int main( int argc , char *argv[] )
{
THD_3dim_dataset *xset , *cset, *mset=NULL ;
int nopt=1 , method=PEARSON , do_autoclip=0 ;
int nvox , nvals , ii, jj, kout, kin, polort=1 ;
int ix1,jy1,kz1, ix2, jy2, kz2 ;
char *prefix = "degree_centrality" ;
byte *mask=NULL;
int nmask , abuc=1 ;
int all_source=0; /* output all source voxels 25 Jun 2010 [rickr] */
char str[32] , *cpt ;
int *imap = NULL ; MRI_vectim *xvectim ;
float (*corfun)(int,float *,float*) = NULL ;
/* djc - add 1d file output for similarity matrix */
FILE *fout1D=NULL;
/* CC - we will have two subbricks: binary and weighted centrality */
int nsubbriks = 2;
int subbrik = 0;
float * bodset;
float * wodset;
int nb_ctr = 0;
/* CC - added flags for thresholding correlations */
double thresh = 0.0;
double othresh = 0.0;
int dothresh = 0;
double sparsity = 0.0;
int dosparsity = 0;
/* variables for calculating degree centrality */
long * binaryDC = NULL;
double * weightedDC = NULL;
/* variables for histogram */
hist_node_head* histogram=NULL;
hist_node* hptr=NULL;
hist_node* pptr=NULL;
int bottom_node_idx = 0;
int totNumCor = 0;
long totPosCor = 0;
int ngoal = 0;
int nretain = 0;
float binwidth = 0.0;
int nhistnodes = 50;
/*----*/
AFNI_SETUP_OMP(0) ; /* 24 Jun 2013 */
if( argc < 2 || strcmp(argv[1],"-help") == 0 ){
printf(
"Usage: 3dDegreeCentrality [options] dset\n"
" Computes voxelwise weighted and binary degree centrality and\n"
" stores the result in a new 3D bucket dataset as floats to\n"
" preserve their values. Degree centrality reflects the strength and\n"
" extent of the correlation of a voxel with every other voxel in\n"
" the brain.\n\n"
" Conceptually the process involves: \n"
" 1. Calculating the correlation between voxel time series for\n"
" every pair of voxels in the brain (as determined by masking)\n"
" 2. Applying a threshold to the resulting correlations to exclude\n"
" those that might have arisen by chance, or to sparsify the\n"
" connectivity graph.\n"
" 3. At each voxel, summarizing its correlation with other voxels\n"
" in the brain, by either counting the number of voxels correlated\n"
" with the seed voxel (binary) or by summing the correlation \n"
" coefficients (weighted).\n"
" Practically the algorithm is ordered differently to optimize for\n"
" computational time and memory usage.\n\n"
" The threshold can be supplied as a correlation coefficient, \n"
" or a sparsity threshold. The sparsity threshold reflects the fraction\n"
" of connections that should be retained after the threshold has been\n"
" applied. To minimize resource consumption, using a sparsity threshold\n"
" involves a two-step procedure. In the first step, a correlation\n"
" coefficient threshold is applied to substantially reduce the number\n"
" of correlations. Next, the remaining correlations are sorted and a\n"
" threshold is calculated so that only the specified fraction of \n"
" possible correlations are above threshold. Due to ties between\n"
" correlations, the fraction of correlations that pass the sparsity\n"
" threshold might be slightly more than the number specified.\n\n"
" Regardless of the thresholding procedure employed, negative \n"
" correlations are excluded from the calculations.\n"
"\n"
"Options:\n"
" -pearson = Correlation is the normal Pearson (product moment)\n"
" correlation coefficient [default].\n"
#if 0
" -spearman = Correlation is the Spearman (rank) correlation\n"
" coefficient.\n"
" -quadrant = Correlation is the quadrant correlation coefficient.\n"
#else
" -spearman AND -quadrant are disabled at this time :-(\n"
#endif
"\n"
" -thresh r = exclude correlations <= r from calculations\n"
" -sparsity s = only use top s percent of correlations in calculations\n"
" s should be an integer between 0 and 100. Uses an\n"
" an adaptive thresholding procedure to reduce memory.\n"
" The speed of determining the adaptive threshold can\n"
" be improved by specifying an initial threshold with\n"
" the -thresh flag.\n"
"\n"
" -polort m = Remove polynomical trend of order 'm', for m=-1..3.\n"
" [default is m=1; removal is by least squares].\n"
" Using m=-1 means no detrending; this is only useful\n"
" for data/information that has been pre-processed.\n"
"\n"
" -autoclip = Clip off low-intensity regions in the dataset,\n"
" -automask = so that the correlation is only computed between\n"
" high-intensity (presumably brain) voxels. The\n"
" mask is determined the same way that 3dAutomask works.\n"
"\n"
" -mask mmm = Mask to define 'in-brain' voxels. Reducing the number\n"
" the number of voxels included in the calculation will\n"
" significantly speedup the calculation. Consider using\n"
" a mask to constrain the calculations to the grey matter\n"
" rather than the whole brain. This is also preferrable\n"
" to using -autoclip or -automask.\n"
"\n"
" -prefix p = Save output into dataset with prefix 'p', this file will\n"
" contain bricks for both 'weighted' or 'degree' centrality\n"
" [default prefix is 'deg_centrality'].\n"
"\n"
" -out1D f = Save information about the above threshold correlations to\n"
" 1D file 'f'. Each row of this file will contain:\n"
" Voxel1 Voxel2 i1 j1 k1 i2 j2 k2 Corr\n"
" Where voxel1 and voxel2 are the 1D indices of the pair of\n"
" voxels, i j k correspond to their 3D coordinates, and Corr\n"
" is the value of the correlation between the voxel time courses.\n"
"\n"
"Notes:\n"
" * The output dataset is a bucket type of floats.\n"
" * The program prints out an estimate of its memory used\n"
" when it ends. It also prints out a progress 'meter'\n"
" to keep you pacified.\n"
"\n"
"-- RWCox - 31 Jan 2002 and 16 Jul 2010\n"
"-- Cameron Craddock - 26 Sept 2015 \n"
) ;
PRINT_AFNI_OMP_USAGE("3dDegreeCentrality",NULL) ;
PRINT_COMPILE_DATE ; exit(0) ;
}
mainENTRY("3dDegreeCentrality main"); machdep(); PRINT_VERSION("3dDegreeCentrality");
AFNI_logger("3dDegreeCentrality",argc,argv);
/*-- option processing --*/
while( nopt < argc && argv[nopt][0] == '-' ){
if( strcmp(argv[nopt],"-time") == 0 ){
abuc = 0 ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-autoclip") == 0 ||
strcmp(argv[nopt],"-automask") == 0 ){
do_autoclip = 1 ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-mask") == 0 ){
mset = THD_open_dataset(argv[++nopt]);
CHECK_OPEN_ERROR(mset,argv[nopt]);
nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-pearson") == 0 ){
method = PEARSON ; nopt++ ; continue ;
}
#if 0
if( strcmp(argv[nopt],"-spearman") == 0 ){
method = SPEARMAN ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-quadrant") == 0 ){
method = QUADRANT ; nopt++ ; continue ;
}
#endif
if( strcmp(argv[nopt],"-eta2") == 0 ){
method = ETA2 ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-prefix") == 0 ){
