SPECT Reconstruction & Filtering
Goals Understand the basics of SPECT Reconstruction Filtered Backprojection Iterative Reconstruction Make informed choices on filter selection and settings Pre vs. Post Filtering Filter Parameters Resolution Recovery 2 /
Acquisition Major components of a tomographic acquisition Matrix : 64 vs 128 Acquisition Zoom Number of projections Time per view, including half-time and half-dose What do these factors have to do with filters? Matrix & zoom determine pixel size. Cardiac studies need to be 6 8 mm/slice for most quantification packages Projections and time per view affect total counts 3 /
Acquisition With a 1.33 zoom: 128 matrix = 3.4mm/pixel 64 matrix = 6.8mm/pixel 4 /
Reconstruction - Simple Backprojection Mathematical method to reconstruct tomographic image Reconstruction of an Image of a point source Each projection sees the point source as a strip or ray-sum Simple backprojection plots the counts per pixel across each ray-sum 5 /
Reconstruction - Simple Backprojection Simple backprojection reconstructs an image by spreading each raysum evenly among the pixels along the projection line This simple method results in streaking or star artifacts 6 /
Reconstruction - Filtered Backprojection Filtered Backprojection modifies the original projection profiles with a filter function before projecting the ray sums into the image plane The streak and star artifacts are further reduced or cancelled when more projections are acquired. A pre-filter is used along with the RAMP filter to produce the final image 7 /
Image Components To choose the optimal filter, need to understand the three major components of any image: Background Generated from the backprojection RAMP filter eliminates most of this Image Data Composed of a series of frequency waves - high, medium, low Frequency depends on type of study Noise Unwanted high frequency contribution to an image due to statistical fluctuation Inversely proportional to the number of counts in the raw data Pre-filter used to remove noise 8 /
Frequency of Image Data High Frequency Large difference in pixel to pixel counts Images with sharp edges May need less filtering Low Frequency Less detail means lower frequency May need more filtering 9 /
High Frequency Noise Low Count Study 25 counts per pixel 20 percent noise High Count Study 100 counts per pixel 10 percent noise Lots High Frequency Noise Less Noise 10 /
Filtering to Optimize Image Data RAMP Filter Removes background Pre-Filter Removes Noise RAMP Filter Pre-Filter Remaining Image Data 11 /
RAMP Filter Only RAMP Filter Removes background Very noisy images 12 /
Filtering to Optimize Image Data Heavy Filter Light Filter High Count Data Lighter Filter Keeps higher frequency components More noise included Sharper image More image data included Low Count Data Heavier Filter More high frequency removed Less noise included Smoother image More image data removed 13 /
Filtering to Optimize Image Data Trade-Off Light Filter Smoothing versus Resolution Heavy Filter What constitutes a light or heavy filter depends on counts in image 14 /
Cutoff Frequency Defines the point at which the filter is defined to be zero Units may be cycles/pixel or cycles/cm Vendors may define filters differently Lower number = smoother image 15 /
Hann Filter Examples Cutoff Frequency: 0.5 cycles/cm Cutoff Frequency: 0.75 cycles/cm Cutoff Frequency: 1.00 cycles/cm 16 /
Butterworth Filter Critical Frequency Defines the point at which the filters starts to roll-off towards zero Units may vary by Vendor A lower number = smoother image (heavier filter) Order or Power Factor Determines the slope of the filter curve A higher number = steeper slope A higher number = smoother image More subtle of an effect than critical frequency 17 /
Butterworth Filter Examples Critical Frequency Critical Frequency: 0.25 cycles/cm Critical Frequency: 0.50 cycles/cm Critical Frequency: 0.75 cycles/cm 18 /
Butterworth Filter Examples Power Factor Critical Frequency: 0.40 cycles/cm Power Factor: 5 Critical Frequency: 0.40 cycles/cm Power Factor: 10 Critical Frequency: 0.40 cycles/cm Power Factor: 20 19 /
