Generative deep learning synthesizes high signal-to-noise ratio sensitivity maps for PET from low count direct normalization data.
Authors
Affiliations (5)
Affiliations (5)
- Radiology, Stanford University, James H. Clark Center, 318 Campus Drive,, Stanford, Stanford , California, 94305-5427, UNITED STATES.
- Stanford University, James H. Clark Center, 318 Campus Drive,, Stanford, Stanford , 94305-5427, UNITED STATES.
- Radiology, Stanford University, 318 Campus Drive, Section E1.1, Room E150, Stanford, California, 94305-6104, UNITED STATES.
- Stanford University, Stanford University, Stanford, California, 94305-6104, UNITED STATES.
- Department of Radiology, Stanford University, James H. Clark Center, 318 Campus Drive,, Stanford, Stanford , California, 94305-5427, UNITED STATES.
Abstract
An accurate and precise normalization procedure is essential to correct
for variations in detector efficiency in reconstructed positron emission tomography
(PET) images. Direct normalization is a conventional approach that requires a large
number of counts per line of response (LOR) from a known normalization source,
which is time-consuming due to the use of relatively low-activity sources. We present a
deep learning framework that generates high signal-to-noise ratio (SNR) normalization
factors and sensitivity maps from low-count direct normalization data acquired with
a PET insert for MRI. We developed an attention-guided Pix2Pix, a conditional
generative adversarial network (cGAN), to maximize detector efficiencies and remove
detector block patterns and associated ring artifacts in the resulting PET images.
Quantitative evaluations were performed by testing the model on the unseen direct
normalization data to reconstruct images of a Hoffman brain phantom, a contrast
phantom, and a uniform cylinder phantom using high-count, low-count (1-15% of
full scan), and synthetic high-count sensitivity maps. The Hoffman brain image
volume normalized using a synthetic sensitivity map with 15% count statistics as input
produced results that closely matched that using the high count normalization data,
with peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM),
and normalized root mean square error (NRMSE) values (mean±standard error) of
30.68±0.31, 0.95±0.002, and 0.35±0.003, respectively. In comparison, the unprocessed
sensitivity map with 15% count statistics yielded substantially worse PSNR, SSIM,
and NRMSE values of 15.93±0.426, 0.54±0.013, and 1.843±0.0280, respectively. This
novel, fast, and effective approach enables high SNR direct normalization of PET image
volumes through deep learning using synthetic correction factors obtained from a short
normalization scan.