def tamarack(imgs : pd.DataFrame, options): # columns of the input df: `img` : MincAtom, `timept` : number, ... # columns of the pride of models : 'timept' : number, 'model' : MincAtom s = Stages() # TODO some assertions that the pride_of_models, if provided, is correct, and that this is intended target type def group_options(options, timepoint): options = copy.deepcopy(options) if options.mbm.lsq6.target_type == TargetType.pride_of_models: options = copy.deepcopy(options) targets = get_closest_model_from_pride_of_models(pride_of_models_dict=get_pride_of_models_mapping( pride_csv=options.mbm.lsq6.target_file, output_dir=options.application.output_directory, pipeline_name=options.application.pipeline_name), time_point=timepoint) options.mbm.lsq6 = options.mbm.lsq6.replace(target_type=TargetType.initial_model, target_file=targets.registration_standard.path) # resolution = (options.registration.resolution # or get_resolution_from_file(targets.registration_standard.path)) # options.registration = options.registration.replace(resolution=resolution) # FIXME use of registration_standard here is quite wrong ... # part of the trouble is that mbm calls registration_targets itself, # so we can't send this RegistrationTargets to `mbm` directly ... # one option: add yet another optional arg to `mbm` ... else: targets = s.defer(registration_targets(lsq6_conf=options.mbm.lsq6, app_conf=options.application, reg_conf=options.registration, first_input_file=imgs.filename.iloc[0])) resolution = (options.registration.resolution or get_resolution_from_file(targets.registration_standard.path)) # This must happen after calling registration_targets otherwise it will resample to options.registration.resolution options.registration = options.registration.replace(resolution=resolution) return options # build all first-level models: first_level_results = ( imgs # TODO 'group' => 'timept' ? .groupby('group', as_index=False) # the usual annoying pattern to do an aggregate with access .aggregate({ 'file' : lambda files: list(files) }) # to the groupby object's keys ... TODO: fix .rename(columns={ 'file' : "files" }) .assign(options=lambda df: df.apply(axis=1, func=lambda row: group_options(options, row.group))) .assign(build_model=lambda df: df.apply(axis=1, func=lambda row: s.defer( mbm(imgs=row.files, options=row.options, prefix="%s" % row.group, output_dir=os.path.join( options.application.output_directory, options.application.pipeline_name + "_first_level", "%s_processed" % row.group))))) .sort_values(by='group') ) if all(first_level_results.options.map(lambda opts: opts.registration.resolution) == first_level_results.options.iloc[0].registration.resolution): options.registration = options.registration.replace( resolution=first_level_results.options.iloc[0].registration.resolution) else: raise ValueError("some first-level models are run at different resolutions, possibly not what you want ...") # construction of the overall inter-average transforms will be done iteratively (for efficiency/aesthetics), # which doesn't really fit the DataFrame mold ... full_hierarchy = get_nonlinear_configuration_from_options( nlin_protocol=options.mbm.nlin.nlin_protocol, reg_method=options.mbm.nlin.reg_method, file_resolution=options.registration.resolution) # FIXME no good can come of this nlin_protocol = full_hierarchy.confs[-1] if isinstance(full_hierarchy, MultilevelANTSConf) else full_hierarchy # first register consecutive averages together: average_registrations = ( first_level_results[:-1] .assign(next_model=list(first_level_results[1:].build_model)) # TODO: we should be able to do lsq6 registration here as well! .assign(xfm=lambda df: df.apply(axis=1, func=lambda row: s.defer( lsq12_nlin(source=row.build_model.avg_img, target=row.next_model.avg_img, lsq12_conf=get_linear_configuration_from_options( options.mbm.lsq12, transform_type=LinearTransType.lsq12, file_resolution=options.registration.resolution), nlin_conf=nlin_protocol))))) # now compose the above transforms to produce transforms from each average to the common average: common_time_pt = options.tamarack.common_time_pt common_model = first_level_results[first_level_results.group == common_time_pt].iloc[0].build_model.avg_img #common = average_registrations[average_registrations.group == common_time_pt].iloc[0] before = average_registrations[average_registrations.group < common_time_pt] # asymmetry in before/after since after = average_registrations[average_registrations.group >= common_time_pt] # we used `next_`, not `previous_` # compose 1st and 2nd level transforms and resample into the common average space: def suffixes(xs): if len(xs) == 0: return [[]] else: ys = suffixes(xs[1:]) return [[xs[0]] + ys[0]] + ys def prefixes(xs): if len(xs) == 0: return [[]] else: ys = prefixes(xs[1:]) return ys + [ys[-1] + [xs[0]]] xfms_to_common = ( first_level_results .assign(uncomposed_xfms=suffixes(list(before.xfm))[:-1] + [None] + prefixes(list(after.xfm))[1:]) .assign(xfm_to_common=lambda df: df.apply(axis=1, func=lambda row: ((lambda x: s.defer(invert_xfmhandler(x)) if row.group >= common_time_pt else x) (s.defer(concat_xfmhandlers(row.uncomposed_xfms, name=("%s_to_common" if row.group < common_time_pt else "%s_from_common") % row.group)))) if row.uncomposed_xfms is not None else None)) .drop('uncomposed_xfms', axis=1)) # TODO None => identity?? # TODO indexing here is not good ... first_level_determinants = pd.concat(list(first_level_results.build_model.apply( lambda x: x.determinants.assign(first_level_avg=x.avg_img))), ignore_index=True) resampled_determinants = ( pd.merge(left=first_level_determinants, right=xfms_to_common.assign(source=lambda df: df.xfm_to_common.apply( lambda x: x.source if x is not None else None)), left_on="first_level_avg", right_on='source') .assign(resampled_log_full_det=lambda df: df.apply(axis=1, func=lambda row: s.defer(mincresample_new(img=row.log_full_det, xfm=row.xfm_to_common.xfm, like=common_model)) if row.xfm_to_common is not None else row.img), resampled_log_nlin_det=lambda df: df.apply(axis=1, func=lambda row: s.defer(mincresample_new(img=row.log_nlin_det, xfm=row.xfm_to_common.xfm, like=common_model)) if row.xfm_to_common is not None else row.img)) ) inverted_overall_xfms = pd.Series({ xfm : (s.defer(concat_xfmhandlers([xfm, row.xfm_to_common])) if row.xfm_to_common is not None else xfm) for _ix, row in xfms_to_common.iterrows() for xfm in row.build_model.xfms.lsq12_nlin_xfm }) overall_xfms = inverted_overall_xfms.apply(lambda x: s.defer(invert_xfmhandler(x))) overall_determinants = determinants_at_fwhms(xfms=overall_xfms, blur_fwhms=options.mbm.stats.stats_kernels, inv_xfms=inverted_overall_xfms) # TODO turn off bootstrap as with two-level code? # TODO combine into one data frame return Result(stages=s, output=Namespace(first_level_results=first_level_results, overall_determinants=overall_determinants, resampled_determinants=resampled_determinants.drop( ['options'], axis=1)))
def tamarack(imgs: pd.DataFrame, options): # columns of the input df: `img` : MincAtom, `timept` : number, ... # columns of the pride of models : 'timept' : number, 'model' : MincAtom s = Stages() # TODO some assertions that the pride_of_models, if provided, is correct, and that this is intended target type def group_options(options, timepoint): options = copy.deepcopy(options) if options.mbm.lsq6.target_type == TargetType.pride_of_models: options = copy.deepcopy(options) targets = get_closest_model_from_pride_of_models( pride_of_models_dict=get_pride_of_models_mapping( pride_csv=options.mbm.lsq6.target_file, output_dir=options.application.output_directory, pipeline_name=options.application.pipeline_name), time_point=timepoint) options.mbm.lsq6 = options.mbm.lsq6.replace( target_type=TargetType.initial_model, target_file=targets.registration_standard.path) # resolution = (options.registration.resolution # or get_resolution_from_file(targets.registration_standard.path)) # options.registration = options.registration.replace(resolution=resolution) # FIXME use of registration_standard here is quite wrong ... # part of the trouble is that mbm calls registration_targets itself, # so we can't send this RegistrationTargets to `mbm` directly ... # one option: add yet another optional arg to `mbm` ... else: targets = s.defer( registration_targets(lsq6_conf=options.mbm.lsq6, app_conf=options.application, reg_conf=options.registration, first_input_file=imgs.filename.iloc[0])) resolution = (options.registration.resolution or get_resolution_from_file( targets.registration_standard.path)) # This must happen after calling registration_targets otherwise it will resample to options.registration.resolution options.registration = options.registration.replace( resolution=resolution) return options # build all first-level models: first_level_results = ( imgs # TODO 'group' => 'timept' ? .groupby('group', as_index=False ) # the usual annoying pattern to do an aggregate with access .aggregate({'file': lambda files: