brain_test = np.expand_dims(brain_test, 0)
deformation = sdn.predict(brain_test)[0, ...]
sample = grid + deformation
sample = np.stack(
(sample[:, :, :, 1], sample[:, :, :, 0], sample[:, :, :, 2]), 3)
warp_seg = interpn((yy, xx, zz),
brain_test_label1,
sample,
method='nearest',
bounds_error=False,
fill_value=0)
dice_after[test_ind, :] = np.array(
[Dice(brain_test_label2 == i, warp_seg == i) for i in label_list])
print('Test sample {}\'s evaluation completed.'.format(test_ind + 1))
# for label_ind, i in enumerate(label_list):
#dice_before[test_ind, label_ind]= Dice(brain_test_label2==i, brain_test_label1 == i)
# seg = transform(deformation, np.expand_dims(np.expand_dims(brain_test_label1==i, 3),0), (res1, res2, res3))
# dice_after[test_ind, label_ind] = Dice(brain_test_label2 == i, np.rint(seg)>0)
#print("ROI {} result appended..".format(i))
# print('Test sample {}\'s evaluation completed.'.format(test_ind+1))
#print("writing the first to log....")
np.save('.npy', dice_after)
#np.save('/global/home/hpc4355/MindBoggle/output/warped.npy', warp_seg)
brain_test_mov = np.load(datapath+test_label[:2]+'.npy')
brain_test_fix = np.load(datapath+test_label[2:]+'.npy')
brain_test = np.stack([brain_test_mov, brain_test_fix], 3)
brain_test_label1 = np.load(labelpath+test_label[:2]+'.npy')
brain_test_label2 = np.load(labelpath+test_label[2:4]+'.npy')
brain_test = np.expand_dims(brain_test, 0)
warped_vol, deformation = sdn_pred.predict(brain_test)
np.save('/home/dkuang/{}.npy'.format(test_label), warped_vol[0,...,0])
for label_ind, i in enumerate(label_list):
seg = transform(deformation, np.expand_dims(np.expand_dims(brain_test_label1==i, 3),0), (res1, res2, res3))
dice_before[test_ind, label_ind] = Dice(brain_test_label2 == i, np.rint(seg)>0)
print(dice_before)
print(np.mean(dice_before, 1))
sdn_refine.compile(loss = ['mse', total_variation_loss],
loss_weights = par['loss_weights'],
optimizer = Adam(lr = par['lr'], decay = 1e-5) )
print(sdn_refine.summary())
gen_train = vol_gen_refine(datapath, predpath,train_list, par['batch_size'])
history = sdn_refine.fit_generator(gen_train, steps_per_epoch = len(train_list)/par['batch_size'], epochs = par['epochs'], use_multiprocessing = True, verbose=1)
dice_after = np.zeros([len(train_list), 56])
for test_ind, test_label in enumerate(train_list):
sdn_warped_img = transform(deformation,
np.expand_dims(brain_test[..., 0], 4),
(res1, res2, res3))
sse_before[ind] = np.sum((brain_test[..., 0] - brain_test[..., 1])**2)
sse_sdn[ind] = np.sum(
(brain_test[0, ..., 1] - sdn_warped_img[0, ..., 0])**2)
sse_utr[ind] = np.sum((brain_test[0, ..., 1] - utr_warped_img)**2)
#print("sse_before: {}".format(sse_before))
#print("sse_sdn: {}".format(sse_sdn))
#print("sse_utr: {}".format(sse_utr))
for label_ind, i in enumerate(label_list):
dice_before[ind, label_ind] = Dice(brain_test_label2 == i,
brain_test_label1 == i)
seg = transform(
deformation,
np.expand_dims(np.expand_dims(brain_test_label1 == i, 3), 0),
(res1, res2, res3))
#seg_utr = transform(disp, np.expand_dims(np.expand_dims(brain_test_label1==i, 3),0), (res1, res2, res3))
dice_after_sdn[ind, label_ind] = Dice(brain_test_label2 == i,
np.rint(seg) > 0)
dice_after_utr[ind, label_ind] = Dice(brain_test_label2 == i,
np.rint(utr_warped_label) == i)
print("{}-th sample done.".format(ind + 1))
#print("utr vs before: {} ROIs are better".format(np.sum(dice_after_utr[0]>dice_before[0])))
#print("sdn vs before: {} ROIs are better".format(np.sum(dice_after_sdn[0]>dice_before[0])))
#print("utr vs sdn: {} ROIs are better".format(np.sum(dice_after_utr[0]>dice_after_sdn[0])))
#Vol_Prob = np.exp(np.mean(label_prob, axis=0)) #other than taking mean?
#Vol_sum = np.sum(Vol_Prob, axis = 3)
#for i in range(len(label_list)):
# Vol_Prob[...,i] = Vol_Prob[...,i]/D_sum[0,...,0]
'''
Probability across samples
'''
maxprob = np.amax(Vol_Prob, axis=3) / np.sum(W)
#np.save(weights_path+'lpba_maxprob_sample.npy', maxprob)
WinningLabel = np.argmax(Vol_Prob, axis=3)
#WinningLabel = WinningLabel*(maxprob>0.02)
'''
Weighted? Joint Fusion?
'''
print('%s seconds used in whole atlas construction.' %
(time.time() - start_time))
np.save(weights_path + 'mb_avg_incept0_temp1_near.npy', ini)
#np.save(weights_path+'mb_prob_incept0.npy', Vol_Prob)
np.save(weights_path + 'mb_maxprob_temp1_near.npy', maxprob)
np.save(weights_path + 'mb_WinningLabel_temp1_near.npy', WinningLabel)
dice = [
Dice(temp_label == label_list[i], WinningLabel == i)
for i in range(1, len(label_list))
]
print(np.mean(dice[1:]))
#print(W)
# Vol_Prob[:-1,:-1,:-1,label_ind] += DJ*interpn((yy, xx, zz), brain_label==i, warped_grid, method='nearest', bounds_error=False, fill_value=0)[:-1,:-1,:-1]
# Vol_Prob = np.exp(np.mean(label_prob, axis=0)) #other than taking mean?
# Vol_sum = np.sum(Vol_Prob, axis = 3)
# for i in range(len(label_list)):
# Vol_Prob[...,i] = Vol_Prob[...,i]/D_sum[0,...,0]
'''
Probability across samples
'''
maxprob = np.amax(Vol_Prob, axis=3)/np.sum(W)
# np.save(weights_path+'lpba_maxprob_sample.npy', maxprob)
WinningLabel = np.argmax(Vol_Prob,axis=3)
# WinningLabel = WinningLabel*(maxprob>0.02)
'''
Weighted? Joint Fusion?
'''
print('%s seconds used in whole atlas construction.'%(time.time() - start_time))
np.save(weights_path+'mb_avg_incept0_temp1_near.npy', ini)
# np.save(weights_path+'mb_prob_incept0.npy', Vol_Prob)
np.save(weights_path+'mb_maxprob_temp1_near.npy', maxprob)
np.save(weights_path+ 'mb_WinningLabel_temp1_near.npy', WinningLabel)
dice = [Dice(temp_label==label_list[i], WinningLabel==i) for i in range(1, len(label_list))]
print(np.mean(dice[1:]))
# print(W)