def write2nifti(file_name, voxel_coord, labels):
"""Write label info into nifti files.
"""
labels = labels + 1
data = np.zeros((91, 109, 91), dtype=np.uint8)
for i in range(len(voxel_coord)):
data[voxel_coord[i][0],
voxel_coord[i][1],
voxel_coord[i][2]] = labels[i]
fsl_dir = os.getenv('FSL_DIR')
std_brain = os.path.join(fsl_dir, 'data', 'standard',
'MNI152_T1_2mm_brain.nii.gz')
header = nib.load(std_brain).get_header()
header['datatype'] = 4
mybase.save2nifti(data, header, file_name)
def make_mask(label_file_list, class_label, output_dir):
"""
Create a label mask derived from a group of label files.
"""
print 'Create a whole-fROI mask ...'
num = len(label_file_list)
for i in range(num):
label_data = nib.load(label_file_list[i]).get_data()
if not i:
mask_data = np.zeros(label_data.shape)
header = nib.load(label_file_list[i]).get_header()
uniq_val = np.unique(label_data)
for val in uniq_val:
if not val in class_label:
label_data[label_data == val] = 0
label_data[label_data > 0] = 1
mask_data += label_data
mask_data[mask_data > 0] = 1
# save to file
mask_file = os.path.join(output_dir, 'mask.nii.gz')
ntbase.save2nifti(mask_data, header, mask_file)
# return data
return mask_data
def make_mask(label_file_list, class_label, output_dir):
"""
Create a label mask derived from a group of label files.
"""
print 'Create a whole-fROI mask ...'
num = len(label_file_list)
for i in range(num):
label_data = nib.load(label_file_list[i]).get_data()
if not i:
mask_data = np.zeros(label_data.shape)
header = nib.load(label_file_list[i]).get_header()
uniq_val = np.unique(label_data)
for val in uniq_val:
if not val in class_label:
label_data[label_data==val] = 0
label_data[label_data>0] = 1
mask_data += label_data
mask_data[mask_data>0] = 1
# save to file
mask_file = os.path.join(output_dir, 'mask.nii.gz')
ntbase.save2nifti(mask_data, header, mask_file)
# return data
return mask_data
# compute Dice
for label_idx in cls_list[1:]:
P = pred_label == label_idx
T = test_label == label_idx
dice_val = mymath.dice_coef(T, P)
print 'Dice for label %s: %f'%(label_idx, dice_val)
if label_idx == 3:
temp_ffa_dice.append(dice_val)
else:
temp_ofa_dice.append(dice_val)
# save to nifti file
output_dir = os.path.join(base_dir, 'ma_202', 'gss_pred_file', 'r_fc')
file_name = os.path.join(output_dir, sessid[i]+'_'+str(num)+'.nii.gz')
mybase.save2nifti(pred_label, header, file_name)
ffa_dice.append(temp_ffa_dice)
ofa_dice.append(temp_ofa_dice)
out_file = 'r_ffa_output.txt'
f = open(out_file, 'w')
str_line = [str(item) for item in selected_num]
str_line = ','.join(str_line)
f.write(str_line + '\n')
for line in ffa_dice:
tmp_line = [str(item) for item in line]
tmp_line = ','.join(tmp_line)
f.write(tmp_line + '\n')
out_file = 'r_ofa_output.txt'
def get_posterior_map(sid_list,
data_dir,
class_label,
forest_list,
classes_list,
spatial_ptn,
save_nifti=True,
probabilistic=True):
"""
Get posterior map of each AF.
