SCDC-Net: Skin Cancer Detection and Classification using Deep Q Neural Network with Hybrid Features

Abstract

Skin cancer is one of the most severe malignancies that may affect individuals of all ages. It can affect people of all races and ethnicities. Thus, early detection and treatment of skin cancer have the potential to save millions of lives. Because of the high degree of similarity between skin lesions, conventional machine learning approaches failed to provide the highest possible detection and classification accuracy in the skin cancer detection and classification (SCDC) system. To achieve the robust performance, this work is presented a SCDC-Net by employing deep learning, transfer learning models with hybrid features. Initially, UG-Net based transfer learning model is developed by combining generative adversarial networks (GAN) and U-Net. Here, UG-Net is used to remove hair from skin lesions and also preprocess enhances lesion. In addition, Hybrid U-Net (HU-Net) is used to localize the disease effected area of skin cancer. Further, various hybrid features are extracted from the segmented output by employing the gray-level cooccurrence matrix (GLCM) based texture characteristics, the discrete wavelet transform (DWT)-based low-level features, and the computation of image statistics for the texture, low-level, and colour features, respectively. In order to ensure the highest possible performance of the system, a deep q neural network (DQNN) was constructed for the classification of skin cancer using GLCM, DWT, and statistical colour features. The simulation results shows that the proposed method resulted in superior subjective and objective performance as compared to state of art approaches on ISIC-2019 public challenge dataset.