Detection and Recognition of Skin Cancer in Dermatoscopy Images

Abstract

Melanoma and basal-cell carcinoma (BCC) are the two most common skin cancers, the death rate of melanoma is very high. If melanoma can be diagnosed early, the survival rate of patients will be greatly improved. But nevus and melanoma have similar appearances and symptoms. In order to reduce the cost for doctors to diagnose skin cancer, we proposed a computer-aided diagnostic system (CAD) that detects and identifies melanoma, nevus, and BCC in dermoscopy images. Firstly, use the hair removal algorithm, Gaussian filter and Wiener filter to remove the noise; Secondly, use the otsu to obtain the lesion area; then extract the texture and color features from the lesion area and use multiset discriminant correlation analysis (MDCA) to fuse the extracted features; finally, skin cancer is classified into melanoma, nevus, and BCC by KNN classification. Our aim is to select suitable features, test the effectiveness of MDCA, and compare the classification results with the methods in recent years. The improved algorithm was tested on the ISIC dataset, which included 469 images of melanoma, 127 images of basal cell carcinoma and 412 images of nevus. Compared with the methods in recent years, the selected features in this study combine with the MDCA method can improve the accuracy rate by 10.34%.