Logo image
Chapter 52 - Parametric Image Generation with Neural Networks
Book chapter

Chapter 52 - Parametric Image Generation with Neural Networks

Michael M. Graham, Steven B. Gillispie, Mark Muzi and Finbarr O'Sullivan
Quantitative Functional Brain Imaging with Positron Emission Tomography, pp.347-352
Academic Press
1998
DOI: 10.1016/B978-012161340-2/50054-8

View Online

Abstract

A trainable artificial neural network performs remarkably in the task of estimating (18F) fluorodeoxyglucose (FDG) parameters from noisy pixel-wise tissue time-activity curves (TACs), even when the plasma TAC is not known. Although training the network takes a relatively long time, once it is trained it can generate parametric images in only a few seconds. The accuracy of the estimates of MRfdg seems to be better than that achieved with graphical analysis and is also better than more conventional methods of parameter optimization when attempted on a pixel-by-pixel basis. The major advantages of this approach are that it is very fast, is relatively noise tolerant, and produces quantitative images of metabolic rate. Disadvantages are that it cannot produce reasonable quality rate constant images, presumably because of the extreme noise seen in single pixel TACs, and that the approach requires a relatively large number of dynamic images.

Details

Metrics

12 Record Views
Logo image