Iterative Reconstruction An alternative to filtered backprojection Can accommodate corrections for attenuation, scatter, and noise and used for half-time/half-dose imaging Can produce higher image quality than FBP in terms of less noise, better resolution, fewer artifacts, and less distortion Computationally intense requires more time to complete FBP IR 20 /
Iterative Reconstruction vs FBP FBP IR 21 /
Iterative Reconstruction Initial estimate of transaxial distribution Creation of projection profiles Comparison of computed profiles to acquired patient profiles Error image used to update image estimate Repeat per iteration 22 /
Types of Iterative Reconstruction MLEM maximum likelihood expectationmaximization Original method used Computationally intense OSEM ordered subset expectation-maximization Projection data is grouped in ordered subsets o Generally groups of projections MLEM is then applied to each of the subsets Faster processing 23 /
Iterative Reconstruction Uses 3D Post-Filter to smooth images and reduce noise If enough counts, may not need post-filter Iterative Reconstruction 2 iterations, 10 subsets No Post-Filter Butterworth Filter Critical Frequency 0.50, Power Factor 10 24 /
Iterative Reconstruction Number of Iterations More does not always mean better Profiles may diverge as iterations increase Subjective More iterations = increased noise 25 /
IR Number of Iterations (Same Post Filter) 2 iterations 10 subsets 6 iterations 10 subsets 4 iterations 10 subsets 8 iterations 10 subsets 26 /
Iterative Reconstruction Number of Subsets As with iterations, more does not always mean better PET lesion detection has been shown to decrease in some studies as subsets increase over 12-14 More subsets = increased noise 27 /
IR Number of Subsets (Same Post Filter) 4 iterations 5 subsets 4 iterations 20 subsets 4 iterations 10 subsets 4 iterations 30 subsets 28 /
Beyond Iterative Reconstruction 29 /
Beyond Iterative Reconstruction Evolution Resolution Recovery Models the collimator-detector response in order to correct for noise and resolution degradation Corrects the blurring effect by including an accurate model of collimatordetector response function in an iterative SPECT reconstruction algorithm Vendor specific algorithm Utilizes intrinsic system response and collimator specific geometric response for each combination of acquisition system, radiopharmaceutical, and acquisition protocol 30 /
Evolution Half-Time Imaging 31 /
Evolution Comparison Filtered Back Projection OSEM Iterative Reconstruction Evolution Images courtesy of Baylor University Medical Center in Dallas 32 /
Filter Considerations Optimal Filter will depend upon: Study Type Organ studied will determine frequency of image data (e.g. Bone vs. In-111 Octreoscan) Radiopharmaceutical used will determine total counts and noise component (Tc-99m Sestamibi vs. I-123 mibg) Patient Body Habitus Patient size will effect counts and noise component Imaging parameters Time per projection Matrix size Equipment Single head vs. dual head Filter description (order/power and units for cutoff frequency are not uniform among vendors) Reconstruction Type Filtered backprojection, iterative reconstruction or resolution recovery Physician preference 33 /
Filter Conclusion There is no perfect filter!!! Filter is a preference Many quantitative programs (i.e. polar maps, gated SPECT ejection fraction) require a specific filter (and imaging protocol) for proper analysis 34 /
Resources Computers in Nuclear Medicine: A Practical Approach, 2 nd Edition, Kai Lee, PhD, Society of Nuclear Medicine, 2005. SPECT: A Primer, 3rd Edition, Robert English, CNMT, Society of Nuclear Medicine, 1995. Effective Use of Computers in Nuclear Medicine, Michael Gelfand, McGraw- Hill, 1988. Characterization of ordered-subsets expectation maximization with 3D postreconstruction Gauss filtering and comparison with filtered backprojection in 99mTc SPECT, Marco Brambilla et al, Annals of Nuclear Medicine Vol. 19, No. 2, 75 82, 2005 Effect of varying number of OSEM subsets on PET lesion detectability, Morey AM1, Kadrmas DJ., J Nucl Med Technol. 2013 Dec;41(4):268-73 Accelerated Image Reconstruction using Ordered Subsets of Projection Data, H. Malcolm Hudson, Richard S. Larkin, IEEE Transactions on Medical Imaging, Vol. XX, No. Y, Month 1994 35 /