Classification of skin lesion with hair and artifacts removal using black-hat morphology and total variation

Abstract

Automatic classification of skin lesion plays vital role in the diagnosis of actual skin cancer type. This classification process requires spatial features information of the skin lesion but dermoscopic images are usually occluded with hair and other artifacts such as shadows and markers etc. This occlusion can affect the classification process which may lead to erroneous diagnosis of skin cancer. In this research an efficient method to enhance dermoscopic images by removing hair and other artifacts using black-hat morphological processing and total variation inpainting technique is proposed. Additionally, to show the impact of proposed enhancement of dermoscopic images, a technique is proposed in an effort to achieve results for skin lesion classification comparable to deep neural networks with as low cost as in Conv 2D by performing two dimensional convolution on images. This system passes through three convolution streams to comprehensively cater information. The proposed model is evaluated on a public Skin Lesion dataset which contains 2000 images. Results depict the improvement in classification accuracy of three skin cancer classes which are Melanoma, Nevus and Seborrheic Keratosis (SK), when hair and artifacts are eliminated by proposed method.