Refining Deep Learning Algorithms for Skin Cancer Detection
Abstract: Skin cancer is a growing major public health problem. Melanoma, which is malignant skin tumor, is the deadliest form of skin cancer. While skin cancer is amenable to early detection by direct inspection, visual similarity with Nevus, which is benign skin tumor, makes the task difficult. In this study, we refine Deep Learning Algorithms towards better classification dermoscopic skin images. The major contribution of this work is to develop a model suitable for less-computational devices.
Image analysis of skin lesions includes three parts, which are lesion segmentation, lesion dermoscopic feature extraction, and lesion classification. For this research our focus is on extracting feature and then classify the images. We used RGB as well as other color models, such as HSV and LIB, in the diagnosis of images. Building up a deep Convolutional Neural Network (CNN) for the task of image classification is done on a GPU-supported workstation.
The images for training and testing are obtained from public dataset of ISIC 2017: Skin Lesion Analysis Towards Melanoma Detection. The experiments are carried out to build a model with higher accuracy of classification, that is calculated by precision, recall, and f-measure. From the training images, the best model obtained an f-measure of 0.87 with seven layer of convolutions features to detect melanoma or nevus from 750 testing images. It is found that shape irregularity of skin lesion is the most significant feature in the undertaken task.
Yau Ka Cheung*
Judging Forms – Official judges only