Degree Name

Master of Science


Medical Radiation Physics - Faculty of Engineering


Introduction: Patient head motion is a well-recognised problem in single photon emission computed tomography (SPECT) of the brain. Motion occurring between or during the acquisition of projections can lead to reconstruction artifacts that compromise accurate patient diagnosis. Although some form of restraint tends to be used in practice, motion incidence and magnitude is still high enough to warrant frequent repeat studies or the application of motion correction. The motivation for this work was the outstanding need for a high-performance motion correction strategy for brain studies. Such a strategy should accurately correct general rigid-body motion. The optimal strategy would also be non-invasive, have a high degree of automation, and be fast, convenient (requiring little or no calibration, patient cooperation, and extra gadgetry), and robust with respect to noise. We describe and implement a fully 3D, non-invasive, data-driven approach that is suitable for use with clinical data and is potentially automatic. The approach is based on a comparison of measured and estimated projection data. Acquired projections are segregated into groups corresponding to discrete locations held by the brain during scanning and the largest group is reconstructed. The position and orientation of this reconstruction is optimised for each remaining group by comparing the measured projections with those generated from the transformed reconstruction. After each optimisation, the current reconstruction estimate is updated with the relevant projections using the ordered-subsets expectation maximisation (OSEM) algorithm. Methods: Three sets of experiments were carried out on different types of data to validate the motion correction procedure and investigate practical aspects of implementing the approach clinically. In the initial set of experiments, seven noisy motion-corrupted projection sets simulating 2-4 head positions were generated from the digital Hoffman brain phantom. The angular location and extent of movement and the magnitude of rotation and translation with respect to each axis was varied for each set. Motion correction was applied to these data using various regimes: with/without attenuation included in the optimisation; with/without a second iteration. Extracted motion parameters were compared with the applied movements. The error between the extracted and applied parameters was quantified in terms of the mean registration error (MRE), an average displacement of the vertices of a box surrounding the brain. Overall improvement from motion correction was quantified in terms of a mean squared difference improvement ratio (MSDR). Corrected, uncorrected, and motion-free slices were also compared visually. For the second group of experiments, three physical Hoffman phantom studies containing single or double movements were obtained. The Polaris motion tracker was used to provide an independent measurement of motion. Motion parameters were extracted using our approach and compared with those measured by the Polaris. An investigation of cost function behaviour was also carried out by mapping the cost function in the neighbourhood of the Polaris solution. The third group of experiments constituted a preliminary clinical validation. Three volunteers underwent a motion-free scan followed by a scan in which they performed one head movement. A fourth volunteer underwent two scans, holding a single (but different) brain location in each. Again the Polaris was used to measure the motion independent of our technique. Data from the fourth volunteer was used to simulate two single-movement studies, facilitating a rigorous quantification of the improvement obtained from motion correction. Optimisations were performed with and without reduced projections, scatter correction, thresholding of background counts, compensation to avoid biasing from truncated data, and pre-smoothing of the acquired data. Results: In the digital phantom experiments, estimated rotations and translations were mostly within 2(degrees) and 1mm of the applied values. The MRE was less than 1 pixel in most cases. Accurate motion estimates could be obtained at over twice the speed by leaving attenuation out of the optimisation stage. Visually, there was a clear reduction in motion-induced artifacts after correction. Most MSDR values were well in excess of 2, and the MSDR tended to increase with increasing corruption. A second iteration of correction did not provide sufficient improvement to warrant the additional time cost. In the physical phantom experiments there was good agreement between the extracted and Polaris measurements for the x and y-rotation and z-translation parameters. A systematic discrepancy existed for the remaining parameters. The discrepancy was reduced for the third dataset (two movements); in this case the corrected study closely resembled that obtained using the Polaris values. Analysis of the cost function indicated that the MSD was fairly insensitive to large rotations whilst being sensitive to typical translations. Discrepancies appeared to be the result of object symmetry. In all of the volunteer studies, sets of motion parameters were obtained that closely followed the trend of the Polaris. In general, however, there was a systematic discrepancy from the actual Polaris values. Scatter correction had little effect on accuracy. Using reduced projections (greater proportion of the image occupied by brain) tended to provide estimates as good or better than using larger projections. Pre-smoothing generally lead to less accurate estimates. For large movements, tracking the plane of truncation was necessary to obtain sensible estimates. Thresholding was important in removing background counts and confining the solution to a sensible portion of the cost function. For all volunteers there was a clear improvement in image symmetry and contrast after using our approach. In certain cases, correction was better than that obtained from the Polaris. Of particular concern is the method used for attaching the head target. Poor attachment can lead to decoupling of target and head movement. For the two semi-simulated studies, the MSD improved by approximately 4 and 2 respectively, whereas the Polaris provided no improvement. Conclusion: We have demonstrated that complex brain movements in simulated and real data can be accurately estimated and corrected using this data-driven approach.