list(files)} ) # to the groupby object's keys ... TODO: fix .rename(columns={ 'file': "files" }).assign(options=lambda df: df.apply( axis=1, func=lambda row: group_options(options, row.group)) ).assign(build_model=lambda df: df.apply( axis=1, func=lambda row: s.defer( mbm(imgs=row.files, options=row.options, prefix="%s" % row.group, output_dir=os.path.join( options.application.output_directory, options .application.pipeline_name + "_first_level", "%s_processed" % row.group)))) ).sort_values(by='group')) if all( first_level_results.options.map( lambda opts: opts.registration.resolution) == first_level_results.options.iloc[0].registration.resolution): options.registration = options.registration.replace( resolution=first_level_results.options.iloc[0].registration. resolution) else: raise ValueError( "some first-level models are run at different resolutions, possibly not what you want ..." ) # construction of the overall inter-average transforms will be done iteratively (for efficiency/aesthetics), # which doesn't really fit the DataFrame mold ... full_hierarchy = get_nonlinear_configuration_from_options( nlin_protocol=options.mbm.nlin.nlin_protocol, reg_method=options.mbm.nlin.reg_method, file_resolution=options.registration.resolution) # FIXME no good can come of this nlin_protocol = full_hierarchy.confs[-1] if isinstance( full_hierarchy, MultilevelANTSConf) else full_hierarchy # first register consecutive averages together: average_registrations = ( first_level_results[:-1].assign( next_model=list(first_level_results[1:].build_model)) # TODO: we should be able to do lsq6 registration here as well! .assign(xfm=lambda df: df.apply( axis=1, func=lambda row: s.defer( lsq12_nlin(source=row.build_model.avg_img, target=row.next_model.avg_img, lsq12_conf=get_linear_configuration_from_options( options.mbm.lsq12, transform_type=LinearTransType.lsq12, file_resolution=options.registration.resolution ), nlin_conf=nlin_protocol))))) # now compose the above transforms to produce transforms from each average to the common average: common_time_pt = options.tamarack.common_time_pt common_model = first_level_results[ first_level_results.group == common_time_pt].iloc[0].build_model.avg_img #common = average_registrations[average_registrations.group == common_time_pt].iloc[0] before = average_registrations[ average_registrations.group < common_time_pt] # asymmetry in before/after since after = average_registrations[ average_registrations.group >= common_time_pt] # we used `next_`, not `previous_` # compose 1st and 2nd level transforms and resample into the common average space: def suffixes(xs): if len(xs) == 0: return [[]] else: ys = suffixes(xs[1:]) return [[xs[0]] + ys[0]] + ys def prefixes(xs): if len(xs) == 0: return [[]] else: ys = prefixes(xs[1:]) return ys + [ys[-1] + [xs[0]]] xfms_to_common = (first_level_results.assign( uncomposed_xfms=suffixes(list(before.xfm))[:-1] + [None] + prefixes(list(after.xfm))[1:]).assign( xfm_to_common=lambda df: df.apply( axis=1, func=lambda row: ((lambda x: s.defer(invert_xfmhandler( x)) if row.group >= common_time_pt else x)(s.defer( concat_xfmhandlers( row.uncomposed_xfms, name=("%s_to_common" if row.group < common_time_pt else "%s_from_common") % row.group)))) if row.uncomposed_xfms is not None else None)).drop( 'uncomposed_xfms', axis=1)) # TODO None => identity?? # TODO indexing here is not good ... first_level_determinants = pd.concat(list( first_level_results.build_model.apply( lambda x: x.determinants.assign(first_level_avg=x.avg_img))), ignore_index=True) resampled_determinants = (pd.merge( left=first_level_determinants, right=xfms_to_common.assign(source=lambda df: df.xfm_to_common.apply( lambda x: x.source if x is not None else None)), left_on="first_level_avg", right_on='source').assign( resampled_log_full_det=lambda df: df.apply( axis=1, func=lambda row: s.defer( mincresample_new(img=row.log_full_det, xfm=row.xfm_to_common.xfm, like=common_model)) if row.xfm_to_common is not None else row.img), resampled_log_nlin_det=lambda df: df.apply( axis=1, func=lambda row: s.defer( mincresample_new(img=row.log_nlin_det, xfm=row.xfm_to_common.xfm, like=common_model)) if row.xfm_to_common is not None else row.img))) inverted_overall_xfms = pd.Series({ xfm: (s.defer(concat_xfmhandlers([xfm, row.xfm_to_common])) if row.xfm_to_common is not None else xfm) for _ix, row in xfms_to_common.iterrows() for xfm in row.build_model.xfms.lsq12_nlin_xfm }) overall_xfms = inverted_overall_xfms.apply( lambda x: s.defer(invert_xfmhandler(x))) overall_determinants = determinants_at_fwhms( xfms=overall_xfms, blur_fwhms=options.mbm.stats.stats_kernels, inv_xfms=inverted_overall_xfms) # TODO turn off bootstrap as with two-level code? # TODO combine into one data frame return Result(stages=s, output=Namespace( first_level_results=first_level_results, overall_determinants=overall_determinants, resampled_determinants=resampled_determinants.drop( ['options'], axis=1)))
def maget(imgs: List[MincAtom], options, prefix, output_dir, build_model_xfms=None): # FIXME prefix, output_dir aren't used !! s = Stages() maget_options = options.maget.maget resolution = options.registration.resolution # TODO or get_resolution_from_file(...) -- only if file always exists! pipeline_sub_dir = os.path.join( options.application.output_directory, options.application.pipeline_name + "_atlases") atlases = get_atlases(maget_options, pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options( options.maget.lsq12, transform_type=LinearTransType.lsq12, file_resolution=resolution) nlin_component = get_nonlinear_component(options.maget.nlin.reg_method) # TODO should this be here or outside `maget` call? #imgs = [s.defer(nlin_component.ToMinc.from_mnc(img)) for img in imgs] #nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.nlin.nlin_protocol, # next(iter(options.maget.nlin.flags_.nlin_protocol)), # reg_method=options.maget.nlin.reg_method, # file_resolution=resolution) if maget_options.mask or maget_options.mask_only: # this used to return alignments but doesn't currently do so masked_img = s.defer( maget_mask( imgs=imgs, maget_options=options.maget, atlases=atlases, pipeline_sub_dir=pipeline_sub_dir + "_masking", # FIXME repeats all alignments!!! resolution=resolution)) # now propagate only the masked form of the images and atlases: imgs = masked_img #atlases = masked_atlases # TODO is this needed? if maget_options.mask_only: # register each input to each atlas, creating a mask return Result( stages=s, output=masked_img) # TODO rename `alignments` to `registrations`?? else: if maget_options.mask: del masked_img # this `del` is just to verify that we don't accidentally use this later, since these potentially # coarser alignments shouldn't be re-used (but if the protocols for masking and alignment are the same, # hash-consing will take care of things), just the masked images they create; can be removed later # if a sensible use is found # images with labels from atlases # N.B.: Even though we've already registered each image to each initial atlas, this happens again here, # but using `nlin_hierarchy` instead of `masking_nlin_hierarchy` as options. # This is not 'work-efficient' in the sense that this computation happens twice (although # hopefully at greater precision the second time!), but the idea is to run a coarse initial # registration to get a mask and then do a better registration with that mask (though I'm not # sure exactly when this is faster than doing a single registration). # This _can_ allow the overall computation to finish more rapidly # (depending on the relative speed of the two alignment methods/parameters, # number of atlases and other templates used, number of cores available, etc.). atlas_labelled_imgs = ( pd.DataFrame({ 'img': img, 'label_file': s.defer( # can't use `label` in a pd.DataFrame index! mincresample_new( img=atlas.labels, xfm=s.defer( lsq12_nlin( source=img, target=atlas, nlin_module=nlin_component, lsq12_conf=lsq12_conf, nlin_options=options.maget.nlin.nlin_protocol, resolution=resolution, #nlin_conf=nlin_hierarchy, resample_source=False)).xfm, like=img, invert=True, interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range', '-labels'))) } for img in imgs for atlas in atlases)) if maget_options.pairwise: def choose_new_templates(ts, n): # currently silly, but we might implement a smarter method ... # FIXME what if there aren't enough other imgs around?! This silently goes weird ... return pd.Series( ts[:n + 1] ) # n+1 instead of n: choose one more since we won't use image as its own template ... # FIXME: the --max-templates flag is ambiguously named ... should be --max-new-templates # (and just use all atlases) # TODO we could have a separate templates_csv (or --template-files f [f ...]) but you can just # run a separate MAGeT pipeline and #if maget_options.templates_csv: # templates = pd.read_csv(maget_options.templates_csv).template #else: templates = pd.DataFrame({ 'template': choose_new_templates(ts=imgs, n=maget_options.max_templates - len(atlases)) }) # note these images are the masked ones if masking was done ... # the templates together with their (multiple) labels from the atlases (this merge just acts as a filter) labelled_templates = pd.merge(left=atlas_labelled_imgs, right=templates, left_on="img", right_on="template").drop('img', axis=1) # images with new labels from the templates imgs_and_templates = pd.merge( #left=atlas_labelled_imgs, left=pd.DataFrame({ "img": imgs }).assign(fake=1), right=labelled_templates.assign(fake=1), on='fake') #left_on='img', right_on='template') # TODO do select here instead of below? #if build_model_xfms is not None: # # use path instead of full mincatom as key in case we're reading these in from a CSV: # xfm_dict = { x.source.path : x.xfm for x in build_model_xfms } template_labelled_imgs = ( imgs_and_templates.rename( columns={'label_file': 'template_label_file'}) # don't register template to itself, since otherwise atlases would vote on that template twice .loc[lambda df: df.index.map(lambda ix: df.img[ix].path != df. template[ix].path)]. assign(label_file=lambda df: df.apply( axis=1, func=lambda row: s.defer( # TODO switch to uses of nlin_component.whatever(...) in several places below? mincresample_new( #nlin_component.Algorithms.resample( img=row.template_label_file, xfm=s.defer( lsq12_nlin( source=row.img, target=row.template, lsq12_conf=lsq12_conf, resolution=resolution, nlin_module=nlin_component, nlin_options=options.maget.nlin. nlin_protocol, #nlin_conf=nlin_hierarchy, resample_source=False)).xfm if build_model_xfms is None # use transforms from model building if we have them: else s.defer( xfmconcat( #nlin_component.Algorithms.concat( [ build_model_xfms[row.img.path], s.defer( xfminvert( #nlin_component.Algorithms.invert( build_model_xfms[row.template. path], subdir="tmp")) ])), like=row.img, invert=True, #use_nn_interpolation=True interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range', '-labels'))))) ) if len(imgs) > 1 else pd.DataFrame({ 'img': [], 'label_file': [] }) # ... as no distinct templates to align if only one image supplied (#320) imgs_with_all_labels = pd.concat([ atlas_labelled_imgs[['img', 'label_file']], template_labelled_imgs[['img', 'label_file']] ], ignore_index=True) else: imgs_with_all_labels = atlas_labelled_imgs #imgs_with_all_labels = imgs_with_all_labels.applymap( # lambda x: s.defer(nlin_component.ToMinc.to_mnc(x))) segmented_imgs = (imgs_with_all_labels.groupby('img').aggregate({ 'label_file': lambda resampled_label_files: list(resampled_label_files) }).rename(columns={ 'label_file': 'label_files' }).reset_index().assign(voted_labels=lambda df: df.apply( axis=1, func=lambda row: s.defer( voxel_vote(label_files=row.label_files, output_dir=os.path.join(row.img.pipeline_sub_dir, row.img.output_sub_dir), name=row.img.filename_wo_ext + "_voted")) )).apply(axis=1, func=lambda row: row.img._replace(labels=row.voted_labels))) return Result(stages=s, output=segmented_imgs)
def mbm(imgs : List[MincAtom], options : MBMConf, prefix : str, output_dir : str = ""): # TODO could also allow pluggable pipeline parts e.g. LSQ6 could be substituted out for the modified LSQ6 # for the kidney tips, etc... # TODO this is tedious and annoyingly similar to the registration chain ... lsq6_dir = os.path.join(output_dir, prefix + "_lsq6") lsq12_dir = os.path.join(output_dir, prefix + "_lsq12") nlin_dir = os.path.join(output_dir, prefix + "_nlin") s = Stages() if len(imgs) == 0: raise ValueError("Please, some files!") # FIXME: why do we have to call registration_targets *outside* of lsq6_nuc_inorm? is it just because of the extra # options required? Also, shouldn't options.registration be a required input (as it contains `input_space`) ...? targets = registration_targets(lsq6_conf=options.mbm.lsq6, app_conf=options.application, first_input_file=imgs[0].path) # TODO this is quite tedious and duplicates stuff in the registration chain ... resolution = (options.registration.resolution or get_resolution_from_file(targets.registration_standard.path)) options.registration = options.registration.replace(resolution=resolution) # FIXME it probably makes most sense if the lsq6 module itself (even within lsq6_nuc_inorm) handles the run_lsq6 # setting (via use of the identity transform) since then this doesn't have to be implemented for every pipeline if options.mbm.lsq6.run_lsq6: lsq6_result = s.defer(lsq6_nuc_inorm(imgs=imgs, resolution=resolution, registration_targets=targets, lsq6_dir=lsq6_dir, lsq6_options=options.mbm.lsq6)) else: # TODO don't actually do this resampling if not required (i.e., if the imgs already have the same grids) identity_xfm = s.defer(param2xfm(out_xfm=FileAtom(name="identity.xfm"))) lsq6_result = [XfmHandler(source=img, target=img, xfm=identity_xfm, resampled=s.defer(mincresample_new(img=img, like=targets.registration_standard, xfm=identity_xfm))) for img in imgs] # what about running nuc/inorm without a linear registration step?? full_hierarchy = get_nonlinear_configuration_from_options(nlin_protocol=options.mbm.nlin.nlin_protocol, reg_method=options.mbm.nlin.reg_method, file_resolution=resolution) lsq12_nlin_result = s.defer(lsq12_nlin_build_model(imgs=[xfm.resampled for xfm in lsq6_result], resolution=resolution, lsq12_dir=lsq12_dir, nlin_dir=nlin_dir, nlin_prefix=prefix, lsq12_conf=options.mbm.lsq12, nlin_conf=full_hierarchy)) inverted_xfms = [s.defer(invert_xfmhandler(xfm)) for xfm in lsq12_nlin_result.output] determinants = s.defer(determinants_at_fwhms( xfms=inverted_xfms, inv_xfms=lsq12_nlin_result.output, blur_fwhms=options.mbm.stats.stats_kernels)) overall_xfms = [s.defer(concat_xfmhandlers([rigid_xfm, lsq12_nlin_xfm])) for rigid_xfm, lsq12_nlin_xfm in zip(lsq6_result, lsq12_nlin_result.output)] output_xfms = (pd.DataFrame({ "rigid_xfm" : lsq6_result, # maybe don't return this if LSQ6 not run?? "lsq12_nlin_xfm" : lsq12_nlin_result.output, "overall_xfm" : overall_xfms })) # we could `merge` the determinants with this table, but preserving information would cause lots of duplication # of the transforms (or storing determinants in more columns, but iterating over dynamically known columns # seems a bit odd ...) # TODO transpose these fields?}) #avg_img=lsq12_nlin_result.avg_img, # inconsistent w/ WithAvgImgs[...]-style outputs # "determinants" : determinants }) #output.avg_img = lsq12_nlin_result.avg_img #output.determinants = determinants # TODO temporary - remove once incorporated properly into `output` proper # TODO add more of lsq12_nlin_result? # FIXME: this needs to go outside of the `mbm` function to avoid being run from within other pipelines (or # those other pipelines need to turn off this option) # TODO return some MAGeT stuff from MBM function ?? # if options.mbm.mbm.run_maget: # import copy # maget_options = copy.deepcopy(options) #Namespace(maget=options) # #maget_options # #maget_options.maget = maget_options.mbm # #maget_options.execution = options.execution # #maget_options.application = options.application # maget_options.maget = options.mbm.maget # del maget_options.mbm # # s.defer(maget([xfm.resampled for xfm in lsq6_result], # options=maget_options, # prefix="%s_MAGeT" % prefix, # output_dir=os.path.join(output_dir, prefix + "_processed"))) # should also move outside `mbm` function ... #if options.mbm.thickness.run_thickness: # if not options.mbm.segmentation.run_maget: # warnings.warn("MAGeT files (atlases, protocols) are needed to run thickness calculation.") # # run MAGeT to segment the nlin average: # import copy # maget_options = copy.deepcopy(options) #Namespace(maget=options) # maget_options.maget = options.mbm.maget # del maget_options.mbm # segmented_avg = s.defer(maget(imgs=[lsq12_nlin_result.avg_img], # options=maget_options, # output_dir=os.path.join(options.application.output_directory, # prefix + "_processed"), # prefix="%s_thickness_MAGeT" % prefix)).ix[0].img # thickness = s.defer(cortical_thickness(xfms=pd.Series(inverted_xfms), atlas=segmented_avg, # label_mapping=FileAtom(options.mbm.thickness.label_mapping), # atlas_fwhm=0.56, thickness_fwhm=0.56)) # TODO magic fwhms # # TODO write CSV -- should `cortical_thickness` do this/return a table? # FIXME: this needs to go outside of the `mbm` function to avoid being run from within other pipelines (or # those other pipelines need to turn off this option) if options.mbm.common_space.do_common_space_registration: warnings.warn("This feature is experimental ...") if not options.mbm.common_space.common_space_model: raise ValueError("No common space template provided!") # TODO allow lsq6 registration as well ... common_space_model = MincAtom(options.mbm.common_space.common_space_model, pipeline_sub_dir=os.path.join(options.application.output_directory, options.application.pipeline_name + "_processed")) # TODO allow different lsq12/nlin config params than the ones used in MBM ... # WEIRD ... see comment in lsq12_nlin code ... nlin_conf = full_hierarchy.confs[-1] if