#!/usr/bin/env python
import os
import shutil
import math
from glob import glob
import cv2
import csv
import random
from PIL import Image, ImageDraw
import dicom
import copy
import numpy as np
import SimpleITK as itk
from skimage import measure
import logging
import cPickle as pickle
# configuration options
DICOM_STRICT = False
SPACING = 0.8
GAP = 5
FAST = 400
if 'SPACING' in os.environ:
SPACING = float(os.environ['SPACING'])
print 'OVERRIDING SPACING = %f' % SPACING
if 'GAP' in os.environ:
GAP = int(os.environ['GAP'])
print 'OVERRIDING GAP = %d' % GAP
def get3c (images, i):
if i < GAP:
return None
if i + GAP >= images.shape[0]:
return None
a = images[i-GAP]
b = images[i]
c = images[i+GAP]
c3 = np.zeros(a.shape + (3,), dtype=np.float32)
c3[:,:,0] = a
c3[:,:,1] = b
c3[:,:,2] = c
return c3
#######################
def trim_loc (array1d, margin=0):
w = np.where(array1d > 0)
x0 = np.min(w)
x1 = np.max(w)+1
return max(0, x0-margin), min(x1+margin, array1d.shape[0])
def try_mkdir (path):
try:
os.makedirs(path)
except:
pass
def try_remove (path):
try:
os.remove(path)
except:
shutil.rmtree(path, ignore_errors=True)
pass
def chunks (l, n):
for i in range(0, len(l), n):
yield l[i:(i+n)]
ROOT = os.path.abspath(os.path.dirname(__file__))
DATA_DIR = os.path.join(ROOT, 'data')
def dicom_error (dcm, msg, level=logging.ERROR):
s = 'DICOM ERROR (%s): %s' % (dcm.filename, msg)
if DICOM_STRICT and level >= logging.ERROR:
raise Exception(s)
else:
logging.log(level, s)
pass
# stores example uid & label information
# Stage.train = [(uid, label)]
# Stage.test = [(uid, 0.5)]
# Stage.examples = [(uid, 0.5)]
def dcm_sanity_check (dcm):
rx, ry, rz, cx, cy, cz = [float(v) for v in dcm.ImageOrientationPatient]
pass
class DICOM:
def __init__ (self, dcm):
self.patient_id = dcm.PatientID
self.study_id = dcm.StudyInstanceUID
self.series_id = dcm.SeriesInstanceUID
self.HU = (float(dcm.RescaleSlope), float(dcm.RescaleIntercept))
# filename as slice ID
self.sid = os.path.splitext(os.path.basename(dcm.filename))[0]
self.dcm = dcm
self.image = dcm.pixel_array
self.shape = dcm.pixel_array.shape
self.pixel_padding = None
try:
self.pixel_padding = int(dcm.PixelPaddingValue)
except:
pass
from dicom.tag import Tag
#tag = Tag(0x0020,0x0032)
#print dcm[tag].value
#print dcm.ImagePositionPatient
#assert dcm[tag] == dcm.ImagePositionPatient
x, y, z = [float(v) for v in dcm.ImagePositionPatient]
self.position = (x, y, z)
rx, ry, rz, cx, cy, cz = [float(v) for v in dcm.ImageOrientationPatient]
self.ori_row = (rx, ry, rz)
self.ori_col = (cx, cy, cz)
x, y = [float(v) for v in dcm.PixelSpacing]
assert x == y
self.spacing = x
# Stage1: 4704 missing SliceLocation
try:
self.location = float(dcm.SliceLocation)
except:
dicom_error(dcm, 'Missing SliceLocation', level=logging.DEBUG)
self.location = self.position[2]
pass
self.bits = dcm.BitsStored
if False:
# Non have SliceThickness
tag = Tag(0x0018, 0x0050)
if not tag in dcm:
dicom_error(dcm, 'Missing SliceThickness', level=logging.WARN)
else:
logging.info('Has SliceThickness: %s' % dcm.filename)
self.thickness = float(dcm[tag].value)
# ???, why the value is as big as 63536
if False:
# Stage1 data:
# 4704 have padding, 126057 not, so skip this
self.padding = None
try:
self.padding = dcm.PixelPaddingValue
except:
dicom_error(dcm, 'Missing PixelPaddingValue', level=logging.WARN)
pass
# sanity check
#if dcm.PatientName != dcm.PatientID:
# dicom_error(dcm, 'PatientName is not dcm.PatientID')
if dcm.Modality != 'CT':
