Classification and Detection of Skin Lesions Through Machine Learning Methods

Abstract

Early detection of skin problems aids in the detection and treatment of disease conditions at an early stage before the disease progresses. Due to recent advances in computer vision and machine learning, researchers are developing tools that can learn from image-based data and detect these lesions autonomously. We present the top studies that are related to the automation of skin lesion detection, segmentation, and classification in this report. We begin with identifying the drawbacks of visual inspections performed by humans and the reasons why health systems must adopt automatic and reliable analysis. Then we review the most recent methods of classifying skin lesions, and how to differentiate between a type of lesion on dermoscope images, ordinary photographs, and other images. We compare both advantages and disadvantages of both pathways, classic machine-learning rules and modern deep networks. Next, the issues of noticing and describing lesions appear, with an emphasis on methods that create definite boundaries to measure their size and shape. Some main approaches to segmentation are outlined together with the challenges they encounter in the world, the common data repositories relied upon by testers, and the grading metrics. These techniques range from deep-learning models, graph tricks, and region-growing schemes. We also stroll through the large databases, benchmark activities, and evaluation criteria that all people use within the sphere. The HAM10000 dataset that was publicly announced is utilized to analyze the skin lesions in this paper, and it provides an accuracy of 97.86% on a proposed Convolutional Neural Network (CNN) model. To spur skin-lesion work in dermatology, we conclude by describing current trends, persistent obstacles, and potentially fruitful research directions.

Keywords: Skin Cancer, Machine Learning, Convolutional Neural Network