prefix = strdup(argv[++nopt]) ;
if( !THD_filename_ok(prefix) ){
ERROR_exit("Illegal value after -prefix!") ;
}
nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-thresh") == 0 ){
double val = (double)strtod(argv[++nopt],&cpt) ;
if( *cpt != '\0' || val >= 1.0 || val < 0.0 ){
ERROR_exit("Illegal value (%f) after -thresh!", val) ;
}
dothresh = 1;
thresh = val ; othresh = val ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-sparsity") == 0 ){
double val = (double)strtod(argv[++nopt],&cpt) ;
if( *cpt != '\0' || val > 100 || val <= 0 ){
ERROR_exit("Illegal value (%f) after -sparsity!", val) ;
}
if( val > 5.0 )
{
WARNING_message("Sparsity %3.2f%% is large and will require alot of memory and time, consider using a smaller value. ", val);
}
dosparsity = 1 ;
sparsity = val ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-polort") == 0 ){
int val = (int)strtod(argv[++nopt],&cpt) ;
if( *cpt != '\0' || val < -1 || val > 3 ){
ERROR_exit("Illegal value after -polort!") ;
}
polort = val ; nopt++ ; continue ;
}
if( strcmp(argv[nopt],"-mem_stat") == 0 ){
MEM_STAT = 1 ; nopt++ ; continue ;
}
if( strncmp(argv[nopt],"-mem_profile",8) == 0 ){
MEM_PROF = 1 ; nopt++ ; continue ;
}
/* check for 1d argument */
if ( strcmp(argv[nopt],"-out1D") == 0 ){
if (!(fout1D = fopen(argv[++nopt], "w"))) {
ERROR_message("Failed to open %s for writing", argv[nopt]);
exit(1);
}
nopt++ ; continue ;
}
ERROR_exit("Illegal option: %s",argv[nopt]) ;
}
/*-- open dataset, check for legality --*/
if( nopt >= argc ) ERROR_exit("Need a dataset on command line!?") ;
xset = THD_open_dataset(argv[nopt]); CHECK_OPEN_ERROR(xset,argv[nopt]);
if( DSET_NVALS(xset) < 3 )
ERROR_exit("Input dataset %s does not have 3 or more sub-bricks!",argv[nopt]) ;
DSET_load(xset) ; CHECK_LOAD_ERROR(xset) ;
/*-- compute mask array, if desired --*/
nvox = DSET_NVOX(xset) ; nvals = DSET_NVALS(xset) ;
INC_MEM_STATS((nvox * nvals * sizeof(double)), "input dset");
PRINT_MEM_STATS("inset");
/* if a mask was specified make sure it is appropriate */
if( mset ){
if( DSET_NVOX(mset) != nvox )
ERROR_exit("Input and mask dataset differ in number of voxels!") ;
mask = THD_makemask(mset, 0, 1.0, 0.0) ;
/* update running memory statistics to reflect loading the image */
INC_MEM_STATS( mset->dblk->total_bytes, "mask dset" );
PRINT_MEM_STATS( "mset load" );
nmask = THD_countmask( nvox , mask ) ;
INC_MEM_STATS( nmask * sizeof(byte), "mask array" );
PRINT_MEM_STATS( "mask" );
INFO_message("%d voxels in -mask dataset",nmask) ;
if( nmask < 2 ) ERROR_exit("Only %d voxels in -mask, exiting...",nmask);
/* update running memory statistics to reflect loading the image */
DEC_MEM_STATS( mset->dblk->total_bytes, "mask dset" );
DSET_unload(mset) ;
PRINT_MEM_STATS( "mset unload" );
}
/* if automasking is requested, handle that now */
else if( do_autoclip ){
mask = THD_automask( xset ) ;
nmask = THD_countmask( nvox , mask ) ;
INFO_message("%d voxels survive -autoclip",nmask) ;
if( nmask < 2 ) ERROR_exit("Only %d voxels in -automask!",nmask);
}
/* otherwise we use all of the voxels in the image */
else {
nmask = nvox ;
INFO_message("computing for all %d voxels",nmask) ;
}
if( method == ETA2 && polort >= 0 )
WARNING_message("Polort for -eta2 should probably be -1...");
/* djc - 1d file out init */
if (fout1D != NULL) {
/* define affine matrix */
mat44 affine_mat = xset->daxes->ijk_to_dicom;
/* print command line statement */
fprintf(fout1D,"#Similarity matrix from command:\n#");
for(ii=0; ii<argc; ++ii) fprintf(fout1D,"%s ", argv[ii]);
/* Print affine matrix */
fprintf(fout1D,"\n");
fprintf(fout1D,"#[ ");
int mi, mj;
for(mi = 0; mi < 4; mi++) {
for(mj = 0; mj < 4; mj++) {
fprintf(fout1D, "%.6f ", affine_mat.m[mi][mj]);
}
}
fprintf(fout1D, "]\n");
/* Print image extents*/
THD_dataxes *xset_daxes = xset->daxes;
fprintf(fout1D, "#Image dimensions:\n");
fprintf(fout1D, "#[%d, %d, %d]\n",
xset_daxes->nxx, xset_daxes->nyy, xset_daxes->nzz);
/* Similarity matrix headers */
fprintf(fout1D,"#Voxel1 Voxel2 i1 j1 k1 i2 j2 k2 Corr\n");
}
/* CC calculate the total number of possible correlations, will be
usefule down the road */
totPosCor = (.5*((float)nmask))*((float)(nmask-1));
/** For the case of Pearson correlation, we make sure the **/
/** data time series have their mean removed (polort >= 0) **/
/** and are normalized, so that correlation = dot product, **/
/** and we can use function zm_THD_pearson_corr for speed. **/
switch( method ){
default:
case PEARSON: corfun = zm_THD_pearson_corr ; break ;
case ETA2: corfun = my_THD_eta_squared ; break ;
}
/*-- create vectim from input dataset --*/
INFO_message("vectim-izing input dataset") ;
/*-- CC added in mask to reduce the size of xvectim -- */
xvectim = THD_dset_to_vectim( xset , mask , 0 ) ;
if( xvectim == NULL ) ERROR_exit("Can't create vectim?!") ;
/*-- CC update our memory stats to reflect vectim -- */
INC_MEM_STATS((xvectim->nvec*sizeof(int)) +
((xvectim->nvec)*(xvectim->nvals))*sizeof(float) +
sizeof(MRI_vectim), "vectim");
PRINT_MEM_STATS( "vectim" );
/*--- CC the vectim contains a mapping between voxel index and mask index,
tap into that here to avoid duplicating memory usage ---*/
if( mask != NULL )
{
imap = xvectim->ivec;
/* --- CC free the mask */
DEC_MEM_STATS( nmask*sizeof(byte), "mask array" );
free(mask); mask=NULL;
PRINT_MEM_STATS( "mask unload" );
}
/* -- CC unloading the dataset to reduce memory usage ?? -- */
DEC_MEM_STATS((DSET_NVOX(xset) * DSET_NVALS(xset) * sizeof(double)), "input dset");
DSET_unload(xset) ;
PRINT_MEM_STATS("inset unload");
/* -- CC configure detrending --*/
if( polort < 0 && method == PEARSON ){
polort = 0; WARNING_message("Pearson correlation always uses polort >= 0");
}
if( polort >= 0 ){
for( ii=0 ; ii < xvectim->nvec ; ii++ ){ /* remove polynomial trend */
DETREND_polort(polort,nvals,VECTIM_PTR(xvectim,ii)) ;
}
}
/* -- this procedure does not change time series that have zero variance -- */
if( method == PEARSON ) THD_vectim_normalize(xvectim) ; /* L2 norm = 1 */
/* -- CC create arrays to hold degree and weighted centrality while
they are being calculated -- */
if( dosparsity == 0 )
{
if( ( binaryDC = (long*)calloc( nmask, sizeof(long) )) == NULL )
{
ERROR_message( "Could not allocate %d byte array for binary DC calculation\n",
nmask*sizeof(long));
}
/* -- update running memory estimate to reflect memory allocation */
INC_MEM_STATS( nmask*sizeof(long), "binary DC array" );
PRINT_MEM_STATS( "binaryDC" );
if( ( weightedDC = (double*)calloc( nmask, sizeof(double) )) == NULL )
{
if (binaryDC){ free(binaryDC); binaryDC = NULL; }
ERROR_message( "Could not allocate %d byte array for weighted DC calculation\n",
nmask*sizeof(double));
}
/* -- update running memory estimate to reflect memory allocation */
