Deep Neural Networks For Accurate Skin Disease Segmentation And Classification

Abstract

In recent years, the skin disease detection and classification are considered as the essential topic to identify the affected people. In literature review, cycle-consistent Generative Adversarial Network (cycle-GAN) is analyzed with the consideration of two step progressive transfer learning and domain adaptation. This cycle-GAN is mainly utilized to pre-trained the images by fully supervised Deep Convolutional Neural Network (DCNN) which is utilized to skin disease classification. This DCNN is not an efficient method for skin images. In this paper, modified SegNet is developed to segment the images during training period and it is augmented with the consideration of cycle-GAN method.  This method operates the dilated convolution operation in its place of general convolution to normally extract the multi-scale contextual features without considering resolution. This extracted feature of multi-scale high resolution is encoded with the assistance of encoder and send to the decoder model. After that, the dropout layer with the addition of Dynamic Conditional Random Fields (DCRFs) to reduce the overfitting issue. Additionally, the dropout layer is defined as the segmented skin images. This segmented image is sent to the ResNet18 to type classification of skin diseases. Hence, this proposed model is defined as segmentation and classification (SegClassNet) model. At lase, the outcomes show that the projected technique attains the mean accuracy of 91.28% for HAM image dataset contrasted to different conventional classification methods.