Full Convolutional Networks for Skin Lesion Detection and Prediction of Melanoma Skin Cancer
We present an algorithm to perform the task of detection of skin lesions and further processing to classify the segmented portion of skin lesions into different classes like Benign and Malignant. The proposed algorithm carries out the segmentation and detection of skin lesion by analyzing the boundaries of lesion regions having exhibited the properties of boundaries. The algorithm consists of different stages of operations like finding the boundaries of skin lesions by analyzing the pixel information gathered from images of a dataset. Further, the proposed algorithm finds the regions exhibiting the properties of boundary or edges of an object. The part one of the research is to determine the portion of regions with edges and boundaries along with convolution and subsampling. Part two of the proposed method is to determine the portion of image with up-sampling and classification by soft-max function. Thereby segmented portion of images are further subjected to tasks like features extraction and classification. The features are extracted from images by segmented portion are processed with different layers similar to VGG-16. The layers and blocks of 16 layers together contribute to the features extraction and features classification of images into different classes of skin lesions. The proposed Full Convolutional Neural networks have achieved an accuracy of 81.33% on a benchmark dataset with few contribution towards robustness. The proposed method has out-performed the results of other existing contemporary methods.