S0 = np.nanmean(X_dw_all[np.squeeze(X_dw_all[:, bvalues == 0] > 0),
bvalues == 0],
axis=0)
valid_id = np.squeeze(X_dw_all[:, bvalues == 0] < (S0 / 3))
X_dw_all[np.squeeze(valid_id), :] = 0
valid_id = np.sum(X_dw_all == 0, axis=1) == 0
X_dw_sel = X_dw_all[valid_id, :]
S0 = np.nanmean(X_dw_sel[:, bvalues == 0], axis=1)
X_dw_sel = X_dw_sel / S0[:, None]
res = [
i for i, val in enumerate(X_dw_sel != X_dw_sel)
if not val.any()
]
net = deep.learn_IVIM(X_dw_sel[res], bvalues, arg)
paramsNN = deep.infer_IVIM(X_dw_sel, bvalues, net)
del net
gofNN = goodness_of_fit(bvalues, paramsNN[0], paramsNN[1],
paramsNN[2], paramsNN[3], X_dw_sel)
names = [
'geof_NN_{ii}_{net}_2'.format(ii=ii, net=run_net),
'Dp_NN_{ii}_{net}_2'.format(ii=ii, net=run_net),
'D_NN_{ii}_{net}_2'.format(ii=ii, net=run_net),
'f_NN_{ii}_{net}_2'.format(ii=ii, net=run_net),
'S0_NN_{ii}_{net}_2'.format(ii=ii, net=run_net)
]
for k in range(len(names)):
img = np.zeros([sx * sy * sz])
if k is 0:
img[valid_id] = gofNN
else:
datmean2=np.nanmean(datatot2,axis=0)
if testdata:
paramslsq = fit_least_squares_S0(bvalues, datmean, S0_output=True)
paramslsq2 = fit_least_squares_S0(bvalues2, datmean2, S0_output=True)
plt.plot(datmean[index])
plt.plot(ivimN(bvalues[index], paramslsq[0]*10, paramslsq[1]*1000, paramslsq[2]*10, paramslsq[3]))
plt.plot(datmean2[index2])
plt.plot(ivimN(bvalues[index], paramslsq2[0]*10, paramslsq2[1]*1000, paramslsq2[2]*10, paramslsq2[3]))
net = deep.learn_IVIM(np.transpose(np.repeat(np.expand_dims(datmean,1),1000,axis=1)), bvalues, arg)
net2 = deep.learn_IVIM(np.transpose(np.repeat(np.expand_dims(datmean2,1),1000,axis=1)), bvalues2, arg)
paramsNN=deep.infer_IVIM(np.expand_dims(datmean,0), bvalues, net)
paramsNN2=deep.infer_IVIM(np.expand_dims(datmean2,0), bvalues2, net2)
plt.plot(datmean[index])
plt.plot(ivimN(bvalues[index], paramsNN[0][0]*10, paramsNN[1][0]*1000, paramsNN[2][0]*10, paramsNN[3][0]))
plt.plot(datmean2[index2])
plt.plot(ivimN(bvalues[index], paramsNN2[0][0]*10, paramsNN2[1][0]*1000, paramsNN2[2][0]*10, paramsNN2[3][0]))
print('least squares fitting\n')
if dolsq:
if not load_lsq:
if segmented:
print('segmented\n')
paramslsq=fit_segmented_array(bvalues,datatot)
X_dw_all = np.reshape(datas, (sx * sy * sz, n_b_values))
valid_id = np.sum(X_dw_all == 0, axis=1) == 0
datatot = np.append(datatot, X_dw_all[valid_id, :], axis=0)
del valid_id, X_dw_all, datas, data
bval = np.delete(bval, np.s_[3::4], 0)
datatot = np.array(np.delete(datatot, np.s_[3::4], 1))
S0 = np.nanmean(datatot[:, bval == 0], axis=1)
datatot = datatot / S0[:, None]
datmean = np.nanmean(datatot, axis=0)
res = [i for i, val in enumerate(datatot != datatot) if not val.any()]
net = deep.learn_IVIM(datatot[res], bval, arg)
torch.save(net, 'network_{nn}.pt'.format(nn=arg.run_net))
paramsNN = deep.infer_IVIM(datatot, bval, net)
del net
gofNN = goodness_of_fit(bval, paramsNN[0], paramsNN[1], paramsNN[2],
paramsNN[3], datatot)
names = [
'geof_NN_{nn}.nii'.format(nn=arg.run_net),
'Dp_NN_{nn}.nii'.format(nn=arg.run_net),
'D_NN_{nn}.nii'.format(nn=arg.run_net),
'f_NN_{nn}.nii'.format(nn=arg.run_net),
'S0_NN_{nn}.nii'.format(nn=arg.run_net)
]
tot = 0
bval = np.array([
0, 0, 0, 0, 700, 700, 700, 700, 1, 1, 1, 1, 5, 5, 5, 5, 100, 100, 100, 100,
300, 300, 300, 300, 10, 10, 10, 10, 0, 0, 0, 0, 20, 20, 20, 20, 500, 500,
def sim(SNR,
b,
arg,
run_net=None,
sims=100000,
num_samples_leval=10000,
