Research On Classification Of Melanoma Based On Deep Learning

Among many skin lesions,melanoma is the most common and deadly skin cancer,and it poses a serious threat to human health.An effective method for treating melanoma is to use dermo-scopicoscopy for early diagnosis,but the melanima diagnosis of traditional visual observation is easily influenced by doctor’s personal subjective experience,and such methods easily lead to misdiagnosis and missed diagnosis.Therefore,using computer technology to improve the diagnostic accuracy of melanoma,as a basis for assisting doctors to conduct early screening of melanoma has very important research value and significance.In the study of the melanoma classification task,it was found that there were some problems such as insufficient training samples,small discrimination between different categories and data imbalance.To solve those problems,using deep learning-based methods such as data augmentation,convolutional neural network,generative adversarial network(GAN)and ensemble learning,the learned classification model could more accurately distinguish melanoma.In this thesis,some main study works and contributions are as follows:(1)To solve the problem of insufficient training sample data,based on improving other methods,two data augmentation methods,random masking and non-random masking,are proposed.Besides,a true and false sample fusion method is proposed by utilizing generative adversarial network(GAN)to generate fake training samples,which are fused with true samples.These methods can effectively improve the model classification effect on the melanoma classification task.(2)To deal with the problems of small discrimination between different categories in melanoma classification tasks,the ResNet-50 is selected to extract features by a series of comparative experiments.The experimental results show that the ResNet-50 is more suitable for melanoma classification than other networks,and it can effectively improves the classification effect.(3)For the data imbalance problem,an ensemble learning method is put forward.The above proposed data augmentation methods,such as random masking,non-random masking,true and false sample fusion,are used to construct a training set with differences.Then,based on convolutional neural network,multiple classification base models are constructed and integrated.The obtained ensemble model can abruptly improve the overall performance in the melanoma classification task.A series of experiments are conducted and the results show that deep learning technology can effectively improve the performance of the model in melanoma classification tasks,the average accuracy value reaches 0.691,which is higher than the challenge champion of the ISIC2016 by nearly 6%.Compared with the results of the ISIC2017 challenge,under the conditions of less than twice training data,the classification results can be ranked the fifth in the challenge,which is very competitive.It means that the proposed method can cope with many challenges in melanoma classification and build a good foundation for constructing a high-accuracy melanoma diagnosis system.