dicom_error(dcm, 'Bad Modality: ' + dcm.Modality)
#if Tag(0x0008,0x103e) in dcm:
if False:
if dcm.SeriesDescription != 'Axial' and dcm.SeriesDescription != 'mediastinal_lymph_nodes' and dcm.SeriesDescription != 'Recon 2: ACRIN LARGE' and dcm.SeriesDescription != 'Recon 3: CHEST-ABD':
dicom_error(dcm, 'Bad SeriesDescription: ' + dcm.SeriesDescription)
#if Tag(0x0008,0x0008) in dcm:
# if not 'AXIAL' in ' '.join(list(dcm.ImageType)).upper():
# dicom_error(dcm, 'Bad image type: ' + list(dcm.ImageType))
ori_type_tag = Tag(0x0010,0x2210)
if ori_type_tag in dcm:
ori_type = dcm[ori_type_tag].value
if 'BIPED' != ori_type:
dicom_error(dcm, 'Bad Anatomical Orientation Type: ' + ori_type)
# location should roughly be position.z
self.funny_slice_location = abs(self.position[2] - self.location) > 10
x, y, z = self.ori_row # should be (1, 0, 0)
if x < 0.9:
dicom_error(dcm, 'Bad row orientation')
x, y, z = self.ori_col # should be (0, 1, 0)
if y < 0.9:
dicom_error(dcm, 'Bad col orientation')
pass
pass
def segment_lung_axial (image, th=123.85, dilate=0.01):
blur = np.copy(image)
for i in range(blur.shape[0]):
cv2.blur(blur[i], (5,5), blur[i])
binary = np.array(blur < th, dtype=np.uint8)
# 0: body
# 1: air & lung
labels = measure.label(binary, background=-1)
# set air (same cc as corners) -> body
bg_labels = set()
for z in [0, -1]:
for y in [0, -1]:
for x in [0, -1]:
bg_labels.add(labels[z, y, x])
bg_labels = list(bg_labels)
print(bg_labels)
if len(bg_labels) > 1:
logging.warn('bg not connected, detected %d components' % len(bg_labels))
pass
for bg_label in bg_labels:
binary[bg_label == labels] = 0
pass
# now binary:
# 0: non-lung & body tissue in lung
# 1: lung & holes in body
for i, sl in enumerate(binary):
#H, W = sl.shape
ll = measure.label(sl, background=-1) # connected components
# biggest CC should be body
vv, cc = np.unique(ll, return_counts=True)
assert len(vv) > 0
body_ll = vv[np.argmax(cc)]
binary[i][ll != body_ll] = 1
pass
# set corner again
labels = measure.label(binary, background=0)
bg_labels = set([0])
for z in [0, -1]:
for y in [0, -1]:
for x in [0, -1]:
bg_labels.add(labels[z, y, x])
val_counts = zip(*np.unique(labels, return_counts=True))
val_counts = [x for x in val_counts if not x[0] in bg_labels] # remove background
val_counts = sorted(val_counts, key=lambda x:-x[1]) # sort by size
th = val_counts[0][1] /4 # 1/4 size of the larged connected component (must be lung)
val = [v for v, c in val_counts if c >= th]
if len(val) >= 3:
logging.warn('more than 2 lungs parts detected %d' % len(val))
binary = np.zeros_like(binary, dtype=np.uint8)
for v in val:
binary[labels == v] = 1
H, W = binary[0].shape
dilate = int(round(math.sqrt(1.0 * H * W) * dilate))
#print("DILATE: ", dilate)
kernel = np.ones((dilate, dilate), dtype=np.int32)
for i in range(binary.shape[0]):
cv2.dilate(binary[i], kernel, binary[i])
#image[binary == 0] = 255
#image = 255 - image
#image[binary == 0] = 255
return binary
#return image #* binary.astype(image.dtype)
AXIAL = 0
SAGITTAL = 1
CORONAL = 2
VIEWS = [AXIAL, SAGITTAL, CORONAL]
VIEW_NAMES = ['axial', 'sagittal', 'coronal']
AXES_ORDERS = ([0, 1, 2], # AXIAL
[2, 1, 0], # SAGITTAL
[1, 0, 2]) # CORONAL
def index_view (I, view):
assert len(I) == 3
a, b, c = AXES_ORDERS[view]
return [I[a], I[b], I[c]]
def strip_pad_512 (n, size=512):
if n >= size:
from_x = (n-size)/2
to_x = 0
n_x = size
shift_x = from_x
else:
from_x = 0
to_x = (size-n)/2
n_x = n
shift_x = -to_x
return from_x, to_x, n_x, shift_x
class CaseBase (object):