INC_MEM_STATS( nmask*sizeof(double), "weighted DC array" );
PRINT_MEM_STATS( "weightedDC" );
}
/* -- CC if we are using a sparsity threshold, build a histogram to calculate the
threshold */
if (dosparsity == 1)
{
/* make sure that there is a bin for correlation values that == 1.0 */
binwidth = (1.005-thresh)/nhistnodes;
/* calculate the number of correlations we wish to retain */
ngoal = nretain = (int)(((double)totPosCor)*((double)sparsity) / 100.0);
/* allocate memory for the histogram bins */
if(( histogram = (hist_node_head*)malloc(nhistnodes*sizeof(hist_node_head))) == NULL )
{
/* if the allocation fails, free all memory and exit */
if (binaryDC){ free(binaryDC); binaryDC = NULL; }
if (weightedDC){ free(weightedDC); weightedDC = NULL; }
ERROR_message( "Could not allocate %d byte array for histogram\n",
nhistnodes*sizeof(hist_node_head));
}
else {
/* -- update running memory estimate to reflect memory allocation */
INC_MEM_STATS( nhistnodes*sizeof(hist_node_head), "hist bins" );
PRINT_MEM_STATS( "hist1" );
}
/* initialize history bins */
for( kout = 0; kout < nhistnodes; kout++ )
{
histogram[ kout ].bin_low = thresh+kout*binwidth;
histogram[ kout ].bin_high = histogram[ kout ].bin_low+binwidth;
histogram[ kout ].nbin = 0;
histogram[ kout ].nodes = NULL;
/*INFO_message("Hist bin %d [%3.3f, %3.3f) [%d, %p]\n",
kout, histogram[ kout ].bin_low, histogram[ kout ].bin_high,
histogram[ kout ].nbin, histogram[ kout ].nodes );*/
}
}
/*-- tell the user what we are about to do --*/
if (dosparsity == 0 )
{
INFO_message( "Calculating degree centrality with threshold = %f.\n", thresh);
}
else
{
INFO_message( "Calculating degree centrality with threshold = %f and sparsity = %3.2f%% (%d)\n",
thresh, sparsity, nretain);
}
/*---------- loop over mask voxels, correlate ----------*/
AFNI_OMP_START ;
#pragma omp parallel if( nmask > 999 )
{
int lii,ljj,lin,lout,ithr,nthr,vstep,vii ;
float *xsar , *ysar ;
hist_node* new_node = NULL ;
hist_node* tptr = NULL ;
hist_node* rptr = NULL ;
int new_node_idx = 0;
double car = 0.0 ;
/*-- get information about who we are --*/
#ifdef USE_OMP
ithr = omp_get_thread_num() ;
nthr = omp_get_num_threads() ;
if( ithr == 0 ) INFO_message("%d OpenMP threads started",nthr) ;
#else
ithr = 0 ; nthr = 1 ;
#endif
/*-- For the progress tracker, we want to print out 50 numbers,
figure out a number of loop iterations that will make this easy */
vstep = (int)( nmask / (nthr*50.0f) + 0.901f ) ; vii = 0 ;
if((MEM_STAT==0) && (ithr == 0 )) fprintf(stderr,"Looping:") ;
#pragma omp for schedule(static, 1)
for( lout=0 ; lout < xvectim->nvec ; lout++ ){ /*----- outer voxel loop -----*/
if( ithr == 0 && vstep > 2 ) /* allow small dsets 16 Jun 2011 [rickr] */
{ vii++ ; if( vii%vstep == vstep/2 && MEM_STAT == 0 ) vstep_print(); }
/* get ref time series from this voxel */
xsar = VECTIM_PTR(xvectim,lout) ;
/* try to make calculation more efficient by only calculating the unique
correlations */
for( lin=(lout+1) ; lin < xvectim->nvec ; lin++ ){ /*----- inner loop over voxels -----*/
/* extract the voxel time series */
ysar = VECTIM_PTR(xvectim,lin) ;
/* now correlate the time series */
car = (double)(corfun(nvals,xsar,ysar)) ;
if ( car <= thresh )
{
continue ;
}
/* update degree centrality values, hopefully the pragma
will handle mutual exclusion */
#pragma omp critical(dataupdate)
{
/* if the correlation is less than threshold, ignore it */
if ( car > thresh )
{
totNumCor += 1;
if ( dosparsity == 0 )
{
binaryDC[lout] += 1; binaryDC[lin] += 1;
weightedDC[lout] += car; weightedDC[lin] += car;
/* print correlation out to the 1D file */
if ( fout1D != NULL )
{
/* determine the i,j,k coords */
ix1 = DSET_index_to_ix(xset,lii) ;
jy1 = DSET_index_to_jy(xset,lii) ;
kz1 = DSET_index_to_kz(xset,lii) ;
ix2 = DSET_index_to_ix(xset,ljj) ;
jy2 = DSET_index_to_jy(xset,ljj) ;
kz2 = DSET_index_to_kz(xset,ljj) ;
/* add source, dest, correlation to 1D file */
fprintf(fout1D, "%d %d %d %d %d %d %d %d %.6f\n",
lii, ljj, ix1, jy1, kz1, ix2, jy2, kz2, car);
}
}
else
{
/* determine the index in the histogram to add the node */
new_node_idx = (int)floor((double)(car-othresh)/(double)binwidth);
if ((new_node_idx > nhistnodes) || (new_node_idx < bottom_node_idx))
{
/* this error should indicate a programming error and should not happen */
WARNING_message("Node index %d is out of range [%d,%d)!",new_node_idx,
bottom_node_idx, nhistnodes);
}
else
{
/* create a node to add to the histogram */
new_node = (hist_node*)calloc(1,sizeof(hist_node));
if( new_node == NULL )
{
/* allocate memory for this node, rather than fiddling with
error handling here, lets just move on */
WARNING_message("Could not allocate a new node!");
}
else
{
/* populate histogram node */
new_node->i = lout;
new_node->j = lin;
new_node->corr = car;
new_node->next = NULL;
/* -- update running memory estimate to reflect memory allocation */
INC_MEM_STATS( sizeof(hist_node), "hist nodes" );
if ((totNumCor % (1024*1024)) == 0) PRINT_MEM_STATS( "hist nodes" );
/* populate histogram */
new_node->next = histogram[new_node_idx].nodes;
histogram[new_node_idx].nodes = new_node;
histogram[new_node_idx].nbin++;
/* see if there are enough correlations in the histogram
for the sparsity */
if ((totNumCor - histogram[bottom_node_idx].nbin) > nretain)
{
/* delete the list of nodes */
rptr = histogram[bottom_node_idx].nodes;
while(rptr != NULL)
{
tptr = rptr;
rptr = rptr->next;
/* check that the ptr is not null before freeing it*/
if(tptr!= NULL)
{
DEC_MEM_STATS( sizeof(hist_node), "hist nodes" );
free(tptr);
}
}
PRINT_MEM_STATS( "unloaded hist nodes - thresh increase" );
histogram[bottom_node_idx].nodes = NULL;
totNumCor -= histogram[bottom_node_idx].nbin;
histogram[bottom_node_idx].nbin=0;
/* get the new threshold */
thresh = (double)histogram[++bottom_node_idx].bin_low;
if(MEM_STAT == 1) INFO_message("Increasing threshold to %3.2f (%d)\n",
thresh,bottom_node_idx);
}
} /* else, newptr != NULL */
} /* else, new_node_idx in range */
} /* else, do_sparsity == 1 */
} /* car > thresh */
} /* this is the end of the critical section */
} /* end of inner loop over voxels */
} /* end of outer loop over ref voxels */
if( ithr == 0 ) fprintf(stderr,".\n") ;
} /* end OpenMP */
AFNI_OMP_END ;
/* update the user so that they know what we are up to */
INFO_message ("AFNI_OMP finished\n");
INFO_message ("Found %d (%3.2f%%) correlations above threshold (%f)\n",
totNumCor, 100.0*((float)totNumCor)/((float)totPosCor), thresh);
/*---------- Finish up ---------*/
/*if( dosparsity == 1 )
{
for( kout = 0; kout < nhistnodes; kout++ )
{
INFO_message("Hist bin %d [%3.3f, %3.3f) [%d, %p]\n",
kout, histogram[ kout ].bin_low, histogram[ kout ].bin_high,
