Research On Benign And Malignant Classification Of Melanoma Based On Feature Similarity Measurement For Codebook Learning

Melanoma is a common malignant tumor of the skin.In the past few decades,its incidence has been increasing rapidly,which has become one of the diseases that seriously threaten human health and even life.The melanoma classification method based on the Bag-of-Features(BoF)model can effectively assist dermatologists to diagnose skin diseases,which has been received extensive attention from experts and scholars in recent years.However,the traditional bag-of-features model is based on the k-means unsupervised clustering algorithm to learn codebook,the cluster centers generated by k-means are irresistibly attracted to denser regions of the feature.This leads to a suboptimal codebook in which most of the clusters are located in dense regions and a few are in sparse regions.In addition,because the k-means algorithm solves the optimal solution through multiple iterations,the entire process is time consuming.Based on the above analysis,in this paper,firstly,we propose a codebook learning algorithm based on feature similarity measurement.Secondly,a melanoma benign and malignant classification algorithm is proposed based on the codebook learning algorithm.The main works are as follows:(1)We propose a codebook learning algorithm based on feature similarity measurement.The algorithm uses linear independence and linear prediction to measure feature similarity,the codewords in the codebook learned by this algorithm are not affected by the feature density,which effectively overcomes the high-density tendency of the cluster centers of the k-means.At the same time,the codebook learning algorithm based on feature similarity measurement has lower time complexity than the k-means algorithm.Through the experimental comparison of two different codebook learning algorithms,it can be concluded that the proposed codebook learning algorithm based on feature similarity measurement can obtain more comprehensive codebook,and improve the classification accuracy and efficiency of melanoma classification algorithm.(2)A melanoma benign and malignant classification algorithm based on feature similarity measurement codebook learning is proposed.In this algorithm,we firstly extract RGB color histogram and SIFT features to describe melanoma images,in particular,in order to reduce the difficulty of codebook learning caused by the direct fusion of features,we use the BoF histogram fusion strategy of different feature descriptors,the combined BoF histogram vectors are then input into a support vector machine(SVM)to complete the classification.(3)In this paper,a variety of evaluation metrics are used to compare proposed method with a variety of existing comparison methods,and a series of comparative experiments are performed on two public available dermoscopy and non-dermoscopic datasets.The experimental results show that compared with other methods,the proposed melanoma classification method presented in this paper has a higher classification accuracy and efficiency.At the same time,we have verified the effectiveness of the feature similarity measurement codebook learning algorithm proposed in this paper.