Dermoscopy Image Detection Based On Integrated Convolutional Neural Networks

Skin cancer has become one of the cancers with high prevalence,and the number of patients worldwide is increasing every year.Melanoma is a common skin cancer with high mortality.The most effective way to treat melanoma is early detection and early treatment.At present,in the field of melanoma detection,it mainly depends on the manual diagnosis of dermatologists by observing the dermoscopic image.Because there are many kinds of melanoma shape,color and texture,and the appearance of malignant melanoma and benign melanoma is very similar,the traditional manual diagnosis has low accuracy,time-consuming and laborious,which hinders the diagnosis of melanoma.With the development of deep learning technology,convolutional neural network model has brought a historic breakthrough to the detection of dermoscopic images,and plays a great role in improving the diagnostic accuracy and work efficiency.In order to solve the problems of intra class difference,inter class similarity and data set imbalance of melanoma and improve the accuracy of the detection system,a dermatoscopy image detection method based on integrated convolution neural network is proposed in this paper.The main research work is as follows:Firstly,a skin mirror image segmentation method based on attention residual block U-Net(ARB-UNet)is studied.This method introduces the convolutional block attention module(CBAM),applies the CBAM module to the "skip connection" of the original u-net model and the expanded residual networks(DRB)module to obtain a new attention residual mechanism(ARB),and selects focal tversky loss as the loss function of the model,experiments show that ARB-UNet can improve the performance of image segmentation.Secondly,the feature extraction of dermoscopy image is studied.The texture feature model and color feature model are mainly studied to represent the features of dermoscopy images.The experimental results show that the classification model based on SIFT texture feature has better classification performance.Finally,a skin mirror image classification method based on integrated fine-tuning convolution neural network is discussed.By reconstructing the full connection layer of the three pretraining models Xception,Res Net50 and Vgg-16,the dropout layer and batch normalization(BN)are introduced to improve the generalization ability of the network model.The pretraining model is transferred and learned on the ISIC2016 skin data set,and the model output is weighted and fused to improve the training accuracy of small samples.The effectiveness and feasibility of the proposed algorithm are verified by experiments.In conclusion,all the research work in this project,which combines deep learning with image medical field,and explores dermoscopy image segmentation,feature extraction and recognition technology based on integrated convolutional neural network,lays the necessary theoretical and experimental foundation for the application of machine learning algorithm in the field of medical diagnosis of skin cancer.