def _calc_loss(self, x_iter, projs, x_initial):
norm = transforms.Normalize(projs[0].mean((1,2)), projs[0].std((1,2)))
percep_l = 0
proj_l = 0
tv_l = 0
ssim_l = 0
if self.w_perceptual_loss > 0.0:
percep_l = self.w_perceptual_loss * self.perceptual(x_iter, x_initial.detach())
if self.w_proj_loss > 0.0:
if self.randomize_projs:
samples = random.sample(
[i for i in range(self.N_PROJ) ],
int(self.randomize_projs*self.N_PROJ)
)
samples.sort()
radon_ = Radon(self.IMAGE_SIZE, theta=self.theta[samples], circle=True)
proj_l = self.w_proj_loss * \
self.mse(norm(radon_(x_iter)[0]),
norm(projs[0,:,:,samples]))
else:
proj_l = self.w_proj_loss * self.mse(norm(self.radon(x_iter)[0]), norm(projs[0]))
if self.w_tv_loss > 0.0:
tv_l = self.w_tv_loss * tv_2d_l2(x_iter[0,0])
if self.w_ssim_loss > 0.0:
ssim_l = self.w_ssim_loss * (1 - self.ssim(x_iter, x_initial.detach() ))
return proj_l + percep_l + tv_l + ssim_l, (proj_l, percep_l, tv_l, ssim_l)
def __init__(self, img_size=256, delta=1, in_nc=1):
super(Due_Generator, self).__init__()
# self.de_recon = Inpaint_Net(in_nc, out_nc, nf, nb, gc, norm_type, act_type, mode)
self.de_sino = PConvUNet(input_channels=in_nc)
self.de_tomo = DenoseNet()
self.angles = np.arange(0., 180., delta)
self.iradon_trans = IRadon(img_size, self.angles)
self.radon_trans = Radon(img_size, self.angles)
def test_radon_iradon_circle_double(self):
img = torch.zeros(1,1,256,256, dtype=torch.double)
img[:, :, 120:130, 120:130] = 1
circle = True
theta = torch.arange(180)
r = Radon(img.shape[2], theta, circle, dtype=torch.double)
ir = IRadon(img.shape[2], theta, circle, dtype=torch.double)
sino = r(img)
reco = ir(sino)
self.assertAlmostEqual(torch.nn.MSELoss()(img, reco).item(), 0, places=3)
def test_radon_iradon_not_circle_lazy_cut_output(self):
img = torch.zeros(1,1,256,256)
img[:, :, 120:130, 120:130] = 1
circle = False
theta = torch.arange(180)
r = Radon(theta=theta, circle=circle)
ir = IRadon(theta=theta, circle=circle, out_size=128)
sino = r(img)
reco = ir(sino)
self.assertAlmostEqual(torch.nn.MSELoss()(img[:,:,64:192,64:192], reco).item(), 0, places=3)
def test_ramp_filter(self):
img = torch.zeros(1, 1, 256, 256)
img[:, :, 120:130, 120:130] = 1
circle = True
theta = torch.arange(180)
r = Radon(img.shape[2], theta, circle)
ir = IRadon(img.shape[2], theta, circle, use_filter=RampFilter())
reco = ir(r(img))
self.assertAlmostEqual(torch.nn.MSELoss()(img, reco).item(),
0,
places=4)
def test_stackgram_istackgram_circle(self):
img = torch.zeros(1, 1, 256, 256)
img[:, :, 120:130, 120:130] = 1
circle = True
theta = torch.arange(180)
r = Radon(img.shape[2], theta, circle)
ir = IRadon(img.shape[2], theta, circle)
sg = Stackgram(img.shape[2], theta, circle)
isg = IStackgram(img.shape[2], theta, circle)
reco = ir(isg(sg(r(img))))
self.assertAlmostEqual(torch.nn.MSELoss()(img, reco).item(),
0,
places=3)
def init_train(self, theta):
self.theta = torch.from_numpy(theta).type(self.DTYPE)
self.n_proj = len(theta)
self.loss_hist = []
self.rmse_hist = []
self.ssim_hist = []
self.psnr_hist = []
self.net = self._get_net(self.net_type)
self.i_iter = 0
self.r = Radon(self.IMAGE_SIZE, theta, True).to(self.DEVICE)
self.masker = Masker(width = 4, mode='interpolate')
self.mse = torch.nn.MSELoss().to(self.DEVICE)
self.optimizer = torch.optim.Adam(self.net.parameters(), lr=self.lr)
self.meta_optimizer = torch.optim.Adam(self.net.parameters(), lr=self.lr*0.1)
if self.weights:
self._load(self.weights)
def test_hann_butterfly_filter(self):
img = torch.zeros(1, 1, 256, 256)
img[:, :, 120:130, 120:130] = 1
circle = True
theta = torch.arange(180)
r = Radon(img.shape[2], theta, circle)
ir = IRadon(img.shape[2],
theta,
circle,
use_filter=HannButterflyFilter(img.shape[2]))
reco = ir(r(img))
self.assertAlmostEqual(torch.nn.MSELoss()(img, reco).item(),
0,
places=3)
# Check that it's close to using HannFilter
ir_og = IRadon(img.shape[2], theta, circle, use_filter=HannFilter())