# self.images
# self.spacing
# self.origin # !!! origin is never transposed!!!
# self.axes
# self.vspacing
def __init__ (self):
self.uid = None
self.path = None
self.images = None # 3-D array
self.spacing = None #
self.origin = None # origin never changes
# under transposing
self.view = None
self.anno = None
# We save the coefficients for normalize to
# Hounsfield Units, and keep that updated
# when normalizing
self.HU = None # (intercept, slope)
self.dcm_z_position = None
self.orig_origin = None
self.orig_spacing = None
self.orig_shape = None
self.pixel_padding = None
pass
def copy_replace_images (self, images):
case = CaseBase()
case.uid = self.uid
case.orig_origin = self.orig_origin
case.orig_spacing = self.orig_spacing
case.orig_shape = self.orig_shape
case.path = self.path
case.images = images
case.spacing = self.spacing
case.view = self.view
case.origin = self.origin
case.anno = self.anno
return case
def normalizeHU (self):
assert not self.HU is None
a, b = self.HU
self.images *= a
self.images += b
self.HU = (1.0, 0)
if not self.pixel_padding is None:
self.pixel_padding = self.pixel_padding * a + b
pass
def transpose_array (self, view, array):
if self.view == view:
return array
elif self.view == AXIAL and view == SAGITTAL:
d1, d2 = 0, 2
elif self.view == AXIAL and view == CORONAL:
d1, d2 = 0, 1
elif self.view == SAGITTAL and view == AXIAL:
d1, d2 = 0, 2
elif self.view == CORONAL and view == AXIAL:
d1, d2 = 0, 1
else:
assert False
return np.swapaxes(array, d1, d2)
def transpose (self, view):
if self.view == view:
return self
elif self.view == AXIAL and view == SAGITTAL:
d1, d2 = 0, 2
elif self.view == AXIAL and view == CORONAL:
d1, d2 = 0, 1
elif self.view == SAGITTAL and view == AXIAL:
d1, d2 = 0, 2
elif self.view == CORONAL and view == AXIAL:
d1, d2 = 0, 1
else:
assert False
case = CaseBase()
case.uid = self.uid
case.orig_origin = self.orig_origin
case.orig_spacing = self.orig_spacing
case.orig_shape = self.orig_shape
case.path = self.path
case.images = np.swapaxes(self.images, d1, d2)
assert isinstance(self.spacing, tuple)
sp = list(self.spacing)
sp[d1], sp[d2] = sp[d2], sp[d1]
case.spacing = tuple(sp)
case.view = view
case.origin = self.origin
case.anno = self.anno
return case
def round512 (self, size=512):
target = np.zeros((size,size,size), dtype=self.images.dtype)
Z, Y, X = self.images.shape
from_z, to_z, n_z, shift_z = strip_pad_512(Z, size=size)
from_y, to_y, n_y, shift_y = strip_pad_512(Y, size=size)
from_x, to_x, n_x, shift_x = strip_pad_512(X, size=size)
target[to_z:(to_z+n_z),
to_y:(to_y+n_y),
to_x:(to_x+n_x)] = self.images[from_z:(from_z+n_z),
from_y:(from_y+n_y),
from_x:(from_x+n_x)]
self.origin[0] += shift_z * self.spacing[0]
self.origin[1] += shift_y * self.spacing[1]
self.origin[2] += shift_x * self.spacing[2]
print("off", to_x, to_y, to_z)
print("len", n_x, n_y, n_z)
print("shi", shift_x, shift_y, shift_z)
self.images = target
pass
def strip (self, mask, margin1=2, margin2=10):
z0, z1 = trim_loc(np.sum(mask, axis=(1,2)), margin=margin1)
y0, y1 = trim_loc(np.sum(mask, axis=(0,2)), margin=margin2)
x0, x1 = trim_loc(np.sum(mask, axis=(0,1)), margin=margin2)
self.origin[0] += z0 * self.spacing[0]
self.origin[1] += y0 * self.spacing[1]
self.origin[2] += x0 * self.spacing[2]
self.images = self.images[z0:z1, y0:y1, x0:x1]
pass
def round_stride (self, stride=16):
T, H, W = self.images.shape[:3]
nT = T / stride * stride
nH = H / stride * stride
nW = W / stride * stride
oT = (T - nT)/2
