Edge Based Skin Cancer Decision Support System Using Machine Learning Algorithms

Skin lesions has grown rapidly in recent years. Detecting and treating skin lesions in its early stages can be extremely beneficial to its recovery. Using Digital image processing (DIP) to extract skin diseases from lesions emerged rapidly due to in accessibility of dermatologist and laboratory facilities in remote areas specifically in under developed countries. Out of many deep learning based feature extraction algorithms, convolutional neural networks (CNN) outperform along with different classifiers. In this study, we focused on edge based classifiers for robust skin lesions classification. To choose the best combination of classifier with existing CNN based feature extraction methods we used HAM-10000 dataset, supervised by ISIC. Different mobile based classifiers such as QDA, KNN and SVM, along with their different hybrid versions were tested in same conditions. Experimental results shows that the cubic kernel function used in the cubic SVM approach allows for more complex decision boundaries to be created for melanoma detection, as the characteristics of melanoma lesions can vary greatly. The cubic kernel function enables the SVM model to capture these complex patterns and make accurate predictions using less computational resources as compared to other mobile based classifiers. Cubic SVM shows the accuracy of 98% while using only 20 seconds of time.