Skin cancer detection with deep learning

• Main Author: Gupta, Jay
• Other Authors: Owen Noel Newton Fernando
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/156442
• Institution: Nanyang Technological University

In the medical community, cancer is defined as a disease when our cells grow uncontrollably in an abnormal process to form tumours. Although some tumours are harmless, others can be fatal if proven to be cancerous (malignant). Out of more than two hundred types of cancers, melanoma and carcinoma are aggressive types of skin cancers, with a less than 66% five-year survival rate for melanoma. They are the fifth and sixth most prevalent form of cancer in the United States and Singapore respectively. Melanoma is a growing concern among the elderly (65 years old and above), and a rising trend is observed among the younger demographics as well. This project aims to develop machine learning models to detect skin cancer with comprehensive benchmarking, and a supplementary smartphone application to monitor and track skin lesions with skin profile assessments to facilitate early detection of skin cancer, thereby, improving the prognosis of patients with higher chances of favourable treatment. An ensemble deep learning model is created by combining InceptionResNetV2, DenseNet201, and both the B4 and B6 variants of the EfficientNet neural network architecture. With transfer learning, the models are trained on the publicly available HAM10000 dataset with an accuracy of 0.94, macro-average F1-score of 0.91, and area under AUC-PR of 0.93 on the test sets. The model is a binary classifier that detects the probability of a lesion being benign (non-cancerous) or malignant (cancerous), which is then deployed, and used in the smartphone application.