isinstance(full_hierarchy, MultilevelMincANTSConf) else full_hierarchy # also weird that we need to call get_linear_configuration_from_options here ... ? lsq12_conf = get_linear_configuration_from_options(conf=options.mbm.lsq12, transform_type=LinearTransType.lsq12, file_resolution=resolution) xfm_to_common = s.defer(lsq12_nlin(source=lsq12_nlin_result.avg_img, target=common_space_model, lsq12_conf=lsq12_conf, nlin_conf=nlin_conf, resample_source=True)) model_common = s.defer(mincresample_new(img=lsq12_nlin_result.avg_img, xfm=xfm_to_common.xfm, like=common_space_model, postfix="_common")) overall_xfms_common = [s.defer(concat_xfmhandlers([rigid_xfm, nlin_xfm, xfm_to_common])) for rigid_xfm, nlin_xfm in zip(lsq6_result, lsq12_nlin_result.output)] xfms_common = [s.defer(concat_xfmhandlers([nlin_xfm, xfm_to_common])) for nlin_xfm in lsq12_nlin_result.output] output_xfms = output_xfms.assign(xfm_common=xfms_common, overall_xfm_common=overall_xfms_common) log_nlin_det_common, log_full_det_common = [dets.map(lambda d: s.defer(mincresample_new( img=d, xfm=xfm_to_common.xfm, like=common_space_model, postfix="_common", extra_flags=("-keep_real_range",), interpolation=Interpolation.nearest_neighbour))) for dets in (determinants.log_nlin_det, determinants.log_full_det)] determinants = determinants.assign(log_nlin_det_common=log_nlin_det_common, log_full_det_common=log_full_det_common) output = Namespace(avg_img=lsq12_nlin_result.avg_img, xfms=output_xfms, determinants=determinants) if options.mbm.common_space.do_common_space_registration: output.model_common = model_common return Result(stages=s, output=output)
def common_space(mbm_result, options): s = Stages() # TODO: the interface of this function (basically a destructive 'id' function) is horrific # TODO: instead, copy the mbm_result here ?? if not options.mbm.common_space.common_space_model: raise ValueError("No common space template provided!") if not options.mbm.common_space.common_space_mask: warnings.warn( "No common space mask provided ... might be OK if your consensus average mask is OK" ) # TODO allow lsq6 registration as well ... common_space_model = MincAtom( options.mbm.common_space.common_space_model, # TODO fix the subdirectories! mask=MincAtom(options.mbm.common_space.common_space_mask, pipeline_sub_dir=os.path.join( options.application.output_directory, options.application.pipeline_name + "_processed")) if options.mbm.common_space.common_space_mask else None, pipeline_sub_dir=os.path.join( options.application.output_directory, options.application.pipeline_name + "_processed")) # TODO allow different lsq12/nlin config params than the ones used in MBM ... # full_hierarchy = get_nonlinear_configuration_from_options(nlin_protocol=options.mbm.nlin.nlin_protocol, # reg_method=options.mbm.nlin.reg_method, # file_resolution=options.registration.resolution) # WEIRD ... see comment in lsq12_nlin code ... # nlin_conf = full_hierarchy.confs[-1] if isinstance(full_hierarchy, MultilevelANTSConf) else full_hierarchy # also weird that we need to call get_linear_configuration_from_options here ... ? # nlin_build_model_component = model_building_with_initial_target_generation( # final_model_building_component=nlin_build_model_component, # prelim_model_building_component=prelim_nlin_build_model_component) # TODO don't use name 'x_module' for something that's technically not a module ... perhaps unit/component? nlin_component = get_nonlinear_component( reg_method=options.mbm.nlin.reg_method) lsq12_conf = get_linear_configuration_from_options( conf=options.mbm.lsq12, transform_type=LinearTransType.lsq12, file_resolution=options.registration.resolution) # N.B.: options.registration.resolution has been *updated* correctly by mbm( ). sigh ... model_to_common = s.defer( lsq12_nlin( source=mbm_result.avg_img, target=common_space_model, lsq12_conf=lsq12_conf, nlin_module=nlin_component, resolution=options.registration.resolution, nlin_options=options.mbm.nlin.nlin_protocol, # =nlin_conf, resample_source=True)) model_common = s.defer( mincresample_new(img=mbm_result.avg_img, xfm=model_to_common.xfm, like=common_space_model, postfix="_common")) overall_xfms_to_common = [ s.defer(concat_xfmhandlers([rigid_xfm, nlin_xfm, model_to_common])) for rigid_xfm, nlin_xfm in zip(mbm_result.xfms.rigid_xfm, mbm_result.xfms.lsq12_nlin_xfm) ] overall_xfms_to_common_inv = [ s.defer(invert_xfmhandler(xfmhandler)) for xfmhandler in [ s.defer(concat_xfmhandlers([rigid_xfm, nlin_xfm, model_to_common])) for rigid_xfm, nlin_xfm in zip(mbm_result.xfms.rigid_xfm, mbm_result.xfms.lsq12_nlin_xfm) ] ] xfms_to_common = [ s.defer(concat_xfmhandlers([nlin_xfm, model_to_common])) for nlin_xfm in mbm_result.xfms.lsq12_nlin_xfm ] mbm_result.xfms = mbm_result.xfms.assign( xfm_to_common=xfms_to_common, overall_xfm_to_common=overall_xfms_to_common) if options.mbm.stats.calc_stats: log_nlin_det_common, log_full_det_common = ([ dets.map(lambda d: s.defer( mincresample_new(img=d, xfm=model_to_common.xfm, like=common_space_model, postfix="_common"))) for dets in (mbm_result.determinants.log_nlin_det, mbm_result.determinants.log_full_det) ]) overall_determinants = s.defer( determinants_at_fwhms(xfms=overall_xfms_to_common_inv, blur_fwhms=options.mbm.stats.stats_kernels)) mbm_result.determinants = \ mbm_result.determinants.assign(log_nlin_det_common=log_nlin_det_common, log_full_det_common=log_full_det_common, log_nlin_overall_det_common=overall_determinants.log_nlin_det, log_full_overall_det_common=overall_determinants.log_full_det ) mbm_result.model_common = model_common return Result(stages=s, output=mbm_result)
def maget_mask(imgs: List[MincAtom], maget_options, resolution: float, pipeline_sub_dir: str, atlases=None): s = Stages() original_imgs = imgs imgs = copy.deepcopy(imgs) original_imgs = pd.Series(original_imgs, index=[img.path for img in original_imgs]) for img in imgs: img.output_sub_dir = os.path.join(img.output_sub_dir, "masking") # TODO dereference maget_options -> maget_options.maget outside maget_mask call? if atlases is None: atlases = get_atlases(maget_options.maget, pipeline_sub_dir=pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options(maget_options.lsq12, LinearTransType.lsq12, resolution) #nlin_module = get_nonlinear_component(reg_method=options.mbm.nlin.reg_method) #masking_nlin_hierarchy = get_nonlinear_configuration_from_options(maget_options.maget.masking_nlin_protocol, # next(iter(maget_options.maget.flags_.masking_nlin_protocol)), # maget_options.maget.mask_method, # resolution) masking_nlin_component = get_nonlinear_component( reg_method=maget_options.maget.mask_method) algorithms = masking_nlin_component.Algorithms #masking_nlin_conf = (masking_nlin_component.parse_protocol_file( # maget_options.maget.masking_nlin_protocol, resolution=resolution) # if maget_options.maget.masking_nlin_protocol is not None # else masking_nlin_component.get_default_conf(resolution=resolution)) # TODO lift outside then delete #masking_imgs = copy.deepcopy(imgs) #for img in masking_imgs: # img.pipeline_sub_dir = os.path.join(img.pipeline_sub_dir, "masking") masking_alignments = pd.DataFrame({ 'img': img, 'atlas': atlas, 'xfm': s.defer( lsq12_nlin( source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_options=maget_options.maget.masking_nlin_protocol, #masking_nlin_conf, resolution=resolution, nlin_module=masking_nlin_component, resample_source=False)) } for img in imgs for atlas in atlases) # propagate a mask to each image using the above `alignments` as follows: # - for each image, voxel_vote on the masks propagated to that image to get a suitable mask # - run mincmath -clobber -mult <img> <voted_mask> to apply the mask to the files masked_img = ( masking_alignments.assign(resampled_mask=lambda df: df.apply( axis=1, func=lambda row: s.defer( algorithms.resample( img=row.atlas.mask, #apply(lambda x: x.mask), xfm=row.xfm.xfm, #apply(lambda x: x.xfm), like=row.img, invert=True, #interpolation=Interpolation.nearest_neighbour, postfix="-input-mask", subdir="tmp", # TODO annoying hack; fix mincresample(_mask) ...: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-input-mask" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), use_nn_interpolation=True)))).groupby('img', as_index=False). aggregate({ 'resampled_mask': lambda masks: list(masks) }).rename(columns={ "resampled_mask": "resampled_masks" }).assign(voted_mask=lambda df: df.apply( axis=1, func=lambda row: # FIXME cannot use mincmath here !!! s.defer( mincmath(op="max", vols=sorted(row.resampled_masks), new_name="%s_max_mask" % row.img.filename_wo_ext, subdir="tmp")))).apply(axis=1, func=lambda row: row.img. _replace(mask=row.voted_mask))) # resample the atlas images back to the input images: # (note: this doesn't modify `masking_alignments`, but only stages additional outputs) masking_alignments.assign(resampled_img=lambda df: df.apply( axis=1, func=lambda row: s.defer( algorithms.resample( img=row.atlas, xfm=row.xfm.xfm, #.apply(lambda x: x.xfm), subdir="tmp", # TODO delete this stupid hack: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-resampled" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), like=row.img, invert=True)))) for img in masked_img: img.output_sub_dir = original_imgs.loc[img.path].output_sub_dir return Result(stages=s, output=masked_img)