oH = (H - nH)/2
oW = (W - nW)/2
self.origin[0] += oT * self.spacing[0]
self.origin[1] += oH * self.spacing[1]
self.origin[2] += oW * self.spacing[2]
self.images = self.images[oT:(oT+nT),oH:(oH+nH),oW:(oW+nW)]
return oT, oH, oW
pass
# consider using scipy.ndimage.interpolation
def rescale (self, slices = None, spacing = None, size = None, method=2):
# if slices != self.images.shape[0], use method:
# 0: adjust slices, so everything is integer and no rounding or approx. is done
# 1: do not change slices, use nearest neighbor
# 2: do not change slices, use interpolation
N, H, W = self.images.shape
case = CaseBase()
case.uid = self.uid
case.orig_origin = self.orig_origin
case.orig_spacing = self.orig_spacing
case.orig_shape = self.orig_shape
case.path = self.path
case.view = self.view
case.origin = self.origin
case.anno = self.anno
case.HU = self.HU
assert (spacing and not size) or (size and not spacing)
if (not slices) or (slices == N):
method = 0
slices = N
step = 1
off = 0
sp1 = self.spacing[0]
elif method == 0: # TODO: need to do actual samping
# origin under this is not correct due to non-0 off
step = int(round(N / slices))
slices = N / step
off = (N - slices * step) / 2
sp1 = self.spacing[0] * step
else:
off = 0
step = float(N -1)/ (slices - 1)
sp1 = self.spacing[0] * step
pass
if spacing:
H = int(round((H-1) * self.spacing[1] / spacing + 1))
W = int(round((W-1) * self.spacing[2] / spacing + 1))
resize = (W, H)
sp2 = spacing
sp3 = spacing
elif size:
sp2 = self.spacing[1] * (H-1) / (size-1)
sp3 = self.spacing[2] * (W-1) / (size-1)
resize = (size, size)
H = size
W = size
else:
resize = None
_, sp2, sp3 = self.spacing
case.spacing = (sp1, sp2, sp3)
case.images = np.zeros((slices, H, W), dtype=np.float32)
for i in range(slices):
if method == 0 or method == 1:
arr = int(round(off))
image = self.images[arr, :, :]
elif method == 2:
L = int(math.floor(off))
R = int(math.ceil(off))
if R <= 0:
image = self.images[0, :, :]
elif L >= N-1:
image = self.images[N-1, :, :]
elif R - L < 0.5: # R == L
image = self.images[L, :, :]
else:
image = (self.images[L, :, :] * (R - off) + self.images[R, :, :] * (off - L)) / (R - L)
pass
if resize:
cv2.resize(image, resize, case.images[i, :, :])
else:
case.images[i, :, :] = image
off += step
return case
def rescale3D (self, spacing):
slices = int(round(self.spacing[0] * (self.images.shape[0] - 1) / spacing + 1))
return self.rescale(slices, spacing, size=None, method=2)
pass
def normalize (self, min=0, max=1, min_th = -1000, max_th = 400):
assert self.images.dtype == np.float32
if not min_th is None:
self.images[self.images < min_th] = min_th
if not max_th is None:
self.images[self.images > max_th] = max_th
m = min_th #np.min(self.images)
M = max_th #np.max(self.images)
scale = (1.0 * max - min)/(M - m)
logging.debug('norm %f %f' % (m, M))
self.images -= m
self.images *= scale
self.images += min
# recalculate HU
# I: original image
# I': new image
# a'I' + b' = aI + b
# I' = (I-m) * scale + min
# = I*scale + (min - m * scale)
# so
# a'I*scale + (min - m * scale)*a' + b' = aI + b
#
# a' = a / scale
# b' = b + a'(m * scale -min)
# = b + a * (m - min/scale)
if self.HU:
a, b = self.HU
#self.HU = (a * (M -m), b + a * m)
self.HU = (a / scale, b + a * (m - min/scale))
pass
def standardize_color (self):
self.normalizeHU()
self.normalize(min_th=-1000,max_th=400,min=0,max=255)
pass
def standardize_color16 (self):
self.normalizeHU()
self.normalize(min_th=-1000,max_th=400,min=0,max=1400)
pass
