Integration of handcrafted and deep neural features for Melanoma classification

Abstract

Deep neural networks (DNNs) are widely utilized to automate medical image in- terpretation in many forms of cancer diagnosis and to support medical specialists with fast data processing. Although man-made characteristics have been used to diagnose since the 1990s, DNN is fairly new in this eld and has shown extremely promising results. The fundamental goal of this study is to detect melanoma cancer in its early stages by obtaining a remarkable outcome with greater accuracy. Our purpose is to address the problem of an increase in skin cancer patients throughout the world, as well as an exponential increase in the danger of mortality from not commencing the diagnosis at an early stage, as a result of late detection. We propose that the research works on handcrafted features and merges the result with deep learning approaches with the initial help with a huge dataset of raw images. The DNN model used in this research has multiple layers with various e ective lter- ing processes called batch normalization and dropout also with added layers named

atten and dense. In this process, images are classi ed to predict melanoma cancer at an early stage with Mean Shift, SIFT, and Gabor separately then the output was ensembled with later added Raw images results to give better accuracy. With an early integration model for separate featured databases and with a late and full integration model for ensemble with various results from the early integrated model we got our results. As a result, this neural network has provided an accuracy of 90% in early models and in late and full integration 86% and 84% respectfully, which is higher than other conventional approaches.