Anatomy Detection and Localization in 3D Medical Images

Abstract

This chapter discusses the use of regression forests for the automaticdetection and simultaneous localization of multiple anatomical regions withinComputed Tomography (CT) and Magnetic Resonance (MR) three-dimensional images.Important applications include: organ-specific tracking of radiation dose overtime; selective retrieval of patient images from radiological database systems;semantic visual navigation; and the initialization of organ-specific imageprocessing operations. We present a continuous parametrization of the anatomylocalization problem, which allows it to be addressed effectively by multivariaterandom regression forests. A single pass of our probabilistic algorithm enablesthe direct mapping from voxels to organ location and size, with training focusingon maximizing the confidence of output predictions. As a by-product, our methodproduces salient anatomical landmarks, i.e. automatically selected “anchor”regions which help localize organs of interest with high confidence. This chapterbuilds upon the work in [80, 277] and demonstrates the flexibility of forests indealing with both CT or multi-channel MR images. Quantitative validation isperformed on two groundtruth labelled databases: i) a database of 400 highlyvariable CT scans, and ii) a database of 33 full-body, multi-channel MR scans. Inboth cases localization errors are shown to be lower and more stable than thosefrom more conventional atlas-based registration approaches. The simplicity of theregressor’s context-rich visual features yield typical run-times of only 4seconds per volume. This anatomy recognition algorithm is now part of thecommercial product Microsoft Amalga Unified Intelligence System.