Usage: 3dECM [options] dset
  Computes voxelwise eigenvector centrality (ECM) and
  stores the result in a new 3D bucket dataset as floats to
  preserve their values. ECM of a voxel reflects the strength
  and extent of a voxel's global connectivity as well as the
  importance of the voxels that it is directly connected to.

  Conceptually the process involves:
      1. Calculating the correlation between voxel time series for
         every pair of voxels in the brain (as determined by masking)
      2. Calculate the eigenvector corresponding to the largest
         eigenvalue of the similarity matrix.

  Guaranteeing that the largest eigenvector is unique and therefore,
  that an ECM solution exists, requires that the similarity matrix
  is strictly positive. This is enforced by either adding one to
  the correlations as in (Lohmann et. al. 2010), or by adding one
  and dividing by two (Wink et al. 2012).

  Calculating the first eigenvector of a whole-brain similarity matrix
  requires alot of system memory and time. 3dECM uses the optimizations
  described in (Wink et al 2012) to improve performance. It additionally
  provides a mechanism for limited the amount of system memory used to
  avoid memory related crashes.

  The perfromance can also be improved by reducing the number of
  connections in the similarity matrix using either a correlation
  or sparsity threshold. The correlation threshold simply removes
  all connections with a correlation less than the threshold. The
  sparsity threshold is a percentage and reflects the fraction of
  the strongest connections that should be retained for analysis.
  Sparsity thresholding uses a histogram approach to 'learn' a
  correlation threshold that would result in the desired level
  of sparsity. Due to ties and virtual ties due to poor precision
  for differentiating connections, the desired level of sparsity
  will not be met exactly, 3dECM will retain more connections than

  Whole brain ECM results in very small voxel values and small
  differences between cortical areas. Reducing the number of
  connections in the analysis improves the voxel values and
  provides greater contrast between cortical areas

.  Lohmann G, Margulies DS, Horstmann A, Pleger B, Lepsien J, et al.
      (2010) Eigenvector Centrality Mapping for Analyzing
      Connectivity Patterns in fMRI Data of the Human Brain. PLoS
      ONE 5(4): e10232. doi: 10.1371/journal.pone.0010232

  Wink, A. M., de Munck, J. C., van der Werf, Y. D., van den Heuvel,
      O. A., & Barkhof, F. (2012). Fast Eigenvector Centrality
      Mapping of Voxel-Wise Connectivity in Functional Magnetic
      Resonance Imaging: Implementation, Validation, and
      Interpretation. Brain Connectivity, 2(5), 265-274.

  -full       = uses the full power method (Lohmann et. al. 2010).
                Enables the use of thresholding and calculating
                thresholded centrality. Uses sparse array to reduce
                memory requirement. Automatically selected if
                -thresh, or -sparsity are used.
  -fecm       = uses a shortcut that substantially speeds up
                computation, but is less flexibile in what can be
                done the similarity matrix. i.e. does not allow
                thresholding correlation coefficients. based on
                fast eigenvector centrality mapping (Wink et. al
                2012). Default when -thresh, or -sparsity
                are NOT used.
  -thresh r   = exclude connections with correlation < r. cannot be
                used with FECM
  -sparsity p = only include the top p% (0 < p <= 100) connectoins in the calculation
                cannot be used with FECM method. (default)
  -do_binary  = perform the ECM calculation on a binarized version of the
                connectivity matrix, this requires a connnectivity or
                sparsity threshold.
  -shift s    = value that should be added to correlation coeffs to
                enforce non-negativity, s >= 0. [default = 0.0, unless
                -fecm is specified in which case the default is 1.0
                (e.g. Wink et al 2012)].
  -scale x    = value that correlation coeffs should be multiplied by
                after shifting, x >= 0 [default = 1.0, unless -fecm is
                specified in which case the default is 0.5 (e.g. Wink et
                al 2012)].
  -eps p      = sets the stopping criterion for the power iteration
                l2|v_old - v_new| < eps*|v_old|. default = .001 (0.1%)
  -max_iter i = sets the maximum number of iterations to use in
                in the power iteration. default = 1000

  -polort m   = Remove polynomical trend of order 'm', for m=0..3.
                [default is m=1; removal is by least squares].
                Using m=0 means that just the mean is removed.

  -autoclip   = Clip off low-intensity regions in the dataset,
  -automask   = so that the correlation is only computed between
                high-intensity (presumably brain) voxels.  The
                mask is determined the same way that 3dAutomask works.

  -mask mmm   = Mask to define 'in-brain' voxels. Reducing the number
                the number of voxels included in the calculation will
                significantly speedup the calculation. Consider using
                a mask to constrain the calculations to the grey matter
                rather than the whole brain. This is also preferrable
                to using -autoclip or -automask.

  -prefix p   = Save output into dataset with prefix 'p'
                [default prefix is 'ecm'].

  -memory G   = Calculating eignevector centrality can consume alot
                of memory. If unchecked this can crash a computer
                or cause it to hang. If the memory hits this limit
                the tool will error out, rather than affecting the
                system [default is 2G].

 * The output dataset is a bucket type of floats.
 * The program prints out an estimate of its memory used
    when it ends.  It also prints out a progress 'meter'
    to keep you pacified.

-- RWCox - 31 Jan 2002 and 16 Jul 2010
-- Cameron Craddock - 13 Nov 2015
-- Daniel Clark - 14 March 2016

* This binary version of 3dECM is compiled using OpenMP, a semi-
   automatic parallelizer software toolkit, which splits the work across
   multiple CPUs/cores on the same shared memory computer.
* OpenMP is NOT like MPI -- it does not work with CPUs connected only
   by a network (e.g., OpenMP doesn't work with 'cluster' setups).
* For implementation and compilation details, please see
* The number of CPU threads used will default to the maximum number on
   your system. You can control this value by setting environment variable
   OMP_NUM_THREADS to some smaller value (including 1).
* Un-setting OMP_NUM_THREADS resets OpenMP back to its default state of
   using all CPUs available.
   ++ However, on some systems, it seems to be necessary to set variable
      OMP_NUM_THREADS explicitly, or you only get one CPU.
   ++ On other systems with many CPUS, you probably want to limit the CPU
      count, since using more than (say) 16 threads is probably useless.
* You must set OMP_NUM_THREADS in the shell BEFORE running the program,
   since OpenMP queries this variable BEFORE the program actually starts.
   ++ You can't usefully set this variable in your ~/.afnirc file or on the
      command line with the '-D' option.
* How many threads are useful? That varies with the program, and how well
   it was coded. You'll have to experiment on your own systems!
* The number of CPUs on this particular computer system is ...... 2.
* The maximum number of CPUs that will be used is now set to .... 2.

++ Compile date = May 11 2021 {AFNI_21.1.07:linux_ubuntu_16_64}