Elsevier

Knowledge-Based Systems

Volume 158, 15 October 2018, Pages 118-135
Knowledge-Based Systems

Intelligent skin cancer detection using enhanced particle swarm optimization

https://doi.org/10.1016/j.knosys.2018.05.042Get rights and content

Highlights

  • We conduct intelligent skin cancer diagnosis using dermoscopic images.

  • An enhanced PSO algorithm is proposed for feature selection.

  • It integrates subswarms, mutation mechanisms and dynamic matrix representations.

  • It follows leaders and avoids enemies in every or randomly selected sub-dimensions.

  • It outperforms other optimization methods and related research significantly.

Abstract

In this research, we undertake intelligent skin cancer diagnosis based on dermoscopic images using a variant of the Particle Swarm Optimization (PSO) algorithm for feature optimization. Since the identification of the most significant discriminative characteristics of the benign and malignant skin lesions plays an important role in robust skin cancer detection, the proposed PSO algorithm is employed for feature optimization. It incorporates not only subswarms, local and global food and enemy signals, attraction and flee operations, and mutation-based local exploitation, but also diverse matrix representations to mitigate premature convergence of the original PSO algorithm. Specifically, two remote swarm leaders, which show similar fitness but low position proximity, are used to lead the subswarm-based search and to enable the exploration of more distinctive search regions. Modified velocity updating strategies are also proposed to enable the particles to follow multiple swarm leaders and avoid the local and global worst individuals, partially (i.e. in randomly selected sub-dimensions) and fully (in every dimension), with an attempt to search for global optima. Probability distribution and dynamic matrix representations are used to diversify the search process. Evaluated with multiple skin lesion and UCI databases and diverse unimodal and multimodal benchmark functions, the proposed PSO variant shows a superior performance over those of other advanced and classical search methods for identifying discriminative features that facilitate benign and malignant lesion classification as well as for solving diverse optimization problems with different landscapes. The Wilcoxon rank sum test is adopted to further ascertain superiority of the proposed algorithm over other methods statistically.

Introduction

Melanoma is an aggressive type of skin cancer that can spread to other organs. Automatic and early diagnosis of melanoma is essential for administering effective treatment and increasing survival chances. Since medical skin cancer diagnosis employs the Asymmetry, Border, Colour, Diameter and Enlargement (ABCDE) guideline, the extraction and identification of such discriminative and significant morphological characteristics play a crucial role in attaining accurate diagnosis rates. However, it is still a challenging task for the retrieval of such distinguishing attributes, owing to the fine-grained variability in the appearance of benign and cancerous skin lesions [1].

This research aims to deal with the above challenge by proposing an enhanced Particle Swarm Optimization (PSO) algorithm for discriminative feature selection in skin cancer diagnosis using dermoscopic images. The main motivations of this research are as follows. Evolutionary algorithms possess powerful search capabilities, and have been widely used for solving various feature selection challenges. Owing to the comparatively simple underlying concepts and relatively few user-defined parameters, PSO has been widely studied for feature selection tasks. Since PSO has a weak exploration capability, and its search process is likely to be trapped in local optima when dealing with multimodal or complex optimization problems, new PSO variants with superior explorative capabilities are required. Therefore, we propose an enhanced PSO model in this research.

Specifically, our proposed PSO model incorporates the subswarm concept, food and enemy signals, attraction and flee operations, mutation-based local exploitation, and diverse matrix representations to mitigate premature convergence of the original PSO algorithm. It shows a great superiority over other methods for the identification of the most significant characteristics of benign and malignant lesion images to facilitate subsequent skin cancer classification. The proposed skin cancer detection system consists of five key stages, i.e. pre-processing, skin lesion segmentation, feature extraction, PSO-based feature selection and classification. The overall system architecture is illustrated in Fig. 1.

The key contributions of this research, which focus on PSO-based feature selection, are as follows.

  • The proposed PSO model first of all divides the original population into two subswarms. Then, two swarm leaders with competitive fitness scores but low position proximity are identified. Each leader leads one subswarm-based search for discriminative lesion feature selection. Since the subswarm-based search is more likely to explore distinctive search regions owing to the low position correlation between the two leaders, it reduces the probability of being trapped in local optima and increases the chances of finding the global optimum.

  • A number of attraction and enemy signals are used for velocity updating in each subswarm. The proposed PSO model enables each particle to follow the leader (i.e. food attraction) and avoid unpromising solutions (i.e. enemies) partially (i.e. in randomly selected sub-dimensions) as well as fully (i.e. in each dimension) to diversify the search process. Three random walks, i.e. Gaussian, Cauchy, and Levy distributions, are used to further enhance the best subswarm solution and to increase exploitation. A dynamic matrix representation of the swarm is also utilized during the search process to increase search diversity. The proposed algorithm shows great efficiency in optimal feature selection for melanoma classification, as well as solving unimodal and multimodal benchmark problems in comparison with other search methods. It is also among the top performers for skin cancer detection in comparison with related research studies reported in the literature.

The paper is organised as follows. Section 2 presents the related studies on skin cancer detection, enhanced PSO variants and diverse evolutionary algorithm-based feature selection techniques. Section 3 presents the proposed PSO model with mutation-based local exploitation as well as attraction and flee based global exploration. A detailed evaluation of the proposed algorithm and other classical methods using skin lesion data sets and other benchmark problems is presented in Section 4. Concluding remarks and suggestions for further research are provided in Section 5.