histogram[ kout ].nbin, histogram[ kout ].nodes );
}
}*/
/*-- create output dataset --*/
cset = EDIT_empty_copy( xset ) ;
/*-- configure the output dataset */
if( abuc ){
EDIT_dset_items( cset ,
ADN_prefix , prefix ,
ADN_nvals , nsubbriks , /* 2 subbricks, degree and weighted centrality */
ADN_ntt , 0 , /* no time axis */
ADN_type , HEAD_ANAT_TYPE ,
ADN_func_type , ANAT_BUCK_TYPE ,
ADN_datum_all , MRI_float ,
ADN_none ) ;
} else {
EDIT_dset_items( cset ,
ADN_prefix , prefix ,
ADN_nvals , nsubbriks , /* 2 subbricks, degree and weighted centrality */
ADN_ntt , nsubbriks , /* num times */
ADN_ttdel , 1.0 , /* fake TR */
ADN_nsl , 0 , /* no slice offsets */
ADN_type , HEAD_ANAT_TYPE ,
ADN_func_type , ANAT_EPI_TYPE ,
ADN_datum_all , MRI_float ,
ADN_none ) ;
}
/* add history information to the hearder */
tross_Make_History( "3dDegreeCentrality" , argc,argv , cset ) ;
ININFO_message("creating output dataset in memory") ;
/* -- Configure the subbriks: Binary Degree Centrality */
subbrik = 0;
EDIT_BRICK_TO_NOSTAT(cset,subbrik) ; /* stat params */
/* CC this sets the subbrik scaling factor, which we will probably want
to do again after we calculate the voxel values */
EDIT_BRICK_FACTOR(cset,subbrik,1.0) ; /* scale factor */
sprintf(str,"Binary Degree Centrality") ;
EDIT_BRICK_LABEL(cset,subbrik,str) ;
EDIT_substitute_brick(cset,subbrik,MRI_float,NULL) ; /* make array */
/* copy measure data into the subbrik */
bodset = DSET_ARRAY(cset,subbrik);
/* -- Configure the subbriks: Weighted Degree Centrality */
subbrik = 1;
EDIT_BRICK_TO_NOSTAT(cset,subbrik) ; /* stat params */
/* CC this sets the subbrik scaling factor, which we will probably want
to do again after we calculate the voxel values */
EDIT_BRICK_FACTOR(cset,subbrik,1.0) ; /* scale factor */
sprintf(str,"Weighted Degree Centrality") ;
EDIT_BRICK_LABEL(cset,subbrik,str) ;
EDIT_substitute_brick(cset,subbrik,MRI_float,NULL) ; /* make array */
/* copy measure data into the subbrik */
wodset = DSET_ARRAY(cset,subbrik);
/* increment memory stats */
INC_MEM_STATS( (DSET_NVOX(cset)*DSET_NVALS(cset)*sizeof(float)), "output dset");
PRINT_MEM_STATS( "outset" );
/* pull the values out of the histogram */
if( dosparsity == 0 )
{
for( kout = 0; kout < nmask; kout++ )
{
if ( imap != NULL )
{
ii = imap[kout] ; /* ii= source voxel (we know that ii is in the mask) */
}
else
{
ii = kout ;
}
if( ii >= DSET_NVOX(cset) )
{
WARNING_message("Avoiding bodset, wodset overflow %d > %d (%s,%d)\n",
ii,DSET_NVOX(cset),__FILE__,__LINE__ );
}
else
{
bodset[ ii ] = (float)(binaryDC[kout]);
wodset[ ii ] = (float)(weightedDC[kout]);
}
}
/* we are done with this memory, and can kill it now*/
if(binaryDC)
{
free(binaryDC);
binaryDC=NULL;
/* -- update running memory estimate to reflect memory allocation */
DEC_MEM_STATS( nmask*sizeof(long), "binary DC array" );
PRINT_MEM_STATS( "binaryDC" );
}
if(weightedDC)
{
free(weightedDC);
weightedDC=NULL;
/* -- update running memory estimate to reflect memory allocation */
DEC_MEM_STATS( nmask*sizeof(double), "weighted DC array" );
PRINT_MEM_STATS( "weightedDC" );
}
}
else
{
/* add in the values from the histogram, this is a two stage procedure:
at first we add in values a whole bin at the time until we get to a point
where we need to add in a partial bin, then we create a new histogram
to sort the values in the bin and then add those bins at a time */
kout = nhistnodes - 1;
while (( histogram[kout].nbin < nretain ) && ( kout >= 0 ))
{
hptr = pptr = histogram[kout].nodes;
while( hptr != NULL )
{
/* determine the indices corresponding to this node */
if ( imap != NULL )
{
ii = imap[hptr->i] ; /* ii= source voxel (we know that ii is in the mask) */
}
else
{
ii = hptr->i ;
}
if ( imap != NULL )
{
jj = imap[hptr->j] ; /* ii= source voxel (we know that ii is in the mask) */
}
else
{
jj = hptr->j ;
}
/* add in the values */
if(( ii >= DSET_NVOX(cset) ) || ( jj >= DSET_NVOX(cset)))
{
if( ii >= DSET_NVOX(cset))
{
WARNING_message("Avoiding bodset, wodset overflow (ii) %d > %d\n (%s,%d)\n",
ii,DSET_NVOX(cset),__FILE__,__LINE__ );
}
if( jj >= DSET_NVOX(cset))
{
WARNING_message("Avoiding bodset, wodset overflow (jj) %d > %d\n (%s,%d)\n",
jj,DSET_NVOX(cset),__FILE__,__LINE__ );
}
}
else
{
bodset[ ii ] += 1.0 ;
wodset[ ii ] += (float)(hptr->corr);
bodset[ jj ] += 1.0 ;
wodset[ jj ] += (float)(hptr->corr);
}
if( fout1D != NULL )
{
/* add source, dest, correlation to 1D file */
ix1 = DSET_index_to_ix(cset,ii) ;
jy1 = DSET_index_to_jy(cset,ii) ;
kz1 = DSET_index_to_kz(cset,ii) ;
ix2 = DSET_index_to_ix(cset,jj) ;
jy2 = DSET_index_to_jy(cset,jj) ;
kz2 = DSET_index_to_kz(cset,jj) ;
fprintf(fout1D, "%d %d %d %d %d %d %d %d %.6f\n",
ii, jj, ix1, jy1, kz1, ix2, jy2, kz2, (float)(hptr->corr));
}
/* increment node pointers */
pptr = hptr;
hptr = hptr->next;
/* delete the node */
if(pptr)
{
/* -- update running memory estimate to reflect memory allocation */
DEC_MEM_STATS(sizeof( hist_node ), "hist nodes" );
/* free the mem */
free(pptr);
pptr=NULL;
}
}
/* decrement the number of correlations we wish to retain */
nretain -= histogram[kout].nbin;
histogram[kout].nodes = NULL;
/* go on to the next bin */
kout--;
}
PRINT_MEM_STATS( "hist1 bins free - inc into output" );
/* if we haven't used all of the correlations that are available, go through and
add a subset of the voxels from the remaining bin */
if(( nretain > 0 ) && (kout >= 0))
{
hist_node_head* histogram2 = NULL;
hist_node_head* histogram2_save = NULL;
int h2nbins = 100;
float h2binwidth = 0.0;
int h2ndx=0;
h2binwidth = (((1.0+binwidth/((float)h2nbins))*histogram[kout].bin_high) - histogram[kout].bin_low) /
((float)h2nbins);
/* allocate the bins */
if(( histogram2 = (hist_node_head*)malloc(h2nbins*sizeof(hist_node_head))) == NULL )
{
if (binaryDC){ free(binaryDC); binaryDC = NULL; }
if (weightedDC){ free(weightedDC); weightedDC = NULL; }
if (histogram){ histogram = free_histogram(histogram, nhistnodes); }
ERROR_message( "Could not allocate %d byte array for histogram2\n",
h2nbins*sizeof(hist_node_head));
}
else {
/* -- update running memory estimate to reflect memory allocation */
histogram2_save = histogram2;
INC_MEM_STATS(( h2nbins*sizeof(hist_node_head )), "hist bins");
PRINT_MEM_STATS( "hist2" );
}
/* initiatize the bins */
for( kin = 0; kin < h2nbins; kin++ )
{
histogram2[ kin ].bin_low = histogram[kout].bin_low + kin*h2binwidth;
histogram2[ kin ].bin_high = histogram2[ kin ].bin_low + h2binwidth;
histogram2[ kin ].nbin = 0;
histogram2[ kin ].nodes = NULL;
/*INFO_message("Hist2 bin %d [%3.3f, %3.3f) [%d, %p]\n",
kin, histogram2[ kin ].bin_low, histogram2[ kin ].bin_high,
histogram2[ kin ].nbin, histogram2[ kin ].nodes );*/
}
/* move correlations from histogram to histgram2 */