# return center coordinate
def world_to_vox (self, world):
# change view
# change origin
z, y, x, r = world
z0, y0, x0 = self.origin
cc = (np.array(world[:3])-np.array(self.origin))
cc = cc[AXES_ORDERS[self.view]]
spacing = np.array(self.spacing)
cc = cc / spacing
rr = r / spacing
#print "xxx", cc[0], rr[0]
return cc, rr
def picpac_anno (self):
# !!! annotation is center & radius instead of orign + size in picpac
if self.anno is None:
return []
ALL = []
nodules = [ self.world_to_vox(world) for world in self.anno]
C, H, W = self.images.shape
for (z, y, x), (zr, yr, xr) in nodules:
first = max(0, int(math.ceil(z - zr)))
last = min(C-1, int(math.floor(z + zr)))
if first > last:
continue
nod = []
x /= W
y /= H
for i in range(first, last + 1):
cos = abs(i - z) / zr
sin = math.sqrt(1 - cos * cos)
cyr = yr * sin / H
cxr = xr * sin / W
nod.append([i, x, y, cxr, cyr])
#print 'ellipse', x, y, xr, yr
#pass
pass
ALL.append(nod)
pass
return ALL
def papaya_box (self, box):
out = [0]*6
assert self.view == AXIAL
for i in range(3):
out[i] = int(round((self.origin[i] + self.spacing[i] * box[i] - self.orig_origin[i]) / self.orig_spacing[i]))
out[i+3] = int(round((self.origin[i] + self.spacing[i] * box[i+3] - self.orig_origin[i]) / self.orig_spacing[i]))
pass
D, _, W = self.orig_shape
out[0], out[3] = D-out[3], D-out[0]
out[2], out[5] = W-out[5], W-out[2]
return out
def save_gif (self, path, anno=False, aug=2, step=1):
# must normalize first to [0, 1]
cube = np.uint8(np.clip(self.images, 0, 255))
frames = [Image.fromarray(cube[i,:,:]) for i in range(0, cube.shape[0], step)]
if anno:
C, H, W = self.images.shape
annos = self.picpac_anno()
for nodule in annos:
for j, x, y, rx, ry in nodule:
x *= W
y *= H
rx *= W * aug
ry *= H * aug
draw = ImageDraw.Draw(frames[j])
draw.ellipse([math.floor(x-rx),
math.floor(y-ry),
math.ceil(x+rx),
math.ceil(y+ry)], outline=255)
del draw
pass
frames[0].save(path, save_all=True, append_images=frames[1:], duration=0.1, loop=0)
pass
def group_zrange (dcms):
zs = [float(dcm.dcm.ImagePositionPatient[2]) for dcm in dcms]
zs = sorted(zs)
gap = 1000000
if len(zs) > 1:
gap = zs[1] - zs[0]
return (zs[0], zs[-1], gap)
def regroup_dcms (dcms):
acq_groups = {}
for dcm in dcms:
an = 0
try:
an = int(dcm.dcm.AcquisitionNumber)
except:
pass
acq_groups.setdefault(an, []).append(dcm)
pass
groups = acq_groups.values()
if len(groups) == 1:
return groups[0]
# we have multiple acquisitions
zrs = [group_zrange(group) for group in groups]
zrs = sorted(zrs, key=lambda x: x[0])
min_gap = min([zr[2] for zr in zrs])
gap_th = 2.0 * min_gap
prev = zrs[0]
bad = False
for zr in zrs[1:]:
gap = zr[0] - prev[1]
if gap < 0 or gap > gap_th:
bad = True
break
if gap != min_gap:
logging.error('bad gap')
prev = zr
if not bad:
logging.error('multiple acquisitions merged')
return dcms
# return the maximal groups
gs = max([len(group) for group in groups])
acq_groups = {k:v for k, v in acq_groups.iteritems() if len(v) == gs}
key = max(acq_groups.keys())
group = acq_groups[key]
print(acq_groups.keys(), key)
logging.error('found conflicting groups. keeping max acq number, %d out of %d dcms' % (len(group), len(dcms)))
return group
# All DiCOMs of a UID, organized
class FsCase (CaseBase):
def __init__ (self, path, regroup = True):
CaseBase.__init__(self)
self.path = path
#self.thumb_path = os.path.join(DATA_DIR, 'thumb', uid)
# load path
dcms = []
for dcm_path in glob(os.path.join(self.path, '*.dcm')):
dcm = dicom.read_file(dcm_path)
try:
