The Effects of PET Reconstruction Parameters on Radiotherapy Response Assessment and an Investigation of SUV-peak Sampling Parameters
dc.contributor.advisor | Turkington, Timothy G | |
dc.contributor.advisor | Das, Shiva K | |
dc.contributor.author | Rankine, Leith John | |
dc.date.accessioned | 2013-05-13T15:37:45Z | |
dc.date.available | 2013-05-13T15:37:45Z | |
dc.date.issued | 2013 | |
dc.department | Medical Physics | |
dc.description.abstract | Purpose: Our primary goal was to examine the effect of PET image reconstruction parameters on baseline and early-treatment FDG-PET/CT quantitative imaging. Early-treatment changes in tumor metabolism in primary tumor and nodes can potentially determine if the patient is responding to therapy, but this assessment can change based on the reconstruction parameters. We investigated the effect of the following reconstruction parameters: number of Ordered-Subset-Expectation-Maximization (OSEM) iterations, post-reconstruction smoothing, and quantitative metrics (SUV-max, SUV-mean, SUV-peak). A concurrent investigation explored in detail the sampling parameters of SUV-peak by way of a Monte Carlo digital phantom study. SUV-peak was proposed as a compromise between SUV-max and SUV-mean, in hope to retain key attractive features of these two metrics (inter-physician independence of SUV-max, noise-averaging of SUV-mean) but reduce unwanted errors (noise dependence of SUV-max, contour-dependence of SUV-mean). Sampling parameters have vaguely been defined, in particular, the scanning resolution (i.e. 1 voxel, 1/2 voxel, 1/4 voxel, etc.) of the SUV-peak spherical ROI . We examined the role that partial-voxel scanning plays in tumor SUV recovery in both noise-free and realistic OS-EM noise environments. Materials and Methods: The response assessment investigation involved 19 patients on an IRB-approved study who underwent 2 baseline PET scans (mean-separation = 11 days) prior to chemoradiotherapy (70 Gy, 2 Gy/fraction). An intra-treatment PET scan was performed early in the course of therapy (10-20 Gy, mean = 14 Gy). The images were reconstructed with varying OS-EM iterations (1-12) and Gaussian post-smoothing (0-7 mm). Patients were analyzed in two separate groups, distinguished by the PET/CT scanner used to acquire data: (1) GE Discovery STE; and (2) Siemens Biograph mCT. For each combination of iterations and smoothing, Bland-Altman analysis was applied to quantitative metrics (SUV-max, SUV-mean, SUV-peak) from the baseline scans to evaluate metabolic variability (repeatability, R = 1.96&sigma). The number and extent of early treatment changes that were significant, i.e., exceeding repeatability, was assessed. An original SUV-peak algorithm was developed, which measures SUV-max and SUV-peak for as small as 1/32 voxel scanning. Two rounds of digital phantoms were generated for the SUV-peak investigation. First, 10,000 spherical tumors were generated at a random matrix location for each diameter 1-4 cm and smoothed with an isotropic Gaussian, FWHM = 0.8 cm, then evaluated using the SUV-peak algorithm. Next, realistic body-sized phantoms were generated with background activity, and 1,000 spherical tumors of activity 4 time the background for each diameter (1-4cm) were placed inside (8 tumors per phantom, location randomized within certain constraints). These images received realistic corrections in projection space for attenuation, spatial resolution, and noise, were reconstructed with an in-house OS-EM algorithm, and then assessed using the SUV-peak algorithm. The mean recovered activity above background and its coefficient of variation were calculated for all metrics for each tumor size, for both simulations. For the realistic noise simulation, various levels of Gaussian smoothing was applied post-reconstruction, the effects summarized in plots showing coefficient of variation vs. mean recovered activity above background - a comparison of the effectiveness of SUV-max and SUV-peak. Results: For the GE Discovery STE 2D cases averaged over all metrics (SUV-max, SUV-mean, SUV-peak) and structures (GTV, LN), repeatability, R, improved with increasing smoothing and decreasing iterations. Individually, SUV-mean repeatability was less affected by the number of iterations, but demonstrated the same relationship with smoothing. SUV-mean outperformed SUV-max and SUV-peak with regards to the number of cases exceeding repeatability, N. Considering R, N, and the sum of relative metric change outside repeatability, &Omega, averaged over all metrics