INFO_message ("Adding %d nodes from histogram to histogram2",histogram[kout].nbin);
while ( histogram[kout].nodes != NULL )
{
hptr = histogram[kout].nodes;
h2ndx = (int)floor((double)(hptr->corr - histogram[kout].bin_low)/(double)h2binwidth);
if(( h2ndx < h2nbins ) && ( h2ndx >= 0 ))
{
histogram[kout].nodes = hptr->next;
hptr->next = histogram2[h2ndx].nodes;
histogram2[h2ndx].nodes = hptr;
histogram2[h2ndx].nbin++;
histogram[kout].nbin--;
}
else
{
WARNING_message("h2ndx %d is not in range [0,%d) :: %.10f,%.10f\n",h2ndx,h2nbins,hptr->corr, histogram[kout].bin_low);
}
}
/* free the remainder of histogram */
{
int nbins_rem = 0;
for(ii = 0; ii < nhistnodes; ii++) nbins_rem+=histogram[ii].nbin;
histogram = free_histogram(histogram, nhistnodes);
PRINT_MEM_STATS( "free remainder of histogram1" );
}
kin = h2nbins - 1;
while (( nretain > 0 ) && ( kin >= 0 ))
{
hptr = pptr = histogram2[kin].nodes;
while( hptr != NULL )
{
/* determine the indices corresponding to this node */
if ( imap != NULL )
{
ii = imap[hptr->i] ;
}
else
{
ii = hptr->i ;
}
if ( imap != NULL )
{
jj = imap[hptr->j] ;
}
else
{
jj = hptr->j ;
}
/* add in the values */
if(( ii >= DSET_NVOX(cset) ) || ( jj >= DSET_NVOX(cset)))
{
if( ii >= DSET_NVOX(cset))
{
WARNING_message("Avoiding bodset, wodset overflow (ii) %d > %d\n (%s,%d)\n",
ii,DSET_NVOX(cset),__FILE__,__LINE__ );
}
if( jj >= DSET_NVOX(cset))
{
WARNING_message("Avoiding bodset, wodset overflow (jj) %d > %d\n (%s,%d)\n",
jj,DSET_NVOX(cset),__FILE__,__LINE__ );
}
}
else
{
bodset[ ii ] += 1.0 ;
wodset[ ii ] += (float)(hptr->corr);
bodset[ jj ] += 1.0 ;
wodset[ jj ] += (float)(hptr->corr);
}
if( fout1D != NULL )
{
/* add source, dest, correlation to 1D file */
ix1 = DSET_index_to_ix(cset,ii) ;
jy1 = DSET_index_to_jy(cset,ii) ;
kz1 = DSET_index_to_kz(cset,ii) ;
ix2 = DSET_index_to_ix(cset,jj) ;
jy2 = DSET_index_to_jy(cset,jj) ;
kz2 = DSET_index_to_kz(cset,jj) ;
fprintf(fout1D, "%d %d %d %d %d %d %d %d %.6f\n",
ii, jj, ix1, jy1, kz1, ix2, jy2, kz2, (float)(hptr->corr));
}
/* increment node pointers */
pptr = hptr;
hptr = hptr->next;
/* delete the node */
if(pptr)
{
free(pptr);
DEC_MEM_STATS(( sizeof(hist_node) ), "hist nodes");
pptr=NULL;
}
}
/* decrement the number of correlations we wish to retain */
nretain -= histogram2[kin].nbin;
histogram2[kin].nodes = NULL;
/* go on to the next bin */
kin--;
}
PRINT_MEM_STATS("hist2 nodes free - incorporated into output");
/* we are finished with histogram2 */
{
histogram2 = free_histogram(histogram2, h2nbins);
/* -- update running memory estimate to reflect memory allocation */
PRINT_MEM_STATS( "free hist2" );
}
if (nretain < 0 )
{
WARNING_message( "Went over sparsity goal %d by %d, with a resolution of %f",
ngoal, -1*nretain, h2binwidth);
}
}
if (nretain > 0 )
{
WARNING_message( "Was not able to meet goal of %d (%3.2f%%) correlations, %d (%3.2f%%) correlations passed the threshold of %3.2f, maybe you need to change the threshold or the desired sparsity?",
ngoal, 100.0*((float)ngoal)/((float)totPosCor), totNumCor, 100.0*((float)totNumCor)/((float)totPosCor), thresh);
}
}
INFO_message("Done..\n") ;
/* update running memory statistics to reflect freeing the vectim */
DEC_MEM_STATS(((xvectim->nvec*sizeof(int)) +
((xvectim->nvec)*(xvectim->nvals))*sizeof(float) +
sizeof(MRI_vectim)), "vectim");
/* toss some trash */
VECTIM_destroy(xvectim) ;
DSET_delete(xset) ;
if(fout1D!=NULL)fclose(fout1D);
PRINT_MEM_STATS( "vectim unload" );
if (weightedDC) free(weightedDC) ; weightedDC = NULL;
if (binaryDC) free(binaryDC) ; binaryDC = NULL;
/* finito */
INFO_message("Writing output dataset to disk [%s bytes]",
commaized_integer_string(cset->dblk->total_bytes)) ;
/* write the dataset */
DSET_write(cset) ;
WROTE_DSET(cset) ;
/* increment our memory stats, since we are relying on the header for this
information, we update the stats before actually freeing the memory */
DEC_MEM_STATS( (DSET_NVOX(cset)*DSET_NVALS(cset)*sizeof(float)), "output dset");
/* free up the output dataset memory */
DSET_unload(cset) ;
DSET_delete(cset) ;
/* force a print */
MEM_STAT = 1;
PRINT_MEM_STATS( "Fin" );
exit(0) ;
}
THD_3dim_dataset * THD_localhistog( int nsar , THD_3dim_dataset **insar ,
int numval , int *rlist , MCW_cluster *nbhd ,
int do_prob , int verb )
{
THD_3dim_dataset *outset=NULL , *inset ;
int nvox=DSET_NVOX(insar[0]) ;
int ids, iv, bb, nnpt=nbhd->num_pt ;
MRI_IMAGE *bbim ; int btyp ;
float **outar , **listar ;
ENTRY("THD_localhistog") ;
/*---- create output dataset ----*/
outset = EDIT_empty_copy(insar[0]) ;
EDIT_dset_items( outset ,
ADN_nvals , numval ,
ADN_datum_all , MRI_float ,
ADN_nsl , 0 ,
ADN_brick_fac , NULL ,
ADN_none ) ;
outar = (float **)malloc(sizeof(float *)*numval) ;
for( bb=0 ; bb < numval ; bb++ ){
EDIT_substitute_brick( outset , bb , MRI_float , NULL ) ;
outar[bb] = DSET_BRICK_ARRAY(outset,bb) ;
}
/*---- make mapping between values and arrays to get those values ----*/
listar = (float **)malloc(sizeof(float *)*TWO16) ;
for( bb=0 ; bb < TWO16 ; bb++ ) listar[bb] = outar[0] ;
for( bb=1 ; bb < numval ; bb++ ){
listar[ rlist[bb] + TWO15 ] = outar[bb] ;
}
/*----------- loop over datasets, add in counts for all voxels -----------*/
for( ids=0 ; ids < nsar ; ids++ ){ /* dataset loop */
inset = insar[ids] ; DSET_load(inset) ;
for( iv=0 ; iv < DSET_NVALS(inset) ; iv++ ){ /* sub-brick loop */
if( verb ) fprintf(stderr,".") ;
bbim = DSET_BRICK(inset,iv) ; btyp = bbim->kind ;
if( nnpt == 1 ){ /* only 1 voxel in nbhd */
int qq,ii,jj,kk,ib,nb ;
switch( bbim->kind ){
case MRI_short:{
short *sar = MRI_SHORT_PTR(bbim) ;
for( qq=0 ; qq < nvox ; qq++ ) listar[sar[qq]+TWO15][qq]++ ;
}
break ;
case MRI_byte:{
byte *bar = MRI_BYTE_PTR(bbim) ;
for( qq=0 ; qq < nvox ; qq++ ) listar[bar[qq]+TWO15][qq]++ ;
}
break ;
case MRI_float:{
float *far = MRI_FLOAT_PTR(bbim) ; short ss ;
for( qq=0 ; qq < nvox ; qq++ ){ ss = SHORTIZE(far[qq]); listar[ss+TWO15][qq]++; }
}
break ;
}
} else { /* multiple voxels in nbhd */
AFNI_OMP_START ;
#pragma omp parallel
{ int qq,ii,jj,kk,ib,nb ; void *nar ; short *sar,ss ; byte *bar ; float *far ;
nar = malloc(sizeof(float)*nnpt) ;
sar = (short *)nar ; bar = (byte *)nar ; far = (float *)nar ;
#pragma omp for
for( qq=0 ; qq < nvox ; qq++ ){ /* qq=voxel index */
ii = DSET_index_to_ix(inset,qq) ;
jj = DSET_index_to_jy(inset,qq) ;
kk = DSET_index_to_kz(inset,qq) ;
nb = mri_get_nbhd_array( bbim , NULL , ii,jj,kk , nbhd , nar ) ;
if( nb == 0 ) continue ;
switch( btyp ){
case MRI_short:
for( ib=0 ; ib < nb ; ib++ ) listar[sar[ib]+TWO15][qq]++ ;
break ;
case MRI_byte:
for( ib=0 ; ib < nb ; ib++ ) listar[bar[ib]+TWO15][qq]++ ;
break ;
case MRI_float:
for( ib=0 ; ib < nb ; ib++ ){ ss = SHORTIZE(far[ib]); listar[ss+TWO15][qq]++; }
