Classification of Skin Lesions by Combining Multilevel Learnings in a DenseNet Architecture

Abstract

Automatic recognition and classification of skin diseases is an area of research that is gaining more and more attention. Unfortunately, most relevant works in the state of the art deal with a binary classification between malignant and non-malignant examples and this limits their use in real contexts where the classification of the specific pathology would be very useful. In this paper, a convolutional neural network (CNN) based on DenseNet architecture has been introduced and exploited for the automatic recognition of seven classes (Melanoma, Melanocytic nevus, Basal cell carcinoma, Actinic keratosis, Benign keratosis, Dermatofibroma, Vascular) of epidermal pathologies starting from dermoscopic images. Specialized network architecture and an innovative multilevel fine-tuning method that generates a set of specialized networks able to provide highly discriminative features have been designed. Finally, an SVM model is used for the final classification of the seven skin lesions. The experiments were carried out using an extended version of the HAM10000 dataset: starting from the publicly available images, geometric transformations such as rotations, flipping and affine were carried out in order to obtain a more balanced dataset.