"""
mask_data = np.array(arlib.load_mask_coord(data_dir))
for i in range(len(sid_list)):
print 'AF from subject %s' % (sid_list[i])
# label the voxels within the mask using atlas forests (AFs)
clf = forest_list[i]
prob = clf.predict_proba(mask_data)
std_prob = np.zeros((prob.shape[0], len(class_label) + 1))
# TODO: construct a std prob table
for cls_idx in range(prob.shape[1]):
if classes_list[i][cls_idx] == 0:
std_prob[..., 0] = prob[..., cls_idx]
else:
tmp_idx = class_label.index(classes_list[i][cls_idx])
std_prob[..., tmp_idx + 1] = prob[..., cls_idx]
# generate final labeling
new_prob = np.zeros(std_prob.shape)
new_prob[range(new_prob.shape[0]), np.argmax(std_prob, axis=1)] = 1
tmp_pred_y = np.argmax(new_prob, axis=1)
pred_y = np.zeros(tmp_pred_y.shape)
for k in range(1, new_prob.shape[1]):
pred_y[tmp_pred_y == k] = class_label[k - 1]
# save posterior map as a nifti file
if save_nifti:
# predicted nifti directory
pred_dir = os.path.join(data_dir, 'posterior_maps')
if not os.path.exists(pred_dir):
os.system('mkdir ' + pred_dir)
fsl_dir = os.getenv('FSL_DIR')
img = nib.load(
os.path.join(fsl_dir, 'data', 'standard',
'MNI152_T1_2mm_brain.nii.gz'))
# save predicted label
header = img.get_header()
coords = mask_data
#pred_data = arlib.write2array(coords, pred_y)
#pred_data = np.around(pred_data)
if probabilistic:
pred_data = np.zeros((91, 109, 91, len(class_label) + 1))
for j in range(len(class_label) + 1):
pred_data[..., j] = np.copy(
arlib.write2array(coords, std_prob[..., j]))
out_file = os.path.join(pred_dir,
sid_list[i] + '_posterior.nii.gz')
else:
pred_data = np.copy(arlib.write2array(coords, pred_y))
out_file = os.path.join(pred_dir,
sid_list[i] + '_label.nii.gz')
mybase.save2nifti(pred_data, header, out_file)
if not probabilistic:
merged_file = os.path.join(pred_dir, 'merged_label.nii.gz')
str_cmd = ['fslmerge', '-a', merged_file]
for subj in sid_list:
temp = os.path.join(pred_dir, subj + '_label.nii.gz')
str_cmd.append(temp)
os.system(' '.join(str_cmd))
def predict(x_mtx,
atlas_num,
out_dir,
out_name,
class_label,
forest_list,
classes_list,
spatial_ptn,
save_nifti=True,
sorted=True,
single_atlas=False):
"""
Predict ROI label for unseen query image.
"""
# mask out voxels which are not activated significantly
x_mtx = np.array(x_mtx)
z_vtr = x_mtx[..., 0].copy()
z_vtr[z_vtr < 2.3] = 0
z_vtr[z_vtr > 0] = 1
smp_mask = z_vtr > 0
#test_x = x_mtx[smp_mask, 0:4]
test_x = x_mtx[smp_mask, 1:4]
# define similarity index
similarity = []
for i in range(len(forest_list)):
#r = normalized_mutual_info_score(spatial_ptn[..., i], z_vtr)
z_vtr = x_mtx[..., 0].copy()
z_vtr[z_vtr < 0] = 0
r = np.corrcoef(spatial_ptn[..., i], z_vtr)[0, 1]
similarity.append(r)
# sort the similarity
sorted_sim_idx = np.argsort(similarity)[::-1]
# label the activation voxels with atlas forests (AFs)
for num in atlas_num:
#print 'atlas number %s'%(num)
if sorted:
if not single_atlas:
selected_atlas = sorted_sim_idx[0:num]
else:
selected_atlas = sorted_sim_idx[(num - 1):num]
else:
selected_atlas = np.random.choice(len(sorted_sim_idx),
num,
replace=False)
pred_prob = None
for idx in selected_atlas:
clf = forest_list[idx]
prob = clf.predict_proba(test_x)
std_prob = np.zeros((prob.shape[0], len(class_label) + 1))
# TODO: construct a std prob table
for cls_idx in range(prob.shape[1]):
if classes_list[idx][cls_idx] == 0:
std_prob[..., 0] = prob[..., cls_idx]
else:
tmp_idx = class_label.index(classes_list[idx][cls_idx])
std_prob[..., tmp_idx + 1] = prob[..., cls_idx]
new_prob = np.zeros(std_prob.shape)
new_prob[range(new_prob.shape[0]), np.argmax(std_prob, axis=1)] = 1
# optional: modulate prediction probability with similarity
#new_prob = new_prob * similarity[clf_idx]
if not isinstance(pred_prob, np.ndarray):
pred_prob = new_prob
else:
pred_prob += new_prob
#print pred_prob.shape
tmp_pred_y = np.argmax(pred_prob, axis=1)
pred_y = np.zeros(tmp_pred_y.shape)
for k in range(1, pred_prob.shape[1]):
pred_y[tmp_pred_y == k] = class_label[k - 1]
# save predicted label as a nifti file
if save_nifti:
# predicted nifti directory
if not os.path.exists(out_dir):
os.system('mkdir ' + out_dir)
fsl_dir = os.getenv('FSL_DIR')
img = nib.load(
os.path.join(fsl_dir, 'data', 'standard',
'MNI152_T1_2mm_brain.nii.gz'))
# save predicted label
header = img.get_header()
#coords = test_x[..., 1:4]
coords = test_x
pred_data = arlib.write2array(coords, pred_y)
pred_data = np.around(pred_data)
out_file = os.path.join(out_dir, out_name)
mybase.save2nifti(np.around(pred_data), header, out_file)
def leave_one_out_test(sid_list,
atlas_num,
data_dir,
class_label,
forest_list,
classes_list,
spatial_ptn,
save_nifti=False,
sorted=True,
single_atlas=False):
"""
Evaluate classifier performance with leave-one-out scheme.
