Overview
The responsible use of modern artificial intelligence in clinical processes requires that human doctors, who are ultimately responsible for making a diagnosis, understand the basis on which artificial intelligence bases its diagnostic proposal. In particular, the same clinical characteristics should be used as specified by evidence-based medicine. In the case of dermatological diseases, this includes, for example, skin redness, scaling, localization or demarcation. Due to phenotypic and other influencing factors, the combinatorial diversity of clinical features and their manifestations is huge. In addition, there are rare diseases and other challenges that make it impossible to create a high-quality and representative database of disease patterns manually annotated with clinical features as a basis for training artificial intelligence. We address this challenge with self-supervised learning techniques and an automatic mapping of visual features to medical taxonomies. The aim of the project is a diagnostic system that is able to explain its diagnosis to human dermatologists based on recognized clinical features.
Translated with DeepL.com (free version)
