MRI_IMAGE * DES_get_psinv( int ntim , int nref , float **ref )
{
MRI_IMAGE *refim , *psinv ; float *refar , *jar ; int ii , jj ;
refim = mri_new(ntim,nref,MRI_float) ;
refar = MRI_FLOAT_PTR(refim) ;
for( jj=0 ; jj < nref ; jj++ ){
jar = refar + jj*ntim ;
for( ii=0 ; ii < ntim ; ii++ ) jar[ii] = ref[jj][ii] ;
}
mri_matrix_psinv_svd(1) ;
psinv = mri_matrix_psinv(refim,NULL,0.0f) ;
mri_free(refim) ;
return psinv ;
}
int main( int argc , char *argv[] )
{
THD_3dim_dataset *inset=NULL ;
byte *mask=NULL ; int mask_nx=0,mask_ny=0,mask_nz=0 , automask=0 , masknum=0 ;
int iarg=1 , verb=1 , ntype=0 , nev,kk,ii,nxyz,nt ;
float na,nb,nc , dx,dy,dz ;
MRI_IMARR *imar=NULL ; int *ivox ; MRI_IMAGE *pim ;
int do_vmean=0 , do_vnorm=0 , sval_itop=0 ;
int polort=-1 ; float *ev ;
MRI_IMARR *ortar ; MRI_IMAGE *ortim ; int nyort=0 ;
float bpass_L=0.0f , bpass_H=0.0f , dtime ; int do_bpass=0 ;
if( argc < 2 || strcmp(argv[1],"-help") == 0 ){
printf(
"Usage: 3dmaskSVD [options] inputdataset\n"
"Author: Zhark the Gloriously Singular\n"
"\n"
"* Computes the principal singular vector of the time series\n"
" vectors extracted from the input dataset over the input mask.\n"
" ++ You can use the '-sval' option to change which singular\n"
" vectors are output.\n"
"* The sign of the output vector is chosen so that the average\n"
" of arctanh(correlation coefficient) over all input data\n"
" vectors (from the mask) is positive.\n"
"* The output vector is normalized: the sum of its components\n"
" squared is 1.\n"
"* You probably want to use 3dDetrend (or something similar) first,\n"
" to get rid of annoying artifacts, such as motion, breathing,\n"
" dark matter interactions with the brain, etc.\n"
" ++ If you are lazy scum like Zhark, you might be able to get\n"
" away with using the '-polort' option.\n"
" ++ In particular, if your data time series has a nonzero mean,\n"
" then you probably want at least '-polort 0' to remove the\n"
" mean, otherwise you'll pretty much just get a constant\n"
" time series as the principal singular vector!\n"
"* An alternative to this program would be 3dmaskdump followed\n"
" by 1dsvd, which could give you all the singular vectors you\n"
" could ever want, and much more -- enough to confuse you for days.\n"
" ++ In particular, although you COULD input a 1D file into\n"
" 3dmaskSVD, the 1dsvd program would make much more sense.\n"
"* This program will be pretty slow if there are over about 2000\n"
" voxels in the mask. It could be made more efficient for\n"
" such cases, but you'll have to give Zhark some 'incentive'.\n"
"* Result vector goes to stdout. Redirect per your pleasures and needs.\n"
"* Also see program 3dLocalSVD if you want to compute the principal\n"
" singular time series vector from a neighborhood of EACH voxel.\n"
" ++ (Which is a pretty slow operation!)\n"
"* http://en.wikipedia.org/wiki/Singular_value_decomposition\n"
"\n"
"-------\n"
"Options:\n"
"-------\n"
" -vnorm = L2 normalize all time series before SVD [recommended!]\n"
" -sval a = output singular vectors 0 .. a [default a=0 = first one only]\n"
" -mask mset = define the mask [default is entire dataset == slow!]\n"
" -automask = you'll have to guess what this option does\n"
" -polort p = if you are lazy and didn't run 3dDetrend (like Zhark)\n"
" -bpass L H = bandpass [mutually exclusive with -polort]\n"
" -ort xx.1D = time series to remove from the data before SVD-ization\n"
" ++ You can give more than 1 '-ort' option\n"
" ++ 'xx.1D' can contain more than 1 column\n"
" -input ddd = alternative way to give the input dataset name\n"
"\n"
"-------\n"
"Example:\n"
"-------\n"
" You have a mask dataset with discrete values 1, 2, ... 77 indicating\n"
" some ROIs; you want to get the SVD from each ROI's time series separately,\n"
" and then put these into 1 big 77 column .1D file. You can do this using\n"
" a csh shell script like the one below:\n"
"\n"
" # Compute the individual SVD vectors\n"
" foreach mm ( `count 1 77` )\n"
" 3dmaskSVD -vnorm -mask mymask+orig\"<${mm}..${mm}>\" epi+orig > qvec${mm}.1D\n"
" end\n"
" # Glue them together into 1 big file, then delete the individual files\n"
" 1dcat qvec*.1D > allvec.1D\n"
" /bin/rm -f qvec*.1D\n"
" # Plot the results to a JPEG file, then compute their correlation matrix\n"
" 1dplot -one -nopush -jpg allvec.jpg allvec.1D\n"
" 1ddot -terse allvec.1D > allvec_COR.1D\n"
"\n"
" [[ If you use the bash shell, you'll have to figure out the syntax ]]\n"
" [[ yourself. Zhark has no sympathy for you bash shell infidels, and ]]\n"
" [[ considers you only slightly better than those lowly Emacs users. ]]\n"
" [[ And do NOT ever even mention 'nedit' in Zhark's august presence! ]]\n"
) ;
