A Lightweight MobileNet-Based Framework for Multi-Class Skin Cancer Classification with Data Augmentation

Abstract

Skin cancer is one of the most common cancers, which has its origin rooted to long duration exposure to the harmful ultraviolet rays from the sun. Though skin cancer is quite prevalent, it is curable if detected accurately at an initial stage. As there are nine variations of skin cancers, dermatologists often scan the skin lesions and evaluate the patient’s clinical data to identify the categories of skin cancer. As the morphologic traits are not apparent to the naked eye, correct diagnosis may be difficult. Therefore, an artificial intelligence-based automated system can be capable of identifying the type of skin cancer with better accuracy. These AI based systems are often resource-intensive, resulting in slow detection. In this research, a machine learning based lightweight system is designed for the early detection of skin cancer from skin images. Ideally, the model should be balanced where all the classes of interest have almost an equal number of images. As the dataset used for this work was not a balanced one, a preprocessing technique called data augmentation is used to improve the performance of the proposed model. After rigorous validation of the proposed model, it is observed that the light-weight MobileNet-based our proposed model could achieve as good as 97% accuracy on the training dataset and a 82% accuracy on the test dataset.