Study On The Segmentation Algorithm Of Dermoscopic Image Based On Deep Learning

Skin pigmentation is one of the skin diseases,which is usually caused by the increase or decrease of pigmentation.Melanoma is the most serious maligmant tumor in skin pigmented lesions.It is easy to be confused with other benign skin pigmented diseases in the early stage,which has delayed the optimal treatment period.If there is no corresponding professional knowledge,it's difficult to distinguish skin pigmented lesions from maligmant ones directly by appearance.Dermoscopic image can provide more color and texture information.At present,doctors make clinical diagnosis through dermoscope image,but this way depends on the professional ability of doctors,and there are some misdiagnoses.The construction of computer-aided diagnosis system for skin diseases can improve the efficiency and accuracy of detection.A more precise segmentation of skin lesions in dermatoscope images can help doctors and systems make better decisions,so as to help patients strive for valuable treatment time.The main contents of this thesis are as follows:(1)In this thesis,a “U” shaped network structure is proposed,which uses ResNet network structure as encoder structure,and uses pyramid pooling module to connect encoder and decoder.The module integrates local information and global information,and improves the sensitivity of the model to different scale features.In this thesis,the swish activation function is introduced.After the replacement of the ReLU activation function,the segmentation performance of the model is improved.Finally,in order to improve the prediction ability of the model at the boundary,this thesis designs a loss function to punish the boundary.It makes the pixels of the boundary get higher weight,and makes the model more accurate in the boundary region segmentation.The Jaccard index value of 0.7512 was obtained on the ISIC2017 dataset by using the model proposed in this thesis and related improvements.(2)In order to further improve the segmentation performance of the model,this thesis proposes corresponding improvement methods in three aspects: preprocessing,data enhancement and learning rate strategy.For the multi-source of isic2017 skin mirror image data set,this thesis uses shadows of grays algorithm to preprocess the color constancy,so that the overall color of the image tends to be consistent.Aiming at the problem that some skin mirror images in the data set are not labeled strictly based on pixels,but have "generality",this thesis proposes a data enhancement method basedon super pixels,which uses linear difference between the original image and its super pixel image to fuse,control the complexity of the model,and improve the generalization ability of the model.In order to make the model converge to a stable minimum value,this thesis uses the periodic cosine learning rate decay strategy.Through periodic restart,the model finds a more stable minimum value under different data set distribution,and improves the generalization ability of the model.In this thesis,the best single model obtained in Chapter 3 is used for experiments.After using the above three methods,the Jaccard index of the base model is increased by 0.89% on the ISIC2017 dataset,and the optimal Jaccard index value of 0.7601 is obtained.