def maget_mask(imgs : List[MincAtom], atlases, options): s = Stages() resample = np.vectorize(mincresample_new, excluded={"extra_flags"}) defer = np.vectorize(s.defer) lsq12_conf = get_linear_configuration_from_options(options.maget.lsq12, LinearTransType.lsq12, options.registration.resolution) masking_nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.maget.masking_nlin_protocol, options.maget.maget.mask_method, options.registration.resolution) masking_alignments = pd.DataFrame({ 'img' : img, 'atlas' : atlas, 'xfm' : s.defer(lsq12_nlin(source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_conf=masking_nlin_hierarchy, resample_source=False))} for img in imgs for atlas in atlases) # propagate a mask to each image using the above `alignments` as follows: # - for each image, voxel_vote on the masks propagated to that image to get a suitable mask # - run mincmath -clobber -mult <img> <voted_mask> to apply the mask to the files masked_img = ( masking_alignments .assign(resampled_mask=lambda df: defer(resample(img=df.atlas.apply(lambda x: x.mask), xfm=df.xfm.apply(lambda x: x.xfm), like=df.img, invert=True, interpolation=Interpolation.nearest_neighbour, postfix="-input-mask", subdir="tmp", # TODO annoying hack; fix mincresample(_mask) ...: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-input-mask" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), extra_flags=("-keep_real_range",)))) .groupby('img', sort=False, as_index=False) # sort=False: just for speed (might also need to implement more comparison methods on `MincAtom`s) .aggregate({'resampled_mask' : lambda masks: list(masks)}) .rename(columns={"resampled_mask" : "resampled_masks"}) .assign(voted_mask=lambda df: df.apply(axis=1, func=lambda row: s.defer(voxel_vote(label_files=row.resampled_masks, name="%s_voted_mask" % row.img.filename_wo_ext, output_dir=os.path.join(row.img.output_sub_dir, "tmp"))))) .assign(masked_img=lambda df: df.apply(axis=1, func=lambda row: s.defer(mincmath(op="mult", # img must precede mask here # for output image range to be correct: vols=[row.img, row.voted_mask], new_name="%s_masked" % row.img.filename_wo_ext, subdir="resampled"))))) #['img'] # resample the atlas images back to the input images: # (note: this doesn't modify `masking_alignments`, but only stages additional outputs) masking_alignments.assign(resampled_img=lambda df: defer(resample(img=df.atlas, xfm=df.xfm.apply(lambda x: x.xfm), subdir="tmp", # TODO delete this stupid hack: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-resampled" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), like=df.img, invert=True))) # replace the table of alignments with a new one with masked images masking_alignments = (pd.merge(left=masking_alignments.assign(unmasked_img=lambda df: df.img), right=masked_img, on=["img"], how="right", sort=False) .assign(img=lambda df: df.masked_img)) return Result(stages=s, output=masking_alignments)
def tissue_vision_pipeline(options): output_dir = options.application.output_directory pipeline_name = options.application.pipeline_name csv = original_csv = get_imgs(options.application) # check_MINC_input_files([img.path for img in imgs]) s = Stages() s.defer(create_quality_control_images(imgs=csv['anatomical_MincAtom'].tolist(), montage_dir=output_dir, montage_output=os.path.join(output_dir, pipeline_name + "_resampled", "input_montage"), auto_range=True, message="input_mincs")) ############################# # Step 1: Run MBM.py to create a consensus average ############################# mbm_result = s.defer(mbm(imgs=csv['anatomical_MincAtom'].tolist(), options=options, prefix=options.application.pipeline_name, output_dir=output_dir, with_maget=False)) #TODO remove transforms = mbm_result.xfms.assign(native_file=lambda df: df.rigid_xfm.apply(lambda x: x.source), lsq6_file=lambda df: df.lsq12_nlin_xfm.apply(lambda x: x.source), lsq6_mask_file=lambda df: df.lsq12_nlin_xfm.apply(lambda x: x.source.mask if x.source.mask else ""), nlin_file=lambda df: df.lsq12_nlin_xfm.apply(lambda x: x.resampled), nlin_mask_file=lambda df: df.lsq12_nlin_xfm.apply(lambda x: x.resampled.mask if x.resampled.mask else ""))\ .applymap(maybe_deref_path) determinants = mbm_result.determinants.drop(["full_det", "nlin_det"], axis=1)\ .applymap(maybe_deref_path) csv = csv.assign(anatomical_lsq6_MincAtom=mbm_result.xfms.lsq12_nlin_xfm.apply(lambda xfm: xfm.source), mbm_lsq6_XfmAtom=mbm_result.xfms.rigid_xfm.apply(lambda x: x.xfm), mbm_lsq12_nlin_XfmAtom=mbm_result.xfms.lsq12_nlin_xfm.apply(lambda x: x.xfm), mbm_full_XfmAtom=mbm_result.xfms.overall_xfm.apply(lambda x: x.xfm)) # x.assign(count_lsq6_MincAtom=lambda df: [x + y for x, y in zip(df["x"], df["y"])]) csv = csv.assign(count_lsq6_MincAtom = lambda df: [s.defer(mincresample_new(img = img, xfm = xfm, like = like)) for img, xfm, like in zip(df["count_MincAtom"], df["mbm_lsq6_XfmAtom"], df["anatomical_lsq6_MincAtom"])]) ############################# # Step 2: Register consensus average to ABI tissuevision Atlas ############################# lsq12_conf = get_linear_configuration_from_options(conf=options.mbm.lsq12, transform_type=LinearTransType.lsq12, file_resolution=options.registration.resolution) nlin_component = get_nonlinear_component(reg_method=options.mbm.nlin.reg_method) atlas_target = MincAtom(name=options.consensus_to_atlas.atlas_target, orig_name=options.consensus_to_atlas.atlas_target, mask=MincAtom(name=options.consensus_to_atlas.atlas_target_mask, orig_name=options.consensus_to_atlas.atlas_target_mask)) atlas_target_label = MincAtom(name=options.consensus_to_atlas.atlas_target_label, orig_name=options.consensus_to_atlas.atlas_target_label, mask=MincAtom(name=options.consensus_to_atlas.atlas_target_mask, orig_name=options.consensus_to_atlas.atlas_target_mask)) lsq12_nlin_result = s.defer(lsq12_nlin(source=mbm_result.avg_img, target=atlas_target, lsq12_conf=lsq12_conf, nlin_module=nlin_component, nlin_options=options.mbm.nlin.nlin_protocol, resolution=options.registration.resolution, resample_source=False )) ############################# # Step 3: Resample count volumes to ABI tissuevision Atlas space and vice versa ############################# csv = csv.assign(lsq6_to_atlas_XfmAtom = lambda df: df['mbm_lsq12_nlin_XfmAtom'].apply(lambda xfm: s.defer(xfmconcat([xfm, lsq12_nlin_result.xfm])))) csv = csv.assign( anatomical_targetspace_MincAtom=lambda df: [s.defer(mincresample_new(img=img, xfm=xfm, like=atlas_target)) for img, xfm in zip(df["anatomical_lsq6_MincAtom"], df["lsq6_to_atlas_XfmAtom"])], count_targetspace_MincAtom=lambda df: [s.defer(mincresample_new(img=img, xfm=xfm, like=atlas_target)) for img, xfm in zip(df["count_lsq6_MincAtom"], df["lsq6_to_atlas_XfmAtom"])], atlas_lsq6space_MincAtom=lambda df: [s.defer(mincresample_new(img=atlas_target_label, xfm=xfm, like=like, invert=True, interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range',))) for xfm, like in zip( df["lsq6_to_atlas_XfmAtom"], df["count_lsq6_MincAtom"])] ) csv.applymap(maybe_deref_path).to_csv("analysis.csv",index=False) s.defer(create_quality_control_images(imgs=csv.count_targetspace_MincAtom.tolist(), montage_dir=output_dir, montage_output=os.path.join(output_dir, pipeline_name + "_resampled", "count_montage"), auto_range=True, message="count_mincs")) return Result(stages=s, output=())
def maget(imgs : List[MincAtom], options, prefix, output_dir): # FIXME prefix, output_dir aren't used !! s = Stages() maget_options = options.maget.maget pipeline_sub_dir = os.path.join(options.application.output_directory, options.application.pipeline_name + "_atlases") if maget_options.atlas_lib is None: raise ValueError("Need some atlases ...") #atlas_dir = os.path.join(output_dir, "input_atlases") ??? # TODO should alternately accept a CSV file ... atlas_library = read_atlas_dir(atlas_lib=maget_options.atlas_lib, pipeline_sub_dir=pipeline_sub_dir) if len(atlas_library) == 0: raise ValueError("No atlases found in specified directory '%s' ..." % options.maget.maget.atlas_lib) num_atlases_needed = min(maget_options.max_templates, len(atlas_library)) # TODO arbitrary; could choose atlases better ... atlases = atlas_library[:num_atlases_needed] # TODO issue a warning if not all atlases used or if more atlases requested than available? # TODO also, doesn't slicing with a higher number (i.e., if max_templates > n) go to the end of the list anyway? lsq12_conf = get_linear_configuration_from_options(options.maget.lsq12, LinearTransType.lsq12, options.registration.resolution) masking_nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.maget.masking_nlin_protocol, options.maget.maget.mask_method, options.registration.resolution) nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.nlin.nlin_protocol, options.maget.nlin.reg_method, options.registration.resolution) resample = np.vectorize(mincresample_new, excluded={"extra_flags"}) defer = np.vectorize(s.defer) # plan the basic registrations between all image-atlas pairs; store the result paths in a table masking_alignments = pd.DataFrame({ 'img' : img, 'atlas' : atlas, 'xfm' : s.defer(lsq12_nlin(source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_conf=masking_nlin_hierarchy, resample_source=False))} for img in imgs for atlas in atlases) if maget_options.mask or maget_options.mask_only: masking_alignments = s.defer(maget_mask(imgs, atlases, options)) masked_atlases = atlases.apply(lambda atlas: s.defer(mincmath(op='mult', vols=[atlas, atlas.mask], subdir="resampled", new_name="%s_masked" % atlas.filename_wo_ext))) # now propagate only the masked form of the images and atlases: imgs = masking_alignments.img atlases = masked_atlases # TODO is this needed? if maget_options.mask_only: # register each input to each atlas, creating a mask return Result(stages=s, output=masking_alignments) # TODO rename `alignments` to `registrations`?? else: del masking_alignments # this `del` is just to verify that we don't accidentally use this later, since my intent is that these # coarser alignments shouldn't be re-used, just the masked images they create; can be removed later # if a sensible use is found if maget_options.pairwise: def choose_new_templates(ts, n): # currently silly, but we might implement a smarter method ... # FIXME what if there aren't enough other imgs around?! This silently goes weird ... return ts[:n+1] # n+1 instead of n: choose one more since we won't use image as its own template ... new_templates = choose_new_templates(ts=imgs, n=maget_options.max_templates) # note these images are the masked ones if masking was done ... # TODO write a function to do these alignments and the image->atlas one above # align the new templates chosen from the images to the initial atlases: new_template_to_atlas_alignments = ( pd.DataFrame({ 'img' : template, 'atlas' : atlas, 'xfm' : s.defer(lsq12_nlin(source=template, target=atlas, lsq12_conf=lsq12_conf, nlin_conf=nlin_hierarchy, resample_source=False))} for template in new_templates for atlas in atlases)) # ... and these atlases are multiplied by their masks (but is this necessary?) # label the new templates from resampling the atlas labels onto them: # TODO now vote on the labels to be used for the new templates ... # TODO extract into procedure? new_templates_labelled = ( new_template_to_atlas_alignments .assign(resampled_labels=lambda df: defer( resample(img=df.atlas.apply(lambda x: x.labels), xfm=df.xfm.apply(lambda x: x.xfm), interpolation=Interpolation.nearest_neighbour, extra_flags=("-keep_real_range",), like=df.img, invert=True))) .groupby('img', sort=False, as_index=False) .aggregate({'resampled_labels' : lambda labels: list(labels)}) .assign(voted_labels=lambda df: df.apply(axis=1, func=lambda row: s.defer(voxel_vote(label_files=row.resampled_labels, name="%s_template_labels" % row.img.filename_wo_ext, output_dir=os.path.join( row.img.pipeline_sub_dir, row.img.output_sub_dir, "labels")))))) # TODO write a procedure for this assign-groupby-aggregate-rename... # FIXME should be in above algebraic manipulation but MincAtoms don't support flexible immutable updating for row in pd.merge(left=new_template_to_atlas_alignments, right=new_templates_labelled, on=["img"], how="right", sort=False).itertuples(): row.img.labels = s.defer(mincresample_new(img=row.voted_labels, xfm=row.xfm.xfm, like=row.img, invert=True, interpolation=Interpolation.nearest_neighbour, #postfix="-input-labels", # this makes names really long ...: # TODO this doesn't work for running MAGeT on the nlin avg: #new_name_wo_ext="%s_on_%s" % # (row.voted_labels.filename_wo_ext, # row.img.filename_wo_ext), #postfix="_labels_via_%s" % row.xfm.xfm.filename_wo_ext, new_name_wo_ext="%s_via_%s" % (row.voted_labels.filename_wo_ext, row.xfm.xfm.filename_wo_ext), extra_flags=("-keep_real_range",))) # now that the new templates have been labelled, combine with the atlases: # FIXME use the masked atlases created earlier ?? all_templates = pd.concat([new_templates_labelled.img, atlases], ignore_index=True) # now take union of the resampled labels from the new templates with labels from the original atlases: #all_alignments = pd.concat([image_to_template_alignments, # alignments.rename(columns={ "atlas" : "template" })], # ignore_index=True, join="inner") else: all_templates = atlases # now register each input to each selected template # N.B.: Even though we've already registered each image to each initial atlas, this happens again here, # but using `nlin_hierarchy` instead of `masking_nlin_hierarchy` as options. # This is not 'work-efficient' in the sense that this computation happens twice (although # hopefully at greater precision the second time!), but the idea is to run a coarse initial # registration to get a mask and then do a better registration with that mask (though I'm not # sure exactly when this is faster than doing a single registration). # This _can_ allow the overall computation to finish more rapidly # (depending on the relative speed of the two alignment methods/parameters, # number of atlases and other templates used, number of cores available, etc.). image_to_template_alignments = ( pd.DataFrame({ "img" : img, "template" : template_img, "xfm" : xfm } for img in imgs # TODO use the masked imgs here? for template_img in all_templates # FIXME delete this one alignment #labelled_templates[labelled_templates.img != img] # since equality is equality of filepaths (a bit dangerous) # TODO is there a more direct/faster way just to delete the template? for xfm in [s.defer(lsq12_nlin(source=img, target=template_img, lsq12_conf=lsq12_conf, nlin_conf=nlin_hierarchy))] ) ) # now do a voxel_vote on all resampled template labels, just as earlier with the masks voted = (image_to_template_alignments .assign(resampled_labels=lambda df: defer(resample(img=df.template.apply(lambda x: x.labels), # FIXME bug: at this point templates from template_alignments # don't have associated labels (i.e., `None`s) -- fatal xfm=df.xfm.apply(lambda x: x.xfm), interpolation=Interpolation.nearest_neighbour, extra_flags=("-keep_real_range",), like=df.img, invert=True))) .groupby('img', sort=False) # TODO the pattern groupby-aggregate(lambda x: list(x))-reset_index-assign is basically a hack # to do a groupby-assign with access to the group name; # see http://stackoverflow.com/a/30224447/849272 for a better solution # (note this pattern occurs several times in MAGeT and two-level code) .aggregate({'resampled_labels' : lambda labels: list(labels)}) .reset_index() .assign(voted_labels=lambda df: defer(np.vectorize(voxel_vote)(label_files=df.resampled_labels, output_dir=df.img.apply( lambda x: os.path.join( x.pipeline_sub_dir, x.output_sub_dir)))))) # TODO doing mincresample -invert separately for the img->atlas xfm for mask, labels is silly # (when Pydpiper's `mincresample` does both automatically)? # blargh, another destructive update ... for row in voted.itertuples(): row.img.labels = row.voted_labels # returning voted_labels as a column is slightly redundant, but possibly useful ... return Result(stages=s, output=voted) # voted.drop("voted_labels", axis=1))
def maget(imgs : List[MincAtom], options, prefix, output_dir, build_model_xfms=None): # FIXME prefix, output_dir aren't used !! s = Stages() maget_options = options.maget.maget resolution = options.registration.resolution # TODO or get_resolution_from_file(...) -- only if file always exists! pipeline_sub_dir = os.path.join(options.application.output_directory, options.application.pipeline_name + "_atlases") atlases = get_atlases(maget_options, pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options(options.maget.lsq12, transform_type=LinearTransType.lsq12, file_resolution=resolution) nlin_component = get_nonlinear_component(options.maget.nlin.reg_method) # TODO should this be here or outside `maget` call? #imgs = [s.defer(nlin_component.ToMinc.from_mnc(img)) for img in imgs] #nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.nlin.nlin_protocol, # next(iter(options.maget.nlin.flags_.nlin_protocol)), # reg_method=options.maget.nlin.reg_method, # file_resolution=resolution) if maget_options.mask or maget_options.mask_only: # this used to return alignments but doesn't currently do so masked_img = s.defer(maget_mask(imgs=imgs, maget_options=options.maget, atlases=atlases, pipeline_sub_dir=pipeline_sub_dir + "_masking", # FIXME repeats all alignments!!! resolution=resolution)) # now propagate only the masked form of the images and atlases: imgs = masked_img #atlases = masked_atlases # TODO is this needed? if maget_options.mask_only: # register each input to each atlas, creating a mask return Result(stages=s, output=masked_img) # TODO rename `alignments` to `registrations`?? else: if maget_options.mask: del masked_img # this `del` is just to verify that we don't accidentally use this later, since these potentially # coarser alignments shouldn't be re-used (but if the protocols for masking and alignment are the same, # hash-consing will take care of things), just the masked images they create; can be removed later # if a sensible use is found # images with labels from atlases # N.B.: Even though we've already registered each image to each initial atlas, this happens again here, # but using `nlin_hierarchy` instead of `masking_nlin_hierarchy` as options. # This is not 'work-efficient' in the sense that this computation happens twice (although # hopefully at greater precision the second time!), but the idea is to run a coarse initial # registration to get a mask and then do a better registration with that mask (though I'm not # sure exactly when this is faster than doing a single registration). # This _can_ allow the overall computation to finish more rapidly # (depending on the relative speed of the two alignment methods/parameters, # number of atlases and other templates used, number of cores available, etc.). atlas_labelled_imgs = ( pd.DataFrame({ 'img' : img, 'label_file' : s.defer( # can't use `label` in a pd.DataFrame index! mincresample_new(img=atlas.labels, xfm=s.defer(lsq12_nlin(source=img, target=atlas, nlin_module=nlin_component, lsq12_conf=lsq12_conf, nlin_options=options.maget.nlin.nlin_protocol, resolution=resolution, #nlin_conf=nlin_hierarchy, resample_source=False)).xfm, like=img, invert=True, interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range', '-labels')))} for img in imgs for atlas in atlases) ) if maget_options.pairwise: def choose_new_templates(ts, n): # currently silly, but we might implement a smarter method ... # FIXME what if there aren't enough other imgs around?! This silently goes weird ... return pd.Series(ts[:n+1]) # n+1 instead of n: choose one more since we won't use image as its own template ... # FIXME: the --max-templates flag is ambiguously named ... should be --max-new-templates # (and just use all atlases) # TODO we could have a separate templates_csv (or --template-files f [f ...]) but you can just # run a separate MAGeT pipeline and #if maget_options.templates_csv: # templates = pd.read_csv(maget_options.templates_csv).template #else: templates = pd.DataFrame({ 'template' : choose_new_templates(ts=imgs, n=maget_options.max_templates - len(atlases))}) # note these images are the masked ones if masking was done ... # the templates together with their (multiple) labels from the atlases (this merge just acts as a filter) labelled_templates = pd.merge(left=atlas_labelled_imgs, right=templates, left_on="img", right_on="template").drop('img', axis=1) # images with new labels from the templates imgs_and_templates = pd.merge(#left=atlas_labelled_imgs, left=pd.DataFrame({ "img" : imgs }).assign(fake=1), right=labelled_templates.assign(fake=1), on='fake') #left_on='img', right_on='template') # TODO do select here instead of below? #if build_model_xfms is not None: # # use path instead of full mincatom as key in case we're reading these in from a CSV: # xfm_dict = { x.source.path : x.xfm for x in build_model_xfms } template_labelled_imgs = ( imgs_and_templates .rename(columns={ 'label_file' : 'template_label_file' }) # don't register template to itself, since otherwise atlases would vote on that template twice .loc[lambda df: df.index.map(lambda ix: df.img[ix].path != df.template[ix].path)] .assign(label_file=lambda df: df.apply(axis=1, func=lambda row: s.defer( # TODO switch to uses of nlin_component.whatever(...) in several places below? mincresample_new( #nlin_component.Algorithms.resample( img=row.template_label_file, xfm=s.defer( lsq12_nlin(source=row.img, target=row.template, lsq12_conf=lsq12_conf, resolution=resolution, nlin_module=nlin_component, nlin_options=options.maget.nlin.nlin_protocol, #nlin_conf=nlin_hierarchy, resample_source=False)).xfm if build_model_xfms is None # use transforms from model building if we have them: else s.defer( xfmconcat( #nlin_component.Algorithms.concat( [build_model_xfms[row.img.path], s.defer( xfminvert( #nlin_component.Algorithms.invert( build_model_xfms[row.template.path], subdir="tmp"))])), like=row.img, invert=True, #use_nn_interpolation=True interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range', '-labels') )))) ) if len(imgs) > 1 else pd.DataFrame({ 'img' : [], 'label_file' : [] }) # ... as no distinct templates to align if only one image supplied (#320) imgs_with_all_labels = pd.concat([atlas_labelled_imgs[['img', 'label_file']], template_labelled_imgs[['img', 'label_file']]], ignore_index=True) else: imgs_with_all_labels = atlas_labelled_imgs #imgs_with_all_labels = imgs_with_all_labels.applymap( # lambda x: s.defer(nlin_component.ToMinc.to_mnc(x))) segmented_imgs = ( imgs_with_all_labels .groupby('img') .aggregate({'label_file' : lambda resampled_label_files: list(resampled_label_files)}) .rename(columns={ 'label_file' : 'label_files' }) .reset_index() .assign(voted_labels=lambda df: df.apply(axis=1, func=lambda row: s.defer(voxel_vote(label_files=row.label_files, output_dir=os.path.join(row.img.pipeline_sub_dir, row.img.output_sub_dir), name=row.img.filename_wo_ext+"_voted")))) .apply(axis=1, func=lambda row: row.img._replace(labels=row.voted_labels)) ) return Result(stages=s, output=segmented_imgs)
def maget_mask(imgs : List[MincAtom], maget_options, resolution : float, pipeline_sub_dir : str, atlases=None): s = Stages() original_imgs = imgs imgs = copy.deepcopy(imgs) original_imgs = pd.Series(original_imgs, index=[img.path for img in original_imgs]) for img in imgs: img.output_sub_dir = os.path.join(img.output_sub_dir, "masking") # TODO dereference maget_options -> maget_options.maget outside maget_mask call? if atlases is None: atlases = get_atlases(maget_options.maget, pipeline_sub_dir=pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options(maget_options.lsq12, LinearTransType.lsq12, resolution) #nlin_module = get_nonlinear_component(reg_method=options.mbm.nlin.reg_method) #masking_nlin_hierarchy = get_nonlinear_configuration_from_options(maget_options.maget.masking_nlin_protocol, # next(iter(maget_options.maget.flags_.masking_nlin_protocol)), # maget_options.maget.mask_method, # resolution) masking_nlin_component = get_nonlinear_component(reg_method=maget_options.maget.mask_method) algorithms = masking_nlin_component.Algorithms #masking_nlin_conf = (masking_nlin_component.parse_protocol_file( # maget_options.maget.masking_nlin_protocol, resolution=resolution) # if maget_options.maget.masking_nlin_protocol is not None # else masking_nlin_component.get_default_conf(resolution=resolution)) # TODO lift outside then delete #masking_imgs = copy.deepcopy(imgs) #for img in masking_imgs: # img.pipeline_sub_dir = os.path.join(img.pipeline_sub_dir, "masking") masking_alignments = pd.DataFrame({ 'img' : img, 'atlas' : atlas, 'xfm' : s.defer( lsq12_nlin(source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_options=maget_options.maget.masking_nlin_protocol, #masking_nlin_conf, resolution=resolution, nlin_module=masking_nlin_component, resample_source=False))} for img in imgs for atlas in atlases) # propagate a mask to each image using the above `alignments` as follows: # - for each image, voxel_vote on the masks propagated to that image to get a suitable mask # - run mincmath -clobber -mult <img> <voted_mask> to apply the mask to the files masked_img = ( masking_alignments .assign(resampled_mask=lambda df: df.apply(axis=1, func=lambda row: s.defer(algorithms.resample(img=row.atlas.mask, #apply(lambda x: x.mask), xfm=row.xfm.xfm, #apply(lambda x: x.xfm), like=row.img, invert=True, #interpolation=Interpolation.nearest_neighbour, postfix="-input-mask", subdir="tmp", # TODO annoying hack; fix mincresample(_mask) ...: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-input-mask" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), use_nn_interpolation=True )))) .groupby('img', as_index=False) .aggregate({'resampled_mask' : lambda masks: list(masks)}) .rename(columns={"resampled_mask" : "resampled_masks"}) .assign(voted_mask=lambda df: df.apply(axis=1, func=lambda row: # FIXME cannot use mincmath here !!! s.defer(mincmath(op="max", vols=sorted(row.resampled_masks), new_name="%s_max_mask" % row.img.filename_wo_ext, subdir="tmp")))) .apply(axis=1, func=lambda row: row.img._replace(mask=row.voted_mask))) # resample the atlas images back to the input images: # (note: this doesn't modify `masking_alignments`, but only stages additional outputs) masking_alignments.assign(resampled_img=lambda df: df.apply(axis=1, func=lambda row: s.defer(algorithms.resample( img=row.atlas, xfm=row.xfm.xfm, #.apply(lambda x: x.xfm), subdir="tmp", # TODO delete this stupid hack: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-resampled" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), like=row.img, invert=True)))) for img in masked_img: img.output_sub_dir = original_imgs.loc[img.path].output_sub_dir return Result(stages=s, output=masked_img)
