A Method for Detecting Skin Cancer Disease Based on Deep Learning in Dermoscopic Images

Cancer, in general, and skin cancer, in particular, is a genetic disease caused by cell division malfunction. It represents a major public health problem due to the large number of deaths caused by it. This is because it is often detected at very advanced stages. Traditional detection methods often involve visual examination by dermatologists. However, these methods are generally one-sided and inaccurate in some cases. Hence, there is an urgent need to develop effective and intelligent methods for early diagnosis to improve skin cancer treatment. This paper presents a solution based on Deep Learning methods, more specifically, on a pre-trained convolutional neural network (CNN) EfficientNet-B0, which we refined by performing an unfreezing technique so that the model could capture the specific features of our skin cancer detection dataset. As an optimizer for our model, we used stochastic gradient descent with momentum with a learning rate of 0.001. This method was implemented in the publicly available ISIC-2019 dermoscopic image database, in which we performed image resizing and cropping, morphological closure, and Gaussian filter application preprocessing. We then proceeded to rebalance the results, followed by augmentation of data. We obtained a classification accuracy of 88%.