Module: Elastic Volume Alignment ()
This module is experimental. Experimental features offer a preview of features to come in future releases of . Note that no technical support will be provided for experimental features.The Elastic Volume Alignment module takes as input a 3D image data set, that is regarded as a set of 2D images, stacked in the Z-direction. The module computes a deformation for each section, based on image intensities, such that it best matches its neighboring sections.
The module implements the method as described in:
V.J. Dercksen, C. Brüß, D. Stalling, S. Gubatz, U. Seiffert and H.-C. Hege (2008): Towards automatic generation of 3D models of biological objects based on serial sections. Visualization in Medicine and Life Sciences, L. Linsen and H. Hagen and B. Hamann (eds.), Springer-Verlag Berlin Heidelberg, pp. 3-25, doi 10.1007/978-3-540-72630-2.
and the references therein, in particular:
J. Modersitzki (2004): Numerical methods for image registration. Oxford University Press, New York.
The first and last section in the stack remain unchanged. The module supports the use of a multi-resolution image pyramid to reduce computation time and increase robustness. The user may provide up to five images of decreasing resolution. In addition to being matched with its direct neighbors, a slice may also be matched with a weighted average of a specified number of neighboring slices. They should ideally increase the robustness by reducing the impact of sections containing artifacts. It is recommended to rigidly align the stack before using this module.
Data [required]
3D image data, representing a set of 2D images, stacked in the Z-direction. This image has the highest resolution in the multi-resolution pyramid (level 0).Data level1 [optional]
Downsampled version of the original image data; second level in the multi-resolution pyramid (level 1).Data level2 [optional]
Downsampled version of the original image data; third level in the multi-resolution pyramid (level 2).Data level3 [optional]
Downsampled version of the original image data; fourth level in the multi-resolution pyramid (level 3).Data level4 [optional]
Downsampled version of the original image data; fifth level in the multi-resolution pyramid (level 4).
Material parameters
Material parameters (Lamé constants). This registration method models the image as a sheet of material that is subject to forces, defined by the image intensities, using the Navier-Lamé equations. The material properties determine how the sheet reacts to these forces. Typically, is set to 0. In this case, the sections become more rigid on increasing value of .Parameter settings
The values of the Material parameters can either be set fixed or estimated automatically. The automatic estimation is described in Modersitzki(2004).Interpolation
Type of image gray-value interpolation.Max step
Maximal displacement of a voxel per iteration.Iterations
The number of iterations for the computation at the highest resolution (level 0) and an increment per level. The total number of iterations at level n is, therefore, iterationsLevel0 + n*increasePerSubLevel.Max. improvement factor
The computation at the current level stops when the relative error is larger than the value specified in this port. A relative error smaller than 1 means that the current step resulted in an improvement of the sum of squared gray-value difference measure (larger than 1 means a decline).Window width
The driving force behind the registration is the difference in intensity values between neighboring sections, quantified by the sum of squared gray-value differences (SSD). In order to reduce the impact of a single distorted neighboring section, the SSD can be computed between a section and a weighted average of a number of neighbors. The number of neighboring sections to consider can be specified for each level separately in this port. The default is one, i.e., only direct neighbors are considered.Weighting
The type of weighting to use when considering multiple neighbors for computing the SSD.Penalize bad slices
Sections that contain severe artifacts should in principle have a larger value of the sum of squared gray-value difference (SSD) value. By turning on this option, such sections are marked as "bad" when their SSD value is larger than the mean SSD of all sections + two standard deviations. When computing the weighted SSD of multiple neighbors, the weight of such sections is decreased by a penalty factor, thus reducing its influence on neighboring slices. Turning on this feature only has an effect when the window width is larger than 1.Penalty factor
The penalty factor for distorted slices when penalize bad slices is turned on.Threads
Number of threads that should be used for computation.Max. memory (MB)
The maximum amount of memory that this module is granted. The expected memory footprint for all input and computed data is determined. When more memory is required, the user is asked to provide a file name for all result data sets that exceed the amount of available memory. Data from these data sets is then read from disk when required. Note, that this out-of-core processing significantly decreases performance.