def _crop_neg(self, index):
N = len(self.neg_cubic) #疑似位置总数目
ids = np.random.randint(0, N)
neg_file = self.neg_cubic[ids]
data = np.load(neg_file) #疑似结点数据文件，字典，包含图像数据以及对应原图上的体素坐标中心
voxel_center = data['center']
patient_id = neg_file.split('/')[-1].split("_")[0]
mhd_file = opt.data_train + patient_id + '.mhd' #结点对应训练集的mhd文件
image_mask, nodule_centers, origin, spacing = make_mask(
mhd_file) #生成mask，并返回其原图的节点信息
img = data['data'].astype(np.float32)
crop_mask = self._crop(image_mask, voxel_center)
return img, crop_mask
def generate_data(tran, emit, T, miss=0., nmasks=1):
""" Generate synthetic data from finite HMM with continuous emissions.
A `miss` fraction of data is marked as missing. `nmasks` sets of
missing data are created.
tran : K x K discrete transition matrix.
emit : K x 1 vector of continuous emission distributions.
T : Number of observations to generate.
miss : Fraction of observations that are missing. Tries to pick evenly
over all states.
nmasks : Number of missing masks to generate.
"""
# doesn't matter for now because we're only dealing with one sequence of
# observations
K = tran.shape[0]
curr_st = 0
obs = []
sts = [0]
obs.append(emit[0].rvs()[0])
for i in xrange(T - 1):
# transition
# pass in probability of 1
curr_st = np.random.choice(K, p=tran[curr_st, :])
sts.append(curr_st)
# emit an observation
point = emit[curr_st].rvs()[0]
obs.append(point)
obs = np.array(obs)
sts = np.array(sts)
masks = None
if miss > 0.:
masks = list()
for i in xrange(nmasks):
masks.append(make_mask(sts, miss))
# Backwards compatability with old tests
if len(masks) == 1:
masks = masks[0]
return obs, sts, masks
def generate_data_smoothing(tran, emit, T, miss=0., left=0, nmasks=1):
""" Generate synthetic data from finite HMM with continuous emissions.
A `miss` fraction of data to the right of index `left` is marked as
missing. `nmasks` sets of missing data are created.
tran : K x K transition matrix (row normalized).
emit : 1-d array-like of size K containing continuous emissions.
T : Length of resulting observation sequence.
miss : Fraction of observations that are missing.
left : Index of leftmost observation that can be missing.
nmasks : Number of missing masks to generate.
"""
# doesn't matter for now because we're only dealing with one sequence of
# observations
K = tran.shape[0]
curr_st = 0
obs = []
sts = [0]
obs.append(emit[0].rvs()[0])
for i in xrange(T - 1):
# transition
# pass in probability of 1
curr_st = np.random.choice(K, p=tran[curr_st, :])
sts.append(curr_st)
# emit an observation
point = emit[curr_st].rvs()[0]
obs.append(point)
obs = np.array(obs)
sts = np.array(sts)
masks = None
if miss > 0.:
masks = list()
for i in xrange(nmasks):
masks.append(make_mask(sts, miss, left))
# Backwards compatability with old tests
if len(masks) == 1:
masks = masks[0]
return obs, sts, masks
def generate_data(tran, emit, T, miss=0., nmasks=1):
""" Generate synthetic data from finite HMM with continuous emissions.
A `miss` fraction of data is marked as missing. `nmasks` sets of
missing data are created.
tran : K x K discrete transition matrix.
emit : K x 1 vector of continuous emission distributions.
T : Number of observations to generate.
miss : Fraction of observations that are missing. Tries to pick evenly
over all states.
nmasks : Number of missing masks to generate.
"""
# doesn't matter for now because we're only dealing with one sequence of
# observations
K = tran.shape[0]
curr_st = 0
obs = []
sts = [0]
obs.append(emit[0].rvs()[0])
for i in xrange(T-1):
# transition
# pass in probability of 1
curr_st = np.random.choice(K, p=tran[curr_st,:])
sts.append(curr_st)
# emit an observation
point = emit[curr_st].rvs()[0]
obs.append(point)
obs = np.array(obs)
sts = np.array(sts)
masks = None
if miss > 0.:
masks = list()
for i in xrange(nmasks):
masks.append(make_mask(sts, miss))
# Backwards compatability with old tests
if len(masks) == 1:
masks = masks[0]
return obs, sts, masks
def generate_data_smoothing(tran, emit, T, miss=0., left=0, nmasks=1):
""" Generate synthetic data from finite HMM with continuous emissions.
A `miss` fraction of data to the right of index `left` is marked as
missing. `nmasks` sets of missing data are created.
tran : K x K transition matrix (row normalized).
emit : 1-d array-like of size K containing continuous emissions.
T : Length of resulting observation sequence.
miss : Fraction of observations that are missing.
left : Index of leftmost observation that can be missing.
nmasks : Number of missing masks to generate.
"""
# doesn't matter for now because we're only dealing with one sequence of
# observations
K = tran.shape[0]
curr_st = 0
obs = []
sts = [0]
obs.append(emit[0].rvs()[0])
for i in xrange(T-1):
# transition
# pass in probability of 1
curr_st = np.random.choice(K, p=tran[curr_st,:])
sts.append(curr_st)
# emit an observation
point = emit[curr_st].rvs()[0]
obs.append(point)
obs = np.array(obs)
sts = np.array(sts)
masks = None
if miss > 0.:
masks = list()
for i in xrange(nmasks):
masks.append(make_mask(sts, miss, left))
# Backwards compatability with old tests
if len(masks) == 1:
masks = masks[0]
return obs, sts, masks
def _crop_pos(self, index):
'''
剪切正样本，从保存的结点文件（64,64,64），剪切（48，48,48）的块
pos_file:/mnt/t/train_nodule_cubic/LKDS-00354_1021.npy,对应于一个结点
patient_id:LKDS-00354,为该结点对应的原始图像
nodule_index:1021 该结点在self.df_node中的index，便于取出该结点的中心
'''
file_index = int(index / opt.seg_ratio)
pos_file = self.pos_cubic[file_index]
nodule_index = int(pos_file.split('/')[-1].split('_')[1].split('.')[0])
df_min = self.df_node.iloc[nodule_index] #结点信息，包含他的中心世界坐标，文件名
patient_id = df_min.seriesuid
world_center = np.array([df_min.coordZ, df_min.coordY, df_min.coordX],
dtype=np.float32)
mhd_file = opt.data_train + patient_id + '.mhd' #结点对应训练集的mhd文件
image_mask, nodule_centers, origin, spacing = make_mask(
mhd_file) #生成mask，并返回其原图的节点信息
img = np.load(pos_file).astype(np.float32)
voxel_center = np.rint((world_center - origin) / spacing)
crop_mask = self._crop(image_mask, voxel_center)
return img, crop_mask
