-
Notifications
You must be signed in to change notification settings - Fork 40
/
luna_p8b.py
149 lines (120 loc) · 6.64 KB
/
luna_p8b.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
import numpy as np
import data_transforms
import data_iterators
import pathfinder
import lasagne as nn
from collections import namedtuple
from functools import partial
import lasagne.layers.dnn as dnn
import theano.tensor as T
import utils
restart_from_save = None
rng = np.random.RandomState(42)
# transformations
p_transform = {'patch_size': (64, 64, 64),
'mm_patch_size': (64, 64, 64),
'pixel_spacing': (1., 1., 1.)
}
p_transform_augment = {
'translation_range_z': [-16, 16],
'translation_range_y': [-16, 16],
'translation_range_x': [-16, 16],
'rotation_range_z': [-180, 180],
'rotation_range_y': [-180, 180],
'rotation_range_x': [-180, 180]
}
# data preparation function
def data_prep_function(data, patch_center, luna_annotations, pixel_spacing, luna_origin, p_transform,
p_transform_augment, **kwargs):
x, patch_annotation_tf, annotations_tf = data_transforms.transform_patch3d(data=data,
luna_annotations=luna_annotations,
patch_center=patch_center,
p_transform=p_transform,
p_transform_augment=p_transform_augment,
pixel_spacing=pixel_spacing,
luna_origin=luna_origin)
x = data_transforms.pixelnormHU(x)
y = data_transforms.make_3d_mask_from_annotations(img_shape=x.shape, annotations=annotations_tf, shape='sphere')
return x, y
data_prep_function_train = partial(data_prep_function, p_transform_augment=p_transform_augment, p_transform=p_transform)
data_prep_function_valid = partial(data_prep_function, p_transform_augment=None, p_transform=p_transform)
# data iterators
batch_size = 4
nbatches_chunk = 8
chunk_size = batch_size * nbatches_chunk
train_valid_ids = utils.load_pkl(pathfinder.LUNA_VALIDATION_SPLIT_PATH)
train_pids, valid_pids = train_valid_ids['train'], train_valid_ids['valid']
train_data_iterator = data_iterators.PatchPositiveLunaDataGenerator(data_path=pathfinder.LUNA_DATA_PATH,
batch_size=chunk_size,
transform_params=p_transform,
data_prep_fun=data_prep_function_train,
rng=rng,
patient_ids=train_pids,
full_batch=True, random=True, infinite=True)
valid_data_iterator = data_iterators.ValidPatchPositiveLunaDataGenerator(data_path=pathfinder.LUNA_DATA_PATH,
transform_params=p_transform,
data_prep_fun=data_prep_function_valid,
patient_ids=valid_pids)
nchunks_per_epoch = train_data_iterator.nsamples / chunk_size
max_nchunks = nchunks_per_epoch * 30
validate_every = int(2. * nchunks_per_epoch)
save_every = int(0.5 * nchunks_per_epoch)
learning_rate_schedule = {
0: 1e-5,
int(max_nchunks * 0.4): 5e-6,
int(max_nchunks * 0.5): 2e-6,
int(max_nchunks * 0.85): 1e-6,
int(max_nchunks * 0.95): 5e-7
}
# model
conv3d = partial(dnn.Conv3DDNNLayer,
filter_size=3,
pad='valid',
W=nn.init.Orthogonal('relu'),
b=nn.init.Constant(0.0),
nonlinearity=nn.nonlinearities.identity)
max_pool3d = partial(dnn.MaxPool3DDNNLayer,
pool_size=2)
def conv_prelu_layer(l_in, n_filters):
l = conv3d(l_in, n_filters)
l = nn.layers.ParametricRectifierLayer(l)
return l
def build_model(patch_size=None):
patch_size = p_transform['patch_size'] if patch_size is None else patch_size
l_in = nn.layers.InputLayer((None, 1,) + patch_size)
l_target = nn.layers.InputLayer((None, 1,) + patch_size)
net = {}
base_n_filters = 128
net['contr_1_1'] = conv_prelu_layer(l_in, base_n_filters)
net['contr_1_2'] = conv_prelu_layer(net['contr_1_1'], base_n_filters)
net['contr_1_3'] = conv_prelu_layer(net['contr_1_2'], base_n_filters)
net['pool1'] = max_pool3d(net['contr_1_3'])
net['encode_1'] = conv_prelu_layer(net['pool1'], base_n_filters)
net['encode_2'] = conv_prelu_layer(net['encode_1'], base_n_filters)
net['encode_3'] = conv_prelu_layer(net['encode_2'], base_n_filters)
net['encode_4'] = conv_prelu_layer(net['encode_3'], base_n_filters)
net['upscale1'] = nn.layers.Upscale3DLayer(net['encode_4'], 2)
net['concat1'] = nn.layers.ConcatLayer([net['upscale1'], net['contr_1_3']],
cropping=(None, None, "center", "center", "center"))
net['dropout_1'] = nn.layers.dropout_channels(net['concat1'], p=0.15)
net['expand_1_1'] = conv_prelu_layer(net['dropout_1'], 2 * base_n_filters)
net['expand_1_2'] = conv_prelu_layer(net['expand_1_1'], 2 * base_n_filters)
net['expand_1_3'] = conv_prelu_layer(net['expand_1_2'], base_n_filters)
l_out = dnn.Conv3DDNNLayer(net['expand_1_3'], num_filters=1,
filter_size=1,
nonlinearity=nn.nonlinearities.sigmoid)
return namedtuple('Model', ['l_in', 'l_out', 'l_target'])(l_in, l_out, l_target)
def build_objective(model, deterministic=False, epsilon=1e-12):
network_predictions = nn.layers.get_output(model.l_out, deterministic=deterministic)[:, 0, :, :, :]
target_values = nn.layers.get_output(model.l_target)[:, 0, :, :, :]
network_predictions, target_values = nn.layers.merge.autocrop([network_predictions, target_values],
[None, 'center', 'center', 'center'])
y_true_f = target_values
y_pred_f = network_predictions
intersection = T.sum(y_true_f * y_pred_f, axis=(1, 2, 3))
dice_batch = (2. * intersection + epsilon) / (
T.sum(y_true_f, axis=(1, 2, 3)) + T.sum(y_pred_f, axis=(1, 2, 3)) + epsilon)
return -1. * T.sum(dice_batch)
def build_updates(train_loss, model, learning_rate):
updates = nn.updates.adam(train_loss, nn.layers.get_all_params(model.l_out), learning_rate)
return updates