PRINT_COMPILE_DATE ; exit(0) ;
}
/*---- official startup ---*/
PRINT_VERSION("3dmaskSVD"); mainENTRY("3dmaskSVD main"); machdep();
AFNI_logger("3dmaskSVD",argc,argv); AUTHOR("Zhark the Singular");
/*---- loop over options ----*/
INIT_IMARR(ortar) ;
mpv_sign_meth = AFNI_yesenv("AFNI_3dmaskSVD_meansign") ;
while( iarg < argc && argv[iarg][0] == '-' ){
if( strcasecmp(argv[iarg],"-bpass") == 0 ){
if( iarg+2 >= argc ) ERROR_exit("need 2 args after -bpass") ;
bpass_L = (float)strtod(argv[++iarg],NULL) ;
bpass_H = (float)strtod(argv[++iarg],NULL) ;
if( bpass_L < 0.0f || bpass_H <= bpass_L )
ERROR_exit("Illegal values after -bpass: %g %g",bpass_L,bpass_H) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-ort") == 0 ){ /* 01 Oct 2009 */
int nx,ny ;
if( ++iarg >= argc ) ERROR_exit("Need argument after '-ort'") ;
ortim = mri_read_1D( argv[iarg] ) ;
if( ortim == NULL ) ERROR_exit("-ort '%s': Can't read 1D file",argv[iarg]) ;
nx = ortim->nx ; ny = ortim->ny ;
if( nx == 1 && ny > 1 ){
MRI_IMAGE *tim=mri_transpose(ortim); mri_free(ortim); ortim = tim; ny = 1;
}
mri_add_name(argv[iarg],ortim) ; ADDTO_IMARR(ortar,ortim) ; nyort += ny ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-polort") == 0 ){
char *qpt ;
if( ++iarg >= argc ) ERROR_exit("Need argument after '-polort'") ;
polort = (int)strtod(argv[iarg],&qpt) ;
if( *qpt != '\0' ) WARNING_message("Illegal non-numeric value after -polort") ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-vnorm") == 0 ){
do_vnorm = 1 ; iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-input") == 0 ){
if( inset != NULL ) ERROR_exit("Can't have two -input options") ;
if( ++iarg >= argc ) ERROR_exit("Need argument after '-input'") ;
inset = THD_open_dataset( argv[iarg] ) ;
CHECK_OPEN_ERROR(inset,argv[iarg]) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-sval") == 0 ){
if( ++iarg >= argc ) ERROR_exit("Need argument after '-sval'") ;
sval_itop = (int)strtod(argv[iarg],NULL) ;
if( sval_itop < 0 ){ sval_itop = 0 ; WARNING_message("'-sval' reset to 0") ; }
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-mask") == 0 ){
THD_3dim_dataset *mset ; int mmm ;
if( ++iarg >= argc ) ERROR_exit("Need argument after '-mask'") ;
if( mask != NULL || automask ) ERROR_exit("Can't have two mask inputs") ;
mset = THD_open_dataset( argv[iarg] ) ;
CHECK_OPEN_ERROR(mset,argv[iarg]) ;
DSET_load(mset) ; CHECK_LOAD_ERROR(mset) ;
mask_nx = DSET_NX(mset); mask_ny = DSET_NY(mset); mask_nz = DSET_NZ(mset);
mask = THD_makemask( mset , 0 , 0.5f, 0.0f ) ; DSET_delete(mset) ;
if( mask == NULL ) ERROR_exit("Can't make mask from dataset '%s'",argv[iarg]) ;
masknum = mmm = THD_countmask( mask_nx*mask_ny*mask_nz , mask ) ;
INFO_message("Number of voxels in mask = %d",mmm) ;
if( mmm < 2 ) ERROR_exit("Mask is too small to process") ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-automask") == 0 ){
if( mask != NULL ) ERROR_exit("Can't have two mask inputs!") ;
automask = 1 ; iarg++ ; continue ;
}
ERROR_exit("Unknown option '%s'",argv[iarg]) ;
} /*--- end of loop over options ---*/
/*---- deal with input dataset ----*/
if( inset == NULL ){
if( iarg >= argc ) ERROR_exit("No input dataset on command line?") ;
inset = THD_open_dataset( argv[iarg] ) ;
CHECK_OPEN_ERROR(inset,argv[iarg]) ;
}
nt = DSET_NVALS(inset) ; /* vector lengths */
if( nt < 9 )
ERROR_exit("Must have at least 9 values per voxel") ;
if( polort+1 >= nt )
ERROR_exit("'-polort %d' too big for time series length = %d",polort,nt) ;
DSET_load(inset) ; CHECK_LOAD_ERROR(inset) ;
nxyz = DSET_NVOX(inset) ;
DSET_UNMSEC(inset) ;
dtime = DSET_TR(inset) ;
if( dtime <= 0.0f ) dtime = 1.0f ;
do_bpass = (bpass_L < bpass_H) ;
if( do_bpass ){
kk = THD_bandpass_OK( nt , dtime , bpass_L , bpass_H , 1 ) ;
if( kk <= 0 ) ERROR_exit("Can't continue since -bpass setup is illegal") ;
polort = -1 ;
}
/*--- deal with the masking ---*/
if( mask != NULL ){
if( mask_nx != DSET_NX(inset) ||
mask_ny != DSET_NY(inset) ||
mask_nz != DSET_NZ(inset) )
ERROR_exit("-mask dataset grid doesn't match input dataset") ;
} else if( automask ){
int mmm ;
mask = THD_automask( inset ) ;
if( mask == NULL )
ERROR_message("Can't create -automask from input dataset?") ;
masknum = mmm = THD_countmask( DSET_NVOX(inset) , mask ) ;
INFO_message("Number of voxels in automask = %d",mmm) ;
if( mmm < 9 ) ERROR_exit("Automask is too small to process") ;
} else {
mask = (byte *)malloc(sizeof(byte)*nxyz) ; masknum = nxyz ;
memset( mask , 1 , sizeof(byte)*nxyz ) ;
INFO_message("Using all %d voxels in dataset",nxyz) ;
}
nev = MIN(nt,masknum) ; /* max possible number of eigenvalues */
if( sval_itop >= nev ){
sval_itop = nev-1 ;
WARNING_message("'-sval' reset to '%d'",sval_itop) ;
}
