Home

Welcome to the microstruktur wiki!

Improvements to make:

• CSD, when response function is static precompute the prepared rotational harmonics upon optimizer instantialization.
• implement classical multi-shell single-model optimizer for CSD alla Tournier 2007. Check equivalent (or very similar) solutions with cvxpy implementation.
• The general-purpose csd optimizer should be a hybrid between classical and cvxpy optimizers. If there is only one model (or volume fractions are fixed for multiple models), the optimization should use the faster classical optimization. If also volume fractions must be optimizer, then the cvxpy solver should be used.
• dhollander16 to return model-free rh coefficient models, possibly using tax14 calibration for wm voxels.
• remove spherical harmonics dependency on dipy. only depend on visualization ondipy.
• when you did a search and replace to put parentheses around the prints you left the .format out. Add test that does the print_acquisition_summary to catch this in the future.
• spherical mean model explanation should include astro-models.
• fitting does not give error if there is no b0.
• ODI and beta_fraction are optimization parameters. The flags to do this should be included in the call for the Watson/BinghamDistributedModel. i.e. if these are turned off they use the regular kappa/beta parameters.
• huge diameters gives NA signal attenuation without error message.
• volume fractions cannot be fixed while using MIX.
• add print_model_summary function. Should output model composition and parameter optimization settings.
• set custom parameter link and set custom replaced parameter. make parameter link a list of 4, and an optimized parameter a list of 5, with the last item being the name appendix for the optimized parameter.
• fod optimizer cannot replay custom parameters now.
• still fix brute2fine fixed parameters.
• callaghan sphere
• van gelderen plane and capped cylinder
• separate parameters links in separate python file in utils.
• rewrite docs for DD1
• Spherical harmonics should always check thst all non linear parameters are fixed.

Issues for later:

• psi brute optimization should take into account that it's circular.
• peaks with corresponding peak intensities.
• mu parameter ranges should not be constrained somehow.
• Brute2Fine should estimate proper grid for mu instead of theta-phi grid in equal steps.
• function "print_relevant_references" that can model-dependently print the references that are related with the current model composition.
• l0.5-norm for mix as in (Zhu, Xinghua, et al. "Model selection and estimation of multi-compartment models in diffusion MRI with a Rician noise model." IPMI. 2013.).
• find optimal parameters for Ns and NSpherePoints for arbitrary model setup.
• make bingham and watson both sh-order and sphere dependent (so less points are sampled at lower sh-orders)
• visualize model using graph nodes from optimized parameters -> linked / preset parameters -> input for models -> combined signal. dask uses http://www.graphviz.org/.
• Implement sparse dictionary fitting using http://spams-devel.gforge.inria.fr/ as in AMICO.
• deep q-space learning Golkov et al. https://sci-hub.tw/10.1109/TMI.2016.2551324
• Analytic gradients in optimization.
• Bingham is currently normalized using spherical mean instead of analytically. The implementation of the generalized hyperconfluent function of Matrix argument is required, see http://www-math.mit.edu/~plamen/files/hyper.pdf. This is also of consequence for closed form watson/bingham dispersed stick implementations, see Appendix A in https://sci-hub.io/10.1016/j.neuroimage.2012.01.056.
• directly estimate bingham / watson in spherical harmonics https://arxiv.org/pdf/1501.04395.pdf
• Implement Matrix-Variate Distribution models and DIAMOND (Scherrer).
• Implement Spherical harmonics as a non-parametric density for micro-environments.

Examples not to be included in first version

• AxCaliber with restricted extra-axonal diffusion [Burcaw et al. 2015]