Article Preview
TopIntroduction
Skin cancer is the most common cancer found in white skin human population. Mutations in skin cell DNA is the main cause of Skin cancer (Apalla, Lallas, Lazaridou, & Ioannides, 2017). This disease can be divided into two main groups, non-melanoma and Melanoma type. Among the non-melanoma two types of skin cancers are mostly found, namely, basal cell carcinoma and squamous cell carcinoma. Melanoma is less frequent but it is more deadly than Non-melanoma. Melanoma is a painful and itchy pigmented, asymmetrical shaped, zigzag bordered, brownish skin lesion that is difficult to heal. The lesion changes its texture and scales up, turns up to a bleeding lesion in the final stage. It is also found that, if the Melanoma skin cancer is detected in later stages, the mortality rate increases (Gupta, Bharadwaj, & Mehrotra, 2016).
Dermatologist uses biopsy as the gold standard test for confirming malignancy in the skin lesion tissue. However, before recommending for biopsy, doctors use several semi quantitative scores on the appearance of the lesion. They are ABCD rule, 7-point checklist (Marks, 2000) etc. These scoring methods have reported detection accuracy around 80% (Morton & Mackie, 1998). Though skin biopsy is an accurate technique but at the same time it is invasive and costly. In the last one and half decades, due to the advancement in image processing, pattern recognition and machine learning, data scientists and researchers have done several work in the domain of computer aided Melanoma detection using skin lesion images. These techniques are non-invasive and reduce both time and cost but increase the efficiency in malignancy detection. In the related work section of this paper, discussions on several similar researches and practical case studies are given. It is found from the literature survey that skin cancer detection accuracy up to 87.5% (Abbes & Sellami, 2019) can be achieved with the University of Waterloo skin cancer dataset (Amelard, Glaister, Wong, & Clausi, 2015). So, scope of improvement is highly possible in terms of accuracy and other performance metric in this domain of research.
This present work gives a research direction for building a low cost but high performance skin cancer detection system using machine learning application. In the present work, several hand crafted image features are extracted from a well known Melanoma dataset (University of Waterloo). The extracted features from the image dataset are ranked by ReliefF and PCA on their relevance and best ranked features are used for classification using single hidden layer MLP. Using those features a 2-class (malignant versus benign) classification strategy is also devised by finding optimized parameter of two-layer MLP using metaheuristic algorithms. Two metaheuristic algorithms, PSO and SA are used to optimize the hidden layer neuron number of the MLP. The authors have selected these two methods as they belong to two different categories of metaheuristic, SA does exploration of the whole search space whereas PSO does concerted exploitation (Wang, Zhang, & Ji, 2015) of the search space. PSO is a simple, computationally efficient, fast convergent, collaborative population-based global searching algorithm with adjustable parameters in the other hand SA is also simple, computationally efficient, physics inspired and relatively free from local optima with good convergent property which is not population based (Beheshti, & Shamsuddin, 2013). The classification results are compared with single layer MLP using best ranked features and it is found that two-layer MLP with optimized neuron gives better result.
The paper is arranged as follows. In the second section some of the works related to the classification of the present image database are discussed. In the third and fourth section research gap found from the literature survey and problem formulation are given respectively. In the fifth section information about several available Melanoma dataset is discussed. In the sixth and seventh sections, details about the feature used in the work and work flow including discussion about different tools used are shown. The eighth section is for result and analysis and finally conclusion of the work is stated in the last section.