CNN supported framework for automatic extraction and evaluation of dermoscopy images

Abstract

Skin Cancer is one of the acute diseases listed under top 5 groups in 2020 report of World Health Organisation. This research aims to propose a Convolutional Neural Network framework to extract and evaluate the suspicious skin region. This framework consists following phases; (i) Image collection and resizing, (ii) Suspicious skin section extraction using VGG-UNet, (iii) Deep-feature extraction, (iv) Handcrafted features mining from the suspicious skin section, (v) serial feature integration, and (vi) Classifier training and validation. This research considered dermoscopy images of International Skin Imaging Collaboration benchmark dataset for the experimental assessment and the result of the proposed framework is separately analysed for segmentation and classification tasks. In this work, benign and malignant class images are considered for the examination and during the classification task, integration of the deep and handcrafted features are considered. The experimental results of this study present a segmentation accuracy of > 98% with UNet and a classification accuracy of > 98% with VGG16 combined with Random Forest classifier.