Section snippets

Related work

In this section, we discuss the related work on computerized skin cancer diagnosis, diverse variants of the PSO algorithm, and evolutionary algorithm-based feature selection methods.

The proposed skin cancer detection system

We propose an intelligent system for benign and malignant skin lesion classification. The proposed system consists of five key stages, i.e. pre-processing, skin lesion segmentation, feature extraction, PSO-based feature optimization and classification. Each key stage, especially the feature selection process, is explained comprehensively, as follows.

Evaluation

To evaluate the proposed PSO variant, we implement several classical search methods for comparison, i.e., PSO [38], Bat Algorithm (BA) [40], Harmony Search (HS) [41], GA [42], Dragonfly Algorithm (DA) [43], Flower Pollination Algorithm (FPA) [44], Moth-Flame Optimization (MFO) [45], Artificial Bee Colony (ABC) [46], Cultural Algorithm (CA) [47], and BBPSO [48]. Several advanced PSO variants are implemented for comparison, including DNLPSO [15], ELPSO [16], AGPSO [17], ThBPSO [26], MS-PSO [27],

Conclusions

In this research, we have described skin lesion classification using PSO-based feature optimization. The proposed PSO model integrates diverse alternative velocity updating strategies in the subswarms to enable a wider exploration of the search space. Two remote swarm leaders have been employed to lead the subswarm-based search to explore distinctive regions. Probability distributions and dynamic matrix representation have also been utilized to increase diversification. The proposed PSO model

Acknowledgements

We appreciate the support for this research received from the European Union (EU) sponsored (Erasmus Mundus) cLINK (Centre of excellence for Learning, Innovation, Networking and Knowledge) project (EU Grant No. 2645).

References (70)

  • P. Moradi et al.

    A hybrid particle swarm optimization for feature subset selection by integrating a novel local search strategy

    Appl. Soft Comput.

    (2016)
  • R. Sheikhpour et al.

    Particle swarm optimization for bandwidth determination and feature selection of kernel density estimation based classifiers in diagnosis of breast cancer

    Appl. Soft Comput.

    (2016)
  • W.-D. Chang

    A modified particle swarm optimization with multiple subpopulations for multimodal function optimization problems

    Appl. Soft Comput.

    (2015)
  • Y. Zhang et al.

    Intelligent affect regression for bodily expressions using hybrid particle swarm optimization and adaptive ensembles

    Expert Syst. Appl.

    (2015)
  • W. Srisukkham et al.

    Intelligent leukaemia diagnosis with bare-bones PSO based feature optimization

    Appl. Soft Comput.

    (2017)
  • L. Zhang et al.

    Feature selection using firefly optimization for classification and regression models

    Decis. Support Syst.

    (2018)
  • L. Zhang et al.

    Classifier ensemble reduction using a modified firefly algorithm: an empirical evaluation

    Expert Syst. Appl.

    (2018)
  • R. Shang et al.

    Subspace learning-based graph regularized feature selection

    Knowl.-Based Syst.

    (2016)
  • R. Shang et al.

    Self-representation based dual-graph regularized feature selection clustering

    Neurocomputing

    (2016)
  • R. Shang et al.

    Global discriminative-based nonnegative spectral clustering

    Pattern Recognit.

    (2016)
  • L. Yang et al.

    Coupled compressed sensing inspired sparse spatial-spectral LSSVM for hyperspectral image classification

    Knowl.-Based Syst.

    (2015)
  • C. Beyan et al.

    Classifying imbalanced data sets using similarity based hierarchical decomposition

    Pattern Recognit.

    (2015)
  • S. Mirjalili

    Moth-flame optimization algorithm: a novel nature-inspired heuristic paradigm

    Knowl.-Based Syst.

    (2015)
  • D. Karaboga et al.

    On the performance of artificial bee colony (ABC) algorithm

    Appl. Soft Comput.

    (2008)
  • J. Derrac et al.

    A practical tutorial on the use of nonparametric statistical tests as a methodology for comparing evolutionary and swarm intelligence algorithms

    Swarm Evol. Comput.

    (2011)
  • A. Pennisi et al.

    Skin lesion image segmentation using Delaunay Triangulation for melanoma detection

    Comput. Med. Imaging Graph.

    (2016)
  • D. Farid et al.

    An adaptive ensemble classifier for mining concept-drifting data streams

    Expert Syst. Appl.

    (2013)
  • P. Kinghorn et al.

    A region-based image caption generator with refined descriptions

    Neurocomputing

    (2018)
  • L. Zhang et al.

    Intelligent facial emotion recognition using moth-firefly optimization

    Knowl.-Based Syst.

    (2016)
  • Y. Zhang et al.

    Adaptive 3D facial action intensity estimation and emotion recognition

    Expert Syst. Appl.

    (2015)
  • A. Esteva et al.

    Dermatologist-level classification of skin cancer with deep neural networks

    Nature

    (2017)
  • Q.U. Ain et al.

    Genetic programming for skin cancer detection in dermoscopic images

  • O. Abuzaghleh et al.

    Noninvasive real-time automated skin lesion analysis system for melanoma early detection and prevention

    IEEE J. Transl. Eng. Health Med.

    (2015)
  • F. Xie et al.

    Melanoma classification on dermoscopy images using a neural network ensemble model

    IEEE Trans. Med. Imaging

    (2017)
  • K. Shimizu et al.

    Four-class classification of skin lesions with task decomposition strategy

    IEEE Trans. Biomed. Eng.

    (2015)
  • Cited by (0)

    View full text