VISE 2017 Summer Instructional Seminar – Provost Research Studio in Ultrasound Imaging Research
to be delivered by Michael Insana, PhD,
Donald Biggar Willett Professor of Engineering,
Professor of Bioengineering,
University of Illinois at Urbana-Champaign
Adventures with machine-learning methods for reconstructing ultrasonic elasticity images
Seminar to be moderated by Brett Byram, Assistant Professor of Biomedical Engineering
Monday, May 22 at noon
Stevenson Center 5326
Abstract:
There is much excitement about the use of convolutional neural networks for image recognition and feature extraction because of the comparatively high diagnostic performance they have demonstrated.聽 However, there has been relatively little use of neural networks in image formation thus far.聽 Ultrasonic elasticity imaging is a great application of machine learning to image formation because this is often a very poorly conditioned mathematical inverse problem.聽 I will describe our efforts at combining finite-element algorithms with unique neural-network structures to estimate mechanical properties of soft tissues.聽 This is a data-driven approach that we refer to as the Autoprogressive Method.聽 It is used to first estimate stresses and strains in 2D or 3D from a time series of sparsely positioned force-displacement measurements. 聽Once networks are trained from exposure to tissue measurements, an informational model is formed.聽 We then probe the model retrospectively to form images of viscoelastic properties.聽 I will describe the many challenges encountered during early development of these methods as we learned to image linear-elastic media and anisotropic tissues.