Deep learning approach for detection of melanoma from skin lesion images

• Main Author: Ku, Hui Sien
• Other Authors: Jagath C. Rajapakse
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: http://hdl.handle.net/10356/77793
• Institution: Nanyang Technological University

Melanoma is a skin cancer type that results in the highest mortality and is increasingly aggressive to affect human health. Early diagnosis and detection of melanoma is crucial to lower its fatality. Since 1990s, many computer-assisted melanoma diagnosis techniques were invented. These approaches have been proven to have competitive diagnosis ability compared with manual dermoscopy due to their ability to go beyond available information to the naked eyes. Convolutional neural network (CNN) is a class of deep neural network that is popularly employed in visual imagery analyzation. This project aims to develop the melanoma detection model to incorporate the function of segmentation and classification. Specifically, this project builds a CNN model to perform classification of skin lesion images into benign and melanoma classes. In addition, it also investigates the use of U-net segmentation technique as the preprocessing step of skin lesion classification to improve the classification performance. The CNN classification model is developed by python using Tensorflow library and the U-net segmentation model is developed using Keras library. The models are trained and tested using International Skin Imaging Collaboration (ISIC) challenge 2018 training dataset. The best developed classification model yields the accuracy of 0.970, sensitivity of 0.764, specificity of 0.966, area under the curve (AUC) of 0.980 and precision of 0.864. The segmentation model reaches the Jaccard Index of 0.80. The two developed models have been integrated to create an Android application to allow the early and easily assessable diagnosis of skin cancer.