• cascading training
• convolutional neural networks
• grad-cam
• image classification
• model blending
• model explanation
• non-melanoma skin cancer
• saliency maps

Non-melanoma skin cancers are frequently occurring cancers distinct from melanoma. Nowadays the workload of dermatologists is becoming impossible to manage, an automated approach to pre-screen patients can be a valid help for them.
We used Convolutional Neural Networks to classify non-melanoma skin cancer, focusing on four kinds of skin cancer: Actinic Keratosis, Basal Cell Carcinoma, Squamous Cell Carcinoma and Seborrheic Keratosis. Firstly, we constructed three data sets, each one of them contains a different type of image among contextual, macro and dermoscopic. For each type of image, we developed a network with different styles of training that we called homogenous training and heterogeneous training. We compared our best model with a State of Art model obtaining in some case better performances. Next, we combined three networks, one for each type of image, to improve predictions.
We also addressed the model explanation, a topic that is getting more and more attention, with the two most used methods: Saliency maps and Grad-CAM.
In the end we introduced a real case scenario, adapting the models we developed to a demo application for iOS and Android.
To complete our study, we also compared a multi-class model with a cascading training of binary models.