Full Convolutional Neural Network Based Dermoscopy Image Segmentation Algorithm

Medical image is an effective media that reflects the structure of biological tissue and of great significance when clinical diagnosis and treatment.As an significant part of medical image analysis,medical image segmentation technology is widely applied to the fields of tumor diagnosis,lesion location and lesion volume measurement,etc.The image segmentation results directly affect the formulation of follow-up disease diagnosis and treatment plans.As a field of medical image segmentation,image segmentation of skin lesions is of great significance for dermatologists to diagnose skin diseases.Deep learning has become a pop-ular research field in the last several years.Image processing algorithms have been widely applied to automatic driving,image recognition,video surveillance and medical image pro-cessing,and achieved remarkable results.Inspired by this,and considering the difficulties in image segmentation of skin lesions,this thesis proposed two deep learning network mod-els based on deep learning technology.Firstly,this thesis proposed an AGU-net model to segment skin lesions images by introducing attention mechanism on the basis of U-net.Sec-ondly,the ASPP module is further introduced to improve the extraction ability of multi-scale context and the segmentation accuracy.This thesis mainly completed the following work:Firstly,the first chapter of this thesis summarizes the background of current medical image segmentation technology and the traditional image segmentation algorithms,and analyze the problems existing in the application of the traditional image segmentation algorithm in medical image field.Secondly,the second chapter in this thesis introduces the development history and related concepts of the current popular deep learning technology,and introduce the common deep learning model and the main practical applications.Then,on the basis of deep learning technology,this thesis proposes an AGU-net model for skin lesion image segmentation.The AGU-net model proposed in this thesis is based on U-net model with more improvements.Considering the defects of skip connection in U-net,this model introduces attention mechanism in the skip connection part,assigns different weights to the features of different positions to indicate the importance of features,in order to enhance the feature expression ability of the U-net algorithm model and improve the segmentation performance.Finally,considering the impact of multi-scale context information on the image segmentation performance,this thesis introduces Atrous Spatial Pyramid Pooling(ASPP)module on the basis of AGU-net to improve the ability to extract information in multi-scale context.