A Proposal for an Explainable Fuzzy-based Deep Learning System for Skin Cancer Prediction

Explainable deep learning (XDL) is a research field that aims to make deep learning predictions or classifications more understandable for humans. Literature shows that deep learning (DL) algorithms are more precise in terms of their prediction than traditional machine learning (ML) algorithms. Nevertheless, they lack the interpretability and explainability that less complex algorithms are more fitted. Nowadays, one of the main goals of XDL research is to make these algorithms highly precise and explainable. There exist different systems and techniques, but only limited research on explaining deep neural networks using soft computing approaches. The purpose of this research is to propose the theoretical fundamentals of an explainable fuzzy-based deep learning (EFBDL) system that is both precise and explainable. The system comprises two main parts. First, the deep network part composed of a convolutional neural network (CNN) based on Inception V4 for image classification, a transfer learning mechanism, and a feature extraction algorithm based on neuron perturbation. Second, a soft computing part comprised of a fuzzy rule-based system (FRBS), a hierarchical network for natural language generation named granular linguistic model of a phenomenon (GLMP), and a human-machine integration methodology for linguistic rules named highly interpretable linguistic knowledge (HILK). The output of the overall system is an explanation of the neural network decision using natural language. This system focuses on preventing skin cancer rather than healing it. Thus, governments could use this kind of system for implementing policies focused on prevention and save in overall treatment costs of the disease.