mri_principal_vector_params( 0 , do_vnorm , sval_itop ) ;
mri_principal_setev(nev) ;
/*-- get data vectors --*/
ivox = (int *)malloc(sizeof(int)*masknum) ;
for( kk=ii=0 ; ii < nxyz ; ii++ ) if( mask[ii] ) ivox[kk++] = ii ;
INFO_message("Extracting data vectors") ;
imar = THD_extract_many_series( masknum, ivox, inset ) ; DSET_unload(inset) ;
if( imar == NULL ) ERROR_exit("Can't get data vector?!") ;
/*-- detrending --*/
if( polort >= 0 || nyort > 0 || do_bpass ){
float **polref=NULL ; float *tsar ;
int nort=IMARR_COUNT(ortar) , nref=0 ;
if( polort >= 0 ){ /* polynomials */
nref = polort+1 ; polref = THD_build_polyref(nref,nt) ;
}
if( nort > 0 ){ /* other orts */
float *oar , *par ; int nx,ny , qq,tt ;
for( kk=0 ; kk < nort ; kk++ ){ /* loop over input -ort files */
ortim = IMARR_SUBIM(ortar,kk) ;
nx = ortim->nx ; ny = ortim->ny ;
if( nx < nt )
ERROR_exit("-ort '%s' length %d shorter than dataset length %d" ,
ortim->name , nx , nt ) ;
polref = (float **)realloc(polref,(nref+ny)*sizeof(float *)) ;
oar = MRI_FLOAT_PTR(ortim) ;
for( qq=0 ; qq < ny ; qq++,oar+=nx ){
par = polref[nref+qq] = (float *)malloc(sizeof(float)*nt) ;
for( tt=0 ; tt < nt ; tt++ ) par[tt] = oar[tt] ;
if( polort == 0 ) THD_const_detrend (nt,par,NULL) ;
else if( polort > 0 ) THD_linear_detrend(nt,par,NULL,NULL) ;
}
nref += ny ;
}
DESTROY_IMARR(ortar) ;
}
if( !do_bpass ){ /* old style ort-ification */
MRI_IMAGE *imq , *imp ; float *qar ;
INFO_message("Detrending data vectors") ;
#if 1
imq = mri_new( nt , nref , MRI_float) ; qar = MRI_FLOAT_PTR(imq) ;
for( kk=0 ; kk < nref ; kk++ )
memcpy( qar+kk*nt , polref[kk] , sizeof(float)*nt ) ;
imp = mri_matrix_psinv( imq , NULL , 1.e-8 ) ;
for( kk=0 ; kk < IMARR_COUNT(imar) ; kk++ ){
mri_matrix_detrend( IMARR_SUBIM(imar,kk) , imq , imp ) ;
}
mri_free(imp) ; mri_free(imq) ;
#else
for( kk=0 ; kk < IMARR_COUNT(imar) ; kk++ ){
tsar = MRI_FLOAT_PTR(IMARR_SUBIM(imar,kk)) ;
THD_generic_detrend_LSQ( nt , tsar , -1 , nref , polref , NULL ) ;
}
#endif
} else { /* bandpass plus (maybe) orts */
float **vec = (float **)malloc(sizeof(float *)*IMARR_COUNT(imar)) ;
INFO_message("Bandpassing data vectors") ;
for( kk=0 ; kk < IMARR_COUNT(imar) ; kk++ )
vec[kk] = MRI_FLOAT_PTR(IMARR_SUBIM(imar,kk)) ;
(void)THD_bandpass_vectors( nt , IMARR_COUNT(imar) , vec ,
dtime , bpass_L , bpass_H ,
2 , nref , polref ) ;
free(vec) ;
}
for( kk=0 ; kk < nref; kk++ ) free(polref[kk]) ;
free(polref) ;
} /* end of detrendization */
/*--- the actual work ---*/
INFO_message("Computing SVD") ;
pim = mri_principal_vector( imar ) ; DESTROY_IMARR(imar) ;
if( pim == NULL ) ERROR_exit("SVD failure!?!") ;
ev = mri_principal_getev() ;
switch(sval_itop+1){
case 1:
INFO_message("First singular value: %g",ev[0]) ; break ;
case 2:
INFO_message("First 2 singular values: %g %g",ev[0],ev[1]) ; break ;
case 3:
INFO_message("First 3 singular values: %g %g %g",ev[0],ev[1],ev[2]) ; break ;
case 4:
INFO_message("First 4 singular values: %g %g %g %g",ev[0],ev[1],ev[2],ev[3]) ; break ;
default:
case 5:
INFO_message("First 5 singular values: %g %g %g %g %g",ev[0],ev[1],ev[2],ev[3],ev[4]) ; break ;
}
mri_write_1D(NULL,pim) ;
exit(0) ;
}
int main( int argc , char *argv[] )
{
THD_3dim_dataset *yset=NULL , *aset=NULL , *mset=NULL , *wset=NULL ;
MRI_IMAGE *fim=NULL, *qim,*tim, *pfim=NULL , *vim , *wim=NULL ;
float *flar , *qar,*tar, *par=NULL , *var , *war=NULL ;
MRI_IMARR *fimar=NULL ;
MRI_IMAGE *aim , *yim ; float *aar , *yar ;
int nt=0 , nxyz=0 , nvox=0 , nparam=0 , nqbase , polort=0 , ii,jj,kk,bb ;
byte *mask=NULL ; int nmask=0 , iarg ;
char *fname_out="-" ; /** equiv to stdout **/
float alpha=0.0f ;
int nfir =0 ; float firwt[5]={0.09f,0.25f,0.32f,0.25f,0.09f} ;
int nmed =0 ;
int nwt =0 ;
#define METHOD_C 3
#define METHOD_K 11
int method = METHOD_C ;
/**--- help the pitiful user? ---**/
if( argc < 2 || strcmp(argv[1],"-help") == 0 ){
printf(
"Usage: 3dInvFMRI [options]\n"
"Program to compute stimulus time series, given a 3D+time dataset\n"
"and an activation map (the inverse of the usual FMRI analysis problem).\n"
"-------------------------------------------------------------------\n"
"OPTIONS:\n"
"\n"
" -data yyy =\n"
" *OR* = Defines input 3D+time dataset [a non-optional option].\n"
" -input yyy =\n"
"\n"
" -map aaa = Defines activation map; 'aaa' should be a bucket dataset,\n"
" each sub-brick of which defines the beta weight map for\n"
" an unknown stimulus time series [also non-optional].\n"
"\n"
" -mapwt www = Defines a weighting factor to use for each element of\n"
" the map. The dataset 'www' can have either 1 sub-brick,\n"
" or the same number as in the -map dataset. In the\n"
" first case, in each voxel, each sub-brick of the map\n"
" gets the same weight in the least squares equations.\n"
