Attention Synergy Convolutional Neural Network For Skin Lesion Segmentation

Skin diseases exist in all regions of the world,occur at all ages,and affect the health of all human beings.Skin disease is not harmful in the early stage,but once it is allowed to develop into skin cancer in the late stage,the death rate will increase rapidly.Thus,the early screening and diagnosis of skin diseases are very important.At present,the diagnosis of skin diseases mainly relies on the naked eye observation of dermatologists.They make judgments by analyzing the color,shape,and other features of skin lesions in dermoscopy images.However,due to the differences in clinical experience and technical level among different dermatologists,this approach is highly uncertain and heavily influenced by subjectivity.Therefore,it is the future trend to develop computer-aided diagnosis technology to help dermatologists diagnose skin diseases objectively and efficiently.Accurate skin lesion segmentation in dermoscopy images is the premise and guarantee for the diagnosis of skin diseases.However,it remains a challenging task due to the fuzzy lesion boundary,the irregular lesion shape,and the existence of various interference factors.Focusing on the above difficulties and challenges,this paper has carried out a series of researches by applying deep learning technology.The main contributions are summarized as follows:1)We propose an Attention Synergy Network(AS-Net)for skin lesion segmentation,which aggregates spatial and channel attention modules to enhance the feature representation ability of the neural network.Specifically,the spatial attention module enhances the lesions in the image and reduces the background noise.The channel attention module selectively enhances discriminant characteristic channels and suppresses less relevant channels.Besides,we further design a synergy module that adaptively adjusts the contributions of the two attention modules to obtain more accurate segmentation results.2)We propose a weighted cross entropy loss function for dermoscopy image segmentation,which can alleviate the problem of different sizes of skin lesions.It guides AS-Net to pay more attention to the skin lesions by increasing the weight of loss to the target,which could achieve ideal segmentation results even in dermoscopy images with a small target proportion.3)We conduct extensive experiments to evaluate the proposed encoder network,attention modules,synergy module,and loss function.Besides,we compare AS-Net with the current mainstream methods on three common datasets,including ISIC2017,ISIC2018,and PH2.The experimental results show that AS-Net outperforms most mainstream methods and achieves the highest scores in DIC and JAC on three different datasets.The segmentation visual effect on indistinguishable samples also indicates that AS-Net can overcome various interference factors in dermoscopy images and achieve accurate skin lesion segmentation.