def getMaskVolumes():
csvLines = utils.readCsv("../trainset_csv/trainNodules_gt.csv")
last_ID = 0
scan = 0
spacing = 0
origin = 0
maskVolumesList = []
# ignore header
for line in csvLines[1:]:
# get image of this patient only one time (there are repeated patient ids)
current_ID = line[0]
if last_ID != current_ID:
print(getFileID(current_ID))
scan, spacing, origin, _ = utils.readMhd(
'../LNDb dataset/dataset/LNDb-' + getFileID(current_ID) +
'.mhd')
spacing = [float(spacing[i]) for i in range(3)]
# find the coordinates of the current nodule (it is done for every line of the csv)
finding_coords = line[4:7]
nodule_x = (float(finding_coords[0]) - float(origin[0])) / float(
spacing[0])
nodule_y = (float(finding_coords[1]) - float(origin[1])) / float(
spacing[1])
nodule_z = (float(finding_coords[2]) - float(origin[2])) / float(
spacing[2])
real_coords = [nodule_x, nodule_y, nodule_z]
# get a mask for the image of this patient (from one of the radiologists that found the current nodule)
radiologists = line[
1] # list of radiologists that found the current nodule
radId = str(
radiologists[0]
) # always choose the mask from the first radiologist in the list
mask, _, _, _ = utils.readMhd('../LNDb dataset/masks/LNDb-' +
getFileID(current_ID) + '_rad' + radId +
'.mhd')
# extract mini cube of the current nodule on the masked scan
mask_volume = utils.extractCube(mask,
spacing,
real_coords,
cube_size=80)
# add mask volumes to the list
maskVolumesList.append(mask_volume)
last_ID = current_ID
return maskVolumesList
def load_image(img_name, centers):
''' loads a single image and returns a list of subimages (cubes around nodule-like
findings)
img_name: valid path to an image from current working directory
centers: a list of the centroids of nodule-like occurrences in img
in world coordinates
'''
scan, spacing, origin, transfmat = readMhd(img_name)
transfmat_toimg, _ = getImgWorldTransfMats(spacing,
transfmat) # just need image
centers = [convertToImgCoord(c, origin, transfmat_toimg) for c in centers]
subimages = [extractCube(scan, spacing, c) for c in centers]
return subimages
def createCube(cubeSize=80):
csvLines = readCsv("csv/trainSet.csv")
header = csvLines[0]
nodules = csvLines[1:]
for i, n in tqdm(enumerate(nodules[:1])):
x = int(n[header.index("x")])
y = int(n[header.index("y")])
z = int(n[header.index("z")])
lnd = int(n[header.index("LNDbID")])
ctr = np.array([x, y, z])
[scan, spacing, _, _] = readMhd('data/LNDb-{:04}.mhd'.format(lnd))
cube = extractCube(scan, spacing, ctr, cubeSize)
np.save("cube/{:0}.npy".format(i), cube)
def getCubes(cubeSize):
csvLines = utils.readCsv("../trainset_csv/trainNodules_gt.csv")
last_ID = 0
scan = 0
spacing = 0
origin = 0
cubeList = []
textures = [row[-1] for row in csvLines]
# delete 1st element (header)
del textures[0]
# ignore header
for line in csvLines[1:]:
current_ID = line[0]
if last_ID != current_ID:
print(getFileID(current_ID))
scan,spacing,origin,_ = utils.readMhd('../LNDb dataset/dataset/LNDb-' + getFileID(current_ID) + '.mhd')
spacing = [float(spacing[i]) for i in range(3)]
finding_coords = line[4:7]
nodule_x = (float(finding_coords[0]) - float(origin[0])) / float(spacing[0])
nodule_y = (float(finding_coords[1]) - float(origin[1])) / float(spacing[1])
nodule_z = (float(finding_coords[2]) - float(origin[2])) / float(spacing[2])
real_coords = [nodule_x, nodule_y, nodule_z]
scan_cube = utils.extractCube(scan, spacing, real_coords, cube_size=cubeSize)
cubeList.append(scan_cube)
# nodule_coords
last_ID = current_ID
return cubeList, textures
for n in nodules:
vol = float(n[header.index('Volume')])
if nodEqDiam(vol)>3: #only get nodule cubes for nodules>3mm
ctr = np.array([float(n[header.index('x')]), float(n[header.index('y')]), float(n[header.index('z')])])
