Research On Lesion Attributes Segmentation Method For Skin Images

Melanoma is the most common malignant tumor in the skin disease,which originates from melanocytes and results from malignant lesions of melanocytes.In recent years,the incidence of melanoma has been increasing in the Chinese population,and early melanoma can be cured by surgical resection,therefore,accurate diagnosis of early melanoma is of great importance to guide clinical treatment.However,the early symptoms of malignant melanoma are not obvious and are difficult to diagnose directly using eye observation.Dermoscopy is widely used to diagnose melanoma.Dermoscopy images can magnify the clarity of skin lesions and provide more information on texture and color,allowing the observation of superficial skin structures that are difficult to observe directly with the eye.Similarly,it is important to identify specific lesion properties for the accurate diagnosis of melanoma.Lesion properties represent specific visual patterns in skin lesions,and each property has its own specific meaning.The five most commonly used clinical attributes are Pigment Network,Streaks,Negative Network,Milia-like Cyst,and Globules.The type of lesion as well as the degree of malignancy can be determined based on the lesion attributes.Computer-aided diagnosis plays an increasingly important role in clinical practice,and by accurately segmenting the properties of skin lesions can better help physicians accurately diagnose melanoma.In this dissertation,we study and analyze the existing dermoscopic image lesion attribute segmentation techniques and construct a new skin lesion attribute segmentation method,the research work is summarized as follows:(1)Most of the existing segmentation methods for dermoscopic images usually segment lesion areas,however,lesion attributes are extremely useful to assist physicians in early diagnosis of melanoma.In this dissertation,we propose a multi-task U-Net based attribute segmentation method for dermoscopy images,a multi-task U-Net model is designed,in which a Channel Context Feature Fusion Module(CCM)and a Dual-domain Attention Module(DA)are introduced.Attention Module(DA),which is called the CDU-Net model in this dissertation.We first fuse channel context features for dermoscopy image features and propose the CCM module,which uses dilated convolution with different dilated rates and stacks them in cascade to increase the receptive field without losing spatial information,and further extracts multi-level features through different scale receptive fields,which helps to reduce the loss of boundary information.In order to obtain more effective information from the feature map,the DA module is proposed.This module consists of spatial domain and channel domain together to realize the information interaction between spatial features and channel features,so that the model has the ability to filter the learning of features,improve the efficiency of network feature extraction,and then improve the accuracy of network segmentation.(2)In order to improve the CDU-Net model for streak attribute segmentation with less data and to further improve the accuracy of melanoma attribute segmentation,this chapter further proposes CDU-Net-SGAN,an optimization method for lesion attribute segmentation of dermoscopy images,using the game idea of generative adversarial networks.We first designed a discriminator network structure based on statistical features to make higher-order constraints on the CDU-Net,a segmentation network of skin lesion region attributes,and trained the generator and discriminator in the model several times alternately until the discriminator could not easily distinguish between the segmentation label map and the output of the segmentation network.After adversarial training,we can obtain an attribute segmentation network with better segmentation effect.Experimental results prove that in the ISIC2018 Task2 dataset,the method proposed in this paper effectively segmented the lesion attributes.Compared with the existing dermoscopic image lesion attribute segmentation algorithms,it has achieved better segmentation results and has higher clinical value.