def maget(imgs : List[MincAtom], options, prefix, output_dir): # FIXME prefix, output_dir aren't used !! s = Stages() maget_options = options.maget.maget resolution = options.registration.resolution # TODO or get_resolution_from_file(...) -- only if file always exists! pipeline_sub_dir = os.path.join(options.application.output_directory, options.application.pipeline_name + "_atlases") if maget_options.atlas_lib is None: raise ValueError("Need some atlases ...") # TODO should alternately accept a CSV file ... atlases = atlases_from_dir(atlas_lib=maget_options.atlas_lib, max_templates=maget_options.max_templates, pipeline_sub_dir=pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options(options.maget.lsq12, transform_type=LinearTransType.lsq12, file_resolution=resolution) nlin_hierarchy = get_nonlinear_configuration_from_options(options.maget.nlin.nlin_protocol, reg_method=options.maget.nlin.reg_method, file_resolution=resolution) if maget_options.mask or maget_options.mask_only: # this used to return alignments but doesn't currently do so masked_img = s.defer(maget_mask(imgs=imgs, maget_options=options.maget, atlases=atlases, pipeline_sub_dir=pipeline_sub_dir + "_masking", # FIXME repeats all alignments!!! resolution=resolution)) # now propagate only the masked form of the images and atlases: imgs = masked_img #atlases = masked_atlases # TODO is this needed? if maget_options.mask_only: # register each input to each atlas, creating a mask return Result(stages=s, output=masked_img) # TODO rename `alignments` to `registrations`?? else: if maget_options.mask: del masked_img # this `del` is just to verify that we don't accidentally use this later, since these potentially # coarser alignments shouldn't be re-used (but if the protocols for masking and alignment are the same, # hash-consing will take care of things), just the masked images they create; can be removed later # if a sensible use is found # images with labels from atlases # N.B.: Even though we've already registered each image to each initial atlas, this happens again here, # but using `nlin_hierarchy` instead of `masking_nlin_hierarchy` as options. # This is not 'work-efficient' in the sense that this computation happens twice (although # hopefully at greater precision the second time!), but the idea is to run a coarse initial # registration to get a mask and then do a better registration with that mask (though I'm not # sure exactly when this is faster than doing a single registration). # This _can_ allow the overall computation to finish more rapidly # (depending on the relative speed of the two alignment methods/parameters, # number of atlases and other templates used, number of cores available, etc.). atlas_labelled_imgs = ( pd.DataFrame({ 'img' : img, 'label_file' : s.defer( # can't use `label` in a pd.DataFrame index! mincresample_new(img=atlas.labels, xfm=s.defer(lsq12_nlin(source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_conf=nlin_hierarchy, resample_source=False)).xfm, like=img, invert=True, interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range',)))} for img in imgs for atlas in atlases) ) if maget_options.pairwise: def choose_new_templates(ts, n): # currently silly, but we might implement a smarter method ... # FIXME what if there aren't enough other imgs around?! This silently goes weird ... return pd.Series(ts[:n+1]) # n+1 instead of n: choose one more since we won't use image as its own template ... # FIXME: the --max-templates flag is ambiguously named ... should be --max-new-templates # (and just use all atlases) templates = pd.DataFrame({ 'template' : choose_new_templates(ts=imgs, n=maget_options.max_templates - len(atlases))}) # note these images are the masked ones if masking was done ... # the templates together with their (multiple) labels from the atlases (this merge just acts as a filter) labelled_templates = pd.merge(left=atlas_labelled_imgs, right=templates, left_on="img", right_on="template").drop('img', axis=1) # images with new labels from the templates imgs_and_templates = pd.merge(#left=atlas_labelled_imgs, left=pd.DataFrame({ "img" : imgs }).assign(fake=1), right=labelled_templates.assign(fake=1), on='fake') #left_on='img', right_on='template') # TODO do select here instead of below? template_labelled_imgs = ( imgs_and_templates .rename(columns={ 'label_file' : 'template_label_file' }) # don't register template to itself, since otherwise atlases would vote on that template twice .select(lambda ix: imgs_and_templates.img[ix].path != imgs_and_templates.template[ix].path) # TODO hardcoded name .assign(label_file=lambda df: df.apply(axis=1, func=lambda row: s.defer(mincresample_new(img=row.template_label_file, xfm=s.defer(lsq12_nlin(source=row.img, target=row.template, lsq12_conf=lsq12_conf, nlin_conf=nlin_hierarchy, resample_source=False)).xfm, like=row.img, invert=True, interpolation=Interpolation.nearest_neighbour, extra_flags=('-keep_real_range',))))) ) imgs_with_all_labels = pd.concat([atlas_labelled_imgs[['img', 'label_file']], template_labelled_imgs[['img', 'label_file']]], ignore_index=True) else: imgs_with_all_labels = atlas_labelled_imgs segmented_imgs = ( imgs_with_all_labels .groupby('img') .aggregate({'label_file' : lambda resampled_label_files: list(resampled_label_files)}) .rename(columns={ 'label_file' : 'label_files' }) .reset_index() .assign(voted_labels=lambda df: df.apply(axis=1, func=lambda row: s.defer(voxel_vote(label_files=row.label_files, output_dir=os.path.join(row.img.pipeline_sub_dir, row.img.output_sub_dir))))) .apply(axis=1, func=lambda row: row.img._replace(labels=row.voted_labels)) ) return Result(stages=s, output=segmented_imgs)
def maget_mask(imgs : List[MincAtom], maget_options, resolution : float, pipeline_sub_dir : str, atlases=None): s = Stages() resample = np.vectorize(mincresample_new, excluded={"extra_flags"}) defer = np.vectorize(s.defer) original_imgs = imgs imgs = copy.deepcopy(imgs) original_imgs = pd.Series(original_imgs, index=[img.path for img in original_imgs]) for img in imgs: img.output_sub_dir = os.path.join(img.output_sub_dir, "masking") # TODO dereference maget_options -> maget_options.maget outside maget_mask call? if atlases is None: if maget_options.maget.atlas_lib is None: raise ValueError("need some atlases for MAGeT-based masking ...") atlases = atlases_from_dir(atlas_lib=maget_options.maget.atlas_lib, max_templates=maget_options.maget.max_templates, pipeline_sub_dir=pipeline_sub_dir) lsq12_conf = get_linear_configuration_from_options(maget_options.lsq12, LinearTransType.lsq12, resolution) masking_nlin_hierarchy = get_nonlinear_configuration_from_options(maget_options.maget.masking_nlin_protocol, maget_options.maget.mask_method, resolution) # TODO lift outside then delete #masking_imgs = copy.deepcopy(imgs) #for img in masking_imgs: # img.pipeline_sub_dir = os.path.join(img.pipeline_sub_dir, "masking") masking_alignments = pd.DataFrame({ 'img' : img, 'atlas' : atlas, 'xfm' : s.defer(lsq12_nlin(source=img, target=atlas, lsq12_conf=lsq12_conf, nlin_conf=masking_nlin_hierarchy, resample_source=False))} for img in imgs for atlas in atlases) # propagate a mask to each image using the above `alignments` as follows: # - for each image, voxel_vote on the masks propagated to that image to get a suitable mask # - run mincmath -clobber -mult <img> <voted_mask> to apply the mask to the files masked_img = ( masking_alignments .assign(resampled_mask=lambda df: defer(resample(img=df.atlas.apply(lambda x: x.mask), xfm=df.xfm.apply(lambda x: x.xfm), like=df.img, invert=True, interpolation=Interpolation.nearest_neighbour, postfix="-input-mask", subdir="tmp", # TODO annoying hack; fix mincresample(_mask) ...: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-input-mask" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), extra_flags=("-keep_real_range",)))) .groupby('img', as_index=False) .aggregate({'resampled_mask' : lambda masks: list(masks)}) .rename(columns={"resampled_mask" : "resampled_masks"}) .assign(voted_mask=lambda df: df.apply(axis=1, func=lambda row: s.defer(mincmath(op="max", vols=sorted(row.resampled_masks), new_name="%s_max_mask" % row.img.filename_wo_ext, subdir="tmp")))) .apply(axis=1, func=lambda row: row.img._replace(mask=row.voted_mask))) # resample the atlas images back to the input images: # (note: this doesn't modify `masking_alignments`, but only stages additional outputs) masking_alignments.assign(resampled_img=lambda df: defer(resample(img=df.atlas, xfm=df.xfm.apply(lambda x: x.xfm), subdir="tmp", # TODO delete this stupid hack: #new_name_wo_ext=df.apply(lambda row: # "%s_to_%s-resampled" % (row.atlas.filename_wo_ext, # row.img.filename_wo_ext), # axis=1), like=df.img, invert=True))) for img in masked_img: img.output_sub_dir = original_imgs.ix[img.path].output_sub_dir return Result(stages=s, output=masked_img)