Reduction of image noise is an important goal in producing the highest quality medical images. A very important
benefit of reducing image noise is the ability to reduce patient exposure while maintaining adequate image quality.
Various methods have been described in the literature for reducing image noise by means of image processing, both
deterministic and statistical. Deterministic methods tend to degrade image resolution or lead to artifacts or non-uniform
noise texture that does not look "natural" to the observer. Statistical methods, including Bayesian estimation, have been
successfully applied to image processing, but may require more time-consuming steps of computing priors.
The approach described in this paper uses a new statistical method we have developed in our laboratory to reduce image
noise. This approach, Correlated-Polarity Noise Reduction (CPNR), makes an estimate of the polarity of noise at a
given pixel, and then subtracts a random value from a normal distribution having a sign that matches the estimated
polarity of the noise in the pixel. For example, if the noise is estimated to be positive in a given pixel, then a random
number that is also positive will be subtracted from that pixel.
The CPNR method reduces the noise in an image by about 20% per iteration, with little negative impact on image
resolution, few artifacts, and final image noise characteristics that appears "normal." Examples of the feasibility of this
approach are presented in application to radiography and CT, but it also has potential utility in tomosynthesis and