"""
# initial output dice value
dice = {}
for idx in class_label:
dice[idx] = []
for i in range(len(sid_list)):
print 'Test subject %s' % (sid_list[i])
test_data = get_subj_sample(sid_list[i], data_dir)
# mask out voxels which are not activated significantly
smp_mask = test_data[..., 0] >= 2.3
#test_x = test_data[smp_mask, 0:4]
test_x = test_data[smp_mask, 1:4]
test_y = test_data[smp_mask, -1]
# define similarity index
similarity = []
atlas_idx = []
for j in range(len(sid_list)):
if i == j:
continue
atlas_idx.append(j)
#r = normalized_mutual_info_score(spatial_ptn[..., i],
# spatial_ptn[..., j])
#r = mymath.mutual_information(spatial_ptn[..., i],
# spatial_ptn[..., j])
r = np.corrcoef(spatial_ptn[..., i], spatial_ptn[..., j])[0, 1]
similarity.append(r)
# sort the similarity
sorted_sim_idx = np.argsort(similarity)[::-1]
#print similarity
#print sorted_sim_idx
# label the activation voxels with atlas forests (AFs)
tmp_dice = {}
for idx in class_label:
tmp_dice[idx] = []
for num in atlas_num:
print 'atlas number %s' % (num)
if sorted:
if not single_atlas:
selected_atlas = sorted_sim_idx[0:num]
else:
selected_atlas = sorted_sim_idx[(num - 1):num]
else:
selected_atlas = np.random.choice(len(sorted_sim_idx),
num,
replace=False)
pred_prob = None
for idx in selected_atlas:
clf = forest_list[atlas_idx[idx]]
prob = clf.predict_proba(test_x)
std_prob = np.zeros((prob.shape[0], len(class_label) + 1))
# TODO: construct a std prob table
for cls_idx in range(prob.shape[1]):
if classes_list[atlas_idx[idx]][cls_idx] == 0:
std_prob[..., 0] = prob[..., cls_idx]
else:
tmp_idx = class_label.index(
classes_list[atlas_idx[idx]][cls_idx])
std_prob[..., tmp_idx + 1] = prob[..., cls_idx]
new_prob = np.zeros(std_prob.shape)
new_prob[range(new_prob.shape[0]),
np.argmax(std_prob, axis=1)] = 1
# optional: modulate prediction probability with similarity
#new_prob = new_prob * similarity[clf_idx]
if not isinstance(pred_prob, np.ndarray):
pred_prob = new_prob
else:
pred_prob += new_prob
#print pred_prob.shape
tmp_pred_y = np.argmax(pred_prob, axis=1)
pred_y = np.zeros(tmp_pred_y.shape)
for k in range(1, pred_prob.shape[1]):
pred_y[tmp_pred_y == k] = class_label[k - 1]
for label_idx in class_label:
P = pred_y == label_idx
T = test_y == label_idx
dice_val = mymath.dice_coef(T, P)
print 'Dice for label %s: %f' % (label_idx, dice_val)
tmp_dice[label_idx].append(dice_val)
# save predicted label as a nifti file
if save_nifti:
# predicted nifti directory
pred_dir = os.path.join(data_dir, str(num) + '_atlas_pred')
if not os.path.exists(pred_dir):
os.system('mkdir ' + pred_dir)
fsl_dir = os.getenv('FSL_DIR')
img = nib.load(
os.path.join(fsl_dir, 'data', 'standard',
'MNI152_T1_2mm_brain.nii.gz'))
# save predicted label
header = img.get_header()
#coords = test_x[..., 1:4]
coords = test_x
pred_data = arlib.write2array(coords, pred_y)
pred_data = np.around(pred_data)
out_file = os.path.join(pred_dir, sid_list[i] + '_pred.nii.gz')
mybase.save2nifti(pred_data, header, out_file)
for idx in class_label:
dice[idx].append(tmp_dice[idx])
for idx in class_label:
print 'Mean Dice for label %s: %s' % (idx, np.mean(dice[idx],
axis=0)[0])
return dice