- Open Access
Texture based skin lesion abruptness quantification to detect malignancy
© The Author(s). 2017
Published: 28 December 2017
Abruptness of pigment patterns at the periphery of a skin lesion is one of the most important dermoscopic features for detection of malignancy. In current clinical setting, abrupt cutoff of a skin lesion determined by an examination of a dermatologist. This process is subjective, nonquantitative, and error-prone. We present an improved computational model to quantitatively measure abruptness of a skin lesion over our previous method. To achieve this, we quantitatively analyze the texture features of a region within the lesion boundary. This region is bounded by an interior border line of the lesion boundary which is determined using level set propagation (LSP) method. This method provides a fast border contraction without a need for extensive boolean operations. Then, we build feature vectors of homogeneity, standard deviation of pixel values, and mean of the pixel values of the region between the contracted border and the original border. These vectors are then classified using neural networks (NN) and SVM classifiers.
As lower homogeneity indicates sharp cutoffs, suggesting melanoma, we carried out our experiments on two dermoscopy image datasets, which consist of 800 benign and 200 malignant melanoma cases. LSP method helped produce better results than Kaya et al., 2016 study. By using texture homogeneity at the periphery of a lesion border determined by LSP, as a classification results, we obtained 87% f1-score and 78% specificity; that we obtained better results than in the previous study. We also compared the performances of two different NN classifiers and support vector machine classifier. The best results obtained using combination of RGB color spaces with the fully-connected multi-hidden layer NN.
Computational results also show that skin lesion abrupt cutoff is a reliable indicator of malignancy. Results show that computational model of texture homogeneity along the periphery of skin lesion borders based on LSP is an effective way of quantitatively measuring abrupt cutoff of a lesion.
Melanoma is one of the deadliest and fastest growing cancer types in the world. In the USA annually 3.5 million skin cancers are diagnosed. Skin cancer is rarely fatal except melanoma which is the 6th common cancer type in the USA . Women 25–29 years of age are the most commonly affected group from melanoma. Ultraviolet tanning devices are listed as known and probable human carcinogens along with plutonium and cigarettes by World Health Organization . In 2017, an estimated 87,110 adults were diagnosed with melanoma in the USA and approximately 9730 were fatal .
Melanoma is a malignancy of melanocytes. Melanocytes are special cells in skin located in its outer epidermis. Since melanoma develops in epidermis, it can be seen by human eye. Early diagnosis and treatment are critical to prevent harm. When caught early, melanoma can be cured through excision operation. However, high rate of false-negative of malignant melanoma is the main challenge for early treatments .
Dermoscopy, a minimal invasive skin imaging technique, is one of the viable methods for detecting melanoma and other pigmented skin proliferations. In the current clinical settings, first step of dermoscopic evaluation is to decide whether the lesion meloanocytic or not. The second step is to find out whether the lesion is benign or malignant. There are commonly accepted protocols to detect malignancy in skin lesions, which are ABCD Rule, 7-point Checklist, Pattern Analysis, Menzies Method, Revised Pattern Analysis, 3-point Checklist, 4-point Checklist, and CASH Algorithm [3, 4].
Celebi et al.  extracted shape, color, and texture features and fed these feature vectors to classifier such that they are ranked using feature selection algorithms to determine the optimal subset size. Their approach yielded a specificity of 92.34% and a sensitivity of 93.33% using 564 images. In their seminal work, Dreiseitl et al.  analyzed the robustness of artificial neural networks (ANN), logistic regression, k-nearest neighbors, decision trees, and support vector machines (SVMs) on classifying common nevi, dysplastic nevi, and melanoma. They addressed three classification problems; dichotomous problem of separating common nevi from melanoma and dysplastic nevi, and the trichotomous problem of genuinely separating all these classes. They reported that on both cases (dichotomous and trichotomous) logistic regression, ANNs, and SVMs showed the same performance, whereas k-nearest neighbor and decision trees performed worse.
