def show_data_augment(index):
images_list = get_all_images_list("train")
# Check index is valid
if index >= len(images_list):
print("Index %s out of range. Max index is %s" %
(index, len(images_list) - 1))
exit(-1)
file_path, filename = images_list[index]
dicom_data = get_dicom_data(file_path)
image = tf.convert_to_tensor(dicom_data.pixel_array, dtype=tf.float32)
image = tf.reshape(image, (1, image_size, image_size, 1))
# Display the final image
fig = plt.figure(figsize=(18, 12))
fig.canvas.set_window_title("Show data augment %s" % filename)
show_subplot(plt, 1, image, "Original Image")
show_subplot(plt, 2, data_augment.get_tall_image(image), "Tall Image")
show_subplot(plt, 3, data_augment.get_large_image(image), "Large Image")
show_subplot(plt, 4, data_augment.get_zoomed_image(image), "Zoomed Image")
show_subplot(plt, 5, data_augment.get_mirror_image(image), "Mirror Image")
show_subplot(plt, 6, data_augment.rotate_and_zoom_image(image),
"Rotated and Zoomed Image")
show_subplot(plt, 7, data_augment.smooth_image(image), "Smooth Image")
show_subplot(plt, 8, data_augment.change_image_brightness(image),
"Bright Image")
plt.show()
def show_true_mask(index):
images_list = get_all_images_list("train")
# Check index is valid
if index >= len(images_list):
print("Index %s out of range. Max index is %s" % (index, len(images_list) - 1))
exit(-1)
file_path, filename = images_list[index]
dicom_data = get_dicom_data(file_path) # True mask is only known for the training dataset
mask = get_true_mask(filename)
if np.max(mask) == 0:
print("No Pneumothorax on %s" % filename)
exit(0)
pixels = np.array(dicom_data.pixel_array) * (1 - .5 * mask)
# Display the final image
fig = plt.figure(figsize=(18, 12))
fig.canvas.set_window_title("Show True Mask of %s" % filename)
plt.imshow(pixels)
plt.show()
def show_prediction(folder, index):
images_list = get_all_images_list(folder)
# Check index is valid
if index >= len(images_list):
print("Index %s out of range. Max index is %s" %
(index, len(images_list) - 1))
exit(-1)
filepath, filename = images_list[index]
dicom_data = get_dicom_data(filepath)
# Display the final image
fig = plt.figure(figsize=(18, 12))
fig.canvas.set_window_title("Show Predicted Mask of %s" % filename)
plt.subplot(1, 2, 1)
pixels = dicom_data.pixel_array
if (folder == "train"):
true_mask = get_true_mask(filename)
pixels = np.array(pixels) * (1 - .5 * true_mask)
plt.imshow(pixels)
plt.subplot(1, 2, 2)
predictions = get_segmentation_prediction(filepath)
plt.imshow(predictions)
plt.show()
def get_segmentation_prediction(filepath):
"Returns the prediction as a numpy array of shape (image_size, image_size) for a given filepath"
dicom_data = get_dicom_data(filepath)
image = format_pixel_array_for_tf(dicom_data.pixel_array)
predicted_logits = unet.predict(image, steps=1)
predictions = tf.convert_to_tensor(predicted_logits, dtype=tf.float32)
predictions = tf.image.resize(predicted_logits, (image_size, image_size),
align_corners=True)
predictions = sess.run(predictions)
predictions = np.reshape(predictions, (image_size, image_size))
return predictions
def get_classification_prediction(filepath):
"""
Returns the prediction as a numpy array of shape (image_size, image_size) for a given filepath\n
Please note that due a bug loading the models, get_classification_prediction currenly uses the body models.
Please fix the bug and use the head models of the hydra
"""
dicom_data = get_dicom_data(filepath)
image = format_pixel_array_for_tf(dicom_data.pixel_array)
predictions = [model.predict(image, steps=1) for model in all_models]
predictions = [p[0][1] for p in predictions]
return predictions
def training_generator(graph, starting_index=0):
"""
Yields a tuple (image, true_mask). Image is a tensor of shape (1, tf_image_size, tf_image_size, 3) and true_mask is a tensor of shape (1, image_size, image_size)\n
Due to the way generators work, it is required to specify the graph to work on
"""
images_list = images_list_root[starting_index:]
for (filepath, filename) in images_list:
with graph.as_default():
dicom_data = get_dicom_data(filepath)
image = format_pixel_array_for_tf(dicom_data.pixel_array)
true_mask = get_true_mask(filename)
true_mask = tf.convert_to_tensor(true_mask, dtype=tf.float32)
true_mask = tf.reshape(true_mask, (1, image_size, image_size, 1))
yield ([image], [true_mask])
def show_data(folder, index):
images_list = get_all_images_list(folder)
# Check index is valid
if index >= len(images_list):
print("Index %s out of range. Max index is %s" % (index, len(images_list) - 1))
exit(-1)
file_path, filename = images_list[index]
dicom_data = get_dicom_data(file_path)
# Display the data and image through matplotlib
plt.imshow(dicom_data.pixel_array)
print("File: %s" % filename)
print(dicom_data)
plt.show()
def training_generator(graph,
data_augment_technique="none",
starting_index=0):
"""
Yields a tuple (image, is_there_pneumothorax). Image is a tensor of shape (1, tf_image_size, tf_image_size, 3) and true_mask is a tensor of shape (1, 2)\n
Due to the way generators work, it is required to specify the graph to work on\n
data_augment_technique is the name of the technique used by data_augmentation (see data_augment.py -> random_data_augment)
"""
images_list = images_list_root[starting_index:]
for (filepath, filename) in images_list:
with graph.as_default():
dicom_data = get_dicom_data(filepath)
image = dicom_data.pixel_array
image = format_pixel_array_for_tf(
image, apply_data_augment_technique=data_augment_technique)
is_there_pneumothorax = get_image_label(filename)
is_there_pneumothorax = [[
1 - is_there_pneumothorax, is_there_pneumothorax
]]
is_there_pneumothorax = tf.convert_to_tensor(
is_there_pneumothorax, dtype=tf.float32)
yield ([image], [is_there_pneumothorax])