Deep Convolutional Neural Network for Skin Lesion Detection with HOG and GLCM features

Abstract

Lesion relates to the unnatural growth of tissues in which skin lesions are most common. It may have two types: benign and malignant. Malignant lesion has the highest death rate compared to benign lesion due to their major metastases. Metastasis is nothing but the capacity to enter distant organs. There are a variety of diagnostic features to differentiate both. The most popular approach is the histological approach through which the impacted tissues are taken and examined underneath a microscope. As it is an invasive technique and the chances of spreading to other natural underlying structures, other diagnostic approaches have come into play. This work is one of them, which is a non-invasive technique and is detected by a computer vision algorithm. To determine where such a lesion is malignant or benign, the extraction feature utilizes a mixture of both the Histograms of oriented gradients (HOG) and the Grayscale Level Co-occurrence Matrix (GLCM) functions. Which are applied to the Convolutional neural network. Convolutional neural network (CNN) obtained an average accuracy of 98.32 percent on a dataset of 4100 images, which is greater than the existing method.