" [default: all weights are 1]\n"
"\n"
" -mask mmm = Defines a mask dataset, to restrict input voxels from\n"
" -data and -map. [default: all voxels are used]\n"
"\n"
" -base fff = Each column of the 1D file 'fff' defines a baseline time\n"
" series; these columns should be the same length as\n"
" number of time points in 'yyy'. Multiple -base options\n"
" can be given.\n"
" -polort pp = Adds polynomials of order 'pp' to the baseline collection.\n"
" The default baseline model is '-polort 0' (constant).\n"
" To specify no baseline model at all, use '-polort -1'.\n"
"\n"
" -out vvv = Name of 1D output file will be 'vvv'.\n"
" [default = '-', which is stdout; probably not good]\n"
"\n"
" -method M = Determines the method to use. 'M' is a single letter:\n"
" -method C = least squares fit to data matrix Y [default]\n"
" -method K = least squares fit to activation matrix A\n"
"\n"
" -alpha aa = Set the 'alpha' factor to 'aa'; alpha is used to penalize\n"
" large values of the output vectors. Default is 0.\n"
" A large-ish value for alpha would be 0.1.\n"
"\n"
" -fir5 = Smooth the results with a 5 point lowpass FIR filter.\n"
" -median5 = Smooth the results with a 5 point median filter.\n"
" [default: no smoothing; only 1 of these can be used]\n"
"-------------------------------------------------------------------\n"
"METHODS:\n"
" Formulate the problem as\n"
" Y = V A' + F C' + errors\n"
" where Y = data matrix (N x M) [from -data]\n"
" V = stimulus (N x p) [to -out]\n"
" A = map matrix (M x p) [from -map]\n"
" F = baseline matrix (N x q) [from -base and -polort]\n"
" C = baseline weights (M x q) [not computed]\n"
" N = time series length = length of -data file\n"
" M = number of voxels in mask\n"
" p = number of stimulus time series to estimate\n"
" = number of parameters in -map file\n"
" q = number of baseline parameters\n"
" and ' = matrix transpose operator\n"
" Next, define matrix Z (Y detrended relative to columns of F) by\n"
" -1\n"
" Z = [I - F(F'F) F'] Y\n"
"-------------------------------------------------------------------\n"
" The method C solution is given by\n"
" -1\n"
" V0 = Z A [A'A]\n"
"\n"
" This solution minimizes the sum of squares over the N*M elements\n"
" of the matrix Y - V A' + F C' (N.B.: A' means A-transpose).\n"
"-------------------------------------------------------------------\n"
" The method K solution is given by\n"
" -1 -1\n"
" W = [Z Z'] Z A and then V = W [W'W]\n"
"\n"
" This solution minimizes the sum of squares of the difference between\n"
" the A(V) predicted from V and the input A, where A(V) is given by\n"
" -1\n"
" A(V) = Z' V [V'V] = Z'W\n"
"-------------------------------------------------------------------\n"
" Technically, the solution is unidentfiable up to an arbitrary\n"
" multiple of the columns of F (i.e., V = V0 + F G, where G is\n"
" an arbitrary q x p matrix); the solution above is the solution\n"
" that is orthogonal to the columns of F.\n"
"\n"
"-- RWCox - March 2006 - purely for experimental purposes!\n"
) ;
printf("\n"
"===================== EXAMPLE USAGE =====================================\n"
"** Step 1: From a training dataset, generate activation map.\n"
" The input dataset has 4 runs, each 108 time points long. 3dDeconvolve\n"
" is used on the first 3 runs (time points 0..323) to generate the\n"
" activation map. There are two visual stimuli (Complex and Simple).\n"
"\n"
" 3dDeconvolve -x1D xout_short_two.1D -input rall_vr+orig'[0..323]' \\\n"
" -num_stimts 2 \\\n"
" -stim_file 1 hrf_complex.1D -stim_label 1 Complex \\\n"
" -stim_file 2 hrf_simple.1D -stim_label 2 Simple \\\n"
" -concat '1D:0,108,216' \\\n"
" -full_first -fout -tout \\\n"
" -bucket func_ht2_short_two -cbucket cbuc_ht2_short_two\n"
"\n"
" N.B.: You may want to de-spike, smooth, and register the 3D+time\n"
" dataset prior to the analysis (as usual). These steps are not\n"
" shown here -- I'm presuming you know how to use AFNI already.\n"
"\n"
"** Step 2: Create a mask of highly activated voxels.\n"
" The F statistic threshold is set to 30, corresponding to a voxel-wise\n"
" p = 1e-12 = very significant. The mask is also lightly clustered, and\n"
" restricted to brain voxels.\n"
"\n"
" 3dAutomask -prefix Amask rall_vr+orig\n"
" 3dcalc -a 'func_ht2_short+orig[0]' -b Amask+orig -datum byte \\\n"
" -nscale -expr 'step(a-30)*b' -prefix STmask300\n"
" 3dmerge -dxyz=1 -1clust 1.1 5 -prefix STmask300c STmask300+orig\n"
"\n"
"** Step 3: Run 3dInvFMRI to estimate the stimulus functions in run #4.\n"
" Run #4 is time points 324..431 of the 3D+time dataset (the -data\n"
" input below). The -map input is the beta weights extracted from\n"
" the -cbucket output of 3dDeconvolve.\n"
