@misc{oai:repo.qst.go.jp:00062110,
 author = {Shidahara, Miho and Ikoma, Youko and Seki, Chie and et.al and 志田原 美保 and 生駒 洋子 and 関 千江},
 month = {May},
 note = {Title
WAVELET DENOISING FOR PARAMETRIC IMAGING OF THE PERIPHERAL BENZODIAZEPINE RECEPTORS WITH F-18-FEDAA1106
\nAuthors
Miho Shidahara, Yoko Ikoma, Chie Seki, Yota Fujimura, Hiroshi Ito, Yuichi Kimura, Tetsuya Suhara and Iwao Kanno 
\nAffiliations
1 Molecular Imaging Center, National Institute of Radiological Sciences, JAPAN.
\nBackground and aims: The statistical noise of time activity curve (TAC) at voxel level causes severe bias and poor precision for estimated binding potential, BP(=k3/k4), images using a nonlinear least square fitting (NLS). The purpose of this study is to evaluate noise reduction capability of wavelet denoising for estimated BP images of the peripheral benzodiazepine receptor (PBR). 18F-FEDAA1106 is a radioligand for PBR, and its BP images should be formed using NLS because no reference regions can be assumed due to the physiological aspects of PBR1. We applied wavelet denoising to simulate data and clinical dynamic image of PBR with 18F-FEDAA1106.
Methods: After administration of 18F-FEDAA1106, three dimensional dynamic PET scans were performed by ECAT EXACT HR+ system (CTI-Siemens, Knoxville, USA) having 41 frames. The wavelet processing was applied in a volume fashion with a three-dimensional discrete dual-tree complex wavelet transformation2 at 4 scales with 112 subbands. The advantage of wavelet denoising is to realize spatially adaptive smoothing. In order to eliminate noise component in wavelet coefficients, real and imaginary coefficients for each sub-band were individually thresholded with NormalShrink3. Simulations were conducted to evaluate the performance of the wavelet denoising for parameter estimation. A simulated dynamic data was consisted of 4 parts of 2 gray matters (K1=0.25, k2=0.078, k3=0.043 or 0.0516, k4=0.0086), a white matter (K1=0.15, k2=0.075, k3=0.043, k4=0.01), and CSF to simulate anatomical structure of the brain (Hoffman brain phantom: 12812855 pixels, 222 mm). Each part had the BPs of 5, 6, 4.3, and 0, respectively. Then the phantom was smoothed with Gaussian filter (2.5x2.5 mm FWHM) and then Gaussian noise was added to mimic exact measurement at the noise level 20%. The BP images derived from wavelet denoising were compared with true BP image using 156 rectangular ROIs (55 pixel). The Wavelet denoising was also applied to clinical data derived from 3 young normal volunteers. Parametric images of BP were formed using voxel-based NLS fitting, and the results were compared with an ordinary ROI averaged estimates1. 
Results: In the simulation studies, estimated BP by denoised image showed better correlation against the true BP values (Fig-1A), although no correlation was observed in the estimates. In clinical data, wavelet denoising improved image quality of the estimates (Fig-1B). Originally estimated BP image includes bias against ROI averaged estimates (Y=1.12X+0.94, R2=0.55), however, estimated BP by denoised image improved relationship with ROI analysis (Y=0.91X+0.95, R2=0.57). 
Conclusions: Wavelet denoising improves bias and variation of pharmacokinetic parameters, especially, BP. 
\nReference 
[1] Fujimura Y, 2006, J Nucl Med, 47, 43-50.
[2] Alpert NM, 2006, NeuroImage, 30, 444-451.
[3] Fourati W, 2005, Inter J on GVIP, 5, 1-9., Brain'07 and BrainPET'07},
 title = {Wavelet denoising for parametric imaging of the peripheral benzodiazepine receptors with 18F-FEDAA1106},
 year = {2007}
}