import numpy as np
import nibabel as nb
import os
import sys
import nighresjava
from ..io import load_volume, save_volume
from ..utils import _output_dir_4saving, _fname_4saving, \
_check_topology_lut_dir, _check_available_memory
[docs]def fuzzy_cmeans(image, clusters=3, max_iterations=50, max_difference=0.01,
smoothing=0.1, fuzziness=2.0, mask_zero=True,
save_data=False, overwrite=False, output_dir=None,
file_name=None):
""" Fuzzy C-means image segmentation
Estimates intensity clusters with spatial smoothness and partial voluming.
Based on the RFCM algorithm of (Pham, 2001).
Parameters
----------
image: niimg
Input image to segment
clusters: int
Number of clusters to estimate (default is 3)
max_iterations: int
Maximum number of iterations to perform (default is 50)
max_difference: float
Maximum difference between steps for stopping (default is 0.01)
smoothing: float
Ratio of spatial smoothness to impose on the clusters (default is 0.1)
fuzziness: float
Scaling of the C-means measure, in [1.0 - 3.0] (default is 2.0)
mask_zero: bool
Whether to ignore zero values (default is true)
save_data: bool
Save output data to file (default is False)
overwrite: bool
Overwrite existing results (default is False)
output_dir: str, optional
Path to desired output directory, will be created if it doesn't exist
file_name: str, optional
Desired base name for output files with file extension
(suffixes will be added)
Returns
----------
dict
Dictionary collecting outputs under the following keys
(suffix of output files in brackets)
* memberships [niimg]: List of membership functions for each cluster in [0,1] (_rfcm-mem#cluster)
* classification (niimg): Hard classification of most likely cluster per voxel (_rfcm-class)
Notes
----------
Original Java module by Pierre-Louis Bazin.
"""
print('\nFuzzy C-means')
# make sure that saving related parameters are correct
if save_data:
output_dir = _output_dir_4saving(output_dir, image)
mem_files = []
for c in range(clusters):
mem_file = os.path.join(output_dir,
_fname_4saving(module=__name__,file_name=file_name,
rootfile=image,
suffix='rfcm-mem'+str(c+1), ))
mem_files.append(mem_file)
classification_file = os.path.join(output_dir,
_fname_4saving(module=__name__,file_name=file_name,
rootfile=image,
suffix='rfcm-class'))
if overwrite is False \
and os.path.isfile(classification_file):
missing = False
for mem_file in mem_files:
if not os.path.isfile(mem_file):
missing = True
if not missing:
print("skip computation (use existing results)")
output = {'classification': classification_file,
'memberships': mem_files}
return output
# start virtual machine, if not already running
try:
mem = _check_available_memory()
nighresjava.initVM(initialheap=mem['init'], maxheap=mem['max'])
except ValueError:
pass
# create instance
rfcm = nighresjava.FuzzyCmeans()
# set parameters
rfcm.setClusterNumber(clusters)
rfcm.setSmoothing(smoothing)
rfcm.setFuzziness(fuzziness)
rfcm.setMaxDist(max_difference)
rfcm.setMaxIter(max_iterations)
# load target image for parameters
print("load: "+str(image))
img = load_volume(image)
data = img.get_data()
affine = img.get_affine()
header = img.get_header()
resolution = [x.item() for x in header.get_zooms()]
dimensions = data.shape
if len(dimensions)>2: rfcm.setDimensions(dimensions[0], dimensions[1], dimensions[2])
else: rfcm.setDimensions(dimensions[0], dimensions[1], 1)
if len(resolution)>2: rfcm.setResolutions(resolution[0], resolution[1], resolution[2])
else: rfcm.setResolutions(resolution[0], resolution[1], resolution[1])
# image
rfcm.setImage(nighresjava.JArray('float')(
(data.flatten('F')).astype(float)))
# execute
try:
if mask_zero: rfcm.initZeroMaskImage()
else: rfcm.initBasicMaskImage()
rfcm.execute()
except:
# if the Java module fails, reraise the error it throws
print("\n The underlying Java code did not execute cleanly: ")
print(sys.exc_info()[0])
raise
return
# reshape output to what nibabel likes
classification_data = np.reshape(np.array(rfcm.getClassification(),
dtype=np.int32), dimensions, 'F')
header['cal_max'] = np.nanmax(classification_data)
classification = nb.Nifti1Image(classification_data, affine, header)
memberships = []
for c in range(clusters):
mem_data = np.reshape(np.array(rfcm.getMembership(c),
dtype=np.float32), dimensions, 'F')
header['cal_max'] = np.nanmax(mem_data)
membership = nb.Nifti1Image(mem_data, affine, header)
memberships.append(membership)
if save_data:
save_volume(classification_file, classification)
for c in range(clusters):
save_volume(mem_files[c], memberships[c])
output= {'classification': classification_file, 'memberships': mem_files}
else:
output= {'classification': classification, 'memberships': memberships}
return output