AI opportunities and challenges

Abstract

Abstract Classical approaches to extract information from medical images include semantic features, defined by human experts, and agonistic features, defined by equations. The concept of machine learning approaches is to use agonistic features, but not based on mathematical description or modeling, but by a black box approach. Principal steps in establishing deep learning based approaches include training of a convolutional neuronal network based on a large dataset, using clinical diagnosis is the gold standard, testing performance in a validation dataset and refining the convolutional neuronal network by adding longitudinal data. In diseases such as diabetic retinopathy the diagnostic performance of such convolutional neural networks is at least as good as the performance of human graders. Up to now artificial intelligence approaches are, however, narrow. They may be superior to humans in performing one task, but they do not address factors that are important for treatment decisions. The present talk will summarize opportunities and challenges in implementing artificial intelligence in healthcare and discuss issues related to reimbursement, liability and patient acceptance.

Ähnliche Arbeiten