Rubegni et al.  extracted texture features, besides color and shape features. Their ANN based approach reached the sensitivity of 96% and specificity 93% on a data set of 558 images containing 217 melanoma cases. Iyatomi et al.  proposed an internet-based system which employs a feature vector consists of shape, texture, and color features. They achieved specificity and sensitivity of 86% using 1200 dermoscopy images. Local methods have also been recently applied for skin lesion classification. Situ et al.  offered a patch-based algorithm which is to use Bag-of-Features approach. They sampled the region of lesion into 16 × 16 grid and extracted Wavelets and Gabor filters as collecting 23 features in total. They compared two different classifiers which are Naïve Bayes and SVM; the best performance they achieved is 82% specificity on a dataset consists of 100 images with 30 melanoma cases.
A considerable number of systems have been proposed for melanoma detection in the last decade. Some of them aim to mimic the procedure that dermatologists pursue for detecting and extracting dermoscopic features, such as granularities , irregular streaks , regression structure , blotches , and blue-white veils . These structures are also used by dermatologists to score the lesion based on seven point-checklist. Leo et al.  described a CAD system that mimics the 7 point-checklist procedure.
However, approaches [5, 7, 15, 16] in the literature dominantly pursued pattern recognition in melanoma detection. Majority of these works are inspired by the ABCD rule , and they extract the features according to the score table of ABCD protocol. Shape features (e.g. irregularity, aspect ratio and maximum diameter, compactness), which refer to both asymmetry and border, color features in several color channels and texture features (e.g., gray level co-occurrence matrix)  are the most common features analyzed when ABCD protocol is used . There are other approaches [15, 18, 19] that used one type of feature for detection of melanoma. Seidenari et al.  aim to distinguish atypical nevi and benign nevi using color statistics in the RGB channel, such as mean, variance, and maximum RGB distance. Their approach reached 86% accuracy, additionally they concluded that there is a remarkable difference in distribution of pigments between the populations they studied. Color histograms have been utilized for discriminating melanomas and atypical or benign nevi [18, 19] with specificity little higher than 80%.
Dermoscopic image analysis
Boundary detection and boundary contour extraction
LSP for lesion border contraction
Shape contraction algorithms play an important role in computer graphics, computer-aided design, manufacturing, CNC machines. We adopted the method studied in a seminal paper of Kimmel et al. . Following set of formulations give the details of this approach.
Numerical-propagation scheme takes central derivatives of x and y in location s, and forward-derivative in time t. However, Langrangian based numerical propagation of a curve given in Eq. 6 is unstable and suffers from aforementioned topological problems, i.e. shocks, self-intersections (a.k.a. self collision). To maintain stability and address topological problems, instead of Langrangian numerical propogation, we use the ‘Eulerian formulation’.
After contracted border is found with LSP method, we calculate texture homogeneity between lesion border and contracted border with various radii sizes.
After border contraction using the LSP and extracting homogeneity features in GLCM, next step is to analyze generated data.
Data analysis and results
After feature extraction step, we categorized dataset according to thickness of layer they are collected from. As mentioned in the abstract, we selected 5, 7, 10, and 15 as the radius of circles between border and contracted border, and the layer is generated by enveloping these circles. In each overlapping circles (patches), we compute the “mean-homogeneity”, “min-homogeneity”, “mean- color value average”, “minimum color value average”, “mean color value standard deviation”, and “minimum color standard deviation”. We performed the experiments on two different color spaces which are RGB and HSV and fed them as input to the NN architectures and SVM.
Dataset provides dermoscopy images which are labeled either as malignant or benign. We are measuring abruptness of lesion along the periphery of the lesion border using homogeneity features to conduct binary classification. Here, we argue that Multi-layer Perceptron-based Neural Networks (MPNN) have ability to compete with SVM, when it is combined with softmax regression.
