KCL logo

Motion Modelling and Analysis Group

School of Biomedical Engineering and Imaging Sciences,

King's College London

Motion Modelling

PET-MR Motion Correction Using a Surrogate-Driven Respiratory Motion Model

(Funded by EU HYPERIMAGE project)

PET imaging is commonly used for detecting cancer and monitoring its progress, but its quantitative accuracy is limited by the problem of motion, such as that caused by respiration. The state-of-the-art in motion correction of PET data is limited to coarse level modelling of the motion and often requires significant extra scanning resources to acquire the necessary imaging data. As a result the translation of motion correction techniques into clinical use has been limited. The recent introduction to the market of simultaneous PET-MR scanners raises the possibility of motion correcting PET data based on motion estimates derived from MR imaging.

Initial work focused on the use of a surrogate-driven motion model. The model was built from dynamic MR images, and captured the relationship between a simple 1-D MR navigator signal and the complex 3-D respiratory motion of the thorax. Once the model was formed, motion estimates could be made by acquiring only the MR navigator and using it as input to the motion model. The resulting motion fields were used to motion correct reconstructed PET gates in simulations.


MR navigator MR motion model Uncorrected PET image Corrected PET image
Surrogate-driven respiratory motion model: (left-to-right) MR navigator used as input to model; MR volume transformed using resulting motion fields; uncorrected PET image; corrected PET image.

This work was then developed to use a more sophisticated motion model and imaging data as the surrogate. Specifically, dimensionality reduction using PCA was applied to a set of motion fields derived from dynamic MR imaging. The resulting eigenvectors represented the respiratory motion model. To apply the model, a fast 2D MR slice was acquired, and the modes of variation of the model were optimised to maximise the match between the acquired slice (known as the surrogate) and an exhale MR volume transformed using the model transformation. Thus, a 2D slice was used to estimate a 3D transformation, constrained by the motions seen during the model formation phase. This technique was also evaluated on simulated PET gates.


2D MR navigator slice PCA registration Uncorrected PET image Corrected PET image
Image-driven PCA respiratory motion model: (left-to-right) Example of a fast 2D MR slice; optimisation of PCA modes of variation to match 2D slice; uncorrected PET image; corrected PET image.

PET-MR Motion Correction Constrained by a Respiratory Motion Model

(Funded by EU SUBLIMA project and EPSRC grant EP/M009319/1)

Subsequent work on the use of motion models for PET-MR motion correction demonstrated that it was not actually necessary to acquire the surrogate data at all in order to apply the model. Given a simple surrogate-driven model such as those used in the work described above, reconstructed PET gates can be transformed to a reference motion state using the optimal motion field produced by the model, i.e. the one that maximises the similarity between the transformed PET gate and the reference PET gate. Effectively, this technique used a PET-PET registration that was constrained by the MR-derived respiratory motion model.


Motion model constrained PET motion correction
MR-derived motion model constrained PET motion correction: (top left) uncorrected; (top right) the proposed technique; (bottom left) unconstrained PET-PET registration; (bottom right) motionless, i.e. theoretical gold standard.

Motion Correction for Cardiac Catheterisations

(Funded by EPSRC grants EP/D061471/1 and EP/H046410/1)

Guidance of minimally invasive cardiac catheterisations can be enhanced by using 'roadmaps' derived from preprocedure imaging, such as MR or CT. However, motion, such as that caused by respiration, can cause a misalignment between the static roadmap and the moving anatomy that it represents. This work aimed to estimate respiratory motion transformations for updating the roadmaps in such procedures. An affine surrogate-driven respiratory motion model was formed from dynamic MR images, using an MR pencil-beam navigator positioned on the diaphraghm as the surrogate. During the procedure, the diaphragm was tracked from X-ray images, and this trace was used as input to the motion model. The model then produced an affine transform that was used to update the roadmap. The technique was evaluated on a number of clinical cases, with accuracy assessed by overlaying a roadmap representing a structure with an X-ray opaque catheter positioned inside it. The alignment between roadmap and anatomy was assessed by comparing the catheter position with the roadmap (see movies below).

Respiratory motion correction for cardiac catheterisations: (left-to-right) coronary sinus overlay; right atrium overlay.


In subsequent work, the motion model was refined to enable it to adapt to changes in breathing pattern, that can occur during some procedures. For example, patients may breathe deeply due to discomfort as the heart is ablated, or breathe more shallowly when they are relaxed due to the effects of sedation. Such breathing pattern changes affect the accuracy of the model, and the new adaptive and predictive model was able to adapt to these changes.


Later, it was shown how intraprocedure ultrasound data could be used as input to the motion model. This innovation enabled the motion information contained in the noisy, low quality but real-time ultrasound data to be exploited, whilst being constrained by a robust preprocedure motion model.


The ultrasound based approach was then extended to make use of a more flexible model, which permitted some deviation from the motions observed during the model formation. This was made possible using a Bayesian approach, in which the motion model defined a prior distribution of motion states, and the intraprocedure ultrasound data defined the likelihood term. By maximising the posterior probability, the most probably motion state was estimated, resulting in improvements in motion estimation accuracy.

US imaging illustration
Bayesian estimation of respiratory motion: illustration of motion estimation using intraprocedure ultrasound.


Population-based motion models are attractive as they obviate the need for a subject-specific scan for model formation. However, to date they have been less accurate and robust than subject-specific models. We aimed to address this weakness by developing an improved personalisation scheme for population-based respiratory motion models. The approach used neighbourhood approximation and learnt anatomical features to enable an accurate motion model to be formed for a new subject using a static anatomical image only. We demonstrated that accuracy was improved compared to the state-of-the-art in population-based models.

Learnt features
Personalisation of population-based respiratory motion models using anatomical features: anatomical features that were found to be correlated with respiratory motion.


Autoadaptive Motion Modelling

(Funded by EPSRC grant EP/H046410/1)

In an extension of the simultaneous groupwise manifold alignment (SGA) scheme, we demonstrated how using SGA with 2D motion fields, rather than 2D images, resulted in a highly novel form of surrogate-driven motion model. In this autoadaptive motion model, the data used to form the model and the surrogate data used to apply the model were both of the same type (i.e. 2D motion fields). This meant that the surrogate data could be automatically added to the model each time the model was applied, enabling it to automatically adapt to changes in breathing pattern without any need for interrupting the application of the model.