Good evening,

My understanding is that the final WMH output (RFREG_pmap) should be a binary output with WMH as 1 and other tissue as 0. However, my outputs are not looking like a binary output, as other tissues are still shown in the final output. I would assume the volume quantification, below, is also incorrect due to this. Is this simply an issue with my visualization method, or is the pmap output incorrect for some reason? This subject shown below is the first individual in your sample dataset.

Study: example Subject: 01
EV: 45100.518722
EV-Deep: 28915.735052
EV-Periventricular: 16184.783671



Thank you very much for your help with this.

Best,
John

Hello. Firstly, I apologize for this very late reply.
Note that REFEG_out is the file containing the final hyperintense regions, and RFREG_pmap is the output of the learning algorithm (before applying the WhyD_quant script that quantifies hyperintense regions). Please check if the RFREG_out is also not making sense ?!
Also, it looks like the image has lot of hyperintense regions, and in general the tissue map itself looks bright.
There are a few things you can try -- changing the default parameters of preprocessing (SPM) steps, changing around the quantification parameters (pmap_cut and clean_th). Please refer to the documentation for what/how these parameters can be changed.
Lemme know if things dont seem to work out. Thanks for your interest!

Hello.
Pardon me, I understood your question incorrectly!
You are right in the semse that the _pmap_ does give the final hyperintense regions.
Because of the higher overall brightness of the tissues, the algorithm is picking up most of the white matter as relevant for hyperintense regions of interest. Further, _pamp_ is a probability map i.e, each voxel corresponds to a signal between 0 and 1 representing how strong the corresponding voxel is interms of being classified as a hyperintensity. Hence, we see the very bright regions along with slightly darker ones (in the ex image you have attached here with).

The solution is still to play around with the preprocessing and quantification hyper-paramters. Changing the bias correction level and smoothing factors in the SPM preprocessing pipeline (refer WHyD_preproc.m) would also be relevant.
Further, changing the pmap_cut and clean_th for quantification would help neglect the darker regions on the _pmap_ image (i.e., while computing EV values the low probable darker regions on _pmap_, which ideally popped up because of the issue of overall brightness being high, will not be used). Please refer to the documentation.

Thank you for your response!

Can I send you a couple of our images and could you maybe help us figure out which parameters we should adjust and what they should be?

Sure. No problem.

Many thanks!! Attached is the data.

Hello,

Thank you very much for your continued help in getting your program to work with my project.
One further question came up in my most recent subject runs:
1) How can we get a visual representation of the thresholded pmap? In other words, how do you decide what to set your Probability Map Cut Value to if the final "pmap" output is not affected by this option? We would like to see how the output will be changed by changing the different hyperparameters ("cleaning threshold" and "probability map cut value").

Thank you very much!

Happy new year!

Is there any progress on our data? Thanks.

Hello.

I have been off during the new years for quite some time. I'm just getting back.
I apologize for the delay, but I will get back as soon as I have some progress looking at the data. Thanks for your patience!

Hi,

Is there any progress on our data? Thanks!

Hi,

I'm sorry to bother you again. My boss is kinda driving me crazy and we're
wondering if you could take a look at our data when you have a time, so we
could set up the right parameters to our dataset. Thanks!

I've got the same problem as you, John.

I've figured, after hours of debugging, that the problem lies in the script WhyD_detect.m of the new version (W2MHS v1.3 [Linux]).

If you use the older version of it (e.g, W2MHS v1.1 [Linux] Source and Training Data) with the following change,



then, you will get the expected WMH map.

Right now, I am trying some hybrid version to figure out where they differ essentially, but with no success at present moment.

Hi, Thiago. Thanks for the information. 

I switched to a toolbox implemented in SPM called "LST" that has the functionality I'm looking for. Maybe take a look at that if you're still having trouble with the W2MHS toolbox.

Best,
John

Hi, John. Thanks for pointing LST out!

Now that everything is working properly, I am using both (W2MHS and LST) in order to compare my results and see which one performs best within my subjects of study.

Best,
Thiago

Hello John and Thiago,
Thanks for pointing out the correction. We were unaware of the discrepancy between the two versions. To that end, we were debugging from preprocessing and other hyper-patameters front all this while, to solve the issues John was facing.
We will update the codes and toolbox as such, based on the pointed out correction, in the coming few days.
Thanks for your interest. We hope the corrected codes will not throw out any more issues.

Hello Thiago,
We will perform a comparison to make sure there are no other discrepancies between the versions, and upload the corrected one soon.
Please feel free to get back if you need any help with regard to the hybrid version you were looking into. Thanks a lot for your interest!