The hidden layer system can include multi-layers within separate instances better and converge the values efficiently. A careful design of a NN is required for obtaining higher accuracy rates in classification. There are some parameters that the user needs to tune  for the best accuracy, such as input layer selection, weights, the number of hidden layers, the number of nodes on each hidden layer, activation function, learning rate, the number of iterations, and cost minimization function. We train our NN with a pair of input feature values and output malignancy values. In our study, in order to solve the malignancy problem of the dataset, we choose two NN architectures; multi-layer perceptron and the fully-connected multi-hidden layer NN.
We obtained results of two different abrupt cutoff feature extraction methods; Kaya et al.  and our LSP based method using the two NN architectures introduced above with same parameters. Optimum results are obtained from the features collected when radius is 10 and on RGB channel. NNs are highly sensitive to hyper-parameter changes, we applied tunings to get optimum results. We empirically determined the iteration numbers as 600, 750, and 1000 without constraining a stoppage criterion. Then, we added the learning rate of 0.0001 to exit the iteration between two consecutive epochs. We applied 10-fold cross-validation to split the data into training and test sets. Since NNs generate random weights between the layers at each time, we run the algorithms 10 times. Consequently, all evaluation metrics are the average of the all results generated in these experiments. Notably, to maintain consistency we used same dataset to test our NN designs.
LSP vs. DS based texture homogeneity feature extraction and classification of lesions with various classifiers: multi-layer perceptron, fully connected multi-hidden layer NN, and SVM. 10-fold cross-validation is used. Results listed here are means of 10 random executions
Feature Extraction- Classification
LSP-Multilayer Perceptron NN
DS-Multi Layer Perceptron NN
LSP-Fully-connected multi-layer NN
DS-Fully-connected multi-hidden layer NN
The parameters of the NN (the multi layer perceptron and the fully-connected multi-hidden layer NN) classifiers and SVM
Number of iteration
Number of run
Number of hidden layer
Number of hidden layer node
Number of hidden layers (If multilayer NN is used)
An improved automated measurement of abrupt cutoff for skin lesions is presented. LSP over dynamic scaling to do lesion border contraction is introduced. Computational results showed that skin lesion abrupt cutoff is a worthy indicator of malignancy. Results show that computational model of texture homogeneity along the periphery of skin lesion borders is an effective tool to quantitatively measure abrupt cutoff of a lesion. A multi-layer perceptron and a fully connected multi-hidden layer NN, and SVM classifiers are used. We obtained 87% f1-score and 78% specificity for correctly classifying lesions with the fully-connected multi-hidden layer NN classifier and LSP based border contraction method.
The publication cost of this article was funded by Arkansas Science and Technology Association Award# 15-B-25 and by the Arkansas INBRE program, with an award# P20 GM103429 from the National Institutes of Health/the National Institute of General Medical Sciences (NIGMS).
Availability of data and materials
Data set is available at https://isic-archive.com/.
About this supplement
This article has been published as part of BMC Bioinformatics Volume 18 Supplement 14, 2017: Proceedings of the 14th Annual MCBIOS conference. The full contents of the supplement are available online at https://bmcbioinformatics.biomedcentral.com/articles/supplements/volume-18-supplement-14.
MB and RE implemented the proposed method as graduate students under SK’s supervision. SK made the overall design of the study and involved algorithm development. SKa implemented dynamic scaling method. TH helped in development of the general comparison testbed, performing data analysis, and statistical measurements. All of the authors read and approved the manuscript.
Ethics approval and consent to participate
Consent for publication
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Cancer Facts and Statistics 2015 | Research | American Cancer Society. [Online]. Available: https://isic-archive.com/. Accessed 23 Mar 2016.