boxed = DICOM(dcm)
except:
print dcm.filename
raise
dcms.append(boxed)
assert dcms[0].spacing == boxed.spacing
assert dcms[0].shape == boxed.shape
assert dcms[0].ori_row == boxed.ori_row
assert dcms[0].ori_col == boxed.ori_col
if dcms[0].pixel_padding != boxed.pixel_padding:
logging.warn('0 padding %s, but now %s, %s' %
(dcms[0].pixel_padding, boxed.pixel_padding, dcm.filename))
#assert dcms[0].HU == boxed.HU
#print boxed.HU
pass
assert len(dcms) >= 2
if regroup:
dcms = regroup_dcms(dcms)
self.pixel_padding = dcms[0].pixel_padding
dcms.sort(key=lambda x: x.position[2])
zs = []
for i in range(1, len(dcms)):
zs.append(dcms[i].position[2] - dcms[i-1].position[2])
pass
zs = np.array(zs)
z_spacing = np.mean(zs)
assert z_spacing > 0
assert np.max(np.abs(zs - z_spacing)) * 1000 < z_spacing
#self.length = dcms[-1].position[2] - dcms[0].position[2]
front = dcms[0]
#self.sizes = (front.shape[0] * front.spacing, front.shape[1] * front.spacing, self.length)
self.dcms = dcms
images = np.zeros((len(dcms),)+front.image.shape, dtype=np.float32)
HU = front.HU
for i in range(len(dcms)):
HU2 = dcms[i].HU
images[i,:,:] = dcms[i].image
if HU2 != HU:
logging.warn("HU: (%d) %s => %s, %s" % (i, HU2, HU, dcms[i].dcm.filename))
images[i, :, :] *= HU2[0] / HU[0]
images[i, :, :] += (HU2[1] - HU[1])/HU[0]
self.dcm_z_position = {}
for dcm in dcms:
name = os.path.splitext(os.path.basename(dcm.dcm.filename))[0]
self.dcm_z_position[name] = dcm.position[2] - front.position[2]
pass
# spacing # z, y, x
self.images = images
self.spacing = (z_spacing, front.spacing, front.spacing)
x, y, z = front.position
self.origin = [z, y, z] #front.location
self.view = AXIAL
self.anno = None
self.HU = HU
self.orig_origin = copy.deepcopy(self.origin)
self.orig_spacing = copy.deepcopy(self.spacing)
self.orig_shape = copy.deepcopy(self.images.shape)
# sanity check
pass
pass
class Case:
def __init__ (self, uid, regroup = True):
self.uid = uid
self.path = os.path.join(DATA_DIR, 'bowl', uid)
if not os.path.exists(self.path):
self.path = os.path.join(DATA_DIR, 'samples', uid)
if not os.path.exists(self.path):
cc = glob(os.path.join(DATA_DIR, 'lymph', 'data', uid, '*/*'))
if len(cc) >= 1:
self.path = cc[0]
assert os.path.exists(self.path)
if len(cc) > 1:
logging.warn('multiple candidates for ' + uid)
else:
cc = glob(os.path.join(DATA_DIR, 'lymph', 'data', '*/*', uid))
if len(cc) >= 1:
self.path = cc[0]
assert os.path.exists(self.path)
if len(cc) > 1:
logging.warn('multiple candidates for ' + uid)
else:
raise Exception('data not found for uid %s' % uid)
pass
FsCase.__init__(self, self.path, regroup)
pass
pass
LUNA_DIR = os.path.join(ROOT, 'data', 'luna')
#LUNA_DIR = os.path.join('data', 'luna')
def load_luna_dir_layout ():
lookup = {}
for i in range(10):
sub = os.path.join(LUNA_DIR, 'subset%d' % i)
for f in glob(os.path.join(sub, '*.mhd')):
bn = os.path.splitext(os.path.basename(f))[0]
#print bn, "=>", sub
lookup[bn] = sub
pass
pass
return lookup
def load_luna_csv (filename):
lines = []
with open(filename, "rb") as f:
csvreader = csv.reader(f)
for line in csvreader:
lines.append(line)
return lines
pass
pass
def load_luna_annotations ():
ALL = {}
with open(os.path.join(LUNA_DIR, 'CSVFILES', 'annotations.csv'), 'r') as f:
f.next()
for l in f:
#print l
uid, x, y, z, d = l.strip().split(',')
x = float(x)
y = float(y)
z = float(z)
r = float(d)/2
ALL.setdefault(uid, []).append((z, y, x, r))
pass
pass
return ALL
def load_luna_meta ():
cache_path = os.path.join(LUNA_DIR, 'meta.pkl')