"\n"
" 3dInvFMRI -mask STmask300c+orig \\\n"
" -data rall_vr+orig'[324..431]' \\\n"
" -map cbuc_ht2_short_two+orig'[6..7]' \\\n"
" -polort 1 -alpha 0.01 -median5 -method K \\\n"
" -out ii300K_short_two.1D\n"
"\n"
" 3dInvFMRI -mask STmask300c+orig \\\n"
" -data rall_vr+orig'[324..431]' \\\n"
" -map cbuc_ht2_short_two+orig'[6..7]' \\\n"
" -polort 1 -alpha 0.01 -median5 -method C \\\n"
" -out ii300C_short_two.1D\n"
"\n"
"** Step 4: Plot the results, and get confused.\n"
"\n"
" 1dplot -ynames VV KK CC -xlabel Run#4 -ylabel ComplexStim \\\n"
" hrf_complex.1D'{324..432}' \\\n"
" ii300K_short_two.1D'[0]' \\\n"
" ii300C_short_two.1D'[0]'\n"
"\n"
" 1dplot -ynames VV KK CC -xlabel Run#4 -ylabel SimpleStim \\\n"
" hrf_simple.1D'{324..432}' \\\n"
" ii300K_short_two.1D'[1]' \\\n"
" ii300C_short_two.1D'[1]'\n"
"\n"
" N.B.: I've found that method K works better if MORE voxels are\n"
" included in the mask (lower threshold) and method C if\n"
" FEWER voxels are included. The above threshold gave 945\n"
" voxels being used to determine the 2 output time series.\n"
"=========================================================================\n"
) ;
PRINT_COMPILE_DATE ; exit(0) ;
}
/**--- bureaucracy ---**/
mainENTRY("3dInvFMRI main"); machdep();
PRINT_VERSION("3dInvFMRI"); AUTHOR("Zhark");
AFNI_logger("3dInvFMRI",argc,argv) ;
/**--- scan command line ---**/
iarg = 1 ;
while( iarg < argc ){
if( strcmp(argv[iarg],"-method") == 0 ){
switch( argv[++iarg][0] ){
default:
WARNING_message("Ignoring illegal -method '%s'",argv[iarg]) ;
break ;
case 'C': method = METHOD_C ; break ;
case 'K': method = METHOD_K ; break ;
}
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-fir5") == 0 ){
if( nmed > 0 ) WARNING_message("Ignoring -fir5 in favor of -median5") ;
else nfir = 5 ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-median5") == 0 ){
if( nfir > 0 ) WARNING_message("Ignoring -median5 in favor of -fir5") ;
else nmed = 5 ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-alpha") == 0 ){
alpha = (float)strtod(argv[++iarg],NULL) ;
if( alpha <= 0.0f ){
alpha = 0.0f ; WARNING_message("-alpha '%s' ignored!",argv[iarg]) ;
}
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-data") == 0 || strcmp(argv[iarg],"-input") == 0 ){
if( yset != NULL ) ERROR_exit("Can't input 2 3D+time datasets") ;
yset = THD_open_dataset(argv[++iarg]) ;
CHECK_OPEN_ERROR(yset,argv[iarg]) ;
nt = DSET_NVALS(yset) ;
if( nt < 2 ) ERROR_exit("Only 1 sub-brick in dataset %s",argv[iarg]) ;
nxyz = DSET_NVOX(yset) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-map") == 0 ){
if( aset != NULL ) ERROR_exit("Can't input 2 -map datasets") ;
aset = THD_open_dataset(argv[++iarg]) ;
CHECK_OPEN_ERROR(aset,argv[iarg]) ;
nparam = DSET_NVALS(aset) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-mapwt") == 0 ){
if( wset != NULL ) ERROR_exit("Can't input 2 -mapwt datasets") ;
wset = THD_open_dataset(argv[++iarg]) ;
CHECK_OPEN_ERROR(wset,argv[iarg]) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-mask") == 0 ){
if( mset != NULL ) ERROR_exit("Can't input 2 -mask datasets") ;
mset = THD_open_dataset(argv[++iarg]) ;
CHECK_OPEN_ERROR(mset,argv[iarg]) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-polort") == 0 ){
char *cpt ;
polort = (int)strtod(argv[++iarg],&cpt) ;
if( *cpt != '\0' ) WARNING_message("Illegal non-numeric value after -polort") ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-out") == 0 ){
fname_out = strdup(argv[++iarg]) ;
if( !THD_filename_ok(fname_out) )
ERROR_exit("Bad -out filename '%s'",fname_out) ;
iarg++ ; continue ;
}
if( strcmp(argv[iarg],"-base") == 0 ){
if( fimar == NULL ) INIT_IMARR(fimar) ;
qim = mri_read_1D( argv[++iarg] ) ;
if( qim == NULL ) ERROR_exit("Can't read 1D file %s",argv[iarg]) ;
ADDTO_IMARR(fimar,qim) ;
iarg++ ; continue ;
}
ERROR_exit("Unrecognized option '%s'",argv[iarg]) ;
}
/**--- finish up processing options ---**/
if( yset == NULL ) ERROR_exit("No input 3D+time dataset?!") ;
if( aset == NULL ) ERROR_exit("No input FMRI -map dataset?!") ;
if( DSET_NVOX(aset) != nxyz )
ERROR_exit("Grid mismatch between -data and -map") ;
INFO_message("Loading dataset for Y") ;
DSET_load(yset); CHECK_LOAD_ERROR(yset) ;
INFO_message("Loading dataset for A") ;
DSET_load(aset); CHECK_LOAD_ERROR(aset) ;
if( wset != NULL ){
if( DSET_NVOX(wset) != nxyz )
ERROR_exit("Grid mismatch between -data and -mapwt") ;
nwt = DSET_NVALS(wset) ;
if( nwt > 1 && nwt != nparam )
ERROR_exit("Wrong number of values=%d in -mapwt; should be 1 or %d",
nwt , nparam ) ;
INFO_message("Loading dataset for mapwt") ;
DSET_load(wset); CHECK_LOAD_ERROR(wset) ;
}
if( mset != NULL ){
if( DSET_NVOX(mset) != nxyz )
