[Gate-users] New CASToR release 2.1

[Simon Stute]

Dear GATE users,

We are pleased to announce that the version 2.1 of CASToR has now been 
released.
It is available on the CASToR website: http://castor-project.org

The main new feature is the availability of penalized reconstructions, 
also called MAP algorithms.

The following algorithms are including:
  - One-Step-Late (from P. J. Green)
  - Penalized Preconditionned Gradient ML (from J. Nuyts et al)
  - BSREM (the version II from S. Ahn and J. Fessler, used by GE)
  - Modified EM for Penalized ML (from A. De Pierro)

The following penalties are including:
  - Markov Random Field (including all standard elements from the 
litterature such as quadratic or relative difference penalties, Bowsher 
weights, etc)
  - Median Root Prior

A complete description of these new features can be found at: 
http://castor-project.org/features

This new version also includes some bug fixes.
The list of changes from the last version can be found in the attached 
text file.

Do not hesitate to send some feedback or to share experience on the 
mailing list, even if positive

And needless to say that all these new algorithms can be used with GATE 
data !

The CASToR collaboration
-------------- next part --------------

## New features
- Add support for penalized reconstructions which includes:
  - A generalization of the penalty concept, including two types of penalty:
    - Customizable Markov-Random-Field (MRF) with:
      --> several types of neighborhoods (boxes, spheres, nearest-neighbors)
      --> proximity factors (constant, euclidian, voxel-based)
      --> similarity factors (including asymmetrical Bowsher)
      --> several potential functions (including quadratic, relative differences, log-cosh, huber, etc)
    - Median-Root penalty (MRP), including several types of neighborhoods
  - Three optimizers able to deal with any type of penalty added to the maximum-likelihood criteria:
    - One-Step-Late, from Green: OSL
    - Penalized Preconditioned Gradient Maximum Likelihood, from Nuyts: PPGML
    - Block-Sequential Regularized Expectation Maximization version II, from Ahn and Fessler: BSREM
  - One convergent optimizer dedicated to the Markov-Random-Field penalty with a quadratic potential function, from De Pierro: DEPIERRO95
  - Self-included documentation with -help-opti and -help-pnlt options.

## Bug corrections, code changes, that can affect the results in some way
- For transmission data, cm-1 is used as the default unit. This factor is added into the mutliplicative terms of the system matrix and not
  anymore applied within the forward model only (so that any call to ForwardProject() or BackwardProject() will take it into account).
- Very minor changes in Landweber and MLEM implementations for log-converted transmission data, in limit cases:
  --> No contribution of the data if the data or the model is strictly less than 1, for both algorithms.
  --> No contribution of the data if the data or the model is higher than blank value, for MLEM.
- In MLTR algorithm, remove the empirical alpha_ratio parameter. With the change of where the cm-1 unit factor is applied (see few lines
  above), it affects the relaxation parameter values that were erroneous in the previous version, due to the ForwardProject() function
  applied onto the alpha image that did not include the cm-1 unit factor. Now the relaxation factor is correct. In case you used the MLTR
  algorithm in the previous version, relaxation factors have to be divided by 10 to get the same results.
- When the time tag of the last events goes beyond the last frame, they are now ignored. Before, they were included in the last frame, which
  was a bug. So in the case of a list-mode datafile where the time stop in the header is less than the time stamp of the last events, there
  will be a difference.

## Small bug fixes
- Several but not listed...