Figure 5.7 bottom row shows the uncorrected/corrected reconstructions for patient 4 in PETmodel, where there are no avid lesions present, but there is a significant motion
artefact at the lung/liver edge in the uncorrected image. This artefact, often known as the ’banana artefact’, occurs where there is a mismatch in respiratory position between MRAC acquisition and emission data. This can occur when the MRAC is acquired at inhale position, rather than exhale as requested by the radiographer. This artefact is removed due to motion correction as the workflow allows the µ-map to be assigned to the most appropriate gate. The hump marked with arrows on the uncorrected line profiles in PETmodel and PETclinical show the top of the liver in the µ-map, at different
points to the actual top of the liver (in emission data).
Processing time for one motion correction PET reconstruction performed offline (using non-optimised MATLAB code and STIR) was ∼ 6 hours in total.
5.7
Discussion
Increases in sharpness and quantitative changes have been demonstrated in PET images with a practical motion correction scheme in a number of patients. These have the potential to improve lesion delineation and quantitation accuracy, and may also contribute to improved lesion detectability. The methodology also demonstrated respiratory artefact reduction, such as evident in patient four at the lung/liver edge, where the common ’banana artefact’ is observed.
A possible source of error exists within the choice of MR sequence and registration scheme used for the motion model. By only examining motion in the sagittal plane, we assume no lateral motion during respiration. Motion in this direction
has been shown in general, to be as low as 1.2 mm for lung lesions
[Seppenwoolde et al., 2002]. For an application of MR motion correction, this could be a source of error as MR can potentially have a much higher spatial resolution (of under 1 mm), whereas PET scanners have a higher typical resolution of around 4.5 mm, so this motion could be considered negligible, considering the application of PET motion correction. A 2D multi-slice acquisition scheme was used to ensure good spatial resolution, but with a gap between slices to keep scan time to a minimum. A full 3D motion model including lateral motion could be found by acquiring
5.7. Discussion 95 Patient 2 Uncorrected Corrected In te n s it y mm PETmodel PETclinical Patient 4 Uncorrected Corrected Patient 3 Uncorrected Corrected 0 2 0 4 0 2 0 3 0 4 0 5 0 6 0 In te n s it y mm 0 5 0 1 0 0 5 0 1 0 0 1 5 0 0 4 0 8 0 1 0 2 0 3 0 4 0 0 5 0 1 0 0 5 0 1 0 0 1 5 0 2 0 0 0 4 0 8 0 1 0 2 0 3 0 4 0 0 2 0 4 0 1 0 2 0 3 0 4 0 5 0 In te n s it y mm In te n s it y mm In te n s it y mm In te n s it y mm Uncorrected Corrected
Figure 5.7: Comparison of uncorrected and motion-corrected reconstructions of PETmodel
(column 1), with line profiles through ROIs for PETmodel(column 2) and PETclinical
(column 3) on patients 2-4. For each image pair, uncorrected images are top with purple line profiles, and corrected images are bottom with dotted green line profiles. Arrows in the bottom row show where the liver starts in the µ-map.
5.7. Discussion 96 contiguous slices [W¨urslin et al., 2013, McClelland et al., 2014] but this would extend the scan time. A general, easy to use, open-source registration scheme was utilised, which was chosen as a practical method, but it cannot deal with non-diffeomorphic transformations (sliding motion), as would occur between the liver and the ribs. Regardless of these possible limitations of the MR sequence and registration scheme we still see promising results, so we continue to use these in the subsequent chapters as the MR sequence is quick and practical in a clinical setting, and both the MR sequence and registration would be easy to implement in other centres.
We have only examined SUV changes in lesions that were already avid and detected in the uncorrected images by the radiologist. Detectability becomes a more important issue for smaller lesions that go undetected in uncorrected images but have the potential to become visible with motion correction. This will be further explored in Chapter 8.
