void run() {
Image::Header input_SH_header (argument[0]);
if (input_SH_header.ndim() != 4)
throw Exception ("input SH image should contain 4 dimensions");
std::vector<ssize_t> strides (4, 0);
strides[3] = 1;
Image::BufferPreload<value_type> input_buf (input_SH_header, strides);
Math::Vector<value_type> responseSH;
responseSH.load (argument[1]);
Math::Vector<value_type> responseRH;
Math::SH::SH2RH (responseRH, responseSH);
Ptr<Image::Buffer<bool> > mask_buf;
Options opt = get_options ("mask");
if (opt.size()) {
mask_buf = new Image::Buffer<bool> (opt[0][0]);
Image::check_dimensions (*mask_buf, input_buf, 0, 3);
}
Image::Header output_SH_header (input_SH_header);
Image::Stride::set_from_command_line (output_SH_header, Image::Stride::contiguous_along_axis (3));
Image::Buffer<value_type> output_SH_buf (argument[2], output_SH_header);
SDeconvFunctor sconv (input_buf, output_SH_buf, mask_buf, responseRH);
Image::ThreadedLoop loop ("performing convolution...", input_buf, 2, 0, 3);
loop.run (sconv);
}
void run()
{
InputBufferType dwi_buffer (argument[0], Image::Stride::contiguous_along_axis (3));
Math::Matrix<cost_value_type> grad = DWI::get_valid_DW_scheme<cost_value_type> (dwi_buffer);
size_t dwi_axis = 3;
while (dwi_buffer.dim (dwi_axis) < 2) ++dwi_axis;
INFO ("assuming DW images are stored along axis " + str (dwi_axis));
Math::Matrix<cost_value_type> bmatrix;
DWI::grad2bmatrix (bmatrix, grad);
Math::Matrix<cost_value_type> binv (bmatrix.columns(), bmatrix.rows());
Math::pinv (binv, bmatrix);
int method = 1;
Options opt = get_options ("method");
if (opt.size()) method = opt[0][0];
opt = get_options ("regularisation");
cost_value_type regularisation = 5000.0;
if (opt.size()) regularisation = opt[0][0];
opt = get_options ("mask");
Ptr<MaskBufferType> mask_buffer;
Ptr<MaskBufferType::voxel_type> mask_vox;
if (opt.size()){
mask_buffer = new MaskBufferType (opt[0][0]);
Image::check_dimensions (*mask_buffer, dwi_buffer, 0, 3);
mask_vox = new MaskBufferType::voxel_type (*mask_buffer);
}
Image::Header dt_header (dwi_buffer);
dt_header.set_ndim (4);
dt_header.dim (3) = 6;
dt_header.datatype() = DataType::Float32;
dt_header.DW_scheme() = grad;
OutputBufferType dt_buffer (argument[1], dt_header);
InputBufferType::voxel_type dwi_vox (dwi_buffer);
OutputBufferType::voxel_type dt_vox (dt_buffer);
Image::ThreadedLoop loop ("estimating tensor components...", dwi_vox, 1, 0, 3);
Processor processor (dwi_vox, dt_vox, mask_vox, bmatrix, binv, method, regularisation, loop.inner_axes()[0], dwi_axis);
loop.run_outer (processor);
}
void run ()
{
Image::Header H (argument[0]);
Image::Info info (H);
info.set_ndim (3);
Image::BufferScratch<bool> mask (info);
auto v_mask = mask.voxel();
std::string mask_path;
Options opt = get_options ("mask");
if (opt.size()) {
mask_path = std::string(opt[0][0]);
Image::Buffer<bool> in (mask_path);
if (!Image::dimensions_match (H, in, 0, 3))
throw Exception ("Input mask image does not match DWI");
if (!(in.ndim() == 3 || (in.ndim() == 4 && in.dim(3) == 1)))
throw Exception ("Input mask image must be a 3D image");
auto v_in = in.voxel();
Image::copy (v_in, v_mask, 0, 3);
} else {
for (auto l = Image::LoopInOrder (v_mask) (v_mask); l; ++l)
v_mask.value() = true;
}
DWI::CSDeconv<float>::Shared shared (H);
const size_t lmax = DWI::lmax_for_directions (shared.DW_dirs);
if (lmax < 4)
throw Exception ("Cannot run dwi2response with lmax less than 4");
shared.lmax = lmax;
Image::BufferPreload<float> dwi (H, Image::Stride::contiguous_along_axis (3));
DWI::Directions::Set directions (1281);
Math::Vector<float> response (lmax/2+1);
response.zero();
{
// Initialise response function
// Use lmax = 2, get the DWI intensity mean and standard deviation within the mask and
// use these as the first two coefficients
auto v_dwi = dwi.voxel();
double sum = 0.0, sq_sum = 0.0;
size_t count = 0;
Image::LoopInOrder loop (dwi, "initialising response function... ", 0, 3);
for (auto l = loop (v_dwi, v_mask); l; ++l) {
if (v_mask.value()) {
for (size_t volume_index = 0; volume_index != shared.dwis.size(); ++volume_index) {
v_dwi[3] = shared.dwis[volume_index];
const float value = v_dwi.value();
sum += value;
sq_sum += Math::pow2 (value);
++count;
}
}
}
response[0] = sum / double (count);
response[1] = - 0.5 * std::sqrt ((sq_sum / double(count)) - Math::pow2 (response[0]));
// Account for scaling in SH basis
response *= std::sqrt (4.0 * Math::pi);
}
INFO ("Initial response function is [" + str(response, 2) + "]");
// Algorithm termination options
opt = get_options ("max_iters");
const size_t max_iters = opt.size() ? int(opt[0][0]) : DWI2RESPONSE_DEFAULT_MAX_ITERS;
opt = get_options ("max_change");
const float max_change = 0.01 * (opt.size() ? float(opt[0][0]) : DWI2RESPONSE_DEFAULT_MAX_CHANGE);
// Should all voxels (potentially within a user-specified mask) be tested at every iteration?