break ;
}
} /* end of voxel loop */
free(nar) ;
} /* end of OpenMP */
AFNI_OMP_END ;
}
} /* end of sub-brick loop */
DSET_unload(inset) ;
} /* end of dataset loop */
if( verb ) fprintf(stderr,"\n") ;
free(listar) ;
/*---- post-process output ---*/
if( do_prob ){
byte **bbar ; int pp ;
if( verb ) INFO_message("Conversion to probabilities") ;
AFNI_OMP_START ;
#pragma omp parallel
{ int qq , ib ; float pfac , val ; byte **bbar ;
#pragma omp for
for( qq=0 ; qq < nvox ; qq++ ){
pfac = 0.0001f ;
for( ib=0 ; ib < numval ; ib++ ) pfac += outar[ib][qq] ;
pfac = 250.0f / pfac ;
for( ib=0 ; ib < numval ; ib++ ){
val = outar[ib][qq]*pfac ; outar[ib][qq] = BYTEIZE(val) ;
}
}
} /* end OpenMP */
AFNI_OMP_END ;
bbar = (byte **)malloc(sizeof(byte *)*numval) ;
for( bb=0 ; bb < numval ; bb++ ){
bbar[bb] = (byte *)malloc(sizeof(byte)*nvox) ;
for( pp=0 ; pp < nvox ; pp++ ) bbar[bb][pp] = (byte)outar[bb][pp] ;
EDIT_substitute_brick(outset,bb,MRI_byte,bbar[bb]) ;
EDIT_BRICK_FACTOR(outset,bb,0.004f) ;
}
free(bbar) ;
} /* end of do_prob */
free(outar) ;
RETURN(outset) ;
}
int main( int argc , char *argv[] )
{
THD_3dim_dataset *dset , *oset=NULL , *tset=NULL ;
int nvals , iv , nxyz , ii,jj,kk , iarg , kz,kzold ;
float cut1=2.5,cut2=4.0 , sq2p,sfac , fq ;
MRI_IMAGE *flim ;
char *prefix="despike" , *tprefix=NULL ;
int corder=-1 , nref , ignore=0 , polort=2 , nuse , nomask=0 ;
int nspike, nbig, nproc ;
float **ref ;
float c21,ic21 , pspike,pbig ;
short *sar , *qar ;
byte *tar , *mask=NULL ;
float *zar , *yar ;
int datum ;
int localedit=0 ; /* 04 Apr 2007 */
int verb=1 ;
int do_NEW = 0 ; /* 29 Nov 2013 */
MRI_IMAGE *NEW_psinv=NULL ;
int dilate = 4 ; /* 04 Dec 2013 */
int ctim = 0 ;
/*----- Read command line -----*/
AFNI_SETUP_OMP(0) ; /* 24 Jun 2013 */
if( argc < 2 || strcmp(argv[1],"-help") == 0 ){
printf("Usage: 3dDespike [options] dataset\n"
"Removes 'spikes' from the 3D+time input dataset and writes\n"
"a new dataset with the spike values replaced by something\n"
"more pleasing to the eye.\n"
"\n"
"Method:\n"
" * L1 fit a smooth-ish curve to each voxel time series\n"
" [see -corder option for description of the curve]\n"
" [see -NEW option for a different & faster fitting method]\n"
" * Compute the MAD of the difference between the curve and\n"
" the data time series (the residuals).\n"
" * Estimate the standard deviation 'sigma' of the residuals\n"
" as sqrt(PI/2)*MAD.\n"
" * For each voxel value, define s = (value-curve)/sigma.\n"
" * Values with s > c1 are replaced with a value that yields\n"
" a modified s' = c1+(c2-c1)*tanh((s-c1)/(c2-c1)).\n"
" * c1 is the threshold value of s for a 'spike' [default c1=2.5].\n"
" * c2 is the upper range of the allowed deviation from the curve:\n"
" s=[c1..infinity) is mapped to s'=[c1..c2) [default c2=4].\n"
"\n"
"Options:\n"
" -ignore I = Ignore the first I points in the time series:\n"
" these values will just be copied to the\n"
" output dataset [default I=0].\n"
" -corder L = Set the curve fit order to L:\n"
" the curve that is fit to voxel data v(t) is\n"
"\n"
" k=L [ (2*PI*k*t) (2*PI*k*t) ]\n"
" f(t) = a+b*t+c*t*t + SUM [ d * sin(--------) + e * cos(--------) ]\n"
" k=1 [ k ( T ) k ( T ) ]\n"
"\n"
" where T = duration of time series;\n"
" the a,b,c,d,e parameters are chosen to minimize\n"
" the sum over t of |v(t)-f(t)| (L1 regression);\n"
" this type of fitting is is insensitive to large\n"
" spikes in the data. The default value of L is\n"
" NT/30, where NT = number of time points.\n"
"\n"
" -cut c1 c2 = Alter default values for the spike cut values\n"
" [default c1=2.5, c2=4.0].\n"
" -prefix pp = Save de-spiked dataset with prefix 'pp'\n"
" [default pp='despike']\n"
" -ssave ttt = Save 'spikiness' measure s for each voxel into a\n"
" 3D+time dataset with prefix 'ttt' [default=no save]\n"
" -nomask = Process all voxels\n"
" [default=use a mask of high-intensity voxels, ]\n"
" [as created via '3dAutomask -dilate 4 dataset'].\n"
" -dilate nd = Dilate 'nd' times (as in 3dAutomask). The default\n"
" value of 'nd' is 4.\n"
" -q[uiet] = Don't print '++' informational messages.\n"
"\n"
" -localedit = Change the editing process to the following:\n"
" If a voxel |s| value is >= c2, then replace\n"
" the voxel value with the average of the two\n"
" nearest non-spike (|s| < c2) values; the first\n"
" one previous and the first one after.\n"
" Note that the c1 cut value is not used here.\n"
"\n"
" -NEW = Use the 'new' method for computing the fit, which\n"
" should be faster than the L1 method for long time\n"
" series (200+ time points); however, the results\n"
" are similar but NOT identical. [29 Nov 2013]\n"
" * You can also make the program use the 'new'\n"
" method by setting the environment variable\n"
" AFNI_3dDespike_NEW\n"
" to the value YES; as in\n"
" setenv AFNI_3dDespike_NEW YES (csh)\n"
" export AFNI_3dDespike_NEW=YES (bash)\n"
" * If this variable is set to YES, you can turn off\n"
" the '-NEW' processing by using the '-OLD' option.\n"
" -->>* For time series more than 500 points long, the\n"
" '-OLD' algorithm is tremendously slow. You should\n"
" use the '-NEW' algorith in such cases.\n"
" ** At some indeterminate point in the future, the '-NEW'\n"
" method will become the default!\n"
" -->>* As of 29 Sep 2016, '-NEW' is the default if there\n"
" is more than 500 points in the time series dataset.\n"
"\n"
" -NEW25 = A slightly more aggressive despiking approach than\n"
" the '-NEW' method.\n"
"\n"
"Caveats:\n"
"* Despiking may interfere with image registration, since head\n"
" movement may produce 'spikes' at the edge of the brain, and\n"
" this information would be used in the registration process.\n"
" This possibility has not been explored or calibrated.\n"
"* [LATER] Actually, it seems like the registration problem\n"
" does NOT happen, and in fact, despiking seems to help!\n"
"* Check your data visually before and after despiking and\n"
" registration!\n"
" [Hint: open 2 AFNI controllers, and turn Time Lock on.]\n"
) ;
PRINT_AFNI_OMP_USAGE("3dDespike",NULL) ;
PRINT_COMPILE_DATE ; exit(0) ;
}
/** AFNI package setup and logging **/
mainENTRY("3dDespike main"); machdep(); AFNI_logger("3dDespike",argc,argv);
PRINT_VERSION("3dDespike") ; AUTHOR("RW Cox") ;
/** parse options **/
if( AFNI_yesenv("AFNI_3dDespike_NEW") ) do_NEW = 1 ; /* 29 Nov 2013 */
iarg = 1 ;
while( iarg < argc && argv[iarg][0] == '-' ){
if( strncmp(argv[iarg],"-q",2) == 0 ){ /* 04 Apr 2007 */
verb = 0 ; iarg++ ; continue ;
}
if( strncmp(argv[iarg],"-v",2) == 0 ){
verb++ ; iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-NEW") == 0 ){ /* 29 Nov 2013 */
do_NEW = 1 ; iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-NEW25") == 0 ){ /* 29 Sep 2016 */
