The Research And Implementation Of Dermatoscope Image Classification Algorithm Based On Ensemble Learning

Skin disease is one of the most common diseases,which brings challenges to clinical diagnosis due to its diversity and complex characteristics.Nowadays,medical workers use dermoscopy to obtain clearer images of skin lesions.Whether from the perspective of prevention,detection or treatment,the algorithm designed to distinguish the specific types of lesions in dermoscopy images meets the needs of today’s society.Besides,due to the good performance of ensemble learning in improving the ability of solving practical problems,this thesis proposes a dermoscopic image classification algorithm based on ensemble learning1.An ensemble classification algorithm of dermoscopy based on image features is proposed.Firstly,the dermoscopic image is preprocessed to balance the data set of the dermoscopic image,and the HSV(Hue,Saturation,Value)color space features are selected.A method based on hierarchical LBP(Local Binary Patterns)and GLCM(Graylevel Co-occurrence Matrix)is proposed to obtain the local and global texture features of the dermoscopic image.Combine with the color feature vector and texture feature vector,and standardize it.The effectiveness of the method is proved by experiments.Then experiments are designed to obtain the optimal parameters of the ensemble algorithm classifier.Finally,the specific structure and algorithm of the dermoscopy image classification model based on image features are described,and the improved dermoscopy image classification model based on weighted image features is proposed,that is,by giving weight value to each basic classifier,it can provide more information for the meta classifier and further improve the classification effect.Finally,extended experiments are carried out based on lit,simple neural network and neural ode.2.An ensemble classification algorithm of dermoscopy based on neural network is proposed.Firstly,color correction is performed on the dermoscopic image to reduce the influence of illumination on the classification effect.Data enhancement is performed to enrich the diversity of the image.Basic classifier and meta classifier are selected and relevant parameters are set.Then,based on the loss function to deal with the imbalance of the dermatoscope image data set,three different algorithms are proposed,and the best algorithm is selected through experiments.In order to improve the effect of classification,an ensemble model of dermoscopy classification based on neural network is proposed,and the basic classifiers are filtered by traversal.And a neural network based color channel ensemble classification algorithm is proposed,and experiments are carried out based on LIT(Learned Intermediate Representation Training),simple neural network and Neural ODE(Neural Ordinary Differential Equations).Finally,the system is built to realize the application of ensemble learning classification algorithm.This thesis sets the parameters of the basic classifier and the meta classifier of ensemble learning through experiments.And through a large number of experiments,this thesis verifies the effectiveness of feature selection and model design in the color featurebased dermoscope classification algorithm,and the effectiveness of color correction,imbalance treatment based on loss function and model design in the neural network integration algorithm,and obtains good experimental results.In this thesis,the seven classification ISIC(International skin imaging collaboration)dermatoscope image open data set is selected as the data set.The classification accuracy of the integration algorithm based on image features is 63.8%,and the classification accuracy of the integration algorithm based on neural network is 87.6%.