Wack DS, Wisniewski S, Erb D, Dandona P, Nabi HA.
Complex singular value decomposition improves the image quality of C11-PIB PET images.
Human Amyloid Imaging 2011 Meeting Abstracts. 2011 Jan 15;
Introduction: C11-PIB PET imaging typically uses multiple frames, which enables the tracer to be followed in
time. However, individual frames of PET data become increasing noisy as the duration decreases. Our method,
based on the Hilbert transform and Singular Value Decomposition (SVD), expresses the entire dynamic PET study
in terms of a few basis components. This provides a dramatic improvement in the appearance of each frame, a
reduction in noise, and should improve parameter estimation. Unlike conventional SVD approaches, our approach
has complex valued components (magnitude and phase). We refer to our approach as complex SVD (CSVD).
Methods: Dynamic PIB PET brain scans were acquired from three participants. Data from each scan were
converted to a 2-dimensional m x n matrix with m representing the number of voxels and n the number of frames.
This matrix was used as input to a Hilbert transform, with the result used in SVD. A noise reduced version of the
PET image is created by constructing the PET image from only the first 4 components from the SVD process.
Evidence of noise reduction was measured as a decrease in the standard deviation (SD) of an ROI placed in the
Results: The average SD of the cerebellum reference ROI was 16% and 10% of the mean values for the
unprocessed and CSVD processed frames, respectively. Mean value differences were within 4%. Notably, the
second phase component image had strong similarities to a grey/white matter image segmentation.
Discussion: Reduction in measured SD implies a reduction in image noise was achieved by our method. Unlike
smoothing, each frame showed dramatic improvement in the definition of structures.
Dynamic PET imaging, typically, requires a tradeoff between image quality and temporal resolution. Our approach
allows an improved temporal resolution, while actually improving image quality of each frame.