do_NEW = 1 ; use_des25 = 1 ; cut1 = 2.5f ; cut2 = 3.2f ; iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-OLD") == 0 ){
do_NEW = 0 ; iarg++ ; continue ;
}
/** -localedit **/
if( strcmp(argv[iarg],"-localedit") == 0 ){ /* 04 Apr 2007 */
localedit = 1 ; iarg++ ; continue ;
}
/** don't use masking **/
if( strcmp(argv[iarg],"-nomask") == 0 ){
nomask = 1 ; iarg++ ; continue ;
}
/** dilation count [04 Dec 2013] **/
if( strcmp(argv[iarg],"-dilate") == 0 ){
dilate = (int)strtod(argv[++iarg],NULL) ;
if( dilate <= 0 ) dilate = 1 ;
else if( dilate > 99 ) dilate = 99 ;
iarg++ ; continue ;
}
/** output dataset prefix **/
if( strcmp(argv[iarg],"-prefix") == 0 ){
prefix = argv[++iarg] ;
if( !THD_filename_ok(prefix) ) ERROR_exit("-prefix is not good");
iarg++ ; continue ;
}
/** ratio dataset prefix **/
if( strcmp(argv[iarg],"-ssave") == 0 ){
tprefix = argv[++iarg] ;
if( !THD_filename_ok(tprefix) ) ERROR_exit("-ssave prefix is not good");
iarg++ ; continue ;
}
/** trigonometric polynomial order **/
if( strcmp(argv[iarg],"-corder") == 0 ){
corder = strtol( argv[++iarg] , NULL , 10 ) ;
if( corder < 0 ) ERROR_exit("Illegal value of -corder");
iarg++ ; continue ;
}
/** how much to ignore at start **/
if( strcmp(argv[iarg],"-ignore") == 0 ){
ignore = strtol( argv[++iarg] , NULL , 10 ) ;
if( ignore < 0 ) ERROR_exit("Illegal value of -ignore");
iarg++ ; continue ;
}
/** thresholds for s ratio **/
if( strcmp(argv[iarg],"-cut") == 0 ){
cut1 = strtod( argv[++iarg] , NULL ) ;
cut2 = strtod( argv[++iarg] , NULL ) ;
if( cut1 < 1.0 || cut2 < cut1+0.5 )
ERROR_exit("Illegal values after -cut");
iarg++ ; continue ;
}
ERROR_exit("Unknown option: %s",argv[iarg]) ;
}
c21 = cut2-cut1 ; ic21 = 1.0/c21 ;
/*----- read input dataset -----*/
if( iarg >= argc ) ERROR_exit("No input dataset!!??");
dset = THD_open_dataset( argv[iarg] ) ;
CHECK_OPEN_ERROR(dset,argv[iarg]) ;
datum = DSET_BRICK_TYPE(dset,0) ;
if( (datum != MRI_short && datum != MRI_float) || !DSET_datum_constant(dset) )
ERROR_exit("Can't process non-short, non-float dataset!") ;
if( verb ) INFO_message("Input data type = %s\n",MRI_TYPE_name[datum]) ;
nvals = DSET_NUM_TIMES(dset) ; nuse = nvals - ignore ;
if( nuse < 15 )
ERROR_exit("Can't use dataset with < 15 time points per voxel!") ;
if( nuse > 500 && !do_NEW ){
INFO_message("Switching to '-NEW' method since number of time points = %d > 500",nuse) ;
do_NEW = 1 ;
}
if( use_des25 && nuse < 99 ) use_des25 = 0 ;
if( verb ) INFO_message("ignoring first %d time points, using last %d",ignore,nuse);
if( corder > 0 && 4*corder+2 > nuse ){
ERROR_exit("-corder %d is too big for NT=%d",corder,nvals) ;
} else if( corder < 0 ){
corder = rint(nuse/30.0) ; if( corder > 50 && !do_NEW ) corder = 50 ;
if( verb ) INFO_message("using %d time points => -corder %d",nuse,corder) ;
} else {
if( verb ) INFO_message("-corder %d set from command line",corder) ;
}
nxyz = DSET_NVOX(dset) ;
if( verb ) INFO_message("Loading dataset %s",argv[iarg]) ;
DSET_load(dset) ; CHECK_LOAD_ERROR(dset) ;
/*-- create automask --*/
if( !nomask ){
mask = THD_automask( dset ) ;
if( verb ){
ii = THD_countmask( DSET_NVOX(dset) , mask ) ;
INFO_message("%d voxels in the automask [out of %d in dataset]",ii,DSET_NVOX(dset)) ;
}
for( ii=0 ; ii < dilate ; ii++ )
THD_mask_dilate( DSET_NX(dset), DSET_NY(dset), DSET_NZ(dset), mask, 3 ) ;
if( verb ){
ii = THD_countmask( DSET_NVOX(dset) , mask ) ;
INFO_message("%d voxels in the dilated automask [out of %d in dataset]",ii,DSET_NVOX(dset)) ;
}
} else {
if( verb ) INFO_message("processing all %d voxels in dataset",DSET_NVOX(dset)) ;
}
/*-- create empty despiked dataset --*/
oset = EDIT_empty_copy( dset ) ;
EDIT_dset_items( oset ,
ADN_prefix , prefix ,
ADN_brick_fac , NULL ,
ADN_datum_all , datum ,
ADN_none ) ;
if( THD_deathcon() && THD_is_file(DSET_HEADNAME(oset)) )
ERROR_exit("output dataset already exists: %s",DSET_HEADNAME(oset));
tross_Copy_History( oset , dset ) ;
tross_Make_History( "3dDespike" , argc , argv , oset ) ;
/* create bricks (will be filled with zeros) */
for( iv=0 ; iv < nvals ; iv++ )
EDIT_substitute_brick( oset , iv , datum , NULL ) ;
/* copy the ignored bricks */
switch( datum ){
case MRI_short:
for( iv=0 ; iv < ignore ; iv++ ){
sar = DSET_ARRAY(oset,iv) ;
qar = DSET_ARRAY(dset,iv) ;
memcpy( sar , qar , DSET_BRICK_BYTES(dset,iv) ) ;
DSET_unload_one(dset,iv) ;
}
break ;
case MRI_float:
for( iv=0 ; iv < ignore ; iv++ ){
zar = DSET_ARRAY(oset,iv) ;
yar = DSET_ARRAY(dset,iv) ;
memcpy( zar , yar , DSET_BRICK_BYTES(dset,iv) ) ;
DSET_unload_one(dset,iv) ;
}
break ;
}
/*-- setup to save a threshold statistic dataset, if desired --*/
if( tprefix != NULL ){
float *fac ;
tset = EDIT_empty_copy( dset ) ;
fac = (float *) malloc( sizeof(float) * nvals ) ;
for( ii=0 ; ii < nvals ; ii++ ) fac[ii] = TFAC ;
EDIT_dset_items( tset ,
ADN_prefix , tprefix ,
ADN_brick_fac , fac ,
ADN_datum_all , MRI_byte ,
ADN_func_type , FUNC_FIM_TYPE ,
ADN_none ) ;
free(fac) ;
tross_Copy_History( tset , dset ) ;
tross_Make_History( "3dDespike" , argc , argv , tset ) ;
#if 0
if( THD_is_file(DSET_HEADNAME(tset)) )
ERROR_exit("-ssave dataset already exists");
#endif
tross_Copy_History( tset , dset ) ;
tross_Make_History( "3dDespike" , argc , argv , tset ) ;
for( iv=0 ; iv < nvals ; iv++ )
EDIT_substitute_brick( tset , iv , MRI_byte , NULL ) ;
}
/*-- setup to find spikes --*/
sq2p = sqrt(0.5*PI) ;
sfac = sq2p / 1.4826f ;
/* make ref functions */
nref = 2*corder+3 ;
ref = (float **) malloc( sizeof(float *) * nref ) ;
for( jj=0 ; jj < nref ; jj++ )
ref[jj] = (float *) malloc( sizeof(float) * nuse ) ;
/* r(t) = 1 */
for( iv=0 ; iv < nuse ; iv++ ) ref[0][iv] = 1.0 ;
jj = 1 ;
/* r(t) = t - tmid */
{ float tm = 0.5 * (nuse-1.0) ; float fac = 2.0 / nuse ;
for( iv=0 ; iv < nuse ; iv++ ) ref[1][iv] = (iv-tm)*fac ;
jj = 2 ;
/* r(t) = (t-tmid)**jj */
for( ; jj <= polort ; jj++ )
for( iv=0 ; iv < nuse ; iv++ )
ref[jj][iv] = pow( (iv-tm)*fac , (double)jj ) ;
}
for( kk=1 ; kk <= corder ; kk++ ){
fq = (2.0*PI*kk)/nuse ;
/* r(t) = sin(2*PI*k*t/N) */
for( iv=0 ; iv < nuse ; iv++ )
ref[jj][iv] = sin(fq*iv) ;
jj++ ;
/* r(t) = cos(2*PI*k*t/N) */
for( iv=0 ; iv < nuse ; iv++ )
ref[jj][iv] = cos(fq*iv) ;
jj++ ;
}
/****** setup for the NEW solution method [29 Nov 2013] ******/
if( do_NEW ){
NEW_psinv = DES_get_psinv(nuse,nref,ref) ;
INFO_message("Procesing time series with NEW model fit algorithm") ;
} else {
INFO_message("Procesing time series with OLD model fit algorithm") ;
}
/*--- loop over voxels and do work ---*/
#define Laplace_t2p(val) ( 1.0 - nifti_stat2cdf( (val), 15, 0.0, 1.4427 , 0.0 ) )
if( verb ){
if( !localedit ){
INFO_message("smash edit thresholds: %.1f .. %.1f MADs",cut1*sq2p,cut2*sq2p) ;