- Siegel RL, Miller KD, Jemal A. Cancer statistics, 2017. CA Cancer J Clin. 2017;67(1):7–30.View ArticlePubMedGoogle Scholar
- Walter FM, Prevost AT, Vasconcelos J, Hall PN, Burrows NP, Morris HC, Kinmonth AL, Emery JD. Using the 7-point checklist as a diagnostic aid for pigmented skin lesions in general practice: a diagnostic validation study. Br J Gen Pract. 2013;63(610):e345–53.View ArticlePubMedPubMed CentralGoogle Scholar
- di Meo N, Stinco G, Bonin S, Gatti A, Trevisini S, Damiani G, Vichi S, Trevisan G. CASH algorithm versus 3-point checklist and its modified version in evaluation of melanocytic pigmented skin lesions: the 4-point checklist. J Dermatol. 2016;43(6):682–5.View ArticlePubMedGoogle Scholar
- Celebi ME, Kingravi HA, Uddin B, Iyatomi H, Aslandogan YA, Stoecker WV, Moss RH. A methodological approach to the classification of dermoscopy images. Comput Med Imaging Graph. 2007;31(6):362–73.View ArticlePubMedPubMed CentralGoogle Scholar
- Dreiseitl S, Ohno-Machado L, Kittler H, Vinterbo S, Billhardt H, Binder M. A comparison of machine learning methods for the diagnosis of pigmented skin lesions. J Biomed Inform. 2001;34(1):28–36.View ArticlePubMedGoogle Scholar
- Rubegni P, Cevenini G, Burroni M, Perotti R, Dell’Eva G, Sbano P, Miracco C, Luzi P, Tosi P, Barbini P, et al. Automated diagnosis of pigmented skin lesions. Int J Cancer. 2002;101(6):576–80.View ArticlePubMedGoogle Scholar
- Iyatomi H, Oka H, Celebi ME, Hashimoto M, Hagiwara M, Tanaka M, Ogawa K. An improved internet-based melanoma screening system with dermatologist-like tumor area extraction algorithm. Comput Med Imaging Graph. 2008;32(7):566–79.View ArticlePubMedGoogle Scholar
- Situ N, Yuan X, Chen J, Zouridakis G. Malignant melanoma detection by bag-of-features classification. In: 2008 30th annual international conference of the IEEE engineering in medicine and biology society: 20–25 Aug. 2008 2008; 2008. p. 3110–3.View ArticleGoogle Scholar
- Stoecker WV, Wronkiewiecz M, Chowdhury R, Stanley RJ, Xu J, Bangert A, Shrestha B, Calcara DA, Rabinovitz HS, Oliviero M, et al. Detection of granularity in dermoscopy images of malignant melanoma using color and texture features. Comput Med Imaging Graph. 2011;35(2):144–7.View ArticlePubMedGoogle Scholar
- Fabbrocin G, Betta G, Leo GD, Liguor C, Paolillo A, Pietrosanto A, Sommella P, Rescigno O, Cacciapuoti S, Pastore F, Vita VD, Mordente I, Ayala F. Epiluminescence image processing for melanocytic skin lesion diagnosis based on 7-point check-list: a preliminary discussion on three parameters. Open Derm J. 2010;4:110–5.View ArticleGoogle Scholar
- Stoecker WV, Gupta K, Stanley RJ, Moss RH, Shrestha B. Detection of asymmetric blotches (asymmetric structureless areas) in dermoscopy images of malignant melanoma using relative color. Skin Res Technol. 2005;11(3):179–84.View ArticlePubMedPubMed CentralGoogle Scholar
- Celebi ME, Iyatomi H, Stoecker WV, Moss RH, Rabinovitz HS, Argenziano G, Soyer HP. Automatic detection of blue-white veil and related structures in Dermoscopy images. Comput Med Imaging Graph. 2008;32(8):670–7.View ArticlePubMedPubMed CentralGoogle Scholar
- Leo GD, Paolillo A, Sommella P, Fabbrocini G. Automatic diagnosis of melanoma: a software system based on the 7-point check-list. In: 2010 43rd Hawaii international conference on system sciences: 5–8 Jan. 2010 2010; 2010. p. 1–10.Google Scholar
