OPTIMAL SMOOTHING OF POISSON DEGRADED NUCLEAR MEDICINE IMAGE DATA (PROCESSING)

The development of a method that removes Poisson noise from nuclear medicine studies will have significant impact on the quantitative analysis and clinical reliability of these data. The primary objective of the work described in this thesis was to develop a linear, non-stationary optimal filter to reduce Poisson noise.

The derived filter is automatically calculated from a large group (library) of similar patient studies representing all similarly acquired studies (the ensemble). Since the library provides the a priori probabilities of the different disease etiologies and acquisition parameters represented in the particular ensemble, a larger library will provide better estimates of the optimal filter values. Furthermore, the filter is designed to minimize the mean square difference between the filtered estimates and their expected image data, providing the optimal trade-off between noise reduction and signal degradation for members of the ensemble.

The filter design was evaluated under controlled conditions using two computer simulated ensembles, devised to represent selected properties of real patient gated blood pool studies. Fortran programs were developed to generate libraries of Poisson degraded simulated studies for each ensemble. These libraries then were used to estimate optimal filters specific to the ensemble. The filters were applied to Poisson degraded members of each ensemble and comparisons of the optimally filtered data with their non-degraded simulated data were performed. Optimal filters calculated from a library of 400 studies were shown to decrease the mean square error in the raw data by 65%. It was also possible to demonstrate that this error can be reduced by 85% if an infinite library is available.

Libraries of previously acquired patient gated blood pool studies then were used to estimate the optimal filters for an ensemble of similarly acquired gated blood pool studies. These filters were applied to studies of 13 patients who received multiple repeat studies at one time. Comparisons of both the filtered and raw data to averages of the repeat studies demonstrated that the optimal filters, calculated from a library of 800 studies, reduce the mean square error in the patient data by 60% (a library of 400 studies resulted in a reduction in the error of 56%). A similar reduction in the error is achieved by averaging 3 identically repeated studies. In addition, a comparison of the results of the optimal filter with the results of a standard Fourier smoothing technique revealed that the optimal filters provide both more accurate and more consistent results.

It is expected that optimally filtered gated blood pool studies will improve quantitative analysis of the data.