King's College London

Research portal

Evaluation of Strategies for PET Motion Correction-Manifold Learning vs. Deep Learning

Research output: Contribution to journalConference paper

Original languageEnglish
Pages (from-to)61-69
JournalLecture Notes in Computer Science
Volume11038
Early online date24 Oct 2018
DOIs
Publication statusE-pub ahead of print - 24 Oct 2018

Documents

King's Authors

Abstract

Image quality in abdominal PET is degraded by respiratory motion. In this paper we compare existing data-driven gating methods for motion correction which are based on manifold learning, with a proposed method in which a convolutional neural network learns estimated motion fields in an end-to-end manner, and then uses those estimated motion fields to motion correct the PET frames. We find that this proposed network approach is unable to outperform manifold learning methods in the literature, in terms of the image quality of the motion corrected volumes. We investigate possible explanations for this negative result and discuss the benefits of these unsupervised approaches which remain the state of the art.

Download statistics

No data available

View graph of relations

© 2018 King's College London | Strand | London WC2R 2LS | England | United Kingdom | Tel +44 (0)20 7836 5454