivity and classification errors

    Lorentzen, H.F., Weismann, K., Larsen, F.G., 2001. Structural asymmetry as a dermatoscopic indicator of malignant melanoma – a latent class analysis of sensitivity and classification errors. Melanoma Research 11, 495–501. https://doi.org/10.1097/00008390-200110000-00009

    work page doi:10.1097/00008390-200110000-00009 2001

  6. [6]

    Determining the asymmetry of skin lesion with fuzzy borders

    Ng, V.T.Y., Fung, B.Y.M., Lee, T.K., 2005. Determining the asymmetry of skin lesion with fuzzy borders. Computers in Biology and Medicine 35, 103–120. https://doi.org/10.1016/j.compbiomed.2003.11.004

    work page doi:10.1016/j.compbiomed.2003.11.004 2005

  7. [7]

    Automatic boundary detection and symmetry calculation in dermoscopy images of skin lesions

    7.Sirakov, N.M., Mete, M., Chakrader, N.S., 2011. Automatic boundary detection and symmetry calculation in dermoscopy images of skin lesions. 2011 18th IEEE International Conference on Image Processing. https://doi.org/10.1109/icip.2011.6115757

    work page arXiv 2011

  8. [8]

    Dermatologist -like feature extraction from skin lesion for improved asymmetry classification in PH2 database

    Chakravorty R, Liang S, Abedini M, Garnavi R. Dermatologist -like feature extraction from skin lesion for improved asymmetry classification in PH2 database. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMB S, vol. 2016- October, Institute of Electrical and Electronics Engineers Inc.; 2016, p. ...

    work page arXiv 2016

Show all 24 references
  1. [9]

    Skin lesions dermatological shape asymmetry measures,

    P. Milczarski, Z. Stawska and P. Maślanka , "Skin lesions dermatological shape asymmetry measures," 2017 9th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS), 2017, pp. 1056 -1061, doi: 10.1109/I...

    2017

  2. [10]

    Quantitative evaluation of binary digital region asymmetry with application to skin lesion detection

    Sancen-Plaza, A., Santiago-Montero, R., Sossa, H., Perez -Pinal, F.J., Martinez-Nolasco, J.J., Padilla-Medina, J.A., 2018. Quantitative evaluation of binary digital region asymmetry with application to skin lesion detection. BMC Medical Informatics and Decision Making 18. http...

    2018 doi

  3. [11]

    Towards the automatic detection of skin lesion shape asymmetry, color variegation and diameter in dermoscopic images

    Ali AR, Li J, O’Shea SJ. Towards the automatic detection of skin lesion shape asymmetry, color variegation and diameter in dermoscopic images. PLoS One 2020;15. https://doi.org/10.1371/journal.pone.0234352

    2020 doi

  4. [12]

    Skin Lesions Asymmetry Estimation Using Artificial Neural Networks

    Damian FA, Moldovanu S, Moraru L. Skin Lesions Asymmetry Estimation Using Artificial Neural Networks. 2021 25th International Conference on System Theory, Control and Computing, ICSTCC 2021 - Proceedings, Institute of Electrical and Electronics Engineers Inc.; 2021, p. 64 –7. ...

    2021

  5. [13]

    Asymmetry analysis of melanoma based on ABCD rule

    Zhang, G., Guo, S., 2021. Asymmetry analysis of melanoma based on ABCD rule. Journal of Physics: Conference Series 1883, 0120 70. https://doi.org/10.1088/1742-6596/1883/1/012070

    2021 doi

  6. [14]

    A novel approach for skin lesion symmetry classification with a deep learning model

    Talavera-Martínez, L., Bibiloni, P., Giacaman, A., Taberner, R., Hernando, L.J.D.P., González -Hidalgo, M., 2022. A novel approach for skin lesion symmetry classification with a deep learning model. Computers in Biology and Medicine 145, 105450. https://doi.org/10.1016/j.compb...

    2022

  7. [15]

    Ferreira, Jorge Marques, Andre R

    Teresa Mendonça, Pedro M. Ferreira, Jorge Marques, Andre R. S. Marcal, Jorge Rozeira. PH² - A dermoscopic image database for research and benchmarking, 35th International Conference of the IEEE Engineering in Medicine and Biology Society, July 3-7, 2013, Osaka, Japan

    2013

  8. [16]

    Imagenet: A large -scale hierarchical image database

    Deng, Jia, et al. "Imagenet: A large -scale hierarchical image database." Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009

    2009

  9. [17]

    Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vis ion and Pattern Recognition (CVPR). https://doi.org/10.1109/cvpr.2016.90

    2016 doi

  10. [18]

    ImageNet classification with deep convolutional neural networks

    Krizhevsky, A., Sutskever, I., Hinton, G.E., 2017. ImageNet classification with deep convolutional neural networks. Communica tions of the ACM 60, 84–90. https://doi.org/10.1145/3065386

    2017 doi

  11. [19]

    Review of deep learning: concepts, CNN architectures, challenges, applications, future directions

    Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., Santamaría, J., Fadhel, M.A., Al-Amidie, M., Farhan, L., 2021. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. Journal of Big Data 8. https://do...

    2021 doi

  12. [20]

    and Darrell, T., 2014, January

    Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E. and Darrell, T., 2014, January. Decaf: A deep convolutio nal activation feature for generic visual recognition. In International conference on machine learning (pp. 647-655). PMLR

    2014

  13. [21]

    Minimizing Average of Loss Functions Using Gradient Descent and Stochastic Gradient Desc ent

    Arefin, M.R., Asadujjaman, M., 2016. Minimizing Average of Loss Functions Using Gradient Descent and Stochastic Gradient Desc ent. Dhaka University Journal of Science 64, 141–145. https://doi.org/10.3329/dujs.v64i2.54490

    2016 doi

  14. [22]

    An improved multiclass LogitBoost using adaptive -one-vs-one

    Sun, P., Reid, M.D., Zhou, J., 2014. An improved multiclass LogitBoost using adaptive -one-vs-one. Machine Learning 97, 295 –326. https://doi.org/10.1007/s10994-014-5434-3

    2014 doi

  15. [23]

    Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M. Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Ko nstantinos Liopyris, Michael Marchetti, Harald Kittler, Allan Halpern: "Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the ...

    2018 arXiv

  16. [24]

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

    Tschandl, P., Rosendahl, C., Kittler, H., 2018. The HAM10000 dataset, a large collection of multi -sourcedermatoscopic images of common pigmented skin lesions. Scientific Data 5. https://doi.org/10.1038/sdata.2018.161

    2018 doi

Pith tools

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