if os.path.exists(cache_path):
with open(cache_path, 'rb') as f:
return pickle.load(f)
logging.warn('loading luna meta data')
meta = (load_luna_dir_layout(),
load_luna_csv(os.path.join(LUNA_DIR, 'CSVFILES', 'candidates.csv')),
load_luna_annotations())
with open(cache_path, 'wb') as f:
pickle.dump(meta, f)
return meta
#LUNA_DIR_LOOKUP, _, LUNA_ANNO = load_luna_meta()
LUNA_DIR_LOOKUP = {}
LUNA_ANNO = {}
def worldToVoxelCoord(worldCoord, origin, spacing):
stretchedVoxelCoord = np.absolute(worldCoord - origin)
voxelCoord = stretchedVoxelCoord / spacing
return voxelCoord
# All DiCOMs of a UID, organized
class LunaCase (CaseBase):
def __init__ (self, uid):
CaseBase.__init__(self)
self.uid = uid
self.path = os.path.join(LUNA_DIR_LOOKUP[uid], uid + '.mhd')
if not os.path.exists(self.path):
raise Exception('data not found for uid %s at %s' % (uid, self.path))
pass
#self.thumb_path = os.path.join(DATA_DIR, 'thumb', uid)
# load path
itkimage = itk.ReadImage(self.path)
self.HU = (1.0, 0.0)
self.images = itk.GetArrayFromImage(itkimage).astype(np.float32)
#print type(self.images), self.images.dtype
self.origin = list(reversed(itkimage.GetOrigin()))
self.spacing = list(reversed(itkimage.GetSpacing()))
self.view = AXIAL
_, a, b = self.spacing
self.anno = LUNA_ANNO.get(uid, None)
assert a == b
# sanity check
pass
pass
def save_mask (path, mask):
shape = np.array(list(mask.shape), dtype=np.uint32)
total = mask.size
totalx = (total +7 )/ 8 * 8
if totalx == total:
padded = mask
else:
padded = np.zeros((totalx,), dtype=np.uint8)
padded[:total] = np.reshape(mask, (total,))
pass
padded = np.reshape(padded, (totalx/8, 8))
print padded.shape
packed = np.packbits(padded)
print packed.shape
np.savez_compressed(path, shape, packed)
pass
def load_mask (path):
import sys
saved = np.load(path)
shape = saved['arr_0']
D, H, W = shape
size = D * H * W
packed = saved['arr_1']
padded = np.unpackbits(packed)
binary = padded[:size]
return np.reshape(binary, [D, H, W])
def is_kaggle (uid):
return len(uid) == 32
def load_case (uid):
if is_kaggle(uid):
return Case(uid)
else:
return LunaCase(uid)
pass
def load_8bit_lungs (uid):
#path = os.path.join('data/cache', uid)
#if os.path.exists(path):
# with open(path, 'rb') as f:
# return pickle.load(f)
case = load_case(uid)
case.standardize_color()
cache = os.path.join('maskcache/mask-123.85-0.01/%s.npz' % case.uid)
binary = None
if os.path.exists(cache) and os.path.getsize(cache) > 0:
# load cache
binary = load_mask(cache)
assert not binary is None
if binary is None:
binary = segment_lung_axial(case.images) #, th=200.85)
save_mask(cache, binary)
pass
case.images[binary==0] = 255
case.images *= -1
case.images += 255
#case = case.rescale3D(1.0)
#with open(path, 'wb') as f:
# pickle.dump(case, f)
#return case
return case
def load_8bit_lungs_noseg (uid):
#path = os.path.join('data/cache', uid)
#if os.path.exists(path):
# with open(path, 'rb') as f:
# return pickle.load(f)
case = load_case(uid)
case.standardize_color()
#case.images = segment_lung_axial(case.images) #, th=200.85)
#case.images *= -1
#case.images += 255
#case = case.rescale3D(1.0)
#with open(path, 'wb') as f:
# pickle.dump(case, f)
#return case
return case
def load_16bit_lungs_noseg (uid):
case = load_case(uid)
case.standardize_color16()
return case
def segment_lung_axial_v2 (image, th):
blur = np.copy(image)
for i in range(blur.shape[0]):
cv2.blur(blur[i], (5,5), blur[i])
binary = np.array(blur < th, dtype=np.uint8)