ERROR_exit("Grid mismatch between -data and -mask") ;
INFO_message("Loading dataset for mask") ;
DSET_load(mset); CHECK_LOAD_ERROR(mset) ;
mask = THD_makemask( mset , 0 , 1.0f,-1.0f ); DSET_delete(mset);
nmask = THD_countmask( nxyz , mask ) ;
if( nmask < 3 ){
WARNING_message("Mask has %d voxels -- ignoring!",nmask) ;
free(mask) ; mask = NULL ; nmask = 0 ;
}
}
nvox = (nmask > 0) ? nmask : nxyz ;
INFO_message("N = time series length = %d",nt ) ;
INFO_message("M = number of voxels = %d",nvox ) ;
INFO_message("p = number of params = %d",nparam) ;
/**--- set up baseline funcs in one array ---*/
nqbase = (polort >= 0 ) ? polort+1 : 0 ;
if( fimar != NULL ){
for( kk=0 ; kk < IMARR_COUNT(fimar) ; kk++ ){
qim = IMARR_SUBIMAGE(fimar,kk) ;
if( qim != NULL && qim->nx != nt )
WARNING_message("-base #%d length=%d; data length=%d",kk+1,qim->nx,nt) ;
nqbase += qim->ny ;
}
}
INFO_message("q = number of baselines = %d",nqbase) ;
#undef F
#define F(i,j) flar[(i)+(j)*nt] /* nt X nqbase */
if( nqbase > 0 ){
fim = mri_new( nt , nqbase , MRI_float ) ; /* F matrix */
flar = MRI_FLOAT_PTR(fim) ;
bb = 0 ;
if( polort >= 0 ){ /** load polynomial baseline **/
double a = 2.0/(nt-1.0) ;
for( jj=0 ; jj <= polort ; jj++ ){
for( ii=0 ; ii < nt ; ii++ )
F(ii,jj) = (float)Plegendre( a*ii-1.0 , jj ) ;
}
bb = polort+1 ;
}
#undef Q
#define Q(i,j) qar[(i)+(j)*qim->nx] /* qim->nx X qim->ny */
if( fimar != NULL ){ /** load -base baseline columns **/
for( kk=0 ; kk < IMARR_COUNT(fimar) ; kk++ ){
qim = IMARR_SUBIMAGE(fimar,kk) ; qar = MRI_FLOAT_PTR(qim) ;
for( jj=0 ; jj < qim->ny ; jj++ ){
for( ii=0 ; ii < nt ; ii++ )
F(ii,bb+jj) = (ii < qim->nx) ? Q(ii,jj) : 0.0f ;
}
bb += qim->ny ;
}
DESTROY_IMARR(fimar) ; fimar=NULL ;
}
/* remove mean from each column after first? */
if( polort >= 0 && nqbase > 1 ){
float sum ;
for( jj=1 ; jj < nqbase ; jj++ ){
sum = 0.0f ;
for( ii=0 ; ii < nt ; ii++ ) sum += F(ii,jj) ;
sum /= nt ;
for( ii=0 ; ii < nt ; ii++ ) F(ii,jj) -= sum ;
}
}
/* compute pseudo-inverse of baseline matrix,
so we can project it out from the data time series */
/* -1 */
/* (F'F) F' matrix */
INFO_message("Computing pseudo-inverse of baseline matrix F") ;
pfim = mri_matrix_psinv(fim,NULL,0.0f) ; par = MRI_FLOAT_PTR(pfim) ;
#undef P
#define P(i,j) par[(i)+(j)*nqbase] /* nqbase X nt */
#if 0
qim = mri_matrix_transpose(pfim) ; /** save to disk? **/
mri_write_1D( "Fpsinv.1D" , qim ) ;
mri_free(qim) ;
#endif
}
/**--- set up map image into aim/aar = A matrix ---**/
#undef GOOD
#define GOOD(i) (mask==NULL || mask[i])
#undef A
#define A(i,j) aar[(i)+(j)*nvox] /* nvox X nparam */
INFO_message("Loading map matrix A") ;
aim = mri_new( nvox , nparam , MRI_float ); aar = MRI_FLOAT_PTR(aim);
for( jj=0 ; jj < nparam ; jj++ ){
for( ii=kk=0 ; ii < nxyz ; ii++ ){
if( GOOD(ii) ){ A(kk,jj) = THD_get_voxel(aset,ii,jj); kk++; }
}}
DSET_unload(aset) ;
/**--- set up map weight into wim/war ---**/
#undef WT
#define WT(i,j) war[(i)+(j)*nvox] /* nvox X nparam */
if( wset != NULL ){
int numneg=0 , numpos=0 ;
float fac ;
INFO_message("Loading map weight matrix") ;
wim = mri_new( nvox , nwt , MRI_float ) ; war = MRI_FLOAT_PTR(wim) ;
for( jj=0 ; jj < nwt ; jj++ ){
for( ii=kk=0 ; ii < nxyz ; ii++ ){
if( GOOD(ii) ){
WT(kk,jj) = THD_get_voxel(wset,ii,jj);
if( WT(kk,jj) > 0.0f ){ numpos++; WT(kk,jj) = sqrt(WT(kk,jj)); }
else if( WT(kk,jj) < 0.0f ){ numneg++; WT(kk,jj) = 0.0f; }
kk++;
}
}}
DSET_unload(wset) ;
if( numpos <= nparam )
WARNING_message("Only %d positive weights found in -wtmap!",numpos) ;
if( numneg > 0 )
WARNING_message("%d negative weights found in -wtmap!",numneg) ;
for( jj=0 ; jj < nwt ; jj++ ){
fac = 0.0f ;
for( kk=0 ; kk < nvox ; kk++ ) if( WT(kk,jj) > fac ) fac = WT(kk,jj) ;
if( fac > 0.0f ){
fac = 1.0f / fac ;
for( kk=0 ; kk < nvox ; kk++ ) WT(kk,jj) *= fac ;
}
}
}
/**--- set up data image into yim/yar = Y matrix ---**/
#undef Y
#define Y(i,j) yar[(i)+(j)*nt] /* nt X nvox */
INFO_message("Loading data matrix Y") ;
yim = mri_new( nt , nvox , MRI_float ); yar = MRI_FLOAT_PTR(yim);
for( ii=0 ; ii < nt ; ii++ ){
for( jj=kk=0 ; jj < nxyz ; jj++ ){
if( GOOD(jj) ){ Y(ii,kk) = THD_get_voxel(yset,jj,ii); kk++; }
}}
DSET_unload(yset) ;
/**--- project baseline out of data image = Z matrix ---**/
if( pfim != NULL ){
#undef T
#define T(i,j) tar[(i)+(j)*nt] /* nt X nvox */
INFO_message("Projecting baseline out of Y") ;
qim = mri_matrix_mult( pfim , yim ) ; /* nqbase X nvox */
tim = mri_matrix_mult( fim , qim ) ; /* nt X nvox */
tar = MRI_FLOAT_PTR(tim) ; /* Y projected onto baseline */
for( jj=0 ; jj < nvox ; jj++ )
for( ii=0 ; ii < nt ; ii++ ) Y(ii,jj) -= T(ii,jj) ;
mri_free(tim); mri_free(qim); mri_free(pfim); mri_free(fim);
}
/***** At this point:
matrix A is in aim,
matrix Z is in yim.
