Research On Classification Algorithms Of Skin Lesions In Dermoscopy Images

Melanoma is one of the skin cancers with the highest fatality rate.The number of patients is increasing every year,especially among adolescents.Fortunately,if detected and treated early,most patients can be cured.The research of automatic melanoma classification in dermoscopy images can help doctors improve the efficiency and accuracy of diagnosis and reduce patient mortality.With the application of convolutional neural networks,deep learning models have a huge advantage over traditional methods and have brought a huge breakthrough to the task of automatic classification of skin lesions.However,due to the diversity of imaging methods and clinical pathology,the skin lesion dermoscopy image have the similarity between the skin lesions and the differences within the category,the original image resolution is too large to make full use of the original image information,the foreground and background of the dataset image are not balanced,and there are still challenges in automatic skin lesion classification.Therefore,for common binary classification tasks of skin lesions,in order to solve the problem of inter-class similarity and intra-class variation between skin lesions,this thesis proposes a new deep learning classification method based on dense connection network,combined with attention mechanism and large margin loss.The network is composed of Dense Block,Transition Block,Attention Module,and Large Margin Loss.For the multi-classification tasks following new datasets in recent years,in order to solve the problems of inter-class similarity and intra-class variation,there are also problems such as image lesion-background imbalance and high image resolution.This thesis proposes a new deep convolutional neural network of skin lesion classification method based on multi-input and attention mechanism,this model uses multi-input strategy and combined attention mechanism to fully the use of image information enables the model to be more focused on the area with discriminative representation.This thesis proposes two novel attention methods based on densely connected networks to automatically classify skin lesions.Binary classification and multiclassification tasks were performed on the ISIC 2017 and HAM10000(ISIC 2018 training set)skin lesion datasets respectively.The experimental results prove the effectiveness of the two methods in the classification of skin lesions.