it's finally here

[etpeterson]

@arokem I finally got this set up and uploaded! I think it works right, though it's still messy. IVIM_fit is the mail file and fitting_diffusion is the supporting file where the important stuff happen. Happy to explain or help, just let me know.

[arokem]

Nice. Thanks for getting that up. What do you think about having a Google Summer of Code student work on integrating this into Dipy? We're in the process of making a list of potential projects over here: https://github.com/nipy/dipy/wiki/Google-Summer-of-Code-2016

[etpeterson]

Sure, it would be great to get it integrated with dipy and it probably would be a good summer project. I could see a few stages of development and even beyond what I've done so far like adding the jacobian to the fitting to speed it up, etc. I can spend some time supervising too but it would be good if you or someone else more familiar with dipy and its development was also involved. How about I write up a description and send it along, kind of like the other examples and you let me know what you think?

[arokem]

Fantastic. If we could put you in as a co-mentor, together with me, that
would be great.

On Wed, Feb 3, 2016 at 9:41 AM, Eric Peterson notifications@github.com
wrote:

Sure, it would be great to get it integrated with dipy and it probably
would be a good summer project. I could see a few stages of development and
even beyond what I've done so far like adding the jacobian to the fitting
to speed it up, etc. I can spend some time supervising too but it would be
good if you or someone else more familiar with dipy and its development was
also involved. How about I write up a description and send it along, kind
of like the other examples and you let me know what you think?

—
Reply to this email directly or view it on GitHub
#1 (comment)
.

[quantshah]

Hi @etpeterson. I have been getting familiar with dipy coding style and diffusion imaging as I started contributing to open source recently. GSoc is a great opportunity to get involved with the open source community and I am interested in contributing to dipy. I worked on #932 and started looking at implementations of Rohde and CHARMED as @arokem suggested. I would like to work on this meanwhile as I gain better understanding of the papers suggested by arokem. I used your code to make IVIM_fit a function and wrote a basic ipython example which outputs an image file - https://github.com/sahmed95/IVIM_fitting/tree/dipy-ivim. Let me know if you can mentor me to integrate this with dipy and if I should open a PR

[etpeterson]

Hi @sahmed95. It would be a huge help to get this going, and I'm happy to mentor and work on this as well. I think a PR would be a great place to start but I'm also fairly new to dipy so @arokem is the real expert with the integration so I'd be happy to hear what he thinks too. IMHO the place to start is a PR and model it after the free water DTI PR even though that is not complete yet (#825).

Regarding the code itself, I think a lot of the options in the code should be removed both because it's excessively complicated and it doesn't fit into dipy as well. I think a two-stage fitting where the high b-values are fit linearly first and then the full model is fit is a good way to start. This two-stage fitting is similar to what #825 has so that's why I mention it.

[quantshah]

Thanks. I have made a separate branch and worked out a simple fit() to run your script as a function. I wanted @arokem and you to know before I started changing the code. @arokem suggested that "it might be worth putting in a few plots. For example, what are the parameters in different parts of the image, how well does the model fit the data, etc."

I will try out a few plots, refactor the code a little more, and open a PR here. Btw the link to #825 is taking me to my own PR. Did u mean #835? RafaelNH's PR?

[etpeterson]

Yes sorry, you're right #835 is the one.

I definitely agree with @arokem on the plotting. It's kind of a delicate fit sometimes so plots are invaluable, especially when the results aren't what you expected.

[quantshah]

Hi, couldn't work much through the weekend but here are some plots - Ipython Notebook using the plotting code @etpeterson wrote in ivim_fit. Each is a plot of the items in np.ndindex(shape3d). Will work more on cleaning up the code this week and then open a PR. @arokem.

[arokem]

The notebook is great -- exactly the kind of reality check we need at this
point. I'll defer to @etpeterson to say whether these plots look as they
should -- I don't have much experience with this particular model. One
thing that occurs to me is that you might want to create some functions
that simulate the IVIM model results without noise. This is invaluable for
testing the model fitting procedure, and is also a good way to teach/learn
about the model in small self-contained scripts. Here, it will also be
useful to make sure that you get the same results as you start refactoring
the model code to match the dipy API (e.g., inheriting from ReconstModel
and ReconstFit).

On Sun, Mar 6, 2016 at 3:42 PM, Shahnawaz Ahmed notifications@github.com
wrote:

Hi, couldn't work much through the weekend but here are some plots - Ipython
Notebook
https://github.com/sahmed95/IVIM_fitting/blob/dipy-ivim/ivim_example.ipynb
using the plotting code @etpeterson https://github.com/etpeterson wrote
in ivim_fit. Each is a plot of the items in np.ndindex(shape3d). Will work
more on cleaning up the code this week and then open a PR. @arokem
https://github.com/arokem.

—
Reply to this email directly or view it on GitHub
#1 (comment)
.

[etpeterson]

Hey @sahmed95, that is great. The plots are really the best way to see if it's fitting well or not. It looks like most of the fits worked but the b-values are for the most part higher than what we really want for IVIM. Here are the values I used for my experiment: 0, 10, 20, 30, 40, 60, 80, 100, 120, 140, 160, 180, 200, 300, 400, 500, 600, 700, 800, 900, 1000. This is because IVIM is focused on slow flow, and diffusion coefficient of free water (more static or slowly diffusing that slow flow) is ~3e-3 so if you plot that out it drops to near zero by b>1000 which means that flowing blood will decay even faster, hence the focus on lower diffusion values.

@arokem had a good suggestion to simulate data and then try to fit it. If you use the b-values above, and reasonable other parameters would be: 0.01<f<0.2, 1e-2<D*<3e-3, and 1e-4<D<1e-5.

[quantshah]

Thanks @arokem @etpeterson for your inputs. Should I try plotting more with the parameters suggested by Eric?

"create some functions that simulate the IVIM model results without noise"

How?

What can be the next step here? Should it be similar to dti or dki fitting? ReconstModel works with gradient table (gtab). Should the IVIM also inherit this and work with gtab?

[etpeterson]

Basically you want to create diffusion signal curves using the equation S(b)=S0(fe^-bD*+(1-f)e^-bD). If you use the values I suggested above then you'll get a signal value S for each b-value. And this you can fit as if it's any other signal coming from an image series. I think you should basically be able to use the fitting functions you created to fit these curves. And hopefully the fitted values should be similar to what was generated. Does that make sense?

[quantshah]

Thanks. It makes sense now after I could get a quick read of the paper you sent. I will generate the data and try plotting.

[quantshah]

@etpeterson @arokem
I generated signals corresponding to the b-values and taking random parameters and got a plot of signal vs bvalues. Link

I am not sure about what S0 values to take so I took random values between 0-1. Next step is fitting.

To fit, I have to get an img from the generated signals similar to an .nii file img as your code takes that as input. Or I can take the part where you fit in begin processing and modify it such that img takes the signal values generated. Not sure how to go about it. Just wanted to know if I am on the right track.

[etpeterson]

You're on the right track.

The S0 value can be anything really, it's just a scaling but I always normalized to the maximum value so S0 was always close to 1, probably 0.95<S0<1.05.

And the other change is that for each parameter set you need to generate the intensities at each b-value. In other words in your generate_signal function you'd input single numbers for the parameters like f, and the vector of b-values so the function would return a vector of values the same size as the b-values.