| In recent years,the incidence of skin cancer,such as melanoma,has been increasing,seriously endangering human’s health and safety.Early detection and treatment of such diseases is an important way to improve the quality of life of patients.Due to the advantages of non-invasiveness,simplicity and improved detection accuracy,dermoscopy has been widely used in the detection of skin lesions including melanoma.However,even professional dermatology clinicians may lead to missed detection and misdiagnosis when making manual diagnosis.Using computer to assist doctors in dermoscopy image segmentation can improve the accuracy and efficiency of skin lesions diagnosis,which has great significance for improving the survival rate of patients.Nowadays,with the development of deep learning,traditional image segmentation algorithms based on manually extracted features are gradually being replaced.Therefore,in this paper,we designed two models based on convolutional neural networks to improve the accuracy of lesion region segmentation in dermoscopy images.The main research contents of this paper are as follows:1.In view of the segmentation difficulties in dermatoscopic images,such as large color and scale changes of lesions,irregular boundary shapes and external interference,we designed a convolutional neural network model named Multi-Scale Attention U-Net based on multi-scale context information,which is aiming at accurate segmentation of lesion areas in dermoscopic images.The network uses U-Net as the baseline model,and the encoding stage builds image pyramids as multi-scale input,which increases the ability of network feature transfer.An antufocus parallel atrous convolution module is embedded at the bottom of the encoder for adaptive extraction and fusion the multi-scale contextual information.In the decoding stage,channel and spatial attention modules are embedded to further improve the feature representation capability of the network.Deep supervision of multiple layers in the decoder,combined with the channel attention module,are used to get the weights of different scale feature maps,also help the network to learn more recognizable features.We verified the effectiveness and accuracy of the network model on the ISIC2017 and ISIC2018 datasets.2.Most network models only focus on the whole lesion area and ignore the boundary information in the skin lesion segmentation task,but both region and boundary are the key to improve the segmentation effect.In this paper,we designed a convolutional neural network model named Context Reverse Attention Network based on reverse attention.The network model is based on the idea of from coarse to fine,which guides the network to focus on mining detailed information such as boundaries.Res Net is used as an encoder to extract feature information of multiple levels,and a rough high-level semantic global map is obtained after restoring the resolution through the decoder,which is used as a rough map for subsequent steps.Context fusion module was designed to fuse low and high features and obtain multi-scale context information,which was then fed into a cascade of reverse attention modules to gradually pay attention to the boundary details of the lesion area,so as to calibrate the inconsistent predictions in the rough map.We verified the accuracy and excellence of the network model on the ISIC2017 and ISIC2018 datasets.In summary,in view of the current problems and challenges in the segmentation of lesions in dermoscopy images,We designed two network models based on convolutional neural networks,and proved the effectiveness of the two network models on the dermoscopy image dataset.. |