Dmin=0.5 / 1000,
Dmax=3.0 / 1000,
fmin=0.05,
fmax=0.4,
Dsmin=0.01,
Dsmax=0.1,
rician=False,
segmented=False):
IVIM_signal_noisy, f, D, Dp = sim_signal(SNR,
b,
sims=sims,
Dmin=Dmin,
Dmax=Dmax,
fmin=fmin,
fmax=fmax,
Dsmin=Dsmin,
Dsmax=Dsmax,
rician=rician)
D = D[:num_samples_leval]
Dp = Dp[:num_samples_leval]
f = f[:num_samples_leval]
if arg.repeats > 1:
paramsNN = np.zeros([arg.repeats, 4, num_samples_leval])
for aa in range(arg.repeats):
start_time = time.time()
net = deep.learn_IVIM(IVIM_signal_noisy, b, arg)
elapsed_time = time.time() - start_time
print('\ntime elapsed for training: {}\n'.format(elapsed_time))
start_time = time.time()
if arg.repeats > 1:
paramsNN[aa] = deep.infer_IVIM(
IVIM_signal_noisy[:num_samples_leval, :], b, net)
else:
paramsNN = deep.infer_IVIM(IVIM_signal_noisy, b, net)
elapsed_time = time.time() - start_time
print('\ntime elapsed for inference: {}\n'.format(elapsed_time))
# ['Ke_NN.nii', 'f_NN.nii', 'tau_NN.nii', 'v_NN.nii']
del net
print('results for NN')
IVIM_signal_noisy = IVIM_signal_noisy[:num_samples_leval, :]
if arg.repeats > 1:
matNN = np.zeros([arg.repeats, 3, 3])
for aa in range(arg.repeats):
matNN[aa] = print_errors(np.squeeze(D), np.squeeze(f),
np.squeeze(Dp), paramsNN[aa])
matNN = np.mean(matNN, axis=0)
stability = np.sqrt(
np.mean(np.square(np.std(paramsNN, axis=0)), axis=1))
stability = stability[[1, 2, 0]] / [
np.mean(D), np.mean(f), np.mean(Dp)
]
paramsNN = paramsNN[0]
else:
matNN = print_errors(np.squeeze(D), np.squeeze(f), np.squeeze(Dp),
paramsNN)
stability = np.zeros(3)
dummy = np.array(paramsNN)
del paramsNN
plt.figure()
plt.plot(D[:1000], Dp[:1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.3)
plt.xlabel('Dt')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/inputDtDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(f[:1000], Dp[:1000], 'rx', markersize=5)
plt.xlim(0, 0.6)
plt.ylim(0, 0.3)
plt.xlabel('f')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/inputfDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(D[:1000], f[:1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.6)
plt.xlabel('Dt')
plt.ylabel('f')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/inputDtf.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(dummy[1, :1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.3)
plt.xlabel('Dt')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/NNDtDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(dummy[2, :1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.6)
plt.ylim(0, 0.3)
plt.xlabel('f')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/NNfDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(dummy[1, :1000], dummy[2, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.6)
plt.xlabel('Dt')
plt.ylabel('f')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/NNDtf.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(Dp[:1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.3)
plt.ylim(0, 0.3)
plt.ylabel('DpNN')
plt.xlabel('Dpin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/DpNNDpin.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(D[:1000], dummy[1, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.005)
plt.ylabel('DtNN')
plt.xlabel('Dtin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/DtNNDtin.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(f[:1000], dummy[2, :1000], 'rx', markersize=5)
