Figure 3.17: Parametric images on a pixel-by-pixel basis of K obtained from Patient 1: (a) slice 10; (b) slice 13; (c) slice 21. Top row shows the images obtained from the raw dynamic images and bottom row shows the images obtained from dynamic images after cluster analysis. The images have been smoothed slightly for better visualization.

on the noise levels inherent in the data which affect, in addition to meaningful parameter estimation, the time required to converge as well as the convergence. Clustering may be useful as a preprocessing step before fast generation of parametric images since only a few characteristic curves which have high statistical significance, need to be fitted as compared to conventional pixel-by-pixel parametric image generation where many thousands of very noisy tissue TACs must be analyzed. The computational advantage and time savings for generation of parametric images (fitting many thousands of kinetic curves versus several curves) are apparent.

Figure 3.17 shows the parametric images of physiological parameters, K, obtained from the neurologic study for Patient 1 in the three selected slices. The top and bottom rows of the images correspond to the results obtained from pixel-by-pixel fitting the TACs in the raw dynamic PET data and data after cluster analysis, respectively. The K images are relatively noisy when compared to the data after cluster analysis because of the high-noise levels of PET data which hampered reliable parametric image generation. However, the visual quality of the K images improves markedly with cluster analysis as a result of the increased SNR of the dynamic images. Low-pass filtering of the original parametric images may improve the SNR but clustering should produce better results because it takes the tissue TACs with similar temporal characteristics for averaging. Meanwhile, low-pass filtering only makes use of the spatial (adjacent pixels) information for filtering and this will only further degrade the spatial resolution. The feasibility of using the kinetic curves extracted by cluster analysis for noninvasive quantification of physiological parameters and parametric imaging has been investigated and some preliminary data have been reported [109]. Some other recent studies can be found elsewhere [110-115].

0 0

Post a comment