An Empirical Study of Deep Learning Frameworks for Melanoma Cancer Detection using Transfer Learning and Data Augmentation

Melanoma is a type of skin cancer that usually develops rapidly and can spread to other parts of the body, causing death or complication of treatment to a large population. Early detection of Melanoma is the key to increase the chances of patients' survival. While accurate Melanoma diagnosis can be performed through clinical examination and dermatology tests, such procedures are usually time-consuming, costly, and severely delayed due to patients' emotional aspects or other obstacles. Recently, machine learning approaches, deep learning, in particular, have shown great potential in diagnosing Melanoma using images captured through the camera. Accurate detection of Melanoma using low-end images with machine learning delivers a solution for rapid screening of Melanoma without clinical visits or experts. Although many deep learning methods can be applied to Melanoma detection, there exists a large variance between their performance, given that their parameters, such as learning rates, optimizers, batch size, etc., always differ. In this paper, we carry out a systematic study to validate a more precise deep learning framework for detecting Melanoma and other types of skin lesions. A generic Convolutional Neural Network (CNN) is performed, and transfer learning using a pre-trained framework is proved to help improve the detection accuracy. In addition, data augmentation is applied, which improves the model performance. A series of parameters for the learning rates, batch size, and optimizers etc., are tested within the models. Our study shows tremendous improvement in Melanoma detection with higher accuracy, which can be very useful for medical experts to provide efficient Melanoma detection to patients.