Solve for V into vim, using the chosen method *****/
switch( method ){
default: ERROR_exit("Illegal method code! WTF?") ; /* Huh? */
/*.....................................................................*/
case METHOD_C:
/**--- compute pseudo-inverse of A map ---**/
INFO_message("Method C: Computing pseudo-inverse of A") ;
if( wim != NULL ) WARNING_message("Ignoring -mapwt dataset") ;
pfim = mri_matrix_psinv(aim,NULL,alpha) ; /* nparam X nvox */
if( pfim == NULL ) ERROR_exit("mri_matrix_psinv() fails") ;
mri_free(aim) ;
/**--- and apply to data to get results ---*/
INFO_message("Computing result V") ;
vim = mri_matrix_multranB( yim , pfim ) ; /* nt x nparam */
mri_free(pfim) ; mri_free(yim) ;
break ;
/*.....................................................................*/
case METHOD_K:
/**--- compute pseudo-inverse of transposed Z ---*/
INFO_message("Method K: Computing pseudo-inverse of Z'") ;
if( nwt > 1 ){
WARNING_message("Ignoring -mapwt dataset: more than 1 sub-brick") ;
nwt = 0 ; mri_free(wim) ; wim = NULL ; war = NULL ;
}
if( nwt == 1 ){
float fac ;
for( kk=0 ; kk < nvox ; kk++ ){
fac = war[kk] ;
for( ii=0 ; ii < nt ; ii++ ) Y(ii,kk) *= fac ;
for( ii=0 ; ii < nparam ; ii++ ) A(kk,ii) *= fac ;
}
}
tim = mri_matrix_transpose(yim) ; mri_free(yim) ;
pfim = mri_matrix_psinv(tim,NULL,alpha) ; mri_free(tim) ;
if( pfim == NULL ) ERROR_exit("mri_matrix_psinv() fails") ;
INFO_message("Computing W") ;
tim = mri_matrix_mult( pfim , aim ) ;
mri_free(aim) ; mri_free(pfim) ;
INFO_message("Computing result V") ;
pfim = mri_matrix_psinv(tim,NULL,0.0f) ; mri_free(tim) ;
vim = mri_matrix_transpose(pfim) ; mri_free(pfim);
break ;
} /* end of switch on method */
if( wim != NULL ) mri_free(wim) ;
/**--- smooth? ---**/
if( nfir > 0 && vim->nx > nfir ){
INFO_message("FIR-5-ing result") ;
var = MRI_FLOAT_PTR(vim) ;
for( jj=0 ; jj < vim->ny ; jj++ )
linear_filter_reflect( nfir,firwt , vim->nx , var + (jj*vim->nx) ) ;
}
if( nmed > 0 && vim->nx > nmed ){
INFO_message("Median-5-ing result") ;
var = MRI_FLOAT_PTR(vim) ;
for( jj=0 ; jj < vim->ny ; jj++ )
median5_filter_reflect( vim->nx , var + (jj*vim->nx) ) ;
}
/**--- write results ---**/
INFO_message("Writing result to '%s'",fname_out) ;
mri_write_1D( fname_out , vim ) ;
exit(0) ;
}
int THD_bandpass_vectors( int nlen , int nvec , float **vec ,
float dt , float fbot , float ftop ,
int qdet , int nort , float **ort )
{
int nfft,nby2 , iv, jbot,jtop , ndof=0 ; register int jj ;
float df , tapr ;
register float *xar, *yar=NULL ;
register complex *zar ; complex Zero={0.0f,0.0f} ;
ENTRY("THD_bandpass_vectors") ;
if( ftop > ICOR_MAX_FTOP && qdet < 0 && (nort <= 0 || ort == NULL) )
RETURN(ndof) ; /* 26 Feb 2010: do nothing at all? */
if( nlen < 9 || nvec < 1 || vec == NULL ){
ERROR_message("bad bandpass data?");
RETURN(ndof);
}
if( bpwrn && dt > 60.0f ) {
WARNING_message("Your bandpass timestep (%f) is high.\n"
" Make sure units are 'sec', not 'msec'.\n"
" This warning will not be repeated." ,
dt);
bpwrn = 0;
}
if( dt <= 0.0f ) dt = 1.0f ;
if( fbot < 0.0f ) fbot = 0.0f ;
if( ftop <= fbot ){
ERROR_message("Bad bandpass frequencies?"); RETURN(ndof);
}
if( nort >= nlen ){
ERROR_message("Too many bandpass orts?") ; RETURN(ndof);
}
/** setup for FFT **/
nfft = (nfft_fixed >= nlen) ? nfft_fixed : csfft_nextup_even(nlen) ;
nby2 = nfft/2 ;
df = 1.0f / (nfft * dt) ; /* frequency resolution */
jbot = (int)rint(fbot/df) ; /* closest freq index to fbot */
jtop = (int)rint(ftop/df) ; /* and to ftop */
if( jtop >= nby2 ) jtop = nby2-1 ; /* can't go past Nyquist! */
if( jbot >= jtop+1 ){
ERROR_message("bandpass: fbot and ftop too close ==> "
"jbot=%d jtop=%d (df=%f)",
jbot,jtop, df) ;
RETURN(ndof) ;
}
/** quadratic detrending first? (should normally be used) **/
switch( qdet ){
case 2:
ndof += 2 ;
for( iv=0 ; iv < nvec ; iv++ )
THD_quadratic_detrend( nlen, vec[iv], NULL,NULL,NULL ) ;
break ;
case 1:
ndof += 1 ;
for( iv=0 ; iv < nvec ; iv++ )
THD_linear_detrend( nlen, vec[iv], NULL,NULL ) ;
