Automatically detecting anatomy

Abstract

The fast, accurate, and robust detection of anatomical structures is an important prerequisite toward the automatic alignment of anatomy or motion correction for image acquisition in the context of magnetic resonance (MR) imaging. This chapter presents a novel approach to support the robust detection of various anatomical structures from fast MR localizer scans based on the principles of multiscale deep reinforcement learning and the concept of multiscale image navigation. Experiments demonstrate that the method achieves a high level of accuracy in automatically detecting the pose of the brain from MR localizer scans. Average position and scale errors under 2 mm and orientation errors under 1 degree are measured against ground truth determined by expert radiologists. Both measures are within the interrater agreement. In addition, the method yields similarly high accuracy also in cases with severe pathologies, such as large tumors.