Deep Learning in Orthopedics: How Do We Build Trust in the Machine?

Abstract

The paradigm of manual radiological assessment in orthopedics is at the point of significant disruption due to major breakthroughs in deep learning algorithms and computational power. Deep learning is a subset of the broader family of artificial intelligence methods that leverages artificial neural networks to assess imaging data sets. Deep learning has already been applied for automatic assessment of plain film radiographs with high degree of success in different orthopedic applications. However, a major limitation of deep learning is that it is a “black-box” algorithm, meaning its process and how it makes decisions are not easily interpretable. Interpretability in a medical setting is of utmost importance to build trust in the machine's outcome and move toward its meaningful integration into healthcare. In this article, we first briefly explain a deep learning algorithm and how it works at a high level. Next, we implement two different methods to enhance interpretability of a deep learning algorithm applied to an orthopedic application for detecting mechanical loosening of total hip replacement implants (1) on a high level using saliency maps and (2) on a lower level using activation maximization. By combining these two visualization methods, we can get a better understanding about deep learning function and shed light on its decision-making process to build more trust in the machine.

Topics & Concepts