reco_og = ir_og(r(img))
self.assertAlmostEqual(torch.nn.MSELoss()(reco, reco_og).item(),
0,
places=4)
from skimage.data import shepp_logan_phantom
from skimage.transform import radon, rescale, iradon
import matplotlib.pyplot as plt
import torch
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
import pydicom
import scipy.ndimage
size = 512 # set size
width = 512 # set size
delt = 1. # smallest angle change
degree = 180. # Constant Scan range from 0 to 180 degree
nnradon = Radon(size, np.arange(0., degree, delt))
nniradon = IRadon(size, np.arange(0., degree,
delt)) # From full degree scan to image area
file = '../RadonAnalyze/dcmdata/000056.dcm'
ctfile = pydicom.dcmread(file)
print(ctfile)
prhalf = 2 * np.arange(2 / width, width / 2, 1, dtype=np.float32) / width
afhalf = 2 * np.arange(width / 2, 2 / width, -1, dtype=np.float32) / width
r_l_filter = np.append(prhalf, afhalf) #设置滤波器，R-L是一种基础的滤波算法
img = ctfile.pixel_array
img[img == -2000] = 0
img = img.astype('float32')
# plt.figure(figsize=(8, 5))# figure 1
def calc(self, projs, theta):
# recon params
self.N_PROJ = len(theta)
self.theta = torch.from_numpy(theta).to(self.DEVICE)
self.radon = Radon(self.IMAGE_SIZE, self.theta, True).to(self.DEVICE)
self.iradon = IRadon(self.IMAGE_SIZE, self.theta, True).to(self.DEVICE)
# start recon
if not os.path.exists(self.log_dir):
os.mkdir(self.log_dir)
net = self._get_net()
x_initial = np_to_torch(self.noisy).type(self.DTYPE)
img_gt_torch = np_to_torch(self.gt).type(self.DTYPE)
net_input = torch.rand(1, self.INPUT_DEPTH,
self.IMAGE_SIZE, self.IMAGE_SIZE).type(self.DTYPE)
net_input_saved = net_input.detach().clone()
noise = net_input.detach().clone()
# Compute number of parameters
s = sum([np.prod(list(p.size())) for p in net.parameters()]);
print ('Number of params: %d' % s)
# Optimizer
optimizer = torch.optim.Adam(net.parameters(), lr=self.lr)
cur_lr = self.lr
projs = np_to_torch(projs).type(self.DTYPE)#self.radon(img_gt_torch).detach().clone()
#np_to_torch(projs).type(self.DTYPE) #
# Iterations
loss_hist = []
rmse_hist = []
ssim_hist = []
ssim_noisy_hist = []
psnr_hist = []
psnr_noisy_hist = []
best_network = None
best_result = None
print('Reconstructing with DIP...')
for i in tqdm(range(self.n_iter)):
# iter
optimizer.zero_grad()
if self.reg_std > 0:
net_input = net_input_saved + (noise.normal_() * self.reg_std)
x_iter = net(net_input)
loss, (proj_l, percep_l, tv_l, ssim_l) = self._calc_loss(x_iter, projs, x_initial)
loss.backward()
optimizer.step()
# metric
if i % self.SHOW_EVERY == 0:
x_iter_npy = np.clip(torch_to_np(x_iter), 0, 1).astype(np.float64)
rmse_hist.append(
mean_squared_error(x_iter_npy, self.gt))
ssim_hist.append(
structural_similarity(x_iter_npy, self.gt, multichannel=False)
)
ssim_noisy_hist.append(
structural_similarity(x_iter_npy, self.noisy, multichannel=False)
)
psnr_hist.append(
peak_signal_noise_ratio(x_iter_npy, self.gt)
)
psnr_noisy_hist.append(
peak_signal_noise_ratio(x_iter_npy, self.noisy)
)
loss_hist.append(loss.item())
print('{}/{}- psnr: {:.3f} - psnr_noisy: {:.3f} - ssim: {:.3f} - ssim_noisy: {:.3f} - rmse: {:.5f} - loss: {:.5f} '.format(
self.name, i, psnr_hist[-1], psnr_noisy_hist[-1], ssim_hist[-1], ssim_noisy_hist[-1], rmse_hist[-1], loss_hist[-1]
))
#print( proj_l.item(), ssim_l.item())
# if psnr_noisy_hist[-1] / max(psnr_noisy_hist) < 0.92:
# print('Falling back to previous checkpoint.')
# for g in optimizer.param_groups:
# g['lr'] = cur_lr / 10.0
# cur_lr = cur_lr / 10.0
# print("optimizer.lr", cur_lr)
# # load network
# for new_param, net_param in zip(best_network, net.parameters()):
# net_param.data.copy_(new_param.cuda())
if i > 2:
if loss_hist[-1] < min(loss_hist[0:-1]):
# save network
best_network = [x.detach().cpu() for x in net.parameters()]
best_result = x_iter_npy.copy()
plot_grid([x_iter_npy], self.FOCUS, save_name=self.log_dir+'/{}.png'.format(i))
self.image_r = best_result
return self.image_r
