def ori_image_prepare(image_path, view, horizontal_flip, parameters):
"""
Loads an image and creates stride_lists
"""
patch_size = parameters['patch_size']
more_patches = parameters['more_patches']
stride_fixed = parameters['stride_fixed']
image = loading.load_image(image_path, view, horizontal_flip)
image = image.astype(float)
loading.standard_normalize_single_image(image)
img_width, img_length = image.shape
width_stride_list = stride_list_generator(img_width, patch_size, more_patches, stride_fixed)
length_stride_list = stride_list_generator(img_length, patch_size, more_patches, stride_fixed)
return image, width_stride_list, length_stride_list
def ori_image_prepare(short_file_path, view, horizontal_flip, parameters):
"""
Loads an image and creates stride_lists
"""
orginal_image_path = parameters['orginal_image_path']
patch_size = parameters['patch_size']
more_patches = parameters['more_patches']
stride_fixed = parameters['stride_fixed']
image_extension = '.hdf5' if parameters['use_hdf5'] else '.png'
image_path = os.path.join(orginal_image_path, short_file_path + image_extension)
image = loading.load_image(image_path, view, horizontal_flip)
image = image.astype(float)
loading.standard_normalize_single_image(image)
img_width, img_length = image.shape
width_stride_list = stride_list_generator(img_width, patch_size, more_patches, stride_fixed)
length_stride_list = stride_list_generator(img_length, patch_size, more_patches, stride_fixed)
return image, width_stride_list, length_stride_list
def run_model(model, exam_list, parameters, turn_on_visualization):
"""
Run the model over images in sample_data.
Save the predictions as csv and visualizations as png.
"""
if (parameters["device_type"] == "gpu") and torch.has_cudnn:
device = torch.device("cuda:{}".format(
parameters["gpu_number"]))
else:
device = torch.device("cpu")
model = model.to(device)
model.eval()
# initialize data holders
pred_dict = {
"image_index": [],
"benign_pred": [],
"malignant_pred": [],
"benign_label": 'no input',
"malignant_label": 'no input'
}
with torch.no_grad():
# iterate through each exam
for datum in tqdm.tqdm(exam_list):
for view in VIEWS.LIST:
short_file_path = datum[view][0]
# load image
# the image is already flipped so no need to do it again
loaded_image = loading.load_image(
image_path=os.path.join(
parameters["image_path"],
short_file_path + ".png"),
view=view,
horizontal_flip=False,
)
loading.standard_normalize_single_image(
loaded_image)
# load segmentation if available
benign_seg_path = os.path.join(
parameters["segmentation_path"],
"{0}_{1}".format(short_file_path,
"benign.png"))
malignant_seg_path = os.path.join(
parameters["segmentation_path"],
"{0}_{1}".format(short_file_path,
"malignant.png"))
benign_seg = np.zeros([1920, 2944], dtype=int)
#benign_seg = None
malignant_seg = np.zeros([1920, 2944], dtype=int)
#malignant_seg = None
if os.path.exists(benign_seg_path):
loaded_seg = loading.load_image(
image_path=benign_seg_path,
view=view,
horizontal_flip=False,
)
benign_seg = loaded_seg
if os.path.exists(malignant_seg_path):
loaded_seg = loading.load_image(
image_path=malignant_seg_path,
view=view,
horizontal_flip=False,
)
malignant_seg = loaded_seg
# convert python 2D array into 4D torch tensor in N,C,H,W format
loaded_image = np.expand_dims(
np.expand_dims(loaded_image, 0), 0).copy()
tensor_batch = torch.Tensor(loaded_image).to(
device)
# forward propagation
output = model(tensor_batch)
pred_numpy = output.data.cpu().numpy()
benign_pred, malignant_pred = pred_numpy[
0, 0], pred_numpy[0, 1]
# save visualization
if turn_on_visualization:
saliency_maps = model.saliency_map.data.cpu(
).numpy()
patch_locations = model.patch_locations
patch_imgs = model.patches
patch_attentions = model.patch_attns[
0, :].data.cpu().numpy()
save_dir = os.path.join(
parameters["output_path"], "visualization",
"{0}.png".format(short_file_path))
#b = np.zeros([2560 , 3328], dtype = int)
visualize_example(loaded_image, saliency_maps,
[benign_seg, malignant_seg],
patch_locations, patch_imgs,
patch_attentions, save_dir,
parameters)
# propagate holders
#benign_label, malignant_label = fetch_cancer_label_by_view(view, datum["cancer_label"])
pred_dict["image_index"].append(short_file_path)
pred_dict["benign_pred"].append(benign_pred)
pred_dict["malignant_pred"].append(malignant_pred)
#pred_dict["benign_label"].append(benign_label)
#pred_dict["malignant_label"].append(malignant_label)
return pd.DataFrame(pred_dict)