The toolbox "LST: Lesion Segmentation Tool" is an open source toolbox for SPM that is able to segment T2 hyperintense lesions in FLAIR images. Originally developed for the segmentation of MS lesions it has has also been proven to be useful for the segmentation of brain lesions in the context of other diseases, such as diabetes mellitus or Alzheimer's disease.
Currently, there are two algorithms implemented for lesion segmentation. The first, a lesion growth algorithm (LGA, Schmidt et al., 2012), requires a T1 image in addition to the FLAIR image. The second algorithm, a lesion prediction algorithm (LPA), requires a FLAIR image only. As a third highlight a pipeline allowing the longitudinal segmentation is implemented. In addition, the toolbox can be used to fill lesions in any image modility. We hope that these algorithms will be able to contribute to current research in MS and other disciplines.
The toolbox was developed by a cooperation of the following organizations: Morphometry Group, Department of Neurology, Technische Universität München (TUM), Munich, Germany; Department of Statistics, Ludwig-Maximilians-University, Munich, Germany; and Structural Brain Mapping Group, Departments of Neurology and Psychiatry, Friedrich-Schiller-University, Jena, Germany. Maintainer of the toolbox is Paul Schmidt.
Heart of the toolbox is the lesion growth algorithm (LGA, Schmidt et al, 2012). This algorithm is able to segment T2-hyperintense lesions from a combination of T1 and FLAIR images. It first segments the T1 image into the three main tissue classes (CSF, GM and WM). This information is then combined with the FLAIR intensities in order to calculate lesion belief maps. By thresholding these maps by a pre-chosen initial threshold (kappa) an initial binary lesion map is obtained which is subsequently grown along voxels that appear hyperintense in the FLAIR image. The result is a lesion probability map.
Since the release of version 2.0.1 the segmentation of the T1 image is obtained by standard SPM routines, thus the VBM toolbox is no longer required.
A disadvantage of this unsupervised algorithm is the choice of the initial threshold. Different kappa-values yield different segmentation results. However, in order to make the process of finding the optimal threshold as easy as possible we implemented a routine that can be used if reference segmentations for a few images are available. Otherwise, visual inspection of the segmentations for a set of kappa-values should be sufficent as well.
As an alternative to the LGA the toolbox provides a further lesion segmentation algorithm, the lesion prediction algorithm (LPA). Its advantages over the LGA is that (a) it only requires a FLAIR image and (b) no parameters need to be set by the user. The method behind this algorithm can be found in Section 6.1 here. The LPA is usually faster and in general more sensitive than the LGA, so give it a try!
The LPA was trained by a logistic regression model with the data of 53 MS patients with severe lesion patterns. Binary lesion maps of these patients were used as response values. As covariates a similar lesion belief map as for the LGA was used as well as a spatial covariate that takes into account voxel specific changes in lesion probability. Such a high dimensional model cannot be estimated by standard procedures, so we used a novel approach for fitting large-scale regression models, for details see here. The parameters of this model fit are then used to segment lesions in new images by providing an estimate for the lesion probability for each voxel.
This algorithm yields useful results even when applied to images obtained on different scanners. However, if your images differ greatly from the images in our training data set the results may be limited. In this case it may be useful to train the algorithm with your data. Please let us know if you are interested in further information.
We implemented a longituidinal pipeline that is able to compare lesion probability maps of multiple time points. This tool labels significant changes that may have been occured between two time points.
The pipeline proceeds by comparing all consecutive time points in an iterative manner. It decides if changes in lesion structure are significant or due to natural variations of the FLAIR signal. Non-significant changes are labeled as lesions in both probability maps, thus, probability lesion maps are corrected within this procedure and may differ from the ones that served as input. As a final result, lesion change labels are produced for all consecutive time points. In these images the three possible cases decrease, no change and increase are labeled by the numbers 1, 2, and 3, repsectively.
In addition, a lesion change plot is constructed. This plot shows the lesion volumes for both time points of all segmented lesions. In this way it is easy to recognize how the lesion structure has been changed, i.e. if the change occured by the appearence of new lesions, the disappearence of old lesions, by the change of already existing lesions, or a combination of these possibilities.
Once lesion probability maps have been calculated they can be used to fill lesions in MR images. The current filling algorithm uses local information instead of global intensity distributions. This allows accurate filling of lesions even in images that are currupted by Bias field.
All main functions in LST are able to automatically produce HTML reports. Although it needs a while to produce these reports we recommend it as it makes it easier to check the results. Reports of different subjects or analyzes can easily be merged.
The LST toolbox is available to the scientific community under the terms of the GNU General Public License. A copy of the GNU General Public License is received along with this toolbox.
We put a lot of effort in the creation of this toolbox, so please cite it accordingly.
You can download the newest version of the toolbox (version 2.0.15) here.
Are you new to LST? Then you may want to have a look at the documentation.
Do you have a question that is not answered here? Feel free to contact us.
Detailed information on how to cite this toolbox can be found in the documentation.
There is no general answer to this question. The LGA is well established and yields good results. The necessity to adjust the initial threshold allows to account for different image protocols. However, choosing the best initial threshold can be time consuming and therefore a real burden. Here, the LPA has a clear advantage since no parameters need to be set by the user. For our data we observed that the LPA is more sensitive with respect to white matter that appears dirty in T1 and slightly hyperintense in FLAIR images. This can be advantageous for lesion filling. However, we also observed that it can be too sensitive in some situations. With respect to processing time the LPA needs about half as much time as the LGA.
As results of both algorithms are not directly comparable we recommend you to stay with either the LGA or the LPA, that is do not mix up the different algorithms within one study.
If you want to use our toolbox for the evaluation of your own lesion segmentation algorithm we kindly ask you to consider both algorihthms, the LGA and the LPA. In addition, please determine the optimal initial threshold for the LGA.
For the actual comparison, please threshold the lesion probability maps with a value of 0.5 and use the resulting binary lesion maps. You can then use the Dice coefficient or other similarity measures for your evaluation.
Yes, see the help of ps_LST_lga, ps_LST_lpa, ps_LST_long, ps_LST_lesfill and ps_LST_doit for details. You may want to execute the commmand 'spm_jobman('initcfg')' beforehand.
The longitudinal toolbox uses the information of two (or more) time points for the segmentation of individual images. Therefore, more information is considered which may result in different segmentations. In addition, the lesion maps for time points 2, ..., m are in the same space as the lesion maps of the first time point.
No. If you'd like to obtain normalized lesion maps we recommend you to proceed as follows: first, segment lesions and fill the coresponding T1 or FLAIR images, depending on the algorithm you used for lesion segmentation. Then, use SPM or VBM routines in order to calculate deformation fields from the filled images and apply these fields to the segmented lesion maps.
Some users experienced the problem that the creation of PNG images for the HTML report failed. Unfortunately, we have no solution to this problem, yet. If you experienced this problem it would be great if you could send us some information about the system you use, i.e. operating system, MATLAB version, error message in the HTML report, do you run MATLAB locally or on a server, ... Thank you!
Yes. The LPA has been trained with data obtained at the Department of Neurology, Technische Universität München. Thus, it is optimized for images that follow the same scanner protocol. If the LPA does not perform well on your data it may be worth it to train the algorithm using a subset of your data. Please contact Paul if you are interested in further information.
Please consult all information we provide here (the about section, the docs as well as the FAQs) before you contact us.