Parametric Examination on Optimized Deep Learning based Melanoma Detection

The occurrence of tumors in the skin grows gradually. In the early stage, examination by an experienced dermatologist will reliably identify malignant skin tumors; however, extensive screening by dermatologists is not feasible for each single patient. Many studies have been undergone to improve dermatologist-level, computer-aided diagnostics are underway to overcome this challenge. After the advent of deep-learning technology, the detection mechanism has become significant in the medical field. This paper intends to suggest a new paradigm for the diagnosis of melanoma, which comprises of four main stages. (i) pre-processing (ii) segmentation (iii) extraction of features and (iv) classification. Pre-processing is mainly carried out, where the input image is susceptible to smoothening and resizing of the edges. Segmentation is subsequently performed through the Otsu thresholding mechanism. The GLCM features along with the Higher Order Uniform Moment Induced- Local Binary Patterns (HOSMI-LBP) features are extracted. The combination of Convolutional Neural Network (CNN) and Neural Network (NN) architectures is used to define the classification of melanoma as either benign or malignant. Two steps of classification is followed in this research work: NN is trained with the derived features; parallely, the input image is categorized explicitly using the optimized CNN framework. Finally, the optimized CNN and NN graded outputs will be averaged and the final result is considered to be the detected output. In specific, the weight and initial rate of CNN is optimized using the Sea Lion Integrated Grey Wolf Algorithm (SLI-GWO) method. To undergo the analysis on the algorithmic performance, this paper makes the parametric analysis of the proposed work is by setting the values of r
1
= 0.2 and r
2
= 0.4 of SLI-GWO