Deep neural networks for classification of dermatological images with multiple skin lesions

Skin cancer is one of the major threats to men's and women's health on a global scale, and as with all other cancers, early diagnosis leads to a high rate of recovery. To reduce the required time for diagnosis, we developed an architecture for the automated classification of dermatological images with multiple skin lesions. The proposed system is based on a classical Unet architecture trained with patches extracted from four images with various skin lesions to identify the areas of interest whose condition is determined by an adapted EfficientNetB5 architecture trained with the HAM10000 dataset. Our results showed that the dermatoscopic image models learned from the HAM10000 dataset can be successfully used to diagnose skin cancer from images with multiple lesions, captured with usual cameras.