Computer-aided Diagnosis Of Multi-source Skin Diseases Images

Malignant melanoma is a malignant cancer with a high degree of malignancy.It has the characteristics of rapid development,rapid metastasis and easy distal metastasis,and it’s incidence is increasing year by year.Early detection and early treatment are the best ways to reduce the harm and mortality of melanoma.However,the early stage of melanoma is hidden and difficult to detect,and melanoma has the characteristics of intra-class dissimilarity and similarity with other skin diseases.As a result,the number of doctors who have the ability to diagnose melanoma grows very slowly,which makes it difficult to match the number of patients with melanoma.In recent years,the rapid development of network and digital image processing technology has provided new ideas for the diagnosis of melanoma.In addition,the development of mobile medical concept provides a basis for the realization of a new diagnostic technology which is non-invasive,simple,accurate,quick to obtain results and can be widely applied.This paper mainly studies the application of new methods such as in-depth learning,integrated learning and transfer learning in computer-aided diagnosis of malignant melanoma.The main research contents and achievements include:1.The current mainstream convolutional neural network was compared and studied.The characteristics of melanoma dermascopic images were explored experimentally.The classification effect of melanoma images was correlated with network structure and network depth,and summarize the selection principles and standards of Dermatology image used in depth neural network.2.Combining the prior knowledge of clinical diagnosis of melanoma,TMME algorithm and ensemble learning method,the DFEC-Net model(Depth-featureensemble-convolution-net deep feature integrated network model)is proposed.Firstly,features are extracted based on prior knowledge and feature maps of boundary,color and texture are constructed.Then,deep learning network ensemble is realized based on ensemble learning idea.Experiments show that the model can significantly improve the sensitivity of malignant melanoma with the same data set.3.The case transferability,model transferability and feature transferability of multi-source skin disease images with different resolution collected by camera/mobilephone and dermoscope are studied.A Transfer strategy is designed according to the features of multi-source skin disease images,which can significantly improve the classification and recognition efficiency of low-resolution small data sets.It provides a new way to solve the problem of medical image small data set and low resolution.