def blocks(self):
test_image, gt = load_vol_brats(self.vol_path[1], slicen=78, pad=8)
prediction = np.argmax(self.model.predict(test_image[None, ...]),
axis=-1)[0]
n_classes = (len(np.unique(prediction)))
corr = np.zeros((n_classes, n_classes))
slices = [78]
intervention_image = np.empty(test_image.shape)
for _modality in range(4):
for i in range(2):
for j in range(2):
try:
intervention_image[:, :, _modality][
test_image.shape[0] // 2 * i:test_image.shape[0] //
2 * (i + 1), test_image.shape[1] // 2 *
j:test_image.shape[1] // 2 * (j + 1)].fill(
np.mean(test_image[gt == 2 * i + j],
axis=0)[_modality])
except Exception as e:
print(e)
prediction_intervention = model.predict(intervention_image[None, ...])
plt.imshow(intervention_image[:, :, 0])
plt.colorbar()
plt.show()
plt.imshow(np.argmax(prediction_intervention, axis=-1)[0],
vmin=0,
vmax=3)
plt.colorbar()
plt.show()
def mean_swap(self,
test_path,
plot=True,
save_path='home/brats/parth/BioExp/results/RCT'):
channel = 3
vol_path = glob(test_path)
test_image, gt = load_vol_brats(vol_path[0], slicen=78, pad=8)
prediction = np.argmax(self.model.predict(test_image[None, ...]),
axis=-1)[0]
n_classes = (len(np.unique(prediction)))
corr = np.zeros((n_classes, n_classes))
slices = [78]
plt.figure(figsize=(20, 20))
for vol in vol_path:
for slicen in slices:
test_image, gt = load_vol_brats(vol, slicen=slicen, pad=8)
prediction = np.argmax(self.model.predict(test_image[None,
...]),
axis=-1)[0]
print("Original Dice Whole:", dice_whole_coef(prediction, gt))
class_dict = {0: 'bg', 1: 'core', 2: 'edema', 3: 'enhancing'}
corr_temp = np.zeros((n_classes, n_classes))
for i in range(n_classes):
for j in range(n_classes):
new_mean = np.mean(test_image[gt == i], axis=0)
old_mean = np.mean(test_image[gt == j], axis=0)
test_image_intervention = np.copy(test_image)
test_image_intervention[gt == j] += (new_mean -
old_mean)
prediction_intervention = np.argmax(self.model.predict(
test_image_intervention[None, ...]),
axis=-1)[0]
if plot == True:
plt.subplot(n_classes, n_classes, 1 + 4 * i + j)
plt.xticks([])
plt.yticks([])
plt.title("{} --> {}".format(
class_dict[j], class_dict[i]))
plt.imshow(prediction_intervention,
cmap=plt.cm.RdBu,
vmin=0,
vmax=3)
plt.colorbar()
corr[i, j] += dice_label_coef(
prediction, gt, (j, )) - dice_label_coef(
prediction_intervention, gt, (j, ))
corr_temp[i, j] += dice_label_coef(
prediction, gt, (j, )) - dice_label_coef(
prediction_intervention, gt, (j, ))
np.set_printoptions(precision=2, suppress=True)
intervention_importance = corr / (len(vol_path) * len(slices))
print(intervention_importance)
os.makedirs(save_path, exist_ok=True)
np.save(save_path + '/mean_swap_all_images.npy',
intervention_importance)
if plot == True:
plt.show()
def __init__(self, model, test_path):
self.model = model
self.vol_path = glob(test_path)
self.test_image, self.gt = load_vol_brats(self.vol_path[3],
slicen=78,
pad=0)
})
model.load_weights(
'/home/parth/Interpretable_ML/saved_models/SimUnet/SimUnet.40_0.060.hdf5'
)
I = intervention(model, '/media/parth/DATA/datasets/brats_2018/val/**')
test_path = glob('/media/parth/DATA/datasets/brats_2018/val/**')
average_change = []
for epsilon in [0.7]: #, 0.07, 0.21, 0.7]:
for i in tqdm(range(len(test_path))):
test_image, gt = load_vol_brats(test_path[i], slicen=78, pad=0)
if len(np.unique(gt)) == 4:
print(len(np.unique(gt)))
# I.blocks('/home/parth/Interpretable_ML/BioExp/sample_vol/brats/**')
adv = I.adverserial(epsilon=epsilon,
mode='gradient',
test_image=test_image,
gt=gt)
if adv[1] > 0:
average_change.append(adv)
print(adv)
print(np.mean(average_change, axis=0))
# I.generate_random_classification(mode='swap')
# I.mean_swap(plot = False)