lnd = int(n[header.index('LNDbID')])
rad = int(n[header.index('RadID')])
# rads = list(map(int,list(n[header.index('RadID')].split(','))))
# radfindings = list(map(int,list(n[header.index('RadFindingID')].split(','))))
finding = int(n[header.index('FindingID')])
# print(lnd,finding,rads,radfindings)
print(lnd,finding,rad,finding)
# Read scan
if lnd!=lndloaded:
[scan,spacing,origin,transfmat] = readMhd('/media/tungthanhlee/SSD/grand_challenge/dataset/LNDB_segmentation/data/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg,transfmat_toworld = getImgWorldTransfMats(spacing,transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr,origin,transfmat_toimg)
# for rad,radfinding in zip(rads,radfindings):
# Read segmentation mask
[mask,_,_,_] = readMhd('/media/tungthanhlee/SSD/grand_challenge/dataset/LNDB_segmentation/masks/LNDb-{:04}_rad{}.mhd'.format(lnd,rad))
# Extract cube around nodule
scan_cube = extractCube(scan,spacing,ctr)
masknod = copy.copy(mask)
# masknod[masknod!=radfinding] = 0
# * Z-Axis position of Slice/Label
# ## 3 adjacent grayscale axial slides
if not os.path.isdir('sliceTs'):
os.mkdir('sliceTs')
tsFiles = []
t_wlds = []
origins = []
id_c = 1
id_n = 0
input_shapes = []
for index in tqdm(range(len(filepaths_imgs))):
sc, sp, o, t = utils.readMhd(filepaths_imgs[index])
t_img, t_wld = utils.getImgWorldTransfMats(sp, t)
z_max = sc.shape[0]
z_max_idx = z_max - (sc.shape[0] % 3)
for id_n in range(int(z_max_idx * .6), int(z_max_idx * .9), 3):
origins.append(o)
input_shapes.append(sc.shape)
slide = sc[int(id_n) - 1:int(id_n) + 2, :, :]
slide = equalize_hist(slide)
slide = anisotropic_diffusion(slide,
voxelspacing=sp,
kappa=20,
gamma=0.01,
if __name__ == "__main__":
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.manual_seed(10)
reader = readCsv('./train_valid_CTs.csv')
data_list = readCsv('./data_list_train_valid.csv')
reader = reader[1:]
data_list = np.array(data_list[1:])
tr_data = list()
tr_label = list()
for number in range(len(reader)): #讀取檔案製作訓練資料集[病人,幾張,3,Y,X]
[scan, spacing, origin, transfmat
] = readMhd('traindata/LNDb-{:04}.mhd'.format(int(reader[number][0])))
[mask, spacing, origin, transfmat] = readMhd(
'trainlabel_1c1t/LNDb-{:04}_w_Text_merged-mask.mhd'.format(
int(reader[number][0])))
for z in range(scan.shape[0]):
tr_data.append(scan[z])
tr_label.append(mask[z])
print("loading finished.")
model = NestedUNet().to(device)
#pre_weight = ""
#model.load_state_dict(torch.load(pre_weight, map_location=device))
print("amount of parameters: {}".format(count_params(model)))
# hyperparameters
learning_rate = 1e-4
float(n[header.index('x')]),
float(n[header.index('y')]),
float(n[header.index('z')])
])
lnd = int(n[header.index('LNDbID')])
rads = list(map(int, list(n[header.index('RadID')].split(','))))
radfindings = list(
map(int, list(n[header.index('RadFindingID')].split(','))))
finding = int(n[header.index('FindingID')])
print(lnd, finding, rads, radfindings)
# Read scan
if lnd != lndloaded:
[scan, spacing, origin,
transfmat] = readMhd('data/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg, transfmat_toworld = getImgWorldTransfMats(
spacing, transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr, origin, transfmat_toimg)
for rad, radfinding in zip(rads, radfindings):
# Read segmentation mask
[mask, _, _,
_] = readMhd('masks/LNDb-{:04}_rad{}.mhd'.format(lnd, rad))
# Extract cube around nodule
scan_cube = extractCube(scan, spacing, ctr)
masknod = copy.copy(mask)
float(n[header.index('z')])
])
lnd = int(n[header.index('LNDbID')])
rads = list(map(int, list(n[header.index('RadID')].split(','))))
radfindings = list(
map(int, list(n[header.index('RadFindingID')].split(','))))
