MASK = nib.load(MASK_PATH).get_data()
TRACT_MASKS = nib.load(TRACT_MASKS_PTH).get_data()
TRACT_MASKS = TRACT_MASKS[:, :, :, 1:] # remove background class
SAVE_PATH = join(DATA_DIR, SUBJ_ID, "pca_stats_unnorm.npz")
ncs = list(range(1, 30))
stats = {
"n_components": [],
"scores": []
}
for nc in ncs:
print("n_components = {}".format(nc).center(100, "-"))
features, _ = dsutils.make_pca_volume(DATA, MASK, n_components=nc, normalize=False)
train_slices = [('sagittal', 72), ('coronal', 87), ('axial', 72)]
train_masks = dsutils.create_data_masks(TRACT_MASKS, train_slices, LABELS)
X_train, y_train, train_coords = dsutils.create_dataset_from_data_mask(features,
train_masks,
labels=LABELS,
multi_label=True)
X_test, y_test, test_coords = dsutils.create_dataset_from_data_mask(features,
TRACT_MASKS,
labels=LABELS,
multi_label=True)
clf = RandomForestClassifier(n_estimators=100,
bootstrap=True,
oob_score=True,
random_state=0,
n_jobs=-1,
max_features='auto',
class_weight='balanced',
# transpose features
if orient == 'coronal':
orient_features = orient_features.transpose(1, 0, 2, 3)
if orient == 'axial':
orient_features = orient_features.transpose(1, 2, 0, 3)
TRACT_MASKS = nib.load(TRACT_MASKS_PTH).get_data()
TRACT_MASKS = TRACT_MASKS[:, :, :, 1:] # remove background class
FEATURES = orient_features
# ---------------------------------------------Train Set----------------------------------------------
train_slices = [('sagittal', 72), ('coronal', 87), ('axial', 72)]
train_masks = dsutils.create_data_masks(TRACT_MASKS,
train_slices,
LABELS,
verbose=False)
X_train, y_train, train_coords = dsutils.create_dataset_from_data_mask(
FEATURES, train_masks, labels=LABELS, multi_label=True)
if ADD_COORDS:
X_train = np.concatenate((X_train, train_coords), axis=1)
# --------------------------------------Random Forest Classifier--------------------------------------
clf = RandomForestClassifier(n_estimators=100,
bootstrap=True,
oob_score=True,
random_state=0,
n_jobs=-1,
max_features='auto',
class_weight='balanced',
max_depth=None,
x = data['data'].to(device)
feature = encoder(x)
orient_feature = feature.detach().cpu().squeeze().permute(1, 2, 0)
idx = int(data['file_name'][0][:-4][-3:])
features[idx] = orient_feature
features = features.numpy()
# transpose features
features = features.transpose(1, 0, 2, 3)
print("Features from epoch={} were created".format(epoch))
# -----------------------------------------Prepare train set------------------------------------------
print('Prepare train set'.center(100, '-'))
train_slices = [('sagittal', 72), ('coronal', 87), ('axial', 72)]
train_masks = dsutils.create_data_masks(ml_masks, train_slices, labels)
X_train, y_train, train_coords = dsutils.create_dataset_from_data_mask(features,
train_masks,
multi_label=True)
print("Trainset shape: ", X_train.shape)
# ------------------------------------------Prepare test set------------------------------------------
print('Prepare test set'.center(100, '-'))
print("Test set is the whole brain volume.")
# test_masks = ml_masks
test_slices = [('sagittal', 71), ('coronal', 86), ('axial', 71)]
test_masks = dsutils.create_data_masks(ml_masks, test_slices, labels)
X_test, y_test, test_coords = dsutils.create_dataset_from_data_mask(features,
test_masks,
multi_label=True)
print("Testset shape: ", X_test.shape)
dmri = nib.load(dmri_path)
print("dMRI shape: ", dmri.shape)
dmri_data = dmri.get_fdata()
# load masks
masks_path = os.path.join(data_dir, subj_id, 'tract_masks')
ml_masks = np.load(os.path.join(masks_path, 'multi_label_mask.npz'))['data']
ml_masks = ml_masks[:, :, :, 1:] # remove background class
print("Mask shape: ", ml_masks.shape)
labels = ['Other', 'CG', 'CST', 'FX', 'CC']
training_slices = [('sagittal', 72), ('coronal', 87), ('axial', 72)]
# make training data
train_masks = dsutils.create_data_masks(ml_masks, training_slices, labels)
X_train, y_train, train_coords = dsutils.create_dataset_from_data_mask(
dmri_data, train_masks, labels=labels, multi_label=True)
# make test data
test_masks = ml_masks
X_test, y_test, test_coords = dsutils.create_dataset_from_data_mask(
dmri_data, test_masks, labels=labels, multi_label=True)
X_train = np.hstack((train_coords, X_train))
X_test = np.hstack((test_coords, X_test))
print("X_train: ", X_train.shape)
print("X_test: ", X_test.shape)
num_features = list(range(1, 289))
best_f1 = 0
best_n = None
print("Preparing test sets.".center(100, "-"))
test_sets = {}
results = {}
orients_num_slices = {"axial": 145, "coronal": 174, "sagittal": 145}
orient_ranges = {
"axial": range(52, 93),
"coronal": range(67, 108),
"sagittal": range(52, 93)
}
for orient, num_slices in orients_num_slices.items():
for i in orient_ranges[orient]:
mask_name = (orient, i)
tmsk = dsutils.create_data_masks(TRACT_MASKS, [mask_name],
LABELS,
verbose=False)
X_test, y_test, test_coords = dsutils.create_dataset_from_data_mask(
FEATURES, tmsk, labels=LABELS, multi_label=True)
if X_test.shape[0] == 0:
continue
if ADD_COORDS:
X_test = np.concatenate((X_test, test_coords), axis=1)
test_sets[mask_name] = {}
test_sets[mask_name]["x"] = X_test
test_sets[mask_name]["y"] = y_test
test_sets[mask_name]["coords"] = test_coords
results[mask_name] = 1
