GEMSe

Description

GEMSe is a tool for the visualization-Guided Exploration of Multi-channel Segmentation algorithms. That means, it provides tools to sample the parameter space of a multi-channel segmentation algorithm based on Principal Component Analysis (PCA), Support Vector Machines (SVM) and the Extended Random Walker (ERW). This module is the implementation to a paper published at EuroVis '16:

Bernhard Fröhler, Torsten Möller, and Christoph Heinzl, “Visualization-Guided Exploration of multi-channel segmentation algorithms”, Computer Graphics Forum (Proceedings Eurographics Conf. Vis. 2016), Vol. 35, No. 3, June 2016.

Usage Scenario

Suppose you have a multi-channel image, as for example from

• A Talbot-Lau Grating Interferometer Computed Tomography device (with Attenuation, Phase Contrast and Dark Field channels)
• An MRI image (with T1, T1c, T2, Flair channels)
• A Hyperspectral Image (with > 100 channels from different wavelength spectrum bands)

You now want to find the most suitable segmentation, taking the information from all channels into consideration. The pipeline of GEMSe to achieve this consists of two steps:

1. Preprocessing: Sampling the parameter space
2. Analysis: Exploring the result and parameter space

Preprocessing (Sampling)

Before sampling can be started, all information channels need to be loaded in open_iA. Each data channel is taken from an individual volume (3D image):

1. Open the first 3D image via File -> Open, then select Tools -> GEMSe -> GEMSe.
2. In the opening "Modality Explorer" window, add all additional channels through the "Add" button
3. Start the sampling by clicking the "Run" button in the "Sampling" row of the "GEMSE Control" widget
4. In the opening dialog, select at least an output directory. If you want you can also adapt the parameter ranges (but typically for the first run, you will not know useful restrictions, so you might want to leave the parameter ranges as they are). The default sampling method is "Latin Hypercubes", which tries to best cover the parameter space. Other sampling methods ("Random" and "Cartesian grid") are also available.
5. Start Sampling by pressing "Run".
6. Depending on the data size, the sampling can take quite some time. Be prepared to let it run over night or over the weekend; the runtime estimation before starting the sampling is quite unreliable; to get a proper estimate, wait for the first round to finish, then check the estimated time remaining.

Analysis

When sampling is done, the result exploration interface will be shown. Also, the Sampler stores all output to the specified output directory, so that you can easily load previously calculated samplings at a later point in time.

A precalculated demo dataset will be available soon.