def data_preprocess(self, input_image, normalization=None, num_channels=0):
if type(input_image) != np.ndarray:
input_image = np.stack(input_image, axis=0)
if normalization == 'CT':
image = mt.medical_normalization(input_image, input_copy=False)
elif normalization == 'CT_chest':
image = mt.medical_normalization(input_image,
max_bound=1000,
min_bound=-1000,
pixel_mean=0.25,
crop=False,
input_copy=False)
elif normalization == 'CT_lung':
image = mt.medical_normalization(input_image,
max_bound=400,
min_bound=-1000,
pixel_mean=0.25,
crop=True,
input_copy=False)
elif normalization == 'image':
image_max = (np.zeros(1, dtype=input_image.dtype) - 1)[0]
#image_max = 65535
image = input_image / float(image_max)
else:
image = input_image
if num_channels > 0:
image = np.expand_dims(image, axis=0)
if num_channels > 1:
image = image.repeat(num_channels, axis=0)
image = torch.from_numpy(image).float()
return image
posbatchend = min(pbi + BATCH_SIZE, positive_train_num)
for pti in range(pbi, posbatchend):
data_index = positive_train_indices[pti]
pfile = tpfiles[data_index]
positive_data = np.load(pfile)
nodule_diameter = 0
if "positive_batch" not in dir():
positive_batch = extract_volumes(
positive_data, nodule_diameter=nodule_diameter)
else:
positive_batch = np.concatenate(
(positive_batch,
extract_volumes(positive_data,
nodule_diameter=nodule_diameter)),
axis=0)
positive_batch = mt.medical_normalization(positive_batch, MAX_BOUND,
MIN_BOUND, PIXEL_MEAN, True)
positive_label = np.zeros(shape=(positive_batch.shape[0], 2),
dtype=float)
positive_label[:, 0] = 1
predictions, accuracies = sess.run([out_fc2, correct_prediction], {
volume_input: positive_batch,
real_label: positive_label
})
if not AUGMENTATION:
for a in range(len(accuracies)):
data_index = positive_train_indices[pbi + a]
pfile = tpfiles[data_index]
if accuracies[a]:
correct_output.write(pfile +
" {}\n".format(predictions[a]))
else:
batchend = min(bi + BATCH_SIZE, test_num)
batch_size = batchend - bi
train_data = np.zeros(
(batch_size, REGION_SIZE, REGION_SIZE, REGION_SIZE), dtype=float)
train_label = np.zeros((batch_size, 2), dtype=float)
for ti in range(bi, batchend):
file = test_files[ti]
malignancy = lt.sample_malignancy(file)
if malignancy >= 0:
train_label[ti - bi][1 - malignancy] = 1
else:
print('unknown malignancy, skip')
continue
data = np.load(file)
data_cropped = mt.crop(data,
(REGION_SIZE, REGION_SIZE, REGION_SIZE))
train_data[ti - bi] = mt.medical_normalization(data_cropped,
MAX_BOUND,
MIN_BOUND,
PIXEL_MEAN,
NORM_CROP,
input_copy=False)
accuracy, prediction = sess.run([batch_accuracy, prediction_fusion], {
volume_input: train_data,
real_label: train_label
})
total_accuracy += accuracy * batch_size
predictions[bi:batchend] = prediction
total_accuracy /= float(test_num)
print("accuracy:{}".format(total_accuracy))
#print("prediction:\n{}" .format(predictions))