and all structures, and normalized, several combinations of reconstruction parameters produced five optimal combinations above set thresholds: 1 iteration with 0.1-3.0 mm smoothing; and 2 iterations with 2.0-3.0 mm smoothing. Current GE 2D reconstruction protocol for HN cases uses 2 iterations and 3.0 mm post-smoothing, which lies on the edge, but within these recommendations. The relationship between repeatability and number of iterations for the 3D cases was more complex; SUV-max demonstrated the best repeatability with 2 iterations, with both SUV-mean and SUV-peak reaching the best repeatability with 4 iterations. The same dependence on smoothing was noted, i.e. increased smoothing gives lover (desirable) repeatability. SUV-mean once again outperformed SUV-max and SUV-peak with regards to the number of cases exceeding repeatability, N. The calculations of N and &Omega averaged over all metrics were limited severely by the low number of cases, damaging the statistical significance of the following recommendation. Three optimal combinations with averaged and normalized R, N, &Omega, above a set threshold are recommended as most effective reconstruction parameter combinations: 4 iterations with 2.0-4.0 mm smoothing. Current Siemens 3D reconstruction protocol for HN cases uses 4 iterations and 3.0 mm post-smoothing, which lies within these recommended parameters. However, no statistically significant conclusions could be drawn from this analysis for this scanner, and performing similar data analysis on a larger patient pool is proposed. The minimum spherical tumor diameter required for full recovery was 3.0-3.5 cm for SUV-peak, and 2.5-3.0 cm for SUV-max. SUV-max was found to overestimate the recovered value of tumors by up to 46% (vs. 10% for SUV-peak); above the minimum diameter for full recovery, SUV-peak values were significantly closer to actual tumor activity. Considering only the realistic noise tumors, the coefficient of variation for SUV-max ranged from 5.5-17.7%, whereas for SUV-peak these values were lower, 2.7-13.2%. Partial-voxel scanning did not substantially affect the coefficient of variation (<0.2%). Comparison of coefficient of variation vs. mean recovered value demonstrated that SUV-max with additional Gaussian smoothing outperforms SUV-peak by up to 0.8% for 1 cm tumors and 0.2% for 4 cm tumors. Other tumor sizes showed little difference between the two metrics. Conclusion: For patients scanned on the GE Discovery STE using the HN protocol (2D acquisition mode), images reconstructed for quantitative analysis may benefit from a low number of OS-EM iterations (≤ 2). Some post-reconstruction smoothing proved to be beneficial (1.0 mm ≤ FWHM ≤ 3.0 mm), however, over-smoothing for the sake of more qualitatively appealing images or improved image quality metric (e.g. SNR, CNR) may prove detrimental to quantitative response assessment analysis. Our results for the Siemens Biograph mCT using the HN protocol (3D acquisition mode) demonstrated favor towards 4 iterations and limited range of smoothing (2.0 mm ≤ FWHM ≤ 4.0 mm). These results are statistically limited, further cases are necessary for any conclusive recommendations on reconstruction parameters. SUV-peak was shown to reduce uncertainties associated with quantitative PET image analysis when compared directly to SUV-max. Above the minimum tumor diameter required for full recovery, SUV-peak also provides a better estimate of the actual tumor activity. However, initial comparisons of SUV-peak and SUV-max over various levels of additional Gaussian smoothing found SUV-max more favorable. Partial-voxel scanning of SUV-peak did not reduce the metric's coefficient of variation in images with realistic noise. Therefore, a phantom investigation is proposed to compare SUV-peak and SUV-max of real scanned images with various levels of post-smoothing, which may conclusively eliminate the need for SUV-peak. | |
dc.identifier.uri | ||
dc.subject | Medical imaging and radiology | |
dc.subject | Biomedical engineering | |
dc.subject | Oncology | |
dc.subject | Head and neck cancer | |
dc.subject | PET | |
dc.subject | Quantitative | |
dc.subject | response assessment | |
dc.subject | SUV | |
dc.subject | SUV-peak | |
dc.title | The Effects of PET Reconstruction Parameters on Radiotherapy Response Assessment and an Investigation of SUV-peak Sampling Parameters | |
dc.type | Master's thesis |
Files
Original bundle
- Name:
- Rankine_duke_0066N_11919.pdf
- Size:
- 2.17 MB
- Format:
- Adobe Portable Document Format