Skin cancer disease analysis using classification mechanism based on 3D feature extraction

Abstract

Dermoscopic image analysis is essential for effective skin cancer diagnosis and classification. Extensive research work has been carried out on dermoscopic image classification for the early detection of skin cancer. However, most of the research works are concentrated on 2D features. Therefore, a 3D lesion establishment mechanism is presented in this work to generate 3D features from the obtained 3D lesions. The objective of this work is to reconstruct 3D lesion image from 2D lesion images and a multispectral reference IR light image. The 3D lesion establishment is achieved by designing an efficient convolutional neural network (CNN) architecture. Details of CNN design architecture are discussed. After reconstruction of 3D lesions, 2D and 3D features are extracted and classification is performed on the obtained 2D and 3D features. Classification performance is evaluated using the images from PH2 database. The mean classification accuracy using K-nearest neighbors (KNN) classifier based on the 3D lesion establishment using the CNN architecture is 98.70%. The performance results are compared against varied classification methods in terms of accuracy, sensitivity, specificity and are proved to be better.