Melanoma and Nevi Classification using Convolution Neural Networks

Abstract

Early identification of melanoma skin cancer is vital for the improvement of patients’ prospects of five year disease free survival. The majority of malignant skin lesions present at a general practice level where a diagnosis is based on a clinical decision algorithm. As a false negative diagnosis is an unacceptable outcome, clinical caution tends to result in a low positive predictive value of as low at 8%. There has been a large burden of surgical excisions that retrospectively prove to have been unnecessary.This paper proposes a method to identify melanomas in dermoscopic images using a convolution neural network (CNN). The proposed method implements transfer learning based on the ResNet50 CNN, pretrained using the ImageNet dataset. Datasets from the ISIC Archive were implemented during training, validation and testing. Further tests were performed on a smaller dataset of images taken from the Dermnet NZ website and from recent clinical cases still awaiting histological results to indicate the trained network’s ability to generalise to real cases. The 86% test accuracy achieved with the proposed method was comparable to the results of prior studies but required significantly less pre-processing actions to classify a lesion and was not dependant on consistent image scaling or the presence of a scale on the image. This method also improved on past research by making use of all of the information present in an image as opposed to focusing on geometric and colour-space based aspects independently.