DIPY IVIM with the variable projection method
Shreyas Fadnavis, a graduate student in Eleftheiros Garyfallidis’s lab has been diligently building a line of research on better ways to model perfusion and diffusion in the same pulsed-gradient spin echo, with a model called IVIM, which stands for intravoxel incoherent motion. This model has been around for a surprisingly long time, with Denis Lebihan’s first papers on the topic going back to the 1980’s and preceding work on DTI. Fueled by the utility of the model in clinical use-cases in oncology, there has been resurgent interest in using IVIM. This motivates improvements to the way that the model is fit. The main challenge is that the model posits that the signal is a mixture of two exponential decay components: one for diffusion, and the other for perfusion. The separation of these components is supposed to be facilitated by the measurement of very low b-values that should only be affected by perfusion, but the fact of the matter is that fitting is still rather challenging.
To address this, Shreyas and Eleftherios developed a method for fitting the model that harnesses the MIX framework to propose better initial conditions for optimization. The method is already implemented in DIPY and is now also described in a forthcoming manuscript, which I spent some time reading, commenting on and making some small suggestions on last week.