Early prediction of skin cancer using deep learning
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
UMT Lahore
Abstract
The skin cancer is considered amongst one of the most fatal diseases, with high morbidity. By 2020, it is estimated that more than 100,000 people have been diagnosed by skin cancer and around 7000 mortalities have been recorded, in United States itself. Majorly, the skin cancers are divided in three main types which are basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and melanoma. In the most cases of the skin cancers, these are squamous cell carcinoma and basal cell carcinoma. Although these are malignant, these are less likely to appear on the other parts of the body if treated at early stages. If not treated early, they can deform locally. The first step in diagnosing a malignant lesion by a dermatologist is a visual examination of the suspicious area of the skin. Accurate diagnosis is important because of the similarity of some lesion types. Manual visualization and inspection through the eye could be hectic and troublesome when the number of samples is large. Therefore, an automated classifier is always required to perform such tasks. As compared to image processing approaches and conventional machine learning, deep learning is more efficient and accurate. Thus, an automated deep learning-based classifier is required for skin cancer identification using images. In the present study, the early prediction of skin cancer has been aimed. Data for 9 skin cancer types have been collected which are basal cell carcinoma, dermatofibroma, melanoma, nevus, vascular lesion, actinic keratosis, seborrheic keratosis, squamous cell carcinoma, and pigmented benign keratosis. Based on this data, a deep learning residual neural network algorithm is incorporated for extensive and deep feature extraction from the images, and their classification. A trained prediction model for skin cancer classification has been developed and validated using
v
cross-validation approaches. Thus, an accuracy of 97.45% has been achieved which prove that the proposed predictor surpasses the existing models and can be served as a time and cost-effective stratagem for identifying skin cancer and its type.