plt.xlim(0, 0.6)
plt.ylim(0, 0.6)
plt.ylabel('fNN')
plt.xlabel('fin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/fNNfin.png'
)
plt.ion()
plt.show()
start_time = time.time()
if segmented:
paramsf = fit.fit_segmented_array(b, IVIM_signal_noisy)
else:
paramsf = fit.fit_least_squares_array(b, IVIM_signal_noisy)
elapsed_time = time.time() - start_time
print('\ntime elapsed for lsqfit: {}\n'.format(elapsed_time))
print('results for lsqfit')
matlsq = print_errors(np.squeeze(D), np.squeeze(f), np.squeeze(Dp),
paramsf)
dummy = np.array(paramsf)
del paramsf, IVIM_signal_noisy
plt.figure()
plt.plot(dummy[1, :1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.3)
plt.xlabel('D')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/LSQDtDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(dummy[2, :1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.6)
plt.ylim(0, 0.3)
plt.xlabel('f')
plt.ylabel('Dp')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/LSQfDp.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(dummy[1, :1000], dummy[2, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.6)
plt.xlabel('D')
plt.ylabel('f')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/LSQDtf.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(Dp[:1000], dummy[0, :1000], 'rx', markersize=5)
plt.xlim(0, 0.3)
plt.ylim(0, 0.3)
plt.ylabel('Dplsq')
plt.xlabel('Dpin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/DpLSQDpin.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(D[:1000], dummy[1, :1000], 'rx', markersize=5)
plt.xlim(0, 0.005)
plt.ylim(0, 0.005)
plt.ylabel('Dtlsq')
plt.xlabel('Dtin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/DtLSQDtin.png'
)
plt.ion()
plt.show()
plt.figure()
plt.plot(f[:1000], dummy[2, :1000], 'rx', markersize=5)
plt.xlim(0, 0.6)
plt.ylim(0, 0.6)
plt.ylabel('flsq')
plt.xlabel('fin')
plt.gcf()
plt.savefig(
'C:/Users/ojgurney-champion/Dropbox/Research/DeepLearning/deep_ivim/Output/fLSQfin.png'
)
plt.ion()
plt.show()
return matlsq, matNN, stability
index2 = np.argsort(bvalues2)
S0 = np.nanmean(datatot[:,bvalues==0],axis=1)
datatot = datatot/S0[:,None]
S0 = np.nanmean(datatot3[:,bvalues==0],axis=1)
datatot3 = datatot3/S0[:,None]
datmean=np.nanmean(datatot,axis=0)
datmean2=np.nanmean(datatot2,axis=0)
res = [i for i, val in enumerate(datatot!=datatot) if not val.any()]
res2 = [i for i, val in enumerate(datatot2!=datatot2) if not val.any()]
paramsNN=np.zeros([20,4,np.shape(datatot3)[0]])
for qq in range(20):
#[net, matNN] = deep.pretrain(bvalues, arg, SNR=15, net=run_net, state=qq, sims=100000)
net = deep.learn_IVIM(datatot[res], bvalues, arg)# net=net)
paramsNN[qq]=deep.infer_IVIM(datatot3, bvalues, net)
#matNN_count[qq]=matNN
del net
# save NN results
names = ['Dp_NN_{net}_rep'.format(net=run_net), 'D_NN_{net}_2'.format(net=run_net), 'f_NN_{net}_2'.format(net=run_net), 'S0_NN_{net}_2'.format(net=run_net)]
multiple = [1000., 1000000., 10000., 1000.]
fold = aa2[0]
ss = 2
data=nib.load('{folder}/CR{fold:02d}/MRI{ss}_{dat}.nii.gz'.format(folder=fold1,fold=fold,ss=ss, dat=dattype))
datas=data.get_data()
sx, sy, sz, n_b_values = datas.shape
data.header.set_data_shape((sx,sy,sz,np.shape(paramsNN)[0]))
X_dw_all = np.reshape(datas, (sx * sy * sz, n_b_values))
valid_id = np.sum(X_dw_all == 0, axis=1) == 0
imgtot=np.zeros([sx,sy,sz,np.shape(paramsNN)[0]])
for k in range(len(names)):