Empirical studies indicate that the NIH Scanditronix PET brain scanner blurs images with a Gaussian point spread function whose full-width at half maximum equals 7.1 mm. PET images are acquired via tomographic reconstruction from projections, using the filtered back projection algorithm. It is these reconstructed PET images that are found to be Gaussian blurred. Deblurring such images offers the possibility of greatly improved diagnostic capability.
A controlled experiment was conducted in the Nuclear Medicine Branch at the National Institutes of Health involving the Hoffman `brain phantom'. This is a device made of plastic in which radioactive sources can be placed at chosen locations. A PET scan of the phantom is very similar to that of an actual human brain. The advantage is that one can compare the resulting PET image with the exact original.
A Hoffman phantom was used that simulated the human brain uptake of -labeled fluorodeoxyglucose, a very common PET radiotracer. In that phantom, the ratios of activity between the gray matter, white matter, and cerebrospinal fluid are . Scans of various duration were obtained. Because of the Poisson process accompanying positron emission, short duration scans are much noisier than long-duration scans. In addition, noise effects are exacerbated by the filtered back projection algorithm, which produces highly correlated noise in the reconstructed image.
In the longest duration scan, (10000 sec), the activity ratios prior to deblurring were found to be , a very substantial departure from the true result. However, after deblurring with the SECB procedure, the correct ratio of was recovered. This indicates that posing the PET deblurring problem as a shift-invariant deconvolution problem, with the empirically determined Gaussian point spread function, is a valid mathematical model. Moreover, SECB deblurring can markedly improve the quantitative accuracy of long-duration PET scans. This remains true for PET scans of 1000 sec duration.
However, for scans of duration less than 600 sec, the presence of highly correlated noise in the blurred image leads the deblurring procedure to produce erroneous results, in that the white matter activity becomes about equal to that of cerebrospinal fluid, and the latter begins to gain apparent areas of activity that are not really there.
The value of this experiment lies in demonstrating that SECB deblurring can greatly improve quantitative PET imaging, provided the level of noise in the blurred image can be controlled. It also pinpoints the source of error in short duration PET scans, namely an artificially high amount of correlated noise resulting from use of the filtered back projection algorithm. Other tomographic reconstruction procedures might be used that do not have this noise enhancing property. Combined with subsequent SECB deblurring, such short duration PET scans can eventually provide quantitatively useful information.