Robust ultrasound image analysis using learning

Abstract

Robust and accurate analysis of clinical ultrasound data is a challenging task due to the complexity of scanned anatomy, noise, shadows, signal dropouts and quantity of the information to be processed. As a result, traditional image analysis relying on the explicit encoding of prior knowledge such as perceptual grouping, variational or generative approaches is usually not enough to capture the complex appearance of ultrasound data. We will discuss a new class of methods that build on recent advances in discriminative machine learning to achieve robust and efficient performance. Image analysis is formulated as a multi-scale learning problem through which object models of increasing complexity are progressively learned. We will demonstrate example applications in Cardiology and OB/GYN.