Segmentation based noise variance estimation from background MRI data

Accurate and precise estimation of the noise variance is often of key importance as an input parameter for posterior image processing tasks. In MR images, background data is well suited for noise estimation since (theoretically) it lacks contributions from object signal. However, background data not only suffers from small contributions of object signal but also from quantization of the intensity values. In this paper, we propose a noise variance estimation method that is insensitive to quantization errors and that is robust against low intensity variations such as low contrast tissues and ghost artifacts. The proposed method starts with an automated background segmentation procedure, and proceeds then by correctly modeling the background's histogram. The model is based on the Rayleigh distribution of the background data and accounts for intensity quantization errors. The noise variance, which is one of the parameters of the model, is then estimated using maximum likelihood estimation. The proposed method is compared with the recently proposed noise estimation methods and is shown to be more accurate.