Enhanced Skin Lesions Classification Using Deep Convolutional Networks

Abstract

The recent development of machine learning and deep learning techniques for medical image analysis has led to the development of intelligent diagnosis systems that can help doctors make a better diagnosis to the patients' diseases. In particular, skin diagnostics is a field where these new techniques can be applied with a high rate of accuracy. This study aims to enhance the accuracy of skin lesions classification based on two factors. The first is deeply trained all layers of implemented pre-trained models. Whereas, the second is balancing the number of images within the seven classes of dataset used. The state-of-the-art convolutional neural networks MobileNet and DenseNet-121 were trained on HAM10000 dataset. The two models pass through three phases; preprocessing, training and evaluation. Firstly, the dataset is down sampled, splitted and augmented to resolve misbalancing problem. Then, both models are deeply trained and finally they are evaluated against baseline models without balancing the classes. Multiple metrics were used to evaluate our models; precision, recall, F1-score, specificity and ROC AUC. In addition, the micro-average and macro-average of all previous metrics to extend to multi-classification. The accuracy of MobileNet and DenseNet-121 reach 82.6% and 71.9% on unseen testing images, respectively on the original dataset (i.e. before balancing the dataset). Whereas, they reach 92.7% and 91.2% on unseen testing images, respectively after balancing the dataset. This enhancement proves the necessity of existence of balanced dataset for training, to have better performance. Furthermore, MobileNet after balancing dataset has out performed the highest accuracy of ISIC 2018 challenge on the same dataset by 4.2%. For that, this model is the recommended one as it is a light-weight model, suitable for mobile applications used by dermatologists and its accuracy is comparably equal to DenseNet121.