ININFO_message(" [ %.3f%% .. %.3f%% of normal distribution]",
200.0*qg(cut1*sfac) , 200.0*qg(cut2*sfac) ) ;
ININFO_message(" [ %.3f%% .. %.3f%% of Laplace distribution]" ,
100.0*Laplace_t2p(cut1) , 100.0*Laplace_t2p(cut2) ) ;
} else {
INFO_message("local edit threshold: %.1f MADS",cut2*sq2p) ;
ININFO_message(" [ %.3f%% of normal distribution]",
200.0*qg(cut2*sfac) ) ;
ININFO_message(" [ %.3f%% of Laplace distribution]",
100.0*Laplace_t2p(cut1) ) ;
}
INFO_message("%d slices to process",DSET_NZ(dset)) ;
}
kzold = -1 ;
nspike = 0 ; nbig = 0 ; nproc = 0 ; ctim = NI_clock_time() ;
AFNI_OMP_START ;
#pragma omp parallel if( nxyz > 6666 )
{ int ii , iv , iu , id , jj ;
float *far , *dar , *var , *fitar , *ssp , *fit , *zar ;
short *sar , *qar ; byte *tar ;
float fsig , fq , cls , snew , val ;
float *NEW_wks=NULL ;
#pragma omp critical (DESPIKE_malloc)
{ far = (float *) malloc( sizeof(float) * nvals ) ;
dar = (float *) malloc( sizeof(float) * nvals ) ;
var = (float *) malloc( sizeof(float) * nvals ) ;
fitar = (float *) malloc( sizeof(float) * nvals ) ;
ssp = (float *) malloc( sizeof(float) * nvals ) ;
fit = (float *) malloc( sizeof(float) * nref ) ;
if( do_NEW ) NEW_wks = (float *)malloc(sizeof(float)*DES_workspace_size(nuse,nref)) ;
}
#ifdef USE_OMP
INFO_message("start OpenMP thread #%d",omp_get_thread_num()) ;
#endif
#pragma omp for
for( ii=0 ; ii < nxyz ; ii++ ){ /* ii = voxel index */
if( mask != NULL && mask[ii] == 0 ) continue ; /* skip this voxel */
#ifndef USE_OMP
kz = DSET_index_to_kz(dset,ii) ; /* starting a new slice */
if( kz != kzold ){
if( verb ){
fprintf(stderr, "++ start slice %2d",kz ) ;
if( nproc > 0 ){
pspike = (100.0*nspike)/nproc ;
pbig = (100.0*nbig )/nproc ;
fprintf(stderr,
"; so far %d data points, %d edits [%.3f%%], %d big edits [%.3f%%]",
nproc,nspike,pspike,nbig,pbig ) ;
}
fprintf(stderr,"\n") ;
}
kzold = kz ;
}
#else
if( verb && ii % 2345 == 1234 ) fprintf(stderr,".") ;
#endif
/*** extract ii-th time series into far[] ***/
switch( datum ){
case MRI_short:
for( iv=0 ; iv < nuse ; iv++ ){
qar = DSET_ARRAY(dset,iv+ignore) ; /* skip ignored data */
far[iv] = (float)qar[ii] ;
}
break ;
case MRI_float:
for( iv=0 ; iv < nuse ; iv++ ){
zar = DSET_ARRAY(dset,iv+ignore) ;
far[iv] = zar[ii] ;
}
break ;
}
AAmemcpy(dar,far,sizeof(float)*nuse) ; /* copy time series into dar[] */
/*** solve for L1 fit ***/
if( do_NEW )
cls = DES_solve( NEW_psinv , far , fit , NEW_wks ) ; /* 29 Nov 2013 */
else
cls = cl1_solve( nuse , nref , far , ref , fit,0 ) ; /* the slow part */
if( cls < 0.0f ){ /* fit failed! */
#if 0
fprintf(stderr,"curve fit fails at voxel %d %d %d\n",
DSET_index_to_ix(dset,ii) ,
DSET_index_to_jy(dset,ii) ,
DSET_index_to_kz(dset,ii) ) ;
#endif
continue ; /* skip this voxel */
}
for( iv=0 ; iv < nuse ; iv++ ){ /* detrend */
val = fit[0]
+ fit[1]*ref[1][iv] /* quadratic part of curve fit */
+ fit[2]*ref[2][iv] ;
for( jj=3 ; jj < nref ; jj++ ) /* rest of curve fit */
val += fit[jj] * ref[jj][iv] ;
fitar[iv] = val ; /* save curve fit value */
var[iv] = dar[iv]-val ; /* remove fitted value = resid */
far[iv] = fabsf(var[iv]) ; /* abs value of resid */
}
/*** compute estimate standard deviation of detrended data ***/
fsig = sq2p * qmed_float(nuse,far) ; /* also mangles far array */
/*** process time series for spikes, editing data in dar[] ***/
if( fsig > 0.0f ){ /* data wasn't fit perfectly */
/* find spikiness for each point in time */
fq = 1.0f / fsig ;
for( iv=0 ; iv < nuse ; iv++ ){
ssp[iv] = fq * var[iv] ; /* spikiness s = how many sigma out */
}
/* save spikiness in -ssave datset */
if( tset != NULL ){
for( iv=0 ; iv < nuse ; iv++ ){
tar = DSET_ARRAY(tset,iv+ignore) ;
snew = ITFAC*fabsf(ssp[iv]) ; /* scale for byte storage */
tar[ii] = BYTEIZE(snew) ; /* cf. mrilib.h */
}
}
/* process values of |s| > cut1, editing dar[] */
for( iv=0 ; iv < nuse ; iv++ ){ /* loop over time points */
if( !localedit ){ /** classic 'smash' edit **/
if( ssp[iv] > cut1 ){
snew = cut1 + c21*mytanh((ssp[iv]-cut1)*ic21) ; /* edit s down */
dar[iv] = fitar[iv] + snew*fsig ;
#pragma omp critical (DESPIKE_counter)
{ nspike++ ; if( ssp[iv] > cut2 ) nbig++ ; }
} else if( ssp[iv] < -cut1 ){
snew = -cut1 + c21*mytanh((ssp[iv]+cut1)*ic21) ; /* edit s up */
dar[iv] = fitar[iv] + snew*fsig ;
#pragma omp critical (DESPIKE_counter)
{ nspike++ ; if( ssp[iv] < -cut2 ) nbig++ ; }
}
} else { /** local edit: 04 Apr 2007 **/
if( ssp[iv] >= cut2 || ssp[iv] <= -cut2 ){
for( iu=iv+1 ; iu < nuse ; iu++ ) /* find non-spike above */
if( ssp[iu] < cut2 && ssp[iu] > -cut2 ) break ;
for( id=iv-1 ; id >= 0 ; id-- ) /* find non-spike below */
if( ssp[id] < cut2 && ssp[id] > -cut2 ) break ;
switch( (id>=0) + 2*(iu<nuse) ){ /* compute replacement val */
case 3: val = 0.5*(dar[iu]+dar[id]); break; /* iu and id OK */
case 2: val = dar[iu] ; break; /* only iu OK */
case 1: val = dar[id] ; break; /* only id OK */
default: val = fitar[iv] ; break; /* shouldn't be */
}
dar[iv] = val ;
#pragma omp critical (DESPIKE_counter)
{ nspike++ ; nbig++ ; }
}
}
} /* end of loop over time points */
#pragma omp atomic
nproc += nuse ; /* number data points processed */
} /* end of processing time series when fsig is positive */
/* put dar[] time series (possibly edited above) into output bricks */
switch( datum ){
case MRI_short:
for( iv=0 ; iv < nuse ; iv++ ){
sar = DSET_ARRAY(oset,iv+ignore) ; /* output brick */
sar[ii] = (short)dar[iv] ; /* original or mutated data */
}
break ;
case MRI_float:
for( iv=0 ; iv < nuse ; iv++ ){
zar = DSET_ARRAY(oset,iv+ignore) ; /* output brick */
zar[ii] = dar[iv] ; /* original or mutated data */
}
break ;
}
} /* end of loop over voxels #ii */
#pragma omp critical (DESPIKE_malloc)
{ free(fit); free(ssp); free(fitar); free(var); free(dar); free(far);
if( do_NEW ) free(NEW_wks) ; }
} /* end OpenMP */
AFNI_OMP_END ;
#ifdef USE_OMP
if( verb ) fprintf(stderr,"\n") ;
#endif
ctim = NI_clock_time() - ctim ;
INFO_message( "Elapsed despike time = %s" , nice_time_string(ctim) ) ;
if( ctim > 345678 && !do_NEW )
ININFO_message("That was SLOW -- try the '-NEW' option for a speedup") ;
#ifdef USE_OMP
if( verb ) fprintf(stderr,"\n") ;
#endif
/*--- finish up ---*/
if( do_NEW ) mri_free(NEW_psinv) ;
DSET_delete(dset) ; /* delete input dataset */
if( verb ){
if( nproc > 0 ){
pspike = (100.0*nspike)/nproc ;
pbig = (100.0*nbig )/nproc ;
INFO_message("FINAL: %d data points, %d edits [%.3f%%], %d big edits [%.3f%%]",
nproc,nspike,pspike,nbig,pbig ) ;
} else {
INFO_message("FINAL: no good voxels found to process!!??") ;
}
}
/* write results */
DSET_write(oset) ;
if( verb ) WROTE_DSET(oset) ;
DSET_delete(oset) ;
if( tset != NULL ){
DSET_write(tset) ;
if( verb ) WROTE_DSET(tset) ;
DSET_delete(tset) ;
}
exit( THD_get_write_error_count() ) ;
}
