Rotation- and reflection-encompassing multispectral nonlocal means denoising filter for magnetic resonance imaging

Image denoising is used extensively in medical image processing. Compared to other advanced filters, nonlocal means (NLM) shows excellent performance while being straightforward to implement. NLM takes advantage of structural redundancy in images by comparing local neighborhoods (local patches) around voxels throughout the whole image. The index voxel is then restored based on a weighted average of all voxels, with the weighting dependent on the degree of similarity. We have extended this by exploiting multispectral (MS) similarity in related images, obtained with a sequence of imaging parameters (1). In addition, we have developed an adaptive user-independent method to define similar voxels (2). The filter exhibits excellent edge preservation and noise reduction in magnetic resonance (MR) images and has been applied to improve parameter determination (3, 4). Here, we introduce a rotation- and reflection-encompassing implementation of our filter, motivated by the fact that patches may be similar only upon rotation. Our results demonstrate the improved performance as compared to our original implementation. This filter has wide applicability in imaging.
1. IEEE TMI 2017;36:181-93
2. MRI 2019;55:133-9
3. JNM 2018;309:121-131
4. JON 2018;28:640-649