opt = get_options ("test_all");
const bool reset_mask = opt.size();
// Single-fibre voxel selection options
opt = get_options ("volume_ratio");
const float volume_ratio = opt.size() ? float(opt[0][0]) : DWI2RESPONSE_DEFAULT_VOLUME_RATIO;
opt = get_options ("dispersion_multiplier");
const float dispersion_multiplier = opt.size() ? float(opt[0][0]) : DWI2RESPONSE_DEFAULT_DISPERSION_MULTIPLIER;
opt = get_options ("integral_multiplier");
const float integral_multiplier = opt.size() ? float(opt[0][0]) : DWI2RESPONSE_DEFAULT_INTEGRAL_STDEV_MULTIPLIER;
SFThresholds thresholds (volume_ratio); // Only threshold the lobe volume ratio for now; other two are not yet used
size_t total_iter = 0;
bool first_pass = true;
size_t prev_sf_count = 0;
{
bool iterate = true;
size_t iter = 0;
ProgressBar progress ("optimising response function... ");
do {
++iter;
{
MR::LogLevelLatch latch (0);
shared.set_response (response);
shared.init();
}
++progress;
if (reset_mask) {
if (mask_path.size()) {
Image::Buffer<bool> in (mask_path);
auto v_in = in.voxel();
Image::copy (v_in, v_mask, 0, 3);
} else {
for (auto l = Image::LoopInOrder(v_mask) (v_mask); l; ++l)
v_mask.value() = true;
}
++progress;
}
std::vector<FODSegResult> seg_results;
{
FODCalcAndSeg processor (dwi, mask, shared, directions, lmax, seg_results);
Image::ThreadedLoop loop (mask, 0, 3);
loop.run (processor);
}
++progress;
if (!first_pass)
thresholds.update (seg_results, dispersion_multiplier, integral_multiplier, iter);
++progress;
Response output (lmax);
mask.zero();
{
SFSelector selector (seg_results, thresholds, mask);
ResponseEstimator estimator (dwi, shared, lmax, output);
Thread::run_queue (selector, FODSegResult(), Thread::multi (estimator));
}
if (!output.get_count())
throw Exception ("Cannot estimate response function; all voxels have been excluded from selection");
const Math::Vector<float> new_response = output.result();
const size_t sf_count = output.get_count();
++progress;
if (App::log_level >= 2)
std::cerr << "\n";
INFO ("Iteration " + str(iter) + ", " + str(sf_count) + " SF voxels, new response function: [" + str(new_response, 2) + "]");
if (sf_count == prev_sf_count) {
INFO ("terminating due to convergence of single-fibre voxel selection");
iterate = false;
}
if (iter == max_iters) {
INFO ("terminating due to completing maximum number of iterations");
iterate = false;
}
bool rf_changed = false;
for (size_t i = 0; i != response.size(); ++i) {
if (std::abs ((new_response[i] - response[i]) / new_response[i]) > max_change)
rf_changed = true;
}
if (!rf_changed) {
INFO ("terminating due to negligible changes in the response function coefficients");
iterate = false;
}
if (!iterate && first_pass) {
iterate = true;
first_pass = false;
INFO ("commencing second-pass of response function estimation");
total_iter = iter;
iter = 0;
}
response = new_response;
prev_sf_count = sf_count;
//v_mask.save ("mask_pass_" + str(first_pass?1:2) + "_iter_" + str(iter) + ".mif");
} while (iterate);
total_iter += iter;
}
CONSOLE ("final response function: [" + str(response, 2) + "] (reached after " + str(total_iter) + " iterations using " + str(prev_sf_count) + " voxels)");
response.save (argument[1]);
opt = get_options ("sf");
if (opt.size())
v_mask.save (std::string (opt[0][0]));
}