finding = int(n[header.index('FindingID')])
noduleTexture = (n[header.index('Text')])
noduleTexture = float(n[header.index('Text')])
#label = label_img(noduleTexture)
print(lnd, finding, rads, radfindings, noduleTexture)
#print("nodule Texture", label)
# Read scan
if lnd != lndloaded:
[scan, spacing, origin,
transfmat] = readMhd('data/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg, transfmat_toworld = getImgWorldTransfMats(
spacing, transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr, origin, transfmat_toimg)
#for rad,radfinding in zip(rads,radfindings):
# Read segmentation mask
#[mask,_,_,_] = readMhd('masks/LNDb-{:04}_rad{}.mhd'.format(lnd,rad))
# Extract cube around nodule
scan_cube = extractCube(scan, spacing, ctr)
"""masknod = copy.copy(mask)
masknod[masknod!=radfinding] = 0
float(n[header.index('y')]),
float(n[header.index('z')])
])
lnd = int(n[header.index('LNDbID')])
rads = list(map(int, list(n[header.index('RadID')].split(','))))
radfindings = list(
map(int, list(n[header.index('RadFindingID')].split(','))))
finding = int(n[header.index('FindingID')])
texts = list(map(float, list(n[header.index('Text')].split(','))))
print(lnd, finding, rads, radfindings)
# Read scan
if lnd != lndloaded:
[scan, spacing, origin,
transfmat] = readMhd('../Dataset/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg, transfmat_toworld = getImgWorldTransfMats(
spacing, transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr, origin, transfmat_toimg)
for rad, radfinding, text in zip(rads, radfindings, texts):
scan_cube = extractCube(scan, spacing, ctr)
# Display mid slices from resampled scan/mask
if dispFlag:
fig, axs = plt.subplots(2, 3)
axs[0, 0].imshow(scan_cube[int(scan_cube.shape[0] / 2), :, :])
# axs[1,0].imshow(mask_cube[int(mask_cube.shape[0]/2),:,:])
if nodEqDiam(vol) > 3: #only get nodule cubes for nodules>3mm
ctr = np.array([
float(n[header.index('x')]),
float(n[header.index('y')]),
float(n[header.index('z')])
])
lnd = int(n[header.index('LNDbID')])
rad = int(n[header.index('RadID')])
finding = int(n[header.index('FindingID')])
print(ctr, lnd, finding, rad)
# Read scan
if lnd != lndloaded:
[scan, spacing, origin,
transfmat] = readMhd('../Dataset/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg, transfmat_toworld = getImgWorldTransfMats(
spacing, transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr, origin, transfmat_toimg)
# Read segmentation mask
[mask, _, _,
_] = readMhd('../Dataset/masks/LNDb-{:04}_rad{}.mhd'.format(lnd, rad))
# Extract cube around nodule
scan_cube = extractCube(scan, spacing, ctr)
masknod = copy.copy(mask)
masknod[masknod != finding] = 0 #is radfinding on other file.
lndloaded = -1
for n in nodules:
vol = float(n[header.index('Volume')])
if nodEqDiam(vol)>3: #only get nodule cubes for nodules>3mm
ctr = np.array([float(n[header.index('x')]), float(n[header.index('y')]), float(n[header.index('z')])])
lnd = int(n[header.index('LNDbID')])
rads = list(map(int,list(n[header.index('RadID')].split(','))))
radfindings = list(map(int,list(n[header.index('RadFindingID')].split(','))))
finding = int(n[header.index('FindingID')])
print(lnd,finding,rads,radfindings)
# Read scan
if lnd!=lndloaded:
[scan,spacing,origin,transfmat] = readMhd('data/LNDb-{:04}.mhd'.format(lnd))
transfmat_toimg,transfmat_toworld = getImgWorldTransfMats(spacing,transfmat)
lndloaded = lnd
# Convert coordinates to image
ctr = convertToImgCoord(ctr,origin,transfmat_toimg)
for rad,radfinding in zip(rads,radfindings):
# Read segmentation mask
[mask,_,_,_] = readMhd('masks/LNDb-{:04}_rad{}.mhd'.format(lnd,rad))
# Extract cube around nodule
scan_cube = extractCube(scan,spacing,ctr)
masknod = copy.copy(mask)
masknod[masknod!=radfinding] = 0
masknod[masknod>0] = 1