Research On Automatic Segmentation Method Of Skin Lesion Image Based On Deep Learning

Skin cancer is a stubborn disease that threatens human life safety,with strong concealment and complex pathology,and the number of cases is increasing year by year.Clinicians rely on their own experience to make judgments about the onset status,which may lead to misdiagnosis.Computer assisted diagnosis technology has the objective and fast characteristics of dividing lesion areas and normal tissues through skin disease image segmentation technology,providing clear disease boundaries,improving the probability of doctors accurately judging pathological types,and reducing the risk of cancer infection for patients.However,the task of skin lesion image segmentation is challenging,as skin lesion images have problems such as a small proportion of key information,excessive background interference information,and blurred lesion edges,making it difficult to accurately perform the segmentation task.In response to the above issues,this article proposes three different skin disease image segmentation methods,with the main work as follows:(1)This article elaborates on the research background and significance of skin disease image segmentation methods,and investigates the mainstream algorithms and improvement plans of domestic and foreign scholars in skin disease image segmentation technology.Introduced the theory of deep learning in the field of segmentation,including basic structures and classic frameworks,and briefly introduced the image segmentation process.(2)Due to the small proportion of key information in skin disease images,segmentation networks are prone to insufficient extraction capabilities.To address this issue,two sets of U-Net structures with structural differentiation were used in series,and a coordinate fusion dual U-shaped skin disease segmentation network was designed.Introduce blueprint convolution at the decoder end of the dual U-shaped network model to improve the training speed of the network.And a coordinate fusion gate was constructed through coordinate attention,which efficiently locates features through the spatiality of feature maps,allowing the transmitted features to be reused on the decoder,promoting deep and sufficient feature fusion in the network.(3)To filter out background noise and reduce the impact of boundary blur,a novel multi-level split convolution module is proposed to lock lesion contours from images of different resolution sizes.Embedding it into the decoder of the skin disease segmentation network is called a multilevel split skin disease segmentation network.The network filters the spatial information of the feature map and enhances the Receptive field of the convolution network through the convolution of holes of different sizes and multi-scale channels.Through ablation experiments,it was verified that the multi-level splitting convolution module can effectively improve the segmentation performance of the model.(4)A skin disease segmentation network based on CNN Transformer dual encoder was designed,which effectively integrates deep and shallow skin disease features through the interaction of encoder information.The use of Transformer can fully extract features from segmented feature maps,which can enhance the sensitivity of the network to key features.By deploying weighted channels on the network bottleneck layer,the ability to filter underlying features is improved.The experimental results show that the segmentation algorithm proposed for skin lesion images is superior to other advanced model algorithms,achieving the expected segmentation effect.