Skin lesion classification using decision trees and random forest algorithms

Abstract

Any superficial skin growth that does not resemble the surrounding area is referred to as skin lesion. It can occur in the form of mole, bump, cyst, rash or other changes that can be classified either as primary or secondary lesion. While primary skin lesions correspond to those changes in color or texture, secondary lesions occur as a primary lesion progression. Skin lesion image segmentation and classification at the early stages can help the patients recover through proper medication and treatment. Many algorithms for segmentation and classification are available in the literature but they all fail to extract lesion boundaries perfectly and classify them with more accuracy. To improve the reliability of the skin image segmentation and classification, we propose to use decision trees and random forest algorithms in this works and compare them with different data sets. The proposed method can generate high-resolution feature maps that can help to preserve the spatial details of the image. While tested against the ISIC 2017 and HAM10000 dataset, we found that the proposed method is more accurate as compared to the existing algorithms in this domain and is also very robust to artifacts or hair fibers present in the skin images.