break ;
case 0:
for( iv=0 ; iv < nvec ; iv++ )
THD_const_detrend( nlen, vec[iv], NULL ) ;
break ;
}
if( qdet > 0 || jbot > 0 || jtop < nby2-1 ){ /* 26 Feb 2010: do the FFTs */
zar = (complex *)malloc(sizeof(complex)*nfft) ; /* work array */
csfft_scale_inverse(1) ; /* scale inverse FFT by 1/nfft */
/** loop over vectors in pairs, FFT-ing and bandpassing **/
ndof += 2 ; /* for 0 and Nyquist freqs */
if( jbot >= 1 ) ndof += 2*jbot - 1 ; /* DOF for low freq */
ndof += 2*(nby2-jtop) - 1 ; /* DOF for high freq */
for( iv=0 ; iv < nvec ; iv+=2 ){
/* load a pair of vectors into zar to double up on FFTs of real data */
xar = vec[iv] ;
if( iv == nvec-1 ){ /* last one has nothing to pair with */
for( jj=0 ; jj < nlen ; jj++ ){ zar[jj].r = xar[jj] ; zar[jj].i = 0.0f ; }
} else {
yar = vec[iv+1] ;
for( jj=0 ; jj < nlen ; jj++ ){ zar[jj].r = xar[jj] ; zar[jj].i = yar[jj] ; }
}
for( jj=nlen ; jj < nfft ; jj++ ) zar[jj] = Zero ; /* zero fill */
csfft_cox( -1 , nfft , zar ) ; /*** the FFT ***/
/* delete unwanted frequencies */
zar[0] = zar[nby2] = Zero ;
tapr = (nort > 0 && ort != NULL) ? 0.05f : 0.5f ;
if( jbot >= 1 ){
zar[jbot].r *= tapr ; zar[jbot].i *= tapr ;
zar[nfft-jbot].r *= tapr ; zar[nfft-jbot].i *= tapr ;
for( jj=1 ; jj < jbot ; jj++ ) zar[jj] = zar[nfft-jj] = Zero ;
}
zar[jtop].r *= tapr ; zar[jtop].i *= tapr ;
zar[nfft-jtop].r *= tapr ; zar[nfft-jtop].i *= tapr ;
for( jj=jtop+1 ; jj < nby2 ; jj++ ) zar[jj] = zar[nfft-jj] = Zero ;
csfft_cox( 1 , nfft , zar ) ; /*** inverse FFT ***/
/* unload vector pair back into original data arrays */
if( iv == nvec-1 ){
for( jj=0 ; jj < nlen ; jj++ ) xar[jj] = zar[jj].r ;
} else {
for( jj=0 ; jj < nlen ; jj++ ){ xar[jj] = zar[jj].r ; yar[jj] = zar[jj].i ; }
}
} /* end of loop over vector pairs: bandpassing now done */
free(zar) ; csfft_scale_inverse(0) ;
} /* end of FFTs */
/*** remove orts? ***/
if( nort > 0 && ort != NULL ){
float **qort = (float **)malloc(sizeof(float *)*nort) ;
MRI_IMAGE *qim , *pim ; float *par, *qar , *rar , *pt,*qt ;
register float sum , xt ; register int kk ;
/* must bandpass copy of orts first -- via recursion
(so we don't re-introduce any of the removed frequencies) */
qim = mri_new( nlen , nort , MRI_float ) ; /** [Q] = nlen X nort **/
qar = MRI_FLOAT_PTR(qim) ; /** load with the orts **/
for( iv=0 ; iv < nort ; iv++ ){
qort[iv] = qar + iv*nlen ;
memcpy( qort[iv] , ort[iv] , sizeof(float)*nlen ) ;
}
(void)THD_bandpass_vectors( nlen, nort, qort, dt, fbot, ftop, qdet, 0,NULL ) ;
free(qort) ;
/* compute pseudo-inverse ([P] = inv{[Q]'[Q]}[Q]') of bandpassed orts */
pim = mri_matrix_psinv( qim , NULL , 1.e-8 ) ; /** [P] = nort X nlen **/
if( pim == NULL ){ /* should not happen */
mri_free(qim) ;
ERROR_message("can't remove bandpass orts?") ;
RETURN(ndof) ;
}
par = MRI_FLOAT_PTR(pim) ; /* nort X nlen matrix */
/* Project the bandpassed orts out of data vectors:
for each vector [y], replace it with [y] - [Q][P][y] ;
this is more efficient than computing the nlen X nlen projection
matrix [I]-[Q][P] and then applying to each vector, since
that would require about nlen*nlen flops per vector, whereas
the 2 matrix multiplications by [P] then [Q] will require
about 2*nlen*nort flops per vector, a clear win if nort < nlen/2. */
rar = (float *)malloc(sizeof(float)*nort) ; /* will hold [P][y] */
for( iv=0 ; iv < nvec ; iv++ ){
xar = vec[iv] ; /* [y] vector */
/* compute nort-vector [r] = [P][y] */
for( jj=0 ; jj < nort ; jj++ ) rar[jj] = 0.0f ; /* initialize to 0 */
for( kk=0 ; kk < nlen ; kk++ ){
pt = par + kk*nort ; xt = xar[kk] ;
for( jj=0 ; jj < nort ; jj++ ) rar[jj] += pt[jj]*xt ;
}
/* now subtract [Q][r] from [y] */
for( jj=0 ; jj < nort ; jj++ ){
qt = qar + jj*nlen ; xt = rar[jj] ;
for( kk=0 ; kk < nlen ; kk++ ) xar[kk] -= qt[kk]*xt ;
}
}
free(rar) ; mri_free(pim) ; mri_free(qim) ;
ndof += nort ;
} /* de-ortification is done */
/** done **/
if( nfft > nlen ){
double fac = ((double)nlen)/(double)nfft ;
ndof = (int)rint(fac*ndof) ;
}
RETURN(ndof) ;
}
