Application of iterative and non-stationary smoothing filters for statistical noise reduction in nuclear medicine.

ABSTRACT

A new method for the enhancement of smoothing filters has been developed. It is an iterative and non-stationary (INS) process by which averaging is limited only to pixels that are 'statistically consistent' with the central pixel of the filter window. Adjustment of filter performance is achieved by the iterative calculation of filtered data variance. INS versions of the 25-point weighted and Butterworth filters have been validated and compared with their classical stationary versions using a numerical phantom and clinical data. The iterative procedure converges well after two or three iterations and produces more smoothing without any loss of contrast or resolution. Qualitative and quantitative results show that, unlike the stationary versions, the INS versions of the filters do not impede the detection of small or low-contrast objects and do preserve the contours of anatomical structures. In 99Tcm myocardial single photon emission tomography, the INS version of the Butterworth filter can be used with the same order and critical frequency, whether data have been recorded at rest (data with a high level of noise) or at stress (data with a low level of noise). In conclusion, INS filters enable noise reduction while maintaining contrast and resolution in filtered images. The procedure can be easily applied to any smoothing filter.