LINDA is an R package for automatic segmentation of chronic stroke
The method is described in Hum Brain Mapp. 2016 Apr;37(4):1405-21.
- Linux or Mac (since Oct 2016 ANTsR can run in Windows 10)
- R v3.0 or above
- The ANTsR package installed in R
- A T1-weighted MRI of a patient with (left hemispheric) stroke
Method 1 (easy, incomplete)
This method does not require any prep on your side, just paste the lines
below and all required packages will (hopefully) be installed (including
ANTsR). However, MNI transformations are not included because Github
does not accept big files. You can still force on the fly registrations
to MNI (ch2) space by setting
saveMNI=TRUE. If you frequently need
outputs in MNI space, use Method 2 below.
install.packages('devtools') devtools::install_github('dorianps/LINDA', upgrade_dependencies=FALSE)
Method 2 (complex, complete)
This method includes the MNI transformations, but works only if you have previously installed ANTsR. To do that, try:
Then download the latest LINDA release v0.5.0 and install from command line:
wget https://github.com/dorianps/LINDA/releases/download/0.5.0/LINDA_v0.5.0.zip unzip LINDA_v0.5.0.zip # this will unzip to LINDA folder R CMD INSTALL LINDA # install the LINDA folder in R
library(LINDA) filename = '/path/to/patient/T1.nii.gz' outputs = linda_predict(filename)
If you don’t specify a filename, a file dialog will allow you to choose the T1 nifti file of the patient.
outputs = linda_predict()
LINDA will run and you will see the timestamp of the various steps.
Results will be saved in a folder named “linda” in the same path where
the T1 is located. The location of these files will be returned in R
(i.e., in the
If the “linda” folder contains segmentations from an earlier run,
processing will stop immediately. Use
cache=FALSE to force overwriting
the old files. If processing is interrupted and restarted, LINDA will
use the existing files in the “linda” folder to start where it was
Available prediction models:
Currently a model trained on 60 patients from Penn is used. The internal prediction engine works with 2mm voxel resolution. This does not mean your images need to be 2mm, any resolution should work.
*(a somewhat slow example run on a single CPU core)
21:18 Starting LINDA v0.5.0 21:18 Creating folder: /data/jux/dpustina/LINDA_package/Sample_ABC/linda 21:18 Loading file: ABC_MPRAGE.nii ... 21:18 Loading template... 21:18 Skull stripping... (long process) 21:46 Saving skull stripped files... 21:46 Computing asymmetry mask... 21:50 Saving asymmetry mask... 21:50 1st round of prediction... 21:50 Running registration: SyN 21:57 Feature calculation 22:02 Lesion segmentation 22:04 Backprojecting prediction... 22:04 Saving prediction... 22:04 2nd round of prediction... 22:04 Running registration: SyN 22:11 Feature calculation 22:16 Lesion segmentation 22:17 Backprojecting prediction... 22:17 3rd round of prediction... 22:17 Running registration: SyNCC 00:44 Feature calculation 00:49 Lesion segmentation 00:50 Backprojecting prediction... 00:50 Saving 3rd final prediction in native space... 00:51 Saving probabilistic prediction in template space... 00:51 Saving probabilistic prediction in native space... 00:51 Skipping data transformation in MNI (ch2) ... 00:51 Done! 3.5 hours
Wonder what to expect? Check individual results from our 60 patients
and 45 patients Georgetown
Our users have reported quite good lesion segmentations for typical large stroke lesions, but less accurate segmentations with tiny lesions.
BrainMask.nii.gz - Brain mask used to skull strip (native space)
N4corrected.nii.gz - Bias corrected, full image (native space)
N4corrected_Brain.nii.gz - Bias corrected, brain only (native space)
N4corrected_Brain_LRflipped.nii.gz - Flipped image used to compute asymmetry mask (native space)
Mask.lesion(1)(2)(3).nii.gz - masks used for registrations (native space)
Prediction(1)(2)(3).nii.gz - lesion predictions after each iteration (template space, but 2mm)
Prediction3_template.nii.gz - final prediction (template space)
Prediction3_native.nii.gz - final prediction (native space)
Prediction3_probability_template.nii.gz - final graded probability (template space)
Prediction3_probability_native.nii.gz - final graded probability (native space)
Reg3_registered_to_template.nii.gz - Subject’s MRI, skull stripped, bias corrected, registered to template (template space)
Reg3_sub_to_template_(affine)(warp) - transformation matrices to register subject to template
Reg3_template_to_sub_(affine)(warp) - transformation matrices to backproject template to subject
Subject_in_MNI.nii.gz - Subject’s MRI, skull stripped, bias corrected, transformed in MNI space
Lesion_in_MNI.nii.gz - Lesion mask in MNI spacethe Console_Output.txt - log file replicating the console output Session_Info.txt - R environment and package versions, useful if you want to reproduce the results in the future.
MNI is a space, not a template. There are many templates in MNI space, most of which do not have the same gyri or sulci at the same coordinates. LINDA uses the CH2 template which is included in LINDA. Please don’t use other MNI templates to overlay results. They may look good from a first look, but they will be wrong. Use the CH2 template that comes with LINDA. In alternative, you can register the CH2 template to another template (i.e., ICBM 2009a) and transform back and forth the results as necessary.
Frequently Asked Questions
- What file formats are accepted?
Nifti images (.nii and .nii.gz) are accepted. Earlier LINDA versions did not accept other formats, but now any format can be read. Note that the Analyze format has unclear left-right orientation.
- Can I use it with right hemispheric lesions?
Yes, but you need to flip the L-R orientation before. After that, the prediction will work just as well.
- Can I use it with bilateral lesions?
It will likely be less accurate. One of the features LINDA uses is the left-right signal asymmetry.
- Can I use it to segment acute and subacute stroke lesions?
No, the current model is trained only on chronic stroke patients. It might be possible to segment acute stroke with models trained on acute data.
- Can I use other images: FLAIR, T2, DWI?
No, the existing model accepts only T1w. In principle, additional models can be built from other modalities (T2, FLAIR, DWI).
- Can I train a model with my own data?
This is perfectly doable, but the training script is not available online (needs some work to adapt it for widepsread use). If you want the example script used for the current model, I can send it easily, just contact me.
- Can I use LINDA to obtain registrations in MNI?
The transfer in MNI is obtained by concatenating transformations “Subject” -> “Penn Template” -> “ch2 MNI template”. Thus there are two sources of potential error. The second source of error is fixed for all subjects because our template has always the same registration on MNI. However, a fixed error is always an error. If you really want precise registration on MNI, I advise to register directly the subject to an MNI template (possibly using a group MNI template rather than the ch2).
- Will you maintain LINDA and publish new models in the future?
There is no plan, time, or funding to do this at this time. If other researchers want to contribute, this can be done easily because LINDA is open source.
The best way to keep track of bugs or failures is to open a New
Issue on the Github system.
If the algorithm proceeds without errors but the automatic segmentation
is erroneous, please send (i) your T1 image and (ii) the segmentation
produced by LINDA in native space. Try also overlaying
Mask.lesion*.nii.gz files on the T1 to check whether the brain mask is
Subscribe to LINDAfirstname.lastname@example.org to send support requests and get notified of new versions.
Other software for lesion studies
Check out the LESYMAP package for lesion to symptom mapping: https://github.com/dorianps/LESYMAP.