def test_create_possibly_padded_sample_for_cropping(crop_size: Any) -> None:
image_size = [4] * 3
image = np.random.uniform(size=[1] + image_size)
labels = np.zeros(shape=[2] + image_size)
mask = np.zeros(shape=image_size, dtype=ImageDataType.MASK.value)
cropped_sample = CroppingDataset.create_possibly_padded_sample_for_cropping(
sample=Sample(image=image, labels=labels, mask=mask, metadata=DummyPatientMetadata),
crop_size=crop_size,
padding_mode=PaddingMode.Zero
)
assert cropped_sample.image.shape[-3:] == crop_size
assert cropped_sample.labels.shape[-3:] == crop_size
assert cropped_sample.mask.shape[-3:] == crop_size
def visualize_random_crops(sample: Sample, config: SegmentationModelBase,
output_folder: Path) -> np.ndarray:
"""
Simulate the effect of sampling random crops (as is done for trainig segmentation models), and store the results
as a Nifti heatmap and as 3 axial/sagittal/coronal slices. The heatmap and the slices are stored in the given
output folder, with filenames that contain the patient ID as the prefix.
:param sample: The patient information from the dataset, with scans and ground truth labels.
:param config: The model configuration.
:param output_folder: The folder into which the heatmap and thumbnails should be written.
:return: A numpy array that has the same size as the image, containing how often each voxel was contained in
"""
output_folder.mkdir(exist_ok=True, parents=True)
sample = CroppingDataset.create_possibly_padded_sample_for_cropping(
sample=sample,
crop_size=config.crop_size,
padding_mode=config.padding_mode)
logging.info(f"Processing sample: {sample.patient_id}")
# Exhaustively sample with random crop function
image_channel0 = sample.image[0]
heatmap = np.zeros(image_channel0.shape, dtype=np.uint16)
# Number of repeats should fit into the range of UInt16, because we will later save the heatmap as an integer
# Nifti file of that datatype.
repeats = 200
for _ in range(repeats):
slicers, _ = augmentation.slicers_for_random_crop(
sample=sample,
crop_size=config.crop_size,
class_weights=config.class_weights)
heatmap[slicers[0], slicers[1], slicers[2]] += 1
is_3dim = heatmap.shape[0] > 1
header = sample.metadata.image_header
if not header:
logging.warning(
f"No image header found for patient {sample.patient_id}. Using default header."
)
header = get_unit_image_header()
if is_3dim:
ct_output_name = str(output_folder / f"{sample.patient_id}_ct.nii.gz")
heatmap_output_name = str(
output_folder / f"{sample.patient_id}_sampled_patches.nii.gz")
io_util.store_as_nifti(image=heatmap,
header=header,
file_name=heatmap_output_name,
image_type=heatmap.dtype,
scale=False)
io_util.store_as_nifti(image=image_channel0,
header=header,
file_name=ct_output_name,
image_type=sample.image.dtype,
scale=False)
heatmap_scaled = heatmap.astype(dtype=np.float) / heatmap.max()
# If the incoming image is effectively a 2D image with degenerate Z dimension, then only plot a single
# axial thumbnail. Otherwise, plot thumbnails for all 3 dimensions.
dimensions = list(range(3)) if is_3dim else [0]
# Center the 3 thumbnails at one of the points where the heatmap attains a maximum. This should ensure that
# the thumbnails are in an area where many of the organs of interest are located.
max_heatmap_index = np.unravel_index(
heatmap.argmax(), heatmap.shape) if is_3dim else (0, 0, 0)
for dimension in dimensions:
plt.clf()
scan_with_transparent_overlay(
scan=image_channel0,
overlay=heatmap_scaled,
dimension=dimension,
position=max_heatmap_index[dimension] if is_3dim else 0,
spacing=header.spacing)
# Construct a filename that has a dimension suffix if we are generating 3 of them. For 2dim images, skip
# the suffix.
thumbnail = f"{sample.patient_id}_sampled_patches"
if is_3dim:
thumbnail += f"_dim{dimension}"
thumbnail += ".png"
resize_and_save(width_inch=5,
height_inch=5,
filename=output_folder / thumbnail)
return heatmap
def main(args: CheckPatchSamplingConfig) -> None:
# Identify paths to inputs and outputs
commandline_args = {
"train_batch_size": 1,
"local_dataset": Path(args.local_dataset)
}
output_folder = Path(args.output_folder)
output_folder.mkdir(parents=True, exist_ok=True)
# Create a config file
config = ModelConfigLoader[SegmentationModelBase](
).create_model_config_from_name(args.model_name,
overrides=commandline_args)
# Set a random seed
ml_util.set_random_seed(config.random_seed)
# Get a dataloader object that checks csv
dataset_splits = config.get_dataset_splits()
# Load a sample using the full image data loader
full_image_dataset = FullImageDataset(config, dataset_splits.train)
for sample_index in range(args.number_samples):
sample = CroppingDataset.create_possibly_padded_sample_for_cropping(
sample=full_image_dataset.get_samples_at_index(
index=sample_index)[0],
crop_size=config.crop_size,
padding_mode=config.padding_mode)
print("Processing sample: ", sample.patient_id)
# Exhaustively sample with random crop function
heatmap = np.zeros(sample.mask.shape, dtype=np.uint16)
for _ in range(args.number_crop_iterations):
cropped_sample, center_point = augmentation.random_crop(
sample=sample,
crop_size=config.crop_size,
class_weights=config.class_weights)
patch_mask = create_mask_for_patch(output_shape=heatmap.shape,
output_dtype=heatmap.dtype,
center=center_point,
crop_size=config.crop_size)
heatmap += patch_mask
ct_output_name = str(output_folder /
"{}_ct.nii.gz".format(int(sample.patient_id)))
heatmap_output_name = str(
output_folder /
"{}_sampled_patches.nii.gz".format(int(sample.patient_id)))
if not sample.metadata.image_header:
raise ValueError("None header expected some header")
io_util.store_as_nifti(image=heatmap,
header=sample.metadata.image_header,
file_name=heatmap_output_name,
image_type=heatmap.dtype,
scale=False)
io_util.store_as_nifti(image=sample.image[0],
header=sample.metadata.image_header,
file_name=ct_output_name,
image_type=sample.image.dtype,
scale=False)