There are a number of limitations with the methodology presented. The three examples of PETclinical for patients 2-4 were acquired between 50-61 minutes prior to
the motion model sequence acquisition. For each patient-specific motion model to be applicable an hour prior to formation, it is assumed that the patient is in the same position in the scanner and breathing style is consistent. Another issue with the respiratory signal is the fact that it is being extracted from two separate PET acquisitions, during the clinical scan then at the end for the motion model building section, and these two signals may not be directly comparable. It is important for the motion model acquired at the end of the scan to be applicable to the respiratory signal gathered throughout the clinical scan. By plotting the two signals for a single patient scan alongside each other it is clear the signals do not have the same amplitude (figure 5.8). As the extracted signals with the PCA method have a zero mean and a scale which is not immediately related to the amount of motion, it is hard to determine whether the drop in amplitude is due to shallower patient breathing, or for another reason, such as a natural drop in count statistics and contrast over time due to washout. With the current methodology, if the amplitude of respiration is larger in PETclinical
than PETmodel then data that falls outside of pre-determined amplitude range are
currently rejected, leading to a loss of count statistics.
5.7. Discussion 97 0 50 100 150 200 250 −10 0 10 20 30 Time (s) P(t) (a) 0 50 100 150 200 −10 0 10 20 30 Time (s) P(t) (b)
Figure 5.8: Respiratory signals from two acquisitions on one patient, separated by 45 minutes
between types of breathing. Other methods have been proposed in the literature that collect data during different types of breathing to account for inter-cycle variability [King et al., 2012], but this increases scan time and can cause extra patient discomfort by forcing different types of breathing. The motion modelling relies on only one image per bin, which may not be a fair representation of that respiratory state, and if a single registration fails it will affect all data within that bin.
Physical relaxation of the patient laying down in the scanner could cause a slight drift in breathing pattern as a scan progresses. To increase the chances of repeatable respiratory motion, respiratory-affected bed position acquisitions should be acquired after localisers and non-respiratory-affected bed positions (head, neck, etc.), allowing the patient to relax before acquisitions of the thorax/abdomen.
The main issues of using motion information gathered much later after the main clinical scan, only using a single respiratory signal, and relying on one image per bin to represent a whole respiratory state will all be minimised in the next chapter, where a continuous motion model methodology is described.
In this chapter we have proposed a practical, anatomy-independent MR-based correction strategy for PET data affected by respiratory motion, and have shown it can improve image quality for PET acquired simultaneously to the motion-capturing MR, and furthermore for PET acquired earlier during a clinical scan whilst any other free-breathing or breath-hold diagnostic MR is being acquired. Quantitatively, mean increase in SUVpeak and SUVmaxwas demonstrated in patients with PET-avid lesions.
The method does not require changing the clinical protocol except for an additional short (1 minute) acquisition at the end of a clinical protocol, and with no external
5.7. Discussion 98 hardware required. We also developed a set of tools required to perform MR image-based PET motion correction, which will be used in the next chapters whilst improving on the methodology.
Chapter 6
Joint PET-MR Respiratory Motion
Models
6.1
Motivation
In this chapter we investigate the use of a continuous correspondence model, rather than using discrete MR bins, as in Chapter 5. There are a number of key advantages of using a continuous motion model.
With the previous method, only one MR image was chosen per slice location per respiratory bin, as a representative of the moving anatomy at a certain point in the respiratory cycle. This proved to be sub-optimal because empty bins may occur where no MR image was acquired at a certain slice location within the defined bin amplitude window. Continuous motion models allow all MR images to be used without the need to discard redundant images, and motion estimates can be made with interpolation at any value of the model surrogate, even for surrogate values at which no MR image was acquired. This also means a higher number of PET gates can be used, minimising intra-gate motion, without risking empty bins.
The continuous motion model allows extrapolation, estimating deformation fields at values of the surrogate signal that were not used as input to the model-building sequence, which has two advantages. First, for motion correction with the previous method, any PET data that fell outside of surrogate signal values that were not used during model-building had to be discarded, leading to less counts and therefore noisier images. With the continuous model, all of the PET data can be used in the image reconstruction by model extrapolation. Secondly, the attenuation µ-map can be used
6.2. Overview 100