Abstract
Preprocessing stage for denoising is a crucial task in image analysis in general, and in computer-aided diagnosis using medical images in particular. Standard acquisition of Magnetic Resonance Images (MRI) presents statistical Rician noise which degrades the performance of the image analysis. This paper presents a new technique to reduce Rician noise of brain MRI. The new method for noise filtering is achieved in the discrete Wavelet Packets Transform (WPT) domain. Four methodologies for thresholding the detail coefficients in the same 2D WPT domain have been experimented considering two scenarios (with and without a previous adaptive Wiener filtering in the spatial domain). Best quantitative and qualitative results have been obtained by the new method presented in this work (specifically tailored for brain MRI), which is adaptive to each subband and dependent on the data. It has been compared with other traditional methods considering the Signal to Noise Ratio (SNR), Normalized Cross Correlation (NCC) and execution time (∼ 0.1 s/slice). A complete dataset of structural (T1-w) brain MRI of the BrainWeb database has been used for experiments. An important aspect is that these experiments with synthetic images proved that the common prior adaptive Wiener filtering often used by many authors is a dispensable procedure.