- Seidenari S, Pellacani G, Grana C. Pigment distribution in melanocytic lesion images: a digital parameter to be employed for computer-aided diagnosis. Skin Res Technol. 2005;11(4):236–41.View ArticlePubMedGoogle Scholar
- Burroni M, Sbano P, Cevenini G, Risulo M, Dell’Eva G, Barbini P, Miracco C, Fimiani M, Andreassi L, Rubegni P. Dysplastic naevus vs. in situ melanoma: digital dermoscopy analysis. Br J Dermatol. 2005;152(4):679–84.View ArticlePubMedGoogle Scholar
- Nachbar F, Stolz W, Merkle T, Cognetta AB, Vogt T, Landthaler M, Bilek P, Braun-Falco O, Plewig G. The ABCD rule of dermatoscopy. High prospective value in the diagnosis of doubtful melanocytic skin lesions. J Am Acad Dermatol. 1994;30:551–9.View ArticlePubMedGoogle Scholar
- Stanley RJ, Stoecker WV, Moss RH. A relative color approach to color discrimination for malignant melanoma detection in dermoscopy images. Skin Res Technol. 2007;13(1):62–72.View ArticlePubMedPubMed CentralGoogle Scholar
- Faziloglu Y, Stanley RJ, Moss RH, Van Stoecker W, McLean RP. Colour histogram analysis for melanoma discrimination in clinical images. Skin Res Technol. 2003;9(2):147–56.View ArticlePubMedPubMed CentralGoogle Scholar
- Gutman D, Codella NCF, Celebi E, Helba B, Marchetti M, Mishra N, Halpern A. Skin Lesion Analysis toward Melanoma Detection: A Challenge at the International Symposium on Biomedical Imaging (ISBI) 2016, hosted by the International Skin Imaging Collaboration (ISIC). CoRR 2016, abs/1605.01397.Google Scholar
- Lio PA, Nghiem P. Interactive atlas of Dermoscopy. J Am Acad Dermatol. 2004;50(5):807–8.View ArticleGoogle Scholar
- Mete M, Sirakov NM. Lesion detection in demoscopy images with novel density-based and active contour approaches. BMC Bioinformatics. 2010;11(6):S23.View ArticlePubMedPubMed CentralGoogle Scholar
- Kaya S, Bayraktar M, Kockara S, Mete M, Halic T, Field HE, Wong HK. Abrupt skin lesion border cutoff measurement for malignancy detection in dermoscopy images. BMC Bioinformatics. 2016;17(13):367.View ArticlePubMedPubMed CentralGoogle Scholar
- Kimmel R, Bruckstein AM. Shape offsets via level sets. Comput Aided Des. 1993;25(3):154–62.View ArticleGoogle Scholar
- H Freeman. On the encoding of arbitrary geometric configurations, IRE Transactions on Electronic Computers EC- 10. 1961 p. 260–268Google Scholar
- Osher S, Sethian JA. Fronts propagating with curvature-dependent speed: algorithms based on Hamilton-Jacobi formulations. J Comput Phys. 1988;79(1):12–49.View ArticleGoogle Scholar
- Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern. 1973;SMC-3(6):610–21.View ArticleGoogle Scholar
- LeCun Y, Bottou L, Orr GB, Müller K-R. Efficient BackProp. In: Neural networks: tricks of the trade - second edition; 2012. p. 9–48.View ArticleGoogle Scholar
- Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, et al. Scikit-learn: machine learning in python. J Mach Learn Res. 2011;12:2825–30.Google Scholar
- Softmax Tutoria; [http://ufldl.stanford.edu/wiki/index.php/Softmax_Regression]. Accessed 9 Nov 2017.
- Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, Devin M, Ghemawat S, Irving G, Isard M, et al. TensorFlow: a system for large-scale machine learning; 2016. p. 265–83.Google Scholar