Harmonic multiresolution estimators for denoising and regularization of SPECT-PET data

This paper presents a study on the development of new multiresolution directional analysis tools for texture denoising of medical images. Multiresolution texture analysis is performed with wavelet packets and brushlet expansions to exploit spatio-temporal coherence and identify persistent anatomical structures while removing uncorrelated noise components. Denoising is performed via thresholding estimators in the transform domain. Denoising performance is evaluated quantitatively on phantom volumes and qualitatively on clinical data sets with SPECT-PET data. We show in this study that these multiresolution directional analysis tools are well adapted to the intrinsic nature of textured data and outperform traditional denoising methods. In the case of spatiotemporal data, we also show that by incorporating the time dimension directly in the analysis, we can bring into play temporal coherence between successive frames to improve denoising performance and enhance moving boundaries and structures.