Google Summer of Code 2016 : Intravoxel Incoherent Motion Techniques in Dipy

Gsoc-final.md

Google Summer of Code 2016

Implementing Intravoxel Incoherent Motion (IVIM) Techniques in Dipy

-Shahnawaz Ahmed (shahnawaz.ahmed95@gmail.com)

Dipy is a python library for analysis of diffusion-weighted MRI (dMRI). Diffusion patterns can reveal microscopic details about tissue architecture and is used in clinical as well as neuroscience research. The intra-voxel incoherent motion (IVIM) model describes diffusion and perfusion in the signal acquired with diffusion MRI as :

S(b) = S0(f e^(- b D*) + (1 - f) e^(-b D))

Commits : https://github.com/sahmed95/dipy/commits/ivim_linear/dipy

PR and discussion : dipy/dipy#1110

Blog : http://sahmed95.blogspot.in/

Application : https://drive.google.com/file/d/0B2sGDIE9AVckUVZZYnZ5U0h2NTQ/view

Comments and future work

Scipy's leastsq is used to fit the model and for versions of Scipy > 0.17 least_squares is used which supports setting of bounds. The decission to use leastsq was taken after a comparison of run time with scipy.optmize.minimize. The time profiling can be seen in this discussion : http://stackoverflow.com/questions/6779383/scipy-difference-between-optimize-fmin-and-optimize-leastsq.

For noisy data, sometimes non-linear least squares fitting fails and for such cases the results from a linear fit are considered. It will be interesting to see how one can deal with noise in the data and get a more robust fitting.

Future work includes developing multiprocessing capabilities for the current multi_voxel decorator in Dipy which can be used to fit a model to a large number of voxels simultaneously and hence can act as a simple unified tool to enable multiprocessing for current and future models.