A Novel Sep-Unet Architecture of Convolutional Neural Networks to Improve Dermoscopic Image Segmentation by Training Parameters Reduction

Abstract

Nowadays, we use dermoscopic images as one of the imaging methods in diagnosis of skin lesions such as skin cancer. But due to the noise and other problems, including hair artifacts around the lesion, this issue requires automatic and reliable segmentation methods. The diversity in the color and structure of the skin lesions is a challenging reason for automatic skin lesion segmentation. In this study, we used convolutional neural networks (CNN) as an efficient method for dermoscopic image segmentation. The main goal of this research is to recommend a novel architecture of deep neural networks for the injured lesion in dermoscopic images which has been improved by the convolutional layers based on the separable layers. By convolutional layers and the specific operations on the kernel of them, the velocity of the algorithm increases and the training parameters decrease. Additionally, we used a suitable preprocessing method to enter the images into the neural network. Suitable structure of the convolutional layers, separable convolutional layers and transposed convolution in the down sampling and up sampling parts, have made the structure of the mentioned neural network. This algorithm is named Sep-unet and could segment the images with 98% dice coefficient.