# 0: body
# 1: air & lung
labels = measure.label(binary, background=-1)
# set air (same cc as corners) -> body
bg_labels = set()
for z in [0, -1]:
for y in [0, -1]:
for x in [0, -1]:
bg_labels.add(labels[z, y, x])
bg_labels = list(bg_labels)
print(bg_labels)
if len(bg_labels) > 1:
logging.warn('bg not connected, detected %d components' % len(bg_labels))
pass
for bg_label in bg_labels:
binary[bg_label == labels] = 0
pass
# now binary:
# 0: non-lung & body tissue in lung
# 1: lung & holes in body
for i, sl in enumerate(binary):
#H, W = sl.shape
ll = measure.label(sl, background=-1) # connected components
# biggest CC should be body
vv, cc = np.unique(ll, return_counts=True)
assert len(vv) > 0
body_ll = vv[np.argmax(cc)]
binary[i][ll != body_ll] = 1
pass
# set corner again
labels = measure.label(binary, background=0)
bg_labels = set([0])
for z in [0, -1]:
for y in [0, -1]:
for x in [0, -1]:
bg_labels.add(labels[z, y, x])
val_counts = zip(*np.unique(labels, return_counts=True))
val_counts = [x for x in val_counts if not x[0] in bg_labels] # remove background
val_counts = sorted(val_counts, key=lambda x:-x[1]) # sort by size
th = val_counts[0][1] /4 # 1/4 size of the larged connected component (must be lung)
val = [v for v, c in val_counts if c >= th]
if len(val) >= 3:
logging.warn('more than 2 lungs parts detected %d' % len(val))
binary = np.zeros_like(binary, dtype=np.uint8)
for v in val:
binary[labels == v] = 1
H, W = binary[0].shape
dilate = int(round(math.sqrt(1.0 * H * W) * dilate))
#print("DILATE: ", dilate)
kernel = np.ones((dilate, dilate), dtype=np.int32)
for i in range(binary.shape[0]):
cv2.dilate(binary[i], kernel, binary[i])
#image[binary == 0] = 255
#image = 255 - image
#image[binary == 0] = 255
return binary
def load_lungs_mask (uid):
cache = os.path.join('maskcache/mask-v2/%s.npz' % case.uid)
binary = None
if os.path.exists(cache) and os.path.getsize(cache) > 0:
# load cache
binary = load_mask(cache)
assert not binary is None
if binary is None:
case = load_case(uid)
case.normalizeHU()
binary = segment_lung_axial_v2(case.images) #, th=200.85)
save_mask(cache, binary)
pass
return binary
#def load_lung_mask (uid):
# #path = os.path.join('data/cache', uid)
# #if os.path.exists(path):
# # with open(path, 'rb') as f:
# # return pickle.load(f)
# case = load_case(uid)
#
# cache = os.path.join('cache/mask-123.85-0.01/%s.npz' % case.uid)
# binary = None
# if os.path.exists(cache) and os.path.getsize(cache) > 0:
# # load cache
# binary = load_mask(cache)
# assert not binary is None
# if binary is None:
# case.standardize_color()
# binary = segment_lung_axial(case.images) #, th=200.85)
# save_mask(cache, binary)
# pass
# case.images = binary.astype(np.float32)
# return case
def load_fts (path):
with open(path, 'rb') as f:
return pickle.load(f)
pass
def patch_clip_range (x, tx, wx, X):
if x < 0: #
wx += x
tx -= x
x = 0
if x + wx > X:
d = x + wx - X
wx -= d
pass
return x, tx, wx
def extract_patch_3c (images, z, y, x, size):
assert len(images.shape) == 3
_, Y, X = images.shape
z = int(round(z))
y = int(round(y))
x = int(round(x))
image = get3c(images, z)
if image is None:
return None
ty = 0
tx = 0
y -= size/2
x -= size/2
wy = size
wx = size
print y, ty, wy, x, tx, wx
y, ty, wy = patch_clip_range(y, ty, wy, Y)
x, tx, wx = patch_clip_range(x, tx, wx, X)
# now do overlap
patch = np.zeros((size, size, 3), dtype=image.dtype)
print y, ty, wy, x, tx, wx
patch[ty:(ty+wy),tx:(tx+wx),:] = image[y:(y+wy),x:(x+wx),:]
return patch
