DIPY: Automatic Denoising and Robust Brain Extraction

DIPY currently uses non-local means approach for denoising MR data, in which we need to estimate the noise variance of the signal which is often a bit troublesome, also it does not make the full use of the directional information in dMRI datasets. This project proposes to use a more robust and efficient method for automatically denoising diffusion MRI and structural MR datasets, using Local-PCA. Along with an accurate implementation of L-PCA and its adaptation to Rician noise, the project will also aim at optimising the implementation of L-PCA using Cython. After implementing L-PCA, a method for robust brain extraction needs to be developed. DIPY’s median OTSU based implementation is known not work so well with non-echo planar diffusion imaging (non-EPI dMRI) data. There are few possible ways to improve this, one is to generate labels and have weighted median OTSU. Another idea involves using some version of patch based segmentation using image library constructed from previously annotated images and using it as a reference for the subsequent extraction.